LLVM  6.0.0svn
LoopStrengthReduce.cpp
Go to the documentation of this file.
1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
12 // on the target.
13 //
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
19 //
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
23 //
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
25 // ...
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
28 //
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
35 //
36 // TODO: More sophistication in the way Formulae are generated and filtered.
37 //
38 // TODO: Handle multiple loops at a time.
39 //
40 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
41 // of a GlobalValue?
42 //
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
45 //
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
52 // use kind.
53 //
54 //===----------------------------------------------------------------------===//
55 
57 #include "llvm/ADT/APInt.h"
58 #include "llvm/ADT/DenseMap.h"
59 #include "llvm/ADT/DenseSet.h"
60 #include "llvm/ADT/Hashing.h"
62 #include "llvm/ADT/STLExtras.h"
63 #include "llvm/ADT/SetVector.h"
65 #include "llvm/ADT/SmallPtrSet.h"
66 #include "llvm/ADT/SmallSet.h"
67 #include "llvm/ADT/SmallVector.h"
68 #include "llvm/Analysis/IVUsers.h"
69 #include "llvm/Analysis/LoopInfo.h"
70 #include "llvm/Analysis/LoopPass.h"
76 #include "llvm/IR/BasicBlock.h"
77 #include "llvm/IR/Constant.h"
78 #include "llvm/IR/Constants.h"
79 #include "llvm/IR/DerivedTypes.h"
80 #include "llvm/IR/Dominators.h"
81 #include "llvm/IR/GlobalValue.h"
82 #include "llvm/IR/IRBuilder.h"
83 #include "llvm/IR/Instruction.h"
84 #include "llvm/IR/Instructions.h"
85 #include "llvm/IR/IntrinsicInst.h"
86 #include "llvm/IR/Module.h"
87 #include "llvm/IR/OperandTraits.h"
88 #include "llvm/IR/Operator.h"
89 #include "llvm/IR/Type.h"
90 #include "llvm/IR/Value.h"
91 #include "llvm/IR/ValueHandle.h"
92 #include "llvm/Pass.h"
93 #include "llvm/Support/Casting.h"
95 #include "llvm/Support/Compiler.h"
96 #include "llvm/Support/Debug.h"
100 #include "llvm/Transforms/Scalar.h"
104 #include <algorithm>
105 #include <cassert>
106 #include <cstddef>
107 #include <cstdint>
108 #include <cstdlib>
109 #include <iterator>
110 #include <map>
111 #include <tuple>
112 #include <utility>
113 
114 using namespace llvm;
115 
116 #define DEBUG_TYPE "loop-reduce"
117 
118 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
119 /// bail out. This threshold is far beyond the number of users that LSR can
120 /// conceivably solve, so it should not affect generated code, but catches the
121 /// worst cases before LSR burns too much compile time and stack space.
122 static const unsigned MaxIVUsers = 200;
123 
124 // Temporary flag to cleanup congruent phis after LSR phi expansion.
125 // It's currently disabled until we can determine whether it's truly useful or
126 // not. The flag should be removed after the v3.0 release.
127 // This is now needed for ivchains.
129  "enable-lsr-phielim", cl::Hidden, cl::init(true),
130  cl::desc("Enable LSR phi elimination"));
131 
132 // The flag adds instruction count to solutions cost comparision.
133 static cl::opt<bool> InsnsCost(
134  "lsr-insns-cost", cl::Hidden, cl::init(true),
135  cl::desc("Add instruction count to a LSR cost model"));
136 
137 // Flag to choose how to narrow complex lsr solution
139  "lsr-exp-narrow", cl::Hidden, cl::init(false),
140  cl::desc("Narrow LSR complex solution using"
141  " expectation of registers number"));
142 
143 // Flag to narrow search space by filtering non-optimal formulae with
144 // the same ScaledReg and Scale.
146  "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
147  cl::desc("Narrow LSR search space by filtering non-optimal formulae"
148  " with the same ScaledReg and Scale"));
149 
150 #ifndef NDEBUG
151 // Stress test IV chain generation.
153  "stress-ivchain", cl::Hidden, cl::init(false),
154  cl::desc("Stress test LSR IV chains"));
155 #else
156 static bool StressIVChain = false;
157 #endif
158 
159 namespace {
160 
161 struct MemAccessTy {
162  /// Used in situations where the accessed memory type is unknown.
163  static const unsigned UnknownAddressSpace = ~0u;
164 
165  Type *MemTy;
166  unsigned AddrSpace;
167 
168  MemAccessTy() : MemTy(nullptr), AddrSpace(UnknownAddressSpace) {}
169 
170  MemAccessTy(Type *Ty, unsigned AS) :
171  MemTy(Ty), AddrSpace(AS) {}
172 
173  bool operator==(MemAccessTy Other) const {
174  return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
175  }
176 
177  bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
178 
179  static MemAccessTy getUnknown(LLVMContext &Ctx,
180  unsigned AS = UnknownAddressSpace) {
181  return MemAccessTy(Type::getVoidTy(Ctx), AS);
182  }
183 };
184 
185 /// This class holds data which is used to order reuse candidates.
186 class RegSortData {
187 public:
188  /// This represents the set of LSRUse indices which reference
189  /// a particular register.
190  SmallBitVector UsedByIndices;
191 
192  void print(raw_ostream &OS) const;
193  void dump() const;
194 };
195 
196 } // end anonymous namespace
197 
198 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
199 void RegSortData::print(raw_ostream &OS) const {
200  OS << "[NumUses=" << UsedByIndices.count() << ']';
201 }
202 
203 LLVM_DUMP_METHOD void RegSortData::dump() const {
204  print(errs()); errs() << '\n';
205 }
206 #endif
207 
208 namespace {
209 
210 /// Map register candidates to information about how they are used.
211 class RegUseTracker {
212  typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
213 
214  RegUsesTy RegUsesMap;
216 
217 public:
218  void countRegister(const SCEV *Reg, size_t LUIdx);
219  void dropRegister(const SCEV *Reg, size_t LUIdx);
220  void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
221 
222  bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
223 
224  const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
225 
226  void clear();
227 
229  typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
230  iterator begin() { return RegSequence.begin(); }
231  iterator end() { return RegSequence.end(); }
232  const_iterator begin() const { return RegSequence.begin(); }
233  const_iterator end() const { return RegSequence.end(); }
234 };
235 
236 } // end anonymous namespace
237 
238 void
239 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
240  std::pair<RegUsesTy::iterator, bool> Pair =
241  RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
242  RegSortData &RSD = Pair.first->second;
243  if (Pair.second)
244  RegSequence.push_back(Reg);
245  RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
246  RSD.UsedByIndices.set(LUIdx);
247 }
248 
249 void
250 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
251  RegUsesTy::iterator It = RegUsesMap.find(Reg);
252  assert(It != RegUsesMap.end());
253  RegSortData &RSD = It->second;
254  assert(RSD.UsedByIndices.size() > LUIdx);
255  RSD.UsedByIndices.reset(LUIdx);
256 }
257 
258 void
259 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
260  assert(LUIdx <= LastLUIdx);
261 
262  // Update RegUses. The data structure is not optimized for this purpose;
263  // we must iterate through it and update each of the bit vectors.
264  for (auto &Pair : RegUsesMap) {
265  SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
266  if (LUIdx < UsedByIndices.size())
267  UsedByIndices[LUIdx] =
268  LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
269  UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
270  }
271 }
272 
273 bool
274 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
275  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
276  if (I == RegUsesMap.end())
277  return false;
278  const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
279  int i = UsedByIndices.find_first();
280  if (i == -1) return false;
281  if ((size_t)i != LUIdx) return true;
282  return UsedByIndices.find_next(i) != -1;
283 }
284 
285 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
286  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
287  assert(I != RegUsesMap.end() && "Unknown register!");
288  return I->second.UsedByIndices;
289 }
290 
291 void RegUseTracker::clear() {
292  RegUsesMap.clear();
293  RegSequence.clear();
294 }
295 
296 namespace {
297 
298 /// This class holds information that describes a formula for computing
299 /// satisfying a use. It may include broken-out immediates and scaled registers.
300 struct Formula {
301  /// Global base address used for complex addressing.
302  GlobalValue *BaseGV;
303 
304  /// Base offset for complex addressing.
305  int64_t BaseOffset;
306 
307  /// Whether any complex addressing has a base register.
308  bool HasBaseReg;
309 
310  /// The scale of any complex addressing.
311  int64_t Scale;
312 
313  /// The list of "base" registers for this use. When this is non-empty. The
314  /// canonical representation of a formula is
315  /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
316  /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
317  /// 3. The reg containing recurrent expr related with currect loop in the
318  /// formula should be put in the ScaledReg.
319  /// #1 enforces that the scaled register is always used when at least two
320  /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
321  /// #2 enforces that 1 * reg is reg.
322  /// #3 ensures invariant regs with respect to current loop can be combined
323  /// together in LSR codegen.
324  /// This invariant can be temporarly broken while building a formula.
325  /// However, every formula inserted into the LSRInstance must be in canonical
326  /// form.
328 
329  /// The 'scaled' register for this use. This should be non-null when Scale is
330  /// not zero.
331  const SCEV *ScaledReg;
332 
333  /// An additional constant offset which added near the use. This requires a
334  /// temporary register, but the offset itself can live in an add immediate
335  /// field rather than a register.
336  int64_t UnfoldedOffset;
337 
338  Formula()
339  : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0),
340  ScaledReg(nullptr), UnfoldedOffset(0) {}
341 
342  void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
343 
344  bool isCanonical(const Loop &L) const;
345 
346  void canonicalize(const Loop &L);
347 
348  bool unscale();
349 
350  bool hasZeroEnd() const;
351 
352  size_t getNumRegs() const;
353  Type *getType() const;
354 
355  void deleteBaseReg(const SCEV *&S);
356 
357  bool referencesReg(const SCEV *S) const;
358  bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
359  const RegUseTracker &RegUses) const;
360 
361  void print(raw_ostream &OS) const;
362  void dump() const;
363 };
364 
365 } // end anonymous namespace
366 
367 /// Recursion helper for initialMatch.
368 static void DoInitialMatch(const SCEV *S, Loop *L,
371  ScalarEvolution &SE) {
372  // Collect expressions which properly dominate the loop header.
373  if (SE.properlyDominates(S, L->getHeader())) {
374  Good.push_back(S);
375  return;
376  }
377 
378  // Look at add operands.
379  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
380  for (const SCEV *S : Add->operands())
381  DoInitialMatch(S, L, Good, Bad, SE);
382  return;
383  }
384 
385  // Look at addrec operands.
386  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
387  if (!AR->getStart()->isZero() && AR->isAffine()) {
388  DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
389  DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
390  AR->getStepRecurrence(SE),
391  // FIXME: AR->getNoWrapFlags()
392  AR->getLoop(), SCEV::FlagAnyWrap),
393  L, Good, Bad, SE);
394  return;
395  }
396 
397  // Handle a multiplication by -1 (negation) if it didn't fold.
398  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
399  if (Mul->getOperand(0)->isAllOnesValue()) {
400  SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
401  const SCEV *NewMul = SE.getMulExpr(Ops);
402 
405  DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
406  const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
407  SE.getEffectiveSCEVType(NewMul->getType())));
408  for (const SCEV *S : MyGood)
409  Good.push_back(SE.getMulExpr(NegOne, S));
410  for (const SCEV *S : MyBad)
411  Bad.push_back(SE.getMulExpr(NegOne, S));
412  return;
413  }
414 
415  // Ok, we can't do anything interesting. Just stuff the whole thing into a
416  // register and hope for the best.
417  Bad.push_back(S);
418 }
419 
420 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
421 /// all loop-invariant and loop-computable values in a single base register.
422 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
425  DoInitialMatch(S, L, Good, Bad, SE);
426  if (!Good.empty()) {
427  const SCEV *Sum = SE.getAddExpr(Good);
428  if (!Sum->isZero())
429  BaseRegs.push_back(Sum);
430  HasBaseReg = true;
431  }
432  if (!Bad.empty()) {
433  const SCEV *Sum = SE.getAddExpr(Bad);
434  if (!Sum->isZero())
435  BaseRegs.push_back(Sum);
436  HasBaseReg = true;
437  }
438  canonicalize(*L);
439 }
440 
441 /// \brief Check whether or not this formula statisfies the canonical
442 /// representation.
443 /// \see Formula::BaseRegs.
444 bool Formula::isCanonical(const Loop &L) const {
445  if (!ScaledReg)
446  return BaseRegs.size() <= 1;
447 
448  if (Scale != 1)
449  return true;
450 
451  if (Scale == 1 && BaseRegs.empty())
452  return false;
453 
454  const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
455  if (SAR && SAR->getLoop() == &L)
456  return true;
457 
458  // If ScaledReg is not a recurrent expr, or it is but its loop is not current
459  // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
460  // loop, we want to swap the reg in BaseRegs with ScaledReg.
461  auto I =
462  find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
463  return isa<const SCEVAddRecExpr>(S) &&
464  (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
465  });
466  return I == BaseRegs.end();
467 }
468 
469 /// \brief Helper method to morph a formula into its canonical representation.
470 /// \see Formula::BaseRegs.
471 /// Every formula having more than one base register, must use the ScaledReg
472 /// field. Otherwise, we would have to do special cases everywhere in LSR
473 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
474 /// On the other hand, 1*reg should be canonicalized into reg.
475 void Formula::canonicalize(const Loop &L) {
476  if (isCanonical(L))
477  return;
478  // So far we did not need this case. This is easy to implement but it is
479  // useless to maintain dead code. Beside it could hurt compile time.
480  assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
481 
482  // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
483  if (!ScaledReg) {
484  ScaledReg = BaseRegs.back();
485  BaseRegs.pop_back();
486  Scale = 1;
487  }
488 
489  // If ScaledReg is an invariant with respect to L, find the reg from
490  // BaseRegs containing the recurrent expr related with Loop L. Swap the
491  // reg with ScaledReg.
492  const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
493  if (!SAR || SAR->getLoop() != &L) {
494  auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
495  [&](const SCEV *S) {
496  return isa<const SCEVAddRecExpr>(S) &&
497  (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
498  });
499  if (I != BaseRegs.end())
500  std::swap(ScaledReg, *I);
501  }
502 }
503 
504 /// \brief Get rid of the scale in the formula.
505 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
506 /// \return true if it was possible to get rid of the scale, false otherwise.
507 /// \note After this operation the formula may not be in the canonical form.
508 bool Formula::unscale() {
509  if (Scale != 1)
510  return false;
511  Scale = 0;
512  BaseRegs.push_back(ScaledReg);
513  ScaledReg = nullptr;
514  return true;
515 }
516 
517 bool Formula::hasZeroEnd() const {
518  if (UnfoldedOffset || BaseOffset)
519  return false;
520  if (BaseRegs.size() != 1 || ScaledReg)
521  return false;
522  return true;
523 }
524 
525 /// Return the total number of register operands used by this formula. This does
526 /// not include register uses implied by non-constant addrec strides.
527 size_t Formula::getNumRegs() const {
528  return !!ScaledReg + BaseRegs.size();
529 }
530 
531 /// Return the type of this formula, if it has one, or null otherwise. This type
532 /// is meaningless except for the bit size.
533 Type *Formula::getType() const {
534  return !BaseRegs.empty() ? BaseRegs.front()->getType() :
535  ScaledReg ? ScaledReg->getType() :
536  BaseGV ? BaseGV->getType() :
537  nullptr;
538 }
539 
540 /// Delete the given base reg from the BaseRegs list.
541 void Formula::deleteBaseReg(const SCEV *&S) {
542  if (&S != &BaseRegs.back())
543  std::swap(S, BaseRegs.back());
544  BaseRegs.pop_back();
545 }
546 
547 /// Test if this formula references the given register.
548 bool Formula::referencesReg(const SCEV *S) const {
549  return S == ScaledReg || is_contained(BaseRegs, S);
550 }
551 
552 /// Test whether this formula uses registers which are used by uses other than
553 /// the use with the given index.
554 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
555  const RegUseTracker &RegUses) const {
556  if (ScaledReg)
557  if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
558  return true;
559  for (const SCEV *BaseReg : BaseRegs)
560  if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
561  return true;
562  return false;
563 }
564 
565 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
566 void Formula::print(raw_ostream &OS) const {
567  bool First = true;
568  if (BaseGV) {
569  if (!First) OS << " + "; else First = false;
570  BaseGV->printAsOperand(OS, /*PrintType=*/false);
571  }
572  if (BaseOffset != 0) {
573  if (!First) OS << " + "; else First = false;
574  OS << BaseOffset;
575  }
576  for (const SCEV *BaseReg : BaseRegs) {
577  if (!First) OS << " + "; else First = false;
578  OS << "reg(" << *BaseReg << ')';
579  }
580  if (HasBaseReg && BaseRegs.empty()) {
581  if (!First) OS << " + "; else First = false;
582  OS << "**error: HasBaseReg**";
583  } else if (!HasBaseReg && !BaseRegs.empty()) {
584  if (!First) OS << " + "; else First = false;
585  OS << "**error: !HasBaseReg**";
586  }
587  if (Scale != 0) {
588  if (!First) OS << " + "; else First = false;
589  OS << Scale << "*reg(";
590  if (ScaledReg)
591  OS << *ScaledReg;
592  else
593  OS << "<unknown>";
594  OS << ')';
595  }
596  if (UnfoldedOffset != 0) {
597  if (!First) OS << " + ";
598  OS << "imm(" << UnfoldedOffset << ')';
599  }
600 }
601 
602 LLVM_DUMP_METHOD void Formula::dump() const {
603  print(errs()); errs() << '\n';
604 }
605 #endif
606 
607 /// Return true if the given addrec can be sign-extended without changing its
608 /// value.
609 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
610  Type *WideTy =
612  return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
613 }
614 
615 /// Return true if the given add can be sign-extended without changing its
616 /// value.
617 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
618  Type *WideTy =
620  return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
621 }
622 
623 /// Return true if the given mul can be sign-extended without changing its
624 /// value.
625 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
626  Type *WideTy =
628  SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
629  return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
630 }
631 
632 /// Return an expression for LHS /s RHS, if it can be determined and if the
633 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
634 /// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
635 /// the multiplication may overflow, which is useful when the result will be
636 /// used in a context where the most significant bits are ignored.
637 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
638  ScalarEvolution &SE,
639  bool IgnoreSignificantBits = false) {
640  // Handle the trivial case, which works for any SCEV type.
641  if (LHS == RHS)
642  return SE.getConstant(LHS->getType(), 1);
643 
644  // Handle a few RHS special cases.
645  const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
646  if (RC) {
647  const APInt &RA = RC->getAPInt();
648  // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
649  // some folding.
650  if (RA.isAllOnesValue())
651  return SE.getMulExpr(LHS, RC);
652  // Handle x /s 1 as x.
653  if (RA == 1)
654  return LHS;
655  }
656 
657  // Check for a division of a constant by a constant.
658  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
659  if (!RC)
660  return nullptr;
661  const APInt &LA = C->getAPInt();
662  const APInt &RA = RC->getAPInt();
663  if (LA.srem(RA) != 0)
664  return nullptr;
665  return SE.getConstant(LA.sdiv(RA));
666  }
667 
668  // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
669  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
670  if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
671  const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
672  IgnoreSignificantBits);
673  if (!Step) return nullptr;
674  const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
675  IgnoreSignificantBits);
676  if (!Start) return nullptr;
677  // FlagNW is independent of the start value, step direction, and is
678  // preserved with smaller magnitude steps.
679  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
680  return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
681  }
682  return nullptr;
683  }
684 
685  // Distribute the sdiv over add operands, if the add doesn't overflow.
686  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
687  if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
689  for (const SCEV *S : Add->operands()) {
690  const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
691  if (!Op) return nullptr;
692  Ops.push_back(Op);
693  }
694  return SE.getAddExpr(Ops);
695  }
696  return nullptr;
697  }
698 
699  // Check for a multiply operand that we can pull RHS out of.
700  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
701  if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
703  bool Found = false;
704  for (const SCEV *S : Mul->operands()) {
705  if (!Found)
706  if (const SCEV *Q = getExactSDiv(S, RHS, SE,
707  IgnoreSignificantBits)) {
708  S = Q;
709  Found = true;
710  }
711  Ops.push_back(S);
712  }
713  return Found ? SE.getMulExpr(Ops) : nullptr;
714  }
715  return nullptr;
716  }
717 
718  // Otherwise we don't know.
719  return nullptr;
720 }
721 
722 /// If S involves the addition of a constant integer value, return that integer
723 /// value, and mutate S to point to a new SCEV with that value excluded.
724 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
725  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
726  if (C->getAPInt().getMinSignedBits() <= 64) {
727  S = SE.getConstant(C->getType(), 0);
728  return C->getValue()->getSExtValue();
729  }
730  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
731  SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
732  int64_t Result = ExtractImmediate(NewOps.front(), SE);
733  if (Result != 0)
734  S = SE.getAddExpr(NewOps);
735  return Result;
736  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
737  SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
738  int64_t Result = ExtractImmediate(NewOps.front(), SE);
739  if (Result != 0)
740  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
741  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
743  return Result;
744  }
745  return 0;
746 }
747 
748 /// If S involves the addition of a GlobalValue address, return that symbol, and
749 /// mutate S to point to a new SCEV with that value excluded.
751  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
752  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
753  S = SE.getConstant(GV->getType(), 0);
754  return GV;
755  }
756  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
757  SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
758  GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
759  if (Result)
760  S = SE.getAddExpr(NewOps);
761  return Result;
762  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
763  SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
764  GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
765  if (Result)
766  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
767  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
769  return Result;
770  }
771  return nullptr;
772 }
773 
774 /// Returns true if the specified instruction is using the specified value as an
775 /// address.
776 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
777  bool isAddress = isa<LoadInst>(Inst);
778  if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
779  if (SI->getPointerOperand() == OperandVal)
780  isAddress = true;
781  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
782  // Addressing modes can also be folded into prefetches and a variety
783  // of intrinsics.
784  switch (II->getIntrinsicID()) {
785  default: break;
786  case Intrinsic::memset:
787  case Intrinsic::prefetch:
788  if (II->getArgOperand(0) == OperandVal)
789  isAddress = true;
790  break;
791  case Intrinsic::memmove:
792  case Intrinsic::memcpy:
793  if (II->getArgOperand(0) == OperandVal ||
794  II->getArgOperand(1) == OperandVal)
795  isAddress = true;
796  break;
797  }
798  } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
799  if (RMW->getPointerOperand() == OperandVal)
800  isAddress = true;
801  } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
802  if (CmpX->getPointerOperand() == OperandVal)
803  isAddress = true;
804  }
805  return isAddress;
806 }
807 
808 /// Return the type of the memory being accessed.
