LLVM  9.0.0svn
LoopStrengthReduce.cpp
Go to the documentation of this file.
1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This transformation analyzes and transforms the induction variables (and
10 // computations derived from them) into forms suitable for efficient execution
11 // on the target.
12 //
13 // This pass performs a strength reduction on array references inside loops that
14 // have as one or more of their components the loop induction variable, it
15 // rewrites expressions to take advantage of scaled-index addressing modes
16 // available on the target, and it performs a variety of other optimizations
17 // related to loop induction variables.
18 //
19 // Terminology note: this code has a lot of handling for "post-increment" or
20 // "post-inc" users. This is not talking about post-increment addressing modes;
21 // it is instead talking about code like this:
22 //
23 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
24 // ...
25 // %i.next = add %i, 1
26 // %c = icmp eq %i.next, %n
27 //
28 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
29 // it's useful to think about these as the same register, with some uses using
30 // the value of the register before the add and some using it after. In this
31 // example, the icmp is a post-increment user, since it uses %i.next, which is
32 // the value of the induction variable after the increment. The other common
33 // case of post-increment users is users outside the loop.
34 //
35 // TODO: More sophistication in the way Formulae are generated and filtered.
36 //
37 // TODO: Handle multiple loops at a time.
38 //
39 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
40 // of a GlobalValue?
41 //
42 // TODO: When truncation is free, truncate ICmp users' operands to make it a
43 // smaller encoding (on x86 at least).
44 //
45 // TODO: When a negated register is used by an add (such as in a list of
46 // multiple base registers, or as the increment expression in an addrec),
47 // we may not actually need both reg and (-1 * reg) in registers; the
48 // negation can be implemented by using a sub instead of an add. The
49 // lack of support for taking this into consideration when making
50 // register pressure decisions is partly worked around by the "Special"
51 // use kind.
52 //
53 //===----------------------------------------------------------------------===//
54 
56 #include "llvm/ADT/APInt.h"
57 #include "llvm/ADT/DenseMap.h"
58 #include "llvm/ADT/DenseSet.h"
59 #include "llvm/ADT/Hashing.h"
61 #include "llvm/ADT/STLExtras.h"
62 #include "llvm/ADT/SetVector.h"
64 #include "llvm/ADT/SmallPtrSet.h"
65 #include "llvm/ADT/SmallSet.h"
66 #include "llvm/ADT/SmallVector.h"
68 #include "llvm/Analysis/IVUsers.h"
70 #include "llvm/Analysis/LoopInfo.h"
71 #include "llvm/Analysis/LoopPass.h"
78 #include "llvm/Config/llvm-config.h"
79 #include "llvm/IR/BasicBlock.h"
80 #include "llvm/IR/Constant.h"
81 #include "llvm/IR/Constants.h"
82 #include "llvm/IR/DerivedTypes.h"
83 #include "llvm/IR/Dominators.h"
84 #include "llvm/IR/GlobalValue.h"
85 #include "llvm/IR/IRBuilder.h"
86 #include "llvm/IR/InstrTypes.h"
87 #include "llvm/IR/Instruction.h"
88 #include "llvm/IR/Instructions.h"
89 #include "llvm/IR/IntrinsicInst.h"
90 #include "llvm/IR/Intrinsics.h"
91 #include "llvm/IR/Module.h"
92 #include "llvm/IR/OperandTraits.h"
93 #include "llvm/IR/Operator.h"
94 #include "llvm/IR/PassManager.h"
95 #include "llvm/IR/Type.h"
96 #include "llvm/IR/Use.h"
97 #include "llvm/IR/User.h"
98 #include "llvm/IR/Value.h"
99 #include "llvm/IR/ValueHandle.h"
100 #include "llvm/Pass.h"
101 #include "llvm/Support/Casting.h"
103 #include "llvm/Support/Compiler.h"
104 #include "llvm/Support/Debug.h"
106 #include "llvm/Support/MathExtras.h"
108 #include "llvm/Transforms/Scalar.h"
109 #include "llvm/Transforms/Utils.h"
111 #include <algorithm>
112 #include <cassert>
113 #include <cstddef>
114 #include <cstdint>
115 #include <cstdlib>
116 #include <iterator>
117 #include <limits>
118 #include <numeric>
119 #include <map>
120 #include <utility>
121 
122 using namespace llvm;
123 
124 #define DEBUG_TYPE "loop-reduce"
125 
126 /// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
127 /// bail out. This threshold is far beyond the number of users that LSR can
128 /// conceivably solve, so it should not affect generated code, but catches the
129 /// worst cases before LSR burns too much compile time and stack space.
130 static const unsigned MaxIVUsers = 200;
131 
132 // Temporary flag to cleanup congruent phis after LSR phi expansion.
133 // It's currently disabled until we can determine whether it's truly useful or
134 // not. The flag should be removed after the v3.0 release.
135 // This is now needed for ivchains.
137  "enable-lsr-phielim", cl::Hidden, cl::init(true),
138  cl::desc("Enable LSR phi elimination"));
139 
140 // The flag adds instruction count to solutions cost comparision.
141 static cl::opt<bool> InsnsCost(
142  "lsr-insns-cost", cl::Hidden, cl::init(true),
143  cl::desc("Add instruction count to a LSR cost model"));
144 
145 // Flag to choose how to narrow complex lsr solution
147  "lsr-exp-narrow", cl::Hidden, cl::init(false),
148  cl::desc("Narrow LSR complex solution using"
149  " expectation of registers number"));
150 
151 // Flag to narrow search space by filtering non-optimal formulae with
152 // the same ScaledReg and Scale.
154  "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
155  cl::desc("Narrow LSR search space by filtering non-optimal formulae"
156  " with the same ScaledReg and Scale"));
157 
159  "lsr-backedge-indexing", cl::Hidden, cl::init(true),
160  cl::desc("Enable the generation of cross iteration indexed memops"));
161 
163  "lsr-complexity-limit", cl::Hidden,
165  cl::desc("LSR search space complexity limit"));
166 
168  "lsr-recursive-setupcost", cl::Hidden, cl::init(true),
169  cl::desc("Enable more thorough lsr setup cost calculation"));
170 
171 #ifndef NDEBUG
172 // Stress test IV chain generation.
174  "stress-ivchain", cl::Hidden, cl::init(false),
175  cl::desc("Stress test LSR IV chains"));
176 #else
177 static bool StressIVChain = false;
178 #endif
179 
180 namespace {
181 
182 struct MemAccessTy {
183  /// Used in situations where the accessed memory type is unknown.
184  static const unsigned UnknownAddressSpace =
186 
187  Type *MemTy = nullptr;
188  unsigned AddrSpace = UnknownAddressSpace;
189 
190  MemAccessTy() = default;
191  MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
192 
193  bool operator==(MemAccessTy Other) const {
194  return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
195  }
196 
197  bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
198 
199  static MemAccessTy getUnknown(LLVMContext &Ctx,
200  unsigned AS = UnknownAddressSpace) {
201  return MemAccessTy(Type::getVoidTy(Ctx), AS);
202  }
203 
204  Type *getType() { return MemTy; }
205 };
206 
207 /// This class holds data which is used to order reuse candidates.
208 class RegSortData {
209 public:
210  /// This represents the set of LSRUse indices which reference
211  /// a particular register.
212  SmallBitVector UsedByIndices;
213 
214  void print(raw_ostream &OS) const;
215  void dump() const;
216 };
217 
218 } // end anonymous namespace
219 
220 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
221 void RegSortData::print(raw_ostream &OS) const {
222  OS << "[NumUses=" << UsedByIndices.count() << ']';
223 }
224 
225 LLVM_DUMP_METHOD void RegSortData::dump() const {
226  print(errs()); errs() << '\n';
227 }
228 #endif
229 
230 namespace {
231 
232 /// Map register candidates to information about how they are used.
233 class RegUseTracker {
234  using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
235 
236  RegUsesTy RegUsesMap;
238 
239 public:
240  void countRegister(const SCEV *Reg, size_t LUIdx);
241  void dropRegister(const SCEV *Reg, size_t LUIdx);
242  void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
243 
244  bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
245 
246  const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
247 
248  void clear();
249 
251  using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
252 
253  iterator begin() { return RegSequence.begin(); }
254  iterator end() { return RegSequence.end(); }
255  const_iterator begin() const { return RegSequence.begin(); }
256  const_iterator end() const { return RegSequence.end(); }
257 };
258 
259 } // end anonymous namespace
260 
261 void
262 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
263  std::pair<RegUsesTy::iterator, bool> Pair =
264  RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
265  RegSortData &RSD = Pair.first->second;
266  if (Pair.second)
267  RegSequence.push_back(Reg);
268  RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
269  RSD.UsedByIndices.set(LUIdx);
270 }
271 
272 void
273 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
274  RegUsesTy::iterator It = RegUsesMap.find(Reg);
275  assert(It != RegUsesMap.end());
276  RegSortData &RSD = It->second;
277  assert(RSD.UsedByIndices.size() > LUIdx);
278  RSD.UsedByIndices.reset(LUIdx);
279 }
280 
281 void
282 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
283  assert(LUIdx <= LastLUIdx);
284 
285  // Update RegUses. The data structure is not optimized for this purpose;
286  // we must iterate through it and update each of the bit vectors.
287  for (auto &Pair : RegUsesMap) {
288  SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
289  if (LUIdx < UsedByIndices.size())
290  UsedByIndices[LUIdx] =
291  LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
292  UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
293  }
294 }
295 
296 bool
297 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
298  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
299  if (I == RegUsesMap.end())
300  return false;
301  const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
302  int i = UsedByIndices.find_first();
303  if (i == -1) return false;
304  if ((size_t)i != LUIdx) return true;
305  return UsedByIndices.find_next(i) != -1;
306 }
307 
308 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
309  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
310  assert(I != RegUsesMap.end() && "Unknown register!");
311  return I->second.UsedByIndices;
312 }
313 
314 void RegUseTracker::clear() {
315  RegUsesMap.clear();
316  RegSequence.clear();
317 }
318 
319 namespace {
320 
321 /// This class holds information that describes a formula for computing
322 /// satisfying a use. It may include broken-out immediates and scaled registers.
323 struct Formula {
324  /// Global base address used for complex addressing.
325  GlobalValue *BaseGV = nullptr;
326 
327  /// Base offset for complex addressing.
328  int64_t BaseOffset = 0;
329 
330  /// Whether any complex addressing has a base register.
331  bool HasBaseReg = false;
332 
333  /// The scale of any complex addressing.
334  int64_t Scale = 0;
335 
336  /// The list of "base" registers for this use. When this is non-empty. The
337  /// canonical representation of a formula is
338  /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
339  /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
340  /// 3. The reg containing recurrent expr related with currect loop in the
341  /// formula should be put in the ScaledReg.
342  /// #1 enforces that the scaled register is always used when at least two
343  /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
344  /// #2 enforces that 1 * reg is reg.
345  /// #3 ensures invariant regs with respect to current loop can be combined
346  /// together in LSR codegen.
347  /// This invariant can be temporarily broken while building a formula.
348  /// However, every formula inserted into the LSRInstance must be in canonical
349  /// form.
351 
352  /// The 'scaled' register for this use. This should be non-null when Scale is
353  /// not zero.
354  const SCEV *ScaledReg = nullptr;
355 
356  /// An additional constant offset which added near the use. This requires a
357  /// temporary register, but the offset itself can live in an add immediate
358  /// field rather than a register.
359  int64_t UnfoldedOffset = 0;
360 
361  Formula() = default;
362 
363  void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
364 
365  bool isCanonical(const Loop &L) const;
366 
367  void canonicalize(const Loop &L);
368 
369  bool unscale();
370 
371  bool hasZeroEnd() const;
372 
373  size_t getNumRegs() const;
374  Type *getType() const;
375 
376  void deleteBaseReg(const SCEV *&S);
377 
378  bool referencesReg(const SCEV *S) const;
379  bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
380  const RegUseTracker &RegUses) const;
381 
382  void print(raw_ostream &OS) const;
383  void dump() const;
384 };
385 
386 } // end anonymous namespace
387 
388 /// Recursion helper for initialMatch.
389 static void DoInitialMatch(const SCEV *S, Loop *L,
392  ScalarEvolution &SE) {
393  // Collect expressions which properly dominate the loop header.
394  if (SE.properlyDominates(S, L->getHeader())) {
395  Good.push_back(S);
396  return;
397  }
398 
399  // Look at add operands.
400  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
401  for (const SCEV *S : Add->operands())
402  DoInitialMatch(S, L, Good, Bad, SE);
403  return;
404  }
405 
406  // Look at addrec operands.
407  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
408  if (!AR->getStart()->isZero() && AR->isAffine()) {
409  DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
410  DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
411  AR->getStepRecurrence(SE),
412  // FIXME: AR->getNoWrapFlags()
413  AR->getLoop(), SCEV::FlagAnyWrap),
414  L, Good, Bad, SE);
415  return;
416  }
417 
418  // Handle a multiplication by -1 (negation) if it didn't fold.
419  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
420  if (Mul->getOperand(0)->isAllOnesValue()) {
421  SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
422  const SCEV *NewMul = SE.getMulExpr(Ops);
423 
426  DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
427  const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
428  SE.getEffectiveSCEVType(NewMul->getType())));
429  for (const SCEV *S : MyGood)
430  Good.push_back(SE.getMulExpr(NegOne, S));
431  for (const SCEV *S : MyBad)
432  Bad.push_back(SE.getMulExpr(NegOne, S));
433  return;
434  }
435 
436  // Ok, we can't do anything interesting. Just stuff the whole thing into a
437  // register and hope for the best.
438  Bad.push_back(S);
439 }
440 
441 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
442 /// all loop-invariant and loop-computable values in a single base register.
443 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
446  DoInitialMatch(S, L, Good, Bad, SE);
447  if (!Good.empty()) {
448  const SCEV *Sum = SE.getAddExpr(Good);
449  if (!Sum->isZero())
450  BaseRegs.push_back(Sum);
451  HasBaseReg = true;
452  }
453  if (!Bad.empty()) {
454  const SCEV *Sum = SE.getAddExpr(Bad);
455  if (!Sum->isZero())
456  BaseRegs.push_back(Sum);
457  HasBaseReg = true;
458  }
459  canonicalize(*L);
460 }
461 
462 /// Check whether or not this formula satisfies the canonical
463 /// representation.
464 /// \see Formula::BaseRegs.
465 bool Formula::isCanonical(const Loop &L) const {
466  if (!ScaledReg)
467  return BaseRegs.size() <= 1;
468 
469  if (Scale != 1)
470  return true;
471 
472  if (Scale == 1 && BaseRegs.empty())
473  return false;
474 
475  const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
476  if (SAR && SAR->getLoop() == &L)
477  return true;
478 
479  // If ScaledReg is not a recurrent expr, or it is but its loop is not current
480  // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
481  // loop, we want to swap the reg in BaseRegs with ScaledReg.
482  auto I =
483  find_if(make_range(BaseRegs.begin(), BaseRegs.end()), [&](const SCEV *S) {
484  return isa<const SCEVAddRecExpr>(S) &&
485  (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
486  });
487  return I == BaseRegs.end();
488 }
489 
490 /// Helper method to morph a formula into its canonical representation.
491 /// \see Formula::BaseRegs.
492 /// Every formula having more than one base register, must use the ScaledReg
493 /// field. Otherwise, we would have to do special cases everywhere in LSR
494 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
495 /// On the other hand, 1*reg should be canonicalized into reg.
496 void Formula::canonicalize(const Loop &L) {
497  if (isCanonical(L))
498  return;
499  // So far we did not need this case. This is easy to implement but it is
500  // useless to maintain dead code. Beside it could hurt compile time.
501  assert(!BaseRegs.empty() && "1*reg => reg, should not be needed.");
502 
503  // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
504  if (!ScaledReg) {
505  ScaledReg = BaseRegs.back();
506  BaseRegs.pop_back();
507  Scale = 1;
508  }
509 
510  // If ScaledReg is an invariant with respect to L, find the reg from
511  // BaseRegs containing the recurrent expr related with Loop L. Swap the
512  // reg with ScaledReg.
513  const SCEVAddRecExpr *SAR = dyn_cast<const SCEVAddRecExpr>(ScaledReg);
514  if (!SAR || SAR->getLoop() != &L) {
515  auto I = find_if(make_range(BaseRegs.begin(), BaseRegs.end()),
516  [&](const SCEV *S) {
517  return isa<const SCEVAddRecExpr>(S) &&
518  (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
519  });
520  if (I != BaseRegs.end())
521  std::swap(ScaledReg, *I);
522  }
523 }
524 
525 /// Get rid of the scale in the formula.
526 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
527 /// \return true if it was possible to get rid of the scale, false otherwise.
528 /// \note After this operation the formula may not be in the canonical form.
529 bool Formula::unscale() {
530  if (Scale != 1)
531  return false;
532  Scale = 0;
533  BaseRegs.push_back(ScaledReg);
534  ScaledReg = nullptr;
535  return true;
536 }
537 
538 bool Formula::hasZeroEnd() const {
539  if (UnfoldedOffset || BaseOffset)
540  return false;
541  if (BaseRegs.size() != 1 || ScaledReg)
542  return false;
543  return true;
544 }
545 
546 /// Return the total number of register operands used by this formula. This does
547 /// not include register uses implied by non-constant addrec strides.
548 size_t Formula::getNumRegs() const {
549  return !!ScaledReg + BaseRegs.size();
550 }
551 
552 /// Return the type of this formula, if it has one, or null otherwise. This type
553 /// is meaningless except for the bit size.
554 Type *Formula::getType() const {
555  return !BaseRegs.empty() ? BaseRegs.front()->getType() :
556  ScaledReg ? ScaledReg->getType() :
557  BaseGV ? BaseGV->getType() :
558  nullptr;
559 }
560 
561 /// Delete the given base reg from the BaseRegs list.
562 void Formula::deleteBaseReg(const SCEV *&S) {
563  if (&S != &BaseRegs.back())
564  std::swap(S, BaseRegs.back());
565  BaseRegs.pop_back();
566 }
567 
568 /// Test if this formula references the given register.
569 bool Formula::referencesReg(const SCEV *S) const {
570  return S == ScaledReg || is_contained(BaseRegs, S);
571 }
572 
573 /// Test whether this formula uses registers which are used by uses other than
574 /// the use with the given index.
575 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
576  const RegUseTracker &RegUses) const {
577  if (ScaledReg)
578  if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
579  return true;
580  for (const SCEV *BaseReg : BaseRegs)
581  if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
582  return true;
583  return false;
584 }
585 
586 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
587 void Formula::print(raw_ostream &OS) const {
588  bool First = true;
589  if (BaseGV) {
590  if (!First) OS << " + "; else First = false;
591  BaseGV->printAsOperand(OS, /*PrintType=*/false);
592  }
593  if (BaseOffset != 0) {
594  if (!First) OS << " + "; else First = false;
595  OS << BaseOffset;
596  }
597  for (const SCEV *BaseReg : BaseRegs) {
598  if (!First) OS << " + "; else First = false;
599  OS << "reg(" << *BaseReg << ')';
600  }
601  if (HasBaseReg && BaseRegs.empty()) {
602  if (!First) OS << " + "; else First = false;
603  OS << "**error: HasBaseReg**";
604  } else if (!HasBaseReg && !BaseRegs.empty()) {
605  if (!First) OS << " + "; else First = false;
606  OS << "**error: !HasBaseReg**";
607  }
608  if (Scale != 0) {
609  if (!First) OS << " + "; else First = false;
610  OS << Scale << "*reg(";
611  if (ScaledReg)
612  OS << *ScaledReg;
613  else
614  OS << "<unknown>";
615  OS << ')';
616  }
617  if (UnfoldedOffset != 0) {
618  if (!First) OS << " + ";
619  OS << "imm(" << UnfoldedOffset << ')';
620  }
621 }
622 
623 LLVM_DUMP_METHOD void Formula::dump() const {
624  print(errs()); errs() << '\n';
625 }
626 #endif
627 
628 /// Return true if the given addrec can be sign-extended without changing its
629 /// value.
630 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
631  Type *WideTy =
633  return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
634 }
635 
636 /// Return true if the given add can be sign-extended without changing its
637 /// value.
638 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
639  Type *WideTy =
641  return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
642 }
643 
644 /// Return true if the given mul can be sign-extended without changing its
645 /// value.
646 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
647  Type *WideTy =
649  SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
650  return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
651 }
652 
653 /// Return an expression for LHS /s RHS, if it can be determined and if the
654 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
655 /// is true, expressions like (X * Y) /s Y are simplified to Y, ignoring that
656 /// the multiplication may overflow, which is useful when the result will be
657 /// used in a context where the most significant bits are ignored.
658 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
659  ScalarEvolution &SE,
660  bool IgnoreSignificantBits = false) {
661  // Handle the trivial case, which works for any SCEV type.
662  if (LHS == RHS)
663  return SE.getConstant(LHS->getType(), 1);
664 
665  // Handle a few RHS special cases.
666  const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
667  if (RC) {
668  const APInt &RA = RC->getAPInt();
669  // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
670  // some folding.
671  if (RA.isAllOnesValue())
672  return SE.getMulExpr(LHS, RC);
673  // Handle x /s 1 as x.
674  if (RA == 1)
675  return LHS;
676  }
677 
678  // Check for a division of a constant by a constant.
679  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
680  if (!RC)
681  return nullptr;
682  const APInt &LA = C->getAPInt();
683  const APInt &RA = RC->getAPInt();
684  if (LA.srem(RA) != 0)
685  return nullptr;
686  return SE.getConstant(LA.sdiv(RA));
687  }
688 
689  // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
690  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
691  if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
692  const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
693  IgnoreSignificantBits);
694  if (!Step) return nullptr;
695  const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
696  IgnoreSignificantBits);
697  if (!Start) return nullptr;
698  // FlagNW is independent of the start value, step direction, and is
699  // preserved with smaller magnitude steps.
700  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
701  return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
702  }
703  return nullptr;
704  }
705 
706  // Distribute the sdiv over add operands, if the add doesn't overflow.
707  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
708  if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
710  for (const SCEV *S : Add->operands()) {
711  const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
712  if (!Op) return nullptr;
713  Ops.push_back(Op);
714  }
715  return SE.getAddExpr(Ops);
716  }
717  return nullptr;
718  }
719 
720  // Check for a multiply operand that we can pull RHS out of.
721  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
722  if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
724  bool Found = false;
725  for (const SCEV *S : Mul->operands()) {
726  if (!Found)
727  if (const SCEV *Q = getExactSDiv(S, RHS, SE,
728  IgnoreSignificantBits)) {
729  S = Q;
730  Found = true;
731  }
732  Ops.push_back(S);
733  }
734  return Found ? SE.getMulExpr(Ops) : nullptr;
735  }
736  return nullptr;
737  }
738 
739  // Otherwise we don't know.
740  return nullptr;
741 }
742 
743 /// If S involves the addition of a constant integer value, return that integer
744 /// value, and mutate S to point to a new SCEV with that value excluded.
745 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
746  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
747  if (C->getAPInt().getMinSignedBits() <= 64) {
748  S = SE.getConstant(C->getType(), 0);
749  return C->getValue()->getSExtValue();
750  }
751  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
752  SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
753  int64_t Result = ExtractImmediate(NewOps.front(), SE);
754  if (Result != 0)
755  S = SE.getAddExpr(NewOps);
756  return Result;
757  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
758  SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
759  int64_t Result = ExtractImmediate(NewOps.front(), SE);
760  if (Result != 0)
761  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
762  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
764  return Result;
765  }
766  return 0;
767 }
768 
769 /// If S involves the addition of a GlobalValue address, return that symbol, and
770 /// mutate S to point to a new SCEV with that value excluded.
