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"
69 #include "llvm/Analysis/IVUsers.h"
71 #include "llvm/Analysis/LoopInfo.h"
72 #include "llvm/Analysis/LoopPass.h"
78 #include "llvm/IR/BasicBlock.h"
79 #include "llvm/IR/Constant.h"
80 #include "llvm/IR/Constants.h"
81 #include "llvm/IR/DerivedTypes.h"
82 #include "llvm/IR/Dominators.h"
83 #include "llvm/IR/GlobalValue.h"
84 #include "llvm/IR/IRBuilder.h"
85 #include "llvm/IR/InstrTypes.h"
86 #include "llvm/IR/Instruction.h"
87 #include "llvm/IR/Instructions.h"
88 #include "llvm/IR/IntrinsicInst.h"
89 #include "llvm/IR/Intrinsics.h"
90 #include "llvm/IR/Module.h"
91 #include "llvm/IR/OperandTraits.h"
92 #include "llvm/IR/Operator.h"
93 #include "llvm/IR/PassManager.h"
94 #include "llvm/IR/Type.h"
95 #include "llvm/IR/Use.h"
96 #include "llvm/IR/User.h"
97 #include "llvm/IR/Value.h"
98 #include "llvm/IR/ValueHandle.h"
99 #include "llvm/Pass.h"
100 #include "llvm/Support/Casting.h"
102 #include "llvm/Support/Compiler.h"
103 #include "llvm/Support/Debug.h"
105 #include "llvm/Support/MathExtras.h"
107 #include "llvm/Transforms/Scalar.h"
110 #include <algorithm>
111 #include <cassert>
112 #include <cstddef>
113 #include <cstdint>
114 #include <cstdlib>
115 #include <iterator>
116 #include <limits>
117 #include <map>
118 #include <utility>
119 
120 using namespace llvm;
121 
122 #define DEBUG_TYPE "loop-reduce"
123 
124 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
125 /// bail out. This threshold is far beyond the number of users that LSR can
126 /// conceivably solve, so it should not affect generated code, but catches the
127 /// worst cases before LSR burns too much compile time and stack space.
128 static const unsigned MaxIVUsers = 200;
129 
130 // Temporary flag to cleanup congruent phis after LSR phi expansion.
131 // It's currently disabled until we can determine whether it's truly useful or
132 // not. The flag should be removed after the v3.0 release.
133 // This is now needed for ivchains.
135  "enable-lsr-phielim", cl::Hidden, cl::init(true),
136  cl::desc("Enable LSR phi elimination"));
137 
138 // The flag adds instruction count to solutions cost comparision.
139 static cl::opt<bool> InsnsCost(
140  "lsr-insns-cost", cl::Hidden, cl::init(true),
141  cl::desc("Add instruction count to a LSR cost model"));
142 
143 // Flag to choose how to narrow complex lsr solution
145  "lsr-exp-narrow", cl::Hidden, cl::init(false),
146  cl::desc("Narrow LSR complex solution using"
147  " expectation of registers number"));
148 
149 // Flag to narrow search space by filtering non-optimal formulae with
150 // the same ScaledReg and Scale.
152  "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
153  cl::desc("Narrow LSR search space by filtering non-optimal formulae"
154  " with the same ScaledReg and Scale"));
155 
156 #ifndef NDEBUG
157 // Stress test IV chain generation.
159  "stress-ivchain", cl::Hidden, cl::init(false),
160  cl::desc("Stress test LSR IV chains"));
161 #else
162 static bool StressIVChain = false;
163 #endif
164 
165 namespace {
166 
167 struct MemAccessTy {
168  /// Used in situations where the accessed memory type is unknown.
169  static const unsigned UnknownAddressSpace =
171 
172  Type *MemTy = nullptr;
173  unsigned AddrSpace = UnknownAddressSpace;
174 
175  MemAccessTy() = default;
176  MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
177 
178  bool operator==(MemAccessTy Other) const {
179  return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
180  }
181 
182  bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
183 
184  static MemAccessTy getUnknown(LLVMContext &Ctx,
185  unsigned AS = UnknownAddressSpace) {
186  return MemAccessTy(Type::getVoidTy(Ctx), AS);
187  }
188 };
189 
190 /// This class holds data which is used to order reuse candidates.
191 class RegSortData {
192 public:
193  /// This represents the set of LSRUse indices which reference
194  /// a particular register.
195  SmallBitVector UsedByIndices;
196 
197  void print(raw_ostream &OS) const;
198  void dump() const;
199 };
200 
201 } // end anonymous namespace
202 
203 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
204 void RegSortData::print(raw_ostream &OS) const {
205  OS << "[NumUses=" << UsedByIndices.count() << ']';
206 }
207 
208 LLVM_DUMP_METHOD void RegSortData::dump() const {
209  print(errs()); errs() << '\n';
210 }
211 #endif
212 
213 namespace {
214 
215 /// Map register candidates to information about how they are used.
216 class RegUseTracker {
217  using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
218 
219  RegUsesTy RegUsesMap;
221 
222 public:
223  void countRegister(const SCEV *Reg, size_t LUIdx);
224  void dropRegister(const SCEV *Reg, size_t LUIdx);
225  void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
226 
227  bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
228 
229  const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
230 
231  void clear();
232 
234  using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
235 
236  iterator begin() { return RegSequence.begin(); }
237  iterator end() { return RegSequence.end(); }
238  const_iterator begin() const { return RegSequence.begin(); }
239  const_iterator end() const { return RegSequence.end(); }
240 };
241 
242 } // end anonymous namespace
243 
244 void
245 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
246  std::pair<RegUsesTy::iterator, bool> Pair =
247  RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
248  RegSortData &RSD = Pair.first->second;
249  if (Pair.second)
250  RegSequence.push_back(Reg);
251  RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
252  RSD.UsedByIndices.set(LUIdx);
253 }
254 
255 void
256 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
257  RegUsesTy::iterator It = RegUsesMap.find(Reg);
258  assert(It != RegUsesMap.end());
259  RegSortData &RSD = It->second;
260  assert(RSD.UsedByIndices.size() > LUIdx);
261  RSD.UsedByIndices.reset(LUIdx);
262 }
263 
264 void
265 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
266  assert(LUIdx <= LastLUIdx);
267 
268  // Update RegUses. The data structure is not optimized for this purpose;
269  // we must iterate through it and update each of the bit vectors.
270  for (auto &Pair : RegUsesMap) {
271  SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
272  if (LUIdx < UsedByIndices.size())
273  UsedByIndices[LUIdx] =
274  LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
275  UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
276  }
277 }
278 
279 bool
280 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
281  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
282  if (I == RegUsesMap.end())
283  return false;
284  const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
285  int i = UsedByIndices.find_first();
286  if (i == -1) return false;
287  if ((size_t)i != LUIdx) return true;
288  return UsedByIndices.find_next(i) != -1;
289 }
290 
291 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
292  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
293  assert(I != RegUsesMap.end() && "Unknown register!");
294  return I->second.UsedByIndices;
295 }
296 
297 void RegUseTracker::clear() {
298  RegUsesMap.clear();
299  RegSequence.clear();
300 }
301 
302 namespace {
303 
304 /// This class holds information that describes a formula for computing
305 /// satisfying a use. It may include broken-out immediates and scaled registers.
306 struct Formula {
307  /// Global base address used for complex addressing.
308  GlobalValue *BaseGV = nullptr;
309 
310  /// Base offset for complex addressing.
311  int64_t BaseOffset = 0;
312 
313  /// Whether any complex addressing has a base register.
314  bool HasBaseReg = false;
315 
316  /// The scale of any complex addressing.
317  int64_t Scale = 0;
318 
319  /// The list of "base" registers for this use. When this is non-empty. The
320  /// canonical representation of a formula is
321  /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
322  /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
323  /// 3. The reg containing recurrent expr related with currect loop in the
324  /// formula should be put in the ScaledReg.
325  /// #1 enforces that the scaled register is always used when at least two
326  /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
327  /// #2 enforces that 1 * reg is reg.
328  /// #3 ensures invariant regs with respect to current loop can be combined
329  /// together in LSR codegen.
330  /// This invariant can be temporarly broken while building a formula.
331  /// However, every formula inserted into the LSRInstance must be in canonical
332  /// form.
334 
335  /// The 'scaled' register for this use. This should be non-null when Scale is
336  /// not zero.
337  const SCEV *ScaledReg = nullptr;
338 
339  /// An additional constant offset which added near the use. This requires a
340  /// temporary register, but the offset itself can live in an add immediate
341  /// field rather than a register.
342  int64_t UnfoldedOffset = 0;
343 
344  Formula() = default;
345 
346  void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
347 
348  bool isCanonical(const Loop &L) const;
349 
350  void canonicalize(const Loop &L);
351 
352  bool unscale();
353 
354  bool hasZeroEnd() const;
355 
356  size_t getNumRegs() const;
357  Type *getType() const;
358 
359  void deleteBaseReg(const SCEV *&S);
360 
361  bool referencesReg(const SCEV *S) const;
362  bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
363  const RegUseTracker &RegUses) const;
364 
365  void print(raw_ostream &OS) const;
366  void dump() const;
367 };
368 
369 } // end anonymous namespace
370 
371 /// Recursion helper for initialMatch.
372 static void DoInitialMatch(const SCEV *S, Loop *L,
375  ScalarEvolution &SE) {
376  // Collect expressions which properly dominate the loop header.
377  if (SE.properlyDominates(S, L->getHeader())) {
378  Good.push_back(S);
379  return;
380  }
381 
382  // Look at add operands.
383  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
384  for (const SCEV *S : Add->operands())
385  DoInitialMatch(S, L, Good, Bad, SE);
386  return;
387  }
388 
389  // Look at addrec operands.
390  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
391  if (!AR->getStart()->isZero() && AR->isAffine()) {
392  DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
393  DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
394  AR->getStepRecurrence(SE),
395  // FIXME: AR->getNoWrapFlags()
396  AR->getLoop(), SCEV::FlagAnyWrap),
397  L, Good, Bad, SE);
398  return;
399  }
400 
401  // Handle a multiplication by -1 (negation) if it didn't fold.
402  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
403  if (Mul->getOperand(0)->isAllOnesValue()) {
404  SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
405  const SCEV *NewMul = SE.getMulExpr(Ops);
406 
409  DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
410  const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
411  SE.getEffectiveSCEVType(NewMul->getType())));
412  for (const SCEV *S : MyGood)
413  Good.push_back(SE.getMulExpr(NegOne, S));
414  for (const SCEV *S : MyBad)
415  Bad.push_back(SE.getMulExpr(NegOne, S));
416  return;
417  }
418 
419  // Ok, we can't do anything interesting. Just stuff the whole thing into a
420  // register and hope for the best.
421  Bad.push_back(S);
422 }
423 
424 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
425 /// all loop-invariant and loop-computable values in a single base register.
426 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
429  DoInitialMatch(S, L, Good, Bad, SE);
430  if (!Good.empty()) {
431  const SCEV *Sum = SE.getAddExpr(Good);
432  if (!Sum->isZero())
433  BaseRegs.push_back(Sum);
434  HasBaseReg = true;
435  }
436  if (!Bad.empty()) {
437  const SCEV *Sum = SE.getAddExpr(Bad);
438  if (!Sum->isZero())
439  BaseRegs.push_back(Sum);
440  HasBaseReg = true;
441  }
442  canonicalize(*L);
443 }
444 
445 /// \brief Check whether or not this formula statisfies the canonical
446 /// representation.
447 /// \see Formula::BaseRegs.
448 bool Formula::isCanonical(const Loop &L) const {
449  if (!ScaledReg)
450  return BaseRegs.size() <= 1;
451 
452  if (Scale != 1)
453  return true;
454 
455  if (Scale == 1 && BaseRegs.empty())
456  return false;
457 
458  const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
459  if (SAR && SAR->getLoop() == &L)
460  return true;
461 
462  // If ScaledReg is not a recurrent expr, or it is but its loop is not current
463  // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
464  // loop, we want to swap the reg in BaseRegs with ScaledReg.
465  auto I =
466  find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
467  return isa<const SCEVAddRecExpr>(S) &&
468  (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
469  });
470  return I == BaseRegs.end();
471 }
472 
473 /// \brief Helper method to morph a formula into its canonical representation.
474 /// \see Formula::BaseRegs.
475 /// Every formula having more than one base register, must use the ScaledReg
476 /// field. Otherwise, we would have to do special cases everywhere in LSR
477 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
478 /// On the other hand, 1*reg should be canonicalized into reg.
479 void Formula::canonicalize(const Loop &L) {
480  if (isCanonical(L))
481  return;
482  // So far we did not need this case. This is easy to implement but it is
483  // useless to maintain dead code. Beside it could hurt compile time.
484  assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
485 
486  // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
487  if (!ScaledReg) {
488  ScaledReg = BaseRegs.back();
489  BaseRegs.pop_back();
490  Scale = 1;
491  }
492 
493  // If ScaledReg is an invariant with respect to L, find the reg from
494  // BaseRegs containing the recurrent expr related with Loop L. Swap the
495  // reg with ScaledReg.
496  const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
497  if (!SAR || SAR->getLoop() != &L) {
498  auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
499  [&](const SCEV *S) {
500  return isa<const SCEVAddRecExpr>(S) &&
501  (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
502  });
503  if (I != BaseRegs.end())
504  std::swap(ScaledReg, *I);
505  }
506 }
507 
508 /// \brief Get rid of the scale in the formula.
509 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
510 /// \return true if it was possible to get rid of the scale, false otherwise.
511 /// \note After this operation the formula may not be in the canonical form.
512 bool Formula::unscale() {
513  if (Scale != 1)
514  return false;
515  Scale = 0;
516  BaseRegs.push_back(ScaledReg);
517  ScaledReg = nullptr;
518  return true;
519 }
520 
521 bool Formula::hasZeroEnd() const {
522  if (UnfoldedOffset || BaseOffset)
523  return false;
524  if (BaseRegs.size() != 1 || ScaledReg)
525  return false;
526  return true;
527 }
528 
529 /// Return the total number of register operands used by this formula. This does
530 /// not include register uses implied by non-constant addrec strides.
531 size_t Formula::getNumRegs() const {
532  return !!ScaledReg + BaseRegs.size();
533 }
534 
535 /// Return the type of this formula, if it has one, or null otherwise. This type
536 /// is meaningless except for the bit size.
537 Type *Formula::getType() const {
538  return !BaseRegs.empty() ? BaseRegs.front()->getType() :
539  ScaledReg ? ScaledReg->getType() :
540  BaseGV ? BaseGV->getType() :
541  nullptr;
542 }
543 
544 /// Delete the given base reg from the BaseRegs list.
545 void Formula::deleteBaseReg(const SCEV *&S) {
546  if (&S != &BaseRegs.back())
547  std::swap(S, BaseRegs.back());
548  BaseRegs.pop_back();
549 }
550 
551 /// Test if this formula references the given register.
552 bool Formula::referencesReg(const SCEV *S) const {
553  return S == ScaledReg || is_contained(BaseRegs, S);
554 }
555 
556 /// Test whether this formula uses registers which are used by uses other than
557 /// the use with the given index.
558 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
559  const RegUseTracker &RegUses) const {
560  if (ScaledReg)
561  if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
562  return true;
563  for (const SCEV *BaseReg : BaseRegs)
564  if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
565  return true;
566  return false;
567 }
568 
569 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
570 void Formula::print(raw_ostream &OS) const {
571  bool First = true;
572  if (BaseGV) {
573  if (!First) OS << " + "; else First = false;
574  BaseGV->printAsOperand(OS, /*PrintType=*/false);
575  }
576  if (BaseOffset != 0) {
577  if (!First) OS << " + "; else First = false;
578  OS << BaseOffset;
579  }
580  for (const SCEV *BaseReg : BaseRegs) {
581  if (!First) OS << " + "; else First = false;
582  OS << "reg(" << *BaseReg << ')';
583  }
584  if (HasBaseReg && BaseRegs.empty()) {
585  if (!First) OS << " + "; else First = false;
586  OS << "**error: HasBaseReg**";
587  } else if (!HasBaseReg && !BaseRegs.empty()) {
588  if (!First) OS << " + "; else First = false;
589  OS << "**error: !HasBaseReg**";
590  }
591  if (Scale != 0) {
592  if (!First) OS << " + "; else First = false;
593  OS << Scale << "*reg(";
594  if (ScaledReg)
595  OS << *ScaledReg;
596  else
597  OS << "<unknown>";
598  OS << ')';
599  }
600  if (UnfoldedOffset != 0) {
601  if (!First) OS << " + ";
602  OS << "imm(" << UnfoldedOffset << ')';
603  }
604 }
605 
606 LLVM_DUMP_METHOD void Formula::dump() const {
607  print(errs()); errs() << '\n';
608 }
609 #endif
610 
611 /// Return true if the given addrec can be sign-extended without changing its
612 /// value.
613 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
614  Type *WideTy =
616  return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
617 }
618 
619 /// Return true if the given add can be sign-extended without changing its
620 /// value.
621 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
622  Type *WideTy =
624  return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
625 }
626 
627 /// Return true if the given mul can be sign-extended without changing its
628 /// value.
629 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
630  Type *WideTy =
632  SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
633  return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
634 }
635 
636 /// Return an expression for LHS /s RHS, if it can be determined and if the
637 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
638 /// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
639 /// the multiplication may overflow, which is useful when the result will be
640 /// used in a context where the most significant bits are ignored.
641 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
642  ScalarEvolution &SE,
643  bool IgnoreSignificantBits = false) {
644  // Handle the trivial case, which works for any SCEV type.
645  if (LHS == RHS)
646  return SE.getConstant(LHS->getType(), 1);
647 
648  // Handle a few RHS special cases.
649  const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
650  if (RC) {
651  const APInt &RA = RC->getAPInt();
652  // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
653  // some folding.
654  if (RA.isAllOnesValue())
655  return SE.getMulExpr(LHS, RC);
656  // Handle x /s 1 as x.
657  if (RA == 1)
658  return LHS;
659  }
660 
661  // Check for a division of a constant by a constant.
662  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
663  if (!RC)
664  return nullptr;
665  const APInt &LA = C->getAPInt();
666  const APInt &RA = RC->getAPInt();
667  if (LA.srem(RA) != 0)
668  return nullptr;
669  return SE.getConstant(LA.sdiv(RA));
670  }
671 
672  // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
673  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
674  if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
675  const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
676  IgnoreSignificantBits);
677  if (!Step) return nullptr;
678  const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
679  IgnoreSignificantBits);
680  if (!Start) return nullptr;
681  // FlagNW is independent of the start value, step direction, and is
682  // preserved with smaller magnitude steps.
683  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
684  return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
685  }
686  return nullptr;
687  }
688 
689  // Distribute the sdiv over add operands, if the add doesn't overflow.
690  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
691  if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
693  for (const SCEV *S : Add->operands()) {
694  const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
695  if (!Op) return nullptr;
696  Ops.push_back(Op);
697  }
698  return SE.getAddExpr(Ops);
699  }
700  return nullptr;
701  }
702 
703  // Check for a multiply operand that we can pull RHS out of.
704  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
705  if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
707  bool Found = false;
708  for (const SCEV *S : Mul->operands()) {
709  if (!Found)
710  if (const SCEV *Q = getExactSDiv(S, RHS, SE,
711  IgnoreSignificantBits)) {
712  S = Q;
713  Found = true;
714  }
715  Ops.push_back(S);
716  }
717  return Found ? SE.getMulExpr(Ops) : nullptr;
718  }
719  return nullptr;
720  }
721 
722  // Otherwise we don't know.
723  return nullptr;
724 }
725 
726 /// If S involves the addition of a constant integer value, return that integer
727 /// value, and mutate S to point to a new SCEV with that value excluded.
728 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
729  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
730  if (C->getAPInt().getMinSignedBits() <= 64) {
731  S = SE.getConstant(C->getType(), 0);
732  return C->getValue()->getSExtValue();
733  }
734  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
735  SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
736  int64_t Result = ExtractImmediate(NewOps.front(), SE);
737  if (Result != 0)
738  S = SE.getAddExpr(NewOps);
739  return Result;
740  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
741  SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
742  int64_t Result = ExtractImmediate(NewOps.front(), SE);
743  if (Result != 0)
744  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
745  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
747  return Result;
748  }
749  return 0;
750 }
751 
752 /// If S involves the addition of a GlobalValue address, return that symbol, and
753 /// mutate S to point to a new SCEV with that value excluded.
755  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
756  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
757  S = SE.getConstant(GV->getType(), 0);
758  return GV;
759  }
760  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
761  SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
762  GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
763  if (Result)
764  S = SE.getAddExpr(NewOps);
765  return Result;
766  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
767  SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
768  GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
769  if (Result)
770  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
771  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
773  return Result;
774  }
775  return nullptr;
776 }
777 
778 /// Returns true if the specified instruction is using the specified value as an
779 /// address.
780 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
781  bool isAddress = isa<LoadInst>(Inst);
782  if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
783  if (SI->getPointerOperand() == OperandVal)
784  isAddress = true;
785  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
786  // Addressing modes can also be folded into prefetches and a variety
787  // of intrinsics.
788  switch (II->getIntrinsicID()) {
789  default: break;
790  case Intrinsic::memset:
791  case Intrinsic::prefetch:
792  if (II->getArgOperand(0) == OperandVal)
793  isAddress = true;
794  break;
795  case Intrinsic::memmove:
796  case Intrinsic::memcpy:
797  if (II->getArgOperand(0) == OperandVal ||
798  II->getArgOperand(1) == OperandVal)
799  isAddress = true;
800  break;
801  }
802  } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
803  if (RMW->getPointerOperand() == OperandVal)
804  isAddress = true;
805  } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
806  if (CmpX->getPointerOperand() == OperandVal)
807  isAddress = true;
808  }
809  return isAddress;
810 }
811 
812 /// Return the type of the memory being accessed.
813 static MemAccessTy getAccessType(const Instruction *Inst) {
814  MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
815  if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
816  AccessTy.MemTy = SI->getOperand(0)->getType();
817  AccessTy.AddrSpace = SI->getPointerAddressSpace();
818  } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
819  AccessTy.AddrSpace = LI->getPointerAddressSpace();
820  } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
821  AccessTy.AddrSpace = RMW->getPointerAddressSpace();
822  } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
823  AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
824  }
825 
826  // All pointers have the same requirements, so canonicalize them to an
827  // arbitrary pointer type to minimize variation.