809 static MemAccessTy getAccessType(const Instruction *Inst) {
810  MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
811  if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
812  AccessTy.MemTy = SI->getOperand(0)->getType();
813  AccessTy.AddrSpace = SI->getPointerAddressSpace();
814  } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
815  AccessTy.AddrSpace = LI->getPointerAddressSpace();
816  } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
817  AccessTy.AddrSpace = RMW->getPointerAddressSpace();
818  } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
819  AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
820  }
821 
822  // All pointers have the same requirements, so canonicalize them to an
823  // arbitrary pointer type to minimize variation.
824  if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
825  AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
826  PTy->getAddressSpace());
827 
828  return AccessTy;
829 }
830 
831 /// Return true if this AddRec is already a phi in its loop.
832 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
833  for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
834  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
835  if (SE.isSCEVable(PN->getType()) &&
836  (SE.getEffectiveSCEVType(PN->getType()) ==
837  SE.getEffectiveSCEVType(AR->getType())) &&
838  SE.getSCEV(PN) == AR)
839  return true;
840  }
841  return false;
842 }
843 
844 /// Check if expanding this expression is likely to incur significant cost. This
845 /// is tricky because SCEV doesn't track which expressions are actually computed
846 /// by the current IR.
847 ///
848 /// We currently allow expansion of IV increments that involve adds,
849 /// multiplication by constants, and AddRecs from existing phis.
850 ///
851 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
852 /// obvious multiple of the UDivExpr.
853 static bool isHighCostExpansion(const SCEV *S,
854  SmallPtrSetImpl<const SCEV*> &Processed,
855  ScalarEvolution &SE) {
856  // Zero/One operand expressions
857  switch (S->getSCEVType()) {
858  case scUnknown:
859  case scConstant:
860  return false;
861  case scTruncate:
862  return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
863  Processed, SE);
864  case scZeroExtend:
865  return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
866  Processed, SE);
867  case scSignExtend:
868  return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
869  Processed, SE);
870  }
871 
872  if (!Processed.insert(S).second)
873  return false;
874 
875  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
876  for (const SCEV *S : Add->operands()) {
877  if (isHighCostExpansion(S, Processed, SE))
878  return true;
879  }
880  return false;
881  }
882 
883  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
884  if (Mul->getNumOperands() == 2) {
885  // Multiplication by a constant is ok
886  if (isa<SCEVConstant>(Mul->getOperand(0)))
887  return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
888 
889  // If we have the value of one operand, check if an existing
890  // multiplication already generates this expression.
891  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
892  Value *UVal = U->getValue();
893  for (User *UR : UVal->users()) {
894  // If U is a constant, it may be used by a ConstantExpr.
895  Instruction *UI = dyn_cast<Instruction>(UR);
896  if (UI && UI->getOpcode() == Instruction::Mul &&
897  SE.isSCEVable(UI->getType())) {
898  return SE.getSCEV(UI) == Mul;
899  }
900  }
901  }
902  }
903  }
904 
905  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
906  if (isExistingPhi(AR, SE))
907  return false;
908  }
909 
910  // Fow now, consider any other type of expression (div/mul/min/max) high cost.
911  return true;
912 }
913 
914 /// If any of the instructions is the specified set are trivially dead, delete
915 /// them and see if this makes any of their operands subsequently dead.
916 static bool
918  bool Changed = false;
919 
920  while (!DeadInsts.empty()) {
921  Value *V = DeadInsts.pop_back_val();
922  Instruction *I = dyn_cast_or_null<Instruction>(V);
923 
924  if (!I || !isInstructionTriviallyDead(I))
925  continue;
926 
927  for (Use &O : I->operands())
928  if (Instruction *U = dyn_cast<Instruction>(O)) {
929  O = nullptr;
930  if (U->use_empty())
931  DeadInsts.emplace_back(U);
932  }
933 
934  I->eraseFromParent();
935  Changed = true;
936  }
937 
938  return Changed;
939 }
940 
941 namespace {
942 
943 class LSRUse;
944 
945 } // end anonymous namespace
946 
947 /// \brief Check if the addressing mode defined by \p F is completely
948 /// folded in \p LU at isel time.
949 /// This includes address-mode folding and special icmp tricks.
950 /// This function returns true if \p LU can accommodate what \p F
951 /// defines and up to 1 base + 1 scaled + offset.
952 /// In other words, if \p F has several base registers, this function may
953 /// still return true. Therefore, users still need to account for
954 /// additional base registers and/or unfolded offsets to derive an
955 /// accurate cost model.
956 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
957  const LSRUse &LU, const Formula &F);
958 // Get the cost of the scaling factor used in F for LU.
959 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
960  const LSRUse &LU, const Formula &F,
961  const Loop &L);
962 
963 namespace {
964 
965 /// This class is used to measure and compare candidate formulae.
966 class Cost {
968 
969 public:
970  Cost() {
971  C.Insns = 0;
972  C.NumRegs = 0;
973  C.AddRecCost = 0;
974  C.NumIVMuls = 0;
975  C.NumBaseAdds = 0;
976  C.ImmCost = 0;
977  C.SetupCost = 0;
978  C.ScaleCost = 0;
979  }
980 
981  bool isLess(Cost &Other, const TargetTransformInfo &TTI);
982 
983  void Lose();
984 
985 #ifndef NDEBUG
986  // Once any of the metrics loses, they must all remain losers.
987  bool isValid() {
988  return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
989  | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
990  || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
991  & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
992  }
993 #endif
994 
995  bool isLoser() {
996  assert(isValid() && "invalid cost");
997  return C.NumRegs == ~0u;
998  }
999 
1000  void RateFormula(const TargetTransformInfo &TTI,
1001  const Formula &F,
1003  const DenseSet<const SCEV *> &VisitedRegs,
1004  const Loop *L,
1005  ScalarEvolution &SE, DominatorTree &DT,
1006  const LSRUse &LU,
1007  SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1008 
1009  void print(raw_ostream &OS) const;
1010  void dump() const;
1011 
1012 private:
1013  void RateRegister(const SCEV *Reg,
1015  const Loop *L,
1016  ScalarEvolution &SE, DominatorTree &DT);
1017  void RatePrimaryRegister(const SCEV *Reg,
1019  const Loop *L,
1020  ScalarEvolution &SE, DominatorTree &DT,
1021  SmallPtrSetImpl<const SCEV *> *LoserRegs);
1022 };
1023 
1024 /// An operand value in an instruction which is to be replaced with some
1025 /// equivalent, possibly strength-reduced, replacement.
1026 struct LSRFixup {
1027  /// The instruction which will be updated.
1028  Instruction *UserInst;
1029 
1030  /// The operand of the instruction which will be replaced. The operand may be
1031  /// used more than once; every instance will be replaced.
1032  Value *OperandValToReplace;
1033 
1034  /// If this user is to use the post-incremented value of an induction
1035  /// variable, this variable is non-null and holds the loop associated with the
1036  /// induction variable.
1037  PostIncLoopSet PostIncLoops;
1038 
1039  /// A constant offset to be added to the LSRUse expression. This allows
1040  /// multiple fixups to share the same LSRUse with different offsets, for
1041  /// example in an unrolled loop.
1042  int64_t Offset;
1043 
1044  bool isUseFullyOutsideLoop(const Loop *L) const;
1045 
1046  LSRFixup();
1047 
1048  void print(raw_ostream &OS) const;
1049  void dump() const;
1050 };
1051 
1052 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1053 /// SmallVectors of const SCEV*.
1054 struct UniquifierDenseMapInfo {
1055  static SmallVector<const SCEV *, 4> getEmptyKey() {
1057  V.push_back(reinterpret_cast<const SCEV *>(-1));
1058  return V;
1059  }
1060 
1061  static SmallVector<const SCEV *, 4> getTombstoneKey() {
1063  V.push_back(reinterpret_cast<const SCEV *>(-2));
1064  return V;
1065  }
1066 
1067  static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1068  return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1069  }
1070 
1071  static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1072  const SmallVector<const SCEV *, 4> &RHS) {
1073  return LHS == RHS;
1074  }
1075 };
1076 
1077 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1078 /// as uses invented by LSR itself. It includes information about what kinds of
1079 /// things can be folded into the user, information about the user itself, and
1080 /// information about how the use may be satisfied. TODO: Represent multiple
1081 /// users of the same expression in common?
1082 class LSRUse {
1083  DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1084 
1085 public:
1086  /// An enum for a kind of use, indicating what types of scaled and immediate
1087  /// operands it might support.
1088  enum KindType {
1089  Basic, ///< A normal use, with no folding.
1090  Special, ///< A special case of basic, allowing -1 scales.
1091  Address, ///< An address use; folding according to TargetLowering
1092  ICmpZero ///< An equality icmp with both operands folded into one.
1093  // TODO: Add a generic icmp too?
1094  };
1095 
1096  typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
1097 
1098  KindType Kind;
1099  MemAccessTy AccessTy;
1100 
1101  /// The list of operands which are to be replaced.
1103 
1104  /// Keep track of the min and max offsets of the fixups.
1105  int64_t MinOffset;
1106  int64_t MaxOffset;
1107 
1108  /// This records whether all of the fixups using this LSRUse are outside of
1109  /// the loop, in which case some special-case heuristics may be used.
1110  bool AllFixupsOutsideLoop;
1111 
1112  /// RigidFormula is set to true to guarantee that this use will be associated
1113  /// with a single formula--the one that initially matched. Some SCEV
1114  /// expressions cannot be expanded. This allows LSR to consider the registers
1115  /// used by those expressions without the need to expand them later after
1116  /// changing the formula.
1117  bool RigidFormula;
1118 
1119  /// This records the widest use type for any fixup using this
1120  /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1121  /// fixup widths to be equivalent, because the narrower one may be relying on
1122  /// the implicit truncation to truncate away bogus bits.
1123  Type *WidestFixupType;
1124 
1125  /// A list of ways to build a value that can satisfy this user. After the
1126  /// list is populated, one of these is selected heuristically and used to
1127  /// formulate a replacement for OperandValToReplace in UserInst.
1128  SmallVector<Formula, 12> Formulae;
1129 
1130  /// The set of register candidates used by all formulae in this LSRUse.
1132 
1133  LSRUse(KindType K, MemAccessTy AT)
1134  : Kind(K), AccessTy(AT), MinOffset(INT64_MAX), MaxOffset(INT64_MIN),
1135  AllFixupsOutsideLoop(true), RigidFormula(false),
1136  WidestFixupType(nullptr) {}
1137 
1138  LSRFixup &getNewFixup() {
1139  Fixups.push_back(LSRFixup());
1140  return Fixups.back();
1141  }
1142 
1143  void pushFixup(LSRFixup &f) {
1144  Fixups.push_back(f);
1145  if (f.Offset > MaxOffset)
1146  MaxOffset = f.Offset;
1147  if (f.Offset < MinOffset)
1148  MinOffset = f.Offset;
1149  }
1150 
1151  bool HasFormulaWithSameRegs(const Formula &F) const;
1152  float getNotSelectedProbability(const SCEV *Reg) const;
1153  bool InsertFormula(const Formula &F, const Loop &L);
1154  void DeleteFormula(Formula &F);
1155  void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1156 
1157  void print(raw_ostream &OS) const;
1158  void dump() const;
1159 };
1160 
1161 } // end anonymous namespace
1162 
1163 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1164  LSRUse::KindType Kind, MemAccessTy AccessTy,
1165  GlobalValue *BaseGV, int64_t BaseOffset,
1166  bool HasBaseReg, int64_t Scale,
1167  Instruction *Fixup = nullptr);
1168 
1169 /// Tally up interesting quantities from the given register.
1170 void Cost::RateRegister(const SCEV *Reg,
1172  const Loop *L,
1173  ScalarEvolution &SE, DominatorTree &DT) {
1174  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1175  // If this is an addrec for another loop, it should be an invariant
1176  // with respect to L since L is the innermost loop (at least
1177  // for now LSR only handles innermost loops).
1178  if (AR->getLoop() != L) {
1179  // If the AddRec exists, consider it's register free and leave it alone.
1180  if (isExistingPhi(AR, SE))
1181  return;
1182 
1183  // It is bad to allow LSR for current loop to add induction variables
1184  // for its sibling loops.
1185  if (!AR->getLoop()->contains(L)) {
1186  Lose();
1187  return;
1188  }
1189 
1190  // Otherwise, it will be an invariant with respect to Loop L.
1191  ++C.NumRegs;
1192  return;
1193  }
1194  C.AddRecCost += 1; /// TODO: This should be a function of the stride.
1195 
1196  // Add the step value register, if it needs one.
1197  // TODO: The non-affine case isn't precisely modeled here.
1198  if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1199  if (!Regs.count(AR->getOperand(1))) {
1200  RateRegister(AR->getOperand(1), Regs, L, SE, DT);
1201  if (isLoser())
1202  return;
1203  }
1204  }
1205  }
1206  ++C.NumRegs;
1207 
1208  // Rough heuristic; favor registers which don't require extra setup
1209  // instructions in the preheader.
1210  if (!isa<SCEVUnknown>(Reg) &&
1211  !isa<SCEVConstant>(Reg) &&
1212  !(isa<SCEVAddRecExpr>(Reg) &&
1213  (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
1214  isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
1215  ++C.SetupCost;
1216 
1217  C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1218  SE.hasComputableLoopEvolution(Reg, L);
1219 }
1220 
1221 /// Record this register in the set. If we haven't seen it before, rate
1222 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1223 /// one of those regs an instant loser.
1224 void Cost::RatePrimaryRegister(const SCEV *Reg,
1226  const Loop *L,
1227  ScalarEvolution &SE, DominatorTree &DT,
1228  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1229  if (LoserRegs && LoserRegs->count(Reg)) {
1230  Lose();
1231  return;
1232  }
1233  if (Regs.insert(Reg).second) {
1234  RateRegister(Reg, Regs, L, SE, DT);
1235  if (LoserRegs && isLoser())
1236  LoserRegs->insert(Reg);
1237  }
1238 }
1239 
1240 void Cost::RateFormula(const TargetTransformInfo &TTI,
1241  const Formula &F,
1243  const DenseSet<const SCEV *> &VisitedRegs,
1244  const Loop *L,
1245  ScalarEvolution &SE, DominatorTree &DT,
1246  const LSRUse &LU,
1247  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1248  assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1249  // Tally up the registers.
1250  unsigned PrevAddRecCost = C.AddRecCost;
1251  unsigned PrevNumRegs = C.NumRegs;
1252  unsigned PrevNumBaseAdds = C.NumBaseAdds;
1253  if (const SCEV *ScaledReg = F.ScaledReg) {
1254  if (VisitedRegs.count(ScaledReg)) {
1255  Lose();
1256  return;
1257  }
1258  RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
1259  if (isLoser())
1260  return;
1261  }
1262  for (const SCEV *BaseReg : F.BaseRegs) {
1263  if (VisitedRegs.count(BaseReg)) {
1264  Lose();
1265  return;
1266  }
1267  RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
1268  if (isLoser())
1269  return;
1270  }
1271 
1272  // Determine how many (unfolded) adds we'll need inside the loop.
1273  size_t NumBaseParts = F.getNumRegs();
1274  if (NumBaseParts > 1)
1275  // Do not count the base and a possible second register if the target
1276  // allows to fold 2 registers.
1277  C.NumBaseAdds +=
1278  NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F)));
1279  C.NumBaseAdds += (F.UnfoldedOffset != 0);
1280 
1281  // Accumulate non-free scaling amounts.
1282  C.ScaleCost += getScalingFactorCost(TTI, LU, F, *L);
1283 
1284  // Tally up the non-zero immediates.
1285  for (const LSRFixup &Fixup : LU.Fixups) {
1286  int64_t O = Fixup.Offset;
1287  int64_t Offset = (uint64_t)O + F.BaseOffset;
1288  if (F.BaseGV)
1289  C.ImmCost += 64; // Handle symbolic values conservatively.
1290  // TODO: This should probably be the pointer size.
1291  else if (Offset != 0)
1292  C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1293 
1294  // Check with target if this offset with this instruction is
1295  // specifically not supported.
1296  if (LU.Kind == LSRUse::Address && Offset != 0 &&
1297  !isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1298  Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1299  C.NumBaseAdds++;
1300  }
1301 
1302  // If we don't count instruction cost exit here.
1303  if (!InsnsCost) {
1304  assert(isValid() && "invalid cost");
1305  return;
1306  }
1307 
1308  // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1309  // additional instruction (at least fill).
1310  unsigned TTIRegNum = TTI.getNumberOfRegisters(false) - 1;
1311  if (C.NumRegs > TTIRegNum) {
1312  // Cost already exceeded TTIRegNum, then only newly added register can add
1313  // new instructions.
1314  if (PrevNumRegs > TTIRegNum)
1315  C.Insns += (C.NumRegs - PrevNumRegs);
1316  else
1317  C.Insns += (C.NumRegs - TTIRegNum);
1318  }
1319 
1320  // If ICmpZero formula ends with not 0, it could not be replaced by
1321  // just add or sub. We'll need to compare final result of AddRec.
1322  // That means we'll need an additional instruction.
1323  // For -10 + {0, +, 1}:
1324  // i = i + 1;
1325  // cmp i, 10
1326  //
1327  // For {-10, +, 1}:
1328  // i = i + 1;
1329  if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd())
1330  C.Insns++;
1331  // Each new AddRec adds 1 instruction to calculation.
1332  C.Insns += (C.AddRecCost - PrevAddRecCost);
1333 
1334  // BaseAdds adds instructions for unfolded registers.
1335  if (LU.Kind != LSRUse::ICmpZero)
1336  C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1337  assert(isValid() && "invalid cost");
1338 }
1339 
1340 /// Set this cost to a losing value.
1341 void Cost::Lose() {
1342  C.Insns = ~0u;
1343  C.NumRegs = ~0u;
1344  C.AddRecCost = ~0u;
1345  C.NumIVMuls = ~0u;
1346  C.NumBaseAdds = ~0u;
1347  C.ImmCost = ~0u;
1348  C.SetupCost = ~0u;
1349  C.ScaleCost = ~0u;
1350 }
1351 
1352 /// Choose the lower cost.
1353 bool Cost::isLess(Cost &Other, const TargetTransformInfo &TTI) {
1354  if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1355  C.Insns != Other.C.Insns)
1356  return C.Insns < Other.C.Insns;
1357  return TTI.isLSRCostLess(C, Other.C);
1358 }
1359 
1360 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1361 void Cost::print(raw_ostream &OS) const {
1362  if (InsnsCost)
1363  OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1364  OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1365  if (C.AddRecCost != 0)
1366  OS << ", with addrec cost " << C.AddRecCost;
1367  if (C.NumIVMuls != 0)
1368  OS << ", plus " << C.NumIVMuls << " IV mul"
1369  << (C.NumIVMuls == 1 ? "" : "s");
1370  if (C.NumBaseAdds != 0)
1371  OS << ", plus " << C.NumBaseAdds << " base add"
1372  << (C.NumBaseAdds == 1 ? "" : "s");
1373  if (C.ScaleCost != 0)
1374  OS << ", plus " << C.ScaleCost << " scale cost";
1375  if (C.ImmCost != 0)
1376  OS << ", plus " << C.ImmCost << " imm cost";
1377  if (C.SetupCost != 0)
1378  OS << ", plus " << C.SetupCost << " setup cost";
1379 }
1380 
1381 LLVM_DUMP_METHOD void Cost::dump() const {
1382  print(errs()); errs() << '\n';
1383 }
1384 #endif
1385 
1386 LSRFixup::LSRFixup()
1387  : UserInst(nullptr), OperandValToReplace(nullptr),
1388  Offset(0) {}
1389 
1390 /// Test whether this fixup always uses its value outside of the given loop.
1391 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1392  // PHI nodes use their value in their incoming blocks.
1393  if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1394  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1395  if (PN->getIncomingValue(i) == OperandValToReplace &&
1396  L->contains(PN->getIncomingBlock(i)))
1397  return false;
1398  return true;
1399  }
1400 
1401  return !L->contains(UserInst);
1402 }
1403 
1404 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1405 void LSRFixup::print(raw_ostream &OS) const {
1406  OS << "UserInst=";
1407  // Store is common and interesting enough to be worth special-casing.
1408  if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1409  OS << "store ";
1410  Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1411  } else if (UserInst->getType()->isVoidTy())
1412  OS << UserInst->getOpcodeName();
1413  else
1414  UserInst->printAsOperand(OS, /*PrintType=*/false);
1415 
1416  OS << ", OperandValToReplace=";
1417  OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1418 
1419  for (const Loop *PIL : PostIncLoops) {
1420  OS << ", PostIncLoop=";
1421  PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1422  }
1423 
1424  if (Offset != 0)
1425  OS << ", Offset=" << Offset;
1426 }
1427 
1428 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1429  print(errs()); errs() << '\n';
1430 }
1431 #endif
1432 
1433 /// Test whether this use as a formula which has the same registers as the given
1434 /// formula.
1435 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1436  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1437  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1438  // Unstable sort by host order ok, because this is only used for uniquifying.
1439  std::sort(Key.begin(), Key.end());
1440  return Uniquifier.count(Key);
1441 }
1442 
1443 /// The function returns a probability of selecting formula without Reg.
1444 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1445  unsigned FNum = 0;
1446  for (const Formula &F : Formulae)
1447  if (F.referencesReg(Reg))
1448  FNum++;
1449  return ((float)(Formulae.size() - FNum)) / Formulae.size();
1450 }
1451 
1452 /// If the given formula has not yet been inserted, add it to the list, and
1453 /// return true. Return false otherwise. The formula must be in canonical form.
1454 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1455  assert(F.isCanonical(L) && "Invalid canonical representation");
1456 
1457  if (!Formulae.empty() && RigidFormula)
1458  return false;
1459 
1460  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1461  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1462  // Unstable sort by host order ok, because this is only used for uniquifying.
1463  std::sort(Key.begin(), Key.end());
1464 
1465  if (!Uniquifier.insert(Key).second)
1466  return false;
1467 
1468  // Using a register to hold the value of 0 is not profitable.
1469  assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1470  "Zero allocated in a scaled register!");
1471 #ifndef NDEBUG
1472  for (const SCEV *BaseReg : F.BaseRegs)
1473  assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1474 #endif
1475 
1476  // Add the formula to the list.
1477  Formulae.push_back(F);
1478 
1479  // Record registers now being used by this use.
1480  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1481  if (F.ScaledReg)
1482  Regs.insert(F.ScaledReg);
1483 
1484  return true;
1485 }
1486 
1487 /// Remove the given formula from this use's list.
1488 void LSRUse::DeleteFormula(Formula &F) {
1489  if (&F != &Formulae.back())
1490  std::swap(F, Formulae.back());
1491  Formulae.pop_back();
1492 }
1493 
1494 /// Recompute the Regs field, and update RegUses.
1495 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1496  // Now that we've filtered out some formulae, recompute the Regs set.