772  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
773  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
774  S = SE.getConstant(GV->getType(), 0);
775  return GV;
776  }
777  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
778  SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
779  GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
780  if (Result)
781  S = SE.getAddExpr(NewOps);
782  return Result;
783  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
784  SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
785  GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
786  if (Result)
787  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
788  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
790  return Result;
791  }
792  return nullptr;
793 }
794 
795 /// Returns true if the specified instruction is using the specified value as an
796 /// address.
797 static bool isAddressUse(const TargetTransformInfo &TTI,
798  Instruction *Inst, Value *OperandVal) {
799  bool isAddress = isa<LoadInst>(Inst);
800  if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
801  if (SI->getPointerOperand() == OperandVal)
802  isAddress = true;
803  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
804  // Addressing modes can also be folded into prefetches and a variety
805  // of intrinsics.
806  switch (II->getIntrinsicID()) {
807  case Intrinsic::memset:
808  case Intrinsic::prefetch:
809  if (II->getArgOperand(0) == OperandVal)
810  isAddress = true;
811  break;
812  case Intrinsic::memmove:
813  case Intrinsic::memcpy:
814  if (II->getArgOperand(0) == OperandVal ||
815  II->getArgOperand(1) == OperandVal)
816  isAddress = true;
817  break;
818  default: {
819  MemIntrinsicInfo IntrInfo;
820  if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
821  if (IntrInfo.PtrVal == OperandVal)
822  isAddress = true;
823  }
824  }
825  }
826  } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
827  if (RMW->getPointerOperand() == OperandVal)
828  isAddress = true;
829  } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
830  if (CmpX->getPointerOperand() == OperandVal)
831  isAddress = true;
832  }
833  return isAddress;
834 }
835 
836 /// Return the type of the memory being accessed.
837 static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
838  Instruction *Inst, Value *OperandVal) {
839  MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
840  if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
841  AccessTy.MemTy = SI->getOperand(0)->getType();
842  AccessTy.AddrSpace = SI->getPointerAddressSpace();
843  } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
844  AccessTy.AddrSpace = LI->getPointerAddressSpace();
845  } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
846  AccessTy.AddrSpace = RMW->getPointerAddressSpace();
847  } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
848  AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
849  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
850  switch (II->getIntrinsicID()) {
851  case Intrinsic::prefetch:
852  case Intrinsic::memset:
853  AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
854  AccessTy.MemTy = OperandVal->getType();
855  break;
856  case Intrinsic::memmove:
857  case Intrinsic::memcpy:
858  AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
859  AccessTy.MemTy = OperandVal->getType();
860  break;
861  default: {
862  MemIntrinsicInfo IntrInfo;
863  if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
864  AccessTy.AddrSpace
865  = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
866  }
867 
868  break;
869  }
870  }
871  }
872 
873  // All pointers have the same requirements, so canonicalize them to an
874  // arbitrary pointer type to minimize variation.
875  if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
876  AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
877  PTy->getAddressSpace());
878 
879  return AccessTy;
880 }
881 
882 /// Return true if this AddRec is already a phi in its loop.
883 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
884  for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
885  if (SE.isSCEVable(PN.getType()) &&
886  (SE.getEffectiveSCEVType(PN.getType()) ==
887  SE.getEffectiveSCEVType(AR->getType())) &&
888  SE.getSCEV(&PN) == AR)
889  return true;
890  }
891  return false;
892 }
893 
894 /// Check if expanding this expression is likely to incur significant cost. This
895 /// is tricky because SCEV doesn't track which expressions are actually computed
896 /// by the current IR.
897 ///
898 /// We currently allow expansion of IV increments that involve adds,
899 /// multiplication by constants, and AddRecs from existing phis.
900 ///
901 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
902 /// obvious multiple of the UDivExpr.
903 static bool isHighCostExpansion(const SCEV *S,
904  SmallPtrSetImpl<const SCEV*> &Processed,
905  ScalarEvolution &SE) {
906  // Zero/One operand expressions
907  switch (S->getSCEVType()) {
908  case scUnknown:
909  case scConstant:
910  return false;
911  case scTruncate:
912  return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
913  Processed, SE);
914  case scZeroExtend:
915  return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
916  Processed, SE);
917  case scSignExtend:
918  return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
919  Processed, SE);
920  }
921 
922  if (!Processed.insert(S).second)
923  return false;
924 
925  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
926  for (const SCEV *S : Add->operands()) {
927  if (isHighCostExpansion(S, Processed, SE))
928  return true;
929  }
930  return false;
931  }
932 
933  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
934  if (Mul->getNumOperands() == 2) {
935  // Multiplication by a constant is ok
936  if (isa<SCEVConstant>(Mul->getOperand(0)))
937  return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
938 
939  // If we have the value of one operand, check if an existing
940  // multiplication already generates this expression.
941  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
942  Value *UVal = U->getValue();
943  for (User *UR : UVal->users()) {
944  // If U is a constant, it may be used by a ConstantExpr.
945  Instruction *UI = dyn_cast<Instruction>(UR);
946  if (UI && UI->getOpcode() == Instruction::Mul &&
947  SE.isSCEVable(UI->getType())) {
948  return SE.getSCEV(UI) == Mul;
949  }
950  }
951  }
952  }
953  }
954 
955  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
956  if (isExistingPhi(AR, SE))
957  return false;
958  }
959 
960  // Fow now, consider any other type of expression (div/mul/min/max) high cost.
961  return true;
962 }
963 
964 /// If any of the instructions in the specified set are trivially dead, delete
965 /// them and see if this makes any of their operands subsequently dead.
966 static bool
968  bool Changed = false;
969 
970  while (!DeadInsts.empty()) {
971  Value *V = DeadInsts.pop_back_val();
972  Instruction *I = dyn_cast_or_null<Instruction>(V);
973 
974  if (!I || !isInstructionTriviallyDead(I))
975  continue;
976 
977  for (Use &O : I->operands())
978  if (Instruction *U = dyn_cast<Instruction>(O)) {
979  O = nullptr;
980  if (U->use_empty())
981  DeadInsts.emplace_back(U);
982  }
983 
984  I->eraseFromParent();
985  Changed = true;
986  }
987 
988  return Changed;
989 }
990 
991 namespace {
992 
993 class LSRUse;
994 
995 } // end anonymous namespace
996 
997 /// Check if the addressing mode defined by \p F is completely
998 /// folded in \p LU at isel time.
999 /// This includes address-mode folding and special icmp tricks.
1000 /// This function returns true if \p LU can accommodate what \p F
1001 /// defines and up to 1 base + 1 scaled + offset.
1002 /// In other words, if \p F has several base registers, this function may
1003 /// still return true. Therefore, users still need to account for
1004 /// additional base registers and/or unfolded offsets to derive an
1005 /// accurate cost model.
1006 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1007  const LSRUse &LU, const Formula &F);
1008 
1009 // Get the cost of the scaling factor used in F for LU.
1010 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1011  const LSRUse &LU, const Formula &F,
1012  const Loop &L);
1013 
1014 namespace {
1015 
1016 /// This class is used to measure and compare candidate formulae.
1017 class Cost {
1018  const Loop *L = nullptr;
1019  ScalarEvolution *SE = nullptr;
1020  const TargetTransformInfo *TTI = nullptr;
1022 
1023 public:
1024  Cost() = delete;
1025  Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI) :
1026  L(L), SE(&SE), TTI(&TTI) {
1027  C.Insns = 0;
1028  C.NumRegs = 0;
1029  C.AddRecCost = 0;
1030  C.NumIVMuls = 0;
1031  C.NumBaseAdds = 0;
1032  C.ImmCost = 0;
1033  C.SetupCost = 0;
1034  C.ScaleCost = 0;
1035  }
1036 
1037  bool isLess(Cost &Other);
1038 
1039  void Lose();
1040 
1041 #ifndef NDEBUG
1042  // Once any of the metrics loses, they must all remain losers.
1043  bool isValid() {
1044  return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1045  | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1046  || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1047  & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1048  }
1049 #endif
1050 
1051  bool isLoser() {
1052  assert(isValid() && "invalid cost");
1053  return C.NumRegs == ~0u;
1054  }
1055 
1056  void RateFormula(const Formula &F,
1058  const DenseSet<const SCEV *> &VisitedRegs,
1059  const LSRUse &LU,
1060  SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1061 
1062  void print(raw_ostream &OS) const;
1063  void dump() const;
1064 
1065 private:
1066  void RateRegister(const Formula &F, const SCEV *Reg,
1068  void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1070  SmallPtrSetImpl<const SCEV *> *LoserRegs);
1071 };
1072 
1073 /// An operand value in an instruction which is to be replaced with some
1074 /// equivalent, possibly strength-reduced, replacement.
1075 struct LSRFixup {
1076  /// The instruction which will be updated.
1077  Instruction *UserInst = nullptr;
1078 
1079  /// The operand of the instruction which will be replaced. The operand may be
1080  /// used more than once; every instance will be replaced.
1081  Value *OperandValToReplace = nullptr;
1082 
1083  /// If this user is to use the post-incremented value of an induction
1084  /// variable, this set is non-empty and holds the loops associated with the
1085  /// induction variable.
1086  PostIncLoopSet PostIncLoops;
1087 
1088  /// A constant offset to be added to the LSRUse expression. This allows
1089  /// multiple fixups to share the same LSRUse with different offsets, for
1090  /// example in an unrolled loop.
1091  int64_t Offset = 0;
1092 
1093  LSRFixup() = default;
1094 
1095  bool isUseFullyOutsideLoop(const Loop *L) const;
1096 
1097  void print(raw_ostream &OS) const;
1098  void dump() const;
1099 };
1100 
1101 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1102 /// SmallVectors of const SCEV*.
1103 struct UniquifierDenseMapInfo {
1104  static SmallVector<const SCEV *, 4> getEmptyKey() {
1106  V.push_back(reinterpret_cast<const SCEV *>(-1));
1107  return V;
1108  }
1109 
1110  static SmallVector<const SCEV *, 4> getTombstoneKey() {
1112  V.push_back(reinterpret_cast<const SCEV *>(-2));
1113  return V;
1114  }
1115 
1116  static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1117  return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1118  }
1119 
1120  static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1121  const SmallVector<const SCEV *, 4> &RHS) {
1122  return LHS == RHS;
1123  }
1124 };
1125 
1126 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1127 /// as uses invented by LSR itself. It includes information about what kinds of
1128 /// things can be folded into the user, information about the user itself, and
1129 /// information about how the use may be satisfied. TODO: Represent multiple
1130 /// users of the same expression in common?
1131 class LSRUse {
1132  DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1133 
1134 public:
1135  /// An enum for a kind of use, indicating what types of scaled and immediate
1136  /// operands it might support.
1137  enum KindType {
1138  Basic, ///< A normal use, with no folding.
1139  Special, ///< A special case of basic, allowing -1 scales.
1140  Address, ///< An address use; folding according to TargetLowering
1141  ICmpZero ///< An equality icmp with both operands folded into one.
1142  // TODO: Add a generic icmp too?
1143  };
1144 
1145  using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1146 
1147  KindType Kind;
1148  MemAccessTy AccessTy;
1149 
1150  /// The list of operands which are to be replaced.
1152 
1153  /// Keep track of the min and max offsets of the fixups.
1154  int64_t MinOffset = std::numeric_limits<int64_t>::max();
1155  int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1156 
1157  /// This records whether all of the fixups using this LSRUse are outside of
1158  /// the loop, in which case some special-case heuristics may be used.
1159  bool AllFixupsOutsideLoop = true;
1160 
1161  /// RigidFormula is set to true to guarantee that this use will be associated
1162  /// with a single formula--the one that initially matched. Some SCEV
1163  /// expressions cannot be expanded. This allows LSR to consider the registers
1164  /// used by those expressions without the need to expand them later after
1165  /// changing the formula.
1166  bool RigidFormula = false;
1167 
1168  /// This records the widest use type for any fixup using this
1169  /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1170  /// fixup widths to be equivalent, because the narrower one may be relying on
1171  /// the implicit truncation to truncate away bogus bits.
1172  Type *WidestFixupType = nullptr;
1173 
1174  /// A list of ways to build a value that can satisfy this user. After the
1175  /// list is populated, one of these is selected heuristically and used to
1176  /// formulate a replacement for OperandValToReplace in UserInst.
1177  SmallVector<Formula, 12> Formulae;
1178 
1179  /// The set of register candidates used by all formulae in this LSRUse.
1181 
1182  LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1183 
1184  LSRFixup &getNewFixup() {
1185  Fixups.push_back(LSRFixup());
1186  return Fixups.back();
1187  }
1188 
1189  void pushFixup(LSRFixup &f) {
1190  Fixups.push_back(f);
1191  if (f.Offset > MaxOffset)
1192  MaxOffset = f.Offset;
1193  if (f.Offset < MinOffset)
1194  MinOffset = f.Offset;
1195  }
1196 
1197  bool HasFormulaWithSameRegs(const Formula &F) const;
1198  float getNotSelectedProbability(const SCEV *Reg) const;
1199  bool InsertFormula(const Formula &F, const Loop &L);
1200  void DeleteFormula(Formula &F);
1201  void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1202 
1203  void print(raw_ostream &OS) const;
1204  void dump() const;
1205 };
1206 
1207 } // end anonymous namespace
1208 
1209 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1210  LSRUse::KindType Kind, MemAccessTy AccessTy,
1211  GlobalValue *BaseGV, int64_t BaseOffset,
1212  bool HasBaseReg, int64_t Scale,
1213  Instruction *Fixup = nullptr);
1214 
1215 static unsigned getSetupCost(const SCEV *Reg) {
1216  if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1217  return 1;
1219  return 0;
1220  if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1221  return getSetupCost(S->getStart());
1222  if (auto S = dyn_cast<SCEVCastExpr>(Reg))
1223  return getSetupCost(S->getOperand());
1224  if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
1225  return std::accumulate(S->op_begin(), S->op_end(), 0,
1226  [](unsigned i, const SCEV *Reg) {
1227  return i + getSetupCost(Reg);
1228  });
1229  if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1230  return getSetupCost(S->getLHS()) + getSetupCost(S->getRHS());
1231  return 0;
1232 }
1233 
1234 /// Tally up interesting quantities from the given register.
1235 void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1237  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1238  // If this is an addrec for another loop, it should be an invariant
1239  // with respect to L since L is the innermost loop (at least
1240  // for now LSR only handles innermost loops).
1241  if (AR->getLoop() != L) {
1242  // If the AddRec exists, consider it's register free and leave it alone.
1243  if (isExistingPhi(AR, *SE))
1244  return;
1245 
1246  // It is bad to allow LSR for current loop to add induction variables
1247  // for its sibling loops.
1248  if (!AR->getLoop()->contains(L)) {
1249  Lose();
1250  return;
1251  }
1252 
1253  // Otherwise, it will be an invariant with respect to Loop L.
1254  ++C.NumRegs;
1255  return;
1256  }
1257 
1258  unsigned LoopCost = 1;
1259  if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1260  TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1261 
1262  // If the step size matches the base offset, we could use pre-indexed
1263  // addressing.
1264  if (TTI->shouldFavorBackedgeIndex(L)) {
1265  if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1266  if (Step->getAPInt() == F.BaseOffset)
1267  LoopCost = 0;
1268  }
1269 
1270  if (TTI->shouldFavorPostInc()) {
1271  const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1272  if (isa<SCEVConstant>(LoopStep)) {
1273  const SCEV *LoopStart = AR->getStart();
1274  if (!isa<SCEVConstant>(LoopStart) &&
1275  SE->isLoopInvariant(LoopStart, L))
1276  LoopCost = 0;
1277  }
1278  }
1279  }
1280  C.AddRecCost += LoopCost;
1281 
1282  // Add the step value register, if it needs one.
1283  // TODO: The non-affine case isn't precisely modeled here.
1284  if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1285  if (!Regs.count(AR->getOperand(1))) {
1286  RateRegister(F, AR->getOperand(1), Regs);
1287  if (isLoser())
1288  return;
1289  }
1290  }
1291  }
1292  ++C.NumRegs;
1293 
1294  // Rough heuristic; favor registers which don't require extra setup
1295  // instructions in the preheader.
1296  C.SetupCost += getSetupCost(Reg);
1297 
1298  C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1299  SE->hasComputableLoopEvolution(Reg, L);
1300 }
1301 
1302 /// Record this register in the set. If we haven't seen it before, rate
1303 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1304 /// one of those regs an instant loser.
1305 void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1307  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1308  if (LoserRegs && LoserRegs->count(Reg)) {
1309  Lose();
1310  return;
1311  }
1312  if (Regs.insert(Reg).second) {
1313  RateRegister(F, Reg, Regs);
1314  if (LoserRegs && isLoser())
1315  LoserRegs->insert(Reg);
1316  }
1317 }
1318 
1319 void Cost::RateFormula(const Formula &F,
1321  const DenseSet<const SCEV *> &VisitedRegs,
1322  const LSRUse &LU,
1323  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1324  assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1325  // Tally up the registers.
1326  unsigned PrevAddRecCost = C.AddRecCost;
1327  unsigned PrevNumRegs = C.NumRegs;
1328  unsigned PrevNumBaseAdds = C.NumBaseAdds;
1329  if (const SCEV *ScaledReg = F.ScaledReg) {
1330  if (VisitedRegs.count(ScaledReg)) {
1331  Lose();
1332  return;
1333  }
1334  RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1335  if (isLoser())
1336  return;
1337  }
1338  for (const SCEV *BaseReg : F.BaseRegs) {
1339  if (VisitedRegs.count(BaseReg)) {
1340  Lose();
1341  return;
1342  }
1343  RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1344  if (isLoser())
1345  return;
1346  }
1347 
1348  // Determine how many (unfolded) adds we'll need inside the loop.
1349  size_t NumBaseParts = F.getNumRegs();
1350  if (NumBaseParts > 1)
1351  // Do not count the base and a possible second register if the target
1352  // allows to fold 2 registers.
1353  C.NumBaseAdds +=
1354  NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1355  C.NumBaseAdds += (F.UnfoldedOffset != 0);
1356 
1357  // Accumulate non-free scaling amounts.
1358  C.ScaleCost += getScalingFactorCost(*TTI, LU, F, *L);
1359 
1360  // Tally up the non-zero immediates.
1361  for (const LSRFixup &Fixup : LU.Fixups) {
1362  int64_t O = Fixup.Offset;
1363  int64_t Offset = (uint64_t)O + F.BaseOffset;
1364  if (F.BaseGV)
1365  C.ImmCost += 64; // Handle symbolic values conservatively.
1366  // TODO: This should probably be the pointer size.
1367  else if (Offset != 0)
1368  C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1369 
1370  // Check with target if this offset with this instruction is
1371  // specifically not supported.
1372  if (LU.Kind == LSRUse::Address && Offset != 0 &&
1373  !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1374  Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1375  C.NumBaseAdds++;
1376  }
1377 
1378  // If we don't count instruction cost exit here.
1379  if (!InsnsCost) {
1380  assert(isValid() && "invalid cost");
1381  return;
1382  }
1383 
1384  // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1385  // additional instruction (at least fill).
1386  unsigned TTIRegNum = TTI->getNumberOfRegisters(false) - 1;
1387  if (C.NumRegs > TTIRegNum) {
1388  // Cost already exceeded TTIRegNum, then only newly added register can add
1389  // new instructions.
1390  if (PrevNumRegs > TTIRegNum)
1391  C.Insns += (C.NumRegs - PrevNumRegs);
1392  else
1393  C.Insns += (C.NumRegs - TTIRegNum);
1394  }
1395 
1396  // If ICmpZero formula ends with not 0, it could not be replaced by
1397  // just add or sub. We'll need to compare final result of AddRec.
1398  // That means we'll need an additional instruction. But if the target can
1399  // macro-fuse a compare with a branch, don't count this extra instruction.
1400  // For -10 + {0, +, 1}:
1401  // i = i + 1;
1402  // cmp i, 10
1403  //
1404  // For {-10, +, 1}:
1405  // i = i + 1;
1406  if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1407  !TTI->canMacroFuseCmp())
1408  C.Insns++;
1409  // Each new AddRec adds 1 instruction to calculation.
1410  C.Insns += (C.AddRecCost - PrevAddRecCost);
1411 
1412  // BaseAdds adds instructions for unfolded registers.
1413  if (LU.Kind != LSRUse::ICmpZero)
1414  C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1415  assert(isValid() && "invalid cost");
1416 }
1417 
1418 /// Set this cost to a losing value.
1419 void Cost::Lose() {
1422  C.AddRecCost = std::numeric_limits<unsigned>::max();
1423  C.NumIVMuls = std::numeric_limits<unsigned>::max();
1424  C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1426  C.SetupCost = std::numeric_limits<unsigned>::max();
1427  C.ScaleCost = std::numeric_limits<unsigned>::max();
1428 }
1429 
1430 /// Choose the lower cost.
1431 bool Cost::isLess(Cost &Other) {
1432  if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1433  C.Insns != Other.C.Insns)
1434  return C.Insns < Other.C.Insns;
1435  return TTI->isLSRCostLess(C, Other.C);
1436 }
1437 
1438 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1439 void Cost::print(raw_ostream &OS) const {
1440  if (InsnsCost)
1441  OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1442  OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1443  if (C.AddRecCost != 0)
1444  OS << ", with addrec cost " << C.AddRecCost;
1445  if (C.NumIVMuls != 0)
1446  OS << ", plus " << C.NumIVMuls << " IV mul"
1447  << (C.NumIVMuls == 1 ? "" : "s");
1448  if (C.NumBaseAdds != 0)
1449  OS << ", plus " << C.NumBaseAdds << " base add"
1450  << (C.NumBaseAdds == 1 ? "" : "s");
1451  if (C.ScaleCost != 0)
1452  OS << ", plus " << C.ScaleCost << " scale cost";
1453  if (C.ImmCost != 0)
1454  OS << ", plus " << C.ImmCost << " imm cost";
1455  if (C.SetupCost != 0)
1456  OS << ", plus " << C.SetupCost << " setup cost";
1457 }
1458 
1459 LLVM_DUMP_METHOD void Cost::dump() const {
1460  print(errs()); errs() << '\n';
1461 }
1462 #endif
1463 
1464 /// Test whether this fixup always uses its value outside of the given loop.
1465 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1466  // PHI nodes use their value in their incoming blocks.
1467  if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1468  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1469  if (PN->getIncomingValue(i) == OperandValToReplace &&
1470  L->contains(PN->getIncomingBlock(i)))
1471  return false;
1472  return true;
1473  }
1474 
1475  return !L->contains(UserInst);
1476 }
1477 
1478 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1479 void LSRFixup::print(raw_ostream &OS) const {
1480  OS << "UserInst=";
1481  // Store is common and interesting enough to be worth special-casing.
1482  if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1483  OS << "store ";
1484  Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1485  } else if (UserInst->getType()->isVoidTy())
1486  OS << UserInst->getOpcodeName();
1487  else
1488  UserInst->printAsOperand(OS, /*PrintType=*/false);
1489 
1490  OS << ", OperandValToReplace=";
1491  OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1492 
1493  for (const Loop *PIL : PostIncLoops) {
1494  OS << ", PostIncLoop=";
1495  PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1496  }
1497 
1498  if (Offset != 0)
1499  OS << ", Offset=" << Offset;
1500 }
1501 
1502 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1503  print(errs()); errs() << '\n';
1504 }
1505 #endif
1506 
1507 /// Test whether this use as a formula which has the same registers as the given
1508 /// formula.