828  if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
829  AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
830  PTy->getAddressSpace());
831 
832  return AccessTy;
833 }
834 
835 /// Return true if this AddRec is already a phi in its loop.
836 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
837  for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
838  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
839  if (SE.isSCEVable(PN->getType()) &&
840  (SE.getEffectiveSCEVType(PN->getType()) ==
841  SE.getEffectiveSCEVType(AR->getType())) &&
842  SE.getSCEV(PN) == AR)
843  return true;
844  }
845  return false;
846 }
847 
848 /// Check if expanding this expression is likely to incur significant cost. This
849 /// is tricky because SCEV doesn't track which expressions are actually computed
850 /// by the current IR.
851 ///
852 /// We currently allow expansion of IV increments that involve adds,
853 /// multiplication by constants, and AddRecs from existing phis.
854 ///
855 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
856 /// obvious multiple of the UDivExpr.
857 static bool isHighCostExpansion(const SCEV *S,
858  SmallPtrSetImpl<const SCEV*> &Processed,
859  ScalarEvolution &SE) {
860  // Zero/One operand expressions
861  switch (S->getSCEVType()) {
862  case scUnknown:
863  case scConstant:
864  return false;
865  case scTruncate:
866  return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
867  Processed, SE);
868  case scZeroExtend:
869  return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
870  Processed, SE);
871  case scSignExtend:
872  return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
873  Processed, SE);
874  }
875 
876  if (!Processed.insert(S).second)
877  return false;
878 
879  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
880  for (const SCEV *S : Add->operands()) {
881  if (isHighCostExpansion(S, Processed, SE))
882  return true;
883  }
884  return false;
885  }
886 
887  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
888  if (Mul->getNumOperands() == 2) {
889  // Multiplication by a constant is ok
890  if (isa<SCEVConstant>(Mul->getOperand(0)))
891  return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
892 
893  // If we have the value of one operand, check if an existing
894  // multiplication already generates this expression.
895  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
896  Value *UVal = U->getValue();
897  for (User *UR : UVal->users()) {
898  // If U is a constant, it may be used by a ConstantExpr.
899  Instruction *UI = dyn_cast<Instruction>(UR);
900  if (UI && UI->getOpcode() == Instruction::Mul &&
901  SE.isSCEVable(UI->getType())) {
902  return SE.getSCEV(UI) == Mul;
903  }
904  }
905  }
906  }
907  }
908 
909  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
910  if (isExistingPhi(AR, SE))
911  return false;
912  }
913 
914  // Fow now, consider any other type of expression (div/mul/min/max) high cost.
915  return true;
916 }
917 
918 /// If any of the instructions is the specified set are trivially dead, delete
919 /// them and see if this makes any of their operands subsequently dead.
920 static bool
922  bool Changed = false;
923 
924  while (!DeadInsts.empty()) {
925  Value *V = DeadInsts.pop_back_val();
926  Instruction *I = dyn_cast_or_null<Instruction>(V);
927 
928  if (!I || !isInstructionTriviallyDead(I))
929  continue;
930 
931  for (Use &O : I->operands())
932  if (Instruction *U = dyn_cast<Instruction>(O)) {
933  O = nullptr;
934  if (U->use_empty())
935  DeadInsts.emplace_back(U);
936  }
937 
938  I->eraseFromParent();
939  Changed = true;
940  }
941 
942  return Changed;
943 }
944 
945 namespace {
946 
947 class LSRUse;
948 
949 } // end anonymous namespace
950 
951 /// \brief Check if the addressing mode defined by \p F is completely
952 /// folded in \p LU at isel time.
953 /// This includes address-mode folding and special icmp tricks.
954 /// This function returns true if \p LU can accommodate what \p F
955 /// defines and up to 1 base + 1 scaled + offset.
956 /// In other words, if \p F has several base registers, this function may
957 /// still return true. Therefore, users still need to account for
958 /// additional base registers and/or unfolded offsets to derive an
959 /// accurate cost model.
960 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
961  const LSRUse &LU, const Formula &F);
962 
963 // Get the cost of the scaling factor used in F for LU.
964 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
965  const LSRUse &LU, const Formula &F,
966  const Loop &L);
967 
968 namespace {
969 
970 /// This class is used to measure and compare candidate formulae.
971 class Cost {
973 
974 public:
975  Cost() {
976  C.Insns = 0;
977  C.NumRegs = 0;
978  C.AddRecCost = 0;
979  C.NumIVMuls = 0;
980  C.NumBaseAdds = 0;
981  C.ImmCost = 0;
982  C.SetupCost = 0;
983  C.ScaleCost = 0;
984  }
985 
986  bool isLess(Cost &Other, const TargetTransformInfo &TTI);
987 
988  void Lose();
989 
990 #ifndef NDEBUG
991  // Once any of the metrics loses, they must all remain losers.
992  bool isValid() {
993  return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
994  | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
995  || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
996  & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
997  }
998 #endif
999 
1000  bool isLoser() {
1001  assert(isValid() && "invalid cost");
1002  return C.NumRegs == ~0u;
1003  }
1004 
1005  void RateFormula(const TargetTransformInfo &TTI,
1006  const Formula &F,
1008  const DenseSet<const SCEV *> &VisitedRegs,
1009  const Loop *L,
1010  ScalarEvolution &SE, DominatorTree &DT,
1011  const LSRUse &LU,
1012  SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1013 
1014  void print(raw_ostream &OS) const;
1015  void dump() const;
1016 
1017 private:
1018  void RateRegister(const SCEV *Reg,
1020  const Loop *L,
1021  ScalarEvolution &SE, DominatorTree &DT);
1022  void RatePrimaryRegister(const SCEV *Reg,
1024  const Loop *L,
1025  ScalarEvolution &SE, DominatorTree &DT,
1026  SmallPtrSetImpl<const SCEV *> *LoserRegs);
1027 };
1028 
1029 /// An operand value in an instruction which is to be replaced with some
1030 /// equivalent, possibly strength-reduced, replacement.
1031 struct LSRFixup {
1032  /// The instruction which will be updated.
1033  Instruction *UserInst = nullptr;
1034 
1035  /// The operand of the instruction which will be replaced. The operand may be
1036  /// used more than once; every instance will be replaced.
1037  Value *OperandValToReplace = nullptr;
1038 
1039  /// If this user is to use the post-incremented value of an induction
1040  /// variable, this set is non-empty and holds the loops associated with the
1041  /// induction variable.
1042  PostIncLoopSet PostIncLoops;
1043 
1044  /// A constant offset to be added to the LSRUse expression. This allows
1045  /// multiple fixups to share the same LSRUse with different offsets, for
1046  /// example in an unrolled loop.
1047  int64_t Offset = 0;
1048 
1049  LSRFixup() = default;
1050 
1051  bool isUseFullyOutsideLoop(const Loop *L) const;
1052 
1053  void print(raw_ostream &OS) const;
1054  void dump() const;
1055 };
1056 
1057 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1058 /// SmallVectors of const SCEV*.
1059 struct UniquifierDenseMapInfo {
1060  static SmallVector<const SCEV *, 4> getEmptyKey() {
1062  V.push_back(reinterpret_cast<const SCEV *>(-1));
1063  return V;
1064  }
1065 
1066  static SmallVector<const SCEV *, 4> getTombstoneKey() {
1068  V.push_back(reinterpret_cast<const SCEV *>(-2));
1069  return V;
1070  }
1071 
1072  static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1073  return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1074  }
1075 
1076  static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1077  const SmallVector<const SCEV *, 4> &RHS) {
1078  return LHS == RHS;
1079  }
1080 };
1081 
1082 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1083 /// as uses invented by LSR itself. It includes information about what kinds of
1084 /// things can be folded into the user, information about the user itself, and
1085 /// information about how the use may be satisfied. TODO: Represent multiple
1086 /// users of the same expression in common?
1087 class LSRUse {
1088  DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1089 
1090 public:
1091  /// An enum for a kind of use, indicating what types of scaled and immediate
1092  /// operands it might support.
1093  enum KindType {
1094  Basic, ///< A normal use, with no folding.
1095  Special, ///< A special case of basic, allowing -1 scales.
1096  Address, ///< An address use; folding according to TargetLowering
1097  ICmpZero ///< An equality icmp with both operands folded into one.
1098  // TODO: Add a generic icmp too?
1099  };
1100 
1101  using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1102 
1103  KindType Kind;
1104  MemAccessTy AccessTy;
1105 
1106  /// The list of operands which are to be replaced.
1108 
1109  /// Keep track of the min and max offsets of the fixups.
1110  int64_t MinOffset = std::numeric_limits<int64_t>::max();
1111  int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1112 
1113  /// This records whether all of the fixups using this LSRUse are outside of
1114  /// the loop, in which case some special-case heuristics may be used.
1115  bool AllFixupsOutsideLoop = true;
1116 
1117  /// RigidFormula is set to true to guarantee that this use will be associated
1118  /// with a single formula--the one that initially matched. Some SCEV
1119  /// expressions cannot be expanded. This allows LSR to consider the registers
1120  /// used by those expressions without the need to expand them later after
1121  /// changing the formula.
1122  bool RigidFormula = false;
1123 
1124  /// This records the widest use type for any fixup using this
1125  /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1126  /// fixup widths to be equivalent, because the narrower one may be relying on
1127  /// the implicit truncation to truncate away bogus bits.
1128  Type *WidestFixupType = nullptr;
1129 
1130  /// A list of ways to build a value that can satisfy this user. After the
1131  /// list is populated, one of these is selected heuristically and used to
1132  /// formulate a replacement for OperandValToReplace in UserInst.
1133  SmallVector<Formula, 12> Formulae;
1134 
1135  /// The set of register candidates used by all formulae in this LSRUse.
1137 
1138  LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1139 
1140  LSRFixup &getNewFixup() {
1141  Fixups.push_back(LSRFixup());
1142  return Fixups.back();
1143  }
1144 
1145  void pushFixup(LSRFixup &f) {
1146  Fixups.push_back(f);
1147  if (f.Offset > MaxOffset)
1148  MaxOffset = f.Offset;
1149  if (f.Offset < MinOffset)
1150  MinOffset = f.Offset;
1151  }
1152 
1153  bool HasFormulaWithSameRegs(const Formula &F) const;
1154  float getNotSelectedProbability(const SCEV *Reg) const;
1155  bool InsertFormula(const Formula &F, const Loop &L);
1156  void DeleteFormula(Formula &F);
1157  void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1158 
1159  void print(raw_ostream &OS) const;
1160  void dump() const;
1161 };
1162 
1163 } // end anonymous namespace
1164 
1165 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1166  LSRUse::KindType Kind, MemAccessTy AccessTy,
1167  GlobalValue *BaseGV, int64_t BaseOffset,
1168  bool HasBaseReg, int64_t Scale,
1169  Instruction *Fixup = nullptr);
1170 
1171 /// Tally up interesting quantities from the given register.
1172 void Cost::RateRegister(const SCEV *Reg,
1174  const Loop *L,
1175  ScalarEvolution &SE, DominatorTree &DT) {
1176  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1177  // If this is an addrec for another loop, it should be an invariant
1178  // with respect to L since L is the innermost loop (at least
1179  // for now LSR only handles innermost loops).
1180  if (AR->getLoop() != L) {
1181  // If the AddRec exists, consider it's register free and leave it alone.
1182  if (isExistingPhi(AR, SE))
1183  return;
1184 
1185  // It is bad to allow LSR for current loop to add induction variables
1186  // for its sibling loops.
1187  if (!AR->getLoop()->contains(L)) {
1188  Lose();
1189  return;
1190  }
1191 
1192  // Otherwise, it will be an invariant with respect to Loop L.
1193  ++C.NumRegs;
1194  return;
1195  }
1196  C.AddRecCost += 1; /// TODO: This should be a function of the stride.
1197 
1198  // Add the step value register, if it needs one.
1199  // TODO: The non-affine case isn't precisely modeled here.
1200  if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1201  if (!Regs.count(AR->getOperand(1))) {
1202  RateRegister(AR->getOperand(1), Regs, L, SE, DT);
1203  if (isLoser())
1204  return;
1205  }
1206  }
1207  }
1208  ++C.NumRegs;
1209 
1210  // Rough heuristic; favor registers which don't require extra setup
1211  // instructions in the preheader.
1212  if (!isa<SCEVUnknown>(Reg) &&
1213  !isa<SCEVConstant>(Reg) &&
1214  !(isa<SCEVAddRecExpr>(Reg) &&
1215  (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
1216  isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
1217  ++C.SetupCost;
1218 
1219  C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1220  SE.hasComputableLoopEvolution(Reg, L);
1221 }
1222 
1223 /// Record this register in the set. If we haven't seen it before, rate
1224 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1225 /// one of those regs an instant loser.
1226 void Cost::RatePrimaryRegister(const SCEV *Reg,
1228  const Loop *L,
1229  ScalarEvolution &SE, DominatorTree &DT,
1230  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1231  if (LoserRegs && LoserRegs->count(Reg)) {
1232  Lose();
1233  return;
1234  }
1235  if (Regs.insert(Reg).second) {
1236  RateRegister(Reg, Regs, L, SE, DT);
1237  if (LoserRegs && isLoser())
1238  LoserRegs->insert(Reg);
1239  }
1240 }
1241 
1242 void Cost::RateFormula(const TargetTransformInfo &TTI,
1243  const Formula &F,
1245  const DenseSet<const SCEV *> &VisitedRegs,
1246  const Loop *L,
1247  ScalarEvolution &SE, DominatorTree &DT,
1248  const LSRUse &LU,
1249  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1250  assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1251  // Tally up the registers.
1252  unsigned PrevAddRecCost = C.AddRecCost;
1253  unsigned PrevNumRegs = C.NumRegs;
1254  unsigned PrevNumBaseAdds = C.NumBaseAdds;
1255  if (const SCEV *ScaledReg = F.ScaledReg) {
1256  if (VisitedRegs.count(ScaledReg)) {
1257  Lose();
1258  return;
1259  }
1260  RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1261  if (isLoser())
1262  return;
1263  }
1264  for (const SCEV *BaseReg : F.BaseRegs) {
1265  if (VisitedRegs.count(BaseReg)) {
1266  Lose();
1267  return;
1268  }
1269  RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1270  if (isLoser())
1271  return;
1272  }
1273 
1274  // Determine how many (unfolded) adds we'll need inside the loop.
1275  size_t NumBaseParts = F.getNumRegs();
1276  if (NumBaseParts > 1)
1277  // Do not count the base and a possible second register if the target
1278  // allows to fold 2 registers.
1279  C.NumBaseAdds +=
1280  NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1281  C.NumBaseAdds += (F.UnfoldedOffset != 0);
1282 
1283  // Accumulate non-free scaling amounts.
1284  C.ScaleCost += getScalingFactorCost(TTI, LU, F, *L);
1285 
1286  // Tally up the non-zero immediates.
1287  for (const LSRFixup &Fixup : LU.Fixups) {
1288  int64_t O = Fixup.Offset;
1289  int64_t Offset = (uint64_t)O + F.BaseOffset;
1290  if (F.BaseGV)
1291  C.ImmCost += 64; // Handle symbolic values conservatively.
1292  // TODO: This should probably be the pointer size.
1293  else if (Offset != 0)
1294  C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1295 
1296  // Check with target if this offset with this instruction is
1297  // specifically not supported.
1298  if (LU.Kind == LSRUse::Address && Offset != 0 &&
1299  !isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1300  Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1301  C.NumBaseAdds++;
1302  }
1303 
1304  // If we don't count instruction cost exit here.
1305  if (!InsnsCost) {
1306  assert(isValid() && "invalid cost");
1307  return;
1308  }
1309 
1310  // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1311  // additional instruction (at least fill).
1312  unsigned TTIRegNum = TTI.getNumberOfRegisters(false) - 1;
1313  if (C.NumRegs > TTIRegNum) {
1314  // Cost already exceeded TTIRegNum, then only newly added register can add
1315  // new instructions.
1316  if (PrevNumRegs > TTIRegNum)
1317  C.Insns += (C.NumRegs - PrevNumRegs);
1318  else
1319  C.Insns += (C.NumRegs - TTIRegNum);
1320  }
1321 
1322  // If ICmpZero formula ends with not 0, it could not be replaced by
1323  // just add or sub. We'll need to compare final result of AddRec.
1324  // That means we'll need an additional instruction.
1325  // For -10 + {0, +, 1}:
1326  // i = i + 1;
1327  // cmp i, 10
1328  //
1329  // For {-10, +, 1}:
1330  // i = i + 1;
1331  if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd())
1332  C.Insns++;
1333  // Each new AddRec adds 1 instruction to calculation.
1334  C.Insns += (C.AddRecCost - PrevAddRecCost);
1335 
1336  // BaseAdds adds instructions for unfolded registers.
1337  if (LU.Kind != LSRUse::ICmpZero)
1338  C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1339  assert(isValid() && "invalid cost");
1340 }
1341 
1342 /// Set this cost to a losing value.
1343 void Cost::Lose() {
1346  C.AddRecCost = std::numeric_limits<unsigned>::max();
1347  C.NumIVMuls = std::numeric_limits<unsigned>::max();
1348  C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1350  C.SetupCost = std::numeric_limits<unsigned>::max();
1351  C.ScaleCost = std::numeric_limits<unsigned>::max();
1352 }
1353 
1354 /// Choose the lower cost.
1355 bool Cost::isLess(Cost &Other, const TargetTransformInfo &TTI) {
1356  if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1357  C.Insns != Other.C.Insns)
1358  return C.Insns < Other.C.Insns;
1359  return TTI.isLSRCostLess(C, Other.C);
1360 }
1361 
1362 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1363 void Cost::print(raw_ostream &OS) const {
1364  if (InsnsCost)
1365  OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1366  OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1367  if (C.AddRecCost != 0)
1368  OS << ", with addrec cost " << C.AddRecCost;
1369  if (C.NumIVMuls != 0)
1370  OS << ", plus " << C.NumIVMuls << " IV mul"
1371  << (C.NumIVMuls == 1 ? "" : "s");
1372  if (C.NumBaseAdds != 0)
1373  OS << ", plus " << C.NumBaseAdds << " base add"
1374  << (C.NumBaseAdds == 1 ? "" : "s");
1375  if (C.ScaleCost != 0)
1376  OS << ", plus " << C.ScaleCost << " scale cost";
1377  if (C.ImmCost != 0)
1378  OS << ", plus " << C.ImmCost << " imm cost";
1379  if (C.SetupCost != 0)
1380  OS << ", plus " << C.SetupCost << " setup cost";
1381 }
1382 
1383 LLVM_DUMP_METHOD void Cost::dump() const {
1384  print(errs()); errs() << '\n';
1385 }
1386 #endif
1387 
1388 /// Test whether this fixup always uses its value outside of the given loop.
1389 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1390  // PHI nodes use their value in their incoming blocks.
1391  if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1392  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1393  if (PN->getIncomingValue(i) == OperandValToReplace &&
1394  L->contains(PN->getIncomingBlock(i)))
1395  return false;
1396  return true;
1397  }
1398 
1399  return !L->contains(UserInst);
1400 }
1401 
1402 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1403 void LSRFixup::print(raw_ostream &OS) const {
1404  OS << "UserInst=";
1405  // Store is common and interesting enough to be worth special-casing.
1406  if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1407  OS << "store ";
1408  Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1409  } else if (UserInst->getType()->isVoidTy())
1410  OS << UserInst->getOpcodeName();
1411  else
1412  UserInst->printAsOperand(OS, /*PrintType=*/false);
1413 
1414  OS << ", OperandValToReplace=";
1415  OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1416 
1417  for (const Loop *PIL : PostIncLoops) {
1418  OS << ", PostIncLoop=";
1419  PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1420  }
1421 
1422  if (Offset != 0)
1423  OS << ", Offset=" << Offset;
1424 }
1425 
1426 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1427  print(errs()); errs() << '\n';
1428 }
1429 #endif
1430 
1431 /// Test whether this use as a formula which has the same registers as the given
1432 /// formula.
1433 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1434  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1435  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1436  // Unstable sort by host order ok, because this is only used for uniquifying.
1437  std::sort(Key.begin(), Key.end());
1438  return Uniquifier.count(Key);
1439 }
1440 
1441 /// The function returns a probability of selecting formula without Reg.
1442 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1443  unsigned FNum = 0;
1444  for (const Formula &F : Formulae)
1445  if (F.referencesReg(Reg))
1446  FNum++;
1447  return ((float)(Formulae.size() - FNum)) / Formulae.size();
1448 }
1449 
1450 /// If the given formula has not yet been inserted, add it to the list, and
1451 /// return true. Return false otherwise. The formula must be in canonical form.
1452 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1453  assert(F.isCanonical(L) && "Invalid canonical representation");
1454 
1455  if (!Formulae.empty() && RigidFormula)
1456  return false;
1457 
1458  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1459  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1460  // Unstable sort by host order ok, because this is only used for uniquifying.
1461  std::sort(Key.begin(), Key.end());
1462 
1463  if (!Uniquifier.insert(Key).second)
1464  return false;
1465 
1466  // Using a register to hold the value of 0 is not profitable.
1467  assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1468  "Zero allocated in a scaled register!");
1469 #ifndef NDEBUG
1470  for (const SCEV *BaseReg : F.BaseRegs)
1471  assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1472 #endif
1473 
1474  // Add the formula to the list.
1475  Formulae.push_back(F);
1476 
1477  // Record registers now being used by this use.
1478  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1479  if (F.ScaledReg)
1480  Regs.insert(F.ScaledReg);
1481 
1482  return true;
1483 }
1484 
1485 /// Remove the given formula from this use's list.
1486 void LSRUse::DeleteFormula(Formula &F) {
1487  if (&F != &Formulae.back())
1488  std::swap(F, Formulae.back());
1489  Formulae.pop_back();
1490 }
1491 
1492 /// Recompute the Regs field, and update RegUses.