1497  SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1498  Regs.clear();
1499  for (const Formula &F : Formulae) {
1500  if (F.ScaledReg) Regs.insert(F.ScaledReg);
1501  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1502  }
1503 
1504  // Update the RegTracker.
1505  for (const SCEV *S : OldRegs)
1506  if (!Regs.count(S))
1507  RegUses.dropRegister(S, LUIdx);
1508 }
1509 
1510 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1511 void LSRUse::print(raw_ostream &OS) const {
1512  OS << "LSR Use: Kind=";
1513  switch (Kind) {
1514  case Basic: OS << "Basic"; break;
1515  case Special: OS << "Special"; break;
1516  case ICmpZero: OS << "ICmpZero"; break;
1517  case Address:
1518  OS << "Address of ";
1519  if (AccessTy.MemTy->isPointerTy())
1520  OS << "pointer"; // the full pointer type could be really verbose
1521  else {
1522  OS << *AccessTy.MemTy;
1523  }
1524 
1525  OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1526  }
1527 
1528  OS << ", Offsets={";
1529  bool NeedComma = false;
1530  for (const LSRFixup &Fixup : Fixups) {
1531  if (NeedComma) OS << ',';
1532  OS << Fixup.Offset;
1533  NeedComma = true;
1534  }
1535  OS << '}';
1536 
1537  if (AllFixupsOutsideLoop)
1538  OS << ", all-fixups-outside-loop";
1539 
1540  if (WidestFixupType)
1541  OS << ", widest fixup type: " << *WidestFixupType;
1542 }
1543 
1544 LLVM_DUMP_METHOD void LSRUse::dump() const {
1545  print(errs()); errs() << '\n';
1546 }
1547 #endif
1548 
1550  LSRUse::KindType Kind, MemAccessTy AccessTy,
1551  GlobalValue *BaseGV, int64_t BaseOffset,
1552  bool HasBaseReg, int64_t Scale,
1553  Instruction *Fixup/*= nullptr*/) {
1554  switch (Kind) {
1555  case LSRUse::Address:
1556  return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1557  HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1558 
1559  case LSRUse::ICmpZero:
1560  // There's not even a target hook for querying whether it would be legal to
1561  // fold a GV into an ICmp.
1562  if (BaseGV)
1563  return false;
1564 
1565  // ICmp only has two operands; don't allow more than two non-trivial parts.
1566  if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1567  return false;
1568 
1569  // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1570  // putting the scaled register in the other operand of the icmp.
1571  if (Scale != 0 && Scale != -1)
1572  return false;
1573 
1574  // If we have low-level target information, ask the target if it can fold an
1575  // integer immediate on an icmp.
1576  if (BaseOffset != 0) {
1577  // We have one of:
1578  // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1579  // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1580  // Offs is the ICmp immediate.
1581  if (Scale == 0)
1582  // The cast does the right thing with INT64_MIN.
1583  BaseOffset = -(uint64_t)BaseOffset;
1584  return TTI.isLegalICmpImmediate(BaseOffset);
1585  }
1586 
1587  // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1588  return true;
1589 
1590  case LSRUse::Basic:
1591  // Only handle single-register values.
1592  return !BaseGV && Scale == 0 && BaseOffset == 0;
1593 
1594  case LSRUse::Special:
1595  // Special case Basic to handle -1 scales.
1596  return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1597  }
1598 
1599  llvm_unreachable("Invalid LSRUse Kind!");
1600 }
1601 
1603  int64_t MinOffset, int64_t MaxOffset,
1604  LSRUse::KindType Kind, MemAccessTy AccessTy,
1605  GlobalValue *BaseGV, int64_t BaseOffset,
1606  bool HasBaseReg, int64_t Scale) {
1607  // Check for overflow.
1608  if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1609  (MinOffset > 0))
1610  return false;
1611  MinOffset = (uint64_t)BaseOffset + MinOffset;
1612  if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1613  (MaxOffset > 0))
1614  return false;
1615  MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1616 
1617  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1618  HasBaseReg, Scale) &&
1619  isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1620  HasBaseReg, Scale);
1621 }
1622 
1624  int64_t MinOffset, int64_t MaxOffset,
1625  LSRUse::KindType Kind, MemAccessTy AccessTy,
1626  const Formula &F, const Loop &L) {
1627  // For the purpose of isAMCompletelyFolded either having a canonical formula
1628  // or a scale not equal to zero is correct.
1629  // Problems may arise from non canonical formulae having a scale == 0.
1630  // Strictly speaking it would best to just rely on canonical formulae.
1631  // However, when we generate the scaled formulae, we first check that the
1632  // scaling factor is profitable before computing the actual ScaledReg for
1633  // compile time sake.
1634  assert((F.isCanonical(L) || F.Scale != 0));
1635  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1636  F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1637 }
1638 
1639 /// Test whether we know how to expand the current formula.
1640 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1641  int64_t MaxOffset, LSRUse::KindType Kind,
1642  MemAccessTy AccessTy, GlobalValue *BaseGV,
1643  int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1644  // We know how to expand completely foldable formulae.
1645  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1646  BaseOffset, HasBaseReg, Scale) ||
1647  // Or formulae that use a base register produced by a sum of base
1648  // registers.
1649  (Scale == 1 &&
1650  isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1651  BaseGV, BaseOffset, true, 0));
1652 }
1653 
1654 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1655  int64_t MaxOffset, LSRUse::KindType Kind,
1656  MemAccessTy AccessTy, const Formula &F) {
1657  return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1658  F.BaseOffset, F.HasBaseReg, F.Scale);
1659 }
1660 
1662  const LSRUse &LU, const Formula &F) {
1663  // Target may want to look at the user instructions.
1664  if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1665  for (const LSRFixup &Fixup : LU.Fixups)
1666  if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1667  (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1668  F.Scale, Fixup.UserInst))
1669  return false;
1670  return true;
1671  }
1672 
1673  return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1674  LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1675  F.Scale);
1676 }
1677 
1678 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1679  const LSRUse &LU, const Formula &F,
1680  const Loop &L) {
1681  if (!F.Scale)
1682  return 0;
1683 
1684  // If the use is not completely folded in that instruction, we will have to
1685  // pay an extra cost only for scale != 1.
1686  if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1687  LU.AccessTy, F, L))
1688  return F.Scale != 1;
1689 
1690  switch (LU.Kind) {
1691  case LSRUse::Address: {
1692  // Check the scaling factor cost with both the min and max offsets.
1693  int ScaleCostMinOffset = TTI.getScalingFactorCost(
1694  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1695  F.Scale, LU.AccessTy.AddrSpace);
1696  int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1697  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1698  F.Scale, LU.AccessTy.AddrSpace);
1699 
1700  assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1701  "Legal addressing mode has an illegal cost!");
1702  return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1703  }
1704  case LSRUse::ICmpZero:
1705  case LSRUse::Basic:
1706  case LSRUse::Special:
1707  // The use is completely folded, i.e., everything is folded into the
1708  // instruction.
1709  return 0;
1710  }
1711 
1712  llvm_unreachable("Invalid LSRUse Kind!");
1713 }
1714 
1715 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1716  LSRUse::KindType Kind, MemAccessTy AccessTy,
1717  GlobalValue *BaseGV, int64_t BaseOffset,
1718  bool HasBaseReg) {
1719  // Fast-path: zero is always foldable.
1720  if (BaseOffset == 0 && !BaseGV) return true;
1721 
1722  // Conservatively, create an address with an immediate and a
1723  // base and a scale.
1724  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1725 
1726  // Canonicalize a scale of 1 to a base register if the formula doesn't
1727  // already have a base register.
1728  if (!HasBaseReg && Scale == 1) {
1729  Scale = 0;
1730  HasBaseReg = true;
1731  }
1732 
1733  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1734  HasBaseReg, Scale);
1735 }
1736 
1737 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1738  ScalarEvolution &SE, int64_t MinOffset,
1739  int64_t MaxOffset, LSRUse::KindType Kind,
1740  MemAccessTy AccessTy, const SCEV *S,
1741  bool HasBaseReg) {
1742  // Fast-path: zero is always foldable.
1743  if (S->isZero()) return true;
1744 
1745  // Conservatively, create an address with an immediate and a
1746  // base and a scale.
1747  int64_t BaseOffset = ExtractImmediate(S, SE);
1748  GlobalValue *BaseGV = ExtractSymbol(S, SE);
1749 
1750  // If there's anything else involved, it's not foldable.
1751  if (!S->isZero()) return false;
1752 
1753  // Fast-path: zero is always foldable.
1754  if (BaseOffset == 0 && !BaseGV) return true;
1755 
1756  // Conservatively, create an address with an immediate and a
1757  // base and a scale.
1758  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1759 
1760  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1761  BaseOffset, HasBaseReg, Scale);
1762 }
1763 
1764 namespace {
1765 
1766 /// An individual increment in a Chain of IV increments. Relate an IV user to
1767 /// an expression that computes the IV it uses from the IV used by the previous
1768 /// link in the Chain.
1769 ///
1770 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1771 /// original IVOperand. The head of the chain's IVOperand is only valid during
1772 /// chain collection, before LSR replaces IV users. During chain generation,
1773 /// IncExpr can be used to find the new IVOperand that computes the same
1774 /// expression.
1775 struct IVInc {
1776  Instruction *UserInst;
1777  Value* IVOperand;
1778  const SCEV *IncExpr;
1779 
1780  IVInc(Instruction *U, Value *O, const SCEV *E):
1781  UserInst(U), IVOperand(O), IncExpr(E) {}
1782 };
1783 
1784 // The list of IV increments in program order. We typically add the head of a
1785 // chain without finding subsequent links.
1786 struct IVChain {
1787  SmallVector<IVInc,1> Incs;
1788  const SCEV *ExprBase;
1789 
1790  IVChain() : ExprBase(nullptr) {}
1791 
1792  IVChain(const IVInc &Head, const SCEV *Base)
1793  : Incs(1, Head), ExprBase(Base) {}
1794 
1795  typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
1796 
1797  // Return the first increment in the chain.
1798  const_iterator begin() const {
1799  assert(!Incs.empty());
1800  return std::next(Incs.begin());
1801  }
1802  const_iterator end() const {
1803  return Incs.end();
1804  }
1805 
1806  // Returns true if this chain contains any increments.
1807  bool hasIncs() const { return Incs.size() >= 2; }
1808 
1809  // Add an IVInc to the end of this chain.
1810  void add(const IVInc &X) { Incs.push_back(X); }
1811 
1812  // Returns the last UserInst in the chain.
1813  Instruction *tailUserInst() const { return Incs.back().UserInst; }
1814 
1815  // Returns true if IncExpr can be profitably added to this chain.
1816  bool isProfitableIncrement(const SCEV *OperExpr,
1817  const SCEV *IncExpr,
1818  ScalarEvolution&);
1819 };
1820 
1821 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1822 /// between FarUsers that definitely cross IV increments and NearUsers that may
1823 /// be used between IV increments.
1824 struct ChainUsers {
1826  SmallPtrSet<Instruction*, 4> NearUsers;
1827 };
1828 
1829 /// This class holds state for the main loop strength reduction logic.
1830 class LSRInstance {
1831  IVUsers &IU;
1832  ScalarEvolution &SE;
1833  DominatorTree &DT;
1834  LoopInfo &LI;
1835  const TargetTransformInfo &TTI;
1836  Loop *const L;
1837  bool Changed;
1838 
1839  /// This is the insert position that the current loop's induction variable
1840  /// increment should be placed. In simple loops, this is the latch block's
1841  /// terminator. But in more complicated cases, this is a position which will
1842  /// dominate all the in-loop post-increment users.
1843  Instruction *IVIncInsertPos;
1844 
1845  /// Interesting factors between use strides.
1846  ///
1847  /// We explicitly use a SetVector which contains a SmallSet, instead of the
1848  /// default, a SmallDenseSet, because we need to use the full range of
1849  /// int64_ts, and there's currently no good way of doing that with
1850  /// SmallDenseSet.
1852 
1853  /// Interesting use types, to facilitate truncation reuse.
1855 
1856  /// The list of interesting uses.
1858 
1859  /// Track which uses use which register candidates.
1860  RegUseTracker RegUses;
1861 
1862  // Limit the number of chains to avoid quadratic behavior. We don't expect to
1863  // have more than a few IV increment chains in a loop. Missing a Chain falls
1864  // back to normal LSR behavior for those uses.
1865  static const unsigned MaxChains = 8;
1866 
1867  /// IV users can form a chain of IV increments.
1869 
1870  /// IV users that belong to profitable IVChains.
1872 
1873  void OptimizeShadowIV();
1874  bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1875  ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1876  void OptimizeLoopTermCond();
1877 
1878  void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1879  SmallVectorImpl<ChainUsers> &ChainUsersVec);
1880  void FinalizeChain(IVChain &Chain);
1881  void CollectChains();
1882  void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1883  SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1884 
1885  void CollectInterestingTypesAndFactors();
1886  void CollectFixupsAndInitialFormulae();
1887 
1888  // Support for sharing of LSRUses between LSRFixups.
1890  UseMapTy UseMap;
1891 
1892  bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1893  LSRUse::KindType Kind, MemAccessTy AccessTy);
1894 
1895  std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1896  MemAccessTy AccessTy);
1897 
1898  void DeleteUse(LSRUse &LU, size_t LUIdx);
1899 
1900  LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1901 
1902  void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1903  void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1904  void CountRegisters(const Formula &F, size_t LUIdx);
1905  bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1906 
1907  void CollectLoopInvariantFixupsAndFormulae();
1908 
1909  void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1910  unsigned Depth = 0);
1911 
1912  void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1913  const Formula &Base, unsigned Depth,
1914  size_t Idx, bool IsScaledReg = false);
1915  void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1916  void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1917  const Formula &Base, size_t Idx,
1918  bool IsScaledReg = false);
1919  void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1920  void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1921  const Formula &Base,
1922  const SmallVectorImpl<int64_t> &Worklist,
1923  size_t Idx, bool IsScaledReg = false);
1924  void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1925  void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1926  void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1927  void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1928  void GenerateCrossUseConstantOffsets();
1929  void GenerateAllReuseFormulae();
1930 
1931  void FilterOutUndesirableDedicatedRegisters();
1932 
1933  size_t EstimateSearchSpaceComplexity() const;
1934  void NarrowSearchSpaceByDetectingSupersets();
1935  void NarrowSearchSpaceByCollapsingUnrolledCode();
1936  void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
1937  void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
1938  void NarrowSearchSpaceByDeletingCostlyFormulas();
1939  void NarrowSearchSpaceByPickingWinnerRegs();
1940  void NarrowSearchSpaceUsingHeuristics();
1941 
1942  void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1943  Cost &SolutionCost,
1945  const Cost &CurCost,
1946  const SmallPtrSet<const SCEV *, 16> &CurRegs,
1947  DenseSet<const SCEV *> &VisitedRegs) const;
1948  void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1949 
1951  HoistInsertPosition(BasicBlock::iterator IP,
1952  const SmallVectorImpl<Instruction *> &Inputs) const;
1954  AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1955  const LSRFixup &LF,
1956  const LSRUse &LU,
1957  SCEVExpander &Rewriter) const;
1958 
1959  Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1960  BasicBlock::iterator IP, SCEVExpander &Rewriter,
1961  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1962  void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
1963  const Formula &F, SCEVExpander &Rewriter,
1964  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1965  void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
1966  SCEVExpander &Rewriter,
1967  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
1968  void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
1969 
1970 public:
1971  LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
1972  LoopInfo &LI, const TargetTransformInfo &TTI);
1973 
1974  bool getChanged() const { return Changed; }
1975 
1976  void print_factors_and_types(raw_ostream &OS) const;
1977  void print_fixups(raw_ostream &OS) const;
1978  void print_uses(raw_ostream &OS) const;
1979  void print(raw_ostream &OS) const;
1980  void dump() const;
1981 };
1982 
1983 } // end anonymous namespace
1984 
1985 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
1986 /// the cast operation.
1987 void LSRInstance::OptimizeShadowIV() {
1988  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1989  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1990  return;
1991 
1992  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1993  UI != E; /* empty */) {
1994  IVUsers::const_iterator CandidateUI = UI;
1995  ++UI;
1996  Instruction *ShadowUse = CandidateUI->getUser();
1997  Type *DestTy = nullptr;
1998  bool IsSigned = false;
1999 
2000  /* If shadow use is a int->float cast then insert a second IV
2001  to eliminate this cast.
2002 
2003  for (unsigned i = 0; i < n; ++i)
2004  foo((double)i);
2005 
2006  is transformed into
2007 
2008  double d = 0.0;
2009  for (unsigned i = 0; i < n; ++i, ++d)
2010  foo(d);
2011  */
2012  if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2013  IsSigned = false;
2014  DestTy = UCast->getDestTy();
2015  }
2016  else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2017  IsSigned = true;
2018  DestTy = SCast->getDestTy();
2019  }
2020  if (!DestTy) continue;
2021 
2022  // If target does not support DestTy natively then do not apply
2023  // this transformation.
2024  if (!TTI.isTypeLegal(DestTy)) continue;
2025 
2026  PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2027  if (!PH) continue;
2028  if (PH->getNumIncomingValues() != 2) continue;
2029 
2030  // If the calculation in integers overflows, the result in FP type will
2031  // differ. So we only can do this transformation if we are guaranteed to not
2032  // deal with overflowing values
2033  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2034  if (!AR) continue;
2035  if (IsSigned && !AR->hasNoSignedWrap()) continue;
2036  if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2037 
2038  Type *SrcTy = PH->getType();
2039  int Mantissa = DestTy->getFPMantissaWidth();
2040  if (Mantissa == -1) continue;
2041  if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2042  continue;
2043 
2044  unsigned Entry, Latch;
2045  if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2046  Entry = 0;
2047  Latch = 1;
2048  } else {
2049  Entry = 1;
2050  Latch = 0;
2051  }
2052 
2054  if (!Init) continue;
2055  Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2056  (double)Init->getSExtValue() :
2057  (double)Init->getZExtValue());
2058 
2059  BinaryOperator *Incr =
2061  if (!Incr) continue;
2062  if (Incr->getOpcode() != Instruction::Add
2063  && Incr->getOpcode() != Instruction::Sub)
2064  continue;
2065 
2066  /* Initialize new IV, double d = 0.0 in above example. */
2067  ConstantInt *C = nullptr;
2068  if (Incr->getOperand(0) == PH)
2069  C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2070  else if (Incr->getOperand(1) == PH)
2071  C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2072  else
2073  continue;
2074 
2075  if (!C) continue;
2076 
2077  // Ignore negative constants, as the code below doesn't handle them
2078  // correctly. TODO: Remove this restriction.
2079  if (!C->getValue().isStrictlyPositive()) continue;
2080 
2081  /* Add new PHINode. */
2082  PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2083 
2084  /* create new increment. '++d' in above example. */
2085  Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2086  BinaryOperator *NewIncr =
2087  BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2088  Instruction::FAdd : Instruction::FSub,
2089  NewPH, CFP, "IV.S.next.", Incr);
2090 
2091  NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2092  NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2093 
2094  /* Remove cast operation */
2095  ShadowUse->replaceAllUsesWith(NewPH);
2096  ShadowUse->eraseFromParent();
2097  Changed = true;
2098  break;
2099  }
2100 }
2101 
2102 /// If Cond has an operand that is an expression of an IV, set the IV user and
2103 /// stride information and return true, otherwise return false.
2104 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2105  for (IVStrideUse &U : IU)
2106  if (U.getUser() == Cond) {
2107  // NOTE: we could handle setcc instructions with multiple uses here, but
2108  // InstCombine does it as well for simple uses, it's not clear that it
2109  // occurs enough in real life to handle.
2110  CondUse = &U;
2111  return true;
2112  }
2113  return false;
2114 }
2115 
2116 /// Rewrite the loop's terminating condition if it uses a max computation.
2117 ///
2118 /// This is a narrow solution to a specific, but acute, problem. For loops
2119 /// like this:
2120 ///
2121 /// i = 0;
2122 /// do {
2123 /// p[i] = 0.0;
2124 /// } while (++i < n);
2125 ///
2126 /// the trip count isn't just 'n', because 'n' might not be positive. And
2127 /// unfortunately this can come up even for loops where the user didn't use
2128 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2129 /// will commonly be lowered like this:
2130 //
2131 /// if (n > 0) {
2132 /// i = 0;
2133 /// do {
2134 /// p[i] = 0.0;
2135 /// } while (++i < n);
2136 /// }
2137 ///
2138 /// and then it's possible for subsequent optimization to obscure the if
2139 /// test in such a way that indvars can't find it.
2140 ///
2141 /// When indvars can't find the if test in loops like this, it creates a
2142 /// max expression, which allows it to give the loop a canonical
2143 /// induction variable:
2144 ///
2145 /// i = 0;
2146 /// max = n < 1 ? 1 : n;
2147 /// do {
2148 /// p[i] = 0.0;
2149 /// } while (++i != max);
2150 ///
2151 /// Canonical induction variables are necessary because the loop passes
2152 /// are designed around them. The most obvious example of this is the
2153 /// LoopInfo analysis, which doesn't remember trip count values. It
2154 /// expects to be able to rediscover the trip count each time it is
2155 /// needed, and it does this using a simple analysis that only succeeds if
2156 /// the loop has a canonical induction variable.
2157 ///
2158 /// However, when it comes time to generate code, the maximum operation
2159 /// can be quite costly, especially if it's inside of an outer loop.
2160 ///
2161 /// This function solves this problem by detecting this type of loop and
2162 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2163 /// the instructions for the maximum computation.
2164 ///
2165 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2166  // Check that the loop matches the pattern we're looking for.
2167  if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2168  Cond->getPredicate() != CmpInst::ICMP_NE)
2169  return Cond;
2170 
2171  SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2172  if (!Sel || !Sel->hasOneUse()) return Cond;
2173 
2174  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2175  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2176  return Cond;
2177  const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2178 
2179  // Add one to the backedge-taken count to get the trip count.
2180  const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2181  if (IterationCount != SE.getSCEV(Sel)) return Cond;
2182 
2183  // Check for a max calculation that matches the pattern. There's no check
2184  // for ICMP_ULE here because the comparison would be with zero, which
2185  // isn't interesting.
2187  const SCEVNAryExpr *Max = nullptr;
2188  if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2189  Pred = ICmpInst::ICMP_SLE;
2190  Max = S;
2191  } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2192  Pred = ICmpInst::ICMP_SLT;
2193  Max = S;
2194  } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2195  Pred = ICmpInst::ICMP_ULT;
2196  Max = U;
2197  } else {
2198  // No match; bail.
2199  return Cond;
2200  }
2201 
2202  // To handle a max with more than two operands, this optimization would
2203  // require additional checking and setup.
2204  if (Max->getNumOperands() != 2)
2205  return Cond;
2206 
2207  const SCEV *MaxLHS = Max->getOperand(0);
2208  const SCEV *MaxRHS = Max->getOperand(1);
2209 
2210  // ScalarEvolution canonicalizes constants to the left. For < and >, look
2211  // for a comparison with 1. For <= and >=, a comparison with zero.