1509 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1510  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1511  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1512  // Unstable sort by host order ok, because this is only used for uniquifying.
1513  llvm::sort(Key);
1514  return Uniquifier.count(Key);
1515 }
1516 
1517 /// The function returns a probability of selecting formula without Reg.
1518 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1519  unsigned FNum = 0;
1520  for (const Formula &F : Formulae)
1521  if (F.referencesReg(Reg))
1522  FNum++;
1523  return ((float)(Formulae.size() - FNum)) / Formulae.size();
1524 }
1525 
1526 /// If the given formula has not yet been inserted, add it to the list, and
1527 /// return true. Return false otherwise. The formula must be in canonical form.
1528 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1529  assert(F.isCanonical(L) && "Invalid canonical representation");
1530 
1531  if (!Formulae.empty() && RigidFormula)
1532  return false;
1533 
1534  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1535  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1536  // Unstable sort by host order ok, because this is only used for uniquifying.
1537  llvm::sort(Key);
1538 
1539  if (!Uniquifier.insert(Key).second)
1540  return false;
1541 
1542  // Using a register to hold the value of 0 is not profitable.
1543  assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1544  "Zero allocated in a scaled register!");
1545 #ifndef NDEBUG
1546  for (const SCEV *BaseReg : F.BaseRegs)
1547  assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1548 #endif
1549 
1550  // Add the formula to the list.
1551  Formulae.push_back(F);
1552 
1553  // Record registers now being used by this use.
1554  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1555  if (F.ScaledReg)
1556  Regs.insert(F.ScaledReg);
1557 
1558  return true;
1559 }
1560 
1561 /// Remove the given formula from this use's list.
1562 void LSRUse::DeleteFormula(Formula &F) {
1563  if (&F != &Formulae.back())
1564  std::swap(F, Formulae.back());
1565  Formulae.pop_back();
1566 }
1567 
1568 /// Recompute the Regs field, and update RegUses.
1569 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1570  // Now that we've filtered out some formulae, recompute the Regs set.
1571  SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1572  Regs.clear();
1573  for (const Formula &F : Formulae) {
1574  if (F.ScaledReg) Regs.insert(F.ScaledReg);
1575  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1576  }
1577 
1578  // Update the RegTracker.
1579  for (const SCEV *S : OldRegs)
1580  if (!Regs.count(S))
1581  RegUses.dropRegister(S, LUIdx);
1582 }
1583 
1584 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1585 void LSRUse::print(raw_ostream &OS) const {
1586  OS << "LSR Use: Kind=";
1587  switch (Kind) {
1588  case Basic: OS << "Basic"; break;
1589  case Special: OS << "Special"; break;
1590  case ICmpZero: OS << "ICmpZero"; break;
1591  case Address:
1592  OS << "Address of ";
1593  if (AccessTy.MemTy->isPointerTy())
1594  OS << "pointer"; // the full pointer type could be really verbose
1595  else {
1596  OS << *AccessTy.MemTy;
1597  }
1598 
1599  OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1600  }
1601 
1602  OS << ", Offsets={";
1603  bool NeedComma = false;
1604  for (const LSRFixup &Fixup : Fixups) {
1605  if (NeedComma) OS << ',';
1606  OS << Fixup.Offset;
1607  NeedComma = true;
1608  }
1609  OS << '}';
1610 
1611  if (AllFixupsOutsideLoop)
1612  OS << ", all-fixups-outside-loop";
1613 
1614  if (WidestFixupType)
1615  OS << ", widest fixup type: " << *WidestFixupType;
1616 }
1617 
1618 LLVM_DUMP_METHOD void LSRUse::dump() const {
1619  print(errs()); errs() << '\n';
1620 }
1621 #endif
1622 
1624  LSRUse::KindType Kind, MemAccessTy AccessTy,
1625  GlobalValue *BaseGV, int64_t BaseOffset,
1626  bool HasBaseReg, int64_t Scale,
1627  Instruction *Fixup/*= nullptr*/) {
1628  switch (Kind) {
1629  case LSRUse::Address:
1630  return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1631  HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1632 
1633  case LSRUse::ICmpZero:
1634  // There's not even a target hook for querying whether it would be legal to
1635  // fold a GV into an ICmp.
1636  if (BaseGV)
1637  return false;
1638 
1639  // ICmp only has two operands; don't allow more than two non-trivial parts.
1640  if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1641  return false;
1642 
1643  // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1644  // putting the scaled register in the other operand of the icmp.
1645  if (Scale != 0 && Scale != -1)
1646  return false;
1647 
1648  // If we have low-level target information, ask the target if it can fold an
1649  // integer immediate on an icmp.
1650  if (BaseOffset != 0) {
1651  // We have one of:
1652  // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1653  // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1654  // Offs is the ICmp immediate.
1655  if (Scale == 0)
1656  // The cast does the right thing with
1657  // std::numeric_limits<int64_t>::min().
1658  BaseOffset = -(uint64_t)BaseOffset;
1659  return TTI.isLegalICmpImmediate(BaseOffset);
1660  }
1661 
1662  // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1663  return true;
1664 
1665  case LSRUse::Basic:
1666  // Only handle single-register values.
1667  return !BaseGV && Scale == 0 && BaseOffset == 0;
1668 
1669  case LSRUse::Special:
1670  // Special case Basic to handle -1 scales.
1671  return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1672  }
1673 
1674  llvm_unreachable("Invalid LSRUse Kind!");
1675 }
1676 
1678  int64_t MinOffset, int64_t MaxOffset,
1679  LSRUse::KindType Kind, MemAccessTy AccessTy,
1680  GlobalValue *BaseGV, int64_t BaseOffset,
1681  bool HasBaseReg, int64_t Scale) {
1682  // Check for overflow.
1683  if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1684  (MinOffset > 0))
1685  return false;
1686  MinOffset = (uint64_t)BaseOffset + MinOffset;
1687  if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1688  (MaxOffset > 0))
1689  return false;
1690  MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1691 
1692  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1693  HasBaseReg, Scale) &&
1694  isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1695  HasBaseReg, Scale);
1696 }
1697 
1699  int64_t MinOffset, int64_t MaxOffset,
1700  LSRUse::KindType Kind, MemAccessTy AccessTy,
1701  const Formula &F, const Loop &L) {
1702  // For the purpose of isAMCompletelyFolded either having a canonical formula
1703  // or a scale not equal to zero is correct.
1704  // Problems may arise from non canonical formulae having a scale == 0.
1705  // Strictly speaking it would best to just rely on canonical formulae.
1706  // However, when we generate the scaled formulae, we first check that the
1707  // scaling factor is profitable before computing the actual ScaledReg for
1708  // compile time sake.
1709  assert((F.isCanonical(L) || F.Scale != 0));
1710  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1711  F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1712 }
1713 
1714 /// Test whether we know how to expand the current formula.
1715 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1716  int64_t MaxOffset, LSRUse::KindType Kind,
1717  MemAccessTy AccessTy, GlobalValue *BaseGV,
1718  int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1719  // We know how to expand completely foldable formulae.
1720  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1721  BaseOffset, HasBaseReg, Scale) ||
1722  // Or formulae that use a base register produced by a sum of base
1723  // registers.
1724  (Scale == 1 &&
1725  isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1726  BaseGV, BaseOffset, true, 0));
1727 }
1728 
1729 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1730  int64_t MaxOffset, LSRUse::KindType Kind,
1731  MemAccessTy AccessTy, const Formula &F) {
1732  return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1733  F.BaseOffset, F.HasBaseReg, F.Scale);
1734 }
1735 
1737  const LSRUse &LU, const Formula &F) {
1738  // Target may want to look at the user instructions.
1739  if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1740  for (const LSRFixup &Fixup : LU.Fixups)
1741  if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1742  (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1743  F.Scale, Fixup.UserInst))
1744  return false;
1745  return true;
1746  }
1747 
1748  return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1749  LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1750  F.Scale);
1751 }
1752 
1753 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1754  const LSRUse &LU, const Formula &F,
1755  const Loop &L) {
1756  if (!F.Scale)
1757  return 0;
1758 
1759  // If the use is not completely folded in that instruction, we will have to
1760  // pay an extra cost only for scale != 1.
1761  if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1762  LU.AccessTy, F, L))
1763  return F.Scale != 1;
1764 
1765  switch (LU.Kind) {
1766  case LSRUse::Address: {
1767  // Check the scaling factor cost with both the min and max offsets.
1768  int ScaleCostMinOffset = TTI.getScalingFactorCost(
1769  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1770  F.Scale, LU.AccessTy.AddrSpace);
1771  int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1772  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1773  F.Scale, LU.AccessTy.AddrSpace);
1774 
1775  assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1776  "Legal addressing mode has an illegal cost!");
1777  return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1778  }
1779  case LSRUse::ICmpZero:
1780  case LSRUse::Basic:
1781  case LSRUse::Special:
1782  // The use is completely folded, i.e., everything is folded into the
1783  // instruction.
1784  return 0;
1785  }
1786 
1787  llvm_unreachable("Invalid LSRUse Kind!");
1788 }
1789 
1790 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1791  LSRUse::KindType Kind, MemAccessTy AccessTy,
1792  GlobalValue *BaseGV, int64_t BaseOffset,
1793  bool HasBaseReg) {
1794  // Fast-path: zero is always foldable.
1795  if (BaseOffset == 0 && !BaseGV) return true;
1796 
1797  // Conservatively, create an address with an immediate and a
1798  // base and a scale.
1799  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1800 
1801  // Canonicalize a scale of 1 to a base register if the formula doesn't
1802  // already have a base register.
1803  if (!HasBaseReg && Scale == 1) {
1804  Scale = 0;
1805  HasBaseReg = true;
1806  }
1807 
1808  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1809  HasBaseReg, Scale);
1810 }
1811 
1812 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1813  ScalarEvolution &SE, int64_t MinOffset,
1814  int64_t MaxOffset, LSRUse::KindType Kind,
1815  MemAccessTy AccessTy, const SCEV *S,
1816  bool HasBaseReg) {
1817  // Fast-path: zero is always foldable.
1818  if (S->isZero()) return true;
1819 
1820  // Conservatively, create an address with an immediate and a
1821  // base and a scale.
1822  int64_t BaseOffset = ExtractImmediate(S, SE);
1823  GlobalValue *BaseGV = ExtractSymbol(S, SE);
1824 
1825  // If there's anything else involved, it's not foldable.
1826  if (!S->isZero()) return false;
1827 
1828  // Fast-path: zero is always foldable.
1829  if (BaseOffset == 0 && !BaseGV) return true;
1830 
1831  // Conservatively, create an address with an immediate and a
1832  // base and a scale.
1833  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1834 
1835  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1836  BaseOffset, HasBaseReg, Scale);
1837 }
1838 
1839 namespace {
1840 
1841 /// An individual increment in a Chain of IV increments. Relate an IV user to
1842 /// an expression that computes the IV it uses from the IV used by the previous
1843 /// link in the Chain.
1844 ///
1845 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1846 /// original IVOperand. The head of the chain's IVOperand is only valid during
1847 /// chain collection, before LSR replaces IV users. During chain generation,
1848 /// IncExpr can be used to find the new IVOperand that computes the same
1849 /// expression.
1850 struct IVInc {
1851  Instruction *UserInst;
1852  Value* IVOperand;
1853  const SCEV *IncExpr;
1854 
1855  IVInc(Instruction *U, Value *O, const SCEV *E)
1856  : UserInst(U), IVOperand(O), IncExpr(E) {}
1857 };
1858 
1859 // The list of IV increments in program order. We typically add the head of a
1860 // chain without finding subsequent links.
1861 struct IVChain {
1862  SmallVector<IVInc, 1> Incs;
1863  const SCEV *ExprBase = nullptr;
1864 
1865  IVChain() = default;
1866  IVChain(const IVInc &Head, const SCEV *Base)
1867  : Incs(1, Head), ExprBase(Base) {}
1868 
1869  using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1870 
1871  // Return the first increment in the chain.
1872  const_iterator begin() const {
1873  assert(!Incs.empty());
1874  return std::next(Incs.begin());
1875  }
1876  const_iterator end() const {
1877  return Incs.end();
1878  }
1879 
1880  // Returns true if this chain contains any increments.
1881  bool hasIncs() const { return Incs.size() >= 2; }
1882 
1883  // Add an IVInc to the end of this chain.
1884  void add(const IVInc &X) { Incs.push_back(X); }
1885 
1886  // Returns the last UserInst in the chain.
1887  Instruction *tailUserInst() const { return Incs.back().UserInst; }
1888 
1889  // Returns true if IncExpr can be profitably added to this chain.
1890  bool isProfitableIncrement(const SCEV *OperExpr,
1891  const SCEV *IncExpr,
1892  ScalarEvolution&);
1893 };
1894 
1895 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1896 /// between FarUsers that definitely cross IV increments and NearUsers that may
1897 /// be used between IV increments.
1898 struct ChainUsers {
1900  SmallPtrSet<Instruction*, 4> NearUsers;
1901 };
1902 
1903 /// This class holds state for the main loop strength reduction logic.
1904 class LSRInstance {
1905  IVUsers &IU;
1906  ScalarEvolution &SE;
1907  DominatorTree &DT;
1908  LoopInfo &LI;
1909  const TargetTransformInfo &TTI;
1910  Loop *const L;
1911  bool FavorBackedgeIndex = false;
1912  bool Changed = false;
1913 
1914  /// This is the insert position that the current loop's induction variable
1915  /// increment should be placed. In simple loops, this is the latch block's
1916  /// terminator. But in more complicated cases, this is a position which will
1917  /// dominate all the in-loop post-increment users.
1918  Instruction *IVIncInsertPos = nullptr;
1919 
1920  /// Interesting factors between use strides.
1921  ///
1922  /// We explicitly use a SetVector which contains a SmallSet, instead of the
1923  /// default, a SmallDenseSet, because we need to use the full range of
1924  /// int64_ts, and there's currently no good way of doing that with
1925  /// SmallDenseSet.
1927 
1928  /// Interesting use types, to facilitate truncation reuse.
1930 
1931  /// The list of interesting uses.
1932  mutable SmallVector<LSRUse, 16> Uses;
1933 
1934  /// Track which uses use which register candidates.
1935  RegUseTracker RegUses;
1936 
1937  // Limit the number of chains to avoid quadratic behavior. We don't expect to
1938  // have more than a few IV increment chains in a loop. Missing a Chain falls
1939  // back to normal LSR behavior for those uses.
1940  static const unsigned MaxChains = 8;
1941 
1942  /// IV users can form a chain of IV increments.
1944 
1945  /// IV users that belong to profitable IVChains.
1947 
1948  void OptimizeShadowIV();
1949  bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1950  ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1951  void OptimizeLoopTermCond();
1952 
1953  void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1954  SmallVectorImpl<ChainUsers> &ChainUsersVec);
1955  void FinalizeChain(IVChain &Chain);
1956  void CollectChains();
1957  void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1958  SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1959 
1960  void CollectInterestingTypesAndFactors();
1961  void CollectFixupsAndInitialFormulae();
1962 
1963  // Support for sharing of LSRUses between LSRFixups.
1965  UseMapTy UseMap;
1966 
1967  bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1968  LSRUse::KindType Kind, MemAccessTy AccessTy);
1969 
1970  std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1971  MemAccessTy AccessTy);
1972 
1973  void DeleteUse(LSRUse &LU, size_t LUIdx);
1974 
1975  LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1976 
1977  void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1978  void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1979  void CountRegisters(const Formula &F, size_t LUIdx);
1980  bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1981 
1982  void CollectLoopInvariantFixupsAndFormulae();
1983 
1984  void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1985  unsigned Depth = 0);
1986 
1987  void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1988  const Formula &Base, unsigned Depth,
1989  size_t Idx, bool IsScaledReg = false);
1990  void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1991  void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1992  const Formula &Base, size_t Idx,
1993  bool IsScaledReg = false);
1994  void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1995  void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1996  const Formula &Base,
1997  const SmallVectorImpl<int64_t> &Worklist,
1998  size_t Idx, bool IsScaledReg = false);
1999  void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2000  void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2001  void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2002  void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2003  void GenerateCrossUseConstantOffsets();
2004  void GenerateAllReuseFormulae();
2005 
2006  void FilterOutUndesirableDedicatedRegisters();
2007 
2008  size_t EstimateSearchSpaceComplexity() const;
2009  void NarrowSearchSpaceByDetectingSupersets();
2010  void NarrowSearchSpaceByCollapsingUnrolledCode();
2011  void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2012  void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2013  void NarrowSearchSpaceByDeletingCostlyFormulas();
2014  void NarrowSearchSpaceByPickingWinnerRegs();
2015  void NarrowSearchSpaceUsingHeuristics();
2016 
2017  void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2018  Cost &SolutionCost,
2020  const Cost &CurCost,
2021  const SmallPtrSet<const SCEV *, 16> &CurRegs,
2022  DenseSet<const SCEV *> &VisitedRegs) const;
2023  void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2024 
2026  HoistInsertPosition(BasicBlock::iterator IP,
2027  const SmallVectorImpl<Instruction *> &Inputs) const;
2029  AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2030  const LSRFixup &LF,
2031  const LSRUse &LU,
2032  SCEVExpander &Rewriter) const;
2033 
2034  Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2035  BasicBlock::iterator IP, SCEVExpander &Rewriter,
2036  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2037  void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2038  const Formula &F, SCEVExpander &Rewriter,
2039  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2040  void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2041  SCEVExpander &Rewriter,
2042  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2043  void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2044 
2045 public:
2046  LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2047  LoopInfo &LI, const TargetTransformInfo &TTI);
2048 
2049  bool getChanged() const { return Changed; }
2050 
2051  void print_factors_and_types(raw_ostream &OS) const;
2052  void print_fixups(raw_ostream &OS) const;
2053  void print_uses(raw_ostream &OS) const;
2054  void print(raw_ostream &OS) const;
2055  void dump() const;
2056 };
2057 
2058 } // end anonymous namespace
2059 
2060 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
2061 /// the cast operation.
2062 void LSRInstance::OptimizeShadowIV() {
2063  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2064  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2065  return;
2066 
2067  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2068  UI != E; /* empty */) {
2069  IVUsers::const_iterator CandidateUI = UI;
2070  ++UI;
2071  Instruction *ShadowUse = CandidateUI->getUser();
2072  Type *DestTy = nullptr;
2073  bool IsSigned = false;
2074 
2075  /* If shadow use is a int->float cast then insert a second IV
2076  to eliminate this cast.
2077 
2078  for (unsigned i = 0; i < n; ++i)
2079  foo((double)i);
2080 
2081  is transformed into
2082 
2083  double d = 0.0;
2084  for (unsigned i = 0; i < n; ++i, ++d)
2085  foo(d);
2086  */
2087  if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2088  IsSigned = false;
2089  DestTy = UCast->getDestTy();
2090  }
2091  else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2092  IsSigned = true;
2093  DestTy = SCast->getDestTy();
2094  }
2095  if (!DestTy) continue;
2096 
2097  // If target does not support DestTy natively then do not apply
2098  // this transformation.
2099  if (!TTI.isTypeLegal(DestTy)) continue;
2100 
2101  PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2102  if (!PH) continue;
2103  if (PH->getNumIncomingValues() != 2) continue;
2104 
2105  // If the calculation in integers overflows, the result in FP type will
2106  // differ. So we only can do this transformation if we are guaranteed to not
2107  // deal with overflowing values
2108  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2109  if (!AR) continue;
2110  if (IsSigned && !AR->hasNoSignedWrap()) continue;
2111  if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2112 
2113  Type *SrcTy = PH->getType();
2114  int Mantissa = DestTy->getFPMantissaWidth();
2115  if (Mantissa == -1) continue;
2116  if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2117  continue;
2118 
2119  unsigned Entry, Latch;
2120  if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2121  Entry = 0;
2122  Latch = 1;
2123  } else {
2124  Entry = 1;
2125  Latch = 0;
2126  }
2127 
2129  if (!Init) continue;
2130  Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2131  (double)Init->getSExtValue() :
2132  (double)Init->getZExtValue());
2133 
2134  BinaryOperator *Incr =
2136  if (!Incr) continue;
2137  if (Incr->getOpcode() != Instruction::Add
2138  && Incr->getOpcode() != Instruction::Sub)
2139  continue;
2140 
2141  /* Initialize new IV, double d = 0.0 in above example. */
2142  ConstantInt *C = nullptr;
2143  if (Incr->getOperand(0) == PH)
2144  C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2145  else if (Incr->getOperand(1) == PH)
2146  C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2147  else
2148  continue;
2149 
2150  if (!C) continue;
2151 
2152  // Ignore negative constants, as the code below doesn't handle them
2153  // correctly. TODO: Remove this restriction.
2154  if (!C->getValue().isStrictlyPositive()) continue;
2155 
2156  /* Add new PHINode. */
2157  PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2158 
2159  /* create new increment. '++d' in above example. */
2160  Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2161  BinaryOperator *NewIncr =
2162  BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2163  Instruction::FAdd : Instruction::FSub,
2164  NewPH, CFP, "IV.S.next.", Incr);
2165 
2166  NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2167  NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2168 
2169  /* Remove cast operation */
2170  ShadowUse->replaceAllUsesWith(NewPH);
2171  ShadowUse->eraseFromParent();
2172  Changed = true;
2173  break;
2174  }
2175 }
2176 
2177 /// If Cond has an operand that is an expression of an IV, set the IV user and
2178 /// stride information and return true, otherwise return false.
2179 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2180  for (IVStrideUse &U : IU)
2181  if (U.getUser() == Cond) {
2182  // NOTE: we could handle setcc instructions with multiple uses here, but
2183  // InstCombine does it as well for simple uses, it's not clear that it
2184  // occurs enough in real life to handle.
2185  CondUse = &U;
2186  return true;
2187  }
2188  return false;
2189 }
2190 
2191 /// Rewrite the loop's terminating condition if it uses a max computation.
2192 ///
2193 /// This is a narrow solution to a specific, but acute, problem. For loops
2194 /// like this:
2195 ///
2196 /// i = 0;
2197 /// do {
2198 /// p[i] = 0.0;
2199 /// } while (++i < n);
2200 ///
2201 /// the trip count isn't just 'n', because 'n' might not be positive. And
2202 /// unfortunately this can come up even for loops where the user didn't use
2203 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2204 /// will commonly be lowered like this:
2205 ///
2206 /// if (n > 0) {
2207 /// i = 0;
2208 /// do {
2209 /// p[i] = 0.0;
2210 /// } while (++i < n);
2211 /// }
2212 ///
2213 /// and then it's possible for subsequent optimization to obscure the if
2214 /// test in such a way that indvars can't find it.
2215 ///
2216 /// When indvars can't find the if test in loops like this, it creates a
2217 /// max expression, which allows it to give the loop a canonical
2218 /// induction variable:
2219 ///
2220 /// i = 0;
2221 /// max = n < 1 ? 1 : n;
2222 /// do {
2223 /// p[i] = 0.0;
2224 /// } while (++i != max);
2225 ///
2226 /// Canonical induction variables are necessary because the loop passes
2227 /// are designed around them. The most obvious example of this is the
2228 /// LoopInfo analysis, which doesn't remember trip count values. It
2229 /// expects to be able to rediscover the trip count each time it is
2230 /// needed, and it does this using a simple analysis that only succeeds if
2231 /// the loop has a canonical induction variable.