1493 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1494  // Now that we've filtered out some formulae, recompute the Regs set.
1495  SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1496  Regs.clear();
1497  for (const Formula &F : Formulae) {
1498  if (F.ScaledReg) Regs.insert(F.ScaledReg);
1499  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1500  }
1501 
1502  // Update the RegTracker.
1503  for (const SCEV *S : OldRegs)
1504  if (!Regs.count(S))
1505  RegUses.dropRegister(S, LUIdx);
1506 }
1507 
1508 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1509 void LSRUse::print(raw_ostream &OS) const {
1510  OS << "LSR Use: Kind=";
1511  switch (Kind) {
1512  case Basic: OS << "Basic"; break;
1513  case Special: OS << "Special"; break;
1514  case ICmpZero: OS << "ICmpZero"; break;
1515  case Address:
1516  OS << "Address of ";
1517  if (AccessTy.MemTy->isPointerTy())
1518  OS << "pointer"; // the full pointer type could be really verbose
1519  else {
1520  OS << *AccessTy.MemTy;
1521  }
1522 
1523  OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1524  }
1525 
1526  OS << ", Offsets={";
1527  bool NeedComma = false;
1528  for (const LSRFixup &Fixup : Fixups) {
1529  if (NeedComma) OS << ',';
1530  OS << Fixup.Offset;
1531  NeedComma = true;
1532  }
1533  OS << '}';
1534 
1535  if (AllFixupsOutsideLoop)
1536  OS << ", all-fixups-outside-loop";
1537 
1538  if (WidestFixupType)
1539  OS << ", widest fixup type: " << *WidestFixupType;
1540 }
1541 
1542 LLVM_DUMP_METHOD void LSRUse::dump() const {
1543  print(errs()); errs() << '\n';
1544 }
1545 #endif
1546 
1548  LSRUse::KindType Kind, MemAccessTy AccessTy,
1549  GlobalValue *BaseGV, int64_t BaseOffset,
1550  bool HasBaseReg, int64_t Scale,
1551  Instruction *Fixup/*= nullptr*/) {
1552  switch (Kind) {
1553  case LSRUse::Address:
1554  return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1555  HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1556 
1557  case LSRUse::ICmpZero:
1558  // There's not even a target hook for querying whether it would be legal to
1559  // fold a GV into an ICmp.
1560  if (BaseGV)
1561  return false;
1562 
1563  // ICmp only has two operands; don't allow more than two non-trivial parts.
1564  if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1565  return false;
1566 
1567  // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1568  // putting the scaled register in the other operand of the icmp.
1569  if (Scale != 0 && Scale != -1)
1570  return false;
1571 
1572  // If we have low-level target information, ask the target if it can fold an
1573  // integer immediate on an icmp.
1574  if (BaseOffset != 0) {
1575  // We have one of:
1576  // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1577  // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1578  // Offs is the ICmp immediate.
1579  if (Scale == 0)
1580  // The cast does the right thing with
1581  // std::numeric_limits<int64_t>::min().
1582  BaseOffset = -(uint64_t)BaseOffset;
1583  return TTI.isLegalICmpImmediate(BaseOffset);
1584  }
1585 
1586  // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1587  return true;
1588 
1589  case LSRUse::Basic:
1590  // Only handle single-register values.
1591  return !BaseGV && Scale == 0 && BaseOffset == 0;
1592 
1593  case LSRUse::Special:
1594  // Special case Basic to handle -1 scales.
1595  return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1596  }
1597 
1598  llvm_unreachable("Invalid LSRUse Kind!");
1599 }
1600 
1602  int64_t MinOffset, int64_t MaxOffset,
1603  LSRUse::KindType Kind, MemAccessTy AccessTy,
1604  GlobalValue *BaseGV, int64_t BaseOffset,
1605  bool HasBaseReg, int64_t Scale) {
1606  // Check for overflow.
1607  if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1608  (MinOffset > 0))
1609  return false;
1610  MinOffset = (uint64_t)BaseOffset + MinOffset;
1611  if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1612  (MaxOffset > 0))
1613  return false;
1614  MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1615 
1616  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1617  HasBaseReg, Scale) &&
1618  isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1619  HasBaseReg, Scale);
1620 }
1621 
1623  int64_t MinOffset, int64_t MaxOffset,
1624  LSRUse::KindType Kind, MemAccessTy AccessTy,
1625  const Formula &F, const Loop &L) {
1626  // For the purpose of isAMCompletelyFolded either having a canonical formula
1627  // or a scale not equal to zero is correct.
1628  // Problems may arise from non canonical formulae having a scale == 0.
1629  // Strictly speaking it would best to just rely on canonical formulae.
1630  // However, when we generate the scaled formulae, we first check that the
1631  // scaling factor is profitable before computing the actual ScaledReg for
1632  // compile time sake.
1633  assert((F.isCanonical(L) || F.Scale != 0));
1634  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1635  F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1636 }
1637 
1638 /// Test whether we know how to expand the current formula.
1639 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1640  int64_t MaxOffset, LSRUse::KindType Kind,
1641  MemAccessTy AccessTy, GlobalValue *BaseGV,
1642  int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1643  // We know how to expand completely foldable formulae.
1644  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1645  BaseOffset, HasBaseReg, Scale) ||
1646  // Or formulae that use a base register produced by a sum of base
1647  // registers.
1648  (Scale == 1 &&
1649  isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1650  BaseGV, BaseOffset, true, 0));
1651 }
1652 
1653 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1654  int64_t MaxOffset, LSRUse::KindType Kind,
1655  MemAccessTy AccessTy, const Formula &F) {
1656  return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1657  F.BaseOffset, F.HasBaseReg, F.Scale);
1658 }
1659 
1661  const LSRUse &LU, const Formula &F) {
1662  // Target may want to look at the user instructions.
1663  if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1664  for (const LSRFixup &Fixup : LU.Fixups)
1665  if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1666  (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1667  F.Scale, Fixup.UserInst))
1668  return false;
1669  return true;
1670  }
1671 
1672  return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1673  LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1674  F.Scale);
1675 }
1676 
1677 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1678  const LSRUse &LU, const Formula &F,
1679  const Loop &L) {
1680  if (!F.Scale)
1681  return 0;
1682 
1683  // If the use is not completely folded in that instruction, we will have to
1684  // pay an extra cost only for scale != 1.
1685  if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1686  LU.AccessTy, F, L))
1687  return F.Scale != 1;
1688 
1689  switch (LU.Kind) {
1690  case LSRUse::Address: {
1691  // Check the scaling factor cost with both the min and max offsets.
1692  int ScaleCostMinOffset = TTI.getScalingFactorCost(
1693  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1694  F.Scale, LU.AccessTy.AddrSpace);
1695  int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1696  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1697  F.Scale, LU.AccessTy.AddrSpace);
1698 
1699  assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1700  "Legal addressing mode has an illegal cost!");
1701  return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1702  }
1703  case LSRUse::ICmpZero:
1704  case LSRUse::Basic:
1705  case LSRUse::Special:
1706  // The use is completely folded, i.e., everything is folded into the
1707  // instruction.
1708  return 0;
1709  }
1710 
1711  llvm_unreachable("Invalid LSRUse Kind!");
1712 }
1713 
1714 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1715  LSRUse::KindType Kind, MemAccessTy AccessTy,
1716  GlobalValue *BaseGV, int64_t BaseOffset,
1717  bool HasBaseReg) {
1718  // Fast-path: zero is always foldable.
1719  if (BaseOffset == 0 && !BaseGV) return true;
1720 
1721  // Conservatively, create an address with an immediate and a
1722  // base and a scale.
1723  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1724 
1725  // Canonicalize a scale of 1 to a base register if the formula doesn't
1726  // already have a base register.
1727  if (!HasBaseReg && Scale == 1) {
1728  Scale = 0;
1729  HasBaseReg = true;
1730  }
1731 
1732  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1733  HasBaseReg, Scale);
1734 }
1735 
1736 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1737  ScalarEvolution &SE, int64_t MinOffset,
1738  int64_t MaxOffset, LSRUse::KindType Kind,
1739  MemAccessTy AccessTy, const SCEV *S,
1740  bool HasBaseReg) {
1741  // Fast-path: zero is always foldable.
1742  if (S->isZero()) return true;
1743 
1744  // Conservatively, create an address with an immediate and a
1745  // base and a scale.
1746  int64_t BaseOffset = ExtractImmediate(S, SE);
1747  GlobalValue *BaseGV = ExtractSymbol(S, SE);
1748 
1749  // If there's anything else involved, it's not foldable.
1750  if (!S->isZero()) return false;
1751 
1752  // Fast-path: zero is always foldable.
1753  if (BaseOffset == 0 && !BaseGV) return true;
1754 
1755  // Conservatively, create an address with an immediate and a
1756  // base and a scale.
1757  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1758 
1759  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1760  BaseOffset, HasBaseReg, Scale);
1761 }
1762 
1763 namespace {
1764 
1765 /// An individual increment in a Chain of IV increments. Relate an IV user to
1766 /// an expression that computes the IV it uses from the IV used by the previous
1767 /// link in the Chain.
1768 ///
1769 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1770 /// original IVOperand. The head of the chain's IVOperand is only valid during
1771 /// chain collection, before LSR replaces IV users. During chain generation,
1772 /// IncExpr can be used to find the new IVOperand that computes the same
1773 /// expression.
1774 struct IVInc {
1775  Instruction *UserInst;
1776  Value* IVOperand;
1777  const SCEV *IncExpr;
1778 
1779  IVInc(Instruction *U, Value *O, const SCEV *E)
1780  : UserInst(U), IVOperand(O), IncExpr(E) {}
1781 };
1782 
1783 // The list of IV increments in program order. We typically add the head of a
1784 // chain without finding subsequent links.
1785 struct IVChain {
1786  SmallVector<IVInc, 1> Incs;
1787  const SCEV *ExprBase = nullptr;
1788 
1789  IVChain() = default;
1790  IVChain(const IVInc &Head, const SCEV *Base)
1791  : Incs(1, Head), ExprBase(Base) {}
1792 
1793  using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1794 
1795  // Return the first increment in the chain.
1796  const_iterator begin() const {
1797  assert(!Incs.empty());
1798  return std::next(Incs.begin());
1799  }
1800  const_iterator end() const {
1801  return Incs.end();
1802  }
1803 
1804  // Returns true if this chain contains any increments.
1805  bool hasIncs() const { return Incs.size() >= 2; }
1806 
1807  // Add an IVInc to the end of this chain.
1808  void add(const IVInc &X) { Incs.push_back(X); }
1809 
1810  // Returns the last UserInst in the chain.
1811  Instruction *tailUserInst() const { return Incs.back().UserInst; }
1812 
1813  // Returns true if IncExpr can be profitably added to this chain.
1814  bool isProfitableIncrement(const SCEV *OperExpr,
1815  const SCEV *IncExpr,
1816  ScalarEvolution&);
1817 };
1818 
1819 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1820 /// between FarUsers that definitely cross IV increments and NearUsers that may
1821 /// be used between IV increments.
1822 struct ChainUsers {
1824  SmallPtrSet<Instruction*, 4> NearUsers;
1825 };
1826 
1827 /// This class holds state for the main loop strength reduction logic.
1828 class LSRInstance {
1829  IVUsers &IU;
1830  ScalarEvolution &SE;
1831  DominatorTree &DT;
1832  LoopInfo &LI;
1833  const TargetTransformInfo &TTI;
1834  Loop *const L;
1835  bool Changed = false;
1836 
1837  /// This is the insert position that the current loop's induction variable
1838  /// increment should be placed. In simple loops, this is the latch block's
1839  /// terminator. But in more complicated cases, this is a position which will
1840  /// dominate all the in-loop post-increment users.
1841  Instruction *IVIncInsertPos = nullptr;
1842 
1843  /// Interesting factors between use strides.
1844  ///
1845  /// We explicitly use a SetVector which contains a SmallSet, instead of the
1846  /// default, a SmallDenseSet, because we need to use the full range of
1847  /// int64_ts, and there's currently no good way of doing that with
1848  /// SmallDenseSet.
1850 
1851  /// Interesting use types, to facilitate truncation reuse.
1853 
1854  /// The list of interesting uses.
1856 
1857  /// Track which uses use which register candidates.
1858  RegUseTracker RegUses;
1859 
1860  // Limit the number of chains to avoid quadratic behavior. We don't expect to
1861  // have more than a few IV increment chains in a loop. Missing a Chain falls
1862  // back to normal LSR behavior for those uses.
1863  static const unsigned MaxChains = 8;
1864 
1865  /// IV users can form a chain of IV increments.
1867 
1868  /// IV users that belong to profitable IVChains.
1870 
1871  void OptimizeShadowIV();
1872  bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1873  ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1874  void OptimizeLoopTermCond();
1875 
1876  void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1877  SmallVectorImpl<ChainUsers> &ChainUsersVec);
1878  void FinalizeChain(IVChain &Chain);
1879  void CollectChains();
1880  void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1881  SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1882 
1883  void CollectInterestingTypesAndFactors();
1884  void CollectFixupsAndInitialFormulae();
1885 
1886  // Support for sharing of LSRUses between LSRFixups.
1888  UseMapTy UseMap;
1889 
1890  bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1891  LSRUse::KindType Kind, MemAccessTy AccessTy);
1892 
1893  std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1894  MemAccessTy AccessTy);
1895 
1896  void DeleteUse(LSRUse &LU, size_t LUIdx);
1897 
1898  LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1899 
1900  void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1901  void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1902  void CountRegisters(const Formula &F, size_t LUIdx);
1903  bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1904 
1905  void CollectLoopInvariantFixupsAndFormulae();
1906 
1907  void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1908  unsigned Depth = 0);
1909 
1910  void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1911  const Formula &Base, unsigned Depth,
1912  size_t Idx, bool IsScaledReg = false);
1913  void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1914  void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1915  const Formula &Base, size_t Idx,
1916  bool IsScaledReg = false);
1917  void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1918  void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1919  const Formula &Base,
1920  const SmallVectorImpl<int64_t> &Worklist,
1921  size_t Idx, bool IsScaledReg = false);
1922  void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1923  void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1924  void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1925  void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1926  void GenerateCrossUseConstantOffsets();
1927  void GenerateAllReuseFormulae();
1928 
1929  void FilterOutUndesirableDedicatedRegisters();
1930 
1931  size_t EstimateSearchSpaceComplexity() const;
1932  void NarrowSearchSpaceByDetectingSupersets();
1933  void NarrowSearchSpaceByCollapsingUnrolledCode();
1934  void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1935  void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
1936  void NarrowSearchSpaceByDeletingCostlyFormulas();
1937  void NarrowSearchSpaceByPickingWinnerRegs();
1938  void NarrowSearchSpaceUsingHeuristics();
1939 
1940  void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1941  Cost &SolutionCost,
1943  const Cost &CurCost,
1944  const SmallPtrSet<const SCEV *, 16> &CurRegs,
1945  DenseSet<const SCEV *> &VisitedRegs) const;
1946  void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1947 
1949  HoistInsertPosition(BasicBlock::iterator IP,
1950  const SmallVectorImpl<Instruction *> &Inputs) const;
1952  AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1953  const LSRFixup &LF,
1954  const LSRUse &LU,
1955  SCEVExpander &Rewriter) const;
1956 
1957  Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1958  BasicBlock::iterator IP, SCEVExpander &Rewriter,
1959  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1960  void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
1961  const Formula &F, SCEVExpander &Rewriter,
1962  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1963  void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1964  SCEVExpander &Rewriter,
1965  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1966  void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
1967 
1968 public:
1969  LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
1970  LoopInfo &LI, const TargetTransformInfo &TTI);
1971 
1972  bool getChanged() const { return Changed; }
1973 
1974  void print_factors_and_types(raw_ostream &OS) const;
1975  void print_fixups(raw_ostream &OS) const;
1976  void print_uses(raw_ostream &OS) const;
1977  void print(raw_ostream &OS) const;
1978  void dump() const;
1979 };
1980 
1981 } // end anonymous namespace
1982 
1983 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
1984 /// the cast operation.
1985 void LSRInstance::OptimizeShadowIV() {
1986  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1987  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1988  return;
1989 
1990  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1991  UI != E; /* empty */) {
1992  IVUsers::const_iterator CandidateUI = UI;
1993  ++UI;
1994  Instruction *ShadowUse = CandidateUI->getUser();
1995  Type *DestTy = nullptr;
1996  bool IsSigned = false;
1997 
1998  /* If shadow use is a int->float cast then insert a second IV
1999  to eliminate this cast.
2000 
2001  for (unsigned i = 0; i < n; ++i)
2002  foo((double)i);
2003 
2004  is transformed into
2005 
2006  double d = 0.0;
2007  for (unsigned i = 0; i < n; ++i, ++d)
2008  foo(d);
2009  */
2010  if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2011  IsSigned = false;
2012  DestTy = UCast->getDestTy();
2013  }
2014  else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2015  IsSigned = true;
2016  DestTy = SCast->getDestTy();
2017  }
2018  if (!DestTy) continue;
2019 
2020  // If target does not support DestTy natively then do not apply
2021  // this transformation.
2022  if (!TTI.isTypeLegal(DestTy)) continue;
2023 
2024  PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2025  if (!PH) continue;
2026  if (PH->getNumIncomingValues() != 2) continue;
2027 
2028  // If the calculation in integers overflows, the result in FP type will
2029  // differ. So we only can do this transformation if we are guaranteed to not
2030  // deal with overflowing values
2031  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2032  if (!AR) continue;
2033  if (IsSigned && !AR->hasNoSignedWrap()) continue;
2034  if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2035 
2036  Type *SrcTy = PH->getType();
2037  int Mantissa = DestTy->getFPMantissaWidth();
2038  if (Mantissa == -1) continue;
2039  if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2040  continue;
2041 
2042  unsigned Entry, Latch;
2043  if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2044  Entry = 0;
2045  Latch = 1;
2046  } else {
2047  Entry = 1;
2048  Latch = 0;
2049  }
2050 
2052  if (!Init) continue;
2053  Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2054  (double)Init->getSExtValue() :
2055  (double)Init->getZExtValue());
2056 
2057  BinaryOperator *Incr =
2059  if (!Incr) continue;
2060  if (Incr->getOpcode() != Instruction::Add
2061  && Incr->getOpcode() != Instruction::Sub)
2062  continue;
2063 
2064  /* Initialize new IV, double d = 0.0 in above example. */
2065  ConstantInt *C = nullptr;
2066  if (Incr->getOperand(0) == PH)
2067  C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2068  else if (Incr->getOperand(1) == PH)
2069  C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2070  else
2071  continue;
2072 
2073  if (!C) continue;
2074 
2075  // Ignore negative constants, as the code below doesn't handle them
2076  // correctly. TODO: Remove this restriction.
2077  if (!C->getValue().isStrictlyPositive()) continue;
2078 
2079  /* Add new PHINode. */
2080  PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2081 
2082  /* create new increment. '++d' in above example. */
2083  Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2084  BinaryOperator *NewIncr =
2085  BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2086  Instruction::FAdd : Instruction::FSub,
2087  NewPH, CFP, "IV.S.next.", Incr);
2088 
2089  NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2090  NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2091 
2092  /* Remove cast operation */
2093  ShadowUse->replaceAllUsesWith(NewPH);
2094  ShadowUse->eraseFromParent();
2095  Changed = true;
2096  break;
2097  }
2098 }
2099 
2100 /// If Cond has an operand that is an expression of an IV, set the IV user and
2101 /// stride information and return true, otherwise return false.
2102 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2103  for (IVStrideUse &U : IU)
2104  if (U.getUser() == Cond) {
2105  // NOTE: we could handle setcc instructions with multiple uses here, but
2106  // InstCombine does it as well for simple uses, it's not clear that it
2107  // occurs enough in real life to handle.
2108  CondUse = &U;
2109  return true;
2110  }
2111  return false;
2112 }
2113 
2114 /// Rewrite the loop's terminating condition if it uses a max computation.
2115 ///
2116 /// This is a narrow solution to a specific, but acute, problem. For loops
2117 /// like this:
2118 ///
2119 /// i = 0;
2120 /// do {
2121 /// p[i] = 0.0;
2122 /// } while (++i < n);
2123 ///
2124 /// the trip count isn't just 'n', because 'n' might not be positive. And
2125 /// unfortunately this can come up even for loops where the user didn't use
2126 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2127 /// will commonly be lowered like this:
2128 ///
2129 /// if (n > 0) {
2130 /// i = 0;
2131 /// do {
2132 /// p[i] = 0.0;
2133 /// } while (++i < n);
2134 /// }
2135 ///
2136 /// and then it's possible for subsequent optimization to obscure the if
2137 /// test in such a way that indvars can't find it.
2138 ///
2139 /// When indvars can't find the if test in loops like this, it creates a
2140 /// max expression, which allows it to give the loop a canonical
2141 /// induction variable:
2142 ///
2143 /// i = 0;
2144 /// max = n < 1 ? 1 : n;
2145 /// do {
2146 /// p[i] = 0.0;
2147 /// } while (++i != max);
2148 ///
2149 /// Canonical induction variables are necessary because the loop passes
2150 /// are designed around them. The most obvious example of this is the
2151 /// LoopInfo analysis, which doesn't remember trip count values. It
2152 /// expects to be able to rediscover the trip count each time it is
2153 /// needed, and it does this using a simple analysis that only succeeds if
2154 /// the loop has a canonical induction variable.
2155 ///
2156 /// However, when it comes time to generate code, the maximum operation
2157 /// can be quite costly, especially if it's inside of an outer loop.
2158 ///
2159 /// This function solves this problem by detecting this type of loop and
2160 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2161 /// the instructions for the maximum computation.
2162 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2163  // Check that the loop matches the pattern we're looking for.
2164  if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2165  Cond->getPredicate() != CmpInst::ICMP_NE)
2166  return Cond;
2167 
2168  SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2169  if (!Sel || !Sel->hasOneUse()) return Cond;
2170 
2171  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2172  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2173  return Cond;
2174  const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2175 
2176  // Add one to the backedge-taken count to get the trip count.
2177  const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2178  if (IterationCount != SE.getSCEV(Sel)) return Cond;
2179 
2180  // Check for a max calculation that matches the pattern. There's no check
2181  // for ICMP_ULE here because the comparison would be with zero, which
2182  // isn't interesting.