2212  if (!MaxLHS ||
2213  (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2214  return Cond;
2215 
2216  // Check the relevant induction variable for conformance to
2217  // the pattern.
2218  const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2219  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2220  if (!AR || !AR->isAffine() ||
2221  AR->getStart() != One ||
2222  AR->getStepRecurrence(SE) != One)
2223  return Cond;
2224 
2225  assert(AR->getLoop() == L &&
2226  "Loop condition operand is an addrec in a different loop!");
2227 
2228  // Check the right operand of the select, and remember it, as it will
2229  // be used in the new comparison instruction.
2230  Value *NewRHS = nullptr;
2231  if (ICmpInst::isTrueWhenEqual(Pred)) {
2232  // Look for n+1, and grab n.
2233  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2234  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2235  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2236  NewRHS = BO->getOperand(0);
2237  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2238  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2239  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2240  NewRHS = BO->getOperand(0);
2241  if (!NewRHS)
2242  return Cond;
2243  } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2244  NewRHS = Sel->getOperand(1);
2245  else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2246  NewRHS = Sel->getOperand(2);
2247  else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2248  NewRHS = SU->getValue();
2249  else
2250  // Max doesn't match expected pattern.
2251  return Cond;
2252 
2253  // Determine the new comparison opcode. It may be signed or unsigned,
2254  // and the original comparison may be either equality or inequality.
2255  if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2256  Pred = CmpInst::getInversePredicate(Pred);
2257 
2258  // Ok, everything looks ok to change the condition into an SLT or SGE and
2259  // delete the max calculation.
2260  ICmpInst *NewCond =
2261  new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2262 
2263  // Delete the max calculation instructions.
2264  Cond->replaceAllUsesWith(NewCond);
2265  CondUse->setUser(NewCond);
2266  Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2267  Cond->eraseFromParent();
2268  Sel->eraseFromParent();
2269  if (Cmp->use_empty())
2270  Cmp->eraseFromParent();
2271  return NewCond;
2272 }
2273 
2274 /// Change loop terminating condition to use the postinc iv when possible.
2275 void
2276 LSRInstance::OptimizeLoopTermCond() {
2278 
2279  // We need a different set of heuristics for rotated and non-rotated loops.
2280  // If a loop is rotated then the latch is also the backedge, so inserting
2281  // post-inc expressions just before the latch is ideal. To reduce live ranges
2282  // it also makes sense to rewrite terminating conditions to use post-inc
2283  // expressions.
2284  //
2285  // If the loop is not rotated then the latch is not a backedge; the latch
2286  // check is done in the loop head. Adding post-inc expressions before the
2287  // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2288  // in the loop body. In this case we do *not* want to use post-inc expressions
2289  // in the latch check, and we want to insert post-inc expressions before
2290  // the backedge.
2291  BasicBlock *LatchBlock = L->getLoopLatch();
2292  SmallVector<BasicBlock*, 8> ExitingBlocks;
2293  L->getExitingBlocks(ExitingBlocks);
2294  if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2295  return LatchBlock != BB;
2296  })) {
2297  // The backedge doesn't exit the loop; treat this as a head-tested loop.
2298  IVIncInsertPos = LatchBlock->getTerminator();
2299  return;
2300  }
2301 
2302  // Otherwise treat this as a rotated loop.
2303  for (BasicBlock *ExitingBlock : ExitingBlocks) {
2304 
2305  // Get the terminating condition for the loop if possible. If we
2306  // can, we want to change it to use a post-incremented version of its
2307  // induction variable, to allow coalescing the live ranges for the IV into
2308  // one register value.
2309 
2310  BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2311  if (!TermBr)
2312  continue;
2313  // FIXME: Overly conservative, termination condition could be an 'or' etc..
2314  if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2315  continue;
2316 
2317  // Search IVUsesByStride to find Cond's IVUse if there is one.
2318  IVStrideUse *CondUse = nullptr;
2319  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2320  if (!FindIVUserForCond(Cond, CondUse))
2321  continue;
2322 
2323  // If the trip count is computed in terms of a max (due to ScalarEvolution
2324  // being unable to find a sufficient guard, for example), change the loop
2325  // comparison to use SLT or ULT instead of NE.
2326  // One consequence of doing this now is that it disrupts the count-down
2327  // optimization. That's not always a bad thing though, because in such
2328  // cases it may still be worthwhile to avoid a max.
2329  Cond = OptimizeMax(Cond, CondUse);
2330 
2331  // If this exiting block dominates the latch block, it may also use
2332  // the post-inc value if it won't be shared with other uses.
2333  // Check for dominance.
2334  if (!DT.dominates(ExitingBlock, LatchBlock))
2335  continue;
2336 
2337  // Conservatively avoid trying to use the post-inc value in non-latch
2338  // exits if there may be pre-inc users in intervening blocks.
2339  if (LatchBlock != ExitingBlock)
2340  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2341  // Test if the use is reachable from the exiting block. This dominator
2342  // query is a conservative approximation of reachability.
2343  if (&*UI != CondUse &&
2344  !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2345  // Conservatively assume there may be reuse if the quotient of their
2346  // strides could be a legal scale.
2347  const SCEV *A = IU.getStride(*CondUse, L);
2348  const SCEV *B = IU.getStride(*UI, L);
2349  if (!A || !B) continue;
2350  if (SE.getTypeSizeInBits(A->getType()) !=
2351  SE.getTypeSizeInBits(B->getType())) {
2352  if (SE.getTypeSizeInBits(A->getType()) >
2353  SE.getTypeSizeInBits(B->getType()))
2354  B = SE.getSignExtendExpr(B, A->getType());
2355  else
2356  A = SE.getSignExtendExpr(A, B->getType());
2357  }
2358  if (const SCEVConstant *D =
2359  dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2360  const ConstantInt *C = D->getValue();
2361  // Stride of one or negative one can have reuse with non-addresses.
2362  if (C->isOne() || C->isMinusOne())
2363  goto decline_post_inc;
2364  // Avoid weird situations.
2365  if (C->getValue().getMinSignedBits() >= 64 ||
2366  C->getValue().isMinSignedValue())
2367  goto decline_post_inc;
2368  // Check for possible scaled-address reuse.
2369  MemAccessTy AccessTy = getAccessType(UI->getUser());
2370  int64_t Scale = C->getSExtValue();
2371  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2372  /*BaseOffset=*/0,
2373  /*HasBaseReg=*/false, Scale,
2374  AccessTy.AddrSpace))
2375  goto decline_post_inc;
2376  Scale = -Scale;
2377  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2378  /*BaseOffset=*/0,
2379  /*HasBaseReg=*/false, Scale,
2380  AccessTy.AddrSpace))
2381  goto decline_post_inc;
2382  }
2383  }
2384 
2385  DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2386  << *Cond << '\n');
2387 
2388  // It's possible for the setcc instruction to be anywhere in the loop, and
2389  // possible for it to have multiple users. If it is not immediately before
2390  // the exiting block branch, move it.
2391  if (&*++BasicBlock::iterator(Cond) != TermBr) {
2392  if (Cond->hasOneUse()) {
2393  Cond->moveBefore(TermBr);
2394  } else {
2395  // Clone the terminating condition and insert into the loopend.
2396  ICmpInst *OldCond = Cond;
2397  Cond = cast<ICmpInst>(Cond->clone());
2398  Cond->setName(L->getHeader()->getName() + ".termcond");
2399  ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2400 
2401  // Clone the IVUse, as the old use still exists!
2402  CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2403  TermBr->replaceUsesOfWith(OldCond, Cond);
2404  }
2405  }
2406 
2407  // If we get to here, we know that we can transform the setcc instruction to
2408  // use the post-incremented version of the IV, allowing us to coalesce the
2409  // live ranges for the IV correctly.
2410  CondUse->transformToPostInc(L);
2411  Changed = true;
2412 
2413  PostIncs.insert(Cond);
2414  decline_post_inc:;
2415  }
2416 
2417  // Determine an insertion point for the loop induction variable increment. It
2418  // must dominate all the post-inc comparisons we just set up, and it must
2419  // dominate the loop latch edge.
2420  IVIncInsertPos = L->getLoopLatch()->getTerminator();
2421  for (Instruction *Inst : PostIncs) {
2422  BasicBlock *BB =
2423  DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2424  Inst->getParent());
2425  if (BB == Inst->getParent())
2426  IVIncInsertPos = Inst;
2427  else if (BB != IVIncInsertPos->getParent())
2428  IVIncInsertPos = BB->getTerminator();
2429  }
2430 }
2431 
2432 /// Determine if the given use can accommodate a fixup at the given offset and
2433 /// other details. If so, update the use and return true.
2434 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2435  bool HasBaseReg, LSRUse::KindType Kind,
2436  MemAccessTy AccessTy) {
2437  int64_t NewMinOffset = LU.MinOffset;
2438  int64_t NewMaxOffset = LU.MaxOffset;
2439  MemAccessTy NewAccessTy = AccessTy;
2440 
2441  // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2442  // something conservative, however this can pessimize in the case that one of
2443  // the uses will have all its uses outside the loop, for example.
2444  if (LU.Kind != Kind)
2445  return false;
2446 
2447  // Check for a mismatched access type, and fall back conservatively as needed.
2448  // TODO: Be less conservative when the type is similar and can use the same
2449  // addressing modes.
2450  if (Kind == LSRUse::Address) {
2451  if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2452  NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2453  AccessTy.AddrSpace);
2454  }
2455  }
2456 
2457  // Conservatively assume HasBaseReg is true for now.
2458  if (NewOffset < LU.MinOffset) {
2459  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2460  LU.MaxOffset - NewOffset, HasBaseReg))
2461  return false;
2462  NewMinOffset = NewOffset;
2463  } else if (NewOffset > LU.MaxOffset) {
2464  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2465  NewOffset - LU.MinOffset, HasBaseReg))
2466  return false;
2467  NewMaxOffset = NewOffset;
2468  }
2469 
2470  // Update the use.
2471  LU.MinOffset = NewMinOffset;
2472  LU.MaxOffset = NewMaxOffset;
2473  LU.AccessTy = NewAccessTy;
2474  return true;
2475 }
2476 
2477 /// Return an LSRUse index and an offset value for a fixup which needs the given
2478 /// expression, with the given kind and optional access type. Either reuse an
2479 /// existing use or create a new one, as needed.
2480 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2481  LSRUse::KindType Kind,
2482  MemAccessTy AccessTy) {
2483  const SCEV *Copy = Expr;
2484  int64_t Offset = ExtractImmediate(Expr, SE);
2485 
2486  // Basic uses can't accept any offset, for example.
2487  if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2488  Offset, /*HasBaseReg=*/ true)) {
2489  Expr = Copy;
2490  Offset = 0;
2491  }
2492 
2493  std::pair<UseMapTy::iterator, bool> P =
2494  UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2495  if (!P.second) {
2496  // A use already existed with this base.
2497  size_t LUIdx = P.first->second;
2498  LSRUse &LU = Uses[LUIdx];
2499  if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2500  // Reuse this use.
2501  return std::make_pair(LUIdx, Offset);
2502  }
2503 
2504  // Create a new use.
2505  size_t LUIdx = Uses.size();
2506  P.first->second = LUIdx;
2507  Uses.push_back(LSRUse(Kind, AccessTy));
2508  LSRUse &LU = Uses[LUIdx];
2509 
2510  LU.MinOffset = Offset;
2511  LU.MaxOffset = Offset;
2512  return std::make_pair(LUIdx, Offset);
2513 }
2514 
2515 /// Delete the given use from the Uses list.
2516 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2517  if (&LU != &Uses.back())
2518  std::swap(LU, Uses.back());
2519  Uses.pop_back();
2520 
2521  // Update RegUses.
2522  RegUses.swapAndDropUse(LUIdx, Uses.size());
2523 }
2524 
2525 /// Look for a use distinct from OrigLU which is has a formula that has the same
2526 /// registers as the given formula.
2527 LSRUse *
2528 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2529  const LSRUse &OrigLU) {
2530  // Search all uses for the formula. This could be more clever.
2531  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2532  LSRUse &LU = Uses[LUIdx];
2533  // Check whether this use is close enough to OrigLU, to see whether it's
2534  // worthwhile looking through its formulae.
2535  // Ignore ICmpZero uses because they may contain formulae generated by
2536  // GenerateICmpZeroScales, in which case adding fixup offsets may
2537  // be invalid.
2538  if (&LU != &OrigLU &&
2539  LU.Kind != LSRUse::ICmpZero &&
2540  LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2541  LU.WidestFixupType == OrigLU.WidestFixupType &&
2542  LU.HasFormulaWithSameRegs(OrigF)) {
2543  // Scan through this use's formulae.
2544  for (const Formula &F : LU.Formulae) {
2545  // Check to see if this formula has the same registers and symbols
2546  // as OrigF.
2547  if (F.BaseRegs == OrigF.BaseRegs &&
2548  F.ScaledReg == OrigF.ScaledReg &&
2549  F.BaseGV == OrigF.BaseGV &&
2550  F.Scale == OrigF.Scale &&
2551  F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2552  if (F.BaseOffset == 0)
2553  return &LU;
2554  // This is the formula where all the registers and symbols matched;
2555  // there aren't going to be any others. Since we declined it, we
2556  // can skip the rest of the formulae and proceed to the next LSRUse.
2557  break;
2558  }
2559  }
2560  }
2561  }
2562 
2563  // Nothing looked good.
2564  return nullptr;
2565 }
2566 
2567 void LSRInstance::CollectInterestingTypesAndFactors() {
2569 
2570  // Collect interesting types and strides.
2572  for (const IVStrideUse &U : IU) {
2573  const SCEV *Expr = IU.getExpr(U);
2574 
2575  // Collect interesting types.
2576  Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2577 
2578  // Add strides for mentioned loops.
2579  Worklist.push_back(Expr);
2580  do {
2581  const SCEV *S = Worklist.pop_back_val();
2582  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2583  if (AR->getLoop() == L)
2584  Strides.insert(AR->getStepRecurrence(SE));
2585  Worklist.push_back(AR->getStart());
2586  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2587  Worklist.append(Add->op_begin(), Add->op_end());
2588  }
2589  } while (!Worklist.empty());
2590  }
2591 
2592  // Compute interesting factors from the set of interesting strides.
2594  I = Strides.begin(), E = Strides.end(); I != E; ++I)
2596  std::next(I); NewStrideIter != E; ++NewStrideIter) {
2597  const SCEV *OldStride = *I;
2598  const SCEV *NewStride = *NewStrideIter;
2599 
2600  if (SE.getTypeSizeInBits(OldStride->getType()) !=
2601  SE.getTypeSizeInBits(NewStride->getType())) {
2602  if (SE.getTypeSizeInBits(OldStride->getType()) >
2603  SE.getTypeSizeInBits(NewStride->getType()))
2604  NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2605  else
2606  OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2607  }
2608  if (const SCEVConstant *Factor =
2609  dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2610  SE, true))) {
2611  if (Factor->getAPInt().getMinSignedBits() <= 64)
2612  Factors.insert(Factor->getAPInt().getSExtValue());
2613  } else if (const SCEVConstant *Factor =
2614  dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2615  NewStride,
2616  SE, true))) {
2617  if (Factor->getAPInt().getMinSignedBits() <= 64)
2618  Factors.insert(Factor->getAPInt().getSExtValue());
2619  }
2620  }
2621 
2622  // If all uses use the same type, don't bother looking for truncation-based
2623  // reuse.
2624  if (Types.size() == 1)
2625  Types.clear();
2626 
2627  DEBUG(print_factors_and_types(dbgs()));
2628 }
2629 
2630 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2631 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2632 /// IVStrideUses, we could partially skip this.
2633 static User::op_iterator
2635  Loop *L, ScalarEvolution &SE) {
2636  for(; OI != OE; ++OI) {
2637  if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2638  if (!SE.isSCEVable(Oper->getType()))
2639  continue;
2640 
2641  if (const SCEVAddRecExpr *AR =
2642  dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2643  if (AR->getLoop() == L)
2644  break;
2645  }
2646  }
2647  }
2648  return OI;
2649 }
2650 
2651 /// IVChain logic must consistenctly peek base TruncInst operands, so wrap it in
2652 /// a convenient helper.
2653 static Value *getWideOperand(Value *Oper) {
2654  if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2655  return Trunc->getOperand(0);
2656  return Oper;
2657 }
2658 
2659 /// Return true if we allow an IV chain to include both types.
2660 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2661  Type *LType = LVal->getType();
2662  Type *RType = RVal->getType();
2663  return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2664  // Different address spaces means (possibly)
2665  // different types of the pointer implementation,
2666  // e.g. i16 vs i32 so disallow that.
2667  (LType->getPointerAddressSpace() ==
2668  RType->getPointerAddressSpace()));
2669 }
2670 
2671 /// Return an approximation of this SCEV expression's "base", or NULL for any
2672 /// constant. Returning the expression itself is conservative. Returning a
2673 /// deeper subexpression is more precise and valid as long as it isn't less
2674 /// complex than another subexpression. For expressions involving multiple
2675 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2676 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2677 /// IVInc==b-a.
2678 ///
2679 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2680 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2681 static const SCEV *getExprBase(const SCEV *S) {
2682  switch (S->getSCEVType()) {
2683  default: // uncluding scUnknown.
2684  return S;
2685  case scConstant:
2686  return nullptr;
2687  case scTruncate:
2688  return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2689  case scZeroExtend:
2690  return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2691  case scSignExtend:
2692  return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2693  case scAddExpr: {
2694  // Skip over scaled operands (scMulExpr) to follow add operands as long as
2695  // there's nothing more complex.
2696  // FIXME: not sure if we want to recognize negation.
2697  const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2698  for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2699  E(Add->op_begin()); I != E; ++I) {
2700  const SCEV *SubExpr = *I;
2701  if (SubExpr->getSCEVType() == scAddExpr)
2702  return getExprBase(SubExpr);
2703 
2704  if (SubExpr->getSCEVType() != scMulExpr)
2705  return SubExpr;
2706  }
2707  return S; // all operands are scaled, be conservative.
2708  }
2709  case scAddRecExpr:
2710  return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2711  }
2712 }
2713 
2714 /// Return true if the chain increment is profitable to expand into a loop
2715 /// invariant value, which may require its own register. A profitable chain
2716 /// increment will be an offset relative to the same base. We allow such offsets
2717 /// to potentially be used as chain increment as long as it's not obviously
2718 /// expensive to expand using real instructions.
2719 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2720  const SCEV *IncExpr,
2721  ScalarEvolution &SE) {
2722  // Aggressively form chains when -stress-ivchain.
2723  if (StressIVChain)
2724  return true;
2725 
2726  // Do not replace a constant offset from IV head with a nonconstant IV
2727  // increment.
2728  if (!isa<SCEVConstant>(IncExpr)) {
2729  const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2730  if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2731  return false;
2732  }
2733 
2734  SmallPtrSet<const SCEV*, 8> Processed;
2735  return !isHighCostExpansion(IncExpr, Processed, SE);
2736 }
2737 
2738 /// Return true if the number of registers needed for the chain is estimated to
2739 /// be less than the number required for the individual IV users. First prohibit
2740 /// any IV users that keep the IV live across increments (the Users set should
2741 /// be empty). Next count the number and type of increments in the chain.
2742 ///
2743 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2744 /// effectively use postinc addressing modes. Only consider it profitable it the
2745 /// increments can be computed in fewer registers when chained.
2746 ///
2747 /// TODO: Consider IVInc free if it's already used in another chains.
2748 static bool
2750  ScalarEvolution &SE, const TargetTransformInfo &TTI) {
2751  if (StressIVChain)
2752  return true;
2753 
2754  if (!Chain.hasIncs())
2755  return false;
2756 
2757  if (!Users.empty()) {
2758  DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2759  for (Instruction *Inst : Users) {
2760  dbgs() << " " << *Inst << "\n";
2761  });
2762  return false;
2763  }
2764  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2765 
2766  // The chain itself may require a register, so intialize cost to 1.
2767  int cost = 1;
2768 
2769  // A complete chain likely eliminates the need for keeping the original IV in
2770  // a register. LSR does not currently know how to form a complete chain unless
2771  // the header phi already exists.
2772  if (isa<PHINode>(Chain.tailUserInst())
2773  && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2774  --cost;
2775  }
2776  const SCEV *LastIncExpr = nullptr;
2777  unsigned NumConstIncrements = 0;
2778  unsigned NumVarIncrements = 0;
2779  unsigned NumReusedIncrements = 0;
2780  for (const IVInc &Inc : Chain) {
2781  if (Inc.IncExpr->isZero())
2782  continue;
2783 
2784  // Incrementing by zero or some constant is neutral. We assume constants can
2785  // be folded into an addressing mode or an add's immediate operand.
2786  if (isa<SCEVConstant>(Inc.IncExpr)) {
2787  ++NumConstIncrements;
2788  continue;
2789  }
2790 
2791  if (Inc.IncExpr == LastIncExpr)
2792  ++NumReusedIncrements;
2793  else
2794  ++NumVarIncrements;
2795 
2796  LastIncExpr = Inc.IncExpr;
2797  }
2798  // An IV chain with a single increment is handled by LSR's postinc
2799  // uses. However, a chain with multiple increments requires keeping the IV's
2800  // value live longer than it needs to be if chained.
2801  if (NumConstIncrements > 1)
2802  --cost;
2803 
2804  // Materializing increment expressions in the preheader that didn't exist in
2805  // the original code may cost a register. For example, sign-extended array
2806  // indices can produce ridiculous increments like this:
2807  // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2808  cost += NumVarIncrements;
2809 
2810  // Reusing variable increments likely saves a register to hold the multiple of
2811  // the stride.
2812  cost -= NumReusedIncrements;
2813 
2814  DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2815  << "\n");
2816 
2817  return cost < 0;
2818 }
2819 
2820 /// Add this IV user to an existing chain or make it the head of a new chain.
2821 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2822  SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2823  // When IVs are used as types of varying widths, they are generally converted
2824  // to a wider type with some uses remaining narrow under a (free) trunc.
2825  Value *const NextIV = getWideOperand(IVOper);
2826  const SCEV *const OperExpr = SE.getSCEV(NextIV);
2827  const SCEV *const OperExprBase = getExprBase(OperExpr);
2828 
2829  // Visit all existing chains. Check if its IVOper can be computed as a
2830  // profitable loop invariant increment from the last link in the Chain.
2831  unsigned ChainIdx = 0, NChains = IVChainVec.size();
2832  const SCEV *LastIncExpr = nullptr;
2833  for (; ChainIdx < NChains; ++ChainIdx) {
2834  IVChain &Chain = IVChainVec[ChainIdx];
2835 
2836  // Prune the solution space aggressively by checking that both IV operands
2837  // are expressions that operate on the same unscaled SCEVUnknown. This
2838  // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2839  // first avoids creating extra SCEV expressions.