2232 ///
2233 /// However, when it comes time to generate code, the maximum operation
2234 /// can be quite costly, especially if it's inside of an outer loop.
2235 ///
2236 /// This function solves this problem by detecting this type of loop and
2237 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2238 /// the instructions for the maximum computation.
2239 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2240  // Check that the loop matches the pattern we're looking for.
2241  if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2242  Cond->getPredicate() != CmpInst::ICMP_NE)
2243  return Cond;
2244 
2245  SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2246  if (!Sel || !Sel->hasOneUse()) return Cond;
2247 
2248  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2249  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2250  return Cond;
2251  const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2252 
2253  // Add one to the backedge-taken count to get the trip count.
2254  const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2255  if (IterationCount != SE.getSCEV(Sel)) return Cond;
2256 
2257  // Check for a max calculation that matches the pattern. There's no check
2258  // for ICMP_ULE here because the comparison would be with zero, which
2259  // isn't interesting.
2261  const SCEVNAryExpr *Max = nullptr;
2262  if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2263  Pred = ICmpInst::ICMP_SLE;
2264  Max = S;
2265  } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2266  Pred = ICmpInst::ICMP_SLT;
2267  Max = S;
2268  } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2269  Pred = ICmpInst::ICMP_ULT;
2270  Max = U;
2271  } else {
2272  // No match; bail.
2273  return Cond;
2274  }
2275 
2276  // To handle a max with more than two operands, this optimization would
2277  // require additional checking and setup.
2278  if (Max->getNumOperands() != 2)
2279  return Cond;
2280 
2281  const SCEV *MaxLHS = Max->getOperand(0);
2282  const SCEV *MaxRHS = Max->getOperand(1);
2283 
2284  // ScalarEvolution canonicalizes constants to the left. For < and >, look
2285  // for a comparison with 1. For <= and >=, a comparison with zero.
2286  if (!MaxLHS ||
2287  (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2288  return Cond;
2289 
2290  // Check the relevant induction variable for conformance to
2291  // the pattern.
2292  const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2293  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2294  if (!AR || !AR->isAffine() ||
2295  AR->getStart() != One ||
2296  AR->getStepRecurrence(SE) != One)
2297  return Cond;
2298 
2299  assert(AR->getLoop() == L &&
2300  "Loop condition operand is an addrec in a different loop!");
2301 
2302  // Check the right operand of the select, and remember it, as it will
2303  // be used in the new comparison instruction.
2304  Value *NewRHS = nullptr;
2305  if (ICmpInst::isTrueWhenEqual(Pred)) {
2306  // Look for n+1, and grab n.
2307  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2308  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2309  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2310  NewRHS = BO->getOperand(0);
2311  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2312  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2313  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2314  NewRHS = BO->getOperand(0);
2315  if (!NewRHS)
2316  return Cond;
2317  } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2318  NewRHS = Sel->getOperand(1);
2319  else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2320  NewRHS = Sel->getOperand(2);
2321  else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2322  NewRHS = SU->getValue();
2323  else
2324  // Max doesn't match expected pattern.
2325  return Cond;
2326 
2327  // Determine the new comparison opcode. It may be signed or unsigned,
2328  // and the original comparison may be either equality or inequality.
2329  if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2330  Pred = CmpInst::getInversePredicate(Pred);
2331 
2332  // Ok, everything looks ok to change the condition into an SLT or SGE and
2333  // delete the max calculation.
2334  ICmpInst *NewCond =
2335  new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2336 
2337  // Delete the max calculation instructions.
2338  Cond->replaceAllUsesWith(NewCond);
2339  CondUse->setUser(NewCond);
2340  Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2341  Cond->eraseFromParent();
2342  Sel->eraseFromParent();
2343  if (Cmp->use_empty())
2344  Cmp->eraseFromParent();
2345  return NewCond;
2346 }
2347 
2348 /// Change loop terminating condition to use the postinc iv when possible.
2349 void
2350 LSRInstance::OptimizeLoopTermCond() {
2352 
2353  // We need a different set of heuristics for rotated and non-rotated loops.
2354  // If a loop is rotated then the latch is also the backedge, so inserting
2355  // post-inc expressions just before the latch is ideal. To reduce live ranges
2356  // it also makes sense to rewrite terminating conditions to use post-inc
2357  // expressions.
2358  //
2359  // If the loop is not rotated then the latch is not a backedge; the latch
2360  // check is done in the loop head. Adding post-inc expressions before the
2361  // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2362  // in the loop body. In this case we do *not* want to use post-inc expressions
2363  // in the latch check, and we want to insert post-inc expressions before
2364  // the backedge.
2365  BasicBlock *LatchBlock = L->getLoopLatch();
2366  SmallVector<BasicBlock*, 8> ExitingBlocks;
2367  L->getExitingBlocks(ExitingBlocks);
2368  if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2369  return LatchBlock != BB;
2370  })) {
2371  // The backedge doesn't exit the loop; treat this as a head-tested loop.
2372  IVIncInsertPos = LatchBlock->getTerminator();
2373  return;
2374  }
2375 
2376  // Otherwise treat this as a rotated loop.
2377  for (BasicBlock *ExitingBlock : ExitingBlocks) {
2378  // Get the terminating condition for the loop if possible. If we
2379  // can, we want to change it to use a post-incremented version of its
2380  // induction variable, to allow coalescing the live ranges for the IV into
2381  // one register value.
2382 
2383  BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2384  if (!TermBr)
2385  continue;
2386  // FIXME: Overly conservative, termination condition could be an 'or' etc..
2387  if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2388  continue;
2389 
2390  // Search IVUsesByStride to find Cond's IVUse if there is one.
2391  IVStrideUse *CondUse = nullptr;
2392  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2393  if (!FindIVUserForCond(Cond, CondUse))
2394  continue;
2395 
2396  // If the trip count is computed in terms of a max (due to ScalarEvolution
2397  // being unable to find a sufficient guard, for example), change the loop
2398  // comparison to use SLT or ULT instead of NE.
2399  // One consequence of doing this now is that it disrupts the count-down
2400  // optimization. That's not always a bad thing though, because in such
2401  // cases it may still be worthwhile to avoid a max.
2402  Cond = OptimizeMax(Cond, CondUse);
2403 
2404  // If this exiting block dominates the latch block, it may also use
2405  // the post-inc value if it won't be shared with other uses.
2406  // Check for dominance.
2407  if (!DT.dominates(ExitingBlock, LatchBlock))
2408  continue;
2409 
2410  // Conservatively avoid trying to use the post-inc value in non-latch
2411  // exits if there may be pre-inc users in intervening blocks.
2412  if (LatchBlock != ExitingBlock)
2413  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2414  // Test if the use is reachable from the exiting block. This dominator
2415  // query is a conservative approximation of reachability.
2416  if (&*UI != CondUse &&
2417  !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2418  // Conservatively assume there may be reuse if the quotient of their
2419  // strides could be a legal scale.
2420  const SCEV *A = IU.getStride(*CondUse, L);
2421  const SCEV *B = IU.getStride(*UI, L);
2422  if (!A || !B) continue;
2423  if (SE.getTypeSizeInBits(A->getType()) !=
2424  SE.getTypeSizeInBits(B->getType())) {
2425  if (SE.getTypeSizeInBits(A->getType()) >
2426  SE.getTypeSizeInBits(B->getType()))
2427  B = SE.getSignExtendExpr(B, A->getType());
2428  else
2429  A = SE.getSignExtendExpr(A, B->getType());
2430  }
2431  if (const SCEVConstant *D =
2432  dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2433  const ConstantInt *C = D->getValue();
2434  // Stride of one or negative one can have reuse with non-addresses.
2435  if (C->isOne() || C->isMinusOne())
2436  goto decline_post_inc;
2437  // Avoid weird situations.
2438  if (C->getValue().getMinSignedBits() >= 64 ||
2439  C->getValue().isMinSignedValue())
2440  goto decline_post_inc;
2441  // Check for possible scaled-address reuse.
2442  if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2443  MemAccessTy AccessTy = getAccessType(
2444  TTI, UI->getUser(), UI->getOperandValToReplace());
2445  int64_t Scale = C->getSExtValue();
2446  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2447  /*BaseOffset=*/0,
2448  /*HasBaseReg=*/false, Scale,
2449  AccessTy.AddrSpace))
2450  goto decline_post_inc;
2451  Scale = -Scale;
2452  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2453  /*BaseOffset=*/0,
2454  /*HasBaseReg=*/false, Scale,
2455  AccessTy.AddrSpace))
2456  goto decline_post_inc;
2457  }
2458  }
2459  }
2460 
2461  LLVM_DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2462  << *Cond << '\n');
2463 
2464  // It's possible for the setcc instruction to be anywhere in the loop, and
2465  // possible for it to have multiple users. If it is not immediately before
2466  // the exiting block branch, move it.
2467  if (&*++BasicBlock::iterator(Cond) != TermBr) {
2468  if (Cond->hasOneUse()) {
2469  Cond->moveBefore(TermBr);
2470  } else {
2471  // Clone the terminating condition and insert into the loopend.
2472  ICmpInst *OldCond = Cond;
2473  Cond = cast<ICmpInst>(Cond->clone());
2474  Cond->setName(L->getHeader()->getName() + ".termcond");
2475  ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2476 
2477  // Clone the IVUse, as the old use still exists!
2478  CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2479  TermBr->replaceUsesOfWith(OldCond, Cond);
2480  }
2481  }
2482 
2483  // If we get to here, we know that we can transform the setcc instruction to
2484  // use the post-incremented version of the IV, allowing us to coalesce the
2485  // live ranges for the IV correctly.
2486  CondUse->transformToPostInc(L);
2487  Changed = true;
2488 
2489  PostIncs.insert(Cond);
2490  decline_post_inc:;
2491  }
2492 
2493  // Determine an insertion point for the loop induction variable increment. It
2494  // must dominate all the post-inc comparisons we just set up, and it must
2495  // dominate the loop latch edge.
2496  IVIncInsertPos = L->getLoopLatch()->getTerminator();
2497  for (Instruction *Inst : PostIncs) {
2498  BasicBlock *BB =
2499  DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2500  Inst->getParent());
2501  if (BB == Inst->getParent())
2502  IVIncInsertPos = Inst;
2503  else if (BB != IVIncInsertPos->getParent())
2504  IVIncInsertPos = BB->getTerminator();
2505  }
2506 }
2507 
2508 /// Determine if the given use can accommodate a fixup at the given offset and
2509 /// other details. If so, update the use and return true.
2510 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2511  bool HasBaseReg, LSRUse::KindType Kind,
2512  MemAccessTy AccessTy) {
2513  int64_t NewMinOffset = LU.MinOffset;
2514  int64_t NewMaxOffset = LU.MaxOffset;
2515  MemAccessTy NewAccessTy = AccessTy;
2516 
2517  // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2518  // something conservative, however this can pessimize in the case that one of
2519  // the uses will have all its uses outside the loop, for example.
2520  if (LU.Kind != Kind)
2521  return false;
2522 
2523  // Check for a mismatched access type, and fall back conservatively as needed.
2524  // TODO: Be less conservative when the type is similar and can use the same
2525  // addressing modes.
2526  if (Kind == LSRUse::Address) {
2527  if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2528  NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2529  AccessTy.AddrSpace);
2530  }
2531  }
2532 
2533  // Conservatively assume HasBaseReg is true for now.
2534  if (NewOffset < LU.MinOffset) {
2535  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2536  LU.MaxOffset - NewOffset, HasBaseReg))
2537  return false;
2538  NewMinOffset = NewOffset;
2539  } else if (NewOffset > LU.MaxOffset) {
2540  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2541  NewOffset - LU.MinOffset, HasBaseReg))
2542  return false;
2543  NewMaxOffset = NewOffset;
2544  }
2545 
2546  // Update the use.
2547  LU.MinOffset = NewMinOffset;
2548  LU.MaxOffset = NewMaxOffset;
2549  LU.AccessTy = NewAccessTy;
2550  return true;
2551 }
2552 
2553 /// Return an LSRUse index and an offset value for a fixup which needs the given
2554 /// expression, with the given kind and optional access type. Either reuse an
2555 /// existing use or create a new one, as needed.
2556 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2557  LSRUse::KindType Kind,
2558  MemAccessTy AccessTy) {
2559  const SCEV *Copy = Expr;
2560  int64_t Offset = ExtractImmediate(Expr, SE);
2561 
2562  // Basic uses can't accept any offset, for example.
2563  if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2564  Offset, /*HasBaseReg=*/ true)) {
2565  Expr = Copy;
2566  Offset = 0;
2567  }
2568 
2569  std::pair<UseMapTy::iterator, bool> P =
2570  UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2571  if (!P.second) {
2572  // A use already existed with this base.
2573  size_t LUIdx = P.first->second;
2574  LSRUse &LU = Uses[LUIdx];
2575  if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2576  // Reuse this use.
2577  return std::make_pair(LUIdx, Offset);
2578  }
2579 
2580  // Create a new use.
2581  size_t LUIdx = Uses.size();
2582  P.first->second = LUIdx;
2583  Uses.push_back(LSRUse(Kind, AccessTy));
2584  LSRUse &LU = Uses[LUIdx];
2585 
2586  LU.MinOffset = Offset;
2587  LU.MaxOffset = Offset;
2588  return std::make_pair(LUIdx, Offset);
2589 }
2590 
2591 /// Delete the given use from the Uses list.
2592 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2593  if (&LU != &Uses.back())
2594  std::swap(LU, Uses.back());
2595  Uses.pop_back();
2596 
2597  // Update RegUses.
2598  RegUses.swapAndDropUse(LUIdx, Uses.size());
2599 }
2600 
2601 /// Look for a use distinct from OrigLU which is has a formula that has the same
2602 /// registers as the given formula.
2603 LSRUse *
2604 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2605  const LSRUse &OrigLU) {
2606  // Search all uses for the formula. This could be more clever.
2607  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2608  LSRUse &LU = Uses[LUIdx];
2609  // Check whether this use is close enough to OrigLU, to see whether it's
2610  // worthwhile looking through its formulae.
2611  // Ignore ICmpZero uses because they may contain formulae generated by
2612  // GenerateICmpZeroScales, in which case adding fixup offsets may
2613  // be invalid.
2614  if (&LU != &OrigLU &&
2615  LU.Kind != LSRUse::ICmpZero &&
2616  LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2617  LU.WidestFixupType == OrigLU.WidestFixupType &&
2618  LU.HasFormulaWithSameRegs(OrigF)) {
2619  // Scan through this use's formulae.
2620  for (const Formula &F : LU.Formulae) {
2621  // Check to see if this formula has the same registers and symbols
2622  // as OrigF.
2623  if (F.BaseRegs == OrigF.BaseRegs &&
2624  F.ScaledReg == OrigF.ScaledReg &&
2625  F.BaseGV == OrigF.BaseGV &&
2626  F.Scale == OrigF.Scale &&
2627  F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2628  if (F.BaseOffset == 0)
2629  return &LU;
2630  // This is the formula where all the registers and symbols matched;
2631  // there aren't going to be any others. Since we declined it, we
2632  // can skip the rest of the formulae and proceed to the next LSRUse.
2633  break;
2634  }
2635  }
2636  }
2637  }
2638 
2639  // Nothing looked good.
2640  return nullptr;
2641 }
2642 
2643 void LSRInstance::CollectInterestingTypesAndFactors() {
2645 
2646  // Collect interesting types and strides.
2648  for (const IVStrideUse &U : IU) {
2649  const SCEV *Expr = IU.getExpr(U);
2650 
2651  // Collect interesting types.
2652  Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2653 
2654  // Add strides for mentioned loops.
2655  Worklist.push_back(Expr);
2656  do {
2657  const SCEV *S = Worklist.pop_back_val();
2658  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2659  if (AR->getLoop() == L)
2660  Strides.insert(AR->getStepRecurrence(SE));
2661  Worklist.push_back(AR->getStart());
2662  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2663  Worklist.append(Add->op_begin(), Add->op_end());
2664  }
2665  } while (!Worklist.empty());
2666  }
2667 
2668  // Compute interesting factors from the set of interesting strides.
2670  I = Strides.begin(), E = Strides.end(); I != E; ++I)
2672  std::next(I); NewStrideIter != E; ++NewStrideIter) {
2673  const SCEV *OldStride = *I;
2674  const SCEV *NewStride = *NewStrideIter;
2675 
2676  if (SE.getTypeSizeInBits(OldStride->getType()) !=
2677  SE.getTypeSizeInBits(NewStride->getType())) {
2678  if (SE.getTypeSizeInBits(OldStride->getType()) >
2679  SE.getTypeSizeInBits(NewStride->getType()))
2680  NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2681  else
2682  OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2683  }
2684  if (const SCEVConstant *Factor =
2685  dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2686  SE, true))) {
2687  if (Factor->getAPInt().getMinSignedBits() <= 64)
2688  Factors.insert(Factor->getAPInt().getSExtValue());
2689  } else if (const SCEVConstant *Factor =
2690  dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2691  NewStride,
2692  SE, true))) {
2693  if (Factor->getAPInt().getMinSignedBits() <= 64)
2694  Factors.insert(Factor->getAPInt().getSExtValue());
2695  }
2696  }
2697 
2698  // If all uses use the same type, don't bother looking for truncation-based
2699  // reuse.
2700  if (Types.size() == 1)
2701  Types.clear();
2702 
2703  LLVM_DEBUG(print_factors_and_types(dbgs()));
2704 }
2705 
2706 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2707 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2708 /// IVStrideUses, we could partially skip this.
2709 static User::op_iterator
2711  Loop *L, ScalarEvolution &SE) {
2712  for(; OI != OE; ++OI) {
2713  if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2714  if (!SE.isSCEVable(Oper->getType()))
2715  continue;
2716 
2717  if (const SCEVAddRecExpr *AR =
2718  dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2719  if (AR->getLoop() == L)
2720  break;
2721  }
2722  }
2723  }
2724  return OI;
2725 }
2726 
2727 /// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2728 /// a convenient helper.
2729 static Value *getWideOperand(Value *Oper) {
2730  if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2731  return Trunc->getOperand(0);
2732  return Oper;
2733 }
2734 
2735 /// Return true if we allow an IV chain to include both types.
2736 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2737  Type *LType = LVal->getType();
2738  Type *RType = RVal->getType();
2739  return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2740  // Different address spaces means (possibly)
2741  // different types of the pointer implementation,
2742  // e.g. i16 vs i32 so disallow that.
2743  (LType->getPointerAddressSpace() ==
2744  RType->getPointerAddressSpace()));
2745 }
2746 
2747 /// Return an approximation of this SCEV expression's "base", or NULL for any
2748 /// constant. Returning the expression itself is conservative. Returning a
2749 /// deeper subexpression is more precise and valid as long as it isn't less
2750 /// complex than another subexpression. For expressions involving multiple
2751 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2752 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2753 /// IVInc==b-a.
2754 ///
2755 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2756 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2757 static const SCEV *getExprBase(const SCEV *S) {
2758  switch (S->getSCEVType()) {
2759  default: // uncluding scUnknown.
2760  return S;
2761  case scConstant:
2762  return nullptr;
2763  case scTruncate:
2764  return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2765  case scZeroExtend:
2766  return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2767  case scSignExtend:
2768  return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2769  case scAddExpr: {
2770  // Skip over scaled operands (scMulExpr) to follow add operands as long as
2771  // there's nothing more complex.
2772  // FIXME: not sure if we want to recognize negation.
2773  const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2774  for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2775  E(Add->op_begin()); I != E; ++I) {
2776  const SCEV *SubExpr = *I;
2777  if (SubExpr->getSCEVType() == scAddExpr)
2778  return getExprBase(SubExpr);
2779 
2780  if (SubExpr->getSCEVType() != scMulExpr)
2781  return SubExpr;
2782  }
2783  return S; // all operands are scaled, be conservative.
2784  }
2785  case scAddRecExpr:
2786  return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2787  }
2788 }
2789 
2790 /// Return true if the chain increment is profitable to expand into a loop
2791 /// invariant value, which may require its own register. A profitable chain
2792 /// increment will be an offset relative to the same base. We allow such offsets
2793 /// to potentially be used as chain increment as long as it's not obviously
2794 /// expensive to expand using real instructions.
2795 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2796  const SCEV *IncExpr,
2797  ScalarEvolution &SE) {
2798  // Aggressively form chains when -stress-ivchain.
2799  if (StressIVChain)
2800  return true;
2801 
2802  // Do not replace a constant offset from IV head with a nonconstant IV
2803  // increment.
2804  if (!isa<SCEVConstant>(IncExpr)) {
2805  const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2806  if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2807  return false;
2808  }
2809 
2810  SmallPtrSet<const SCEV*, 8> Processed;
2811  return !isHighCostExpansion(IncExpr, Processed, SE);
2812 }
2813 
2814 /// Return true if the number of registers needed for the chain is estimated to
2815 /// be less than the number required for the individual IV users. First prohibit
2816 /// any IV users that keep the IV live across increments (the Users set should
2817 /// be empty). Next count the number and type of increments in the chain.
2818 ///
2819 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2820 /// effectively use postinc addressing modes. Only consider it profitable it the
2821 /// increments can be computed in fewer registers when chained.
2822 ///
2823 /// TODO: Consider IVInc free if it's already used in another chains.
2824 static bool
2826  ScalarEvolution &SE) {
2827  if (StressIVChain)
2828  return true;
2829 
2830  if (!Chain.hasIncs())
2831  return false;
2832 
2833  if (!Users.empty()) {
2834  LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2835  for (Instruction *Inst
2836  : Users) { dbgs() << " " << *Inst << "\n"; });
2837  return false;
2838  }
2839  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2840 
2841  // The chain itself may require a register, so intialize cost to 1.
2842  int cost = 1;
2843 
2844  // A complete chain likely eliminates the need for keeping the original IV in
2845  // a register. LSR does not currently know how to form a complete chain unless
2846  // the header phi already exists.
2847  if (isa<PHINode>(Chain.tailUserInst())
2848  && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2849  --cost;
2850  }
2851  const SCEV *LastIncExpr = nullptr;
2852  unsigned NumConstIncrements = 0;
2853  unsigned NumVarIncrements = 0;
2854  unsigned NumReusedIncrements = 0;
2855  for (const IVInc &Inc : Chain) {
2856  if (Inc.IncExpr->isZero())
2857  continue;
2858 
2859  // Incrementing by zero or some constant is neutral. We assume constants can
2860  // be folded into an addressing mode or an add's immediate operand.
2861  if (isa<SCEVConstant>(Inc.IncExpr)) {
2862  ++NumConstIncrements;
2863  continue;
2864  }
2865 
2866  if (Inc.IncExpr == LastIncExpr)
2867  ++NumReusedIncrements;
2868  else
2869  ++NumVarIncrements;
2870 
2871  LastIncExpr = Inc.IncExpr;
2872  }
2873  // An IV chain with a single increment is handled by LSR's postinc
2874  // uses. However, a chain with multiple increments requires keeping the IV's
2875  // value live longer than it needs to be if chained.
2876  if (NumConstIncrements > 1)
2877  --cost;
2878 
2879  // Materializing increment expressions in the preheader that didn't exist in
2880  // the original code may cost a register. For example, sign-extended array
2881  // indices can produce ridiculous increments like this:
2882  // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2883  cost += NumVarIncrements;
2884 
2885  // Reusing variable increments likely saves a register to hold the multiple of
2886  // the stride.