2184  const SCEVNAryExpr *Max = nullptr;
2185  if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2186  Pred = ICmpInst::ICMP_SLE;
2187  Max = S;
2188  } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2189  Pred = ICmpInst::ICMP_SLT;
2190  Max = S;
2191  } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2192  Pred = ICmpInst::ICMP_ULT;
2193  Max = U;
2194  } else {
2195  // No match; bail.
2196  return Cond;
2197  }
2198 
2199  // To handle a max with more than two operands, this optimization would
2200  // require additional checking and setup.
2201  if (Max->getNumOperands() != 2)
2202  return Cond;
2203 
2204  const SCEV *MaxLHS = Max->getOperand(0);
2205  const SCEV *MaxRHS = Max->getOperand(1);
2206 
2207  // ScalarEvolution canonicalizes constants to the left. For < and >, look
2208  // for a comparison with 1. For <= and >=, a comparison with zero.
2209  if (!MaxLHS ||
2210  (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2211  return Cond;
2212 
2213  // Check the relevant induction variable for conformance to
2214  // the pattern.
2215  const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2216  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2217  if (!AR || !AR->isAffine() ||
2218  AR->getStart() != One ||
2219  AR->getStepRecurrence(SE) != One)
2220  return Cond;
2221 
2222  assert(AR->getLoop() == L &&
2223  "Loop condition operand is an addrec in a different loop!");
2224 
2225  // Check the right operand of the select, and remember it, as it will
2226  // be used in the new comparison instruction.
2227  Value *NewRHS = nullptr;
2228  if (ICmpInst::isTrueWhenEqual(Pred)) {
2229  // Look for n+1, and grab n.
2230  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2231  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2232  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2233  NewRHS = BO->getOperand(0);
2234  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2235  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2236  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2237  NewRHS = BO->getOperand(0);
2238  if (!NewRHS)
2239  return Cond;
2240  } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2241  NewRHS = Sel->getOperand(1);
2242  else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2243  NewRHS = Sel->getOperand(2);
2244  else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2245  NewRHS = SU->getValue();
2246  else
2247  // Max doesn't match expected pattern.
2248  return Cond;
2249 
2250  // Determine the new comparison opcode. It may be signed or unsigned,
2251  // and the original comparison may be either equality or inequality.
2252  if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2253  Pred = CmpInst::getInversePredicate(Pred);
2254 
2255  // Ok, everything looks ok to change the condition into an SLT or SGE and
2256  // delete the max calculation.
2257  ICmpInst *NewCond =
2258  new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2259 
2260  // Delete the max calculation instructions.
2261  Cond->replaceAllUsesWith(NewCond);
2262  CondUse->setUser(NewCond);
2263  Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2264  Cond->eraseFromParent();
2265  Sel->eraseFromParent();
2266  if (Cmp->use_empty())
2267  Cmp->eraseFromParent();
2268  return NewCond;
2269 }
2270 
2271 /// Change loop terminating condition to use the postinc iv when possible.
2272 void
2273 LSRInstance::OptimizeLoopTermCond() {
2275 
2276  // We need a different set of heuristics for rotated and non-rotated loops.
2277  // If a loop is rotated then the latch is also the backedge, so inserting
2278  // post-inc expressions just before the latch is ideal. To reduce live ranges
2279  // it also makes sense to rewrite terminating conditions to use post-inc
2280  // expressions.
2281  //
2282  // If the loop is not rotated then the latch is not a backedge; the latch
2283  // check is done in the loop head. Adding post-inc expressions before the
2284  // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2285  // in the loop body. In this case we do *not* want to use post-inc expressions
2286  // in the latch check, and we want to insert post-inc expressions before
2287  // the backedge.
2288  BasicBlock *LatchBlock = L->getLoopLatch();
2289  SmallVector<BasicBlock*, 8> ExitingBlocks;
2290  L->getExitingBlocks(ExitingBlocks);
2291  if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2292  return LatchBlock != BB;
2293  })) {
2294  // The backedge doesn't exit the loop; treat this as a head-tested loop.
2295  IVIncInsertPos = LatchBlock->getTerminator();
2296  return;
2297  }
2298 
2299  // Otherwise treat this as a rotated loop.
2300  for (BasicBlock *ExitingBlock : ExitingBlocks) {
2301  // Get the terminating condition for the loop if possible. If we
2302  // can, we want to change it to use a post-incremented version of its
2303  // induction variable, to allow coalescing the live ranges for the IV into
2304  // one register value.
2305 
2306  BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2307  if (!TermBr)
2308  continue;
2309  // FIXME: Overly conservative, termination condition could be an 'or' etc..
2310  if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2311  continue;
2312 
2313  // Search IVUsesByStride to find Cond's IVUse if there is one.
2314  IVStrideUse *CondUse = nullptr;
2315  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2316  if (!FindIVUserForCond(Cond, CondUse))
2317  continue;
2318 
2319  // If the trip count is computed in terms of a max (due to ScalarEvolution
2320  // being unable to find a sufficient guard, for example), change the loop
2321  // comparison to use SLT or ULT instead of NE.
2322  // One consequence of doing this now is that it disrupts the count-down
2323  // optimization. That's not always a bad thing though, because in such
2324  // cases it may still be worthwhile to avoid a max.
2325  Cond = OptimizeMax(Cond, CondUse);
2326 
2327  // If this exiting block dominates the latch block, it may also use
2328  // the post-inc value if it won't be shared with other uses.
2329  // Check for dominance.
2330  if (!DT.dominates(ExitingBlock, LatchBlock))
2331  continue;
2332 
2333  // Conservatively avoid trying to use the post-inc value in non-latch
2334  // exits if there may be pre-inc users in intervening blocks.
2335  if (LatchBlock != ExitingBlock)
2336  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2337  // Test if the use is reachable from the exiting block. This dominator
2338  // query is a conservative approximation of reachability.
2339  if (&*UI != CondUse &&
2340  !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2341  // Conservatively assume there may be reuse if the quotient of their
2342  // strides could be a legal scale.
2343  const SCEV *A = IU.getStride(*CondUse, L);
2344  const SCEV *B = IU.getStride(*UI, L);
2345  if (!A || !B) continue;
2346  if (SE.getTypeSizeInBits(A->getType()) !=
2347  SE.getTypeSizeInBits(B->getType())) {
2348  if (SE.getTypeSizeInBits(A->getType()) >
2349  SE.getTypeSizeInBits(B->getType()))
2350  B = SE.getSignExtendExpr(B, A->getType());
2351  else
2352  A = SE.getSignExtendExpr(A, B->getType());
2353  }
2354  if (const SCEVConstant *D =
2355  dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2356  const ConstantInt *C = D->getValue();
2357  // Stride of one or negative one can have reuse with non-addresses.
2358  if (C->isOne() || C->isMinusOne())
2359  goto decline_post_inc;
2360  // Avoid weird situations.
2361  if (C->getValue().getMinSignedBits() >= 64 ||
2362  C->getValue().isMinSignedValue())
2363  goto decline_post_inc;
2364  // Check for possible scaled-address reuse.
2365  MemAccessTy AccessTy = getAccessType(UI->getUser());
2366  int64_t Scale = C->getSExtValue();
2367  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2368  /*BaseOffset=*/0,
2369  /*HasBaseReg=*/false, Scale,
2370  AccessTy.AddrSpace))
2371  goto decline_post_inc;
2372  Scale = -Scale;
2373  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2374  /*BaseOffset=*/0,
2375  /*HasBaseReg=*/false, Scale,
2376  AccessTy.AddrSpace))
2377  goto decline_post_inc;
2378  }
2379  }
2380 
2381  DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2382  << *Cond << '\n');
2383 
2384  // It's possible for the setcc instruction to be anywhere in the loop, and
2385  // possible for it to have multiple users. If it is not immediately before
2386  // the exiting block branch, move it.
2387  if (&*++BasicBlock::iterator(Cond) != TermBr) {
2388  if (Cond->hasOneUse()) {
2389  Cond->moveBefore(TermBr);
2390  } else {
2391  // Clone the terminating condition and insert into the loopend.
2392  ICmpInst *OldCond = Cond;
2393  Cond = cast<ICmpInst>(Cond->clone());
2394  Cond->setName(L->getHeader()->getName() + ".termcond");
2395  ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2396 
2397  // Clone the IVUse, as the old use still exists!
2398  CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2399  TermBr->replaceUsesOfWith(OldCond, Cond);
2400  }
2401  }
2402 
2403  // If we get to here, we know that we can transform the setcc instruction to
2404  // use the post-incremented version of the IV, allowing us to coalesce the
2405  // live ranges for the IV correctly.
2406  CondUse->transformToPostInc(L);
2407  Changed = true;
2408 
2409  PostIncs.insert(Cond);
2410  decline_post_inc:;
2411  }
2412 
2413  // Determine an insertion point for the loop induction variable increment. It
2414  // must dominate all the post-inc comparisons we just set up, and it must
2415  // dominate the loop latch edge.
2416  IVIncInsertPos = L->getLoopLatch()->getTerminator();
2417  for (Instruction *Inst : PostIncs) {
2418  BasicBlock *BB =
2419  DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2420  Inst->getParent());
2421  if (BB == Inst->getParent())
2422  IVIncInsertPos = Inst;
2423  else if (BB != IVIncInsertPos->getParent())
2424  IVIncInsertPos = BB->getTerminator();
2425  }
2426 }
2427 
2428 /// Determine if the given use can accommodate a fixup at the given offset and
2429 /// other details. If so, update the use and return true.
2430 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2431  bool HasBaseReg, LSRUse::KindType Kind,
2432  MemAccessTy AccessTy) {
2433  int64_t NewMinOffset = LU.MinOffset;
2434  int64_t NewMaxOffset = LU.MaxOffset;
2435  MemAccessTy NewAccessTy = AccessTy;
2436 
2437  // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2438  // something conservative, however this can pessimize in the case that one of
2439  // the uses will have all its uses outside the loop, for example.
2440  if (LU.Kind != Kind)
2441  return false;
2442 
2443  // Check for a mismatched access type, and fall back conservatively as needed.
2444  // TODO: Be less conservative when the type is similar and can use the same
2445  // addressing modes.
2446  if (Kind == LSRUse::Address) {
2447  if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2448  NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2449  AccessTy.AddrSpace);
2450  }
2451  }
2452 
2453  // Conservatively assume HasBaseReg is true for now.
2454  if (NewOffset < LU.MinOffset) {
2455  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2456  LU.MaxOffset - NewOffset, HasBaseReg))
2457  return false;
2458  NewMinOffset = NewOffset;
2459  } else if (NewOffset > LU.MaxOffset) {
2460  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2461  NewOffset - LU.MinOffset, HasBaseReg))
2462  return false;
2463  NewMaxOffset = NewOffset;
2464  }
2465 
2466  // Update the use.
2467  LU.MinOffset = NewMinOffset;
2468  LU.MaxOffset = NewMaxOffset;
2469  LU.AccessTy = NewAccessTy;
2470  return true;
2471 }
2472 
2473 /// Return an LSRUse index and an offset value for a fixup which needs the given
2474 /// expression, with the given kind and optional access type. Either reuse an
2475 /// existing use or create a new one, as needed.
2476 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2477  LSRUse::KindType Kind,
2478  MemAccessTy AccessTy) {
2479  const SCEV *Copy = Expr;
2480  int64_t Offset = ExtractImmediate(Expr, SE);
2481 
2482  // Basic uses can't accept any offset, for example.
2483  if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2484  Offset, /*HasBaseReg=*/ true)) {
2485  Expr = Copy;
2486  Offset = 0;
2487  }
2488 
2489  std::pair<UseMapTy::iterator, bool> P =
2490  UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2491  if (!P.second) {
2492  // A use already existed with this base.
2493  size_t LUIdx = P.first->second;
2494  LSRUse &LU = Uses[LUIdx];
2495  if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2496  // Reuse this use.
2497  return std::make_pair(LUIdx, Offset);
2498  }
2499 
2500  // Create a new use.
2501  size_t LUIdx = Uses.size();
2502  P.first->second = LUIdx;
2503  Uses.push_back(LSRUse(Kind, AccessTy));
2504  LSRUse &LU = Uses[LUIdx];
2505 
2506  LU.MinOffset = Offset;
2507  LU.MaxOffset = Offset;
2508  return std::make_pair(LUIdx, Offset);
2509 }
2510 
2511 /// Delete the given use from the Uses list.
2512 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2513  if (&LU != &Uses.back())
2514  std::swap(LU, Uses.back());
2515  Uses.pop_back();
2516 
2517  // Update RegUses.
2518  RegUses.swapAndDropUse(LUIdx, Uses.size());
2519 }
2520 
2521 /// Look for a use distinct from OrigLU which is has a formula that has the same
2522 /// registers as the given formula.
2523 LSRUse *
2524 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2525  const LSRUse &OrigLU) {
2526  // Search all uses for the formula. This could be more clever.
2527  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2528  LSRUse &LU = Uses[LUIdx];
2529  // Check whether this use is close enough to OrigLU, to see whether it's
2530  // worthwhile looking through its formulae.
2531  // Ignore ICmpZero uses because they may contain formulae generated by
2532  // GenerateICmpZeroScales, in which case adding fixup offsets may
2533  // be invalid.
2534  if (&LU != &OrigLU &&
2535  LU.Kind != LSRUse::ICmpZero &&
2536  LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2537  LU.WidestFixupType == OrigLU.WidestFixupType &&
2538  LU.HasFormulaWithSameRegs(OrigF)) {
2539  // Scan through this use's formulae.
2540  for (const Formula &F : LU.Formulae) {
2541  // Check to see if this formula has the same registers and symbols
2542  // as OrigF.
2543  if (F.BaseRegs == OrigF.BaseRegs &&
2544  F.ScaledReg == OrigF.ScaledReg &&
2545  F.BaseGV == OrigF.BaseGV &&
2546  F.Scale == OrigF.Scale &&
2547  F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2548  if (F.BaseOffset == 0)
2549  return &LU;
2550  // This is the formula where all the registers and symbols matched;
2551  // there aren't going to be any others. Since we declined it, we
2552  // can skip the rest of the formulae and proceed to the next LSRUse.
2553  break;
2554  }
2555  }
2556  }
2557  }
2558 
2559  // Nothing looked good.
2560  return nullptr;
2561 }
2562 
2563 void LSRInstance::CollectInterestingTypesAndFactors() {
2565 
2566  // Collect interesting types and strides.
2568  for (const IVStrideUse &U : IU) {
2569  const SCEV *Expr = IU.getExpr(U);
2570 
2571  // Collect interesting types.
2572  Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2573 
2574  // Add strides for mentioned loops.
2575  Worklist.push_back(Expr);
2576  do {
2577  const SCEV *S = Worklist.pop_back_val();
2578  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2579  if (AR->getLoop() == L)
2580  Strides.insert(AR->getStepRecurrence(SE));
2581  Worklist.push_back(AR->getStart());
2582  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2583  Worklist.append(Add->op_begin(), Add->op_end());
2584  }
2585  } while (!Worklist.empty());
2586  }
2587 
2588  // Compute interesting factors from the set of interesting strides.
2590  I = Strides.begin(), E = Strides.end(); I != E; ++I)
2592  std::next(I); NewStrideIter != E; ++NewStrideIter) {
2593  const SCEV *OldStride = *I;
2594  const SCEV *NewStride = *NewStrideIter;
2595 
2596  if (SE.getTypeSizeInBits(OldStride->getType()) !=
2597  SE.getTypeSizeInBits(NewStride->getType())) {
2598  if (SE.getTypeSizeInBits(OldStride->getType()) >
2599  SE.getTypeSizeInBits(NewStride->getType()))
2600  NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2601  else
2602  OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2603  }
2604  if (const SCEVConstant *Factor =
2605  dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2606  SE, true))) {
2607  if (Factor->getAPInt().getMinSignedBits() <= 64)
2608  Factors.insert(Factor->getAPInt().getSExtValue());
2609  } else if (const SCEVConstant *Factor =
2610  dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2611  NewStride,
2612  SE, true))) {
2613  if (Factor->getAPInt().getMinSignedBits() <= 64)
2614  Factors.insert(Factor->getAPInt().getSExtValue());
2615  }
2616  }
2617 
2618  // If all uses use the same type, don't bother looking for truncation-based
2619  // reuse.
2620  if (Types.size() == 1)
2621  Types.clear();
2622 
2623  DEBUG(print_factors_and_types(dbgs()));
2624 }
2625 
2626 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2627 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2628 /// IVStrideUses, we could partially skip this.
2629 static User::op_iterator
2631  Loop *L, ScalarEvolution &SE) {
2632  for(; OI != OE; ++OI) {
2633  if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2634  if (!SE.isSCEVable(Oper->getType()))
2635  continue;
2636 
2637  if (const SCEVAddRecExpr *AR =
2638  dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2639  if (AR->getLoop() == L)
2640  break;
2641  }
2642  }
2643  }
2644  return OI;
2645 }
2646 
2647 /// IVChain logic must consistenctly peek base TruncInst operands, so wrap it in
2648 /// a convenient helper.
2649 static Value *getWideOperand(Value *Oper) {
2650  if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2651  return Trunc->getOperand(0);
2652  return Oper;
2653 }
2654 
2655 /// Return true if we allow an IV chain to include both types.
2656 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2657  Type *LType = LVal->getType();
2658  Type *RType = RVal->getType();
2659  return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2660  // Different address spaces means (possibly)
2661  // different types of the pointer implementation,
2662  // e.g. i16 vs i32 so disallow that.
2663  (LType->getPointerAddressSpace() ==
2664  RType->getPointerAddressSpace()));
2665 }
2666 
2667 /// Return an approximation of this SCEV expression's "base", or NULL for any
2668 /// constant. Returning the expression itself is conservative. Returning a
2669 /// deeper subexpression is more precise and valid as long as it isn't less
2670 /// complex than another subexpression. For expressions involving multiple
2671 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2672 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2673 /// IVInc==b-a.
2674 ///
2675 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2676 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2677 static const SCEV *getExprBase(const SCEV *S) {
2678  switch (S->getSCEVType()) {
2679  default: // uncluding scUnknown.
2680  return S;
2681  case scConstant:
2682  return nullptr;
2683  case scTruncate:
2684  return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2685  case scZeroExtend:
2686  return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2687  case scSignExtend:
2688  return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2689  case scAddExpr: {
2690  // Skip over scaled operands (scMulExpr) to follow add operands as long as
2691  // there's nothing more complex.
2692  // FIXME: not sure if we want to recognize negation.
2693  const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2694  for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2695  E(Add->op_begin()); I != E; ++I) {
2696  const SCEV *SubExpr = *I;
2697  if (SubExpr->getSCEVType() == scAddExpr)
2698  return getExprBase(SubExpr);
2699 
2700  if (SubExpr->getSCEVType() != scMulExpr)
2701  return SubExpr;
2702  }
2703  return S; // all operands are scaled, be conservative.
2704  }
2705  case scAddRecExpr:
2706  return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2707  }
2708 }
2709 
2710 /// Return true if the chain increment is profitable to expand into a loop
2711 /// invariant value, which may require its own register. A profitable chain
2712 /// increment will be an offset relative to the same base. We allow such offsets
2713 /// to potentially be used as chain increment as long as it's not obviously
2714 /// expensive to expand using real instructions.
2715 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2716  const SCEV *IncExpr,
2717  ScalarEvolution &SE) {
2718  // Aggressively form chains when -stress-ivchain.
2719  if (StressIVChain)
2720  return true;
2721 
2722  // Do not replace a constant offset from IV head with a nonconstant IV
2723  // increment.
2724  if (!isa<SCEVConstant>(IncExpr)) {
2725  const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2726  if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2727  return false;
2728  }
2729 
2730  SmallPtrSet<const SCEV*, 8> Processed;
2731  return !isHighCostExpansion(IncExpr, Processed, SE);
2732 }
2733 
2734 /// Return true if the number of registers needed for the chain is estimated to
2735 /// be less than the number required for the individual IV users. First prohibit
2736 /// any IV users that keep the IV live across increments (the Users set should
2737 /// be empty). Next count the number and type of increments in the chain.
2738 ///
2739 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2740 /// effectively use postinc addressing modes. Only consider it profitable it the
2741 /// increments can be computed in fewer registers when chained.
2742 ///
2743 /// TODO: Consider IVInc free if it's already used in another chains.
2744 static bool
2746  ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2747  if (StressIVChain)
2748  return true;
2749 
2750  if (!Chain.hasIncs())
2751  return false;
2752 
2753  if (!Users.empty()) {
2754  DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2755  for (Instruction *Inst : Users) {
2756  dbgs() << " " << *Inst << "\n";
2757  });
2758  return false;
2759  }
2760  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2761 
2762  // The chain itself may require a register, so intialize cost to 1.
2763  int cost = 1;
2764 
2765  // A complete chain likely eliminates the need for keeping the original IV in
2766  // a register. LSR does not currently know how to form a complete chain unless
2767  // the header phi already exists.
2768  if (isa<PHINode>(Chain.tailUserInst())
2769  && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2770  --cost;
2771  }
2772  const SCEV *LastIncExpr = nullptr;
2773  unsigned NumConstIncrements = 0;
2774  unsigned NumVarIncrements = 0;
2775  unsigned NumReusedIncrements = 0;
2776  for (const IVInc &Inc : Chain) {
2777  if (Inc.IncExpr->isZero())
2778  continue;
2779 
2780  // Incrementing by zero or some constant is neutral. We assume constants can
2781  // be folded into an addressing mode or an add's immediate operand.
2782  if (isa<SCEVConstant>(Inc.IncExpr)) {
2783  ++NumConstIncrements;
2784  continue;
2785  }
2786 
2787  if (Inc.IncExpr == LastIncExpr)
2788  ++NumReusedIncrements;
2789  else
2790  ++NumVarIncrements;
2791 
2792  LastIncExpr = Inc.IncExpr;
2793  }
2794  // An IV chain with a single increment is handled by LSR's postinc
2795  // uses. However, a chain with multiple increments requires keeping the IV's
2796  // value live longer than it needs to be if chained.