2840  if (!StressIVChain && Chain.ExprBase != OperExprBase)
2841  continue;
2842 
2843  Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2844  if (!isCompatibleIVType(PrevIV, NextIV))
2845  continue;
2846 
2847  // A phi node terminates a chain.
2848  if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2849  continue;
2850 
2851  // The increment must be loop-invariant so it can be kept in a register.
2852  const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2853  const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2854  if (!SE.isLoopInvariant(IncExpr, L))
2855  continue;
2856 
2857  if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2858  LastIncExpr = IncExpr;
2859  break;
2860  }
2861  }
2862  // If we haven't found a chain, create a new one, unless we hit the max. Don't
2863  // bother for phi nodes, because they must be last in the chain.
2864  if (ChainIdx == NChains) {
2865  if (isa<PHINode>(UserInst))
2866  return;
2867  if (NChains >= MaxChains && !StressIVChain) {
2868  DEBUG(dbgs() << "IV Chain Limit\n");
2869  return;
2870  }
2871  LastIncExpr = OperExpr;
2872  // IVUsers may have skipped over sign/zero extensions. We don't currently
2873  // attempt to form chains involving extensions unless they can be hoisted
2874  // into this loop's AddRec.
2875  if (!isa<SCEVAddRecExpr>(LastIncExpr))
2876  return;
2877  ++NChains;
2878  IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2879  OperExprBase));
2880  ChainUsersVec.resize(NChains);
2881  DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2882  << ") IV=" << *LastIncExpr << "\n");
2883  } else {
2884  DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2885  << ") IV+" << *LastIncExpr << "\n");
2886  // Add this IV user to the end of the chain.
2887  IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2888  }
2889  IVChain &Chain = IVChainVec[ChainIdx];
2890 
2891  SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2892  // This chain's NearUsers become FarUsers.
2893  if (!LastIncExpr->isZero()) {
2894  ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2895  NearUsers.end());
2896  NearUsers.clear();
2897  }
2898 
2899  // All other uses of IVOperand become near uses of the chain.
2900  // We currently ignore intermediate values within SCEV expressions, assuming
2901  // they will eventually be used be the current chain, or can be computed
2902  // from one of the chain increments. To be more precise we could
2903  // transitively follow its user and only add leaf IV users to the set.
2904  for (User *U : IVOper->users()) {
2905  Instruction *OtherUse = dyn_cast<Instruction>(U);
2906  if (!OtherUse)
2907  continue;
2908  // Uses in the chain will no longer be uses if the chain is formed.
2909  // Include the head of the chain in this iteration (not Chain.begin()).
2910  IVChain::const_iterator IncIter = Chain.Incs.begin();
2911  IVChain::const_iterator IncEnd = Chain.Incs.end();
2912  for( ; IncIter != IncEnd; ++IncIter) {
2913  if (IncIter->UserInst == OtherUse)
2914  break;
2915  }
2916  if (IncIter != IncEnd)
2917  continue;
2918 
2919  if (SE.isSCEVable(OtherUse->getType())
2920  && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2921  && IU.isIVUserOrOperand(OtherUse)) {
2922  continue;
2923  }
2924  NearUsers.insert(OtherUse);
2925  }
2926 
2927  // Since this user is part of the chain, it's no longer considered a use
2928  // of the chain.
2929  ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
2930 }
2931 
2932 /// Populate the vector of Chains.
2933 ///
2934 /// This decreases ILP at the architecture level. Targets with ample registers,
2935 /// multiple memory ports, and no register renaming probably don't want
2936 /// this. However, such targets should probably disable LSR altogether.
2937 ///
2938 /// The job of LSR is to make a reasonable choice of induction variables across
2939 /// the loop. Subsequent passes can easily "unchain" computation exposing more
2940 /// ILP *within the loop* if the target wants it.
2941 ///
2942 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
2943 /// will not reorder memory operations, it will recognize this as a chain, but
2944 /// will generate redundant IV increments. Ideally this would be corrected later
2945 /// by a smart scheduler:
2946 /// = A[i]
2947 /// = A[i+x]
2948 /// A[i] =
2949 /// A[i+x] =
2950 ///
2951 /// TODO: Walk the entire domtree within this loop, not just the path to the
2952 /// loop latch. This will discover chains on side paths, but requires
2953 /// maintaining multiple copies of the Chains state.
2954 void LSRInstance::CollectChains() {
2955  DEBUG(dbgs() << "Collecting IV Chains.\n");
2956  SmallVector<ChainUsers, 8> ChainUsersVec;
2957 
2958  SmallVector<BasicBlock *,8> LatchPath;
2959  BasicBlock *LoopHeader = L->getHeader();
2960  for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
2961  Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
2962  LatchPath.push_back(Rung->getBlock());
2963  }
2964  LatchPath.push_back(LoopHeader);
2965 
2966  // Walk the instruction stream from the loop header to the loop latch.
2967  for (BasicBlock *BB : reverse(LatchPath)) {
2968  for (Instruction &I : *BB) {
2969  // Skip instructions that weren't seen by IVUsers analysis.
2970  if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
2971  continue;
2972 
2973  // Ignore users that are part of a SCEV expression. This way we only
2974  // consider leaf IV Users. This effectively rediscovers a portion of
2975  // IVUsers analysis but in program order this time.
2976  if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
2977  continue;
2978 
2979  // Remove this instruction from any NearUsers set it may be in.
2980  for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
2981  ChainIdx < NChains; ++ChainIdx) {
2982  ChainUsersVec[ChainIdx].NearUsers.erase(&I);
2983  }
2984  // Search for operands that can be chained.
2985  SmallPtrSet<Instruction*, 4> UniqueOperands;
2986  User::op_iterator IVOpEnd = I.op_end();
2987  User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
2988  while (IVOpIter != IVOpEnd) {
2989  Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
2990  if (UniqueOperands.insert(IVOpInst).second)
2991  ChainInstruction(&I, IVOpInst, ChainUsersVec);
2992  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
2993  }
2994  } // Continue walking down the instructions.
2995  } // Continue walking down the domtree.
2996  // Visit phi backedges to determine if the chain can generate the IV postinc.
2997  for (BasicBlock::iterator I = L->getHeader()->begin();
2998  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2999  if (!SE.isSCEVable(PN->getType()))
3000  continue;
3001 
3002  Instruction *IncV =
3004  if (IncV)
3005  ChainInstruction(PN, IncV, ChainUsersVec);
3006  }
3007  // Remove any unprofitable chains.
3008  unsigned ChainIdx = 0;
3009  for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3010  UsersIdx < NChains; ++UsersIdx) {
3011  if (!isProfitableChain(IVChainVec[UsersIdx],
3012  ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3013  continue;
3014  // Preserve the chain at UsesIdx.
3015  if (ChainIdx != UsersIdx)
3016  IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3017  FinalizeChain(IVChainVec[ChainIdx]);
3018  ++ChainIdx;
3019  }
3020  IVChainVec.resize(ChainIdx);
3021 }
3022 
3023 void LSRInstance::FinalizeChain(IVChain &Chain) {
3024  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3025  DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
3026 
3027  for (const IVInc &Inc : Chain) {
3028  DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n");
3029  auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3030  assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
3031  IVIncSet.insert(UseI);
3032  }
3033 }
3034 
3035 /// Return true if the IVInc can be folded into an addressing mode.
3036 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3037  Value *Operand, const TargetTransformInfo &TTI) {
3038  const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3039  if (!IncConst || !isAddressUse(UserInst, Operand))
3040  return false;
3041 
3042  if (IncConst->getAPInt().getMinSignedBits() > 64)
3043  return false;
3044 
3045  MemAccessTy AccessTy = getAccessType(UserInst);
3046  int64_t IncOffset = IncConst->getValue()->getSExtValue();
3047  if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3048  IncOffset, /*HaseBaseReg=*/false))
3049  return false;
3050 
3051  return true;
3052 }
3053 
3054 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3055 /// user's operand from the previous IV user's operand.
3056 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3057  SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3058  // Find the new IVOperand for the head of the chain. It may have been replaced
3059  // by LSR.
3060  const IVInc &Head = Chain.Incs[0];
3061  User::op_iterator IVOpEnd = Head.UserInst->op_end();
3062  // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3063  User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3064  IVOpEnd, L, SE);
3065  Value *IVSrc = nullptr;
3066  while (IVOpIter != IVOpEnd) {
3067  IVSrc = getWideOperand(*IVOpIter);
3068 
3069  // If this operand computes the expression that the chain needs, we may use
3070  // it. (Check this after setting IVSrc which is used below.)
3071  //
3072  // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3073  // narrow for the chain, so we can no longer use it. We do allow using a
3074  // wider phi, assuming the LSR checked for free truncation. In that case we
3075  // should already have a truncate on this operand such that
3076  // getSCEV(IVSrc) == IncExpr.
3077  if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3078  || SE.getSCEV(IVSrc) == Head.IncExpr) {
3079  break;
3080  }
3081  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3082  }
3083  if (IVOpIter == IVOpEnd) {
3084  // Gracefully give up on this chain.
3085  DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3086  return;
3087  }
3088 
3089  DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3090  Type *IVTy = IVSrc->getType();
3091  Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3092  const SCEV *LeftOverExpr = nullptr;
3093  for (const IVInc &Inc : Chain) {
3094  Instruction *InsertPt = Inc.UserInst;
3095  if (isa<PHINode>(InsertPt))
3096  InsertPt = L->getLoopLatch()->getTerminator();
3097 
3098  // IVOper will replace the current IV User's operand. IVSrc is the IV
3099  // value currently held in a register.
3100  Value *IVOper = IVSrc;
3101  if (!Inc.IncExpr->isZero()) {
3102  // IncExpr was the result of subtraction of two narrow values, so must
3103  // be signed.
3104  const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3105  LeftOverExpr = LeftOverExpr ?
3106  SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3107  }
3108  if (LeftOverExpr && !LeftOverExpr->isZero()) {
3109  // Expand the IV increment.
3110  Rewriter.clearPostInc();
3111  Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3112  const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3113  SE.getUnknown(IncV));
3114  IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3115 
3116  // If an IV increment can't be folded, use it as the next IV value.
3117  if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3118  assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3119  IVSrc = IVOper;
3120  LeftOverExpr = nullptr;
3121  }
3122  }
3123  Type *OperTy = Inc.IVOperand->getType();
3124  if (IVTy != OperTy) {
3125  assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3126  "cannot extend a chained IV");
3127  IRBuilder<> Builder(InsertPt);
3128  IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3129  }
3130  Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3131  DeadInsts.emplace_back(Inc.IVOperand);
3132  }
3133  // If LSR created a new, wider phi, we may also replace its postinc. We only
3134  // do this if we also found a wide value for the head of the chain.
3135  if (isa<PHINode>(Chain.tailUserInst())) {
3136  for (BasicBlock::iterator I = L->getHeader()->begin();
3137  PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
3138  if (!isCompatibleIVType(Phi, IVSrc))
3139  continue;
3140  Instruction *PostIncV = dyn_cast<Instruction>(
3141  Phi->getIncomingValueForBlock(L->getLoopLatch()));
3142  if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3143  continue;
3144  Value *IVOper = IVSrc;
3145  Type *PostIncTy = PostIncV->getType();
3146  if (IVTy != PostIncTy) {
3147  assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3148  IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3149  Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3150  IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3151  }
3152  Phi->replaceUsesOfWith(PostIncV, IVOper);
3153  DeadInsts.emplace_back(PostIncV);
3154  }
3155  }
3156 }
3157 
3158 void LSRInstance::CollectFixupsAndInitialFormulae() {
3159  for (const IVStrideUse &U : IU) {
3160  Instruction *UserInst = U.getUser();
3161  // Skip IV users that are part of profitable IV Chains.
3162  User::op_iterator UseI =
3163  find(UserInst->operands(), U.getOperandValToReplace());
3164  assert(UseI != UserInst->op_end() && "cannot find IV operand");
3165  if (IVIncSet.count(UseI)) {
3166  DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3167  continue;
3168  }
3169 
3170  LSRUse::KindType Kind = LSRUse::Basic;
3171  MemAccessTy AccessTy;
3172  if (isAddressUse(UserInst, U.getOperandValToReplace())) {
3173  Kind = LSRUse::Address;
3174  AccessTy = getAccessType(UserInst);
3175  }
3176 
3177  const SCEV *S = IU.getExpr(U);
3178  PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3179 
3180  // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3181  // (N - i == 0), and this allows (N - i) to be the expression that we work
3182  // with rather than just N or i, so we can consider the register
3183  // requirements for both N and i at the same time. Limiting this code to
3184  // equality icmps is not a problem because all interesting loops use
3185  // equality icmps, thanks to IndVarSimplify.
3186  if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst))
3187  if (CI->isEquality()) {
3188  // Swap the operands if needed to put the OperandValToReplace on the
3189  // left, for consistency.
3190  Value *NV = CI->getOperand(1);
3191  if (NV == U.getOperandValToReplace()) {
3192  CI->setOperand(1, CI->getOperand(0));
3193  CI->setOperand(0, NV);
3194  NV = CI->getOperand(1);
3195  Changed = true;
3196  }
3197 
3198  // x == y --> x - y == 0
3199  const SCEV *N = SE.getSCEV(NV);
3200  if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3201  // S is normalized, so normalize N before folding it into S
3202  // to keep the result normalized.
3203  N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3204  Kind = LSRUse::ICmpZero;
3205  S = SE.getMinusSCEV(N, S);
3206  }
3207 
3208  // -1 and the negations of all interesting strides (except the negation
3209  // of -1) are now also interesting.
3210  for (size_t i = 0, e = Factors.size(); i != e; ++i)
3211  if (Factors[i] != -1)
3212  Factors.insert(-(uint64_t)Factors[i]);
3213  Factors.insert(-1);
3214  }
3215 
3216  // Get or create an LSRUse.
3217  std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3218  size_t LUIdx = P.first;
3219  int64_t Offset = P.second;
3220  LSRUse &LU = Uses[LUIdx];
3221 
3222  // Record the fixup.
3223  LSRFixup &LF = LU.getNewFixup();
3224  LF.UserInst = UserInst;
3225  LF.OperandValToReplace = U.getOperandValToReplace();
3226  LF.PostIncLoops = TmpPostIncLoops;
3227  LF.Offset = Offset;
3228  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3229 
3230  if (!LU.WidestFixupType ||
3231  SE.getTypeSizeInBits(LU.WidestFixupType) <
3232  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3233  LU.WidestFixupType = LF.OperandValToReplace->getType();
3234 
3235  // If this is the first use of this LSRUse, give it a formula.
3236  if (LU.Formulae.empty()) {
3237  InsertInitialFormula(S, LU, LUIdx);
3238  CountRegisters(LU.Formulae.back(), LUIdx);
3239  }
3240  }
3241 
3242  DEBUG(print_fixups(dbgs()));
3243 }
3244 
3245 /// Insert a formula for the given expression into the given use, separating out
3246 /// loop-variant portions from loop-invariant and loop-computable portions.
3247 void
3248 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3249  // Mark uses whose expressions cannot be expanded.
3250  if (!isSafeToExpand(S, SE))
3251  LU.RigidFormula = true;
3252 
3253  Formula F;
3254  F.initialMatch(S, L, SE);
3255  bool Inserted = InsertFormula(LU, LUIdx, F);
3256  assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3257 }
3258 
3259 /// Insert a simple single-register formula for the given expression into the
3260 /// given use.
3261 void
3262 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3263  LSRUse &LU, size_t LUIdx) {
3264  Formula F;
3265  F.BaseRegs.push_back(S);
3266  F.HasBaseReg = true;
3267  bool Inserted = InsertFormula(LU, LUIdx, F);
3268  assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3269 }
3270 
3271 /// Note which registers are used by the given formula, updating RegUses.
3272 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3273  if (F.ScaledReg)
3274  RegUses.countRegister(F.ScaledReg, LUIdx);
3275  for (const SCEV *BaseReg : F.BaseRegs)
3276  RegUses.countRegister(BaseReg, LUIdx);
3277 }
3278 
3279 /// If the given formula has not yet been inserted, add it to the list, and
3280 /// return true. Return false otherwise.
3281 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3282  // Do not insert formula that we will not be able to expand.
3283  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3284  "Formula is illegal");
3285 
3286  if (!LU.InsertFormula(F, *L))
3287  return false;
3288 
3289  CountRegisters(F, LUIdx);
3290  return true;
3291 }
3292 
3293 /// Check for other uses of loop-invariant values which we're tracking. These
3294 /// other uses will pin these values in registers, making them less profitable
3295 /// for elimination.
3296 /// TODO: This currently misses non-constant addrec step registers.
3297 /// TODO: Should this give more weight to users inside the loop?
3298 void
3299 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3300  SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3302 
3303  while (!Worklist.empty()) {
3304  const SCEV *S = Worklist.pop_back_val();
3305 
3306  // Don't process the same SCEV twice
3307  if (!Visited.insert(S).second)
3308  continue;
3309 
3310  if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3311  Worklist.append(N->op_begin(), N->op_end());
3312  else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3313  Worklist.push_back(C->getOperand());
3314  else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3315  Worklist.push_back(D->getLHS());
3316  Worklist.push_back(D->getRHS());
3317  } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3318  const Value *V = US->getValue();
3319  if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3320  // Look for instructions defined outside the loop.
3321  if (L->contains(Inst)) continue;
3322  } else if (isa<UndefValue>(V))
3323  // Undef doesn't have a live range, so it doesn't matter.
3324  continue;
3325  for (const Use &U : V->uses()) {
3326  const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3327  // Ignore non-instructions.
3328  if (!UserInst)
3329  continue;
3330  // Ignore instructions in other functions (as can happen with
3331  // Constants).
3332  if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3333  continue;
3334  // Ignore instructions not dominated by the loop.
3335  const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3336  UserInst->getParent() :
3337  cast<PHINode>(UserInst)->getIncomingBlock(
3338  PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3339  if (!DT.dominates(L->getHeader(), UseBB))
3340  continue;
3341  // Don't bother if the instruction is in a BB which ends in an EHPad.
3342  if (UseBB->getTerminator()->isEHPad())
3343  continue;
3344  // Don't bother rewriting PHIs in catchswitch blocks.
3345  if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3346  continue;
3347  // Ignore uses which are part of other SCEV expressions, to avoid
3348  // analyzing them multiple times.
3349  if (SE.isSCEVable(UserInst->getType())) {
3350  const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3351  // If the user is a no-op, look through to its uses.
3352  if (!isa<SCEVUnknown>(UserS))
3353  continue;
3354  if (UserS == US) {
3355  Worklist.push_back(
3356  SE.getUnknown(const_cast<Instruction *>(UserInst)));
3357  continue;
3358  }
3359  }
3360  // Ignore icmp instructions which are already being analyzed.
3361  if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3362  unsigned OtherIdx = !U.getOperandNo();
3363  Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3364  if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3365  continue;
3366  }
3367 
3368  std::pair<size_t, int64_t> P = getUse(
3369  S, LSRUse::Basic, MemAccessTy());
3370  size_t LUIdx = P.first;
3371  int64_t Offset = P.second;
3372  LSRUse &LU = Uses[LUIdx];
3373  LSRFixup &LF = LU.getNewFixup();
3374  LF.UserInst = const_cast<Instruction *>(UserInst);
3375  LF.OperandValToReplace = U;
3376  LF.Offset = Offset;
3377  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3378  if (!LU.WidestFixupType ||
3379  SE.getTypeSizeInBits(LU.WidestFixupType) <
3380  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3381  LU.WidestFixupType = LF.OperandValToReplace->getType();
3382  InsertSupplementalFormula(US, LU, LUIdx);
3383  CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3384  break;
3385  }
3386  }
3387  }
3388 }
3389 
3390 /// Split S into subexpressions which can be pulled out into separate
3391 /// registers. If C is non-null, multiply each subexpression by C.
3392 ///
3393 /// Return remainder expression after factoring the subexpressions captured by
3394 /// Ops. If Ops is complete, return NULL.
3395 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3397  const Loop *L,
3398  ScalarEvolution &SE,
3399  unsigned Depth = 0) {
3400  // Arbitrarily cap recursion to protect compile time.
3401  if (Depth >= 3)
3402  return S;
3403 
3404  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3405  // Break out add operands.
3406  for (const SCEV *S : Add->operands()) {
3407  const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3408  if (Remainder)
3409  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3410  }
3411  return nullptr;
3412  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3413  // Split a non-zero base out of an addrec.
3414  if (AR->getStart()->isZero() || !AR->isAffine())
3415  return S;
3416 
3417  const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3418  C, Ops, L, SE, Depth+1);
3419  // Split the non-zero AddRec unless it is part of a nested recurrence that
3420  // does not pertain to this loop.
3421  if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3422  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3423  Remainder = nullptr;
3424  }
3425  if (Remainder != AR->getStart()) {
3426  if (!Remainder)
3427  Remainder = SE.getConstant(AR->getType(), 0);
3428  return SE.getAddRecExpr(Remainder,
3429  AR->getStepRecurrence(SE),
3430  AR->getLoop(),
3431  //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3433  }
3434  } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3435  // Break (C * (a + b + c)) into C*a + C*b + C*c.
3436  if (Mul->getNumOperands() != 2)
3437  return S;
3438  if (const SCEVConstant *Op0 =
3439  dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3440  C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3441  const SCEV *Remainder =
3442  CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3443  if (Remainder)
3444  Ops.push_back(SE.getMulExpr(C, Remainder));
3445  return nullptr;
3446  }
3447  }
3448  return S;
3449 }
3450 
3451 /// \brief Helper function for LSRInstance::GenerateReassociations.
3452 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3453  const Formula &Base,
3454  unsigned Depth, size_t Idx,
3455  bool IsScaledReg) {
3456  const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3458  const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3459  if (Remainder)
3460  AddOps.push_back(Remainder);
3461 
3462  if (AddOps.size() == 1)
3463  return;
3464 
3466  JE = AddOps.end();
3467  J != JE; ++J) {
3468 
3469  // Loop-variant "unknown" values are uninteresting; we won't be able to
3470  // do anything meaningful with them.
3471  if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3472  continue;
3473 
3474  // Don't pull a constant into a register if the constant could be folded
3475  // into an immediate field.
3476  if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3477  LU.AccessTy, *J, Base.getNumRegs() > 1))
3478  continue;
3479 
3480  // Collect all operands except *J.
3481  SmallVector<const SCEV *, 8> InnerAddOps(
3482  ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3483  InnerAddOps.append(std::next(J),
3484  ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3485 
3486  // Don't leave just a constant behind in a register if the constant could
3487  // be folded into an immediate field.
3488  if (InnerAddOps.size() == 1 &&
3489  isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3490  LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3491  continue;
3492 
3493  const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3494  if (InnerSum->isZero())
3495  continue;
3496  Formula F = Base;
3497 
3498  // Add the remaining pieces of the add back into the new formula.