2887  cost -= NumReusedIncrements;
2888 
2889  LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2890  << "\n");
2891 
2892  return cost < 0;
2893 }
2894 
2895 /// Add this IV user to an existing chain or make it the head of a new chain.
2896 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2897  SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2898  // When IVs are used as types of varying widths, they are generally converted
2899  // to a wider type with some uses remaining narrow under a (free) trunc.
2900  Value *const NextIV = getWideOperand(IVOper);
2901  const SCEV *const OperExpr = SE.getSCEV(NextIV);
2902  const SCEV *const OperExprBase = getExprBase(OperExpr);
2903 
2904  // Visit all existing chains. Check if its IVOper can be computed as a
2905  // profitable loop invariant increment from the last link in the Chain.
2906  unsigned ChainIdx = 0, NChains = IVChainVec.size();
2907  const SCEV *LastIncExpr = nullptr;
2908  for (; ChainIdx < NChains; ++ChainIdx) {
2909  IVChain &Chain = IVChainVec[ChainIdx];
2910 
2911  // Prune the solution space aggressively by checking that both IV operands
2912  // are expressions that operate on the same unscaled SCEVUnknown. This
2913  // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2914  // first avoids creating extra SCEV expressions.
2915  if (!StressIVChain && Chain.ExprBase != OperExprBase)
2916  continue;
2917 
2918  Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2919  if (!isCompatibleIVType(PrevIV, NextIV))
2920  continue;
2921 
2922  // A phi node terminates a chain.
2923  if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2924  continue;
2925 
2926  // The increment must be loop-invariant so it can be kept in a register.
2927  const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2928  const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2929  if (!SE.isLoopInvariant(IncExpr, L))
2930  continue;
2931 
2932  if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2933  LastIncExpr = IncExpr;
2934  break;
2935  }
2936  }
2937  // If we haven't found a chain, create a new one, unless we hit the max. Don't
2938  // bother for phi nodes, because they must be last in the chain.
2939  if (ChainIdx == NChains) {
2940  if (isa<PHINode>(UserInst))
2941  return;
2942  if (NChains >= MaxChains && !StressIVChain) {
2943  LLVM_DEBUG(dbgs() << "IV Chain Limit\n");
2944  return;
2945  }
2946  LastIncExpr = OperExpr;
2947  // IVUsers may have skipped over sign/zero extensions. We don't currently
2948  // attempt to form chains involving extensions unless they can be hoisted
2949  // into this loop's AddRec.
2950  if (!isa<SCEVAddRecExpr>(LastIncExpr))
2951  return;
2952  ++NChains;
2953  IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2954  OperExprBase));
2955  ChainUsersVec.resize(NChains);
2956  LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2957  << ") IV=" << *LastIncExpr << "\n");
2958  } else {
2959  LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2960  << ") IV+" << *LastIncExpr << "\n");
2961  // Add this IV user to the end of the chain.
2962  IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2963  }
2964  IVChain &Chain = IVChainVec[ChainIdx];
2965 
2966  SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2967  // This chain's NearUsers become FarUsers.
2968  if (!LastIncExpr->isZero()) {
2969  ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2970  NearUsers.end());
2971  NearUsers.clear();
2972  }
2973 
2974  // All other uses of IVOperand become near uses of the chain.
2975  // We currently ignore intermediate values within SCEV expressions, assuming
2976  // they will eventually be used be the current chain, or can be computed
2977  // from one of the chain increments. To be more precise we could
2978  // transitively follow its user and only add leaf IV users to the set.
2979  for (User *U : IVOper->users()) {
2980  Instruction *OtherUse = dyn_cast<Instruction>(U);
2981  if (!OtherUse)
2982  continue;
2983  // Uses in the chain will no longer be uses if the chain is formed.
2984  // Include the head of the chain in this iteration (not Chain.begin()).
2985  IVChain::const_iterator IncIter = Chain.Incs.begin();
2986  IVChain::const_iterator IncEnd = Chain.Incs.end();
2987  for( ; IncIter != IncEnd; ++IncIter) {
2988  if (IncIter->UserInst == OtherUse)
2989  break;
2990  }
2991  if (IncIter != IncEnd)
2992  continue;
2993 
2994  if (SE.isSCEVable(OtherUse->getType())
2995  && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2996  && IU.isIVUserOrOperand(OtherUse)) {
2997  continue;
2998  }
2999  NearUsers.insert(OtherUse);
3000  }
3001 
3002  // Since this user is part of the chain, it's no longer considered a use
3003  // of the chain.
3004  ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3005 }
3006 
3007 /// Populate the vector of Chains.
3008 ///
3009 /// This decreases ILP at the architecture level. Targets with ample registers,
3010 /// multiple memory ports, and no register renaming probably don't want
3011 /// this. However, such targets should probably disable LSR altogether.
3012 ///
3013 /// The job of LSR is to make a reasonable choice of induction variables across
3014 /// the loop. Subsequent passes can easily "unchain" computation exposing more
3015 /// ILP *within the loop* if the target wants it.
3016 ///
3017 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
3018 /// will not reorder memory operations, it will recognize this as a chain, but
3019 /// will generate redundant IV increments. Ideally this would be corrected later
3020 /// by a smart scheduler:
3021 /// = A[i]
3022 /// = A[i+x]
3023 /// A[i] =
3024 /// A[i+x] =
3025 ///
3026 /// TODO: Walk the entire domtree within this loop, not just the path to the
3027 /// loop latch. This will discover chains on side paths, but requires
3028 /// maintaining multiple copies of the Chains state.
3029 void LSRInstance::CollectChains() {
3030  LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n");
3031  SmallVector<ChainUsers, 8> ChainUsersVec;
3032 
3033  SmallVector<BasicBlock *,8> LatchPath;
3034  BasicBlock *LoopHeader = L->getHeader();
3035  for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3036  Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3037  LatchPath.push_back(Rung->getBlock());
3038  }
3039  LatchPath.push_back(LoopHeader);
3040 
3041  // Walk the instruction stream from the loop header to the loop latch.
3042  for (BasicBlock *BB : reverse(LatchPath)) {
3043  for (Instruction &I : *BB) {
3044  // Skip instructions that weren't seen by IVUsers analysis.
3045  if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3046  continue;
3047 
3048  // Ignore users that are part of a SCEV expression. This way we only
3049  // consider leaf IV Users. This effectively rediscovers a portion of
3050  // IVUsers analysis but in program order this time.
3051  if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3052  continue;
3053 
3054  // Remove this instruction from any NearUsers set it may be in.
3055  for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3056  ChainIdx < NChains; ++ChainIdx) {
3057  ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3058  }
3059  // Search for operands that can be chained.
3060  SmallPtrSet<Instruction*, 4> UniqueOperands;
3061  User::op_iterator IVOpEnd = I.op_end();
3062  User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3063  while (IVOpIter != IVOpEnd) {
3064  Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3065  if (UniqueOperands.insert(IVOpInst).second)
3066  ChainInstruction(&I, IVOpInst, ChainUsersVec);
3067  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3068  }
3069  } // Continue walking down the instructions.
3070  } // Continue walking down the domtree.
3071  // Visit phi backedges to determine if the chain can generate the IV postinc.
3072  for (PHINode &PN : L->getHeader()->phis()) {
3073  if (!SE.isSCEVable(PN.getType()))
3074  continue;
3075 
3076  Instruction *IncV =
3077  dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3078  if (IncV)
3079  ChainInstruction(&PN, IncV, ChainUsersVec);
3080  }
3081  // Remove any unprofitable chains.
3082  unsigned ChainIdx = 0;
3083  for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3084  UsersIdx < NChains; ++UsersIdx) {
3085  if (!isProfitableChain(IVChainVec[UsersIdx],
3086  ChainUsersVec[UsersIdx].FarUsers, SE))
3087  continue;
3088  // Preserve the chain at UsesIdx.
3089  if (ChainIdx != UsersIdx)
3090  IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3091  FinalizeChain(IVChainVec[ChainIdx]);
3092  ++ChainIdx;
3093  }
3094  IVChainVec.resize(ChainIdx);
3095 }
3096 
3097 void LSRInstance::FinalizeChain(IVChain &Chain) {
3098  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3099  LLVM_DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
3100 
3101  for (const IVInc &Inc : Chain) {
3102  LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n");
3103  auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3104  assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
3105  IVIncSet.insert(UseI);
3106  }
3107 }
3108 
3109 /// Return true if the IVInc can be folded into an addressing mode.
3110 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3111  Value *Operand, const TargetTransformInfo &TTI) {
3112  const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3113  if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3114  return false;
3115 
3116  if (IncConst->getAPInt().getMinSignedBits() > 64)
3117  return false;
3118 
3119  MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3120  int64_t IncOffset = IncConst->getValue()->getSExtValue();
3121  if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3122  IncOffset, /*HaseBaseReg=*/false))
3123  return false;
3124 
3125  return true;
3126 }
3127 
3128 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3129 /// user's operand from the previous IV user's operand.
3130 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3131  SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3132  // Find the new IVOperand for the head of the chain. It may have been replaced
3133  // by LSR.
3134  const IVInc &Head = Chain.Incs[0];
3135  User::op_iterator IVOpEnd = Head.UserInst->op_end();
3136  // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3137  User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3138  IVOpEnd, L, SE);
3139  Value *IVSrc = nullptr;
3140  while (IVOpIter != IVOpEnd) {
3141  IVSrc = getWideOperand(*IVOpIter);
3142 
3143  // If this operand computes the expression that the chain needs, we may use
3144  // it. (Check this after setting IVSrc which is used below.)
3145  //
3146  // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3147  // narrow for the chain, so we can no longer use it. We do allow using a
3148  // wider phi, assuming the LSR checked for free truncation. In that case we
3149  // should already have a truncate on this operand such that
3150  // getSCEV(IVSrc) == IncExpr.
3151  if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3152  || SE.getSCEV(IVSrc) == Head.IncExpr) {
3153  break;
3154  }
3155  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3156  }
3157  if (IVOpIter == IVOpEnd) {
3158  // Gracefully give up on this chain.
3159  LLVM_DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3160  return;
3161  }
3162 
3163  LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3164  Type *IVTy = IVSrc->getType();
3165  Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3166  const SCEV *LeftOverExpr = nullptr;
3167  for (const IVInc &Inc : Chain) {
3168  Instruction *InsertPt = Inc.UserInst;
3169  if (isa<PHINode>(InsertPt))
3170  InsertPt = L->getLoopLatch()->getTerminator();
3171 
3172  // IVOper will replace the current IV User's operand. IVSrc is the IV
3173  // value currently held in a register.
3174  Value *IVOper = IVSrc;
3175  if (!Inc.IncExpr->isZero()) {
3176  // IncExpr was the result of subtraction of two narrow values, so must
3177  // be signed.
3178  const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3179  LeftOverExpr = LeftOverExpr ?
3180  SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3181  }
3182  if (LeftOverExpr && !LeftOverExpr->isZero()) {
3183  // Expand the IV increment.
3184  Rewriter.clearPostInc();
3185  Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3186  const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3187  SE.getUnknown(IncV));
3188  IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3189 
3190  // If an IV increment can't be folded, use it as the next IV value.
3191  if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3192  assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3193  IVSrc = IVOper;
3194  LeftOverExpr = nullptr;
3195  }
3196  }
3197  Type *OperTy = Inc.IVOperand->getType();
3198  if (IVTy != OperTy) {
3199  assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3200  "cannot extend a chained IV");
3201  IRBuilder<> Builder(InsertPt);
3202  IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3203  }
3204  Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3205  DeadInsts.emplace_back(Inc.IVOperand);
3206  }
3207  // If LSR created a new, wider phi, we may also replace its postinc. We only
3208  // do this if we also found a wide value for the head of the chain.
3209  if (isa<PHINode>(Chain.tailUserInst())) {
3210  for (PHINode &Phi : L->getHeader()->phis()) {
3211  if (!isCompatibleIVType(&Phi, IVSrc))
3212  continue;
3213  Instruction *PostIncV = dyn_cast<Instruction>(
3214  Phi.getIncomingValueForBlock(L->getLoopLatch()));
3215  if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3216  continue;
3217  Value *IVOper = IVSrc;
3218  Type *PostIncTy = PostIncV->getType();
3219  if (IVTy != PostIncTy) {
3220  assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3221  IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3222  Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3223  IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3224  }
3225  Phi.replaceUsesOfWith(PostIncV, IVOper);
3226  DeadInsts.emplace_back(PostIncV);
3227  }
3228  }
3229 }
3230 
3231 void LSRInstance::CollectFixupsAndInitialFormulae() {
3232  for (const IVStrideUse &U : IU) {
3233  Instruction *UserInst = U.getUser();
3234  // Skip IV users that are part of profitable IV Chains.
3235  User::op_iterator UseI =
3236  find(UserInst->operands(), U.getOperandValToReplace());
3237  assert(UseI != UserInst->op_end() && "cannot find IV operand");
3238  if (IVIncSet.count(UseI)) {
3239  LLVM_DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3240  continue;
3241  }
3242 
3243  LSRUse::KindType Kind = LSRUse::Basic;
3244  MemAccessTy AccessTy;
3245  if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3246  Kind = LSRUse::Address;
3247  AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3248  }
3249 
3250  const SCEV *S = IU.getExpr(U);
3251  PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3252 
3253  // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3254  // (N - i == 0), and this allows (N - i) to be the expression that we work
3255  // with rather than just N or i, so we can consider the register
3256  // requirements for both N and i at the same time. Limiting this code to
3257  // equality icmps is not a problem because all interesting loops use
3258  // equality icmps, thanks to IndVarSimplify.
3259  if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst))
3260  if (CI->isEquality()) {
3261  // Swap the operands if needed to put the OperandValToReplace on the
3262  // left, for consistency.
3263  Value *NV = CI->getOperand(1);
3264  if (NV == U.getOperandValToReplace()) {
3265  CI->setOperand(1, CI->getOperand(0));
3266  CI->setOperand(0, NV);
3267  NV = CI->getOperand(1);
3268  Changed = true;
3269  }
3270 
3271  // x == y --> x - y == 0
3272  const SCEV *N = SE.getSCEV(NV);
3273  if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3274  // S is normalized, so normalize N before folding it into S
3275  // to keep the result normalized.
3276  N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3277  Kind = LSRUse::ICmpZero;
3278  S = SE.getMinusSCEV(N, S);
3279  }
3280 
3281  // -1 and the negations of all interesting strides (except the negation
3282  // of -1) are now also interesting.
3283  for (size_t i = 0, e = Factors.size(); i != e; ++i)
3284  if (Factors[i] != -1)
3285  Factors.insert(-(uint64_t)Factors[i]);
3286  Factors.insert(-1);
3287  }
3288 
3289  // Get or create an LSRUse.
3290  std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3291  size_t LUIdx = P.first;
3292  int64_t Offset = P.second;
3293  LSRUse &LU = Uses[LUIdx];
3294 
3295  // Record the fixup.
3296  LSRFixup &LF = LU.getNewFixup();
3297  LF.UserInst = UserInst;
3298  LF.OperandValToReplace = U.getOperandValToReplace();
3299  LF.PostIncLoops = TmpPostIncLoops;
3300  LF.Offset = Offset;
3301  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3302 
3303  if (!LU.WidestFixupType ||
3304  SE.getTypeSizeInBits(LU.WidestFixupType) <
3305  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3306  LU.WidestFixupType = LF.OperandValToReplace->getType();
3307 
3308  // If this is the first use of this LSRUse, give it a formula.
3309  if (LU.Formulae.empty()) {
3310  InsertInitialFormula(S, LU, LUIdx);
3311  CountRegisters(LU.Formulae.back(), LUIdx);
3312  }
3313  }
3314 
3315  LLVM_DEBUG(print_fixups(dbgs()));
3316 }
3317 
3318 /// Insert a formula for the given expression into the given use, separating out
3319 /// loop-variant portions from loop-invariant and loop-computable portions.
3320 void
3321 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3322  // Mark uses whose expressions cannot be expanded.
3323  if (!isSafeToExpand(S, SE))
3324  LU.RigidFormula = true;
3325 
3326  Formula F;
3327  F.initialMatch(S, L, SE);
3328  bool Inserted = InsertFormula(LU, LUIdx, F);
3329  assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3330 }
3331 
3332 /// Insert a simple single-register formula for the given expression into the
3333 /// given use.
3334 void
3335 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3336  LSRUse &LU, size_t LUIdx) {
3337  Formula F;
3338  F.BaseRegs.push_back(S);
3339  F.HasBaseReg = true;
3340  bool Inserted = InsertFormula(LU, LUIdx, F);
3341  assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3342 }
3343 
3344 /// Note which registers are used by the given formula, updating RegUses.
3345 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3346  if (F.ScaledReg)
3347  RegUses.countRegister(F.ScaledReg, LUIdx);
3348  for (const SCEV *BaseReg : F.BaseRegs)
3349  RegUses.countRegister(BaseReg, LUIdx);
3350 }
3351 
3352 /// If the given formula has not yet been inserted, add it to the list, and
3353 /// return true. Return false otherwise.
3354 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3355  // Do not insert formula that we will not be able to expand.
3356  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3357  "Formula is illegal");
3358 
3359  if (!LU.InsertFormula(F, *L))
3360  return false;
3361 
3362  CountRegisters(F, LUIdx);
3363  return true;
3364 }
3365 
3366 /// Check for other uses of loop-invariant values which we're tracking. These
3367 /// other uses will pin these values in registers, making them less profitable
3368 /// for elimination.
3369 /// TODO: This currently misses non-constant addrec step registers.
3370 /// TODO: Should this give more weight to users inside the loop?
3371 void
3372 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3373  SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3375 
3376  while (!Worklist.empty()) {
3377  const SCEV *S = Worklist.pop_back_val();
3378 
3379  // Don't process the same SCEV twice
3380  if (!Visited.insert(S).second)
3381  continue;
3382 
3383  if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3384  Worklist.append(N->op_begin(), N->op_end());
3385  else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3386  Worklist.push_back(C->getOperand());
3387  else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3388  Worklist.push_back(D->getLHS());
3389  Worklist.push_back(D->getRHS());
3390  } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3391  const Value *V = US->getValue();
3392  if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3393  // Look for instructions defined outside the loop.
3394  if (L->contains(Inst)) continue;
3395  } else if (isa<UndefValue>(V))
3396  // Undef doesn't have a live range, so it doesn't matter.
3397  continue;
3398  for (const Use &U : V->uses()) {
3399  const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3400  // Ignore non-instructions.
3401  if (!UserInst)
3402  continue;
3403  // Ignore instructions in other functions (as can happen with
3404  // Constants).
3405  if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3406  continue;
3407  // Ignore instructions not dominated by the loop.
3408  const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3409  UserInst->getParent() :
3410  cast<PHINode>(UserInst)->getIncomingBlock(
3411  PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3412  if (!DT.dominates(L->getHeader(), UseBB))
3413  continue;
3414  // Don't bother if the instruction is in a BB which ends in an EHPad.
3415  if (UseBB->getTerminator()->isEHPad())
3416  continue;
3417  // Don't bother rewriting PHIs in catchswitch blocks.
3418  if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3419  continue;
3420  // Ignore uses which are part of other SCEV expressions, to avoid
3421  // analyzing them multiple times.
3422  if (SE.isSCEVable(UserInst->getType())) {
3423  const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3424  // If the user is a no-op, look through to its uses.
3425  if (!isa<SCEVUnknown>(UserS))
3426  continue;
3427  if (UserS == US) {
3428  Worklist.push_back(
3429  SE.getUnknown(const_cast<Instruction *>(UserInst)));
3430  continue;
3431  }
3432  }
3433  // Ignore icmp instructions which are already being analyzed.
3434  if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3435  unsigned OtherIdx = !U.getOperandNo();
3436  Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3437  if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3438  continue;
3439  }
3440 
3441  std::pair<size_t, int64_t> P = getUse(
3442  S, LSRUse::Basic, MemAccessTy());
3443  size_t LUIdx = P.first;
3444  int64_t Offset = P.second;
3445  LSRUse &LU = Uses[LUIdx];
3446  LSRFixup &LF = LU.getNewFixup();
3447  LF.UserInst = const_cast<Instruction *>(UserInst);
3448  LF.OperandValToReplace = U;
3449  LF.Offset = Offset;
3450  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3451  if (!LU.WidestFixupType ||
3452  SE.getTypeSizeInBits(LU.WidestFixupType) <
3453  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3454  LU.WidestFixupType = LF.OperandValToReplace->getType();
3455  InsertSupplementalFormula(US, LU, LUIdx);
3456  CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3457  break;
3458  }
3459  }
3460  }
3461 }
3462 
3463 /// Split S into subexpressions which can be pulled out into separate
3464 /// registers. If C is non-null, multiply each subexpression by C.
3465 ///
3466 /// Return remainder expression after factoring the subexpressions captured by
3467 /// Ops. If Ops is complete, return NULL.
3468 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3470  const Loop *L,
3471  ScalarEvolution &SE,
3472  unsigned Depth = 0) {
3473  // Arbitrarily cap recursion to protect compile time.
3474  if (Depth >= 3)
3475  return S;
3476 
3477  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3478  // Break out add operands.
3479  for (const SCEV *S : Add->operands()) {
3480  const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3481  if (Remainder)
3482  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3483  }
3484  return nullptr;
3485  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3486  // Split a non-zero base out of an addrec.
3487  if (AR->getStart()->isZero() || !AR->isAffine())
3488  return S;
3489 
3490  const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3491  C, Ops, L, SE, Depth+1);
3492  // Split the non-zero AddRec unless it is part of a nested recurrence that
3493  // does not pertain to this loop.
3494  if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3495  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3496  Remainder = nullptr;
3497  }
3498  if (Remainder != AR->getStart()) {
3499  if (!Remainder)
3500  Remainder = SE.getConstant(AR->getType(), 0);
3501  return SE.getAddRecExpr(Remainder,
3502  AR->getStepRecurrence(SE),
3503  AR->getLoop(),
3504  //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3506  }
3507  } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3508  // Break (C * (a + b + c)) into C*a + C*b + C*c.
3509  if (Mul->getNumOperands() != 2)
3510  return S;
3511  if (const SCEVConstant *Op0 =
3512  dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3513  C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3514  const SCEV *Remainder =
3515  CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3516  if (Remainder)
3517  Ops.push_back(SE.getMulExpr(C, Remainder));
3518  return nullptr;
3519  }
3520  }
3521  return S;
3522 }
3523 
3524 /// Return true if the SCEV represents a value that may end up as a
3525 /// post-increment operation.
3527  LSRUse &LU, const SCEV *S, const Loop *L,
3528  ScalarEvolution &SE) {
3529  if (LU.Kind != LSRUse::Address ||
3530  !LU.AccessTy.getType()->isIntOrIntVectorTy())
3531  return false;
3532  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3533  if (!AR)
3534  return false;
3535  const SCEV *LoopStep = AR->getStepRecurrence(SE);
3536  if (!isa<SCEVConstant>(LoopStep))
3537  return false;
3538  if (LU.AccessTy.getType()->getScalarSizeInBits() !=
3539  LoopStep->getType()->getScalarSizeInBits())
3540  return false;
3541  // Check if a post-indexed load/store can be used.
3542  if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3543  TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3544  const SCEV *LoopStart = AR->getStart();
3545  if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3546  return true;
3547  }
3548  return false;
3549 }
3550 
3551 /// Helper function for LSRInstance::GenerateReassociations.