2797  if (NumConstIncrements > 1)
2798  --cost;
2799 
2800  // Materializing increment expressions in the preheader that didn't exist in
2801  // the original code may cost a register. For example, sign-extended array
2802  // indices can produce ridiculous increments like this:
2803  // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2804  cost += NumVarIncrements;
2805 
2806  // Reusing variable increments likely saves a register to hold the multiple of
2807  // the stride.
2808  cost -= NumReusedIncrements;
2809 
2810  DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2811  << "\n");
2812 
2813  return cost < 0;
2814 }
2815 
2816 /// Add this IV user to an existing chain or make it the head of a new chain.
2817 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2818  SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2819  // When IVs are used as types of varying widths, they are generally converted
2820  // to a wider type with some uses remaining narrow under a (free) trunc.
2821  Value *const NextIV = getWideOperand(IVOper);
2822  const SCEV *const OperExpr = SE.getSCEV(NextIV);
2823  const SCEV *const OperExprBase = getExprBase(OperExpr);
2824 
2825  // Visit all existing chains. Check if its IVOper can be computed as a
2826  // profitable loop invariant increment from the last link in the Chain.
2827  unsigned ChainIdx = 0, NChains = IVChainVec.size();
2828  const SCEV *LastIncExpr = nullptr;
2829  for (; ChainIdx < NChains; ++ChainIdx) {
2830  IVChain &Chain = IVChainVec[ChainIdx];
2831 
2832  // Prune the solution space aggressively by checking that both IV operands
2833  // are expressions that operate on the same unscaled SCEVUnknown. This
2834  // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2835  // first avoids creating extra SCEV expressions.
2836  if (!StressIVChain && Chain.ExprBase != OperExprBase)
2837  continue;
2838 
2839  Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2840  if (!isCompatibleIVType(PrevIV, NextIV))
2841  continue;
2842 
2843  // A phi node terminates a chain.
2844  if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2845  continue;
2846 
2847  // The increment must be loop-invariant so it can be kept in a register.
2848  const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2849  const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2850  if (!SE.isLoopInvariant(IncExpr, L))
2851  continue;
2852 
2853  if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2854  LastIncExpr = IncExpr;
2855  break;
2856  }
2857  }
2858  // If we haven't found a chain, create a new one, unless we hit the max. Don't
2859  // bother for phi nodes, because they must be last in the chain.
2860  if (ChainIdx == NChains) {
2861  if (isa<PHINode>(UserInst))
2862  return;
2863  if (NChains >= MaxChains && !StressIVChain) {
2864  DEBUG(dbgs() << "IV Chain Limit\n");
2865  return;
2866  }
2867  LastIncExpr = OperExpr;
2868  // IVUsers may have skipped over sign/zero extensions. We don't currently
2869  // attempt to form chains involving extensions unless they can be hoisted
2870  // into this loop's AddRec.
2871  if (!isa<SCEVAddRecExpr>(LastIncExpr))
2872  return;
2873  ++NChains;
2874  IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2875  OperExprBase));
2876  ChainUsersVec.resize(NChains);
2877  DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2878  << ") IV=" << *LastIncExpr << "\n");
2879  } else {
2880  DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2881  << ") IV+" << *LastIncExpr << "\n");
2882  // Add this IV user to the end of the chain.
2883  IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2884  }
2885  IVChain &Chain = IVChainVec[ChainIdx];
2886 
2887  SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2888  // This chain's NearUsers become FarUsers.
2889  if (!LastIncExpr->isZero()) {
2890  ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2891  NearUsers.end());
2892  NearUsers.clear();
2893  }
2894 
2895  // All other uses of IVOperand become near uses of the chain.
2896  // We currently ignore intermediate values within SCEV expressions, assuming
2897  // they will eventually be used be the current chain, or can be computed
2898  // from one of the chain increments. To be more precise we could
2899  // transitively follow its user and only add leaf IV users to the set.
2900  for (User *U : IVOper->users()) {
2901  Instruction *OtherUse = dyn_cast<Instruction>(U);
2902  if (!OtherUse)
2903  continue;
2904  // Uses in the chain will no longer be uses if the chain is formed.
2905  // Include the head of the chain in this iteration (not Chain.begin()).
2906  IVChain::const_iterator IncIter = Chain.Incs.begin();
2907  IVChain::const_iterator IncEnd = Chain.Incs.end();
2908  for( ; IncIter != IncEnd; ++IncIter) {
2909  if (IncIter->UserInst == OtherUse)
2910  break;
2911  }
2912  if (IncIter != IncEnd)
2913  continue;
2914 
2915  if (SE.isSCEVable(OtherUse->getType())
2916  && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2917  && IU.isIVUserOrOperand(OtherUse)) {
2918  continue;
2919  }
2920  NearUsers.insert(OtherUse);
2921  }
2922 
2923  // Since this user is part of the chain, it's no longer considered a use
2924  // of the chain.
2925  ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2926 }
2927 
2928 /// Populate the vector of Chains.
2929 ///
2930 /// This decreases ILP at the architecture level. Targets with ample registers,
2931 /// multiple memory ports, and no register renaming probably don't want
2932 /// this. However, such targets should probably disable LSR altogether.
2933 ///
2934 /// The job of LSR is to make a reasonable choice of induction variables across
2935 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2936 /// ILP *within the loop* if the target wants it.
2937 ///
2938 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2939 /// will not reorder memory operations, it will recognize this as a chain, but
2940 /// will generate redundant IV increments. Ideally this would be corrected later
2941 /// by a smart scheduler:
2942 /// = A[i]
2943 /// = A[i+x]
2944 /// A[i] =
2945 /// A[i+x] =
2946 ///
2947 /// TODO: Walk the entire domtree within this loop, not just the path to the
2948 /// loop latch. This will discover chains on side paths, but requires
2949 /// maintaining multiple copies of the Chains state.
2950 void LSRInstance::CollectChains() {
2951  DEBUG(dbgs() << "Collecting IV Chains.\n");
2952  SmallVector<ChainUsers, 8> ChainUsersVec;
2953 
2954  SmallVector<BasicBlock *,8> LatchPath;
2955  BasicBlock *LoopHeader = L->getHeader();
2956  for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2957  Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2958  LatchPath.push_back(Rung->getBlock());
2959  }
2960  LatchPath.push_back(LoopHeader);
2961 
2962  // Walk the instruction stream from the loop header to the loop latch.
2963  for (BasicBlock *BB : reverse(LatchPath)) {
2964  for (Instruction &I : *BB) {
2965  // Skip instructions that weren't seen by IVUsers analysis.
2966  if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
2967  continue;
2968 
2969  // Ignore users that are part of a SCEV expression. This way we only
2970  // consider leaf IV Users. This effectively rediscovers a portion of
2971  // IVUsers analysis but in program order this time.
2972  if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
2973  continue;
2974 
2975  // Remove this instruction from any NearUsers set it may be in.
2976  for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2977  ChainIdx < NChains; ++ChainIdx) {
2978  ChainUsersVec[ChainIdx].NearUsers.erase(&I);
2979  }
2980  // Search for operands that can be chained.
2981  SmallPtrSet<Instruction*, 4> UniqueOperands;
2982  User::op_iterator IVOpEnd = I.op_end();
2983  User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
2984  while (IVOpIter != IVOpEnd) {
2985  Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2986  if (UniqueOperands.insert(IVOpInst).second)
2987  ChainInstruction(&I, IVOpInst, ChainUsersVec);
2988  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2989  }
2990  } // Continue walking down the instructions.
2991  } // Continue walking down the domtree.
2992  // Visit phi backedges to determine if the chain can generate the IV postinc.
2993  for (BasicBlock::iterator I = L->getHeader()->begin();
2994  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2995  if (!SE.isSCEVable(PN->getType()))
2996  continue;
2997 
2998  Instruction *IncV =
3000  if (IncV)
3001  ChainInstruction(PN, IncV, ChainUsersVec);
3002  }
3003  // Remove any unprofitable chains.
3004  unsigned ChainIdx = 0;
3005  for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3006  UsersIdx < NChains; ++UsersIdx) {
3007  if (!isProfitableChain(IVChainVec[UsersIdx],
3008  ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3009  continue;
3010  // Preserve the chain at UsesIdx.
3011  if (ChainIdx != UsersIdx)
3012  IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3013  FinalizeChain(IVChainVec[ChainIdx]);
3014  ++ChainIdx;
3015  }
3016  IVChainVec.resize(ChainIdx);
3017 }
3018 
3019 void LSRInstance::FinalizeChain(IVChain &Chain) {
3020  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3021  DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
3022 
3023  for (const IVInc &Inc : Chain) {
3024  DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n");
3025  auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3026  assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
3027  IVIncSet.insert(UseI);
3028  }
3029 }
3030 
3031 /// Return true if the IVInc can be folded into an addressing mode.
3032 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3033  Value *Operand, const TargetTransformInfo &TTI) {
3034  const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3035  if (!IncConst || !isAddressUse(UserInst, Operand))
3036  return false;
3037 
3038  if (IncConst->getAPInt().getMinSignedBits() > 64)
3039  return false;
3040 
3041  MemAccessTy AccessTy = getAccessType(UserInst);
3042  int64_t IncOffset = IncConst->getValue()->getSExtValue();
3043  if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3044  IncOffset, /*HaseBaseReg=*/false))
3045  return false;
3046 
3047  return true;
3048 }
3049 
3050 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3051 /// user's operand from the previous IV user's operand.
3052 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3053  SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3054  // Find the new IVOperand for the head of the chain. It may have been replaced
3055  // by LSR.
3056  const IVInc &Head = Chain.Incs[0];
3057  User::op_iterator IVOpEnd = Head.UserInst->op_end();
3058  // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3059  User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3060  IVOpEnd, L, SE);
3061  Value *IVSrc = nullptr;
3062  while (IVOpIter != IVOpEnd) {
3063  IVSrc = getWideOperand(*IVOpIter);
3064 
3065  // If this operand computes the expression that the chain needs, we may use
3066  // it. (Check this after setting IVSrc which is used below.)
3067  //
3068  // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3069  // narrow for the chain, so we can no longer use it. We do allow using a
3070  // wider phi, assuming the LSR checked for free truncation. In that case we
3071  // should already have a truncate on this operand such that
3072  // getSCEV(IVSrc) == IncExpr.
3073  if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3074  || SE.getSCEV(IVSrc) == Head.IncExpr) {
3075  break;
3076  }
3077  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3078  }
3079  if (IVOpIter == IVOpEnd) {
3080  // Gracefully give up on this chain.
3081  DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3082  return;
3083  }
3084 
3085  DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3086  Type *IVTy = IVSrc->getType();
3087  Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3088  const SCEV *LeftOverExpr = nullptr;
3089  for (const IVInc &Inc : Chain) {
3090  Instruction *InsertPt = Inc.UserInst;
3091  if (isa<PHINode>(InsertPt))
3092  InsertPt = L->getLoopLatch()->getTerminator();
3093 
3094  // IVOper will replace the current IV User's operand. IVSrc is the IV
3095  // value currently held in a register.
3096  Value *IVOper = IVSrc;
3097  if (!Inc.IncExpr->isZero()) {
3098  // IncExpr was the result of subtraction of two narrow values, so must
3099  // be signed.
3100  const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3101  LeftOverExpr = LeftOverExpr ?
3102  SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3103  }
3104  if (LeftOverExpr && !LeftOverExpr->isZero()) {
3105  // Expand the IV increment.
3106  Rewriter.clearPostInc();
3107  Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3108  const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3109  SE.getUnknown(IncV));
3110  IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3111 
3112  // If an IV increment can't be folded, use it as the next IV value.
3113  if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3114  assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3115  IVSrc = IVOper;
3116  LeftOverExpr = nullptr;
3117  }
3118  }
3119  Type *OperTy = Inc.IVOperand->getType();
3120  if (IVTy != OperTy) {
3121  assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3122  "cannot extend a chained IV");
3123  IRBuilder<> Builder(InsertPt);
3124  IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3125  }
3126  Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3127  DeadInsts.emplace_back(Inc.IVOperand);
3128  }
3129  // If LSR created a new, wider phi, we may also replace its postinc. We only
3130  // do this if we also found a wide value for the head of the chain.
3131  if (isa<PHINode>(Chain.tailUserInst())) {
3132  for (BasicBlock::iterator I = L->getHeader()->begin();
3133  PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
3134  if (!isCompatibleIVType(Phi, IVSrc))
3135  continue;
3136  Instruction *PostIncV = dyn_cast<Instruction>(
3137  Phi->getIncomingValueForBlock(L->getLoopLatch()));
3138  if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3139  continue;
3140  Value *IVOper = IVSrc;
3141  Type *PostIncTy = PostIncV->getType();
3142  if (IVTy != PostIncTy) {
3143  assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3144  IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3145  Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3146  IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3147  }
3148  Phi->replaceUsesOfWith(PostIncV, IVOper);
3149  DeadInsts.emplace_back(PostIncV);
3150  }
3151  }
3152 }
3153 
3154 void LSRInstance::CollectFixupsAndInitialFormulae() {
3155  for (const IVStrideUse &U : IU) {
3156  Instruction *UserInst = U.getUser();
3157  // Skip IV users that are part of profitable IV Chains.
3158  User::op_iterator UseI =
3159  find(UserInst->operands(), U.getOperandValToReplace());
3160  assert(UseI != UserInst->op_end() && "cannot find IV operand");
3161  if (IVIncSet.count(UseI)) {
3162  DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3163  continue;
3164  }
3165 
3166  LSRUse::KindType Kind = LSRUse::Basic;
3167  MemAccessTy AccessTy;
3168  if (isAddressUse(UserInst, U.getOperandValToReplace())) {
3169  Kind = LSRUse::Address;
3170  AccessTy = getAccessType(UserInst);
3171  }
3172 
3173  const SCEV *S = IU.getExpr(U);
3174  PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3175 
3176  // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3177  // (N - i == 0), and this allows (N - i) to be the expression that we work
3178  // with rather than just N or i, so we can consider the register
3179  // requirements for both N and i at the same time. Limiting this code to
3180  // equality icmps is not a problem because all interesting loops use
3181  // equality icmps, thanks to IndVarSimplify.
3182  if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst))
3183  if (CI->isEquality()) {
3184  // Swap the operands if needed to put the OperandValToReplace on the
3185  // left, for consistency.
3186  Value *NV = CI->getOperand(1);
3187  if (NV == U.getOperandValToReplace()) {
3188  CI->setOperand(1, CI->getOperand(0));
3189  CI->setOperand(0, NV);
3190  NV = CI->getOperand(1);
3191  Changed = true;
3192  }
3193 
3194  // x == y --> x - y == 0
3195  const SCEV *N = SE.getSCEV(NV);
3196  if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3197  // S is normalized, so normalize N before folding it into S
3198  // to keep the result normalized.
3199  N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3200  Kind = LSRUse::ICmpZero;
3201  S = SE.getMinusSCEV(N, S);
3202  }
3203 
3204  // -1 and the negations of all interesting strides (except the negation
3205  // of -1) are now also interesting.
3206  for (size_t i = 0, e = Factors.size(); i != e; ++i)
3207  if (Factors[i] != -1)
3208  Factors.insert(-(uint64_t)Factors[i]);
3209  Factors.insert(-1);
3210  }
3211 
3212  // Get or create an LSRUse.
3213  std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3214  size_t LUIdx = P.first;
3215  int64_t Offset = P.second;
3216  LSRUse &LU = Uses[LUIdx];
3217 
3218  // Record the fixup.
3219  LSRFixup &LF = LU.getNewFixup();
3220  LF.UserInst = UserInst;
3221  LF.OperandValToReplace = U.getOperandValToReplace();
3222  LF.PostIncLoops = TmpPostIncLoops;
3223  LF.Offset = Offset;
3224  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3225 
3226  if (!LU.WidestFixupType ||
3227  SE.getTypeSizeInBits(LU.WidestFixupType) <
3228  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3229  LU.WidestFixupType = LF.OperandValToReplace->getType();
3230 
3231  // If this is the first use of this LSRUse, give it a formula.
3232  if (LU.Formulae.empty()) {
3233  InsertInitialFormula(S, LU, LUIdx);
3234  CountRegisters(LU.Formulae.back(), LUIdx);
3235  }
3236  }
3237 
3238  DEBUG(print_fixups(dbgs()));
3239 }
3240 
3241 /// Insert a formula for the given expression into the given use, separating out
3242 /// loop-variant portions from loop-invariant and loop-computable portions.
3243 void
3244 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3245  // Mark uses whose expressions cannot be expanded.
3246  if (!isSafeToExpand(S, SE))
3247  LU.RigidFormula = true;
3248 
3249  Formula F;
3250  F.initialMatch(S, L, SE);
3251  bool Inserted = InsertFormula(LU, LUIdx, F);
3252  assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3253 }
3254 
3255 /// Insert a simple single-register formula for the given expression into the
3256 /// given use.
3257 void
3258 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3259  LSRUse &LU, size_t LUIdx) {
3260  Formula F;
3261  F.BaseRegs.push_back(S);
3262  F.HasBaseReg = true;
3263  bool Inserted = InsertFormula(LU, LUIdx, F);
3264  assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3265 }
3266 
3267 /// Note which registers are used by the given formula, updating RegUses.
3268 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3269  if (F.ScaledReg)
3270  RegUses.countRegister(F.ScaledReg, LUIdx);
3271  for (const SCEV *BaseReg : F.BaseRegs)
3272  RegUses.countRegister(BaseReg, LUIdx);
3273 }
3274 
3275 /// If the given formula has not yet been inserted, add it to the list, and
3276 /// return true. Return false otherwise.
3277 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3278  // Do not insert formula that we will not be able to expand.
3279  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3280  "Formula is illegal");
3281 
3282  if (!LU.InsertFormula(F, *L))
3283  return false;
3284 
3285  CountRegisters(F, LUIdx);
3286  return true;
3287 }
3288 
3289 /// Check for other uses of loop-invariant values which we're tracking. These
3290 /// other uses will pin these values in registers, making them less profitable
3291 /// for elimination.
3292 /// TODO: This currently misses non-constant addrec step registers.
3293 /// TODO: Should this give more weight to users inside the loop?
3294 void
3295 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3296  SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3298 
3299  while (!Worklist.empty()) {
3300  const SCEV *S = Worklist.pop_back_val();
3301 
3302  // Don't process the same SCEV twice
3303  if (!Visited.insert(S).second)
3304  continue;
3305 
3306  if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3307  Worklist.append(N->op_begin(), N->op_end());
3308  else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3309  Worklist.push_back(C->getOperand());
3310  else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3311  Worklist.push_back(D->getLHS());
3312  Worklist.push_back(D->getRHS());
3313  } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3314  const Value *V = US->getValue();
3315  if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3316  // Look for instructions defined outside the loop.
3317  if (L->contains(Inst)) continue;
3318  } else if (isa<UndefValue>(V))
3319  // Undef doesn't have a live range, so it doesn't matter.
3320  continue;
3321  for (const Use &U : V->uses()) {
3322  const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3323  // Ignore non-instructions.
3324  if (!UserInst)
3325  continue;
3326  // Ignore instructions in other functions (as can happen with
3327  // Constants).
3328  if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3329  continue;
3330  // Ignore instructions not dominated by the loop.
3331  const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3332  UserInst->getParent() :
3333  cast<PHINode>(UserInst)->getIncomingBlock(
3334  PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3335  if (!DT.dominates(L->getHeader(), UseBB))
3336  continue;
3337  // Don't bother if the instruction is in a BB which ends in an EHPad.
3338  if (UseBB->getTerminator()->isEHPad())
3339  continue;
3340  // Don't bother rewriting PHIs in catchswitch blocks.
3341  if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3342  continue;
3343  // Ignore uses which are part of other SCEV expressions, to avoid
3344  // analyzing them multiple times.
3345  if (SE.isSCEVable(UserInst->getType())) {
3346  const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3347  // If the user is a no-op, look through to its uses.
3348  if (!isa<SCEVUnknown>(UserS))
3349  continue;
3350  if (UserS == US) {
3351  Worklist.push_back(
3352  SE.getUnknown(const_cast<Instruction *>(UserInst)));
3353  continue;
3354  }
3355  }
3356  // Ignore icmp instructions which are already being analyzed.
3357  if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3358  unsigned OtherIdx = !U.getOperandNo();
3359  Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3360  if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3361  continue;
3362  }
3363 
3364  std::pair<size_t, int64_t> P = getUse(
3365  S, LSRUse::Basic, MemAccessTy());
3366  size_t LUIdx = P.first;
3367  int64_t Offset = P.second;
3368  LSRUse &LU = Uses[LUIdx];
3369  LSRFixup &LF = LU.getNewFixup();
3370  LF.UserInst = const_cast<Instruction *>(UserInst);
3371  LF.OperandValToReplace = U;
3372  LF.Offset = Offset;
3373  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3374  if (!LU.WidestFixupType ||
3375  SE.getTypeSizeInBits(LU.WidestFixupType) <
3376  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3377  LU.WidestFixupType = LF.OperandValToReplace->getType();
3378  InsertSupplementalFormula(US, LU, LUIdx);
3379  CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3380  break;
3381  }
3382  }
3383  }
3384 }
3385 
3386 /// Split S into subexpressions which can be pulled out into separate
3387 /// registers. If C is non-null, multiply each subexpression by C.
3388 ///
3389 /// Return remainder expression after factoring the subexpressions captured by
3390 /// Ops. If Ops is complete, return NULL.
3391 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3393  const Loop *L,
3394  ScalarEvolution &SE,
3395  unsigned Depth = 0) {
3396  // Arbitrarily cap recursion to protect compile time.
3397  if (Depth >= 3)
3398  return S;
3399 
3400  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3401  // Break out add operands.
3402  for (const SCEV *S : Add->operands()) {
3403  const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3404  if (Remainder)
3405  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3406  }
3407  return nullptr;
3408  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3409  // Split a non-zero base out of an addrec.
3410  if (AR->getStart()->isZero() || !AR->isAffine())
3411  return S;
3412 
3413  const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3414  C, Ops, L, SE, Depth+1);
3415  // Split the non-zero AddRec unless it is part of a nested recurrence that
3416  // does not pertain to this loop.