3499  const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3500  if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3501  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3502  InnerSumSC->getValue()->getZExtValue())) {
3503  F.UnfoldedOffset =
3504  (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3505  if (IsScaledReg)
3506  F.ScaledReg = nullptr;
3507  else
3508  F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3509  } else if (IsScaledReg)
3510  F.ScaledReg = InnerSum;
3511  else
3512  F.BaseRegs[Idx] = InnerSum;
3513 
3514  // Add J as its own register, or an unfolded immediate.
3515  const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3516  if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3517  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3518  SC->getValue()->getZExtValue()))
3519  F.UnfoldedOffset =
3520  (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3521  else
3522  F.BaseRegs.push_back(*J);
3523  // We may have changed the number of register in base regs, adjust the
3524  // formula accordingly.
3525  F.canonicalize(*L);
3526 
3527  if (InsertFormula(LU, LUIdx, F))
3528  // If that formula hadn't been seen before, recurse to find more like
3529  // it.
3530  GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1);
3531  }
3532 }
3533 
3534 /// Split out subexpressions from adds and the bases of addrecs.
3535 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3536  Formula Base, unsigned Depth) {
3537  assert(Base.isCanonical(*L) && "Input must be in the canonical form");
3538  // Arbitrarily cap recursion to protect compile time.
3539  if (Depth >= 3)
3540  return;
3541 
3542  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3543  GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3544 
3545  if (Base.Scale == 1)
3546  GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3547  /* Idx */ -1, /* IsScaledReg */ true);
3548 }
3549 
3550 /// Generate a formula consisting of all of the loop-dominating registers added
3551 /// into a single register.
3552 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3553  Formula Base) {
3554  // This method is only interesting on a plurality of registers.
3555  if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1)
3556  return;
3557 
3558  // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3559  // processing the formula.
3560  Base.unscale();
3561  Formula F = Base;
3562  F.BaseRegs.clear();
3564  for (const SCEV *BaseReg : Base.BaseRegs) {
3565  if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3566  !SE.hasComputableLoopEvolution(BaseReg, L))
3567  Ops.push_back(BaseReg);
3568  else
3569  F.BaseRegs.push_back(BaseReg);
3570  }
3571  if (Ops.size() > 1) {
3572  const SCEV *Sum = SE.getAddExpr(Ops);
3573  // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3574  // opportunity to fold something. For now, just ignore such cases
3575  // rather than proceed with zero in a register.
3576  if (!Sum->isZero()) {
3577  F.BaseRegs.push_back(Sum);
3578  F.canonicalize(*L);
3579  (void)InsertFormula(LU, LUIdx, F);
3580  }
3581  }
3582 }
3583 
3584 /// \brief Helper function for LSRInstance::GenerateSymbolicOffsets.
3585 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3586  const Formula &Base, size_t Idx,
3587  bool IsScaledReg) {
3588  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3589  GlobalValue *GV = ExtractSymbol(G, SE);
3590  if (G->isZero() || !GV)
3591  return;
3592  Formula F = Base;
3593  F.BaseGV = GV;
3594  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3595  return;
3596  if (IsScaledReg)
3597  F.ScaledReg = G;
3598  else
3599  F.BaseRegs[Idx] = G;
3600  (void)InsertFormula(LU, LUIdx, F);
3601 }
3602 
3603 /// Generate reuse formulae using symbolic offsets.
3604 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3605  Formula Base) {
3606  // We can't add a symbolic offset if the address already contains one.
3607  if (Base.BaseGV) return;
3608 
3609  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3610  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3611  if (Base.Scale == 1)
3612  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3613  /* IsScaledReg */ true);
3614 }
3615 
3616 /// \brief Helper function for LSRInstance::GenerateConstantOffsets.
3617 void LSRInstance::GenerateConstantOffsetsImpl(
3618  LSRUse &LU, unsigned LUIdx, const Formula &Base,
3619  const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3620  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3621  for (int64_t Offset : Worklist) {
3622  Formula F = Base;
3623  F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3624  if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3625  LU.AccessTy, F)) {
3626  // Add the offset to the base register.
3627  const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3628  // If it cancelled out, drop the base register, otherwise update it.
3629  if (NewG->isZero()) {
3630  if (IsScaledReg) {
3631  F.Scale = 0;
3632  F.ScaledReg = nullptr;
3633  } else
3634  F.deleteBaseReg(F.BaseRegs[Idx]);
3635  F.canonicalize(*L);
3636  } else if (IsScaledReg)
3637  F.ScaledReg = NewG;
3638  else
3639  F.BaseRegs[Idx] = NewG;
3640 
3641  (void)InsertFormula(LU, LUIdx, F);
3642  }
3643  }
3644 
3645  int64_t Imm = ExtractImmediate(G, SE);
3646  if (G->isZero() || Imm == 0)
3647  return;
3648  Formula F = Base;
3649  F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3650  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3651  return;
3652  if (IsScaledReg)
3653  F.ScaledReg = G;
3654  else
3655  F.BaseRegs[Idx] = G;
3656  (void)InsertFormula(LU, LUIdx, F);
3657 }
3658 
3659 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3660 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3661  Formula Base) {
3662  // TODO: For now, just add the min and max offset, because it usually isn't
3663  // worthwhile looking at everything inbetween.
3664  SmallVector<int64_t, 2> Worklist;
3665  Worklist.push_back(LU.MinOffset);
3666  if (LU.MaxOffset != LU.MinOffset)
3667  Worklist.push_back(LU.MaxOffset);
3668 
3669  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3670  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3671  if (Base.Scale == 1)
3672  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3673  /* IsScaledReg */ true);
3674 }
3675 
3676 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3677 /// == y -> x*c == y*c.
3678 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3679  Formula Base) {
3680  if (LU.Kind != LSRUse::ICmpZero) return;
3681 
3682  // Determine the integer type for the base formula.
3683  Type *IntTy = Base.getType();
3684  if (!IntTy) return;
3685  if (SE.getTypeSizeInBits(IntTy) > 64) return;
3686 
3687  // Don't do this if there is more than one offset.
3688  if (LU.MinOffset != LU.MaxOffset) return;
3689 
3690  // Check if transformation is valid. It is illegal to multiply pointer.
3691  if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3692  return;
3693  for (const SCEV *BaseReg : Base.BaseRegs)
3694  if (BaseReg->getType()->isPointerTy())
3695  return;
3696  assert(!Base.BaseGV && "ICmpZero use is not legal!");
3697 
3698  // Check each interesting stride.
3699  for (int64_t Factor : Factors) {
3700  // Check that the multiplication doesn't overflow.
3701  if (Base.BaseOffset == INT64_MIN && Factor == -1)
3702  continue;
3703  int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3704  if (NewBaseOffset / Factor != Base.BaseOffset)
3705  continue;
3706  // If the offset will be truncated at this use, check that it is in bounds.
3707  if (!IntTy->isPointerTy() &&
3708  !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3709  continue;
3710 
3711  // Check that multiplying with the use offset doesn't overflow.
3712  int64_t Offset = LU.MinOffset;
3713  if (Offset == INT64_MIN && Factor == -1)
3714  continue;
3715  Offset = (uint64_t)Offset * Factor;
3716  if (Offset / Factor != LU.MinOffset)
3717  continue;
3718  // If the offset will be truncated at this use, check that it is in bounds.
3719  if (!IntTy->isPointerTy() &&
3720  !ConstantInt::isValueValidForType(IntTy, Offset))
3721  continue;
3722 
3723  Formula F = Base;
3724  F.BaseOffset = NewBaseOffset;
3725 
3726  // Check that this scale is legal.
3727  if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3728  continue;
3729 
3730  // Compensate for the use having MinOffset built into it.
3731  F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3732 
3733  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3734 
3735  // Check that multiplying with each base register doesn't overflow.
3736  for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3737  F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3738  if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3739  goto next;
3740  }
3741 
3742  // Check that multiplying with the scaled register doesn't overflow.
3743  if (F.ScaledReg) {
3744  F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3745  if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3746  continue;
3747  }
3748 
3749  // Check that multiplying with the unfolded offset doesn't overflow.
3750  if (F.UnfoldedOffset != 0) {
3751  if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
3752  continue;
3753  F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3754  if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3755  continue;
3756  // If the offset will be truncated, check that it is in bounds.
3757  if (!IntTy->isPointerTy() &&
3758  !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3759  continue;
3760  }
3761 
3762  // If we make it here and it's legal, add it.
3763  (void)InsertFormula(LU, LUIdx, F);
3764  next:;
3765  }
3766 }
3767 
3768 /// Generate stride factor reuse formulae by making use of scaled-offset address
3769 /// modes, for example.
3770 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3771  // Determine the integer type for the base formula.
3772  Type *IntTy = Base.getType();
3773  if (!IntTy) return;
3774 
3775  // If this Formula already has a scaled register, we can't add another one.
3776  // Try to unscale the formula to generate a better scale.
3777  if (Base.Scale != 0 && !Base.unscale())
3778  return;
3779 
3780  assert(Base.Scale == 0 && "unscale did not did its job!");
3781 
3782  // Check each interesting stride.
3783  for (int64_t Factor : Factors) {
3784  Base.Scale = Factor;
3785  Base.HasBaseReg = Base.BaseRegs.size() > 1;
3786  // Check whether this scale is going to be legal.
3787  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3788  Base)) {
3789  // As a special-case, handle special out-of-loop Basic users specially.
3790  // TODO: Reconsider this special case.
3791  if (LU.Kind == LSRUse::Basic &&
3792  isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3793  LU.AccessTy, Base) &&
3794  LU.AllFixupsOutsideLoop)
3795  LU.Kind = LSRUse::Special;
3796  else
3797  continue;
3798  }
3799  // For an ICmpZero, negating a solitary base register won't lead to
3800  // new solutions.
3801  if (LU.Kind == LSRUse::ICmpZero &&
3802  !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3803  continue;
3804  // For each addrec base reg, if its loop is current loop, apply the scale.
3805  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3806  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3807  if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3808  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3809  if (FactorS->isZero())
3810  continue;
3811  // Divide out the factor, ignoring high bits, since we'll be
3812  // scaling the value back up in the end.
3813  if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3814  // TODO: This could be optimized to avoid all the copying.
3815  Formula F = Base;
3816  F.ScaledReg = Quotient;
3817  F.deleteBaseReg(F.BaseRegs[i]);
3818  // The canonical representation of 1*reg is reg, which is already in
3819  // Base. In that case, do not try to insert the formula, it will be
3820  // rejected anyway.
3821  if (F.Scale == 1 && (F.BaseRegs.empty() ||
3822  (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3823  continue;
3824  // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
3825  // non canonical Formula with ScaledReg's loop not being L.
3826  if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
3827  F.canonicalize(*L);
3828  (void)InsertFormula(LU, LUIdx, F);
3829  }
3830  }
3831  }
3832  }
3833 }
3834 
3835 /// Generate reuse formulae from different IV types.
3836 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
3837  // Don't bother truncating symbolic values.
3838  if (Base.BaseGV) return;
3839 
3840  // Determine the integer type for the base formula.
3841  Type *DstTy = Base.getType();
3842  if (!DstTy) return;
3843  DstTy = SE.getEffectiveSCEVType(DstTy);
3844 
3845  for (Type *SrcTy : Types) {
3846  if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
3847  Formula F = Base;
3848 
3849  if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
3850  for (const SCEV *&BaseReg : F.BaseRegs)
3851  BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
3852 
3853  // TODO: This assumes we've done basic processing on all uses and
3854  // have an idea what the register usage is.
3855  if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
3856  continue;
3857 
3858  F.canonicalize(*L);
3859  (void)InsertFormula(LU, LUIdx, F);
3860  }
3861  }
3862 }
3863 
3864 namespace {
3865 
3866 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
3867 /// modifications so that the search phase doesn't have to worry about the data
3868 /// structures moving underneath it.
3869 struct WorkItem {
3870  size_t LUIdx;
3871  int64_t Imm;
3872  const SCEV *OrigReg;
3873 
3874  WorkItem(size_t LI, int64_t I, const SCEV *R)
3875  : LUIdx(LI), Imm(I), OrigReg(R) {}
3876 
3877  void print(raw_ostream &OS) const;
3878  void dump() const;
3879 };
3880 
3881 } // end anonymous namespace
3882 
3883 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
3884 void WorkItem::print(raw_ostream &OS) const {
3885  OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
3886  << " , add offset " << Imm;
3887 }
3888 
3889 LLVM_DUMP_METHOD void WorkItem::dump() const {
3890  print(errs()); errs() << '\n';
3891 }
3892 #endif
3893 
3894 /// Look for registers which are a constant distance apart and try to form reuse
3895 /// opportunities between them.
3896 void LSRInstance::GenerateCrossUseConstantOffsets() {
3897  // Group the registers by their value without any added constant offset.
3898  typedef std::map<int64_t, const SCEV *> ImmMapTy;
3900  DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
3902  for (const SCEV *Use : RegUses) {
3903  const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
3904  int64_t Imm = ExtractImmediate(Reg, SE);
3905  auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
3906  if (Pair.second)
3907  Sequence.push_back(Reg);
3908  Pair.first->second.insert(std::make_pair(Imm, Use));
3909  UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
3910  }
3911 
3912  // Now examine each set of registers with the same base value. Build up
3913  // a list of work to do and do the work in a separate step so that we're
3914  // not adding formulae and register counts while we're searching.
3915  SmallVector<WorkItem, 32> WorkItems;
3916  SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
3917  for (const SCEV *Reg : Sequence) {
3918  const ImmMapTy &Imms = Map.find(Reg)->second;
3919 
3920  // It's not worthwhile looking for reuse if there's only one offset.
3921  if (Imms.size() == 1)
3922  continue;
3923 
3924  DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
3925  for (const auto &Entry : Imms)
3926  dbgs() << ' ' << Entry.first;
3927  dbgs() << '\n');
3928 
3929  // Examine each offset.
3930  for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
3931  J != JE; ++J) {
3932  const SCEV *OrigReg = J->second;
3933 
3934  int64_t JImm = J->first;
3935  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
3936 
3937  if (!isa<SCEVConstant>(OrigReg) &&
3938  UsedByIndicesMap[Reg].count() == 1) {
3939  DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
3940  continue;
3941  }
3942 
3943  // Conservatively examine offsets between this orig reg a few selected
3944  // other orig regs.
3945  ImmMapTy::const_iterator OtherImms[] = {
3946  Imms.begin(), std::prev(Imms.end()),
3947  Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
3948  2)
3949  };
3950  for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
3951  ImmMapTy::const_iterator M = OtherImms[i];
3952  if (M == J || M == JE) continue;
3953 
3954  // Compute the difference between the two.
3955  int64_t Imm = (uint64_t)JImm - M->first;
3956  for (unsigned LUIdx : UsedByIndices.set_bits())
3957  // Make a memo of this use, offset, and register tuple.
3958  if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
3959  WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
3960  }
3961  }
3962  }
3963 
3964  Map.clear();
3965  Sequence.clear();
3966  UsedByIndicesMap.clear();
3967  UniqueItems.clear();
3968 
3969  // Now iterate through the worklist and add new formulae.
3970  for (const WorkItem &WI : WorkItems) {
3971  size_t LUIdx = WI.LUIdx;
3972  LSRUse &LU = Uses[LUIdx];
3973  int64_t Imm = WI.Imm;
3974  const SCEV *OrigReg = WI.OrigReg;
3975 
3976  Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
3977  const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
3978  unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
3979 
3980  // TODO: Use a more targeted data structure.
3981  for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
3982  Formula F = LU.Formulae[L];
3983  // FIXME: The code for the scaled and unscaled registers looks
3984  // very similar but slightly different. Investigate if they
3985  // could be merged. That way, we would not have to unscale the
3986  // Formula.
3987  F.unscale();
3988  // Use the immediate in the scaled register.
3989  if (F.ScaledReg == OrigReg) {
3990  int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
3991  // Don't create 50 + reg(-50).
3992  if (F.referencesReg(SE.getSCEV(
3993  ConstantInt::get(IntTy, -(uint64_t)Offset))))
3994  continue;
3995  Formula NewF = F;
3996  NewF.BaseOffset = Offset;
3997  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3998  NewF))
3999  continue;
4000  NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4001 
4002  // If the new scale is a constant in a register, and adding the constant
4003  // value to the immediate would produce a value closer to zero than the
4004  // immediate itself, then the formula isn't worthwhile.
4005  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4006  if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4007  (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4008  .ule(std::abs(NewF.BaseOffset)))
4009  continue;
4010 
4011  // OK, looks good.
4012  NewF.canonicalize(*this->L);
4013  (void)InsertFormula(LU, LUIdx, NewF);
4014  } else {
4015  // Use the immediate in a base register.
4016  for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4017  const SCEV *BaseReg = F.BaseRegs[N];
4018  if (BaseReg != OrigReg)
4019  continue;
4020  Formula NewF = F;
4021  NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4022  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4023  LU.Kind, LU.AccessTy, NewF)) {
4024  if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4025  continue;
4026  NewF = F;
4027  NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4028  }
4029  NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4030 
4031  // If the new formula has a constant in a register, and adding the
4032  // constant value to the immediate would produce a value closer to
4033  // zero than the immediate itself, then the formula isn't worthwhile.
4034  for (const SCEV *NewReg : NewF.BaseRegs)
4035  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4036  if ((C->getAPInt() + NewF.BaseOffset)
4037  .abs()
4038  .slt(std::abs(NewF.BaseOffset)) &&
4039  (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4040  countTrailingZeros<uint64_t>(NewF.BaseOffset))
4041  goto skip_formula;
4042 
4043  // Ok, looks good.
4044  NewF.canonicalize(*this->L);
4045  (void)InsertFormula(LU, LUIdx, NewF);
4046  break;
4047  skip_formula:;
4048  }
4049  }
4050  }
4051  }
4052 }
4053 
4054 /// Generate formulae for each use.
4055 void
4056 LSRInstance::GenerateAllReuseFormulae() {
4057  // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4058  // queries are more precise.
4059  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4060  LSRUse &LU = Uses[LUIdx];
4061  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4062  GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4063  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4064  GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4065  }
4066  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4067  LSRUse &LU = Uses[LUIdx];
4068  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4069  GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4070  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4071  GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4072  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4073  GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4074  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4075  GenerateScales(LU, LUIdx, LU.Formulae[i]);
4076  }
4077  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4078  LSRUse &LU = Uses[LUIdx];
4079  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4080  GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4081  }
4082 
4083  GenerateCrossUseConstantOffsets();
4084 
4085  DEBUG(dbgs() << "\n"
4086  "After generating reuse formulae:\n";
4087  print_uses(dbgs()));
4088 }
4089 
4090 /// If there are multiple formulae with the same set of registers used
4091 /// by other uses, pick the best one and delete the others.
4092 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4093  DenseSet<const SCEV *> VisitedRegs;
4096 #ifndef NDEBUG
4097  bool ChangedFormulae = false;
4098 #endif
4099 
4100  // Collect the best formula for each unique set of shared registers. This
4101  // is reset for each use.
4102  typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
4103  BestFormulaeTy;
4104  BestFormulaeTy BestFormulae;
4105 
4106  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4107  LSRUse &LU = Uses[LUIdx];
4108  DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4109 
4110  bool Any = false;
4111  for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4112  FIdx != NumForms; ++FIdx) {
4113  Formula &F = LU.Formulae[FIdx];
4114 
4115  // Some formulas are instant losers. For example, they may depend on
4116  // nonexistent AddRecs from other loops. These need to be filtered
4117  // immediately, otherwise heuristics could choose them over others leading
4118  // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4119  // avoids the need to recompute this information across formulae using the
4120  // same bad AddRec. Passing LoserRegs is also essential unless we remove
4121  // the corresponding bad register from the Regs set.
4122  Cost CostF;
4123  Regs.clear();
4124  CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, SE, DT, LU, &LoserRegs);
4125  if (CostF.isLoser()) {
4126  // During initial formula generation, undesirable formulae are generated
4127  // by uses within other loops that have some non-trivial address mode or
4128  // use the postinc form of the IV. LSR needs to provide these formulae
4129  // as the basis of rediscovering the desired formula that uses an AddRec
4130  // corresponding to the existing phi. Once all formulae have been
4131  // generated, these initial losers may be pruned.
4132  DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
4133  dbgs() << "\n");
4134  }
4135  else {
4137  for (const SCEV *Reg : F.BaseRegs) {
4138  if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4139  Key.push_back(Reg);
4140  }
4141  if (F.ScaledReg &&
4142  RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4143  Key.push_back(F.ScaledReg);
4144  // Unstable sort by host order ok, because this is only used for
4145  // uniquifying.
4146  std::sort(Key.begin(), Key.end());
4147 
4148  std::pair<BestFormulaeTy::const_iterator, bool> P =
4149  BestFormulae.insert(std::make_pair(Key, FIdx));
4150  if (P.second)
4151  continue;
4152 
4153  Formula &Best = LU.Formulae[P.first->second];
4154 
4155  Cost CostBest;
4156  Regs.clear();
4157  CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, SE, DT, LU);
4158  if (CostF.isLess(CostBest, TTI))
4159  std::swap(F, Best);
4160  DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4161  dbgs() << "\n"
4162  " in favor of formula "; Best.print(dbgs());
4163  dbgs() << '\n');
4164  }
4165 #ifndef NDEBUG
4166  ChangedFormulae = true;
4167 #endif
4168  LU.DeleteFormula(F);
4169  --FIdx;
4170  --NumForms;
4171  Any = true;
4172  }
4173 
4174  // Now that we've filtered out some formulae, recompute the Regs set.
4175  if (Any)
4176  LU.RecomputeRegs(LUIdx, RegUses);
4177 
4178  // Reset this to prepare for the next use.
4179  BestFormulae.clear();
4180  }
4181 
4182  DEBUG(if (ChangedFormulae) {
4183  dbgs() << "\n"
4184  "After filtering out undesirable candidates:\n";
4185  print_uses(dbgs());
4186  });
4187 }
4188 
4189 // This is a rough guess that seems to work fairly well.
4190 static const size_t ComplexityLimit = UINT16_MAX;
4191 
4192 /// Estimate the worst-case number of solutions the solver might have to
4193 /// consider. It almost never considers this many solutions because it prune the
4194 /// search space, but the pruning isn't always sufficient.
4195 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4196  size_t Power = 1;
4197  for (const LSRUse &LU : Uses) {
4198  size_t FSize = LU.Formulae.size();
4199  if (FSize >= ComplexityLimit) {
4200  Power = ComplexityLimit;
4201  break;
4202  }
4203  Power *= FSize;
4204  if (Power >= ComplexityLimit)
4205  break;
4206  }
4207  return Power;
4208 }
4209 
4210 /// When one formula uses a superset of the registers of another formula, it
4211 /// won't help reduce register pressure (though it may not necessarily hurt
4212 /// register pressure); remove it to simplify the system.