3552 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3553  const Formula &Base,
3554  unsigned Depth, size_t Idx,
3555  bool IsScaledReg) {
3556  const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3557  // Don't generate reassociations for the base register of a value that
3558  // may generate a post-increment operator. The reason is that the
3559  // reassociations cause extra base+register formula to be created,
3560  // and possibly chosen, but the post-increment is more efficient.
3561  if (TTI.shouldFavorPostInc() && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3562  return;
3564  const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3565  if (Remainder)
3566  AddOps.push_back(Remainder);
3567 
3568  if (AddOps.size() == 1)
3569  return;
3570 
3572  JE = AddOps.end();
3573  J != JE; ++J) {
3574  // Loop-variant "unknown" values are uninteresting; we won't be able to
3575  // do anything meaningful with them.
3576  if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3577  continue;
3578 
3579  // Don't pull a constant into a register if the constant could be folded
3580  // into an immediate field.
3581  if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3582  LU.AccessTy, *J, Base.getNumRegs() > 1))
3583  continue;
3584 
3585  // Collect all operands except *J.
3586  SmallVector<const SCEV *, 8> InnerAddOps(
3587  ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3588  InnerAddOps.append(std::next(J),
3589  ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3590 
3591  // Don't leave just a constant behind in a register if the constant could
3592  // be folded into an immediate field.
3593  if (InnerAddOps.size() == 1 &&
3594  isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3595  LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3596  continue;
3597 
3598  const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3599  if (InnerSum->isZero())
3600  continue;
3601  Formula F = Base;
3602 
3603  // Add the remaining pieces of the add back into the new formula.
3604  const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3605  if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3606  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3607  InnerSumSC->getValue()->getZExtValue())) {
3608  F.UnfoldedOffset =
3609  (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3610  if (IsScaledReg)
3611  F.ScaledReg = nullptr;
3612  else
3613  F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3614  } else if (IsScaledReg)
3615  F.ScaledReg = InnerSum;
3616  else
3617  F.BaseRegs[Idx] = InnerSum;
3618 
3619  // Add J as its own register, or an unfolded immediate.
3620  const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3621  if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3622  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3623  SC->getValue()->getZExtValue()))
3624  F.UnfoldedOffset =
3625  (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3626  else
3627  F.BaseRegs.push_back(*J);
3628  // We may have changed the number of register in base regs, adjust the
3629  // formula accordingly.
3630  F.canonicalize(*L);
3631 
3632  if (InsertFormula(LU, LUIdx, F))
3633  // If that formula hadn't been seen before, recurse to find more like
3634  // it.
3635  // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3636  // Because just Depth is not enough to bound compile time.
3637  // This means that every time AddOps.size() is greater 16^x we will add
3638  // x to Depth.
3639  GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3640  Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3641  }
3642 }
3643 
3644 /// Split out subexpressions from adds and the bases of addrecs.
3645 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3646  Formula Base, unsigned Depth) {
3647  assert(Base.isCanonical(*L) && "Input must be in the canonical form");
3648  // Arbitrarily cap recursion to protect compile time.
3649  if (Depth >= 3)
3650  return;
3651 
3652  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3653  GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3654 
3655  if (Base.Scale == 1)
3656  GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3657  /* Idx */ -1, /* IsScaledReg */ true);
3658 }
3659 
3660 /// Generate a formula consisting of all of the loop-dominating registers added
3661 /// into a single register.
3662 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3663  Formula Base) {
3664  // This method is only interesting on a plurality of registers.
3665  if (Base.BaseRegs.size() + (Base.Scale == 1) +
3666  (Base.UnfoldedOffset != 0) <= 1)
3667  return;
3668 
3669  // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3670  // processing the formula.
3671  Base.unscale();
3673  Formula NewBase = Base;
3674  NewBase.BaseRegs.clear();
3675  Type *CombinedIntegerType = nullptr;
3676  for (const SCEV *BaseReg : Base.BaseRegs) {
3677  if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3678  !SE.hasComputableLoopEvolution(BaseReg, L)) {
3679  if (!CombinedIntegerType)
3680  CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3681  Ops.push_back(BaseReg);
3682  }
3683  else
3684  NewBase.BaseRegs.push_back(BaseReg);
3685  }
3686 
3687  // If no register is relevant, we're done.
3688  if (Ops.size() == 0)
3689  return;
3690 
3691  // Utility function for generating the required variants of the combined
3692  // registers.
3693  auto GenerateFormula = [&](const SCEV *Sum) {
3694  Formula F = NewBase;
3695 
3696  // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3697  // opportunity to fold something. For now, just ignore such cases
3698  // rather than proceed with zero in a register.
3699  if (Sum->isZero())
3700  return;
3701 
3702  F.BaseRegs.push_back(Sum);
3703  F.canonicalize(*L);
3704  (void)InsertFormula(LU, LUIdx, F);
3705  };
3706 
3707  // If we collected at least two registers, generate a formula combining them.
3708  if (Ops.size() > 1) {
3709  SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3710  GenerateFormula(SE.getAddExpr(OpsCopy));
3711  }
3712 
3713  // If we have an unfolded offset, generate a formula combining it with the
3714  // registers collected.
3715  if (NewBase.UnfoldedOffset) {
3716  assert(CombinedIntegerType && "Missing a type for the unfolded offset");
3717  Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3718  true));
3719  NewBase.UnfoldedOffset = 0;
3720  GenerateFormula(SE.getAddExpr(Ops));
3721  }
3722 }
3723 
3724 /// Helper function for LSRInstance::GenerateSymbolicOffsets.
3725 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3726  const Formula &Base, size_t Idx,
3727  bool IsScaledReg) {
3728  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3729  GlobalValue *GV = ExtractSymbol(G, SE);
3730  if (G->isZero() || !GV)
3731  return;
3732  Formula F = Base;
3733  F.BaseGV = GV;
3734  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3735  return;
3736  if (IsScaledReg)
3737  F.ScaledReg = G;
3738  else
3739  F.BaseRegs[Idx] = G;
3740  (void)InsertFormula(LU, LUIdx, F);
3741 }
3742 
3743 /// Generate reuse formulae using symbolic offsets.
3744 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3745  Formula Base) {
3746  // We can't add a symbolic offset if the address already contains one.
3747  if (Base.BaseGV) return;
3748 
3749  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3750  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3751  if (Base.Scale == 1)
3752  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3753  /* IsScaledReg */ true);
3754 }
3755 
3756 /// Helper function for LSRInstance::GenerateConstantOffsets.
3757 void LSRInstance::GenerateConstantOffsetsImpl(
3758  LSRUse &LU, unsigned LUIdx, const Formula &Base,
3759  const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3760 
3761  auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3762  Formula F = Base;
3763  F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3764 
3765  if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3766  LU.AccessTy, F)) {
3767  // Add the offset to the base register.
3768  const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3769  // If it cancelled out, drop the base register, otherwise update it.
3770  if (NewG->isZero()) {
3771  if (IsScaledReg) {
3772  F.Scale = 0;
3773  F.ScaledReg = nullptr;
3774  } else
3775  F.deleteBaseReg(F.BaseRegs[Idx]);
3776  F.canonicalize(*L);
3777  } else if (IsScaledReg)
3778  F.ScaledReg = NewG;
3779  else
3780  F.BaseRegs[Idx] = NewG;
3781 
3782  (void)InsertFormula(LU, LUIdx, F);
3783  }
3784  };
3785 
3786  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3787 
3788  // With constant offsets and constant steps, we can generate pre-inc
3789  // accesses by having the offset equal the step. So, for access #0 with a
3790  // step of 8, we generate a G - 8 base which would require the first access
3791  // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3792  // for itself and hopefully becomes the base for other accesses. This means
3793  // means that a single pre-indexed access can be generated to become the new
3794  // base pointer for each iteration of the loop, resulting in no extra add/sub
3795  // instructions for pointer updating.
3796  if (FavorBackedgeIndex && LU.Kind == LSRUse::Address) {
3797  if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3798  if (auto *StepRec =
3799  dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3800  const APInt &StepInt = StepRec->getAPInt();
3801  int64_t Step = StepInt.isNegative() ?
3802  StepInt.getSExtValue() : StepInt.getZExtValue();
3803 
3804  for (int64_t Offset : Worklist) {
3805  Offset -= Step;
3806  GenerateOffset(G, Offset);
3807  }
3808  }
3809  }
3810  }
3811  for (int64_t Offset : Worklist)
3812  GenerateOffset(G, Offset);
3813 
3814  int64_t Imm = ExtractImmediate(G, SE);
3815  if (G->isZero() || Imm == 0)
3816  return;
3817  Formula F = Base;
3818  F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3819  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3820  return;
3821  if (IsScaledReg)
3822  F.ScaledReg = G;
3823  else
3824  F.BaseRegs[Idx] = G;
3825  (void)InsertFormula(LU, LUIdx, F);
3826 }
3827 
3828 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3829 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3830  Formula Base) {
3831  // TODO: For now, just add the min and max offset, because it usually isn't
3832  // worthwhile looking at everything inbetween.
3833  SmallVector<int64_t, 2> Worklist;
3834  Worklist.push_back(LU.MinOffset);
3835  if (LU.MaxOffset != LU.MinOffset)
3836  Worklist.push_back(LU.MaxOffset);
3837 
3838  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3839  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3840  if (Base.Scale == 1)
3841  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3842  /* IsScaledReg */ true);
3843 }
3844 
3845 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3846 /// == y -> x*c == y*c.
3847 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3848  Formula Base) {
3849  if (LU.Kind != LSRUse::ICmpZero) return;
3850 
3851  // Determine the integer type for the base formula.
3852  Type *IntTy = Base.getType();
3853  if (!IntTy) return;
3854  if (SE.getTypeSizeInBits(IntTy) > 64) return;
3855 
3856  // Don't do this if there is more than one offset.
3857  if (LU.MinOffset != LU.MaxOffset) return;
3858 
3859  // Check if transformation is valid. It is illegal to multiply pointer.
3860  if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3861  return;
3862  for (const SCEV *BaseReg : Base.BaseRegs)
3863  if (BaseReg->getType()->isPointerTy())
3864  return;
3865  assert(!Base.BaseGV && "ICmpZero use is not legal!");
3866 
3867  // Check each interesting stride.
3868  for (int64_t Factor : Factors) {
3869  // Check that the multiplication doesn't overflow.
3870  if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3871  continue;
3872  int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3873  if (NewBaseOffset / Factor != Base.BaseOffset)
3874  continue;
3875  // If the offset will be truncated at this use, check that it is in bounds.
3876  if (!IntTy->isPointerTy() &&
3877  !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3878  continue;
3879 
3880  // Check that multiplying with the use offset doesn't overflow.
3881  int64_t Offset = LU.MinOffset;
3882  if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3883  continue;
3884  Offset = (uint64_t)Offset * Factor;
3885  if (Offset / Factor != LU.MinOffset)
3886  continue;
3887  // If the offset will be truncated at this use, check that it is in bounds.
3888  if (!IntTy->isPointerTy() &&
3889  !ConstantInt::isValueValidForType(IntTy, Offset))
3890  continue;
3891 
3892  Formula F = Base;
3893  F.BaseOffset = NewBaseOffset;
3894 
3895  // Check that this scale is legal.
3896  if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3897  continue;
3898 
3899  // Compensate for the use having MinOffset built into it.
3900  F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3901 
3902  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3903 
3904  // Check that multiplying with each base register doesn't overflow.
3905  for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3906  F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3907  if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3908  goto next;
3909  }
3910 
3911  // Check that multiplying with the scaled register doesn't overflow.
3912  if (F.ScaledReg) {
3913  F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3914  if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3915  continue;
3916  }
3917 
3918  // Check that multiplying with the unfolded offset doesn't overflow.
3919  if (F.UnfoldedOffset != 0) {
3920  if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3921  Factor == -1)
3922  continue;
3923  F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3924  if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3925  continue;
3926  // If the offset will be truncated, check that it is in bounds.
3927  if (!IntTy->isPointerTy() &&
3928  !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3929  continue;
3930  }
3931 
3932  // If we make it here and it's legal, add it.
3933  (void)InsertFormula(LU, LUIdx, F);
3934  next:;
3935  }
3936 }
3937 
3938 /// Generate stride factor reuse formulae by making use of scaled-offset address
3939 /// modes, for example.
3940 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3941  // Determine the integer type for the base formula.
3942  Type *IntTy = Base.getType();
3943  if (!IntTy) return;
3944 
3945  // If this Formula already has a scaled register, we can't add another one.
3946  // Try to unscale the formula to generate a better scale.
3947  if (Base.Scale != 0 && !Base.unscale())
3948  return;
3949 
3950  assert(Base.Scale == 0 && "unscale did not did its job!");
3951 
3952  // Check each interesting stride.
3953  for (int64_t Factor : Factors) {
3954  Base.Scale = Factor;
3955  Base.HasBaseReg = Base.BaseRegs.size() > 1;
3956  // Check whether this scale is going to be legal.
3957  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3958  Base)) {
3959  // As a special-case, handle special out-of-loop Basic users specially.
3960  // TODO: Reconsider this special case.
3961  if (LU.Kind == LSRUse::Basic &&
3962  isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3963  LU.AccessTy, Base) &&
3964  LU.AllFixupsOutsideLoop)
3965  LU.Kind = LSRUse::Special;
3966  else
3967  continue;
3968  }
3969  // For an ICmpZero, negating a solitary base register won't lead to
3970  // new solutions.
3971  if (LU.Kind == LSRUse::ICmpZero &&
3972  !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3973  continue;
3974  // For each addrec base reg, if its loop is current loop, apply the scale.
3975  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3976  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3977  if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3978  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3979  if (FactorS->isZero())
3980  continue;
3981  // Divide out the factor, ignoring high bits, since we'll be
3982  // scaling the value back up in the end.
3983  if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3984  // TODO: This could be optimized to avoid all the copying.
3985  Formula F = Base;
3986  F.ScaledReg = Quotient;
3987  F.deleteBaseReg(F.BaseRegs[i]);
3988  // The canonical representation of 1*reg is reg, which is already in
3989  // Base. In that case, do not try to insert the formula, it will be
3990  // rejected anyway.
3991  if (F.Scale == 1 && (F.BaseRegs.empty() ||
3992  (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3993  continue;
3994  // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
3995  // non canonical Formula with ScaledReg's loop not being L.
3996  if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
3997  F.canonicalize(*L);
3998  (void)InsertFormula(LU, LUIdx, F);
3999  }
4000  }
4001  }
4002  }
4003 }
4004 
4005 /// Generate reuse formulae from different IV types.
4006 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4007  // Don't bother truncating symbolic values.
4008  if (Base.BaseGV) return;
4009 
4010  // Determine the integer type for the base formula.
4011  Type *DstTy = Base.getType();
4012  if (!DstTy) return;
4013  DstTy = SE.getEffectiveSCEVType(DstTy);
4014 
4015  for (Type *SrcTy : Types) {
4016  if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4017  Formula F = Base;
4018 
4019  // Sometimes SCEV is able to prove zero during ext transform. It may
4020  // happen if SCEV did not do all possible transforms while creating the
4021  // initial node (maybe due to depth limitations), but it can do them while
4022  // taking ext.
4023  if (F.ScaledReg) {
4024  const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4025  if (NewScaledReg->isZero())
4026  continue;
4027  F.ScaledReg = NewScaledReg;
4028  }
4029  bool HasZeroBaseReg = false;
4030  for (const SCEV *&BaseReg : F.BaseRegs) {
4031  const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4032  if (NewBaseReg->isZero()) {
4033  HasZeroBaseReg = true;
4034  break;
4035  }
4036  BaseReg = NewBaseReg;
4037  }
4038  if (HasZeroBaseReg)
4039  continue;
4040 
4041  // TODO: This assumes we've done basic processing on all uses and
4042  // have an idea what the register usage is.
4043  if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4044  continue;
4045 
4046  F.canonicalize(*L);
4047  (void)InsertFormula(LU, LUIdx, F);
4048  }
4049  }
4050 }
4051 
4052 namespace {
4053 
4054 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4055 /// modifications so that the search phase doesn't have to worry about the data
4056 /// structures moving underneath it.
4057 struct WorkItem {
4058  size_t LUIdx;
4059  int64_t Imm;
4060  const SCEV *OrigReg;
4061 
4062  WorkItem(size_t LI, int64_t I, const SCEV *R)
4063  : LUIdx(LI), Imm(I), OrigReg(R) {}
4064 
4065  void print(raw_ostream &OS) const;
4066  void dump() const;
4067 };
4068 
4069 } // end anonymous namespace
4070 
4071 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4072 void WorkItem::print(raw_ostream &OS) const {
4073  OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4074  << " , add offset " << Imm;
4075 }
4076 
4077 LLVM_DUMP_METHOD void WorkItem::dump() const {
4078  print(errs()); errs() << '\n';
4079 }
4080 #endif
4081 
4082 /// Look for registers which are a constant distance apart and try to form reuse
4083 /// opportunities between them.
4084 void LSRInstance::GenerateCrossUseConstantOffsets() {
4085  // Group the registers by their value without any added constant offset.
4086  using ImmMapTy = std::map<int64_t, const SCEV *>;
4087 
4089  DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4091  for (const SCEV *Use : RegUses) {
4092  const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4093  int64_t Imm = ExtractImmediate(Reg, SE);
4094  auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4095  if (Pair.second)
4096  Sequence.push_back(Reg);
4097  Pair.first->second.insert(std::make_pair(Imm, Use));
4098  UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4099  }
4100 
4101  // Now examine each set of registers with the same base value. Build up
4102  // a list of work to do and do the work in a separate step so that we're
4103  // not adding formulae and register counts while we're searching.
4104  SmallVector<WorkItem, 32> WorkItems;
4105  SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4106  for (const SCEV *Reg : Sequence) {
4107  const ImmMapTy &Imms = Map.find(Reg)->second;
4108 
4109  // It's not worthwhile looking for reuse if there's only one offset.
4110  if (Imms.size() == 1)
4111  continue;
4112 
4113  LLVM_DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
4114  for (const auto &Entry
4115  : Imms) dbgs()
4116  << ' ' << Entry.first;
4117  dbgs() << '\n');
4118 
4119  // Examine each offset.
4120  for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4121  J != JE; ++J) {
4122  const SCEV *OrigReg = J->second;
4123 
4124  int64_t JImm = J->first;
4125  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4126 
4127  if (!isa<SCEVConstant>(OrigReg) &&
4128  UsedByIndicesMap[Reg].count() == 1) {
4129  LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg
4130  << '\n');
4131  continue;
4132  }
4133 
4134  // Conservatively examine offsets between this orig reg a few selected
4135  // other orig regs.
4136  int64_t First = Imms.begin()->first;
4137  int64_t Last = std::prev(Imms.end())->first;
4138  // Compute (First + Last) / 2 without overflow using the fact that
4139  // First + Last = 2 * (First + Last) + (First ^ Last).
4140  int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4141  // If the result is negative and First is odd and Last even (or vice versa),
4142  // we rounded towards -inf. Add 1 in that case, to round towards 0.
4143  Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4144  ImmMapTy::const_iterator OtherImms[] = {
4145  Imms.begin(), std::prev(Imms.end()),
4146  Imms.lower_bound(Avg)};
4147  for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4148  ImmMapTy::const_iterator M = OtherImms[i];
4149  if (M == J || M == JE) continue;
4150 
4151  // Compute the difference between the two.
4152  int64_t Imm = (uint64_t)JImm - M->first;
4153  for (unsigned LUIdx : UsedByIndices.set_bits())
4154  // Make a memo of this use, offset, and register tuple.
4155  if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4156  WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4157  }
4158  }
4159  }
4160 
4161  Map.clear();
4162  Sequence.clear();
4163  UsedByIndicesMap.clear();
4164  UniqueItems.clear();
4165 
4166  // Now iterate through the worklist and add new formulae.
4167  for (const WorkItem &WI : WorkItems) {
4168  size_t LUIdx = WI.LUIdx;
4169  LSRUse &LU = Uses[LUIdx];
4170  int64_t Imm = WI.Imm;
4171  const SCEV *OrigReg = WI.OrigReg;
4172 
4173  Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4174  const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4175  unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4176 
4177  // TODO: Use a more targeted data structure.
4178  for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4179  Formula F = LU.Formulae[L];
4180  // FIXME: The code for the scaled and unscaled registers looks
4181  // very similar but slightly different. Investigate if they
4182  // could be merged. That way, we would not have to unscale the
4183  // Formula.
4184  F.unscale();
4185  // Use the immediate in the scaled register.
4186  if (F.ScaledReg == OrigReg) {
4187  int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4188  // Don't create 50 + reg(-50).
4189  if (F.referencesReg(SE.getSCEV(
4190  ConstantInt::get(IntTy, -(uint64_t)Offset))))
4191  continue;
4192  Formula NewF = F;
4193  NewF.BaseOffset = Offset;
4194  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4195  NewF))
4196  continue;
4197  NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4198 
4199  // If the new scale is a constant in a register, and adding the constant
4200  // value to the immediate would produce a value closer to zero than the
4201  // immediate itself, then the formula isn't worthwhile.
4202  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4203  if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4204  (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4205  .ule(std::abs(NewF.BaseOffset)))
4206  continue;
4207 
4208  // OK, looks good.
4209  NewF.canonicalize(*this->L);
4210  (void)InsertFormula(LU, LUIdx, NewF);
4211  } else {
4212  // Use the immediate in a base register.
4213  for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4214  const SCEV *BaseReg = F.BaseRegs[N];
4215  if (BaseReg != OrigReg)
4216  continue;
4217  Formula NewF = F;
4218  NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4219  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4220  LU.Kind, LU.AccessTy, NewF)) {
4221  if (TTI.shouldFavorPostInc() &&
4222  mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4223  continue;
4224  if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4225  continue;
4226  NewF = F;
4227  NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4228  }
4229  NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4230 
4231  // If the new formula has a constant in a register, and adding the
4232  // constant value to the immediate would produce a value closer to
4233  // zero than the immediate itself, then the formula isn't worthwhile.
4234  for (const SCEV *NewReg : NewF.BaseRegs)
4235  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4236  if ((C->getAPInt() + NewF.BaseOffset)
4237  .abs()
4238  .slt(std::abs(NewF.BaseOffset)) &&
4239  (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4240  countTrailingZeros<uint64_t>(NewF.BaseOffset))
4241  goto skip_formula;
4242 
4243  // Ok, looks good.
4244  NewF.canonicalize(*this->L);
4245  (void)InsertFormula(LU, LUIdx, NewF);
4246  break;
4247  skip_formula:;
4248  }
4249  }
4250  }
4251  }
4252 }
4253 
4254 /// Generate formulae for each use.
4255 void
4256 LSRInstance::GenerateAllReuseFormulae() {
4257  // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4258  // queries are more precise.
4259  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4260  LSRUse &LU = Uses[LUIdx];
4261  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4262  GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4263  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4264  GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4265  }
4266  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4267  LSRUse &LU = Uses[LUIdx];
4268  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4269  GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4270  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4271  GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4272  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4273  GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4274  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4275  GenerateScales(LU, LUIdx, LU.Formulae[i]);
4276  }
4277  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4278  LSRUse &LU = Uses[LUIdx];
4279  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4280  GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4281  }
4282 
4283  GenerateCrossUseConstantOffsets();
4284 
4285  LLVM_DEBUG(dbgs() << "\n"
4286  "After generating reuse formulae:\n";
4287  print_uses(dbgs()));
4288 }
4289 
4290 /// If there are multiple formulae with the same set of registers used
4291 /// by other uses, pick the best one and delete the others.