3417  if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3418  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3419  Remainder = nullptr;
3420  }
3421  if (Remainder != AR->getStart()) {
3422  if (!Remainder)
3423  Remainder = SE.getConstant(AR->getType(), 0);
3424  return SE.getAddRecExpr(Remainder,
3425  AR->getStepRecurrence(SE),
3426  AR->getLoop(),
3427  //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3429  }
3430  } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3431  // Break (C * (a + b + c)) into C*a + C*b + C*c.
3432  if (Mul->getNumOperands() != 2)
3433  return S;
3434  if (const SCEVConstant *Op0 =
3435  dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3436  C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3437  const SCEV *Remainder =
3438  CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3439  if (Remainder)
3440  Ops.push_back(SE.getMulExpr(C, Remainder));
3441  return nullptr;
3442  }
3443  }
3444  return S;
3445 }
3446 
3447 /// \brief Helper function for LSRInstance::GenerateReassociations.
3448 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3449  const Formula &Base,
3450  unsigned Depth, size_t Idx,
3451  bool IsScaledReg) {
3452  const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3454  const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3455  if (Remainder)
3456  AddOps.push_back(Remainder);
3457 
3458  if (AddOps.size() == 1)
3459  return;
3460 
3462  JE = AddOps.end();
3463  J != JE; ++J) {
3464  // Loop-variant "unknown" values are uninteresting; we won't be able to
3465  // do anything meaningful with them.
3466  if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3467  continue;
3468 
3469  // Don't pull a constant into a register if the constant could be folded
3470  // into an immediate field.
3471  if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3472  LU.AccessTy, *J, Base.getNumRegs() > 1))
3473  continue;
3474 
3475  // Collect all operands except *J.
3476  SmallVector<const SCEV *, 8> InnerAddOps(
3477  ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3478  InnerAddOps.append(std::next(J),
3479  ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3480 
3481  // Don't leave just a constant behind in a register if the constant could
3482  // be folded into an immediate field.
3483  if (InnerAddOps.size() == 1 &&
3484  isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3485  LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3486  continue;
3487 
3488  const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3489  if (InnerSum->isZero())
3490  continue;
3491  Formula F = Base;
3492 
3493  // Add the remaining pieces of the add back into the new formula.
3494  const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3495  if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3496  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3497  InnerSumSC->getValue()->getZExtValue())) {
3498  F.UnfoldedOffset =
3499  (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3500  if (IsScaledReg)
3501  F.ScaledReg = nullptr;
3502  else
3503  F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3504  } else if (IsScaledReg)
3505  F.ScaledReg = InnerSum;
3506  else
3507  F.BaseRegs[Idx] = InnerSum;
3508 
3509  // Add J as its own register, or an unfolded immediate.
3510  const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3511  if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3512  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3513  SC->getValue()->getZExtValue()))
3514  F.UnfoldedOffset =
3515  (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3516  else
3517  F.BaseRegs.push_back(*J);
3518  // We may have changed the number of register in base regs, adjust the
3519  // formula accordingly.
3520  F.canonicalize(*L);
3521 
3522  if (InsertFormula(LU, LUIdx, F))
3523  // If that formula hadn't been seen before, recurse to find more like
3524  // it.
3525  GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3526  }
3527 }
3528 
3529 /// Split out subexpressions from adds and the bases of addrecs.
3530 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3531  Formula Base, unsigned Depth) {
3532  assert(Base.isCanonical(*L) && "Input must be in the canonical form");
3533  // Arbitrarily cap recursion to protect compile time.
3534  if (Depth >= 3)
3535  return;
3536 
3537  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3538  GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3539 
3540  if (Base.Scale == 1)
3541  GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3542  /* Idx */ -1, /* IsScaledReg */ true);
3543 }
3544 
3545 /// Generate a formula consisting of all of the loop-dominating registers added
3546 /// into a single register.
3547 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3548  Formula Base) {
3549  // This method is only interesting on a plurality of registers.
3550  if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3551  return;
3552 
3553  // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3554  // processing the formula.
3555  Base.unscale();
3556  Formula F = Base;
3557  F.BaseRegs.clear();
3559  for (const SCEV *BaseReg : Base.BaseRegs) {
3560  if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3561  !SE.hasComputableLoopEvolution(BaseReg, L))
3562  Ops.push_back(BaseReg);
3563  else
3564  F.BaseRegs.push_back(BaseReg);
3565  }
3566  if (Ops.size() > 1) {
3567  const SCEV *Sum = SE.getAddExpr(Ops);
3568  // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3569  // opportunity to fold something. For now, just ignore such cases
3570  // rather than proceed with zero in a register.
3571  if (!Sum->isZero()) {
3572  F.BaseRegs.push_back(Sum);
3573  F.canonicalize(*L);
3574  (void)InsertFormula(LU, LUIdx, F);
3575  }
3576  }
3577 }
3578 
3579 /// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3580 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3581  const Formula &Base, size_t Idx,
3582  bool IsScaledReg) {
3583  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3584  GlobalValue *GV = ExtractSymbol(G, SE);
3585  if (G->isZero() || !GV)
3586  return;
3587  Formula F = Base;
3588  F.BaseGV = GV;
3589  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3590  return;
3591  if (IsScaledReg)
3592  F.ScaledReg = G;
3593  else
3594  F.BaseRegs[Idx] = G;
3595  (void)InsertFormula(LU, LUIdx, F);
3596 }
3597 
3598 /// Generate reuse formulae using symbolic offsets.
3599 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3600  Formula Base) {
3601  // We can't add a symbolic offset if the address already contains one.
3602  if (Base.BaseGV) return;
3603 
3604  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3605  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3606  if (Base.Scale == 1)
3607  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3608  /* IsScaledReg */ true);
3609 }
3610 
3611 /// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3612 void LSRInstance::GenerateConstantOffsetsImpl(
3613  LSRUse &LU, unsigned LUIdx, const Formula &Base,
3614  const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3615  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3616  for (int64_t Offset : Worklist) {
3617  Formula F = Base;
3618  F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3619  if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3620  LU.AccessTy, F)) {
3621  // Add the offset to the base register.
3622  const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3623  // If it cancelled out, drop the base register, otherwise update it.
3624  if (NewG->isZero()) {
3625  if (IsScaledReg) {
3626  F.Scale = 0;
3627  F.ScaledReg = nullptr;
3628  } else
3629  F.deleteBaseReg(F.BaseRegs[Idx]);
3630  F.canonicalize(*L);
3631  } else if (IsScaledReg)
3632  F.ScaledReg = NewG;
3633  else
3634  F.BaseRegs[Idx] = NewG;
3635 
3636  (void)InsertFormula(LU, LUIdx, F);
3637  }
3638  }
3639 
3640  int64_t Imm = ExtractImmediate(G, SE);
3641  if (G->isZero() || Imm == 0)
3642  return;
3643  Formula F = Base;
3644  F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3645  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3646  return;
3647  if (IsScaledReg)
3648  F.ScaledReg = G;
3649  else
3650  F.BaseRegs[Idx] = G;
3651  (void)InsertFormula(LU, LUIdx, F);
3652 }
3653 
3654 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3655 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3656  Formula Base) {
3657  // TODO: For now, just add the min and max offset, because it usually isn't
3658  // worthwhile looking at everything inbetween.
3659  SmallVector<int64_t, 2> Worklist;
3660  Worklist.push_back(LU.MinOffset);
3661  if (LU.MaxOffset != LU.MinOffset)
3662  Worklist.push_back(LU.MaxOffset);
3663 
3664  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3665  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3666  if (Base.Scale == 1)
3667  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3668  /* IsScaledReg */ true);
3669 }
3670 
3671 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3672 /// == y -> x*c == y*c.
3673 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3674  Formula Base) {
3675  if (LU.Kind != LSRUse::ICmpZero) return;
3676 
3677  // Determine the integer type for the base formula.
3678  Type *IntTy = Base.getType();
3679  if (!IntTy) return;
3680  if (SE.getTypeSizeInBits(IntTy) > 64) return;
3681 
3682  // Don't do this if there is more than one offset.
3683  if (LU.MinOffset != LU.MaxOffset) return;
3684 
3685  // Check if transformation is valid. It is illegal to multiply pointer.
3686  if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3687  return;
3688  for (const SCEV *BaseReg : Base.BaseRegs)
3689  if (BaseReg->getType()->isPointerTy())
3690  return;
3691  assert(!Base.BaseGV && "ICmpZero use is not legal!");
3692 
3693  // Check each interesting stride.
3694  for (int64_t Factor : Factors) {
3695  // Check that the multiplication doesn't overflow.
3696  if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3697  continue;
3698  int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3699  if (NewBaseOffset / Factor != Base.BaseOffset)
3700  continue;
3701  // If the offset will be truncated at this use, check that it is in bounds.
3702  if (!IntTy->isPointerTy() &&
3703  !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3704  continue;
3705 
3706  // Check that multiplying with the use offset doesn't overflow.
3707  int64_t Offset = LU.MinOffset;
3708  if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3709  continue;
3710  Offset = (uint64_t)Offset * Factor;
3711  if (Offset / Factor != LU.MinOffset)
3712  continue;
3713  // If the offset will be truncated at this use, check that it is in bounds.
3714  if (!IntTy->isPointerTy() &&
3715  !ConstantInt::isValueValidForType(IntTy, Offset))
3716  continue;
3717 
3718  Formula F = Base;
3719  F.BaseOffset = NewBaseOffset;
3720 
3721  // Check that this scale is legal.
3722  if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3723  continue;
3724 
3725  // Compensate for the use having MinOffset built into it.
3726  F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3727 
3728  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3729 
3730  // Check that multiplying with each base register doesn't overflow.
3731  for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3732  F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3733  if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3734  goto next;
3735  }
3736 
3737  // Check that multiplying with the scaled register doesn't overflow.
3738  if (F.ScaledReg) {
3739  F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3740  if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3741  continue;
3742  }
3743 
3744  // Check that multiplying with the unfolded offset doesn't overflow.
3745  if (F.UnfoldedOffset != 0) {
3746  if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3747  Factor == -1)
3748  continue;
3749  F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3750  if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3751  continue;
3752  // If the offset will be truncated, check that it is in bounds.
3753  if (!IntTy->isPointerTy() &&
3754  !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3755  continue;
3756  }
3757 
3758  // If we make it here and it's legal, add it.
3759  (void)InsertFormula(LU, LUIdx, F);
3760  next:;
3761  }
3762 }
3763 
3764 /// Generate stride factor reuse formulae by making use of scaled-offset address
3765 /// modes, for example.
3766 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3767  // Determine the integer type for the base formula.
3768  Type *IntTy = Base.getType();
3769  if (!IntTy) return;
3770 
3771  // If this Formula already has a scaled register, we can't add another one.
3772  // Try to unscale the formula to generate a better scale.
3773  if (Base.Scale != 0 && !Base.unscale())
3774  return;
3775 
3776  assert(Base.Scale == 0 && "unscale did not did its job!");
3777 
3778  // Check each interesting stride.
3779  for (int64_t Factor : Factors) {
3780  Base.Scale = Factor;
3781  Base.HasBaseReg = Base.BaseRegs.size() > 1;
3782  // Check whether this scale is going to be legal.
3783  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3784  Base)) {
3785  // As a special-case, handle special out-of-loop Basic users specially.
3786  // TODO: Reconsider this special case.
3787  if (LU.Kind == LSRUse::Basic &&
3788  isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3789  LU.AccessTy, Base) &&
3790  LU.AllFixupsOutsideLoop)
3791  LU.Kind = LSRUse::Special;
3792  else
3793  continue;
3794  }
3795  // For an ICmpZero, negating a solitary base register won't lead to
3796  // new solutions.
3797  if (LU.Kind == LSRUse::ICmpZero &&
3798  !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3799  continue;
3800  // For each addrec base reg, if its loop is current loop, apply the scale.
3801  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3802  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3803  if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3804  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3805  if (FactorS->isZero())
3806  continue;
3807  // Divide out the factor, ignoring high bits, since we'll be
3808  // scaling the value back up in the end.
3809  if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3810  // TODO: This could be optimized to avoid all the copying.
3811  Formula F = Base;
3812  F.ScaledReg = Quotient;
3813  F.deleteBaseReg(F.BaseRegs[i]);
3814  // The canonical representation of 1*reg is reg, which is already in
3815  // Base. In that case, do not try to insert the formula, it will be
3816  // rejected anyway.
3817  if (F.Scale == 1 && (F.BaseRegs.empty() ||
3818  (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3819  continue;
3820  // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
3821  // non canonical Formula with ScaledReg's loop not being L.
3822  if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
3823  F.canonicalize(*L);
3824  (void)InsertFormula(LU, LUIdx, F);
3825  }
3826  }
3827  }
3828  }
3829 }
3830 
3831 /// Generate reuse formulae from different IV types.
3832 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3833  // Don't bother truncating symbolic values.
3834  if (Base.BaseGV) return;
3835 
3836  // Determine the integer type for the base formula.
3837  Type *DstTy = Base.getType();
3838  if (!DstTy) return;
3839  DstTy = SE.getEffectiveSCEVType(DstTy);
3840 
3841  for (Type *SrcTy : Types) {
3842  if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3843  Formula F = Base;
3844 
3845  if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
3846  for (const SCEV *&BaseReg : F.BaseRegs)
3847  BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
3848 
3849  // TODO: This assumes we've done basic processing on all uses and
3850  // have an idea what the register usage is.
3851  if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3852  continue;
3853 
3854  F.canonicalize(*L);
3855  (void)InsertFormula(LU, LUIdx, F);
3856  }
3857  }
3858 }
3859 
3860 namespace {
3861 
3862 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
3863 /// modifications so that the search phase doesn't have to worry about the data
3864 /// structures moving underneath it.
3865 struct WorkItem {
3866  size_t LUIdx;
3867  int64_t Imm;
3868  const SCEV *OrigReg;
3869 
3870  WorkItem(size_t LI, int64_t I, const SCEV *R)
3871  : LUIdx(LI), Imm(I), OrigReg(R) {}
3872 
3873  void print(raw_ostream &OS) const;
3874  void dump() const;
3875 };
3876 
3877 } // end anonymous namespace
3878 
3879 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3880 void WorkItem::print(raw_ostream &OS) const {
3881  OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3882  << " , add offset " << Imm;
3883 }
3884 
3885 LLVM_DUMP_METHOD void WorkItem::dump() const {
3886  print(errs()); errs() << '\n';
3887 }
3888 #endif
3889 
3890 /// Look for registers which are a constant distance apart and try to form reuse
3891 /// opportunities between them.
3892 void LSRInstance::GenerateCrossUseConstantOffsets() {
3893  // Group the registers by their value without any added constant offset.
3894  using ImmMapTy = std::map<int64_t, const SCEV *>;
3895 
3897  DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3899  for (const SCEV *Use : RegUses) {
3900  const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
3901  int64_t Imm = ExtractImmediate(Reg, SE);
3902  auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
3903  if (Pair.second)
3904  Sequence.push_back(Reg);
3905  Pair.first->second.insert(std::make_pair(Imm, Use));
3906  UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
3907  }
3908 
3909  // Now examine each set of registers with the same base value. Build up
3910  // a list of work to do and do the work in a separate step so that we're
3911  // not adding formulae and register counts while we're searching.
3912  SmallVector<WorkItem, 32> WorkItems;
3913  SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3914  for (const SCEV *Reg : Sequence) {
3915  const ImmMapTy &Imms = Map.find(Reg)->second;
3916 
3917  // It's not worthwhile looking for reuse if there's only one offset.
3918  if (Imms.size() == 1)
3919  continue;
3920 
3921  DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3922  for (const auto &Entry : Imms)
3923  dbgs() << ' ' << Entry.first;
3924  dbgs() << '\n');
3925 
3926  // Examine each offset.
3927  for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3928  J != JE; ++J) {
3929  const SCEV *OrigReg = J->second;
3930 
3931  int64_t JImm = J->first;
3932  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3933 
3934  if (!isa<SCEVConstant>(OrigReg) &&
3935  UsedByIndicesMap[Reg].count() == 1) {
3936  DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3937  continue;
3938  }
3939 
3940  // Conservatively examine offsets between this orig reg a few selected
3941  // other orig regs.
3942  ImmMapTy::const_iterator OtherImms[] = {
3943  Imms.begin(), std::prev(Imms.end()),
3944  Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3945  2)
3946  };
3947  for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3948  ImmMapTy::const_iterator M = OtherImms[i];
3949  if (M == J || M == JE) continue;
3950 
3951  // Compute the difference between the two.
3952  int64_t Imm = (uint64_t)JImm - M->first;
3953  for (unsigned LUIdx : UsedByIndices.set_bits())
3954  // Make a memo of this use, offset, and register tuple.
3955  if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3956  WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3957  }
3958  }
3959  }
3960 
3961  Map.clear();
3962  Sequence.clear();
3963  UsedByIndicesMap.clear();
3964  UniqueItems.clear();
3965 
3966  // Now iterate through the worklist and add new formulae.
3967  for (const WorkItem &WI : WorkItems) {
3968  size_t LUIdx = WI.LUIdx;
3969  LSRUse &LU = Uses[LUIdx];
3970  int64_t Imm = WI.Imm;
3971  const SCEV *OrigReg = WI.OrigReg;
3972 
3973  Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3974  const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3975  unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3976 
3977  // TODO: Use a more targeted data structure.
3978  for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3979  Formula F = LU.Formulae[L];
3980  // FIXME: The code for the scaled and unscaled registers looks
3981  // very similar but slightly different. Investigate if they
3982  // could be merged. That way, we would not have to unscale the
3983  // Formula.
3984  F.unscale();
3985  // Use the immediate in the scaled register.
3986  if (F.ScaledReg == OrigReg) {
3987  int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3988  // Don't create 50 + reg(-50).
3989  if (F.referencesReg(SE.getSCEV(
3990  ConstantInt::get(IntTy, -(uint64_t)Offset))))
3991  continue;
3992  Formula NewF = F;
3993  NewF.BaseOffset = Offset;
3994  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3995  NewF))
3996  continue;
3997  NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
3998 
3999  // If the new scale is a constant in a register, and adding the constant
4000  // value to the immediate would produce a value closer to zero than the
4001  // immediate itself, then the formula isn't worthwhile.
4002  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4003  if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4004  (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4005  .ule(std::abs(NewF.BaseOffset)))
4006  continue;
4007 
4008  // OK, looks good.
4009  NewF.canonicalize(*this->L);
4010  (void)InsertFormula(LU, LUIdx, NewF);
4011  } else {
4012  // Use the immediate in a base register.
4013  for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4014  const SCEV *BaseReg = F.BaseRegs[N];
4015  if (BaseReg != OrigReg)
4016  continue;
4017  Formula NewF = F;
4018  NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4019  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4020  LU.Kind, LU.AccessTy, NewF)) {
4021  if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4022  continue;
4023  NewF = F;
4024  NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4025  }
4026  NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4027 
4028  // If the new formula has a constant in a register, and adding the
4029  // constant value to the immediate would produce a value closer to
4030  // zero than the immediate itself, then the formula isn't worthwhile.
4031  for (const SCEV *NewReg : NewF.BaseRegs)
4032  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4033  if ((C->getAPInt() + NewF.BaseOffset)
4034  .abs()
4035  .slt(std::abs(NewF.BaseOffset)) &&
4036  (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4037  countTrailingZeros<uint64_t>(NewF.BaseOffset))
4038  goto skip_formula;
4039 
4040  // Ok, looks good.
4041  NewF.canonicalize(*this->L);
4042  (void)InsertFormula(LU, LUIdx, NewF);
4043  break;
4044  skip_formula:;
4045  }
4046  }
4047  }
4048  }
4049 }
4050 
4051 /// Generate formulae for each use.
4052 void
4053 LSRInstance::GenerateAllReuseFormulae() {
4054  // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4055  // queries are more precise.
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  GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4060  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4061  GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4062  }
4063  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4064  LSRUse &LU = Uses[LUIdx];
4065  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4066  GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4067  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4068  GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4069  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4070  GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4071  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4072  GenerateScales(LU, LUIdx, LU.Formulae[i]);
4073  }
4074  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4075  LSRUse &LU = Uses[LUIdx];
4076  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4077  GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4078  }
4079 
4080  GenerateCrossUseConstantOffsets();
4081 
4082  DEBUG(dbgs() << "\n"
4083  "After generating reuse formulae:\n";
4084  print_uses(dbgs()));
4085 }
4086 
4087 /// If there are multiple formulae with the same set of registers used
4088 /// by other uses, pick the best one and delete the others.
4089 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4090  DenseSet<const SCEV *> VisitedRegs;
4093 #ifndef NDEBUG
4094  bool ChangedFormulae = false;
4095 #endif
4096 
4097  // Collect the best formula for each unique set of shared registers. This
4098  // is reset for each use.
4099  using BestFormulaeTy =
4100  DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4101 
4102  BestFormulaeTy BestFormulae;
4103 
4104  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4105  LSRUse &LU = Uses[LUIdx];
4106  DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4107 
4108  bool Any = false;
4109  for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4110  FIdx != NumForms; ++FIdx) {
4111  Formula &F = LU.Formulae[FIdx];
4112 
4113  // Some formulas are instant losers. For example, they may depend on
4114  // nonexistent AddRecs from other loops. These need to be filtered
4115  // immediately, otherwise heuristics could choose them over others leading
4116  // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4117  // avoids the need to recompute this information across formulae using the
4118  // same bad AddRec. Passing LoserRegs is also essential unless we remove
4119  // the corresponding bad register from the Regs set.
4120  Cost CostF;
4121  Regs.clear();
4122  CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, SE, DT, LU, &LoserRegs);
4123  if (CostF.isLoser()) {
4124  // During initial formula generation, undesirable formulae are generated
4125  // by uses within other loops that have some non-trivial address mode or
4126  // use the postinc form of the IV. LSR needs to provide these formulae
4127  // as the basis of rediscovering the desired formula that uses an AddRec
4128  // corresponding to the existing phi. Once all formulae have been
4129  // generated, these initial losers may be pruned.