4213 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4214  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4215  DEBUG(dbgs() << "The search space is too complex.\n");
4216 
4217  DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4218  "which use a superset of registers used by other "
4219  "formulae.\n");
4220 
4221  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4222  LSRUse &LU = Uses[LUIdx];
4223  bool Any = false;
4224  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4225  Formula &F = LU.Formulae[i];
4226  // Look for a formula with a constant or GV in a register. If the use
4227  // also has a formula with that same value in an immediate field,
4228  // delete the one that uses a register.
4230  I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4231  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4232  Formula NewF = F;
4233  NewF.BaseOffset += C->getValue()->getSExtValue();
4234  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4235  (I - F.BaseRegs.begin()));
4236  if (LU.HasFormulaWithSameRegs(NewF)) {
4237  DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4238  LU.DeleteFormula(F);
4239  --i;
4240  --e;
4241  Any = true;
4242  break;
4243  }
4244  } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4245  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4246  if (!F.BaseGV) {
4247  Formula NewF = F;
4248  NewF.BaseGV = GV;
4249  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4250  (I - F.BaseRegs.begin()));
4251  if (LU.HasFormulaWithSameRegs(NewF)) {
4252  DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4253  dbgs() << '\n');
4254  LU.DeleteFormula(F);
4255  --i;
4256  --e;
4257  Any = true;
4258  break;
4259  }
4260  }
4261  }
4262  }
4263  }
4264  if (Any)
4265  LU.RecomputeRegs(LUIdx, RegUses);
4266  }
4267 
4268  DEBUG(dbgs() << "After pre-selection:\n";
4269  print_uses(dbgs()));
4270  }
4271 }
4272 
4273 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4274 /// allocate a single register for them.
4275 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4276  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4277  return;
4278 
4279  DEBUG(dbgs() << "The search space is too complex.\n"
4280  "Narrowing the search space by assuming that uses separated "
4281  "by a constant offset will use the same registers.\n");
4282 
4283  // This is especially useful for unrolled loops.
4284 
4285  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4286  LSRUse &LU = Uses[LUIdx];
4287  for (const Formula &F : LU.Formulae) {
4288  if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4289  continue;
4290 
4291  LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4292  if (!LUThatHas)
4293  continue;
4294 
4295  if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4296  LU.Kind, LU.AccessTy))
4297  continue;
4298 
4299  DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4300 
4301  LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4302 
4303  // Transfer the fixups of LU to LUThatHas.
4304  for (LSRFixup &Fixup : LU.Fixups) {
4305  Fixup.Offset += F.BaseOffset;
4306  LUThatHas->pushFixup(Fixup);
4307  DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4308  }
4309 
4310  // Delete formulae from the new use which are no longer legal.
4311  bool Any = false;
4312  for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4313  Formula &F = LUThatHas->Formulae[i];
4314  if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4315  LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4316  DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4317  dbgs() << '\n');
4318  LUThatHas->DeleteFormula(F);
4319  --i;
4320  --e;
4321  Any = true;
4322  }
4323  }
4324 
4325  if (Any)
4326  LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4327 
4328  // Delete the old use.
4329  DeleteUse(LU, LUIdx);
4330  --LUIdx;
4331  --NumUses;
4332  break;
4333  }
4334  }
4335 
4336  DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4337 }
4338 
4339 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4340 /// we've done more filtering, as it may be able to find more formulae to
4341 /// eliminate.
4342 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4343  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4344  DEBUG(dbgs() << "The search space is too complex.\n");
4345 
4346  DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4347  "undesirable dedicated registers.\n");
4348 
4349  FilterOutUndesirableDedicatedRegisters();
4350 
4351  DEBUG(dbgs() << "After pre-selection:\n";
4352  print_uses(dbgs()));
4353  }
4354 }
4355 
4356 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4357 /// Pick the best one and delete the others.
4358 /// This narrowing heuristic is to keep as many formulae with different
4359 /// Scale and ScaledReg pair as possible while narrowing the search space.
4360 /// The benefit is that it is more likely to find out a better solution
4361 /// from a formulae set with more Scale and ScaledReg variations than
4362 /// a formulae set with the same Scale and ScaledReg. The picking winner
4363 /// reg heurstic will often keep the formulae with the same Scale and
4364 /// ScaledReg and filter others, and we want to avoid that if possible.
4365 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4366  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4367  return;
4368 
4369  DEBUG(dbgs() << "The search space is too complex.\n"
4370  "Narrowing the search space by choosing the best Formula "
4371  "from the Formulae with the same Scale and ScaledReg.\n");
4372 
4373  // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4374  typedef DenseMap<std::pair<const SCEV *, int64_t>, size_t> BestFormulaeTy;
4375  BestFormulaeTy BestFormulae;
4376 #ifndef NDEBUG
4377  bool ChangedFormulae = false;
4378 #endif
4379  DenseSet<const SCEV *> VisitedRegs;
4381 
4382  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4383  LSRUse &LU = Uses[LUIdx];
4384  DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
4385 
4386  // Return true if Formula FA is better than Formula FB.
4387  auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4388  // First we will try to choose the Formula with fewer new registers.
4389  // For a register used by current Formula, the more the register is
4390  // shared among LSRUses, the less we increase the register number
4391  // counter of the formula.
4392  size_t FARegNum = 0;
4393  for (const SCEV *Reg : FA.BaseRegs) {
4394  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4395  FARegNum += (NumUses - UsedByIndices.count() + 1);
4396  }
4397  size_t FBRegNum = 0;
4398  for (const SCEV *Reg : FB.BaseRegs) {
4399  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4400  FBRegNum += (NumUses - UsedByIndices.count() + 1);
4401  }
4402  if (FARegNum != FBRegNum)
4403  return FARegNum < FBRegNum;
4404 
4405  // If the new register numbers are the same, choose the Formula with
4406  // less Cost.
4407  Cost CostFA, CostFB;
4408  Regs.clear();
4409  CostFA.RateFormula(TTI, FA, Regs, VisitedRegs, L, SE, DT, LU);
4410  Regs.clear();
4411  CostFB.RateFormula(TTI, FB, Regs, VisitedRegs, L, SE, DT, LU);
4412  return CostFA.isLess(CostFB, TTI);
4413  };
4414 
4415  bool Any = false;
4416  for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4417  ++FIdx) {
4418  Formula &F = LU.Formulae[FIdx];
4419  if (!F.ScaledReg)
4420  continue;
4421  auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4422  if (P.second)
4423  continue;
4424 
4425  Formula &Best = LU.Formulae[P.first->second];
4426  if (IsBetterThan(F, Best))
4427  std::swap(F, Best);
4428  DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4429  dbgs() << "\n"
4430  " in favor of formula ";
4431  Best.print(dbgs()); dbgs() << '\n');
4432 #ifndef NDEBUG
4433  ChangedFormulae = true;
4434 #endif
4435  LU.DeleteFormula(F);
4436  --FIdx;
4437  --NumForms;
4438  Any = true;
4439  }
4440  if (Any)
4441  LU.RecomputeRegs(LUIdx, RegUses);
4442 
4443  // Reset this to prepare for the next use.
4444  BestFormulae.clear();
4445  }
4446 
4447  DEBUG(if (ChangedFormulae) {
4448  dbgs() << "\n"
4449  "After filtering out undesirable candidates:\n";
4450  print_uses(dbgs());
4451  });
4452 }
4453 
4454 /// The function delete formulas with high registers number expectation.
4455 /// Assuming we don't know the value of each formula (already delete
4456 /// all inefficient), generate probability of not selecting for each
4457 /// register.
4458 /// For example,
4459 /// Use1:
4460 /// reg(a) + reg({0,+,1})
4461 /// reg(a) + reg({-1,+,1}) + 1
4462 /// reg({a,+,1})
4463 /// Use2:
4464 /// reg(b) + reg({0,+,1})
4465 /// reg(b) + reg({-1,+,1}) + 1
4466 /// reg({b,+,1})
4467 /// Use3:
4468 /// reg(c) + reg(b) + reg({0,+,1})
4469 /// reg(c) + reg({b,+,1})
4470 ///
4471 /// Probability of not selecting
4472 /// Use1 Use2 Use3
4473 /// reg(a) (1/3) * 1 * 1
4474 /// reg(b) 1 * (1/3) * (1/2)
4475 /// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4476 /// reg({-1,+,1}) (2/3) * (2/3) * 1
4477 /// reg({a,+,1}) (2/3) * 1 * 1
4478 /// reg({b,+,1}) 1 * (2/3) * (2/3)
4479 /// reg(c) 1 * 1 * 0
4480 ///
4481 /// Now count registers number mathematical expectation for each formula:
4482 /// Note that for each use we exclude probability if not selecting for the use.
4483 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4484 /// probabilty 1/3 of not selecting for Use1).
4485 /// Use1:
4486 /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4487 /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4488 /// reg({a,+,1}) 1
4489 /// Use2:
4490 /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4491 /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4492 /// reg({b,+,1}) 2/3
4493 /// Use3:
4494 /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4495 /// reg(c) + reg({b,+,1}) 1 + 2/3
4496 
4497 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4498  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4499  return;
4500  // Ok, we have too many of formulae on our hands to conveniently handle.
4501  // Use a rough heuristic to thin out the list.
4502 
4503  // Set of Regs wich will be 100% used in final solution.
4504  // Used in each formula of a solution (in example above this is reg(c)).
4505  // We can skip them in calculations.
4507  DEBUG(dbgs() << "The search space is too complex.\n");
4508 
4509  // Map each register to probability of not selecting
4511  for (const SCEV *Reg : RegUses) {
4512  if (UniqRegs.count(Reg))
4513  continue;
4514  float PNotSel = 1;
4515  for (const LSRUse &LU : Uses) {
4516  if (!LU.Regs.count(Reg))
4517  continue;
4518  float P = LU.getNotSelectedProbability(Reg);
4519  if (P != 0.0)
4520  PNotSel *= P;
4521  else
4522  UniqRegs.insert(Reg);
4523  }
4524  RegNumMap.insert(std::make_pair(Reg, PNotSel));
4525  }
4526 
4527  DEBUG(dbgs() << "Narrowing the search space by deleting costly formulas\n");
4528 
4529  // Delete formulas where registers number expectation is high.
4530  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4531  LSRUse &LU = Uses[LUIdx];
4532  // If nothing to delete - continue.
4533  if (LU.Formulae.size() < 2)
4534  continue;
4535  // This is temporary solution to test performance. Float should be
4536  // replaced with round independent type (based on integers) to avoid
4537  // different results for different target builds.
4538  float FMinRegNum = LU.Formulae[0].getNumRegs();
4539  float FMinARegNum = LU.Formulae[0].getNumRegs();
4540  size_t MinIdx = 0;
4541  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4542  Formula &F = LU.Formulae[i];
4543  float FRegNum = 0;
4544  float FARegNum = 0;
4545  for (const SCEV *BaseReg : F.BaseRegs) {
4546  if (UniqRegs.count(BaseReg))
4547  continue;
4548  FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4549  if (isa<SCEVAddRecExpr>(BaseReg))
4550  FARegNum +=
4551  RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4552  }
4553  if (const SCEV *ScaledReg = F.ScaledReg) {
4554  if (!UniqRegs.count(ScaledReg)) {
4555  FRegNum +=
4556  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4557  if (isa<SCEVAddRecExpr>(ScaledReg))
4558  FARegNum +=
4559  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4560  }
4561  }
4562  if (FMinRegNum > FRegNum ||
4563  (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4564  FMinRegNum = FRegNum;
4565  FMinARegNum = FARegNum;
4566  MinIdx = i;
4567  }
4568  }
4569  DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());
4570  dbgs() << " with min reg num " << FMinRegNum << '\n');
4571  if (MinIdx != 0)
4572  std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4573  while (LU.Formulae.size() != 1) {
4574  DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());
4575  dbgs() << '\n');
4576  LU.Formulae.pop_back();
4577  }
4578  LU.RecomputeRegs(LUIdx, RegUses);
4579  assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
4580  Formula &F = LU.Formulae[0];
4581  DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n');
4582  // When we choose the formula, the regs become unique.
4583  UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4584  if (F.ScaledReg)
4585  UniqRegs.insert(F.ScaledReg);
4586  }
4587  DEBUG(dbgs() << "After pre-selection:\n";
4588  print_uses(dbgs()));
4589 }
4590 
4591 
4592 /// Pick a register which seems likely to be profitable, and then in any use
4593 /// which has any reference to that register, delete all formulae which do not
4594 /// reference that register.
4595 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4596  // With all other options exhausted, loop until the system is simple
4597  // enough to handle.
4599  while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4600  // Ok, we have too many of formulae on our hands to conveniently handle.
4601  // Use a rough heuristic to thin out the list.
4602  DEBUG(dbgs() << "The search space is too complex.\n");
4603 
4604  // Pick the register which is used by the most LSRUses, which is likely
4605  // to be a good reuse register candidate.
4606  const SCEV *Best = nullptr;
4607  unsigned BestNum = 0;
4608  for (const SCEV *Reg : RegUses) {
4609  if (Taken.count(Reg))
4610  continue;
4611  if (!Best) {
4612  Best = Reg;
4613  BestNum = RegUses.getUsedByIndices(Reg).count();
4614  } else {
4615  unsigned Count = RegUses.getUsedByIndices(Reg).count();
4616  if (Count > BestNum) {
4617  Best = Reg;
4618  BestNum = Count;
4619  }
4620  }
4621  }
4622 
4623  DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4624  << " will yield profitable reuse.\n");
4625  Taken.insert(Best);
4626 
4627  // In any use with formulae which references this register, delete formulae
4628  // which don't reference it.
4629  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4630  LSRUse &LU = Uses[LUIdx];
4631  if (!LU.Regs.count(Best)) continue;
4632 
4633  bool Any = false;
4634  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4635  Formula &F = LU.Formulae[i];
4636  if (!F.referencesReg(Best)) {
4637  DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4638  LU.DeleteFormula(F);
4639  --e;
4640  --i;
4641  Any = true;
4642  assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4643  continue;
4644  }
4645  }
4646 
4647  if (Any)
4648  LU.RecomputeRegs(LUIdx, RegUses);
4649  }
4650 
4651  DEBUG(dbgs() << "After pre-selection:\n";
4652  print_uses(dbgs()));
4653  }
4654 }
4655 
4656 /// If there are an extraordinary number of formulae to choose from, use some
4657 /// rough heuristics to prune down the number of formulae. This keeps the main
4658 /// solver from taking an extraordinary amount of time in some worst-case
4659 /// scenarios.
4660 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4661  NarrowSearchSpaceByDetectingSupersets();
4662  NarrowSearchSpaceByCollapsingUnrolledCode();
4663  NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4664  if (FilterSameScaledReg)
4665  NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4666  if (LSRExpNarrow)
4667  NarrowSearchSpaceByDeletingCostlyFormulas();
4668  else
4669  NarrowSearchSpaceByPickingWinnerRegs();
4670 }
4671 
4672 /// This is the recursive solver.
4673 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4674  Cost &SolutionCost,
4676  const Cost &CurCost,
4677  const SmallPtrSet<const SCEV *, 16> &CurRegs,
4678  DenseSet<const SCEV *> &VisitedRegs) const {
4679  // Some ideas:
4680  // - prune more:
4681  // - use more aggressive filtering
4682  // - sort the formula so that the most profitable solutions are found first
4683  // - sort the uses too
4684  // - search faster:
4685  // - don't compute a cost, and then compare. compare while computing a cost
4686  // and bail early.
4687  // - track register sets with SmallBitVector
4688 
4689  const LSRUse &LU = Uses[Workspace.size()];
4690 
4691  // If this use references any register that's already a part of the
4692  // in-progress solution, consider it a requirement that a formula must
4693  // reference that register in order to be considered. This prunes out
4694  // unprofitable searching.
4696  for (const SCEV *S : CurRegs)
4697  if (LU.Regs.count(S))
4698  ReqRegs.insert(S);
4699 
4701  Cost NewCost;
4702  for (const Formula &F : LU.Formulae) {
4703  // Ignore formulae which may not be ideal in terms of register reuse of
4704  // ReqRegs. The formula should use all required registers before
4705  // introducing new ones.
4706  int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4707  for (const SCEV *Reg : ReqRegs) {
4708  if ((F.ScaledReg && F.ScaledReg == Reg) ||
4709  is_contained(F.BaseRegs, Reg)) {
4710  --NumReqRegsToFind;
4711  if (NumReqRegsToFind == 0)
4712  break;
4713  }
4714  }
4715  if (NumReqRegsToFind != 0) {
4716  // If none of the formulae satisfied the required registers, then we could
4717  // clear ReqRegs and try again. Currently, we simply give up in this case.
4718  continue;
4719  }
4720 
4721  // Evaluate the cost of the current formula. If it's already worse than
4722  // the current best, prune the search at that point.
4723  NewCost = CurCost;
4724  NewRegs = CurRegs;
4725  NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, SE, DT, LU);
4726  if (NewCost.isLess(SolutionCost, TTI)) {
4727  Workspace.push_back(&F);
4728  if (Workspace.size() != Uses.size()) {
4729  SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4730  NewRegs, VisitedRegs);
4731  if (F.getNumRegs() == 1 && Workspace.size() == 1)
4732  VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4733  } else {
4734  DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4735  dbgs() << ".\n Regs:";
4736  for (const SCEV *S : NewRegs)
4737  dbgs() << ' ' << *S;
4738  dbgs() << '\n');
4739 
4740  SolutionCost = NewCost;
4741  Solution = Workspace;
4742  }
4743  Workspace.pop_back();
4744  }
4745  }
4746 }
4747 
4748 /// Choose one formula from each use. Return the results in the given Solution
4749 /// vector.
4750 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4752  Cost SolutionCost;
4753  SolutionCost.Lose();
4754  Cost CurCost;
4756  DenseSet<const SCEV *> VisitedRegs;
4757  Workspace.reserve(Uses.size());
4758 
4759  // SolveRecurse does all the work.
4760  SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4761  CurRegs, VisitedRegs);
4762  if (Solution.empty()) {
4763  DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4764  return;
4765  }
4766 
4767  // Ok, we've now made all our decisions.
4768  DEBUG(dbgs() << "\n"
4769  "The chosen solution requires "; SolutionCost.print(dbgs());
4770  dbgs() << ":\n";
4771  for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4772  dbgs() << " ";
4773  Uses[i].print(dbgs());
4774  dbgs() << "\n"
4775  " ";
4776  Solution[i]->print(dbgs());
4777  dbgs() << '\n';
4778  });
4779 
4780  assert(Solution.size() == Uses.size() && "Malformed solution!");
4781 }
4782 
4783 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4784 /// we can go while still being dominated by the input positions. This helps
4785 /// canonicalize the insert position, which encourages sharing.
4787 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4789  const {
4790  Instruction *Tentative = &*IP;
4791  while (true) {
4792  bool AllDominate = true;
4793  Instruction *BetterPos = nullptr;
4794  // Don't bother attempting to insert before a catchswitch, their basic block
4795  // cannot have other non-PHI instructions.
4796  if (isa<CatchSwitchInst>(Tentative))
4797  return IP;
4798 
4799  for (Instruction *Inst : Inputs) {
4800  if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
4801  AllDominate = false;
4802  break;
4803  }
4804  // Attempt to find an insert position in the middle of the block,
4805  // instead of at the end, so that it can be used for other expansions.
4806  if (Tentative->getParent() == Inst->getParent() &&
4807  (!BetterPos || !DT.dominates(Inst, BetterPos)))
4808  BetterPos = &*std::next(BasicBlock::iterator(Inst));
4809  }
4810  if (!AllDominate)
4811  break;
4812  if (BetterPos)
4813  IP = BetterPos->getIterator();
4814  else
4815  IP = Tentative->getIterator();
4816 
4817  const Loop *IPLoop = LI.getLoopFor(IP->getParent());
4818  unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
4819 
4820  BasicBlock *IDom;
4821  for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
4822  if (!Rung) return IP;
4823  Rung = Rung->getIDom();
4824  if (!Rung) return IP;
4825  IDom = Rung->getBlock();
4826 
4827  // Don't climb into a loop though.
4828  const Loop *IDomLoop = LI.getLoopFor(IDom);
4829  unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
4830  if (IDomDepth <= IPLoopDepth &&
4831  (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
4832  break;
4833  }
4834 
4835  Tentative = IDom->getTerminator();
4836  }
4837 
4838  return IP;
4839 }
4840 
4841 /// Determine an input position which will be dominated by the operands and
4842 /// which will dominate the result.
4844 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
4845  const LSRFixup &LF,
4846  const LSRUse &LU,
4847  SCEVExpander &Rewriter) const {
4848  // Collect some instructions which must be dominated by the
4849  // expanding replacement. These must be dominated by any operands that
4850  // will be required in the expansion.
4852  if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
4853  Inputs.push_back(I);
4854  if (LU.Kind == LSRUse::ICmpZero)
4855  if (Instruction *I =
4856  dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
4857  Inputs.push_back(I);
4858  if (LF.PostIncLoops.count(L)) {
4859  if (LF.isUseFullyOutsideLoop(L))
4860  Inputs.push_back(L->getLoopLatch()->getTerminator());
4861  else
4862  Inputs.push_back(IVIncInsertPos);
4863  }
4864  // The expansion must also be dominated by the increment positions of any
4865  // loops it for which it is using post-inc mode.
4866  for (const Loop *PIL : LF.PostIncLoops) {
4867  if (PIL == L) continue;
4868 
4869  // Be dominated by the loop exit.
4870  SmallVector<BasicBlock *, 4> ExitingBlocks;
4871  PIL->getExitingBlocks(ExitingBlocks);
4872  if (!ExitingBlocks.empty()) {
4873  BasicBlock *BB = ExitingBlocks[0];
4874  for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
4875  BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
4876  Inputs.push_back(BB->getTerminator());
4877  }
4878  }
4879 
4880  assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
4881  && !isa<DbgInfoIntrinsic>(LowestIP) &&
4882  "Insertion point must be a normal instruction");
4883 
4884  // Then, climb up the immediate dominator tree as far as we can go while
4885  // still being dominated by the input positions.
4886  BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
4887 
4888  // Don't insert instructions before PHI nodes.
4889  while (isa<PHINode>(IP)) ++IP;
4890 
4891  // Ignore landingpad instructions.
4892  while (IP->isEHPad()) ++IP;
4893 
4894  // Ignore debug intrinsics.
4895  while (isa<DbgInfoIntrinsic>(IP)) ++IP;
4896 
4897  // Set IP below instructions recently inserted by SCEVExpander. This keeps the
4898  // IP consistent across expansions and allows the previously inserted
4899  // instructions to be reused by subsequent expansion.
4900  while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
4901  ++IP;
4902 
4903  return IP;
4904 }
4905 
4906 /// Emit instructions for the leading candidate expression for this LSRUse (this
4907 /// is called "expanding").
4908 Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
4909  const Formula &F, BasicBlock::iterator IP,
4910  SCEVExpander &Rewriter,
4911  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
4912  if (LU.RigidFormula)
4913  return LF.OperandValToReplace;
4914 
4915  // Determine an input position which will be dominated by the operands and
4916  // which will dominate the result.
4917  IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
4918  Rewriter.setInsertPoint(&*IP);
4919 
4920  // Inform the Rewriter if we have a post-increment use, so that it can
4921  // perform an advantageous expansion.
4922  Rewriter.setPostInc(LF.PostIncLoops);
4923 
4924  // This is the type that the user actually needs.
4925  Type *OpTy = LF.OperandValToReplace->getType();
4926  // This will be the type that we'll initially expand to.
4927  Type *Ty = F.getType();
4928  if (!Ty)
4929  // No type known; just expand directly to the ultimate type.
4930  Ty = OpTy;
4931  else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
4932  // Expand directly to the ultimate type if it's the right size.
4933  Ty = OpTy;
4934  // This is the type to do integer arithmetic in.
4935  Type *IntTy = SE.getEffectiveSCEVType(Ty);
4936 
4937  // Build up a list of operands to add together to form the full base.
4939 
4940  // Expand the BaseRegs portion.
4941  for (const SCEV *Reg : F.BaseRegs) {
4942  assert(!Reg->isZero() && "Zero allocated in a base register!");
4943 
4944  // If we're expanding for a post-inc user, make the post-inc adjustment.
4945  Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
4946  Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
4947  }
4948 
4949  // Expand the ScaledReg portion.
4950  Value *ICmpScaledV = nullptr;
4951  if (F.Scale != 0) {
4952  const SCEV *ScaledS = F.ScaledReg;
4953 
4954  // If we're expanding for a post-inc user, make the post-inc adjustment.
4955  PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
4956  ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
4957 
4958  if (LU.Kind == LSRUse::ICmpZero) {
4959  // Expand ScaleReg as if it was part of the base regs.
4960  if (F.Scale == 1)
4961  Ops.push_back(
4962  SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
4963  else {
4964  // An interesting way of "folding" with an icmp is to use a negated
4965  // scale, which we'll implement by inserting it into the other operand
4966  // of the icmp.
4967  assert(F.Scale == -1 &&
4968  "The only scale supported by ICmpZero uses is -1!");
4969  ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
4970  }
4971  } else {
4972  // Otherwise just expand the scaled register and an explicit scale,
4973  // which is expected to be matched as part of the address.
4974 
4975  // Flush the operand list to suppress SCEVExpander hoisting address modes.
4976  // Unless the addressing mode will not be folded.
4977  if (!Ops.empty() && LU.Kind == LSRUse::Address &&
4978  isAMCompletelyFolded(TTI, LU, F)) {
4979  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4980  Ops.clear();
4981  Ops.push_back(SE.getUnknown(FullV));
4982  }
4983  ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
4984  if (F.Scale != 1)
4985  ScaledS =
4986  SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
4987  Ops.push_back(ScaledS);
4988  }
4989  }
4990 
4991  // Expand the GV portion.
4992  if (F.BaseGV) {
4993  // Flush the operand list to suppress SCEVExpander hoisting.
4994  if (!Ops.empty()) {
4995  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
4996  Ops.clear();
4997  Ops.push_back(SE.getUnknown(FullV));
4998  }
4999  Ops.push_back(SE.getUnknown(F.BaseGV));
5000  }
5001 
5002  // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5003  // unfolded offsets. LSR assumes they both live next to their uses.
5004  if (!Ops.empty()) {
5005  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5006  Ops.clear();
5007  Ops.push_back(SE.getUnknown(FullV));
5008  }
5009 
5010  // Expand the immediate portion.
5011  int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5012  if (Offset != 0) {
5013  if (LU.Kind == LSRUse::ICmpZero) {
5014  // The other interesting way of "folding" with an ICmpZero is to use a
5015  // negated immediate.
5016  if (!ICmpScaledV)
5017  ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5018  else {
5019  Ops.push_back(SE.getUnknown(ICmpScaledV));
5020  ICmpScaledV = ConstantInt::get(IntTy, Offset);
5021  }
5022  } else {
5023  // Just add the immediate values. These again are expected to be matched
5024  // as part of the address.
5025  Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5026  }
5027  }
5028 
5029  // Expand the unfolded offset portion.
5030  int64_t UnfoldedOffset = F.UnfoldedOffset;
5031  if (UnfoldedOffset != 0) {
5032  // Just add the immediate values.
5034  UnfoldedOffset)));
5035  }
5036 
5037  // Emit instructions summing all the operands.
5038  const SCEV *FullS = Ops.empty() ?
5039  SE.getConstant(IntTy, 0) :
5040  SE.getAddExpr(Ops);
5041  Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5042 
5043  // We're done expanding now, so reset the rewriter.
5044  Rewriter.clearPostInc();
5045 
5046  // An ICmpZero Formula represents an ICmp which we're handling as a
5047  // comparison against zero. Now that we've expanded an expression for that
5048  // form, update the ICmp's other operand.
5049  if (LU.Kind == LSRUse::ICmpZero) {
5050  ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5051  DeadInsts.emplace_back(CI->getOperand(1));
5052  assert(!F.BaseGV && "ICmp does not support folding a global value and "
5053  "a scale at the same time!");
5054  if (F.Scale == -1) {
5055  if (ICmpScaledV->getType() != OpTy) {
5056  Instruction *Cast =
5057  CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5058  OpTy, false),
5059  ICmpScaledV, OpTy, "tmp", CI);
5060  ICmpScaledV = Cast;
5061  }
5062  CI->setOperand(1, ICmpScaledV);
5063  } else {
5064  // A scale of 1 means that the scale has been expanded as part of the
5065  // base regs.
5066  assert((F.Scale == 0 || F.Scale == 1) &&
5067  "ICmp does not support folding a global value and "
5068  "a scale at the same time!");
5070  -(uint64_t)Offset);
5071  if (C->getType() != OpTy)
5073  OpTy, false),
5074  C, OpTy);
5075 
5076  CI->setOperand(1, C);
5077  }
5078  }
5079 
5080  return FullV;
5081 }
5082 
5083 /// Helper for Rewrite. PHI nodes are special because the use of their operands
5084 /// effectively happens in their predecessor blocks, so the expression may need
5085 /// to be expanded in multiple places.
5086 void LSRInstance::RewriteForPHI(
5087  PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5088  SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5090  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5091  if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5092  BasicBlock *BB = PN->getIncomingBlock(i);
5093 
5094  // If this is a critical edge, split the edge so that we do not insert
5095  // the code on all predecessor/successor paths. We do this unless this
5096  // is the canonical backedge for this loop, which complicates post-inc
5097  // users.
5098  if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5099  !isa<IndirectBrInst>(BB->getTerminator()) &&
5100  !isa<CatchSwitchInst>(BB->getTerminator())) {
5101  BasicBlock *Parent = PN->getParent();
5102  Loop *PNLoop = LI.getLoopFor(Parent);
5103  if (!PNLoop || Parent != PNLoop->getHeader()) {
5104  // Split the critical edge.
5105  BasicBlock *NewBB = nullptr;
5106  if (!Parent->isLandingPad()) {
5107  NewBB = SplitCriticalEdge(BB, Parent,
5109  .setMergeIdenticalEdges()
5110  .setDontDeleteUselessPHIs());
5111  } else {
5113  SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5114  NewBB = NewBBs[0];
5115  }
5116  // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5117  // phi predecessors are identical. The simple thing to do is skip
5118  // splitting in this case rather than complicate the API.
5119  if (NewBB) {
5120  // If PN is outside of the loop and BB is in the loop, we want to
5121  // move the block to be immediately before the PHI block, not
5122  // immediately after BB.
5123  if (L->contains(BB) && !L->contains(PN))
5124  NewBB->moveBefore(PN->getParent());
5125 
5126  // Splitting the edge can reduce the number of PHI entries we have.
5127  e = PN->getNumIncomingValues();
5128  BB = NewBB;
5129  i = PN->getBasicBlockIndex(BB);
5130  }
5131  }
5132  }
5133 
5134  std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5135  Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5136  if (!Pair.second)
5137  PN->setIncomingValue(i, Pair.first->second);
5138  else {
5139  Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5140  Rewriter, DeadInsts);
5141 
5142  // If this is reuse-by-noop-cast, insert the noop cast.
5143  Type *OpTy = LF.OperandValToReplace->getType();
5144  if (FullV->getType() != OpTy)
5145  FullV =
5147  OpTy, false),
5148  FullV, LF.OperandValToReplace->getType(),
5149  "tmp", BB->getTerminator());
5150 
5151  PN->setIncomingValue(i, FullV);
5152  Pair.first->second = FullV;
5153  }
5154  }
5155 }
5156 
5157 /// Emit instructions for the leading candidate expression for this LSRUse (this
5158 /// is called "expanding"), and update the UserInst to reference the newly
5159 /// expanded value.
5160 void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5161  const Formula &F, SCEVExpander &Rewriter,
5162  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5163  // First, find an insertion point that dominates UserInst. For PHI nodes,
5164  // find the nearest block which dominates all the relevant uses.
5165  if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5166  RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5167  } else {
5168  Value *FullV =
5169  Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5170 
5171  // If this is reuse-by-noop-cast, insert the noop cast.
5172  Type *OpTy = LF.OperandValToReplace->getType();
5173  if (FullV->getType() != OpTy) {
5174  Instruction *Cast =
5175  CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5176  FullV, OpTy, "tmp", LF.UserInst);
5177  FullV = Cast;
5178  }
5179 
5180  // Update the user. ICmpZero is handled specially here (for now) because
5181  // Expand may have updated one of the operands of the icmp already, and
5182  // its new value may happen to be equal to LF.OperandValToReplace, in
5183  // which case doing replaceUsesOfWith leads to replacing both operands
5184  // with the same value. TODO: Reorganize this.
5185  if (LU.Kind == LSRUse::ICmpZero)
5186  LF.UserInst->setOperand(0, FullV);
5187  else
5188  LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5189  }
5190 
5191  DeadInsts.emplace_back(LF.OperandValToReplace);
5192 }
5193 
5194 /// Rewrite all the fixup locations with new values, following the chosen
5195 /// solution.
5196 void LSRInstance::ImplementSolution(
5197  const SmallVectorImpl<const Formula *> &Solution) {
5198  // Keep track of instructions we may have made dead, so that
5199  // we can remove them after we are done working.
5201 
5203  "lsr");
5204 #ifndef NDEBUG
5205  Rewriter.setDebugType(DEBUG_TYPE);
5206 #endif
5207  Rewriter.disableCanonicalMode();
5208  Rewriter.enableLSRMode();
5209  Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5210 
5211  // Mark phi nodes that terminate chains so the expander tries to reuse them.
5212  for (const IVChain &Chain : IVChainVec) {
5213  if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5214  Rewriter.setChainedPhi(PN);
5215  }
5216 
5217  // Expand the new value definitions and update the users.
5218  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5219  for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5220  Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5221  Changed = true;
5222  }
5223 
5224  for (const IVChain &Chain : IVChainVec) {
5225  GenerateIVChain(Chain, Rewriter, DeadInsts);
5226  Changed = true;
5227  }
5228  // Clean up after ourselves. This must be done before deleting any
5229  // instructions.
5230  Rewriter.clear();
5231 
5232  Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5233 }
5234 
5235 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5236  DominatorTree &DT, LoopInfo &LI,
5237  const TargetTransformInfo &TTI)
5238  : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L), Changed(false),
5239  IVIncInsertPos(nullptr) {
5240  // If LoopSimplify form is not available, stay out of trouble.
5241  if (!L->isLoopSimplifyForm())
5242  return;
5243 
5244  // If there's no interesting work to be done, bail early.
5245  if (IU.empty()) return;
5246 
5247  // If there's too much analysis to be done, bail early. We won't be able to
5248  // model the problem anyway.
5249  unsigned NumUsers = 0;
5250  for (const IVStrideUse &U : IU) {
5251  if (++NumUsers > MaxIVUsers) {
5252  (void)U;
5253  DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U << "\n");
5254  return;
5255  }
5256  // Bail out if we have a PHI on an EHPad that gets a value from a
5257  // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5258  // no good place to stick any instructions.
5259  if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5260  auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5261  if (isa<FuncletPadInst>(FirstNonPHI) ||
5262  isa<CatchSwitchInst>(FirstNonPHI))
5263  for (BasicBlock *PredBB : PN->blocks())
5264  if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5265  return;
5266  }
5267  }
5268 
5269 #ifndef NDEBUG
5270  // All dominating loops must have preheaders, or SCEVExpander may not be able
5271  // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5272  //
5273  // IVUsers analysis should only create users that are dominated by simple loop
5274  // headers. Since this loop should dominate all of its users, its user list
5275  // should be empty if this loop itself is not within a simple loop nest.
5276  for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5277  Rung; Rung = Rung->getIDom()) {
5278  BasicBlock *BB = Rung->getBlock();
5279  const Loop *DomLoop = LI.getLoopFor(BB);
5280  if (DomLoop && DomLoop->getHeader() == BB) {
5281  assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
5282  }
5283  }
5284 #endif // DEBUG
5285 
5286  DEBUG(dbgs() << "\nLSR on loop ";
5287  L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
5288  dbgs() << ":\n");
5289 
5290  // First, perform some low-level loop optimizations.
5291  OptimizeShadowIV();
5292  OptimizeLoopTermCond();
5293 
5294  // If loop preparation eliminates all interesting IV users, bail.
5295  if (IU.empty()) return;
5296 
5297  // Skip nested loops until we can model them better with formulae.
5298  if (!L->empty()) {
5299  DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
5300  return;
5301  }
5302 
5303  // Start collecting data and preparing for the solver.
5304  CollectChains();
5305  CollectInterestingTypesAndFactors();
5306  CollectFixupsAndInitialFormulae();
5307  CollectLoopInvariantFixupsAndFormulae();
5308 
5309  assert(!Uses.empty() && "IVUsers reported at least one use");
5310  DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
5311  print_uses(dbgs()));
5312 
5313  // Now use the reuse data to generate a bunch of interesting ways
5314  // to formulate the values needed for the uses.
5315  GenerateAllReuseFormulae();
5316 
5317  FilterOutUndesirableDedicatedRegisters();
5318  NarrowSearchSpaceUsingHeuristics();
5319 
5321  Solve(Solution);
5322 
5323  // Release memory that is no longer needed.
5324  Factors.clear();
5325  Types.clear();
5326  RegUses.clear();
5327 
5328  if (Solution.empty())
5329  return;
5330 
5331 #ifndef NDEBUG
5332  // Formulae should be legal.
5333  for (const LSRUse &LU : Uses) {
5334  for (const Formula &F : LU.Formulae)
5335  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
5336  F) && "Illegal formula generated!");
5337  };
5338 #endif
5339 
5340  // Now that we've decided what we want, make it so.
5341  ImplementSolution(Solution);
5342 }
5343 
5344 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5345 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5346  if (Factors.empty() && Types.empty()) return;
5347 
5348  OS << "LSR has identified the following interesting factors and types: ";
5349  bool First = true;
5350 
5351  for (int64_t Factor : Factors) {
5352  if (!First) OS << ", ";
5353  First = false;
5354  OS << '*' << Factor;
5355  }
5356 
5357  for (Type *Ty : Types) {
5358  if (!First) OS << ", ";
5359  First = false;
5360  OS << '(' << *Ty << ')';
5361  }
5362  OS << '\n';
5363 }
5364 
5365 void LSRInstance::print_fixups(raw_ostream &OS) const {
5366  OS << "LSR is examining the following fixup sites:\n";
5367  for (const LSRUse &LU : Uses)
5368  for (const LSRFixup &LF : LU.Fixups) {
5369  dbgs() << " ";
5370  LF.print(OS);
5371  OS << '\n';
5372  }
5373 }
5374 
5375 void LSRInstance::print_uses(raw_ostream &OS) const {
5376  OS << "LSR is examining the following uses:\n";
5377  for (const LSRUse &LU : Uses) {
5378  dbgs() << " ";
5379  LU.print(OS);
5380  OS << '\n';
5381  for (const Formula &F : LU.Formulae) {
5382  OS << " ";
5383  F.print(OS);
5384  OS << '\n';
5385  }
5386  }
5387 }
5388 
5389 void LSRInstance::print(raw_ostream &OS) const {
5390  print_factors_and_types(OS);
5391  print_fixups(OS);
5392  print_uses(OS);
5393 }
5394 
5395 LLVM_DUMP_METHOD void LSRInstance::dump() const {
5396  print(errs()); errs() << '\n';
5397 }
5398 #endif
5399 
5400 namespace {
5401 
5402 class LoopStrengthReduce : public LoopPass {
5403 public:
5404  static char ID; // Pass ID, replacement for typeid
5405 
5406  LoopStrengthReduce();
5407 
5408 private:
5409  bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5410  void getAnalysisUsage(AnalysisUsage &AU) const override;
5411 };
5412 
5413 } // end anonymous namespace
5414 
5415 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5417 }
5418 
5419 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5420  // We split critical edges, so we change the CFG. However, we do update
5421  // many analyses if they are around.
5423 
5431  // Requiring LoopSimplify a second time here prevents IVUsers from running
5432  // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5437 }
5438 
5440  DominatorTree &DT, LoopInfo &LI,
5441  const TargetTransformInfo &TTI) {
5442  bool Changed = false;
5443 
5444  // Run the main LSR transformation.
5445  Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5446 
5447  // Remove any extra phis created by processing inner loops.
5448  Changed |= DeleteDeadPHIs(L->getHeader());
5449  if (EnablePhiElim && L->isLoopSimplifyForm()) {
5451  const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5452  SCEVExpander Rewriter(SE, DL, "lsr");
5453 #ifndef NDEBUG
5454  Rewriter.setDebugType(DEBUG_TYPE);
5455 #endif
5456  unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5457  if (numFolded) {
5458  Changed = true;
5460  DeleteDeadPHIs(L->getHeader());
5461  }
5462  }
5463  return Changed;
5464 }
5465 
5466 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5467  if (skipLoop(L))
5468  return false;
5469 
5470  auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5471  auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5472  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5473  auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5474  const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5475  *L->getHeader()->getParent());
5476  return ReduceLoopStrength(L, IU, SE, DT, LI, TTI);
5477 }
5478 
5481  LPMUpdater &) {
5482  if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5483  AR.DT, AR.LI, AR.TTI))
5484  return PreservedAnalyses::all();
5485 
5487 }
5488 
5489 char LoopStrengthReduce::ID = 0;
5490 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5491  "Loop Strength Reduction", false, false)
5497 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5498 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5499  "Loop Strength Reduction", false, false)
5500 
5501 Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }
Pass interface - Implemented by all &#39;passes&#39;.
Definition: Pass.h:81
static bool isValueValidForType(Type *Ty, uint64_t V)
This static method returns true if the type Ty is big enough to represent the value V...
Definition: Constants.cpp:1172
uint64_t CallInst * C
DomTreeNodeBase< NodeT > * getNode(NodeT *BB) const
getNode - return the (Post)DominatorTree node for the specified basic block.
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:69
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:109
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:243
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
iterator_range< use_iterator > uses()
Definition: Value.h:350
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
raw_ostream & errs()
This returns a reference to a raw_ostream for standard error.
BlockT * getLoopLatch() const
If there is a single latch block for this loop, return it.
Definition: LoopInfoImpl.h:157
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
PreservedAnalyses getLoopPassPreservedAnalyses()
Returns the minimum set of Analyses that all loop passes must preserve.
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:234
DiagnosticInfoOptimizationBase::Argument NV
typename SuperClass::const_iterator const_iterator
Definition: SmallVector.h:329
static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE)
Return true if this AddRec is already a phi in its loop.
Pass * createLoopStrengthReducePass()
This is a &#39;bitvector&#39; (really, a variable-sized bit array), optimized for the case when the array is ...
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
const SCEV * getConstant(ConstantInt *V)
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
bool isLegalAddImmediate(int64_t Imm) const
Return true if the specified immediate is legal add immediate, that is the target has add instruction...
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds...
Definition: Compiler.h:449
size_type size() const
Determine the number of elements in the SetVector.
Definition: SetVector.h:78
This header provides classes for managing a pipeline of passes over loops in LLVM IR...
Type * getEffectiveSCEVType(Type *Ty) const
Return a type with the same bitwidth as the given type and which represents how SCEV will treat the g...
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
APInt sdiv(const APInt &RHS) const
Signed division function for APInt.
Definition: APInt.cpp:1590
an instruction that atomically checks whether a specified value is in a memory location, and, if it is, stores a new value there.
Definition: Instructions.h:514
const SCEV * normalizeForPostIncUse(const SCEV *S, const PostIncLoopSet &Loops, ScalarEvolution &SE)
Normalize S to be post-increment for all loops present in Loops.
static const size_t ComplexityLimit
Implements a dense probed hash-table based set.
Definition: DenseSet.h:221
unsigned getLoopDepth() const
Return the nesting level of this loop.
Definition: LoopInfo.h:99
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:885
static const unsigned MaxIVUsers
MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for bail out...
bool properlyDominates(const SCEV *S, const BasicBlock *BB)
Return true if elements that makes up the given SCEV properly dominate the specified basic block...
void initializeLoopStrengthReducePass(PassRegistry &)
uint64_t getTypeSizeInBits(Type *Ty) const
Return the size in bits of the specified type, for which isSCEVable must return true.
void setDebugType(const char *s)
NodeT * findNearestCommonDominator(NodeT *A, NodeT *B) const
findNearestCommonDominator - Find nearest common dominator basic block for basic block A and B...
bool isLoopInvariant(const SCEV *S, const Loop *L)
Return true if the value of the given SCEV is unchanging in the specified loop.
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:816
void resize(unsigned N, bool t=false)
Grow or shrink the bitvector.
static void dump(StringRef Title, SpillInfo const &Spills)
Definition: CoroFrame.cpp:285
static bool isCanonical(const MDString *S)
bool properlyDominates(const DomTreeNodeBase< NodeT > *A, const DomTreeNodeBase< NodeT > *B) const
properlyDominates - Returns true iff A dominates B and A != B.
The adaptor from a function pass to a loop pass computes these analyses and makes them available to t...
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
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:336
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:958
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:631
bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg, int64_t Scale, unsigned AddrSpace=0, Instruction *I=nullptr) const
Return true if the addressing mode represented by AM is legal for this target, for a load/store of th...
void clearPostInc()
Disable all post-inc expansion.
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
static User::op_iterator findIVOperand(User::op_iterator OI, User::op_iterator OE, Loop *L, ScalarEvolution &SE)
Helper for CollectChains that finds an IV operand (computed by an AddRec in this loop) within [OI...
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
LLVMContext & getContext() const
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:664
Reg
All possible values of the reg field in the ModR/M byte.