4292 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4293  DenseSet<const SCEV *> VisitedRegs;
4296 #ifndef NDEBUG
4297  bool ChangedFormulae = false;
4298 #endif
4299 
4300  // Collect the best formula for each unique set of shared registers. This
4301  // is reset for each use.
4302  using BestFormulaeTy =
4303  DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4304 
4305  BestFormulaeTy BestFormulae;
4306 
4307  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4308  LSRUse &LU = Uses[LUIdx];
4309  LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
4310  dbgs() << '\n');
4311 
4312  bool Any = false;
4313  for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4314  FIdx != NumForms; ++FIdx) {
4315  Formula &F = LU.Formulae[FIdx];
4316 
4317  // Some formulas are instant losers. For example, they may depend on
4318  // nonexistent AddRecs from other loops. These need to be filtered
4319  // immediately, otherwise heuristics could choose them over others leading
4320  // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4321  // avoids the need to recompute this information across formulae using the
4322  // same bad AddRec. Passing LoserRegs is also essential unless we remove
4323  // the corresponding bad register from the Regs set.
4324  Cost CostF(L, SE, TTI);
4325  Regs.clear();
4326  CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4327  if (CostF.isLoser()) {
4328  // During initial formula generation, undesirable formulae are generated
4329  // by uses within other loops that have some non-trivial address mode or
4330  // use the postinc form of the IV. LSR needs to provide these formulae
4331  // as the basis of rediscovering the desired formula that uses an AddRec
4332  // corresponding to the existing phi. Once all formulae have been
4333  // generated, these initial losers may be pruned.
4334  LLVM_DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
4335  dbgs() << "\n");
4336  }
4337  else {
4339  for (const SCEV *Reg : F.BaseRegs) {
4340  if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4341  Key.push_back(Reg);
4342  }
4343  if (F.ScaledReg &&
4344  RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4345  Key.push_back(F.ScaledReg);
4346  // Unstable sort by host order ok, because this is only used for
4347  // uniquifying.
4348  llvm::sort(Key);
4349 
4350  std::pair<BestFormulaeTy::const_iterator, bool> P =
4351  BestFormulae.insert(std::make_pair(Key, FIdx));
4352  if (P.second)
4353  continue;
4354 
4355  Formula &Best = LU.Formulae[P.first->second];
4356 
4357  Cost CostBest(L, SE, TTI);
4358  Regs.clear();
4359  CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4360  if (CostF.isLess(CostBest))
4361  std::swap(F, Best);
4362  LLVM_DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4363  dbgs() << "\n"
4364  " in favor of formula ";
4365  Best.print(dbgs()); dbgs() << '\n');
4366  }
4367 #ifndef NDEBUG
4368  ChangedFormulae = true;
4369 #endif
4370  LU.DeleteFormula(F);
4371  --FIdx;
4372  --NumForms;
4373  Any = true;
4374  }
4375 
4376  // Now that we've filtered out some formulae, recompute the Regs set.
4377  if (Any)
4378  LU.RecomputeRegs(LUIdx, RegUses);
4379 
4380  // Reset this to prepare for the next use.
4381  BestFormulae.clear();
4382  }
4383 
4384  LLVM_DEBUG(if (ChangedFormulae) {
4385  dbgs() << "\n"
4386  "After filtering out undesirable candidates:\n";
4387  print_uses(dbgs());
4388  });
4389 }
4390 
4391 /// Estimate the worst-case number of solutions the solver might have to
4392 /// consider. It almost never considers this many solutions because it prune the
4393 /// search space, but the pruning isn't always sufficient.
4394 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4395  size_t Power = 1;
4396  for (const LSRUse &LU : Uses) {
4397  size_t FSize = LU.Formulae.size();
4398  if (FSize >= ComplexityLimit) {
4399  Power = ComplexityLimit;
4400  break;
4401  }
4402  Power *= FSize;
4403  if (Power >= ComplexityLimit)
4404  break;
4405  }
4406  return Power;
4407 }
4408 
4409 /// When one formula uses a superset of the registers of another formula, it
4410 /// won't help reduce register pressure (though it may not necessarily hurt
4411 /// register pressure); remove it to simplify the system.
4412 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4413  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4414  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4415 
4416  LLVM_DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4417  "which use a superset of registers used by other "
4418  "formulae.\n");
4419 
4420  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4421  LSRUse &LU = Uses[LUIdx];
4422  bool Any = false;
4423  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4424  Formula &F = LU.Formulae[i];
4425  // Look for a formula with a constant or GV in a register. If the use
4426  // also has a formula with that same value in an immediate field,
4427  // delete the one that uses a register.
4429  I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4430  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4431  Formula NewF = F;
4432  //FIXME: Formulas should store bitwidth to do wrapping properly.
4433  // See PR41034.
4434  NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4435  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4436  (I - F.BaseRegs.begin()));
4437  if (LU.HasFormulaWithSameRegs(NewF)) {
4438  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4439  dbgs() << '\n');
4440  LU.DeleteFormula(F);
4441  --i;
4442  --e;
4443  Any = true;
4444  break;
4445  }
4446  } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4447  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4448  if (!F.BaseGV) {
4449  Formula NewF = F;
4450  NewF.BaseGV = GV;
4451  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4452  (I - F.BaseRegs.begin()));
4453  if (LU.HasFormulaWithSameRegs(NewF)) {
4454  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4455  dbgs() << '\n');
4456  LU.DeleteFormula(F);
4457  --i;
4458  --e;
4459  Any = true;
4460  break;
4461  }
4462  }
4463  }
4464  }
4465  }
4466  if (Any)
4467  LU.RecomputeRegs(LUIdx, RegUses);
4468  }
4469 
4470  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4471  }
4472 }
4473 
4474 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4475 /// allocate a single register for them.
4476 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4477  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4478  return;
4479 
4480  LLVM_DEBUG(
4481  dbgs() << "The search space is too complex.\n"
4482  "Narrowing the search space by assuming that uses separated "
4483  "by a constant offset will use the same registers.\n");
4484 
4485  // This is especially useful for unrolled loops.
4486 
4487  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4488  LSRUse &LU = Uses[LUIdx];
4489  for (const Formula &F : LU.Formulae) {
4490  if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4491  continue;
4492 
4493  LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4494  if (!LUThatHas)
4495  continue;
4496 
4497  if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4498  LU.Kind, LU.AccessTy))
4499  continue;
4500 
4501  LLVM_DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4502 
4503  LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4504 
4505  // Transfer the fixups of LU to LUThatHas.
4506  for (LSRFixup &Fixup : LU.Fixups) {
4507  Fixup.Offset += F.BaseOffset;
4508  LUThatHas->pushFixup(Fixup);
4509  LLVM_DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4510  }
4511 
4512  // Delete formulae from the new use which are no longer legal.
4513  bool Any = false;
4514  for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4515  Formula &F = LUThatHas->Formulae[i];
4516  if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4517  LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4518  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4519  LUThatHas->DeleteFormula(F);
4520  --i;
4521  --e;
4522  Any = true;
4523  }
4524  }
4525 
4526  if (Any)
4527  LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4528 
4529  // Delete the old use.
4530  DeleteUse(LU, LUIdx);
4531  --LUIdx;
4532  --NumUses;
4533  break;
4534  }
4535  }
4536 
4537  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4538 }
4539 
4540 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4541 /// we've done more filtering, as it may be able to find more formulae to
4542 /// eliminate.
4543 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4544  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4545  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4546 
4547  LLVM_DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4548  "undesirable dedicated registers.\n");
4549 
4550  FilterOutUndesirableDedicatedRegisters();
4551 
4552  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4553  }
4554 }
4555 
4556 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4557 /// Pick the best one and delete the others.
4558 /// This narrowing heuristic is to keep as many formulae with different
4559 /// Scale and ScaledReg pair as possible while narrowing the search space.
4560 /// The benefit is that it is more likely to find out a better solution
4561 /// from a formulae set with more Scale and ScaledReg variations than
4562 /// a formulae set with the same Scale and ScaledReg. The picking winner
4563 /// reg heuristic will often keep the formulae with the same Scale and
4564 /// ScaledReg and filter others, and we want to avoid that if possible.
4565 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4566  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4567  return;
4568 
4569  LLVM_DEBUG(
4570  dbgs() << "The search space is too complex.\n"
4571  "Narrowing the search space by choosing the best Formula "
4572  "from the Formulae with the same Scale and ScaledReg.\n");
4573 
4574  // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4575  using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4576 
4577  BestFormulaeTy BestFormulae;
4578 #ifndef NDEBUG
4579  bool ChangedFormulae = false;
4580 #endif
4581  DenseSet<const SCEV *> VisitedRegs;
4583 
4584  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4585  LSRUse &LU = Uses[LUIdx];
4586  LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
4587  dbgs() << '\n');
4588 
4589  // Return true if Formula FA is better than Formula FB.
4590  auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4591  // First we will try to choose the Formula with fewer new registers.
4592  // For a register used by current Formula, the more the register is
4593  // shared among LSRUses, the less we increase the register number
4594  // counter of the formula.
4595  size_t FARegNum = 0;
4596  for (const SCEV *Reg : FA.BaseRegs) {
4597  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4598  FARegNum += (NumUses - UsedByIndices.count() + 1);
4599  }
4600  size_t FBRegNum = 0;
4601  for (const SCEV *Reg : FB.BaseRegs) {
4602  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4603  FBRegNum += (NumUses - UsedByIndices.count() + 1);
4604  }
4605  if (FARegNum != FBRegNum)
4606  return FARegNum < FBRegNum;
4607 
4608  // If the new register numbers are the same, choose the Formula with
4609  // less Cost.
4610  Cost CostFA(L, SE, TTI);
4611  Cost CostFB(L, SE, TTI);
4612  Regs.clear();
4613  CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4614  Regs.clear();
4615  CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4616  return CostFA.isLess(CostFB);
4617  };
4618 
4619  bool Any = false;
4620  for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4621  ++FIdx) {
4622  Formula &F = LU.Formulae[FIdx];
4623  if (!F.ScaledReg)
4624  continue;
4625  auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4626  if (P.second)
4627  continue;
4628 
4629  Formula &Best = LU.Formulae[P.first->second];
4630  if (IsBetterThan(F, Best))
4631  std::swap(F, Best);
4632  LLVM_DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4633  dbgs() << "\n"
4634  " in favor of formula ";
4635  Best.print(dbgs()); dbgs() << '\n');
4636 #ifndef NDEBUG
4637  ChangedFormulae = true;
4638 #endif
4639  LU.DeleteFormula(F);
4640  --FIdx;
4641  --NumForms;
4642  Any = true;
4643  }
4644  if (Any)
4645  LU.RecomputeRegs(LUIdx, RegUses);
4646 
4647  // Reset this to prepare for the next use.
4648  BestFormulae.clear();
4649  }
4650 
4651  LLVM_DEBUG(if (ChangedFormulae) {
4652  dbgs() << "\n"
4653  "After filtering out undesirable candidates:\n";
4654  print_uses(dbgs());
4655  });
4656 }
4657 
4658 /// The function delete formulas with high registers number expectation.
4659 /// Assuming we don't know the value of each formula (already delete
4660 /// all inefficient), generate probability of not selecting for each
4661 /// register.
4662 /// For example,
4663 /// Use1:
4664 /// reg(a) + reg({0,+,1})
4665 /// reg(a) + reg({-1,+,1}) + 1
4666 /// reg({a,+,1})
4667 /// Use2:
4668 /// reg(b) + reg({0,+,1})
4669 /// reg(b) + reg({-1,+,1}) + 1
4670 /// reg({b,+,1})
4671 /// Use3:
4672 /// reg(c) + reg(b) + reg({0,+,1})
4673 /// reg(c) + reg({b,+,1})
4674 ///
4675 /// Probability of not selecting
4676 /// Use1 Use2 Use3
4677 /// reg(a) (1/3) * 1 * 1
4678 /// reg(b) 1 * (1/3) * (1/2)
4679 /// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4680 /// reg({-1,+,1}) (2/3) * (2/3) * 1
4681 /// reg({a,+,1}) (2/3) * 1 * 1
4682 /// reg({b,+,1}) 1 * (2/3) * (2/3)
4683 /// reg(c) 1 * 1 * 0
4684 ///
4685 /// Now count registers number mathematical expectation for each formula:
4686 /// Note that for each use we exclude probability if not selecting for the use.
4687 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4688 /// probabilty 1/3 of not selecting for Use1).
4689 /// Use1:
4690 /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4691 /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4692 /// reg({a,+,1}) 1
4693 /// Use2:
4694 /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4695 /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4696 /// reg({b,+,1}) 2/3
4697 /// Use3:
4698 /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4699 /// reg(c) + reg({b,+,1}) 1 + 2/3
4700 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4701  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4702  return;
4703  // Ok, we have too many of formulae on our hands to conveniently handle.
4704  // Use a rough heuristic to thin out the list.
4705 
4706  // Set of Regs wich will be 100% used in final solution.
4707  // Used in each formula of a solution (in example above this is reg(c)).
4708  // We can skip them in calculations.
4710  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4711 
4712  // Map each register to probability of not selecting
4714  for (const SCEV *Reg : RegUses) {
4715  if (UniqRegs.count(Reg))
4716  continue;
4717  float PNotSel = 1;
4718  for (const LSRUse &LU : Uses) {
4719  if (!LU.Regs.count(Reg))
4720  continue;
4721  float P = LU.getNotSelectedProbability(Reg);
4722  if (P != 0.0)
4723  PNotSel *= P;
4724  else
4725  UniqRegs.insert(Reg);
4726  }
4727  RegNumMap.insert(std::make_pair(Reg, PNotSel));
4728  }
4729 
4730  LLVM_DEBUG(
4731  dbgs() << "Narrowing the search space by deleting costly formulas\n");
4732 
4733  // Delete formulas where registers number expectation is high.
4734  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4735  LSRUse &LU = Uses[LUIdx];
4736  // If nothing to delete - continue.
4737  if (LU.Formulae.size() < 2)
4738  continue;
4739  // This is temporary solution to test performance. Float should be
4740  // replaced with round independent type (based on integers) to avoid
4741  // different results for different target builds.
4742  float FMinRegNum = LU.Formulae[0].getNumRegs();
4743  float FMinARegNum = LU.Formulae[0].getNumRegs();
4744  size_t MinIdx = 0;
4745  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4746  Formula &F = LU.Formulae[i];
4747  float FRegNum = 0;
4748  float FARegNum = 0;
4749  for (const SCEV *BaseReg : F.BaseRegs) {
4750  if (UniqRegs.count(BaseReg))
4751  continue;
4752  FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4753  if (isa<SCEVAddRecExpr>(BaseReg))
4754  FARegNum +=
4755  RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4756  }
4757  if (const SCEV *ScaledReg = F.ScaledReg) {
4758  if (!UniqRegs.count(ScaledReg)) {
4759  FRegNum +=
4760  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4761  if (isa<SCEVAddRecExpr>(ScaledReg))
4762  FARegNum +=
4763  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4764  }
4765  }
4766  if (FMinRegNum > FRegNum ||
4767  (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4768  FMinRegNum = FRegNum;
4769  FMinARegNum = FARegNum;
4770  MinIdx = i;
4771  }
4772  }
4773  LLVM_DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());
4774  dbgs() << " with min reg num " << FMinRegNum << '\n');
4775  if (MinIdx != 0)
4776  std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4777  while (LU.Formulae.size() != 1) {
4778  LLVM_DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());
4779  dbgs() << '\n');
4780  LU.Formulae.pop_back();
4781  }
4782  LU.RecomputeRegs(LUIdx, RegUses);
4783  assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
4784  Formula &F = LU.Formulae[0];
4785  LLVM_DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n');
4786  // When we choose the formula, the regs become unique.
4787  UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4788  if (F.ScaledReg)
4789  UniqRegs.insert(F.ScaledReg);
4790  }
4791  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4792 }
4793 
4794 /// Pick a register which seems likely to be profitable, and then in any use
4795 /// which has any reference to that register, delete all formulae which do not
4796 /// reference that register.
4797 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4798  // With all other options exhausted, loop until the system is simple
4799  // enough to handle.
4801  while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4802  // Ok, we have too many of formulae on our hands to conveniently handle.
4803  // Use a rough heuristic to thin out the list.
4804  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4805 
4806  // Pick the register which is used by the most LSRUses, which is likely
4807  // to be a good reuse register candidate.
4808  const SCEV *Best = nullptr;
4809  unsigned BestNum = 0;
4810  for (const SCEV *Reg : RegUses) {
4811  if (Taken.count(Reg))
4812  continue;
4813  if (!Best) {
4814  Best = Reg;
4815  BestNum = RegUses.getUsedByIndices(Reg).count();
4816  } else {
4817  unsigned Count = RegUses.getUsedByIndices(Reg).count();
4818  if (Count > BestNum) {
4819  Best = Reg;
4820  BestNum = Count;
4821  }
4822  }
4823  }
4824 
4825  LLVM_DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4826  << " will yield profitable reuse.\n");
4827  Taken.insert(Best);
4828 
4829  // In any use with formulae which references this register, delete formulae
4830  // which don't reference it.
4831  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4832  LSRUse &LU = Uses[LUIdx];
4833  if (!LU.Regs.count(Best)) continue;
4834 
4835  bool Any = false;
4836  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4837  Formula &F = LU.Formulae[i];
4838  if (!F.referencesReg(Best)) {
4839  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4840  LU.DeleteFormula(F);
4841  --e;
4842  --i;
4843  Any = true;
4844  assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4845  continue;
4846  }
4847  }
4848 
4849  if (Any)
4850  LU.RecomputeRegs(LUIdx, RegUses);
4851  }
4852 
4853  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4854  }
4855 }
4856 
4857 /// If there are an extraordinary number of formulae to choose from, use some
4858 /// rough heuristics to prune down the number of formulae. This keeps the main
4859 /// solver from taking an extraordinary amount of time in some worst-case
4860 /// scenarios.
4861 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4862  NarrowSearchSpaceByDetectingSupersets();
4863  NarrowSearchSpaceByCollapsingUnrolledCode();
4864  NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4865  if (FilterSameScaledReg)
4866  NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4867  if (LSRExpNarrow)
4868  NarrowSearchSpaceByDeletingCostlyFormulas();
4869  else
4870  NarrowSearchSpaceByPickingWinnerRegs();
4871 }
4872 
4873 /// This is the recursive solver.
4874 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4875  Cost &SolutionCost,
4877  const Cost &CurCost,
4878  const SmallPtrSet<const SCEV *, 16> &CurRegs,
4879  DenseSet<const SCEV *> &VisitedRegs) const {
4880  // Some ideas:
4881  // - prune more:
4882  // - use more aggressive filtering
4883  // - sort the formula so that the most profitable solutions are found first
4884  // - sort the uses too
4885  // - search faster:
4886  // - don't compute a cost, and then compare. compare while computing a cost
4887  // and bail early.
4888  // - track register sets with SmallBitVector
4889 
4890  const LSRUse &LU = Uses[Workspace.size()];
4891 
4892  // If this use references any register that's already a part of the
4893  // in-progress solution, consider it a requirement that a formula must
4894  // reference that register in order to be considered. This prunes out
4895  // unprofitable searching.
4897  for (const SCEV *S : CurRegs)
4898  if (LU.Regs.count(S))
4899  ReqRegs.insert(S);
4900 
4902  Cost NewCost(L, SE, TTI);
4903  for (const Formula &F : LU.Formulae) {
4904  // Ignore formulae which may not be ideal in terms of register reuse of
4905  // ReqRegs. The formula should use all required registers before
4906  // introducing new ones.
4907  int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4908  for (const SCEV *Reg : ReqRegs) {
4909  if ((F.ScaledReg && F.ScaledReg == Reg) ||
4910  is_contained(F.BaseRegs, Reg)) {
4911  --NumReqRegsToFind;
4912  if (NumReqRegsToFind == 0)
4913  break;
4914  }
4915  }
4916  if (NumReqRegsToFind != 0) {
4917  // If none of the formulae satisfied the required registers, then we could
4918  // clear ReqRegs and try again. Currently, we simply give up in this case.
4919  continue;
4920  }
4921 
4922  // Evaluate the cost of the current formula. If it's already worse than
4923  // the current best, prune the search at that point.
4924  NewCost = CurCost;
4925  NewRegs = CurRegs;
4926  NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
4927  if (NewCost.isLess(SolutionCost)) {
4928  Workspace.push_back(&F);
4929  if (Workspace.size() != Uses.size()) {
4930  SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4931  NewRegs, VisitedRegs);
4932  if (F.getNumRegs() == 1 && Workspace.size() == 1)
4933  VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4934  } else {
4935  LLVM_DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4936  dbgs() << ".\nRegs:\n";
4937  for (const SCEV *S : NewRegs) dbgs()
4938  << "- " << *S << "\n";
4939  dbgs() << '\n');
4940 
4941  SolutionCost = NewCost;
4942  Solution = Workspace;
4943  }
4944  Workspace.pop_back();
4945  }
4946  }
4947 }
4948 
4949 /// Choose one formula from each use. Return the results in the given Solution
4950 /// vector.
4951 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4953  Cost SolutionCost(L, SE, TTI);
4954  SolutionCost.Lose();
4955  Cost CurCost(L, SE, TTI);
4957  DenseSet<const SCEV *> VisitedRegs;
4958  Workspace.reserve(Uses.size());
4959 
4960  // SolveRecurse does all the work.
4961  SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4962  CurRegs, VisitedRegs);
4963  if (Solution.empty()) {
4964  LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4965  return;
4966  }
4967 
4968  // Ok, we've now made all our decisions.
4969  LLVM_DEBUG(dbgs() << "\n"
4970  "The chosen solution requires ";
4971  SolutionCost.print(dbgs()); dbgs() << ":\n";
4972  for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4973  dbgs() << " ";
4974  Uses[i].print(dbgs());
4975  dbgs() << "\n"
4976  " ";
4977  Solution[i]->print(dbgs());
4978  dbgs() << '\n';
4979  });
4980 
4981  assert(Solution.size() == Uses.size() && "Malformed solution!");
4982 }
4983 
4984 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4985 /// we can go while still being dominated by the input positions. This helps
4986 /// canonicalize the insert position, which encourages sharing.
4988 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4989  const SmallVectorImpl<Instruction *> &Inputs)
4990  const {
4991  Instruction *Tentative = &*IP;
4992  while (true) {
4993  bool AllDominate = true;
4994  Instruction *BetterPos = nullptr;
4995  // Don't bother attempting to insert before a catchswitch, their basic block
4996  // cannot have other non-PHI instructions.
4997  if (isa<CatchSwitchInst>(Tentative))
4998  return IP;
4999 
5000  for (Instruction *Inst : Inputs) {
5001  if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5002  AllDominate = false;
5003  break;
5004  }
5005  // Attempt to find an insert position in the middle of the block,
5006  // instead of at the end, so that it can be used for other expansions.
5007  if (Tentative->getParent() == Inst->getParent() &&
5008  (!BetterPos || !DT.dominates(Inst, BetterPos)))
5009  BetterPos = &*std::next(BasicBlock::iterator(Inst));
5010  }
5011  if (!AllDominate)
5012  break;
5013  if (BetterPos)
5014  IP = BetterPos->getIterator();
5015  else
5016  IP = Tentative->getIterator();
5017 
5018  const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5019  unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5020 
5021  BasicBlock *IDom;
5022  for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5023  if (!Rung) return IP;
5024  Rung = Rung->getIDom();
5025  if (!Rung) return IP;
5026  IDom = Rung->getBlock();
5027 
5028  // Don't climb into a loop though.