4130  DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
4131  dbgs() << "\n");
4132  }
4133  else {
4135  for (const SCEV *Reg : F.BaseRegs) {
4136  if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4137  Key.push_back(Reg);
4138  }
4139  if (F.ScaledReg &&
4140  RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4141  Key.push_back(F.ScaledReg);
4142  // Unstable sort by host order ok, because this is only used for
4143  // uniquifying.
4144  std::sort(Key.begin(), Key.end());
4145 
4146  std::pair<BestFormulaeTy::const_iterator, bool> P =
4147  BestFormulae.insert(std::make_pair(Key, FIdx));
4148  if (P.second)
4149  continue;
4150 
4151  Formula &Best = LU.Formulae[P.first->second];
4152 
4153  Cost CostBest;
4154  Regs.clear();
4155  CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, SE, DT, LU);
4156  if (CostF.isLess(CostBest, TTI))
4157  std::swap(F, Best);
4158  DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4159  dbgs() << "\n"
4160  " in favor of formula "; Best.print(dbgs());
4161  dbgs() << '\n');
4162  }
4163 #ifndef NDEBUG
4164  ChangedFormulae = true;
4165 #endif
4166  LU.DeleteFormula(F);
4167  --FIdx;
4168  --NumForms;
4169  Any = true;
4170  }
4171 
4172  // Now that we've filtered out some formulae, recompute the Regs set.
4173  if (Any)
4174  LU.RecomputeRegs(LUIdx, RegUses);
4175 
4176  // Reset this to prepare for the next use.
4177  BestFormulae.clear();
4178  }
4179 
4180  DEBUG(if (ChangedFormulae) {
4181  dbgs() << "\n"
4182  "After filtering out undesirable candidates:\n";
4183  print_uses(dbgs());
4184  });
4185 }
4186 
4187 // This is a rough guess that seems to work fairly well.
4189 
4190 /// Estimate the worst-case number of solutions the solver might have to
4191 /// consider. It almost never considers this many solutions because it prune the
4192 /// search space, but the pruning isn't always sufficient.
4193 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4194  size_t Power = 1;
4195  for (const LSRUse &LU : Uses) {
4196  size_t FSize = LU.Formulae.size();
4197  if (FSize >= ComplexityLimit) {
4198  Power = ComplexityLimit;
4199  break;
4200  }
4201  Power *= FSize;
4202  if (Power >= ComplexityLimit)
4203  break;
4204  }
4205  return Power;
4206 }
4207 
4208 /// When one formula uses a superset of the registers of another formula, it
4209 /// won't help reduce register pressure (though it may not necessarily hurt
4210 /// register pressure); remove it to simplify the system.
4211 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4212  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4213  DEBUG(dbgs() << "The search space is too complex.\n");
4214 
4215  DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4216  "which use a superset of registers used by other "
4217  "formulae.\n");
4218 
4219  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4220  LSRUse &LU = Uses[LUIdx];
4221  bool Any = false;
4222  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4223  Formula &F = LU.Formulae[i];
4224  // Look for a formula with a constant or GV in a register. If the use
4225  // also has a formula with that same value in an immediate field,
4226  // delete the one that uses a register.
4228  I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4229  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4230  Formula NewF = F;
4231  NewF.BaseOffset += C->getValue()->getSExtValue();
4232  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4233  (I - F.BaseRegs.begin()));
4234  if (LU.HasFormulaWithSameRegs(NewF)) {
4235  DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4236  LU.DeleteFormula(F);
4237  --i;
4238  --e;
4239  Any = true;
4240  break;
4241  }
4242  } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4243  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4244  if (!F.BaseGV) {
4245  Formula NewF = F;
4246  NewF.BaseGV = GV;
4247  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4248  (I - F.BaseRegs.begin()));
4249  if (LU.HasFormulaWithSameRegs(NewF)) {
4250  DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4251  dbgs() << '\n');
4252  LU.DeleteFormula(F);
4253  --i;
4254  --e;
4255  Any = true;
4256  break;
4257  }
4258  }
4259  }
4260  }
4261  }
4262  if (Any)
4263  LU.RecomputeRegs(LUIdx, RegUses);
4264  }
4265 
4266  DEBUG(dbgs() << "After pre-selection:\n";
4267  print_uses(dbgs()));
4268  }
4269 }
4270 
4271 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4272 /// allocate a single register for them.
4273 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4274  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4275  return;
4276 
4277  DEBUG(dbgs() << "The search space is too complex.\n"
4278  "Narrowing the search space by assuming that uses separated "
4279  "by a constant offset will use the same registers.\n");
4280 
4281  // This is especially useful for unrolled loops.
4282 
4283  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4284  LSRUse &LU = Uses[LUIdx];
4285  for (const Formula &F : LU.Formulae) {
4286  if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4287  continue;
4288 
4289  LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4290  if (!LUThatHas)
4291  continue;
4292 
4293  if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4294  LU.Kind, LU.AccessTy))
4295  continue;
4296 
4297  DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4298 
4299  LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4300 
4301  // Transfer the fixups of LU to LUThatHas.
4302  for (LSRFixup &Fixup : LU.Fixups) {
4303  Fixup.Offset += F.BaseOffset;
4304  LUThatHas->pushFixup(Fixup);
4305  DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4306  }
4307 
4308  // Delete formulae from the new use which are no longer legal.
4309  bool Any = false;
4310  for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4311  Formula &F = LUThatHas->Formulae[i];
4312  if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4313  LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4314  DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4315  dbgs() << '\n');
4316  LUThatHas->DeleteFormula(F);
4317  --i;
4318  --e;
4319  Any = true;
4320  }
4321  }
4322 
4323  if (Any)
4324  LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4325 
4326  // Delete the old use.
4327  DeleteUse(LU, LUIdx);
4328  --LUIdx;
4329  --NumUses;
4330  break;
4331  }
4332  }
4333 
4334  DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4335 }
4336 
4337 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4338 /// we've done more filtering, as it may be able to find more formulae to
4339 /// eliminate.
4340 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4341  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4342  DEBUG(dbgs() << "The search space is too complex.\n");
4343 
4344  DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4345  "undesirable dedicated registers.\n");
4346 
4347  FilterOutUndesirableDedicatedRegisters();
4348 
4349  DEBUG(dbgs() << "After pre-selection:\n";
4350  print_uses(dbgs()));
4351  }
4352 }
4353 
4354 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4355 /// Pick the best one and delete the others.
4356 /// This narrowing heuristic is to keep as many formulae with different
4357 /// Scale and ScaledReg pair as possible while narrowing the search space.
4358 /// The benefit is that it is more likely to find out a better solution
4359 /// from a formulae set with more Scale and ScaledReg variations than
4360 /// a formulae set with the same Scale and ScaledReg. The picking winner
4361 /// reg heurstic will often keep the formulae with the same Scale and
4362 /// ScaledReg and filter others, and we want to avoid that if possible.
4363 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4364  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4365  return;
4366 
4367  DEBUG(dbgs() << "The search space is too complex.\n"
4368  "Narrowing the search space by choosing the best Formula "
4369  "from the Formulae with the same Scale and ScaledReg.\n");
4370 
4371  // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4372  using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4373 
4374  BestFormulaeTy BestFormulae;
4375 #ifndef NDEBUG
4376  bool ChangedFormulae = false;
4377 #endif
4378  DenseSet<const SCEV *> VisitedRegs;
4380 
4381  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4382  LSRUse &LU = Uses[LUIdx];
4383  DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4384 
4385  // Return true if Formula FA is better than Formula FB.
4386  auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4387  // First we will try to choose the Formula with fewer new registers.
4388  // For a register used by current Formula, the more the register is
4389  // shared among LSRUses, the less we increase the register number
4390  // counter of the formula.
4391  size_t FARegNum = 0;
4392  for (const SCEV *Reg : FA.BaseRegs) {
4393  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4394  FARegNum += (NumUses - UsedByIndices.count() + 1);
4395  }
4396  size_t FBRegNum = 0;
4397  for (const SCEV *Reg : FB.BaseRegs) {
4398  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4399  FBRegNum += (NumUses - UsedByIndices.count() + 1);
4400  }
4401  if (FARegNum != FBRegNum)
4402  return FARegNum < FBRegNum;
4403 
4404  // If the new register numbers are the same, choose the Formula with
4405  // less Cost.
4406  Cost CostFA, CostFB;
4407  Regs.clear();
4408  CostFA.RateFormula(TTI, FA, Regs, VisitedRegs, L, SE, DT, LU);
4409  Regs.clear();
4410  CostFB.RateFormula(TTI, FB, Regs, VisitedRegs, L, SE, DT, LU);
4411  return CostFA.isLess(CostFB, TTI);
4412  };
4413 
4414  bool Any = false;
4415  for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4416  ++FIdx) {
4417  Formula &F = LU.Formulae[FIdx];
4418  if (!F.ScaledReg)
4419  continue;
4420  auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4421  if (P.second)
4422  continue;
4423 
4424  Formula &Best = LU.Formulae[P.first->second];
4425  if (IsBetterThan(F, Best))
4426  std::swap(F, Best);
4427  DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4428  dbgs() << "\n"
4429  " in favor of formula ";
4430  Best.print(dbgs()); dbgs() << '\n');
4431 #ifndef NDEBUG
4432  ChangedFormulae = true;
4433 #endif
4434  LU.DeleteFormula(F);
4435  --FIdx;
4436  --NumForms;
4437  Any = true;
4438  }
4439  if (Any)
4440  LU.RecomputeRegs(LUIdx, RegUses);
4441 
4442  // Reset this to prepare for the next use.
4443  BestFormulae.clear();
4444  }
4445 
4446  DEBUG(if (ChangedFormulae) {
4447  dbgs() << "\n"
4448  "After filtering out undesirable candidates:\n";
4449  print_uses(dbgs());
4450  });
4451 }
4452 
4453 /// The function delete formulas with high registers number expectation.
4454 /// Assuming we don't know the value of each formula (already delete
4455 /// all inefficient), generate probability of not selecting for each
4456 /// register.
4457 /// For example,
4458 /// Use1:
4459 /// reg(a) + reg({0,+,1})
4460 /// reg(a) + reg({-1,+,1}) + 1
4461 /// reg({a,+,1})
4462 /// Use2:
4463 /// reg(b) + reg({0,+,1})
4464 /// reg(b) + reg({-1,+,1}) + 1
4465 /// reg({b,+,1})
4466 /// Use3:
4467 /// reg(c) + reg(b) + reg({0,+,1})
4468 /// reg(c) + reg({b,+,1})
4469 ///
4470 /// Probability of not selecting
4471 /// Use1 Use2 Use3
4472 /// reg(a) (1/3) * 1 * 1
4473 /// reg(b) 1 * (1/3) * (1/2)
4474 /// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4475 /// reg({-1,+,1}) (2/3) * (2/3) * 1
4476 /// reg({a,+,1}) (2/3) * 1 * 1
4477 /// reg({b,+,1}) 1 * (2/3) * (2/3)
4478 /// reg(c) 1 * 1 * 0
4479 ///
4480 /// Now count registers number mathematical expectation for each formula:
4481 /// Note that for each use we exclude probability if not selecting for the use.
4482 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4483 /// probabilty 1/3 of not selecting for Use1).
4484 /// Use1:
4485 /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4486 /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4487 /// reg({a,+,1}) 1
4488 /// Use2:
4489 /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4490 /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4491 /// reg({b,+,1}) 2/3
4492 /// Use3:
4493 /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4494 /// reg(c) + reg({b,+,1}) 1 + 2/3
4495 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4496  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4497  return;
4498  // Ok, we have too many of formulae on our hands to conveniently handle.
4499  // Use a rough heuristic to thin out the list.
4500 
4501  // Set of Regs wich will be 100% used in final solution.
4502  // Used in each formula of a solution (in example above this is reg(c)).
4503  // We can skip them in calculations.
4505  DEBUG(dbgs() << "The search space is too complex.\n");
4506 
4507  // Map each register to probability of not selecting
4509  for (const SCEV *Reg : RegUses) {
4510  if (UniqRegs.count(Reg))
4511  continue;
4512  float PNotSel = 1;
4513  for (const LSRUse &LU : Uses) {
4514  if (!LU.Regs.count(Reg))
4515  continue;
4516  float P = LU.getNotSelectedProbability(Reg);
4517  if (P != 0.0)
4518  PNotSel *= P;
4519  else
4520  UniqRegs.insert(Reg);
4521  }
4522  RegNumMap.insert(std::make_pair(Reg, PNotSel));
4523  }
4524 
4525  DEBUG(dbgs() << "Narrowing the search space by deleting costly formulas\n");
4526 
4527  // Delete formulas where registers number expectation is high.
4528  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4529  LSRUse &LU = Uses[LUIdx];
4530  // If nothing to delete - continue.
4531  if (LU.Formulae.size() < 2)
4532  continue;
4533  // This is temporary solution to test performance. Float should be
4534  // replaced with round independent type (based on integers) to avoid
4535  // different results for different target builds.
4536  float FMinRegNum = LU.Formulae[0].getNumRegs();
4537  float FMinARegNum = LU.Formulae[0].getNumRegs();
4538  size_t MinIdx = 0;
4539  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4540  Formula &F = LU.Formulae[i];
4541  float FRegNum = 0;
4542  float FARegNum = 0;
4543  for (const SCEV *BaseReg : F.BaseRegs) {
4544  if (UniqRegs.count(BaseReg))
4545  continue;
4546  FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4547  if (isa<SCEVAddRecExpr>(BaseReg))
4548  FARegNum +=
4549  RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4550  }
4551  if (const SCEV *ScaledReg = F.ScaledReg) {
4552  if (!UniqRegs.count(ScaledReg)) {
4553  FRegNum +=
4554  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4555  if (isa<SCEVAddRecExpr>(ScaledReg))
4556  FARegNum +=
4557  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4558  }
4559  }
4560  if (FMinRegNum > FRegNum ||
4561  (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4562  FMinRegNum = FRegNum;
4563  FMinARegNum = FARegNum;
4564  MinIdx = i;
4565  }
4566  }
4567  DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());
4568  dbgs() << " with min reg num " << FMinRegNum << '\n');
4569  if (MinIdx != 0)
4570  std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4571  while (LU.Formulae.size() != 1) {
4572  DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());
4573  dbgs() << '\n');
4574  LU.Formulae.pop_back();
4575  }
4576  LU.RecomputeRegs(LUIdx, RegUses);
4577  assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
4578  Formula &F = LU.Formulae[0];
4579  DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n');
4580  // When we choose the formula, the regs become unique.
4581  UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4582  if (F.ScaledReg)
4583  UniqRegs.insert(F.ScaledReg);
4584  }
4585  DEBUG(dbgs() << "After pre-selection:\n";
4586  print_uses(dbgs()));
4587 }
4588 
4589 /// Pick a register which seems likely to be profitable, and then in any use
4590 /// which has any reference to that register, delete all formulae which do not
4591 /// reference that register.
4592 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4593  // With all other options exhausted, loop until the system is simple
4594  // enough to handle.
4596  while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4597  // Ok, we have too many of formulae on our hands to conveniently handle.
4598  // Use a rough heuristic to thin out the list.
4599  DEBUG(dbgs() << "The search space is too complex.\n");
4600 
4601  // Pick the register which is used by the most LSRUses, which is likely
4602  // to be a good reuse register candidate.
4603  const SCEV *Best = nullptr;
4604  unsigned BestNum = 0;
4605  for (const SCEV *Reg : RegUses) {
4606  if (Taken.count(Reg))
4607  continue;
4608  if (!Best) {
4609  Best = Reg;
4610  BestNum = RegUses.getUsedByIndices(Reg).count();
4611  } else {
4612  unsigned Count = RegUses.getUsedByIndices(Reg).count();
4613  if (Count > BestNum) {
4614  Best = Reg;
4615  BestNum = Count;
4616  }
4617  }
4618  }
4619 
4620  DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4621  << " will yield profitable reuse.\n");
4622  Taken.insert(Best);
4623 
4624  // In any use with formulae which references this register, delete formulae
4625  // which don't reference it.
4626  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4627  LSRUse &LU = Uses[LUIdx];
4628  if (!LU.Regs.count(Best)) continue;
4629 
4630  bool Any = false;
4631  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4632  Formula &F = LU.Formulae[i];
4633  if (!F.referencesReg(Best)) {
4634  DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4635  LU.DeleteFormula(F);
4636  --e;
4637  --i;
4638  Any = true;
4639  assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4640  continue;
4641  }
4642  }
4643 
4644  if (Any)
4645  LU.RecomputeRegs(LUIdx, RegUses);
4646  }
4647 
4648  DEBUG(dbgs() << "After pre-selection:\n";
4649  print_uses(dbgs()));
4650  }
4651 }
4652 
4653 /// If there are an extraordinary number of formulae to choose from, use some
4654 /// rough heuristics to prune down the number of formulae. This keeps the main
4655 /// solver from taking an extraordinary amount of time in some worst-case
4656 /// scenarios.
4657 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4658  NarrowSearchSpaceByDetectingSupersets();
4659  NarrowSearchSpaceByCollapsingUnrolledCode();
4660  NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4661  if (FilterSameScaledReg)
4662  NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4663  if (LSRExpNarrow)
4664  NarrowSearchSpaceByDeletingCostlyFormulas();
4665  else
4666  NarrowSearchSpaceByPickingWinnerRegs();
4667 }
4668 
4669 /// This is the recursive solver.
4670 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4671  Cost &SolutionCost,
4673  const Cost &CurCost,
4674  const SmallPtrSet<const SCEV *, 16> &CurRegs,
4675  DenseSet<const SCEV *> &VisitedRegs) const {
4676  // Some ideas:
4677  // - prune more:
4678  // - use more aggressive filtering
4679  // - sort the formula so that the most profitable solutions are found first
4680  // - sort the uses too
4681  // - search faster:
4682  // - don't compute a cost, and then compare. compare while computing a cost
4683  // and bail early.
4684  // - track register sets with SmallBitVector
4685 
4686  const LSRUse &LU = Uses[Workspace.size()];
4687 
4688  // If this use references any register that's already a part of the
4689  // in-progress solution, consider it a requirement that a formula must
4690  // reference that register in order to be considered. This prunes out
4691  // unprofitable searching.
4693  for (const SCEV *S : CurRegs)
4694  if (LU.Regs.count(S))
4695  ReqRegs.insert(S);
4696 
4698  Cost NewCost;
4699  for (const Formula &F : LU.Formulae) {
4700  // Ignore formulae which may not be ideal in terms of register reuse of
4701  // ReqRegs. The formula should use all required registers before
4702  // introducing new ones.
4703  int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4704  for (const SCEV *Reg : ReqRegs) {
4705  if ((F.ScaledReg && F.ScaledReg == Reg) ||
4706  is_contained(F.BaseRegs, Reg)) {
4707  --NumReqRegsToFind;
4708  if (NumReqRegsToFind == 0)
4709  break;
4710  }
4711  }
4712  if (NumReqRegsToFind != 0) {
4713  // If none of the formulae satisfied the required registers, then we could
4714  // clear ReqRegs and try again. Currently, we simply give up in this case.
4715  continue;
4716  }
4717 
4718  // Evaluate the cost of the current formula. If it's already worse than
4719  // the current best, prune the search at that point.
4720  NewCost = CurCost;
4721  NewRegs = CurRegs;
4722  NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, SE, DT, LU);
4723  if (NewCost.isLess(SolutionCost, TTI)) {
4724  Workspace.push_back(&F);
4725  if (Workspace.size() != Uses.size()) {
4726  SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4727  NewRegs, VisitedRegs);
4728  if (F.getNumRegs() == 1 && Workspace.size() == 1)
4729  VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4730  } else {
4731  DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4732  dbgs() << ".\n Regs:";
4733  for (const SCEV *S : NewRegs)
4734  dbgs() << ' ' << *S;
4735  dbgs() << '\n');
4736 
4737  SolutionCost = NewCost;
4738  Solution = Workspace;
4739  }
4740  Workspace.pop_back();
4741  }
4742  }
4743 }
4744 
4745 /// Choose one formula from each use. Return the results in the given Solution
4746 /// vector.
4747 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4749  Cost SolutionCost;
4750  SolutionCost.Lose();
4751  Cost CurCost;
4753  DenseSet<const SCEV *> VisitedRegs;
4754  Workspace.reserve(Uses.size());
4755 
4756  // SolveRecurse does all the work.
4757  SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4758  CurRegs, VisitedRegs);
4759  if (Solution.empty()) {
4760  DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4761  return;
4762  }
4763 
4764  // Ok, we've now made all our decisions.
4765  DEBUG(dbgs() << "\n"
4766  "The chosen solution requires "; SolutionCost.print(dbgs());
4767  dbgs() << ":\n";
4768  for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4769  dbgs() << " ";
4770  Uses[i].print(dbgs());
4771  dbgs() << "\n"
4772  " ";
4773  Solution[i]->print(dbgs());
4774  dbgs() << '\n';
4775  });
4776 
4777  assert(Solution.size() == Uses.size() && "Malformed solution!");
4778 }
4779 
4780 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4781 /// we can go while still being dominated by the input positions. This helps
4782 /// canonicalize the insert position, which encourages sharing.
4784 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4785  const SmallVectorImpl<Instruction *> &Inputs)
4786  const {
4787  Instruction *Tentative = &*IP;
4788  while (true) {
4789  bool AllDominate = true;
4790  Instruction *BetterPos = nullptr;
4791  // Don't bother attempting to insert before a catchswitch, their basic block
4792  // cannot have other non-PHI instructions.
4793  if (isa<CatchSwitchInst>(Tentative))
4794  return IP;
4795 
4796  for (Instruction *Inst : Inputs) {
4797  if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4798  AllDominate = false;
4799  break;
4800  }
4801  // Attempt to find an insert position in the middle of the block,
4802  // instead of at the end, so that it can be used for other expansions.
4803  if (Tentative->getParent() == Inst->getParent() &&
4804  (!BetterPos || !DT.dominates(Inst, BetterPos)))
4805  BetterPos = &*std::next(BasicBlock::iterator(Inst));
4806  }
4807  if (!AllDominate)
4808  break;
4809  if (BetterPos)
4810  IP = BetterPos->getIterator();
4811  else
4812  IP = Tentative->getIterator();
4813 
4814  const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4815  unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4816 
4817  BasicBlock *IDom;
4818  for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4819  if (!Rung) return IP;
4820  Rung = Rung->getIDom();
4821  if (!Rung) return IP;
4822  IDom = Rung->getBlock();
4823 
4824  // Don't climb into a loop though.
4825  const Loop *IDomLoop = LI.getLoopFor(IDom);
4826  unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4827  if (IDomDepth <= IPLoopDepth &&
4828  (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4829  break;
4830  }
4831 
4832  Tentative = IDom->getTerminator();
4833  }
4834 
4835  return IP;
4836 }
4837 
4838 /// Determine an input position which will be dominated by the operands and
4839 /// which will dominate the result.
4841 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4842  const LSRFixup &LF,
4843  const LSRUse &LU,
4844  SCEVExpander &Rewriter) const {
4845  // Collect some instructions which must be dominated by the
4846  // expanding replacement. These must be dominated by any operands that
4847  // will be required in the expansion.
4849  if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4850  Inputs.push_back(I);
4851  if (LU.Kind == LSRUse::ICmpZero)
4852  if (Instruction *I =
4853  dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4854  Inputs.push_back(I);
4855  if (LF.PostIncLoops.count(L)) {
4856  if (LF.isUseFullyOutsideLoop(L))
4857  Inputs.push_back(L->getLoopLatch()->getTerminator());
4858  else
4859  Inputs.push_back(IVIncInsertPos);
4860  }
4861  // The expansion must also be dominated by the increment positions of any
4862  // loops it for which it is using post-inc mode.
4863  for (const Loop *PIL : LF.PostIncLoops) {
4864  if (PIL == L) continue;
4865 
4866  // Be dominated by the loop exit.
4867  SmallVector<BasicBlock *, 4> ExitingBlocks;
4868  PIL->getExitingBlocks(ExitingBlocks);
4869  if (!ExitingBlocks.empty()) {
4870  BasicBlock *BB = ExitingBlocks[0];
4871  for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4872  BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4873  Inputs.push_back(BB->getTerminator());
4874  }
4875  }
4876 
4877  assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
4878  && !isa<DbgInfoIntrinsic>(LowestIP) &&
4879  "Insertion point must be a normal instruction");
4880 
4881  // Then, climb up the immediate dominator tree as far as we can go while
4882  // still being dominated by the input positions.
4883  BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4884 
4885  // Don't insert instructions before PHI nodes.
4886  while (isa<PHINode>(IP)) ++IP;
4887 
4888  // Ignore landingpad instructions.
4889  while (IP->isEHPad()) ++IP;
4890 
4891  // Ignore debug intrinsics.
4892  while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4893 
4894  // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4895  // IP consistent across expansions and allows the previously inserted
4896  // instructions to be reused by subsequent expansion.
4897  while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
4898  ++IP;
4899 
4900  return IP;
4901 }
4902 
4903 /// Emit instructions for the leading candidate expression for this LSRUse (this
4904 /// is called "expanding").
4905 Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
4906  const Formula &F, BasicBlock::iterator IP,
4907  SCEVExpander &Rewriter,
4908  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
4909  if (LU.RigidFormula)
4910  return LF.OperandValToReplace;
4911 
4912  // Determine an input position which will be dominated by the operands and
4913  // which will dominate the result.
4914  IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4915  Rewriter.setInsertPoint(&*IP);
4916 
4917  // Inform the Rewriter if we have a post-increment use, so that it can
4918  // perform an advantageous expansion.
4919  Rewriter.setPostInc(LF.PostIncLoops);
4920 
4921  // This is the type that the user actually needs.
4922  Type *OpTy = LF.OperandValToReplace->getType();
4923  // This will be the type that we'll initially expand to.
4924  Type *Ty = F.getType();
4925  if (!Ty)
4926  // No type known; just expand directly to the ultimate type.
4927  Ty = OpTy;
4928  else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4929  // Expand directly to the ultimate type if it's the right size.
4930  Ty = OpTy;
4931  // This is the type to do integer arithmetic in.
4932  Type *IntTy = SE.getEffectiveSCEVType(Ty);
4933 
4934  // Build up a list of operands to add together to form the full base.
4936 
4937  // Expand the BaseRegs portion.
4938  for (const SCEV *Reg : F.BaseRegs) {
4939  assert(!Reg->isZero() && "Zero allocated in a base register!");
4940 
4941  // If we're expanding for a post-inc user, make the post-inc adjustment.
4942  Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
4943  Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
4944  }
4945 
4946  // Expand the ScaledReg portion.
4947  Value *ICmpScaledV = nullptr;
4948  if (F.Scale != 0) {
4949  const SCEV *ScaledS = F.ScaledReg;
4950 
4951  // If we're expanding for a post-inc user, make the post-inc adjustment.
4952  PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4953  ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
4954 
4955  if (LU.Kind == LSRUse::ICmpZero) {
4956  // Expand ScaleReg as if it was part of the base regs.
4957  if (F.Scale == 1)
4958  Ops.push_back(
4959  SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
4960  else {
4961  // An interesting way of "folding" with an icmp is to use a negated
4962  // scale, which we'll implement by inserting it into the other operand
4963  // of the icmp.
4964  assert(F.Scale == -1 &&
4965  "The only scale supported by ICmpZero uses is -1!");
4966  ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
4967  }
4968  } else {
4969  // Otherwise just expand the scaled register and an explicit scale,
4970  // which is expected to be matched as part of the address.
4971 
4972  // Flush the operand list to suppress SCEVExpander hoisting address modes.
4973  // Unless the addressing mode will not be folded.
4974  if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4975  isAMCompletelyFolded(TTI, LU, F)) {
4976  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4977  Ops.clear();
4978  Ops.push_back(SE.getUnknown(FullV));
4979  }
4980  ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
4981  if (F.Scale != 1)
4982  ScaledS =
4983  SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
4984  Ops.push_back(ScaledS);
4985  }
4986  }
4987 
4988  // Expand the GV portion.
4989  if (F.BaseGV) {
4990  // Flush the operand list to suppress SCEVExpander hoisting.
4991  if (!Ops.empty()) {
4992  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4993  Ops.clear();
4994  Ops.push_back(SE.getUnknown(FullV));
4995  }
4996  Ops.push_back(SE.getUnknown(F.BaseGV));
4997  }
4998 
4999  // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5000  // unfolded offsets. LSR assumes they both live next to their uses.
5001  if (!Ops.empty()) {
5002  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5003  Ops.clear();
5004  Ops.push_back(SE.getUnknown(FullV));
5005  }
5006 
5007  // Expand the immediate portion.
5008  int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5009  if (Offset != 0) {
5010  if (LU.Kind == LSRUse::ICmpZero) {
5011  // The other interesting way of "folding" with an ICmpZero is to use a
5012  // negated immediate.
5013  if (!ICmpScaledV)
5014  ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5015  else {
5016  Ops.push_back(SE.getUnknown(ICmpScaledV));
5017  ICmpScaledV = ConstantInt::get(IntTy, Offset);
5018  }
5019  } else {
5020  // Just add the immediate values. These again are expected to be matched
5021  // as part of the address.
5022  Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5023  }
5024  }
5025 
5026  // Expand the unfolded offset portion.
5027  int64_t UnfoldedOffset = F.UnfoldedOffset;
5028  if (UnfoldedOffset != 0) {
5029  // Just add the immediate values.
5031  UnfoldedOffset)));
5032  }
5033 
5034  // Emit instructions summing all the operands.
5035  const SCEV *FullS = Ops.empty() ?
5036  SE.getConstant(IntTy, 0) :
5037  SE.getAddExpr(Ops);
5038  Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5039 
5040  // We're done expanding now, so reset the rewriter.
5041  Rewriter.clearPostInc();
5042 
5043  // An ICmpZero Formula represents an ICmp which we're handling as a
5044  // comparison against zero. Now that we've expanded an expression for that
5045  // form, update the ICmp's other operand.
5046  if (LU.Kind == LSRUse::ICmpZero) {
5047  ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5048  DeadInsts.emplace_back(CI->getOperand(1));
5049  assert(!F.BaseGV && "ICmp does not support folding a global value and "
5050  "a scale at the same time!");
5051  if (F.Scale == -1) {
5052  if (ICmpScaledV->getType() != OpTy) {
5053  Instruction *Cast =
5054  CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5055  OpTy, false),
5056  ICmpScaledV, OpTy, "tmp", CI);
5057  ICmpScaledV = Cast;
5058  }
5059  CI->setOperand(1, ICmpScaledV);
5060  } else {
5061  // A scale of 1 means that the scale has been expanded as part of the
5062  // base regs.
5063  assert((F.Scale == 0 || F.Scale == 1) &&
5064  "ICmp does not support folding a global value and "
5065  "a scale at the same time!");
5067  -(uint64_t)Offset);
5068  if (C->getType() != OpTy)
5070  OpTy, false),
5071  C, OpTy);
5072 
5073  CI->setOperand(1, C);
5074  }
5075  }
5076 
5077  return FullV;
5078 }
5079 
5080 /// Helper for Rewrite. PHI nodes are special because the use of their operands
5081 /// effectively happens in their predecessor blocks, so the expression may need
5082 /// to be expanded in multiple places.
5083 void LSRInstance::RewriteForPHI(
5084  PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5085  SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5087  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5088  if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5089  BasicBlock *BB = PN->getIncomingBlock(i);
5090 
5091  // If this is a critical edge, split the edge so that we do not insert
5092  // the code on all predecessor/successor paths. We do this unless this
5093  // is the canonical backedge for this loop, which complicates post-inc
5094  // users.
5095  if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5096  !isa<IndirectBrInst>(BB->getTerminator()) &&
5097  !isa<CatchSwitchInst>(BB->getTerminator())) {
5098  BasicBlock *Parent = PN->getParent();
5099  Loop *PNLoop = LI.getLoopFor(Parent);
5100  if (!PNLoop || Parent != PNLoop->getHeader()) {
5101  // Split the critical edge.
5102  BasicBlock *NewBB = nullptr;
5103  if (!Parent->isLandingPad()) {
5104  NewBB = SplitCriticalEdge(BB, Parent,
5106  .setMergeIdenticalEdges()
5107  .setDontDeleteUselessPHIs());
5108  } else {
5110  SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5111  NewBB = NewBBs[0];
5112  }
5113  // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5114  // phi predecessors are identical. The simple thing to do is skip
5115  // splitting in this case rather than complicate the API.
5116  if (NewBB) {
5117  // If PN is outside of the loop and BB is in the loop, we want to
5118  // move the block to be immediately before the PHI block, not
5119  // immediately after BB.
5120  if (L->contains(BB) && !L->contains(PN))
5121  NewBB->moveBefore(PN->getParent());
5122 
5123  // Splitting the edge can reduce the number of PHI entries we have.
5124  e = PN->getNumIncomingValues();
5125  BB = NewBB;
5126  i = PN->getBasicBlockIndex(BB);
5127  }
5128  }
5129  }
5130 
5131  std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5132  Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5133  if (!Pair.second)
5134  PN->setIncomingValue(i, Pair.first->second);
5135  else {
5136  Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5137  Rewriter, DeadInsts);
5138 
5139  // If this is reuse-by-noop-cast, insert the noop cast.
5140  Type *OpTy = LF.OperandValToReplace->getType();
5141  if (FullV->getType() != OpTy)
5142  FullV =
5144  OpTy, false),
5145  FullV, LF.OperandValToReplace->getType(),
5146  "tmp", BB->getTerminator());
5147 
5148  PN->setIncomingValue(i, FullV);
5149  Pair.first->second = FullV;
5150  }
5151  }
5152 }
5153 
5154 /// Emit instructions for the leading candidate expression for this LSRUse (this
5155 /// is called "expanding"), and update the UserInst to reference the newly
5156 /// expanded value.
5157 void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5158  const Formula &F, SCEVExpander &Rewriter,
5159  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5160  // First, find an insertion point that dominates UserInst. For PHI nodes,
5161  // find the nearest block which dominates all the relevant uses.
5162  if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5163  RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5164  } else {
5165  Value *FullV =
5166  Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5167 
5168  // If this is reuse-by-noop-cast, insert the noop cast.
5169  Type *OpTy = LF.OperandValToReplace->getType();
5170  if (FullV->getType() != OpTy) {
5171  Instruction *Cast =
5172  CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5173  FullV, OpTy, "tmp", LF.UserInst);
5174  FullV = Cast;
5175  }
5176 
5177  // Update the user. ICmpZero is handled specially here (for now) because
5178  // Expand may have updated one of the operands of the icmp already, and
5179  // its new value may happen to be equal to LF.OperandValToReplace, in
5180  // which case doing replaceUsesOfWith leads to replacing both operands
5181  // with the same value. TODO: Reorganize this.
5182  if (LU.Kind == LSRUse::ICmpZero)
5183  LF.UserInst->setOperand(0, FullV);
5184  else
5185  LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5186  }
5187 
5188  DeadInsts.emplace_back(LF.OperandValToReplace);
5189 }
5190 
5191 /// Rewrite all the fixup locations with new values, following the chosen
5192 /// solution.
5193 void LSRInstance::ImplementSolution(
5194  const SmallVectorImpl<const Formula *> &Solution) {
5195  // Keep track of instructions we may have made dead, so that
5196  // we can remove them after we are done working.
5198 
5200  "lsr");
5201 #ifndef NDEBUG
5202  Rewriter.setDebugType(DEBUG_TYPE);
5203 #endif
5204  Rewriter.disableCanonicalMode();
5205  Rewriter.enableLSRMode();
5206  Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5207 
5208  // Mark phi nodes that terminate chains so the expander tries to reuse them.
5209  for (const IVChain &Chain : IVChainVec) {
5210  if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5211  Rewriter.setChainedPhi(PN);
5212  }
5213 
5214  // Expand the new value definitions and update the users.
5215  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5216  for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5217  Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5218  Changed = true;
5219  }
5220 
5221  for (const IVChain &Chain : IVChainVec) {
5222  GenerateIVChain(Chain, Rewriter, DeadInsts);
5223  Changed = true;
5224  }
5225  // Clean up after ourselves. This must be done before deleting any
5226  // instructions.
5227  Rewriter.clear();
5228 
5229  Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5230 }
5231 
5232 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5233  DominatorTree &DT, LoopInfo &LI,
5234  const TargetTransformInfo &TTI)
5235  : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L) {
5236  // If LoopSimplify form is not available, stay out of trouble.
5237  if (!L->isLoopSimplifyForm())
5238  return;
5239 
5240  // If there's no interesting work to be done, bail early.
5241  if (IU.empty()) return;
5242 
5243  // If there's too much analysis to be done, bail early. We won't be able to
5244  // model the problem anyway.
5245  unsigned NumUsers = 0;
5246  for (const IVStrideUse &U : IU) {
5247  if (++NumUsers > MaxIVUsers) {
5248  (void)U;
5249  DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U << "\n");
5250  return;
5251  }
5252  // Bail out if we have a PHI on an EHPad that gets a value from a
5253  // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5254  // no good place to stick any instructions.
5255  if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5256  auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5257  if (isa<FuncletPadInst>(FirstNonPHI) ||
5258  isa<CatchSwitchInst>(FirstNonPHI))
5259  for (BasicBlock *PredBB : PN->blocks())
5260  if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5261  return;
5262  }
5263  }
5264 
5265 #ifndef NDEBUG
5266  // All dominating loops must have preheaders, or SCEVExpander may not be able
5267  // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5268  //
5269  // IVUsers analysis should only create users that are dominated by simple loop
5270  // headers. Since this loop should dominate all of its users, its user list
5271  // should be empty if this loop itself is not within a simple loop nest.
5272  for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5273  Rung; Rung = Rung->getIDom()) {
5274  BasicBlock *BB = Rung->getBlock();
5275  const Loop *DomLoop = LI.getLoopFor(BB);
5276  if (DomLoop && DomLoop->getHeader() == BB) {
5277  assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
5278  }
5279  }
5280 #endif // DEBUG
5281 
5282  DEBUG(dbgs() << "\nLSR on loop ";
5283  L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
5284  dbgs() << ":\n");
5285 
5286  // First, perform some low-level loop optimizations.
5287  OptimizeShadowIV();
5288  OptimizeLoopTermCond();
5289 
5290  // If loop preparation eliminates all interesting IV users, bail.
5291  if (IU.empty()) return;
5292 
5293  // Skip nested loops until we can model them better with formulae.
5294  if (!L->empty()) {
5295  DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
5296  return;
5297  }
5298 
5299  // Start collecting data and preparing for the solver.
5300  CollectChains();
5301  CollectInterestingTypesAndFactors();
5302  CollectFixupsAndInitialFormulae();
5303  CollectLoopInvariantFixupsAndFormulae();
5304 
5305  assert(!Uses.empty() && "IVUsers reported at least one use");
5306  DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
5307  print_uses(dbgs()));
5308 
5309  // Now use the reuse data to generate a bunch of interesting ways
5310  // to formulate the values needed for the uses.
5311  GenerateAllReuseFormulae();
5312 
5313  FilterOutUndesirableDedicatedRegisters();
5314  NarrowSearchSpaceUsingHeuristics();
5315 
5317  Solve(Solution);
5318 
5319  // Release memory that is no longer needed.
5320  Factors.clear();
5321  Types.clear();
5322  RegUses.clear();
5323 
5324  if (Solution.empty())
5325  return;
5326 
5327 #ifndef NDEBUG
5328  // Formulae should be legal.
5329  for (const LSRUse &LU : Uses) {
5330  for (const Formula &F : LU.Formulae)
5331  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
5332  F) && "Illegal formula generated!");
5333  };
5334 #endif
5335 
5336  // Now that we've decided what we want, make it so.
5337  ImplementSolution(Solution);
5338 }
5339 
5340 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5341 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5342  if (Factors.empty() && Types.empty()) return;
5343 
5344  OS << "LSR has identified the following interesting factors and types: ";
5345  bool First = true;
5346 
5347  for (int64_t Factor : Factors) {
5348  if (!First) OS << ", ";
5349  First = false;
5350  OS << '*' << Factor;
5351  }
5352 
5353  for (Type *Ty : Types) {
5354  if (!First) OS << ", ";
5355  First = false;
5356  OS << '(' << *Ty << ')';
5357  }
5358  OS << '\n';
5359 }
5360 
5361 void LSRInstance::print_fixups(raw_ostream &OS) const {
5362  OS << "LSR is examining the following fixup sites:\n";
5363  for (const LSRUse &LU : Uses)
5364  for (const LSRFixup &LF : LU.Fixups) {
5365  dbgs() << " ";
5366  LF.print(OS);
5367  OS << '\n';
5368  }
5369 }
5370 
5371 void LSRInstance::print_uses(raw_ostream &OS) const {
5372  OS << "LSR is examining the following uses:\n";
5373  for (const LSRUse &LU : Uses) {
5374  dbgs() << " ";
5375  LU.print(OS);
5376  OS << '\n';
5377  for (const Formula &F : LU.Formulae) {
5378  OS << " ";
5379  F.print(OS);
5380  OS << '\n';
5381  }
5382  }
5383 }
5384 
5385 void LSRInstance::print(raw_ostream &OS) const {
5386  print_factors_and_types(OS);
5387  print_fixups(OS);
5388  print_uses(OS);
5389 }
5390 
5391 LLVM_DUMP_METHOD void LSRInstance::dump() const {
5392  print(errs()); errs() << '\n';
5393 }
5394 #endif
5395 
5396 namespace {
5397 
5398 class LoopStrengthReduce : public LoopPass {
5399 public:
5400  static char ID; // Pass ID, replacement for typeid
5401 
5402  LoopStrengthReduce();
5403 
5404 private:
5405  bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5406  void getAnalysisUsage(AnalysisUsage &AU) const override;
5407 };
5408 
5409 } // end anonymous namespace
5410 
5411 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5412  initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
5413 }
5414 
5415 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5416  // We split critical edges, so we change the CFG. However, we do update
5417  // many analyses if they are around.
5419 
5427  // Requiring LoopSimplify a second time here prevents IVUsers from running
5428  // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5433 }
5434 
5436  DominatorTree &DT, LoopInfo &LI,
5437  const TargetTransformInfo &TTI) {
5438  bool Changed = false;
5439 
5440  // Run the main LSR transformation.
5441  Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5442 
5443  // Remove any extra phis created by processing inner loops.
5444  Changed |= DeleteDeadPHIs(L->getHeader());
5445  if (EnablePhiElim && L->isLoopSimplifyForm()) {
5447  const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5448  SCEVExpander Rewriter(SE, DL, "lsr");
5449 #ifndef NDEBUG
5450  Rewriter.setDebugType(DEBUG_TYPE);
5451 #endif
5452  unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5453  if (numFolded) {
5454  Changed = true;
5456  DeleteDeadPHIs(L->getHeader());
5457  }
5458  }
5459  return Changed;
5460 }
5461 
5462 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5463  if (skipLoop(L))
5464  return false;
5465 
5466  auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5467  auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5468  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5469  auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5470  const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5471  *L->getHeader()->getParent());
5472  return ReduceLoopStrength(L, IU, SE, DT, LI, TTI);
5473 }
5474 
5475 PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
5477  LPMUpdater &) {
5478  if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5479  AR.DT, AR.LI, AR.TTI))
5480  return PreservedAnalyses::all();
5481 
5483 }
5484 
5485 char LoopStrengthReduce::ID = 0;
5486 
5487 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5488  "Loop Strength Reduction", false, false)
5494 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5495 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5496  "Loop Strength Reduction", false, false)
5497 
5498 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:244
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
iterator_range< use_iterator > uses()
Definition: Value.h:356
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:157
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:235
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
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:92
This provides a very simple, boring adaptor for a begin and end iterator into a range type...
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:106
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:878
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:767
void resize(unsigned N, bool t=false)
Grow or shrink the bitvector.
static void dump(StringRef Title, SpillInfo const &Spills)
Definition: CoroFrame.cpp:283
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...
F(f)
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
This defines the Use class.
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:344
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
SI optimize exec mask operations pre RA
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:191
#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:951
static cl::opt< bool > InsnsCost("lsr-insns-cost", cl::Hidden, cl::init(true), cl::desc("Add instruction count to a LSR cost model"))
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:659
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...