5029  const Loop *IDomLoop = LI.getLoopFor(IDom);
5030  unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5031  if (IDomDepth <= IPLoopDepth &&
5032  (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5033  break;
5034  }
5035 
5036  Tentative = IDom->getTerminator();
5037  }
5038 
5039  return IP;
5040 }
5041 
5042 /// Determine an input position which will be dominated by the operands and
5043 /// which will dominate the result.
5045 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
5046  const LSRFixup &LF,
5047  const LSRUse &LU,
5048  SCEVExpander &Rewriter) const {
5049  // Collect some instructions which must be dominated by the
5050  // expanding replacement. These must be dominated by any operands that
5051  // will be required in the expansion.
5053  if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5054  Inputs.push_back(I);
5055  if (LU.Kind == LSRUse::ICmpZero)
5056  if (Instruction *I =
5057  dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5058  Inputs.push_back(I);
5059  if (LF.PostIncLoops.count(L)) {
5060  if (LF.isUseFullyOutsideLoop(L))
5061  Inputs.push_back(L->getLoopLatch()->getTerminator());
5062  else
5063  Inputs.push_back(IVIncInsertPos);
5064  }
5065  // The expansion must also be dominated by the increment positions of any
5066  // loops it for which it is using post-inc mode.
5067  for (const Loop *PIL : LF.PostIncLoops) {
5068  if (PIL == L) continue;
5069 
5070  // Be dominated by the loop exit.
5071  SmallVector<BasicBlock *, 4> ExitingBlocks;
5072  PIL->getExitingBlocks(ExitingBlocks);
5073  if (!ExitingBlocks.empty()) {
5074  BasicBlock *BB = ExitingBlocks[0];
5075  for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5076  BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5077  Inputs.push_back(BB->getTerminator());
5078  }
5079  }
5080 
5081  assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
5082  && !isa<DbgInfoIntrinsic>(LowestIP) &&
5083  "Insertion point must be a normal instruction");
5084 
5085  // Then, climb up the immediate dominator tree as far as we can go while
5086  // still being dominated by the input positions.
5087  BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5088 
5089  // Don't insert instructions before PHI nodes.
5090  while (isa<PHINode>(IP)) ++IP;
5091 
5092  // Ignore landingpad instructions.
5093  while (IP->isEHPad()) ++IP;
5094 
5095  // Ignore debug intrinsics.
5096  while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5097 
5098  // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5099  // IP consistent across expansions and allows the previously inserted
5100  // instructions to be reused by subsequent expansion.
5101  while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5102  ++IP;
5103 
5104  return IP;
5105 }
5106 
5107 /// Emit instructions for the leading candidate expression for this LSRUse (this
5108 /// is called "expanding").
5109 Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5110  const Formula &F, BasicBlock::iterator IP,
5111  SCEVExpander &Rewriter,
5112  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5113  if (LU.RigidFormula)
5114  return LF.OperandValToReplace;
5115 
5116  // Determine an input position which will be dominated by the operands and
5117  // which will dominate the result.
5118  IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
5119  Rewriter.setInsertPoint(&*IP);
5120 
5121  // Inform the Rewriter if we have a post-increment use, so that it can
5122  // perform an advantageous expansion.
5123  Rewriter.setPostInc(LF.PostIncLoops);
5124 
5125  // This is the type that the user actually needs.
5126  Type *OpTy = LF.OperandValToReplace->getType();
5127  // This will be the type that we'll initially expand to.
5128  Type *Ty = F.getType();
5129  if (!Ty)
5130  // No type known; just expand directly to the ultimate type.
5131  Ty = OpTy;
5132  else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5133  // Expand directly to the ultimate type if it's the right size.
5134  Ty = OpTy;
5135  // This is the type to do integer arithmetic in.
5136  Type *IntTy = SE.getEffectiveSCEVType(Ty);
5137 
5138  // Build up a list of operands to add together to form the full base.
5140 
5141  // Expand the BaseRegs portion.
5142  for (const SCEV *Reg : F.BaseRegs) {
5143  assert(!Reg->isZero() && "Zero allocated in a base register!");
5144 
5145  // If we're expanding for a post-inc user, make the post-inc adjustment.
5146  Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5147  Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5148  }
5149 
5150  // Expand the ScaledReg portion.
5151  Value *ICmpScaledV = nullptr;
5152  if (F.Scale != 0) {
5153  const SCEV *ScaledS = F.ScaledReg;
5154 
5155  // If we're expanding for a post-inc user, make the post-inc adjustment.
5156  PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5157  ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5158 
5159  if (LU.Kind == LSRUse::ICmpZero) {
5160  // Expand ScaleReg as if it was part of the base regs.
5161  if (F.Scale == 1)
5162  Ops.push_back(
5163  SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5164  else {
5165  // An interesting way of "folding" with an icmp is to use a negated
5166  // scale, which we'll implement by inserting it into the other operand
5167  // of the icmp.
5168  assert(F.Scale == -1 &&
5169  "The only scale supported by ICmpZero uses is -1!");
5170  ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5171  }
5172  } else {
5173  // Otherwise just expand the scaled register and an explicit scale,
5174  // which is expected to be matched as part of the address.
5175 
5176  // Flush the operand list to suppress SCEVExpander hoisting address modes.
5177  // Unless the addressing mode will not be folded.
5178  if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5179  isAMCompletelyFolded(TTI, LU, F)) {
5180  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5181  Ops.clear();
5182  Ops.push_back(SE.getUnknown(FullV));
5183  }
5184  ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5185  if (F.Scale != 1)
5186  ScaledS =
5187  SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5188  Ops.push_back(ScaledS);
5189  }
5190  }
5191 
5192  // Expand the GV portion.
5193  if (F.BaseGV) {
5194  // Flush the operand list to suppress SCEVExpander hoisting.
5195  if (!Ops.empty()) {
5196  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5197  Ops.clear();
5198  Ops.push_back(SE.getUnknown(FullV));
5199  }
5200  Ops.push_back(SE.getUnknown(F.BaseGV));
5201  }
5202 
5203  // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5204  // unfolded offsets. LSR assumes they both live next to their uses.
5205  if (!Ops.empty()) {
5206  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5207  Ops.clear();
5208  Ops.push_back(SE.getUnknown(FullV));
5209  }
5210 
5211  // Expand the immediate portion.
5212  int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5213  if (Offset != 0) {
5214  if (LU.Kind == LSRUse::ICmpZero) {
5215  // The other interesting way of "folding" with an ICmpZero is to use a
5216  // negated immediate.
5217  if (!ICmpScaledV)
5218  ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5219  else {
5220  Ops.push_back(SE.getUnknown(ICmpScaledV));
5221  ICmpScaledV = ConstantInt::get(IntTy, Offset);
5222  }
5223  } else {
5224  // Just add the immediate values. These again are expected to be matched
5225  // as part of the address.
5226  Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5227  }
5228  }
5229 
5230  // Expand the unfolded offset portion.
5231  int64_t UnfoldedOffset = F.UnfoldedOffset;
5232  if (UnfoldedOffset != 0) {
5233  // Just add the immediate values.
5235  UnfoldedOffset)));
5236  }
5237 
5238  // Emit instructions summing all the operands.
5239  const SCEV *FullS = Ops.empty() ?
5240  SE.getConstant(IntTy, 0) :
5241  SE.getAddExpr(Ops);
5242  Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5243 
5244  // We're done expanding now, so reset the rewriter.
5245  Rewriter.clearPostInc();
5246 
5247  // An ICmpZero Formula represents an ICmp which we're handling as a
5248  // comparison against zero. Now that we've expanded an expression for that
5249  // form, update the ICmp's other operand.
5250  if (LU.Kind == LSRUse::ICmpZero) {
5251  ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5252  DeadInsts.emplace_back(CI->getOperand(1));
5253  assert(!F.BaseGV && "ICmp does not support folding a global value and "
5254  "a scale at the same time!");
5255  if (F.Scale == -1) {
5256  if (ICmpScaledV->getType() != OpTy) {
5257  Instruction *Cast =
5258  CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5259  OpTy, false),
5260  ICmpScaledV, OpTy, "tmp", CI);
5261  ICmpScaledV = Cast;
5262  }
5263  CI->setOperand(1, ICmpScaledV);
5264  } else {
5265  // A scale of 1 means that the scale has been expanded as part of the
5266  // base regs.
5267  assert((F.Scale == 0 || F.Scale == 1) &&
5268  "ICmp does not support folding a global value and "
5269  "a scale at the same time!");
5271  -(uint64_t)Offset);
5272  if (C->getType() != OpTy)
5274  OpTy, false),
5275  C, OpTy);
5276 
5277  CI->setOperand(1, C);
5278  }
5279  }
5280 
5281  return FullV;
5282 }
5283 
5284 /// Helper for Rewrite. PHI nodes are special because the use of their operands
5285 /// effectively happens in their predecessor blocks, so the expression may need
5286 /// to be expanded in multiple places.
5287 void LSRInstance::RewriteForPHI(
5288  PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5289  SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5291  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5292  if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5293  bool needUpdateFixups = false;
5294  BasicBlock *BB = PN->getIncomingBlock(i);
5295 
5296  // If this is a critical edge, split the edge so that we do not insert
5297  // the code on all predecessor/successor paths. We do this unless this
5298  // is the canonical backedge for this loop, which complicates post-inc
5299  // users.
5300  if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5301  !isa<IndirectBrInst>(BB->getTerminator()) &&
5302  !isa<CatchSwitchInst>(BB->getTerminator())) {
5303  BasicBlock *Parent = PN->getParent();
5304  Loop *PNLoop = LI.getLoopFor(Parent);
5305  if (!PNLoop || Parent != PNLoop->getHeader()) {
5306  // Split the critical edge.
5307  BasicBlock *NewBB = nullptr;
5308  if (!Parent->isLandingPad()) {
5309  NewBB = SplitCriticalEdge(BB, Parent,
5311  .setMergeIdenticalEdges()
5312  .setKeepOneInputPHIs());
5313  } else {
5315  SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5316  NewBB = NewBBs[0];
5317  }
5318  // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5319  // phi predecessors are identical. The simple thing to do is skip
5320  // splitting in this case rather than complicate the API.
5321  if (NewBB) {
5322  // If PN is outside of the loop and BB is in the loop, we want to
5323  // move the block to be immediately before the PHI block, not
5324  // immediately after BB.
5325  if (L->contains(BB) && !L->contains(PN))
5326  NewBB->moveBefore(PN->getParent());
5327 
5328  // Splitting the edge can reduce the number of PHI entries we have.
5329  e = PN->getNumIncomingValues();
5330  BB = NewBB;
5331  i = PN->getBasicBlockIndex(BB);
5332 
5333  needUpdateFixups = true;
5334  }
5335  }
5336  }
5337 
5338  std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5339  Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5340  if (!Pair.second)
5341  PN->setIncomingValue(i, Pair.first->second);
5342  else {
5343  Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5344  Rewriter, DeadInsts);
5345 
5346  // If this is reuse-by-noop-cast, insert the noop cast.
5347  Type *OpTy = LF.OperandValToReplace->getType();
5348  if (FullV->getType() != OpTy)
5349  FullV =
5351  OpTy, false),
5352  FullV, LF.OperandValToReplace->getType(),
5353  "tmp", BB->getTerminator());
5354 
5355  PN->setIncomingValue(i, FullV);
5356  Pair.first->second = FullV;
5357  }
5358 
5359  // If LSR splits critical edge and phi node has other pending
5360  // fixup operands, we need to update those pending fixups. Otherwise
5361  // formulae will not be implemented completely and some instructions
5362  // will not be eliminated.
5363  if (needUpdateFixups) {
5364  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5365  for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
5366  // If fixup is supposed to rewrite some operand in the phi
5367  // that was just updated, it may be already moved to
5368  // another phi node. Such fixup requires update.
5369  if (Fixup.UserInst == PN) {
5370  // Check if the operand we try to replace still exists in the
5371  // original phi.
5372  bool foundInOriginalPHI = false;
5373  for (const auto &val : PN->incoming_values())
5374  if (val == Fixup.OperandValToReplace) {
5375  foundInOriginalPHI = true;
5376  break;
5377  }
5378 
5379  // If fixup operand found in original PHI - nothing to do.
5380  if (foundInOriginalPHI)
5381  continue;
5382 
5383  // Otherwise it might be moved to another PHI and requires update.
5384  // If fixup operand not found in any of the incoming blocks that
5385  // means we have already rewritten it - nothing to do.
5386  for (const auto &Block : PN->blocks())
5387  for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5388  ++I) {
5389  PHINode *NewPN = cast<PHINode>(I);
5390  for (const auto &val : NewPN->incoming_values())
5391  if (val == Fixup.OperandValToReplace)
5392  Fixup.UserInst = NewPN;
5393  }
5394  }
5395  }
5396  }
5397 }
5398 
5399 /// Emit instructions for the leading candidate expression for this LSRUse (this
5400 /// is called "expanding"), and update the UserInst to reference the newly
5401 /// expanded value.
5402 void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5403  const Formula &F, SCEVExpander &Rewriter,
5404  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5405  // First, find an insertion point that dominates UserInst. For PHI nodes,
5406  // find the nearest block which dominates all the relevant uses.
5407  if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5408  RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5409  } else {
5410  Value *FullV =
5411  Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5412 
5413  // If this is reuse-by-noop-cast, insert the noop cast.
5414  Type *OpTy = LF.OperandValToReplace->getType();
5415  if (FullV->getType() != OpTy) {
5416  Instruction *Cast =
5417  CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5418  FullV, OpTy, "tmp", LF.UserInst);
5419  FullV = Cast;
5420  }
5421 
5422  // Update the user. ICmpZero is handled specially here (for now) because
5423  // Expand may have updated one of the operands of the icmp already, and
5424  // its new value may happen to be equal to LF.OperandValToReplace, in
5425  // which case doing replaceUsesOfWith leads to replacing both operands
5426  // with the same value. TODO: Reorganize this.
5427  if (LU.Kind == LSRUse::ICmpZero)
5428  LF.UserInst->setOperand(0, FullV);
5429  else
5430  LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5431  }
5432 
5433  DeadInsts.emplace_back(LF.OperandValToReplace);
5434 }
5435 
5436 /// Rewrite all the fixup locations with new values, following the chosen
5437 /// solution.
5438 void LSRInstance::ImplementSolution(
5439  const SmallVectorImpl<const Formula *> &Solution) {
5440  // Keep track of instructions we may have made dead, so that
5441  // we can remove them after we are done working.
5443 
5445  "lsr");
5446 #ifndef NDEBUG
5447  Rewriter.setDebugType(DEBUG_TYPE);
5448 #endif
5449  Rewriter.disableCanonicalMode();
5450  Rewriter.enableLSRMode();
5451  Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5452 
5453  // Mark phi nodes that terminate chains so the expander tries to reuse them.
5454  for (const IVChain &Chain : IVChainVec) {
5455  if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5456  Rewriter.setChainedPhi(PN);
5457  }
5458 
5459  // Expand the new value definitions and update the users.
5460  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5461  for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5462  Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5463  Changed = true;
5464  }
5465 
5466  for (const IVChain &Chain : IVChainVec) {
5467  GenerateIVChain(Chain, Rewriter, DeadInsts);
5468  Changed = true;
5469  }
5470  // Clean up after ourselves. This must be done before deleting any
5471  // instructions.
5472  Rewriter.clear();
5473 
5474  Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5475 }
5476 
5477 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5478  DominatorTree &DT, LoopInfo &LI,
5479  const TargetTransformInfo &TTI)
5480  : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L),
5481  FavorBackedgeIndex(EnableBackedgeIndexing &&
5482  TTI.shouldFavorBackedgeIndex(L)) {
5483  // If LoopSimplify form is not available, stay out of trouble.
5484  if (!L->isLoopSimplifyForm())
5485  return;
5486 
5487  // If there's no interesting work to be done, bail early.
5488  if (IU.empty()) return;
5489 
5490  // If there's too much analysis to be done, bail early. We won't be able to
5491  // model the problem anyway.
5492  unsigned NumUsers = 0;
5493  for (const IVStrideUse &U : IU) {
5494  if (++NumUsers > MaxIVUsers) {
5495  (void)U;
5496  LLVM_DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U
5497  << "\n");
5498  return;
5499  }
5500  // Bail out if we have a PHI on an EHPad that gets a value from a
5501  // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5502  // no good place to stick any instructions.
5503  if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5504  auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5505  if (isa<FuncletPadInst>(FirstNonPHI) ||
5506  isa<CatchSwitchInst>(FirstNonPHI))
5507  for (BasicBlock *PredBB : PN->blocks())
5508  if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5509  return;
5510  }
5511  }
5512 
5513 #ifndef NDEBUG
5514  // All dominating loops must have preheaders, or SCEVExpander may not be able
5515  // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5516  //
5517  // IVUsers analysis should only create users that are dominated by simple loop
5518  // headers. Since this loop should dominate all of its users, its user list
5519  // should be empty if this loop itself is not within a simple loop nest.
5520  for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5521  Rung; Rung = Rung->getIDom()) {
5522  BasicBlock *BB = Rung->getBlock();
5523  const Loop *DomLoop = LI.getLoopFor(BB);
5524  if (DomLoop && DomLoop->getHeader() == BB) {
5525  assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
5526  }
5527  }
5528 #endif // DEBUG
5529 
5530  LLVM_DEBUG(dbgs() << "\nLSR on loop ";
5531  L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
5532  dbgs() << ":\n");
5533 
5534  // First, perform some low-level loop optimizations.
5535  OptimizeShadowIV();
5536  OptimizeLoopTermCond();
5537 
5538  // If loop preparation eliminates all interesting IV users, bail.
5539  if (IU.empty()) return;
5540 
5541  // Skip nested loops until we can model them better with formulae.
5542  if (!L->empty()) {
5543  LLVM_DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
5544  return;
5545  }
5546 
5547  // Start collecting data and preparing for the solver.
5548  CollectChains();
5549  CollectInterestingTypesAndFactors();
5550  CollectFixupsAndInitialFormulae();
5551  CollectLoopInvariantFixupsAndFormulae();
5552 
5553  if (Uses.empty())
5554  return;
5555 
5556  LLVM_DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
5557  print_uses(dbgs()));
5558 
5559  // Now use the reuse data to generate a bunch of interesting ways
5560  // to formulate the values needed for the uses.
5561  GenerateAllReuseFormulae();
5562 
5563  FilterOutUndesirableDedicatedRegisters();
5564  NarrowSearchSpaceUsingHeuristics();
5565 
5567  Solve(Solution);
5568 
5569  // Release memory that is no longer needed.
5570  Factors.clear();
5571  Types.clear();
5572  RegUses.clear();
5573 
5574  if (Solution.empty())
5575  return;
5576 
5577 #ifndef NDEBUG
5578  // Formulae should be legal.
5579  for (const LSRUse &LU : Uses) {
5580  for (const Formula &F : LU.Formulae)
5581  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
5582  F) && "Illegal formula generated!");
5583  };
5584 #endif
5585 
5586  // Now that we've decided what we want, make it so.
5587  ImplementSolution(Solution);
5588 }
5589 
5590 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5591 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5592  if (Factors.empty() && Types.empty()) return;
5593 
5594  OS << "LSR has identified the following interesting factors and types: ";
5595  bool First = true;
5596 
5597  for (int64_t Factor : Factors) {
5598  if (!First) OS << ", ";
5599  First = false;
5600  OS << '*' << Factor;
5601  }
5602 
5603  for (Type *Ty : Types) {
5604  if (!First) OS << ", ";
5605  First = false;
5606  OS << '(' << *Ty << ')';
5607  }
5608  OS << '\n';
5609 }
5610 
5611 void LSRInstance::print_fixups(raw_ostream &OS) const {
5612  OS << "LSR is examining the following fixup sites:\n";
5613  for (const LSRUse &LU : Uses)
5614  for (const LSRFixup &LF : LU.Fixups) {
5615  dbgs() << " ";
5616  LF.print(OS);
5617  OS << '\n';
5618  }
5619 }
5620 
5621 void LSRInstance::print_uses(raw_ostream &OS) const {
5622  OS << "LSR is examining the following uses:\n";
5623  for (const LSRUse &LU : Uses) {
5624  dbgs() << " ";
5625  LU.print(OS);
5626  OS << '\n';
5627  for (const Formula &F : LU.Formulae) {
5628  OS << " ";
5629  F.print(OS);
5630  OS << '\n';
5631  }
5632  }
5633 }
5634 
5635 void LSRInstance::print(raw_ostream &OS) const {
5636  print_factors_and_types(OS);
5637  print_fixups(OS);
5638  print_uses(OS);
5639 }
5640 
5641 LLVM_DUMP_METHOD void LSRInstance::dump() const {
5642  print(errs()); errs() << '\n';
5643 }
5644 #endif
5645 
5646 namespace {
5647 
5648 class LoopStrengthReduce : public LoopPass {
5649 public:
5650  static char ID; // Pass ID, replacement for typeid
5651 
5652  LoopStrengthReduce();
5653 
5654 private:
5655  bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5656  void getAnalysisUsage(AnalysisUsage &AU) const override;
5657 };
5658 
5659 } // end anonymous namespace
5660 
5661 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5663 }
5664 
5665 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5666  // We split critical edges, so we change the CFG. However, we do update
5667  // many analyses if they are around.
5669 
5677  // Requiring LoopSimplify a second time here prevents IVUsers from running
5678  // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5683 }
5684 
5686  DominatorTree &DT, LoopInfo &LI,
5687  const TargetTransformInfo &TTI) {
5688  bool Changed = false;
5689 
5690  // Run the main LSR transformation.
5691  Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5692 
5693  // Remove any extra phis created by processing inner loops.
5694  Changed |= DeleteDeadPHIs(L->getHeader());
5695  if (EnablePhiElim && L->isLoopSimplifyForm()) {
5697  const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5698  SCEVExpander Rewriter(SE, DL, "lsr");
5699 #ifndef NDEBUG
5700  Rewriter.setDebugType(DEBUG_TYPE);
5701 #endif
5702  unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5703  if (numFolded) {
5704  Changed = true;
5706  DeleteDeadPHIs(L->getHeader());
5707  }
5708  }
5709  return Changed;
5710 }
5711 
5712 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5713  if (skipLoop(L))
5714  return false;
5715 
5716  auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5717  auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5718  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5719  auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5720  const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5721  *L->getHeader()->getParent());
5722  return ReduceLoopStrength(L, IU, SE, DT, LI, TTI);
5723 }
5724 
5727  LPMUpdater &) {
5728  if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5729  AR.DT, AR.LI, AR.TTI))
5730  return PreservedAnalyses::all();
5731 
5733 }
5734 
5735 char LoopStrengthReduce::ID = 0;
5736 
5737 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5738  "Loop Strength Reduction", false, false)
5744 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5745 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5746  "Loop Strength Reduction", false, false)
5747 
5748 Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }
Pass interface - Implemented by all &#39;passes&#39;.
Definition: Pass.h:80
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:1276
uint64_t CallInst * C
bool isIndexedStoreLegal(enum MemIndexedMode Mode, Type *Ty) const
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:67
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:233
reference emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:645
Definition: Any.h:26
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
iterator_range< use_iterator > uses()
Definition: Value.h:354
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass