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-setupcost-depth-limit", cl::Hidden, cl::init(7),
169  cl::desc("The limit on recursion depth for LSRs setup cost"));
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, unsigned Depth) {
1216  if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1217  return 1;
1218  if (Depth == 0)
1219  return 0;
1220  if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1221  return getSetupCost(S->getStart(), Depth - 1);
1222  if (auto S = dyn_cast<SCEVCastExpr>(Reg))
1223  return getSetupCost(S->getOperand(), Depth - 1);
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, Depth - 1);
1228  });
1229  if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1230  return getSetupCost(S->getLHS(), Depth - 1) +
1231  getSetupCost(S->getRHS(), Depth - 1);
1232  return 0;
1233 }
1234 
1235 /// Tally up interesting quantities from the given register.
1236 void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1238  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1239  // If this is an addrec for another loop, it should be an invariant
1240  // with respect to L since L is the innermost loop (at least
1241  // for now LSR only handles innermost loops).
1242  if (AR->getLoop() != L) {
1243  // If the AddRec exists, consider it's register free and leave it alone.
1244  if (isExistingPhi(AR, *SE))
1245  return;
1246 
1247  // It is bad to allow LSR for current loop to add induction variables
1248  // for its sibling loops.
1249  if (!AR->getLoop()->contains(L)) {
1250  Lose();
1251  return;
1252  }
1253 
1254  // Otherwise, it will be an invariant with respect to Loop L.
1255  ++C.NumRegs;
1256  return;
1257  }
1258 
1259  unsigned LoopCost = 1;
1260  if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1261  TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1262 
1263  // If the step size matches the base offset, we could use pre-indexed
1264  // addressing.
1265  if (TTI->shouldFavorBackedgeIndex(L)) {
1266  if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1267  if (Step->getAPInt() == F.BaseOffset)
1268  LoopCost = 0;
1269  }
1270 
1271  if (TTI->shouldFavorPostInc()) {
1272  const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1273  if (isa<SCEVConstant>(LoopStep)) {
1274  const SCEV *LoopStart = AR->getStart();
1275  if (!isa<SCEVConstant>(LoopStart) &&
1276  SE->isLoopInvariant(LoopStart, L))
1277  LoopCost = 0;
1278  }
1279  }
1280  }
1281  C.AddRecCost += LoopCost;
1282 
1283  // Add the step value register, if it needs one.
1284  // TODO: The non-affine case isn't precisely modeled here.
1285  if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1286  if (!Regs.count(AR->getOperand(1))) {
1287  RateRegister(F, AR->getOperand(1), Regs);
1288  if (isLoser())
1289  return;
1290  }
1291  }
1292  }
1293  ++C.NumRegs;
1294 
1295  // Rough heuristic; favor registers which don't require extra setup
1296  // instructions in the preheader.
1297  C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
1298  // Ensure we don't, even with the recusion limit, produce invalid costs.
1299  C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);
1300 
1301  C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1302  SE->hasComputableLoopEvolution(Reg, L);
1303 }
1304 
1305 /// Record this register in the set. If we haven't seen it before, rate
1306 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1307 /// one of those regs an instant loser.
1308 void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1310  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1311  if (LoserRegs && LoserRegs->count(Reg)) {
1312  Lose();
1313  return;
1314  }
1315  if (Regs.insert(Reg).second) {
1316  RateRegister(F, Reg, Regs);
1317  if (LoserRegs && isLoser())
1318  LoserRegs->insert(Reg);
1319  }
1320 }
1321 
1322 void Cost::RateFormula(const Formula &F,
1324  const DenseSet<const SCEV *> &VisitedRegs,
1325  const LSRUse &LU,
1326  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1327  assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1328  // Tally up the registers.
1329  unsigned PrevAddRecCost = C.AddRecCost;
1330  unsigned PrevNumRegs = C.NumRegs;
1331  unsigned PrevNumBaseAdds = C.NumBaseAdds;
1332  if (const SCEV *ScaledReg = F.ScaledReg) {
1333  if (VisitedRegs.count(ScaledReg)) {
1334  Lose();
1335  return;
1336  }
1337  RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1338  if (isLoser())
1339  return;
1340  }
1341  for (const SCEV *BaseReg : F.BaseRegs) {
1342  if (VisitedRegs.count(BaseReg)) {
1343  Lose();
1344  return;
1345  }
1346  RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1347  if (isLoser())
1348  return;
1349  }
1350 
1351  // Determine how many (unfolded) adds we'll need inside the loop.
1352  size_t NumBaseParts = F.getNumRegs();
1353  if (NumBaseParts > 1)
1354  // Do not count the base and a possible second register if the target
1355  // allows to fold 2 registers.
1356  C.NumBaseAdds +=
1357  NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1358  C.NumBaseAdds += (F.UnfoldedOffset != 0);
1359 
1360  // Accumulate non-free scaling amounts.
1361  C.ScaleCost += getScalingFactorCost(*TTI, LU, F, *L);
1362 
1363  // Tally up the non-zero immediates.
1364  for (const LSRFixup &Fixup : LU.Fixups) {
1365  int64_t O = Fixup.Offset;
1366  int64_t Offset = (uint64_t)O + F.BaseOffset;
1367  if (F.BaseGV)
1368  C.ImmCost += 64; // Handle symbolic values conservatively.
1369  // TODO: This should probably be the pointer size.
1370  else if (Offset != 0)
1371  C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1372 
1373  // Check with target if this offset with this instruction is
1374  // specifically not supported.
1375  if (LU.Kind == LSRUse::Address && Offset != 0 &&
1376  !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1377  Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1378  C.NumBaseAdds++;
1379  }
1380 
1381  // If we don't count instruction cost exit here.
1382  if (!InsnsCost) {
1383  assert(isValid() && "invalid cost");
1384  return;
1385  }
1386 
1387  // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1388  // additional instruction (at least fill).
1389  unsigned TTIRegNum = TTI->getNumberOfRegisters(false) - 1;
1390  if (C.NumRegs > TTIRegNum) {
1391  // Cost already exceeded TTIRegNum, then only newly added register can add
1392  // new instructions.
1393  if (PrevNumRegs > TTIRegNum)
1394  C.Insns += (C.NumRegs - PrevNumRegs);
1395  else
1396  C.Insns += (C.NumRegs - TTIRegNum);
1397  }
1398 
1399  // If ICmpZero formula ends with not 0, it could not be replaced by
1400  // just add or sub. We'll need to compare final result of AddRec.
1401  // That means we'll need an additional instruction. But if the target can
1402  // macro-fuse a compare with a branch, don't count this extra instruction.
1403  // For -10 + {0, +, 1}:
1404  // i = i + 1;
1405  // cmp i, 10
1406  //
1407  // For {-10, +, 1}:
1408  // i = i + 1;
1409  if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1410  !TTI->canMacroFuseCmp())
1411  C.Insns++;
1412  // Each new AddRec adds 1 instruction to calculation.
1413  C.Insns += (C.AddRecCost - PrevAddRecCost);
1414 
1415  // BaseAdds adds instructions for unfolded registers.
1416  if (LU.Kind != LSRUse::ICmpZero)
1417  C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1418  assert(isValid() && "invalid cost");
1419 }
1420 
1421 /// Set this cost to a losing value.
1422 void Cost::Lose() {
1425  C.AddRecCost = std::numeric_limits<unsigned>::max();
1426  C.NumIVMuls = std::numeric_limits<unsigned>::max();
1427  C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1429  C.SetupCost = std::numeric_limits<unsigned>::max();
1430  C.ScaleCost = std::numeric_limits<unsigned>::max();
1431 }
1432 
1433 /// Choose the lower cost.
1434 bool Cost::isLess(Cost &Other) {
1435  if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1436  C.Insns != Other.C.Insns)
1437  return C.Insns < Other.C.Insns;
1438  return TTI->isLSRCostLess(C, Other.C);
1439 }
1440 
1441 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1442 void Cost::print(raw_ostream &OS) const {
1443  if (InsnsCost)
1444  OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1445  OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1446  if (C.AddRecCost != 0)
1447  OS << ", with addrec cost " << C.AddRecCost;
1448  if (C.NumIVMuls != 0)
1449  OS << ", plus " << C.NumIVMuls << " IV mul"
1450  << (C.NumIVMuls == 1 ? "" : "s");
1451  if (C.NumBaseAdds != 0)
1452  OS << ", plus " << C.NumBaseAdds << " base add"
1453  << (C.NumBaseAdds == 1 ? "" : "s");
1454  if (C.ScaleCost != 0)
1455  OS << ", plus " << C.ScaleCost << " scale cost";
1456  if (C.ImmCost != 0)
1457  OS << ", plus " << C.ImmCost << " imm cost";
1458  if (C.SetupCost != 0)
1459  OS << ", plus " << C.SetupCost << " setup cost";
1460 }
1461 
1462 LLVM_DUMP_METHOD void Cost::dump() const {
1463  print(errs()); errs() << '\n';
1464 }
1465 #endif
1466 
1467 /// Test whether this fixup always uses its value outside of the given loop.
1468 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1469  // PHI nodes use their value in their incoming blocks.
1470  if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1471  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1472  if (PN->getIncomingValue(i) == OperandValToReplace &&
1473  L->contains(PN->getIncomingBlock(i)))
1474  return false;
1475  return true;
1476  }
1477 
1478  return !L->contains(UserInst);
1479 }
1480 
1481 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1482 void LSRFixup::print(raw_ostream &OS) const {
1483  OS << "UserInst=";
1484  // Store is common and interesting enough to be worth special-casing.
1485  if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1486  OS << "store ";
1487  Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1488  } else if (UserInst->getType()->isVoidTy())
1489  OS << UserInst->getOpcodeName();
1490  else
1491  UserInst->printAsOperand(OS, /*PrintType=*/false);
1492 
1493  OS << ", OperandValToReplace=";
1494  OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1495 
1496  for (const Loop *PIL : PostIncLoops) {
1497  OS << ", PostIncLoop=";
1498  PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1499  }
1500 
1501  if (Offset != 0)
1502  OS << ", Offset=" << Offset;
1503 }
1504 
1505 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1506  print(errs()); errs() << '\n';
1507 }
1508 #endif
1509 
1510 /// Test whether this use as a formula which has the same registers as the given
1511 /// formula.
1512 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1513  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1514  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1515  // Unstable sort by host order ok, because this is only used for uniquifying.
1516  llvm::sort(Key);
1517  return Uniquifier.count(Key);
1518 }
1519 
1520 /// The function returns a probability of selecting formula without Reg.
1521 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1522  unsigned FNum = 0;
1523  for (const Formula &F : Formulae)
1524  if (F.referencesReg(Reg))
1525  FNum++;
1526  return ((float)(Formulae.size() - FNum)) / Formulae.size();
1527 }
1528 
1529 /// If the given formula has not yet been inserted, add it to the list, and
1530 /// return true. Return false otherwise. The formula must be in canonical form.
1531 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1532  assert(F.isCanonical(L) && "Invalid canonical representation");
1533 
1534  if (!Formulae.empty() && RigidFormula)
1535  return false;
1536 
1537  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1538  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1539  // Unstable sort by host order ok, because this is only used for uniquifying.
1540  llvm::sort(Key);
1541 
1542  if (!Uniquifier.insert(Key).second)
1543  return false;
1544 
1545  // Using a register to hold the value of 0 is not profitable.
1546  assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1547  "Zero allocated in a scaled register!");
1548 #ifndef NDEBUG
1549  for (const SCEV *BaseReg : F.BaseRegs)
1550  assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1551 #endif
1552 
1553  // Add the formula to the list.
1554  Formulae.push_back(F);
1555 
1556  // Record registers now being used by this use.
1557  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1558  if (F.ScaledReg)
1559  Regs.insert(F.ScaledReg);
1560 
1561  return true;
1562 }
1563 
1564 /// Remove the given formula from this use's list.
1565 void LSRUse::DeleteFormula(Formula &F) {
1566  if (&F != &Formulae.back())
1567  std::swap(F, Formulae.back());
1568  Formulae.pop_back();
1569 }
1570 
1571 /// Recompute the Regs field, and update RegUses.
1572 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1573  // Now that we've filtered out some formulae, recompute the Regs set.
1574  SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1575  Regs.clear();
1576  for (const Formula &F : Formulae) {
1577  if (F.ScaledReg) Regs.insert(F.ScaledReg);
1578  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1579  }
1580 
1581  // Update the RegTracker.
1582  for (const SCEV *S : OldRegs)
1583  if (!Regs.count(S))
1584  RegUses.dropRegister(S, LUIdx);
1585 }
1586 
1587 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1588 void LSRUse::print(raw_ostream &OS) const {
1589  OS << "LSR Use: Kind=";
1590  switch (Kind) {
1591  case Basic: OS << "Basic"; break;
1592  case Special: OS << "Special"; break;
1593  case ICmpZero: OS << "ICmpZero"; break;
1594  case Address:
1595  OS << "Address of ";
1596  if (AccessTy.MemTy->isPointerTy())
1597  OS << "pointer"; // the full pointer type could be really verbose
1598  else {
1599  OS << *AccessTy.MemTy;
1600  }
1601 
1602  OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1603  }
1604 
1605  OS << ", Offsets={";
1606  bool NeedComma = false;
1607  for (const LSRFixup &Fixup : Fixups) {
1608  if (NeedComma) OS << ',';
1609  OS << Fixup.Offset;
1610  NeedComma = true;
1611  }
1612  OS << '}';
1613 
1614  if (AllFixupsOutsideLoop)
1615  OS << ", all-fixups-outside-loop";
1616 
1617  if (WidestFixupType)
1618  OS << ", widest fixup type: " << *WidestFixupType;
1619 }
1620 
1621 LLVM_DUMP_METHOD void LSRUse::dump() const {
1622  print(errs()); errs() << '\n';
1623 }
1624 #endif
1625 
1627  LSRUse::KindType Kind, MemAccessTy AccessTy,
1628  GlobalValue *BaseGV, int64_t BaseOffset,
1629  bool HasBaseReg, int64_t Scale,
1630  Instruction *Fixup/*= nullptr*/) {
1631  switch (Kind) {
1632  case LSRUse::Address:
1633  return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1634  HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1635 
1636  case LSRUse::ICmpZero:
1637  // There's not even a target hook for querying whether it would be legal to
1638  // fold a GV into an ICmp.
1639  if (BaseGV)
1640  return false;
1641 
1642  // ICmp only has two operands; don't allow more than two non-trivial parts.
1643  if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1644  return false;
1645 
1646  // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1647  // putting the scaled register in the other operand of the icmp.
1648  if (Scale != 0 && Scale != -1)
1649  return false;
1650 
1651  // If we have low-level target information, ask the target if it can fold an
1652  // integer immediate on an icmp.
1653  if (BaseOffset != 0) {
1654  // We have one of:
1655  // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1656  // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1657  // Offs is the ICmp immediate.
1658  if (Scale == 0)
1659  // The cast does the right thing with
1660  // std::numeric_limits<int64_t>::min().
1661  BaseOffset = -(uint64_t)BaseOffset;
1662  return TTI.isLegalICmpImmediate(BaseOffset);
1663  }
1664 
1665  // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1666  return true;
1667 
1668  case LSRUse::Basic:
1669  // Only handle single-register values.
1670  return !BaseGV && Scale == 0 && BaseOffset == 0;
1671 
1672  case LSRUse::Special:
1673  // Special case Basic to handle -1 scales.
1674  return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1675  }
1676 
1677  llvm_unreachable("Invalid LSRUse Kind!");
1678 }
1679 
1681  int64_t MinOffset, int64_t MaxOffset,
1682  LSRUse::KindType Kind, MemAccessTy AccessTy,
1683  GlobalValue *BaseGV, int64_t BaseOffset,
1684  bool HasBaseReg, int64_t Scale) {
1685  // Check for overflow.
1686  if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1687  (MinOffset > 0))
1688  return false;
1689  MinOffset = (uint64_t)BaseOffset + MinOffset;
1690  if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1691  (MaxOffset > 0))
1692  return false;
1693  MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1694 
1695  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1696  HasBaseReg, Scale) &&
1697  isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1698  HasBaseReg, Scale);
1699 }
1700 
1702  int64_t MinOffset, int64_t MaxOffset,
1703  LSRUse::KindType Kind, MemAccessTy AccessTy,
1704  const Formula &F, const Loop &L) {
1705  // For the purpose of isAMCompletelyFolded either having a canonical formula
1706  // or a scale not equal to zero is correct.
1707  // Problems may arise from non canonical formulae having a scale == 0.
1708  // Strictly speaking it would best to just rely on canonical formulae.
1709  // However, when we generate the scaled formulae, we first check that the
1710  // scaling factor is profitable before computing the actual ScaledReg for
1711  // compile time sake.
1712  assert((F.isCanonical(L) || F.Scale != 0));
1713  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1714  F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1715 }
1716 
1717 /// Test whether we know how to expand the current formula.
1718 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1719  int64_t MaxOffset, LSRUse::KindType Kind,
1720  MemAccessTy AccessTy, GlobalValue *BaseGV,
1721  int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1722  // We know how to expand completely foldable formulae.
1723  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1724  BaseOffset, HasBaseReg, Scale) ||
1725  // Or formulae that use a base register produced by a sum of base
1726  // registers.
1727  (Scale == 1 &&
1728  isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1729  BaseGV, BaseOffset, true, 0));
1730 }
1731 
1732 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1733  int64_t MaxOffset, LSRUse::KindType Kind,
1734  MemAccessTy AccessTy, const Formula &F) {
1735  return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1736  F.BaseOffset, F.HasBaseReg, F.Scale);
1737 }
1738 
1740  const LSRUse &LU, const Formula &F) {
1741  // Target may want to look at the user instructions.
1742  if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1743  for (const LSRFixup &Fixup : LU.Fixups)
1744  if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1745  (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1746  F.Scale, Fixup.UserInst))
1747  return false;
1748  return true;
1749  }
1750 
1751  return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1752  LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1753  F.Scale);
1754 }
1755 
1756 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
1757  const LSRUse &LU, const Formula &F,
1758  const Loop &L) {
1759  if (!F.Scale)
1760  return 0;
1761 
1762  // If the use is not completely folded in that instruction, we will have to
1763  // pay an extra cost only for scale != 1.
1764  if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1765  LU.AccessTy, F, L))
1766  return F.Scale != 1;
1767 
1768  switch (LU.Kind) {
1769  case LSRUse::Address: {
1770  // Check the scaling factor cost with both the min and max offsets.
1771  int ScaleCostMinOffset = TTI.getScalingFactorCost(
1772  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1773  F.Scale, LU.AccessTy.AddrSpace);
1774  int ScaleCostMaxOffset = TTI.getScalingFactorCost(
1775  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1776  F.Scale, LU.AccessTy.AddrSpace);
1777 
1778  assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
1779  "Legal addressing mode has an illegal cost!");
1780  return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1781  }
1782  case LSRUse::ICmpZero:
1783  case LSRUse::Basic:
1784  case LSRUse::Special:
1785  // The use is completely folded, i.e., everything is folded into the
1786  // instruction.
1787  return 0;
1788  }
1789 
1790  llvm_unreachable("Invalid LSRUse Kind!");
1791 }
1792 
1793 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1794  LSRUse::KindType Kind, MemAccessTy AccessTy,
1795  GlobalValue *BaseGV, int64_t BaseOffset,
1796  bool HasBaseReg) {
1797  // Fast-path: zero is always foldable.
1798  if (BaseOffset == 0 && !BaseGV) return true;
1799 
1800  // Conservatively, create an address with an immediate and a
1801  // base and a scale.
1802  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1803 
1804  // Canonicalize a scale of 1 to a base register if the formula doesn't
1805  // already have a base register.
1806  if (!HasBaseReg && Scale == 1) {
1807  Scale = 0;
1808  HasBaseReg = true;
1809  }
1810 
1811  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1812  HasBaseReg, Scale);
1813 }
1814 
1815 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
1816  ScalarEvolution &SE, int64_t MinOffset,
1817  int64_t MaxOffset, LSRUse::KindType Kind,
1818  MemAccessTy AccessTy, const SCEV *S,
1819  bool HasBaseReg) {
1820  // Fast-path: zero is always foldable.
1821  if (S->isZero()) return true;
1822 
1823  // Conservatively, create an address with an immediate and a
1824  // base and a scale.
1825  int64_t BaseOffset = ExtractImmediate(S, SE);
1826  GlobalValue *BaseGV = ExtractSymbol(S, SE);
1827 
1828  // If there's anything else involved, it's not foldable.
1829  if (!S->isZero()) return false;
1830 
1831  // Fast-path: zero is always foldable.
1832  if (BaseOffset == 0 && !BaseGV) return true;
1833 
1834  // Conservatively, create an address with an immediate and a
1835  // base and a scale.
1836  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1837 
1838  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1839  BaseOffset, HasBaseReg, Scale);
1840 }
1841 
1842 namespace {
1843 
1844 /// An individual increment in a Chain of IV increments. Relate an IV user to
1845 /// an expression that computes the IV it uses from the IV used by the previous
1846 /// link in the Chain.
1847 ///
1848 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1849 /// original IVOperand. The head of the chain's IVOperand is only valid during
1850 /// chain collection, before LSR replaces IV users. During chain generation,
1851 /// IncExpr can be used to find the new IVOperand that computes the same
1852 /// expression.
1853 struct IVInc {
1854  Instruction *UserInst;
1855  Value* IVOperand;
1856  const SCEV *IncExpr;
1857 
1858  IVInc(Instruction *U, Value *O, const SCEV *E)
1859  : UserInst(U), IVOperand(O), IncExpr(E) {}
1860 };
1861 
1862 // The list of IV increments in program order. We typically add the head of a
1863 // chain without finding subsequent links.
1864 struct IVChain {
1865  SmallVector<IVInc, 1> Incs;
1866  const SCEV *ExprBase = nullptr;
1867 
1868  IVChain() = default;
1869  IVChain(const IVInc &Head, const SCEV *Base)
1870  : Incs(1, Head), ExprBase(Base) {}
1871 
1872  using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1873 
1874  // Return the first increment in the chain.
1875  const_iterator begin() const {
1876  assert(!Incs.empty());
1877  return std::next(Incs.begin());
1878  }
1879  const_iterator end() const {
1880  return Incs.end();
1881  }
1882 
1883  // Returns true if this chain contains any increments.
1884  bool hasIncs() const { return Incs.size() >= 2; }
1885 
1886  // Add an IVInc to the end of this chain.
1887  void add(const IVInc &X) { Incs.push_back(X); }
1888 
1889  // Returns the last UserInst in the chain.
1890  Instruction *tailUserInst() const { return Incs.back().UserInst; }
1891 
1892  // Returns true if IncExpr can be profitably added to this chain.
1893  bool isProfitableIncrement(const SCEV *OperExpr,
1894  const SCEV *IncExpr,
1895  ScalarEvolution&);
1896 };
1897 
1898 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1899 /// between FarUsers that definitely cross IV increments and NearUsers that may
1900 /// be used between IV increments.
1901 struct ChainUsers {
1903  SmallPtrSet<Instruction*, 4> NearUsers;
1904 };
1905 
1906 /// This class holds state for the main loop strength reduction logic.
1907 class LSRInstance {
1908  IVUsers &IU;
1909  ScalarEvolution &SE;
1910  DominatorTree &DT;
1911  LoopInfo &LI;
1912  const TargetTransformInfo &TTI;
1913  Loop *const L;
1914  bool FavorBackedgeIndex = false;
1915  bool Changed = false;
1916 
1917  /// This is the insert position that the current loop's induction variable
1918  /// increment should be placed. In simple loops, this is the latch block's
1919  /// terminator. But in more complicated cases, this is a position which will
1920  /// dominate all the in-loop post-increment users.
1921  Instruction *IVIncInsertPos = nullptr;
1922 
1923  /// Interesting factors between use strides.
1924  ///
1925  /// We explicitly use a SetVector which contains a SmallSet, instead of the
1926  /// default, a SmallDenseSet, because we need to use the full range of
1927  /// int64_ts, and there's currently no good way of doing that with
1928  /// SmallDenseSet.
1930 
1931  /// Interesting use types, to facilitate truncation reuse.
1933 
1934  /// The list of interesting uses.
1935  mutable SmallVector<LSRUse, 16> Uses;
1936 
1937  /// Track which uses use which register candidates.
1938  RegUseTracker RegUses;
1939 
1940  // Limit the number of chains to avoid quadratic behavior. We don't expect to
1941  // have more than a few IV increment chains in a loop. Missing a Chain falls
1942  // back to normal LSR behavior for those uses.
1943  static const unsigned MaxChains = 8;
1944 
1945  /// IV users can form a chain of IV increments.
1947 
1948  /// IV users that belong to profitable IVChains.
1950 
1951  void OptimizeShadowIV();
1952  bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1953  ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1954  void OptimizeLoopTermCond();
1955 
1956  void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1957  SmallVectorImpl<ChainUsers> &ChainUsersVec);
1958  void FinalizeChain(IVChain &Chain);
1959  void CollectChains();
1960  void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
1961  SmallVectorImpl<WeakTrackingVH> &DeadInsts);
1962 
1963  void CollectInterestingTypesAndFactors();
1964  void CollectFixupsAndInitialFormulae();
1965 
1966  // Support for sharing of LSRUses between LSRFixups.
1968  UseMapTy UseMap;
1969 
1970  bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
1971  LSRUse::KindType Kind, MemAccessTy AccessTy);
1972 
1973  std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
1974  MemAccessTy AccessTy);
1975 
1976  void DeleteUse(LSRUse &LU, size_t LUIdx);
1977 
1978  LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
1979 
1980  void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1981  void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1982  void CountRegisters(const Formula &F, size_t LUIdx);
1983  bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1984 
1985  void CollectLoopInvariantFixupsAndFormulae();
1986 
1987  void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1988  unsigned Depth = 0);
1989 
1990  void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
1991  const Formula &Base, unsigned Depth,
1992  size_t Idx, bool IsScaledReg = false);
1993  void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1994  void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1995  const Formula &Base, size_t Idx,
1996  bool IsScaledReg = false);
1997  void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1998  void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
1999  const Formula &Base,
2000  const SmallVectorImpl<int64_t> &Worklist,
2001  size_t Idx, bool IsScaledReg = false);
2002  void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2003  void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2004  void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2005  void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2006  void GenerateCrossUseConstantOffsets();
2007  void GenerateAllReuseFormulae();
2008 
2009  void FilterOutUndesirableDedicatedRegisters();
2010 
2011  size_t EstimateSearchSpaceComplexity() const;
2012  void NarrowSearchSpaceByDetectingSupersets();
2013  void NarrowSearchSpaceByCollapsingUnrolledCode();
2014  void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2015  void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2016  void NarrowSearchSpaceByDeletingCostlyFormulas();
2017  void NarrowSearchSpaceByPickingWinnerRegs();
2018  void NarrowSearchSpaceUsingHeuristics();
2019 
2020  void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2021  Cost &SolutionCost,
2023  const Cost &CurCost,
2024  const SmallPtrSet<const SCEV *, 16> &CurRegs,
2025  DenseSet<const SCEV *> &VisitedRegs) const;
2026  void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2027 
2029  HoistInsertPosition(BasicBlock::iterator IP,
2030  const SmallVectorImpl<Instruction *> &Inputs) const;
2032  AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2033  const LSRFixup &LF,
2034  const LSRUse &LU,
2035  SCEVExpander &Rewriter) const;
2036 
2037  Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2038  BasicBlock::iterator IP, SCEVExpander &Rewriter,
2039  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2040  void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2041  const Formula &F, SCEVExpander &Rewriter,
2042  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2043  void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2044  SCEVExpander &Rewriter,
2045  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2046  void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2047 
2048 public:
2049  LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2050  LoopInfo &LI, const TargetTransformInfo &TTI);
2051 
2052  bool getChanged() const { return Changed; }
2053 
2054  void print_factors_and_types(raw_ostream &OS) const;
2055  void print_fixups(raw_ostream &OS) const;
2056  void print_uses(raw_ostream &OS) const;
2057  void print(raw_ostream &OS) const;
2058  void dump() const;
2059 };
2060 
2061 } // end anonymous namespace
2062 
2063 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
2064 /// the cast operation.
2065 void LSRInstance::OptimizeShadowIV() {
2066  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2067  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2068  return;
2069 
2070  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2071  UI != E; /* empty */) {
2072  IVUsers::const_iterator CandidateUI = UI;
2073  ++UI;
2074  Instruction *ShadowUse = CandidateUI->getUser();
2075  Type *DestTy = nullptr;
2076  bool IsSigned = false;
2077 
2078  /* If shadow use is a int->float cast then insert a second IV
2079  to eliminate this cast.
2080 
2081  for (unsigned i = 0; i < n; ++i)
2082  foo((double)i);
2083 
2084  is transformed into
2085 
2086  double d = 0.0;
2087  for (unsigned i = 0; i < n; ++i, ++d)
2088  foo(d);
2089  */
2090  if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2091  IsSigned = false;
2092  DestTy = UCast->getDestTy();
2093  }
2094  else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2095  IsSigned = true;
2096  DestTy = SCast->getDestTy();
2097  }
2098  if (!DestTy) continue;
2099 
2100  // If target does not support DestTy natively then do not apply
2101  // this transformation.
2102  if (!TTI.isTypeLegal(DestTy)) continue;
2103 
2104  PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2105  if (!PH) continue;
2106  if (PH->getNumIncomingValues() != 2) continue;
2107 
2108  // If the calculation in integers overflows, the result in FP type will
2109  // differ. So we only can do this transformation if we are guaranteed to not
2110  // deal with overflowing values
2111  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2112  if (!AR) continue;
2113  if (IsSigned && !AR->hasNoSignedWrap()) continue;
2114  if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2115 
2116  Type *SrcTy = PH->getType();
2117  int Mantissa = DestTy->getFPMantissaWidth();
2118  if (Mantissa == -1) continue;
2119  if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2120  continue;
2121 
2122  unsigned Entry, Latch;
2123  if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2124  Entry = 0;
2125  Latch = 1;
2126  } else {
2127  Entry = 1;
2128  Latch = 0;
2129  }
2130 
2132  if (!Init) continue;
2133  Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2134  (double)Init->getSExtValue() :
2135  (double)Init->getZExtValue());
2136 
2137  BinaryOperator *Incr =
2139  if (!Incr) continue;
2140  if (Incr->getOpcode() != Instruction::Add
2141  && Incr->getOpcode() != Instruction::Sub)
2142  continue;
2143 
2144  /* Initialize new IV, double d = 0.0 in above example. */
2145  ConstantInt *C = nullptr;
2146  if (Incr->getOperand(0) == PH)
2147  C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2148  else if (Incr->getOperand(1) == PH)
2149  C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2150  else
2151  continue;
2152 
2153  if (!C) continue;
2154 
2155  // Ignore negative constants, as the code below doesn't handle them
2156  // correctly. TODO: Remove this restriction.
2157  if (!C->getValue().isStrictlyPositive()) continue;
2158 
2159  /* Add new PHINode. */
2160  PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2161 
2162  /* create new increment. '++d' in above example. */
2163  Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2164  BinaryOperator *NewIncr =
2165  BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
2166  Instruction::FAdd : Instruction::FSub,
2167  NewPH, CFP, "IV.S.next.", Incr);
2168 
2169  NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2170  NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2171 
2172  /* Remove cast operation */
2173  ShadowUse->replaceAllUsesWith(NewPH);
2174  ShadowUse->eraseFromParent();
2175  Changed = true;
2176  break;
2177  }
2178 }
2179 
2180 /// If Cond has an operand that is an expression of an IV, set the IV user and
2181 /// stride information and return true, otherwise return false.
2182 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2183  for (IVStrideUse &U : IU)
2184  if (U.getUser() == Cond) {
2185  // NOTE: we could handle setcc instructions with multiple uses here, but
2186  // InstCombine does it as well for simple uses, it's not clear that it
2187  // occurs enough in real life to handle.
2188  CondUse = &U;
2189  return true;
2190  }
2191  return false;
2192 }
2193 
2194 /// Rewrite the loop's terminating condition if it uses a max computation.
2195 ///
2196 /// This is a narrow solution to a specific, but acute, problem. For loops
2197 /// like this:
2198 ///
2199 /// i = 0;
2200 /// do {
2201 /// p[i] = 0.0;
2202 /// } while (++i < n);
2203 ///
2204 /// the trip count isn't just 'n', because 'n' might not be positive. And
2205 /// unfortunately this can come up even for loops where the user didn't use
2206 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2207 /// will commonly be lowered like this:
2208 ///
2209 /// if (n > 0) {
2210 /// i = 0;
2211 /// do {
2212 /// p[i] = 0.0;
2213 /// } while (++i < n);
2214 /// }
2215 ///
2216 /// and then it's possible for subsequent optimization to obscure the if
2217 /// test in such a way that indvars can't find it.
2218 ///
2219 /// When indvars can't find the if test in loops like this, it creates a
2220 /// max expression, which allows it to give the loop a canonical
2221 /// induction variable:
2222 ///
2223 /// i = 0;
2224 /// max = n < 1 ? 1 : n;
2225 /// do {
2226 /// p[i] = 0.0;
2227 /// } while (++i != max);
2228 ///
2229 /// Canonical induction variables are necessary because the loop passes
2230 /// are designed around them. The most obvious example of this is the
2231 /// LoopInfo analysis, which doesn't remember trip count values. It
2232 /// expects to be able to rediscover the trip count each time it is
2233 /// needed, and it does this using a simple analysis that only succeeds if
2234 /// the loop has a canonical induction variable.
2235 ///
2236 /// However, when it comes time to generate code, the maximum operation
2237 /// can be quite costly, especially if it's inside of an outer loop.
2238 ///
2239 /// This function solves this problem by detecting this type of loop and
2240 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2241 /// the instructions for the maximum computation.
2242 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2243  // Check that the loop matches the pattern we're looking for.
2244  if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2245  Cond->getPredicate() != CmpInst::ICMP_NE)
2246  return Cond;
2247 
2248  SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2249  if (!Sel || !Sel->hasOneUse()) return Cond;
2250 
2251  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2252  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2253  return Cond;
2254  const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2255 
2256  // Add one to the backedge-taken count to get the trip count.
2257  const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2258  if (IterationCount != SE.getSCEV(Sel)) return Cond;
2259 
2260  // Check for a max calculation that matches the pattern. There's no check
2261  // for ICMP_ULE here because the comparison would be with zero, which
2262  // isn't interesting.
2264  const SCEVNAryExpr *Max = nullptr;
2265  if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2266  Pred = ICmpInst::ICMP_SLE;
2267  Max = S;
2268  } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2269  Pred = ICmpInst::ICMP_SLT;
2270  Max = S;
2271  } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2272  Pred = ICmpInst::ICMP_ULT;
2273  Max = U;
2274  } else {
2275  // No match; bail.
2276  return Cond;
2277  }
2278 
2279  // To handle a max with more than two operands, this optimization would
2280  // require additional checking and setup.
2281  if (Max->getNumOperands() != 2)
2282  return Cond;
2283 
2284  const SCEV *MaxLHS = Max->getOperand(0);
2285  const SCEV *MaxRHS = Max->getOperand(1);
2286 
2287  // ScalarEvolution canonicalizes constants to the left. For < and >, look
2288  // for a comparison with 1. For <= and >=, a comparison with zero.
2289  if (!MaxLHS ||
2290  (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2291  return Cond;
2292 
2293  // Check the relevant induction variable for conformance to
2294  // the pattern.
2295  const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2296  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2297  if (!AR || !AR->isAffine() ||
2298  AR->getStart() != One ||
2299  AR->getStepRecurrence(SE) != One)
2300  return Cond;
2301 
2302  assert(AR->getLoop() == L &&
2303  "Loop condition operand is an addrec in a different loop!");
2304 
2305  // Check the right operand of the select, and remember it, as it will
2306  // be used in the new comparison instruction.
2307  Value *NewRHS = nullptr;
2308  if (ICmpInst::isTrueWhenEqual(Pred)) {
2309  // Look for n+1, and grab n.
2310  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2311  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2312  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2313  NewRHS = BO->getOperand(0);
2314  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2315  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2316  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2317  NewRHS = BO->getOperand(0);
2318  if (!NewRHS)
2319  return Cond;
2320  } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2321  NewRHS = Sel->getOperand(1);
2322  else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2323  NewRHS = Sel->getOperand(2);
2324  else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2325  NewRHS = SU->getValue();
2326  else
2327  // Max doesn't match expected pattern.
2328  return Cond;
2329 
2330  // Determine the new comparison opcode. It may be signed or unsigned,
2331  // and the original comparison may be either equality or inequality.
2332  if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2333  Pred = CmpInst::getInversePredicate(Pred);
2334 
2335  // Ok, everything looks ok to change the condition into an SLT or SGE and
2336  // delete the max calculation.
2337  ICmpInst *NewCond =
2338  new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2339 
2340  // Delete the max calculation instructions.
2341  Cond->replaceAllUsesWith(NewCond);
2342  CondUse->setUser(NewCond);
2343  Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2344  Cond->eraseFromParent();
2345  Sel->eraseFromParent();
2346  if (Cmp->use_empty())
2347  Cmp->eraseFromParent();
2348  return NewCond;
2349 }
2350 
2351 /// Change loop terminating condition to use the postinc iv when possible.
2352 void
2353 LSRInstance::OptimizeLoopTermCond() {
2355 
2356  // We need a different set of heuristics for rotated and non-rotated loops.
2357  // If a loop is rotated then the latch is also the backedge, so inserting
2358  // post-inc expressions just before the latch is ideal. To reduce live ranges
2359  // it also makes sense to rewrite terminating conditions to use post-inc
2360  // expressions.
2361  //
2362  // If the loop is not rotated then the latch is not a backedge; the latch
2363  // check is done in the loop head. Adding post-inc expressions before the
2364  // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2365  // in the loop body. In this case we do *not* want to use post-inc expressions
2366  // in the latch check, and we want to insert post-inc expressions before
2367  // the backedge.
2368  BasicBlock *LatchBlock = L->getLoopLatch();
2369  SmallVector<BasicBlock*, 8> ExitingBlocks;
2370  L->getExitingBlocks(ExitingBlocks);
2371  if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2372  return LatchBlock != BB;
2373  })) {
2374  // The backedge doesn't exit the loop; treat this as a head-tested loop.
2375  IVIncInsertPos = LatchBlock->getTerminator();
2376  return;
2377  }
2378 
2379  // Otherwise treat this as a rotated loop.
2380  for (BasicBlock *ExitingBlock : ExitingBlocks) {
2381  // Get the terminating condition for the loop if possible. If we
2382  // can, we want to change it to use a post-incremented version of its
2383  // induction variable, to allow coalescing the live ranges for the IV into
2384  // one register value.
2385 
2386  BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2387  if (!TermBr)
2388  continue;
2389  // FIXME: Overly conservative, termination condition could be an 'or' etc..
2390  if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2391  continue;
2392 
2393  // Search IVUsesByStride to find Cond's IVUse if there is one.
2394  IVStrideUse *CondUse = nullptr;
2395  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2396  if (!FindIVUserForCond(Cond, CondUse))
2397  continue;
2398 
2399  // If the trip count is computed in terms of a max (due to ScalarEvolution
2400  // being unable to find a sufficient guard, for example), change the loop
2401  // comparison to use SLT or ULT instead of NE.
2402  // One consequence of doing this now is that it disrupts the count-down
2403  // optimization. That's not always a bad thing though, because in such
2404  // cases it may still be worthwhile to avoid a max.
2405  Cond = OptimizeMax(Cond, CondUse);
2406 
2407  // If this exiting block dominates the latch block, it may also use
2408  // the post-inc value if it won't be shared with other uses.
2409  // Check for dominance.
2410  if (!DT.dominates(ExitingBlock, LatchBlock))
2411  continue;
2412 
2413  // Conservatively avoid trying to use the post-inc value in non-latch
2414  // exits if there may be pre-inc users in intervening blocks.
2415  if (LatchBlock != ExitingBlock)
2416  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2417  // Test if the use is reachable from the exiting block. This dominator
2418  // query is a conservative approximation of reachability.
2419  if (&*UI != CondUse &&
2420  !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2421  // Conservatively assume there may be reuse if the quotient of their
2422  // strides could be a legal scale.
2423  const SCEV *A = IU.getStride(*CondUse, L);
2424  const SCEV *B = IU.getStride(*UI, L);
2425  if (!A || !B) continue;
2426  if (SE.getTypeSizeInBits(A->getType()) !=
2427  SE.getTypeSizeInBits(B->getType())) {
2428  if (SE.getTypeSizeInBits(A->getType()) >
2429  SE.getTypeSizeInBits(B->getType()))
2430  B = SE.getSignExtendExpr(B, A->getType());
2431  else
2432  A = SE.getSignExtendExpr(A, B->getType());
2433  }
2434  if (const SCEVConstant *D =
2435  dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2436  const ConstantInt *C = D->getValue();
2437  // Stride of one or negative one can have reuse with non-addresses.
2438  if (C->isOne() || C->isMinusOne())
2439  goto decline_post_inc;
2440  // Avoid weird situations.
2441  if (C->getValue().getMinSignedBits() >= 64 ||
2442  C->getValue().isMinSignedValue())
2443  goto decline_post_inc;
2444  // Check for possible scaled-address reuse.
2445  if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2446  MemAccessTy AccessTy = getAccessType(
2447  TTI, UI->getUser(), UI->getOperandValToReplace());
2448  int64_t Scale = C->getSExtValue();
2449  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2450  /*BaseOffset=*/0,
2451  /*HasBaseReg=*/false, Scale,
2452  AccessTy.AddrSpace))
2453  goto decline_post_inc;
2454  Scale = -Scale;
2455  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2456  /*BaseOffset=*/0,
2457  /*HasBaseReg=*/false, Scale,
2458  AccessTy.AddrSpace))
2459  goto decline_post_inc;
2460  }
2461  }
2462  }
2463 
2464  LLVM_DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2465  << *Cond << '\n');
2466 
2467  // It's possible for the setcc instruction to be anywhere in the loop, and
2468  // possible for it to have multiple users. If it is not immediately before
2469  // the exiting block branch, move it.
2470  if (&*++BasicBlock::iterator(Cond) != TermBr) {
2471  if (Cond->hasOneUse()) {
2472  Cond->moveBefore(TermBr);
2473  } else {
2474  // Clone the terminating condition and insert into the loopend.
2475  ICmpInst *OldCond = Cond;
2476  Cond = cast<ICmpInst>(Cond->clone());
2477  Cond->setName(L->getHeader()->getName() + ".termcond");
2478  ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2479 
2480  // Clone the IVUse, as the old use still exists!
2481  CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2482  TermBr->replaceUsesOfWith(OldCond, Cond);
2483  }
2484  }
2485 
2486  // If we get to here, we know that we can transform the setcc instruction to
2487  // use the post-incremented version of the IV, allowing us to coalesce the
2488  // live ranges for the IV correctly.
2489  CondUse->transformToPostInc(L);
2490  Changed = true;
2491 
2492  PostIncs.insert(Cond);
2493  decline_post_inc:;
2494  }
2495 
2496  // Determine an insertion point for the loop induction variable increment. It
2497  // must dominate all the post-inc comparisons we just set up, and it must
2498  // dominate the loop latch edge.
2499  IVIncInsertPos = L->getLoopLatch()->getTerminator();
2500  for (Instruction *Inst : PostIncs) {
2501  BasicBlock *BB =
2502  DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2503  Inst->getParent());
2504  if (BB == Inst->getParent())
2505  IVIncInsertPos = Inst;
2506  else if (BB != IVIncInsertPos->getParent())
2507  IVIncInsertPos = BB->getTerminator();
2508  }
2509 }
2510 
2511 /// Determine if the given use can accommodate a fixup at the given offset and
2512 /// other details. If so, update the use and return true.
2513 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2514  bool HasBaseReg, LSRUse::KindType Kind,
2515  MemAccessTy AccessTy) {
2516  int64_t NewMinOffset = LU.MinOffset;
2517  int64_t NewMaxOffset = LU.MaxOffset;
2518  MemAccessTy NewAccessTy = AccessTy;
2519 
2520  // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2521  // something conservative, however this can pessimize in the case that one of
2522  // the uses will have all its uses outside the loop, for example.
2523  if (LU.Kind != Kind)
2524  return false;
2525 
2526  // Check for a mismatched access type, and fall back conservatively as needed.
2527  // TODO: Be less conservative when the type is similar and can use the same
2528  // addressing modes.
2529  if (Kind == LSRUse::Address) {
2530  if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2531  NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2532  AccessTy.AddrSpace);
2533  }
2534  }
2535 
2536  // Conservatively assume HasBaseReg is true for now.
2537  if (NewOffset < LU.MinOffset) {
2538  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2539  LU.MaxOffset - NewOffset, HasBaseReg))
2540  return false;
2541  NewMinOffset = NewOffset;
2542  } else if (NewOffset > LU.MaxOffset) {
2543  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2544  NewOffset - LU.MinOffset, HasBaseReg))
2545  return false;
2546  NewMaxOffset = NewOffset;
2547  }
2548 
2549  // Update the use.
2550  LU.MinOffset = NewMinOffset;
2551  LU.MaxOffset = NewMaxOffset;
2552  LU.AccessTy = NewAccessTy;
2553  return true;
2554 }
2555 
2556 /// Return an LSRUse index and an offset value for a fixup which needs the given
2557 /// expression, with the given kind and optional access type. Either reuse an
2558 /// existing use or create a new one, as needed.
2559 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2560  LSRUse::KindType Kind,
2561  MemAccessTy AccessTy) {
2562  const SCEV *Copy = Expr;
2563  int64_t Offset = ExtractImmediate(Expr, SE);
2564 
2565  // Basic uses can't accept any offset, for example.
2566  if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2567  Offset, /*HasBaseReg=*/ true)) {
2568  Expr = Copy;
2569  Offset = 0;
2570  }
2571 
2572  std::pair<UseMapTy::iterator, bool> P =
2573  UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2574  if (!P.second) {
2575  // A use already existed with this base.
2576  size_t LUIdx = P.first->second;
2577  LSRUse &LU = Uses[LUIdx];
2578  if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2579  // Reuse this use.
2580  return std::make_pair(LUIdx, Offset);
2581  }
2582 
2583  // Create a new use.
2584  size_t LUIdx = Uses.size();
2585  P.first->second = LUIdx;
2586  Uses.push_back(LSRUse(Kind, AccessTy));
2587  LSRUse &LU = Uses[LUIdx];
2588 
2589  LU.MinOffset = Offset;
2590  LU.MaxOffset = Offset;
2591  return std::make_pair(LUIdx, Offset);
2592 }
2593 
2594 /// Delete the given use from the Uses list.
2595 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2596  if (&LU != &Uses.back())
2597  std::swap(LU, Uses.back());
2598  Uses.pop_back();
2599 
2600  // Update RegUses.
2601  RegUses.swapAndDropUse(LUIdx, Uses.size());
2602 }
2603 
2604 /// Look for a use distinct from OrigLU which is has a formula that has the same
2605 /// registers as the given formula.
2606 LSRUse *
2607 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2608  const LSRUse &OrigLU) {
2609  // Search all uses for the formula. This could be more clever.
2610  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2611  LSRUse &LU = Uses[LUIdx];
2612  // Check whether this use is close enough to OrigLU, to see whether it's
2613  // worthwhile looking through its formulae.
2614  // Ignore ICmpZero uses because they may contain formulae generated by
2615  // GenerateICmpZeroScales, in which case adding fixup offsets may
2616  // be invalid.
2617  if (&LU != &OrigLU &&
2618  LU.Kind != LSRUse::ICmpZero &&
2619  LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2620  LU.WidestFixupType == OrigLU.WidestFixupType &&
2621  LU.HasFormulaWithSameRegs(OrigF)) {
2622  // Scan through this use's formulae.
2623  for (const Formula &F : LU.Formulae) {
2624  // Check to see if this formula has the same registers and symbols
2625  // as OrigF.
2626  if (F.BaseRegs == OrigF.BaseRegs &&
2627  F.ScaledReg == OrigF.ScaledReg &&
2628  F.BaseGV == OrigF.BaseGV &&
2629  F.Scale == OrigF.Scale &&
2630  F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2631  if (F.BaseOffset == 0)
2632  return &LU;
2633  // This is the formula where all the registers and symbols matched;
2634  // there aren't going to be any others. Since we declined it, we
2635  // can skip the rest of the formulae and proceed to the next LSRUse.
2636  break;
2637  }
2638  }
2639  }
2640  }
2641 
2642  // Nothing looked good.
2643  return nullptr;
2644 }
2645 
2646 void LSRInstance::CollectInterestingTypesAndFactors() {
2648 
2649  // Collect interesting types and strides.
2651  for (const IVStrideUse &U : IU) {
2652  const SCEV *Expr = IU.getExpr(U);
2653 
2654  // Collect interesting types.
2655  Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2656 
2657  // Add strides for mentioned loops.
2658  Worklist.push_back(Expr);
2659  do {
2660  const SCEV *S = Worklist.pop_back_val();
2661  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2662  if (AR->getLoop() == L)
2663  Strides.insert(AR->getStepRecurrence(SE));
2664  Worklist.push_back(AR->getStart());
2665  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2666  Worklist.append(Add->op_begin(), Add->op_end());
2667  }
2668  } while (!Worklist.empty());
2669  }
2670 
2671  // Compute interesting factors from the set of interesting strides.
2673  I = Strides.begin(), E = Strides.end(); I != E; ++I)
2675  std::next(I); NewStrideIter != E; ++NewStrideIter) {
2676  const SCEV *OldStride = *I;
2677  const SCEV *NewStride = *NewStrideIter;
2678 
2679  if (SE.getTypeSizeInBits(OldStride->getType()) !=
2680  SE.getTypeSizeInBits(NewStride->getType())) {
2681  if (SE.getTypeSizeInBits(OldStride->getType()) >
2682  SE.getTypeSizeInBits(NewStride->getType()))
2683  NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2684  else
2685  OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2686  }
2687  if (const SCEVConstant *Factor =
2688  dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2689  SE, true))) {
2690  if (Factor->getAPInt().getMinSignedBits() <= 64)
2691  Factors.insert(Factor->getAPInt().getSExtValue());
2692  } else if (const SCEVConstant *Factor =
2693  dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2694  NewStride,
2695  SE, true))) {
2696  if (Factor->getAPInt().getMinSignedBits() <= 64)
2697  Factors.insert(Factor->getAPInt().getSExtValue());
2698  }
2699  }
2700 
2701  // If all uses use the same type, don't bother looking for truncation-based
2702  // reuse.
2703  if (Types.size() == 1)
2704  Types.clear();
2705 
2706  LLVM_DEBUG(print_factors_and_types(dbgs()));
2707 }
2708 
2709 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2710 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2711 /// IVStrideUses, we could partially skip this.
2712 static User::op_iterator
2714  Loop *L, ScalarEvolution &SE) {
2715  for(; OI != OE; ++OI) {
2716  if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2717  if (!SE.isSCEVable(Oper->getType()))
2718  continue;
2719 
2720  if (const SCEVAddRecExpr *AR =
2721  dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2722  if (AR->getLoop() == L)
2723  break;
2724  }
2725  }
2726  }
2727  return OI;
2728 }
2729 
2730 /// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2731 /// a convenient helper.
2732 static Value *getWideOperand(Value *Oper) {
2733  if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2734  return Trunc->getOperand(0);
2735  return Oper;
2736 }
2737 
2738 /// Return true if we allow an IV chain to include both types.
2739 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2740  Type *LType = LVal->getType();
2741  Type *RType = RVal->getType();
2742  return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2743  // Different address spaces means (possibly)
2744  // different types of the pointer implementation,
2745  // e.g. i16 vs i32 so disallow that.
2746  (LType->getPointerAddressSpace() ==
2747  RType->getPointerAddressSpace()));
2748 }
2749 
2750 /// Return an approximation of this SCEV expression's "base", or NULL for any
2751 /// constant. Returning the expression itself is conservative. Returning a
2752 /// deeper subexpression is more precise and valid as long as it isn't less
2753 /// complex than another subexpression. For expressions involving multiple
2754 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2755 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2756 /// IVInc==b-a.
2757 ///
2758 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2759 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2760 static const SCEV *getExprBase(const SCEV *S) {
2761  switch (S->getSCEVType()) {
2762  default: // uncluding scUnknown.
2763  return S;
2764  case scConstant:
2765  return nullptr;
2766  case scTruncate:
2767  return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2768  case scZeroExtend:
2769  return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2770  case scSignExtend:
2771  return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2772  case scAddExpr: {
2773  // Skip over scaled operands (scMulExpr) to follow add operands as long as
2774  // there's nothing more complex.
2775  // FIXME: not sure if we want to recognize negation.
2776  const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2777  for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
2778  E(Add->op_begin()); I != E; ++I) {
2779  const SCEV *SubExpr = *I;
2780  if (SubExpr->getSCEVType() == scAddExpr)
2781  return getExprBase(SubExpr);
2782 
2783  if (SubExpr->getSCEVType() != scMulExpr)
2784  return SubExpr;
2785  }
2786  return S; // all operands are scaled, be conservative.
2787  }
2788  case scAddRecExpr:
2789  return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2790  }
2791 }
2792 
2793 /// Return true if the chain increment is profitable to expand into a loop
2794 /// invariant value, which may require its own register. A profitable chain
2795 /// increment will be an offset relative to the same base. We allow such offsets
2796 /// to potentially be used as chain increment as long as it's not obviously
2797 /// expensive to expand using real instructions.
2798 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2799  const SCEV *IncExpr,
2800  ScalarEvolution &SE) {
2801  // Aggressively form chains when -stress-ivchain.
2802  if (StressIVChain)
2803  return true;
2804 
2805  // Do not replace a constant offset from IV head with a nonconstant IV
2806  // increment.
2807  if (!isa<SCEVConstant>(IncExpr)) {
2808  const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2809  if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2810  return false;
2811  }
2812 
2813  SmallPtrSet<const SCEV*, 8> Processed;
2814  return !isHighCostExpansion(IncExpr, Processed, SE);
2815 }
2816 
2817 /// Return true if the number of registers needed for the chain is estimated to
2818 /// be less than the number required for the individual IV users. First prohibit
2819 /// any IV users that keep the IV live across increments (the Users set should
2820 /// be empty). Next count the number and type of increments in the chain.
2821 ///
2822 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2823 /// effectively use postinc addressing modes. Only consider it profitable it the
2824 /// increments can be computed in fewer registers when chained.
2825 ///
2826 /// TODO: Consider IVInc free if it's already used in another chains.
2827 static bool
2829  ScalarEvolution &SE) {
2830  if (StressIVChain)
2831  return true;
2832 
2833  if (!Chain.hasIncs())
2834  return false;
2835 
2836  if (!Users.empty()) {
2837  LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2838  for (Instruction *Inst
2839  : Users) { dbgs() << " " << *Inst << "\n"; });
2840  return false;
2841  }
2842  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2843 
2844  // The chain itself may require a register, so intialize cost to 1.
2845  int cost = 1;
2846 
2847  // A complete chain likely eliminates the need for keeping the original IV in
2848  // a register. LSR does not currently know how to form a complete chain unless
2849  // the header phi already exists.
2850  if (isa<PHINode>(Chain.tailUserInst())
2851  && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2852  --cost;
2853  }
2854  const SCEV *LastIncExpr = nullptr;
2855  unsigned NumConstIncrements = 0;
2856  unsigned NumVarIncrements = 0;
2857  unsigned NumReusedIncrements = 0;
2858  for (const IVInc &Inc : Chain) {
2859  if (Inc.IncExpr->isZero())
2860  continue;
2861 
2862  // Incrementing by zero or some constant is neutral. We assume constants can
2863  // be folded into an addressing mode or an add's immediate operand.
2864  if (isa<SCEVConstant>(Inc.IncExpr)) {
2865  ++NumConstIncrements;
2866  continue;
2867  }
2868 
2869  if (Inc.IncExpr == LastIncExpr)
2870  ++NumReusedIncrements;
2871  else
2872  ++NumVarIncrements;
2873 
2874  LastIncExpr = Inc.IncExpr;
2875  }
2876  // An IV chain with a single increment is handled by LSR's postinc
2877  // uses. However, a chain with multiple increments requires keeping the IV's
2878  // value live longer than it needs to be if chained.
2879  if (NumConstIncrements > 1)
2880  --cost;
2881 
2882  // Materializing increment expressions in the preheader that didn't exist in
2883  // the original code may cost a register. For example, sign-extended array
2884  // indices can produce ridiculous increments like this:
2885  // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2886  cost += NumVarIncrements;
2887 
2888  // Reusing variable increments likely saves a register to hold the multiple of
2889  // the stride.
2890  cost -= NumReusedIncrements;
2891 
2892  LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2893  << "\n");
2894 
2895  return cost < 0;
2896 }
2897 
2898 /// Add this IV user to an existing chain or make it the head of a new chain.
2899 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2900  SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2901  // When IVs are used as types of varying widths, they are generally converted
2902  // to a wider type with some uses remaining narrow under a (free) trunc.
2903  Value *const NextIV = getWideOperand(IVOper);
2904  const SCEV *const OperExpr = SE.getSCEV(NextIV);
2905  const SCEV *const OperExprBase = getExprBase(OperExpr);
2906 
2907  // Visit all existing chains. Check if its IVOper can be computed as a
2908  // profitable loop invariant increment from the last link in the Chain.
2909  unsigned ChainIdx = 0, NChains = IVChainVec.size();
2910  const SCEV *LastIncExpr = nullptr;
2911  for (; ChainIdx < NChains; ++ChainIdx) {
2912  IVChain &Chain = IVChainVec[ChainIdx];
2913 
2914  // Prune the solution space aggressively by checking that both IV operands
2915  // are expressions that operate on the same unscaled SCEVUnknown. This
2916  // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2917  // first avoids creating extra SCEV expressions.
2918  if (!StressIVChain && Chain.ExprBase != OperExprBase)
2919  continue;
2920 
2921  Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2922  if (!isCompatibleIVType(PrevIV, NextIV))
2923  continue;
2924 
2925  // A phi node terminates a chain.
2926  if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2927  continue;
2928 
2929  // The increment must be loop-invariant so it can be kept in a register.
2930  const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2931  const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2932  if (!SE.isLoopInvariant(IncExpr, L))
2933  continue;
2934 
2935  if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2936  LastIncExpr = IncExpr;
2937  break;
2938  }
2939  }
2940  // If we haven't found a chain, create a new one, unless we hit the max. Don't
2941  // bother for phi nodes, because they must be last in the chain.
2942  if (ChainIdx == NChains) {
2943  if (isa<PHINode>(UserInst))
2944  return;
2945  if (NChains >= MaxChains && !StressIVChain) {
2946  LLVM_DEBUG(dbgs() << "IV Chain Limit\n");
2947  return;
2948  }
2949  LastIncExpr = OperExpr;
2950  // IVUsers may have skipped over sign/zero extensions. We don't currently
2951  // attempt to form chains involving extensions unless they can be hoisted
2952  // into this loop's AddRec.
2953  if (!isa<SCEVAddRecExpr>(LastIncExpr))
2954  return;
2955  ++NChains;
2956  IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
2957  OperExprBase));
2958  ChainUsersVec.resize(NChains);
2959  LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
2960  << ") IV=" << *LastIncExpr << "\n");
2961  } else {
2962  LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
2963  << ") IV+" << *LastIncExpr << "\n");
2964  // Add this IV user to the end of the chain.
2965  IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
2966  }
2967  IVChain &Chain = IVChainVec[ChainIdx];
2968 
2969  SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
2970  // This chain's NearUsers become FarUsers.
2971  if (!LastIncExpr->isZero()) {
2972  ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
2973  NearUsers.end());
2974  NearUsers.clear();
2975  }
2976 
2977  // All other uses of IVOperand become near uses of the chain.
2978  // We currently ignore intermediate values within SCEV expressions, assuming
2979  // they will eventually be used be the current chain, or can be computed
2980  // from one of the chain increments. To be more precise we could
2981  // transitively follow its user and only add leaf IV users to the set.
2982  for (User *U : IVOper->users()) {
2983  Instruction *OtherUse = dyn_cast<Instruction>(U);
2984  if (!OtherUse)
2985  continue;
2986  // Uses in the chain will no longer be uses if the chain is formed.
2987  // Include the head of the chain in this iteration (not Chain.begin()).
2988  IVChain::const_iterator IncIter = Chain.Incs.begin();
2989  IVChain::const_iterator IncEnd = Chain.Incs.end();
2990  for( ; IncIter != IncEnd; ++IncIter) {
2991  if (IncIter->UserInst == OtherUse)
2992  break;
2993  }
2994  if (IncIter != IncEnd)
2995  continue;
2996 
2997  if (SE.isSCEVable(OtherUse->getType())
2998  && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
2999  && IU.isIVUserOrOperand(OtherUse)) {
3000  continue;
3001  }
3002  NearUsers.insert(OtherUse);
3003  }
3004 
3005  // Since this user is part of the chain, it's no longer considered a use
3006  // of the chain.
3007  ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3008 }
3009 
3010 /// Populate the vector of Chains.
3011 ///
3012 /// This decreases ILP at the architecture level. Targets with ample registers,
3013 /// multiple memory ports, and no register renaming probably don't want
3014 /// this. However, such targets should probably disable LSR altogether.
3015 ///
3016 /// The job of LSR is to make a reasonable choice of induction variables across
3017 /// the loop. Subsequent passes can easily "unchain" computation exposing more
3018 /// ILP *within the loop* if the target wants it.
3019 ///
3020 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
3021 /// will not reorder memory operations, it will recognize this as a chain, but
3022 /// will generate redundant IV increments. Ideally this would be corrected later
3023 /// by a smart scheduler:
3024 /// = A[i]
3025 /// = A[i+x]
3026 /// A[i] =
3027 /// A[i+x] =
3028 ///
3029 /// TODO: Walk the entire domtree within this loop, not just the path to the
3030 /// loop latch. This will discover chains on side paths, but requires
3031 /// maintaining multiple copies of the Chains state.
3032 void LSRInstance::CollectChains() {
3033  LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n");
3034  SmallVector<ChainUsers, 8> ChainUsersVec;
3035 
3036  SmallVector<BasicBlock *,8> LatchPath;
3037  BasicBlock *LoopHeader = L->getHeader();
3038  for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3039  Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3040  LatchPath.push_back(Rung->getBlock());
3041  }
3042  LatchPath.push_back(LoopHeader);
3043 
3044  // Walk the instruction stream from the loop header to the loop latch.
3045  for (BasicBlock *BB : reverse(LatchPath)) {
3046  for (Instruction &I : *BB) {
3047  // Skip instructions that weren't seen by IVUsers analysis.
3048  if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3049  continue;
3050 
3051  // Ignore users that are part of a SCEV expression. This way we only
3052  // consider leaf IV Users. This effectively rediscovers a portion of
3053  // IVUsers analysis but in program order this time.
3054  if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3055  continue;
3056 
3057  // Remove this instruction from any NearUsers set it may be in.
3058  for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3059  ChainIdx < NChains; ++ChainIdx) {
3060  ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3061  }
3062  // Search for operands that can be chained.
3063  SmallPtrSet<Instruction*, 4> UniqueOperands;
3064  User::op_iterator IVOpEnd = I.op_end();
3065  User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3066  while (IVOpIter != IVOpEnd) {
3067  Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3068  if (UniqueOperands.insert(IVOpInst).second)
3069  ChainInstruction(&I, IVOpInst, ChainUsersVec);
3070  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3071  }
3072  } // Continue walking down the instructions.
3073  } // Continue walking down the domtree.
3074  // Visit phi backedges to determine if the chain can generate the IV postinc.
3075  for (PHINode &PN : L->getHeader()->phis()) {
3076  if (!SE.isSCEVable(PN.getType()))
3077  continue;
3078 
3079  Instruction *IncV =
3080  dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3081  if (IncV)
3082  ChainInstruction(&PN, IncV, ChainUsersVec);
3083  }
3084  // Remove any unprofitable chains.
3085  unsigned ChainIdx = 0;
3086  for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3087  UsersIdx < NChains; ++UsersIdx) {
3088  if (!isProfitableChain(IVChainVec[UsersIdx],
3089  ChainUsersVec[UsersIdx].FarUsers, SE))
3090  continue;
3091  // Preserve the chain at UsesIdx.
3092  if (ChainIdx != UsersIdx)
3093  IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3094  FinalizeChain(IVChainVec[ChainIdx]);
3095  ++ChainIdx;
3096  }
3097  IVChainVec.resize(ChainIdx);
3098 }
3099 
3100 void LSRInstance::FinalizeChain(IVChain &Chain) {
3101  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3102  LLVM_DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
3103 
3104  for (const IVInc &Inc : Chain) {
3105  LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n");
3106  auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3107  assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
3108  IVIncSet.insert(UseI);
3109  }
3110 }
3111 
3112 /// Return true if the IVInc can be folded into an addressing mode.
3113 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3114  Value *Operand, const TargetTransformInfo &TTI) {
3115  const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3116  if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3117  return false;
3118 
3119  if (IncConst->getAPInt().getMinSignedBits() > 64)
3120  return false;
3121 
3122  MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3123  int64_t IncOffset = IncConst->getValue()->getSExtValue();
3124  if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3125  IncOffset, /*HaseBaseReg=*/false))
3126  return false;
3127 
3128  return true;
3129 }
3130 
3131 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3132 /// user's operand from the previous IV user's operand.
3133 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3134  SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3135  // Find the new IVOperand for the head of the chain. It may have been replaced
3136  // by LSR.
3137  const IVInc &Head = Chain.Incs[0];
3138  User::op_iterator IVOpEnd = Head.UserInst->op_end();
3139  // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3140  User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3141  IVOpEnd, L, SE);
3142  Value *IVSrc = nullptr;
3143  while (IVOpIter != IVOpEnd) {
3144  IVSrc = getWideOperand(*IVOpIter);
3145 
3146  // If this operand computes the expression that the chain needs, we may use
3147  // it. (Check this after setting IVSrc which is used below.)
3148  //
3149  // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3150  // narrow for the chain, so we can no longer use it. We do allow using a
3151  // wider phi, assuming the LSR checked for free truncation. In that case we
3152  // should already have a truncate on this operand such that
3153  // getSCEV(IVSrc) == IncExpr.
3154  if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3155  || SE.getSCEV(IVSrc) == Head.IncExpr) {
3156  break;
3157  }
3158  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3159  }
3160  if (IVOpIter == IVOpEnd) {
3161  // Gracefully give up on this chain.
3162  LLVM_DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3163  return;
3164  }
3165 
3166  LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3167  Type *IVTy = IVSrc->getType();
3168  Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3169  const SCEV *LeftOverExpr = nullptr;
3170  for (const IVInc &Inc : Chain) {
3171  Instruction *InsertPt = Inc.UserInst;
3172  if (isa<PHINode>(InsertPt))
3173  InsertPt = L->getLoopLatch()->getTerminator();
3174 
3175  // IVOper will replace the current IV User's operand. IVSrc is the IV
3176  // value currently held in a register.
3177  Value *IVOper = IVSrc;
3178  if (!Inc.IncExpr->isZero()) {
3179  // IncExpr was the result of subtraction of two narrow values, so must
3180  // be signed.
3181  const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3182  LeftOverExpr = LeftOverExpr ?
3183  SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3184  }
3185  if (LeftOverExpr && !LeftOverExpr->isZero()) {
3186  // Expand the IV increment.
3187  Rewriter.clearPostInc();
3188  Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3189  const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3190  SE.getUnknown(IncV));
3191  IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3192 
3193  // If an IV increment can't be folded, use it as the next IV value.
3194  if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3195  assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3196  IVSrc = IVOper;
3197  LeftOverExpr = nullptr;
3198  }
3199  }
3200  Type *OperTy = Inc.IVOperand->getType();
3201  if (IVTy != OperTy) {
3202  assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3203  "cannot extend a chained IV");
3204  IRBuilder<> Builder(InsertPt);
3205  IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3206  }
3207  Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3208  DeadInsts.emplace_back(Inc.IVOperand);
3209  }
3210  // If LSR created a new, wider phi, we may also replace its postinc. We only
3211  // do this if we also found a wide value for the head of the chain.
3212  if (isa<PHINode>(Chain.tailUserInst())) {
3213  for (PHINode &Phi : L->getHeader()->phis()) {
3214  if (!isCompatibleIVType(&Phi, IVSrc))
3215  continue;
3216  Instruction *PostIncV = dyn_cast<Instruction>(
3217  Phi.getIncomingValueForBlock(L->getLoopLatch()));
3218  if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3219  continue;
3220  Value *IVOper = IVSrc;
3221  Type *PostIncTy = PostIncV->getType();
3222  if (IVTy != PostIncTy) {
3223  assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3224  IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3225  Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3226  IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3227  }
3228  Phi.replaceUsesOfWith(PostIncV, IVOper);
3229  DeadInsts.emplace_back(PostIncV);
3230  }
3231  }
3232 }
3233 
3234 void LSRInstance::CollectFixupsAndInitialFormulae() {
3235  for (const IVStrideUse &U : IU) {
3236  Instruction *UserInst = U.getUser();
3237  // Skip IV users that are part of profitable IV Chains.
3238  User::op_iterator UseI =
3239  find(UserInst->operands(), U.getOperandValToReplace());
3240  assert(UseI != UserInst->op_end() && "cannot find IV operand");
3241  if (IVIncSet.count(UseI)) {
3242  LLVM_DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3243  continue;
3244  }
3245 
3246  LSRUse::KindType Kind = LSRUse::Basic;
3247  MemAccessTy AccessTy;
3248  if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3249  Kind = LSRUse::Address;
3250  AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3251  }
3252 
3253  const SCEV *S = IU.getExpr(U);
3254  PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3255 
3256  // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3257  // (N - i == 0), and this allows (N - i) to be the expression that we work
3258  // with rather than just N or i, so we can consider the register
3259  // requirements for both N and i at the same time. Limiting this code to
3260  // equality icmps is not a problem because all interesting loops use
3261  // equality icmps, thanks to IndVarSimplify.
3262  if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst))
3263  if (CI->isEquality()) {
3264  // Swap the operands if needed to put the OperandValToReplace on the
3265  // left, for consistency.
3266  Value *NV = CI->getOperand(1);
3267  if (NV == U.getOperandValToReplace()) {
3268  CI->setOperand(1, CI->getOperand(0));
3269  CI->setOperand(0, NV);
3270  NV = CI->getOperand(1);
3271  Changed = true;
3272  }
3273 
3274  // x == y --> x - y == 0
3275  const SCEV *N = SE.getSCEV(NV);
3276  if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
3277  // S is normalized, so normalize N before folding it into S
3278  // to keep the result normalized.
3279  N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3280  Kind = LSRUse::ICmpZero;
3281  S = SE.getMinusSCEV(N, S);
3282  }
3283 
3284  // -1 and the negations of all interesting strides (except the negation
3285  // of -1) are now also interesting.
3286  for (size_t i = 0, e = Factors.size(); i != e; ++i)
3287  if (Factors[i] != -1)
3288  Factors.insert(-(uint64_t)Factors[i]);
3289  Factors.insert(-1);
3290  }
3291 
3292  // Get or create an LSRUse.
3293  std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3294  size_t LUIdx = P.first;
3295  int64_t Offset = P.second;
3296  LSRUse &LU = Uses[LUIdx];
3297 
3298  // Record the fixup.
3299  LSRFixup &LF = LU.getNewFixup();
3300  LF.UserInst = UserInst;
3301  LF.OperandValToReplace = U.getOperandValToReplace();
3302  LF.PostIncLoops = TmpPostIncLoops;
3303  LF.Offset = Offset;
3304  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3305 
3306  if (!LU.WidestFixupType ||
3307  SE.getTypeSizeInBits(LU.WidestFixupType) <
3308  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3309  LU.WidestFixupType = LF.OperandValToReplace->getType();
3310 
3311  // If this is the first use of this LSRUse, give it a formula.
3312  if (LU.Formulae.empty()) {
3313  InsertInitialFormula(S, LU, LUIdx);
3314  CountRegisters(LU.Formulae.back(), LUIdx);
3315  }
3316  }
3317 
3318  LLVM_DEBUG(print_fixups(dbgs()));
3319 }
3320 
3321 /// Insert a formula for the given expression into the given use, separating out
3322 /// loop-variant portions from loop-invariant and loop-computable portions.
3323 void
3324 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3325  // Mark uses whose expressions cannot be expanded.
3326  if (!isSafeToExpand(S, SE))
3327  LU.RigidFormula = true;
3328 
3329  Formula F;
3330  F.initialMatch(S, L, SE);
3331  bool Inserted = InsertFormula(LU, LUIdx, F);
3332  assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3333 }
3334 
3335 /// Insert a simple single-register formula for the given expression into the
3336 /// given use.
3337 void
3338 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3339  LSRUse &LU, size_t LUIdx) {
3340  Formula F;
3341  F.BaseRegs.push_back(S);
3342  F.HasBaseReg = true;
3343  bool Inserted = InsertFormula(LU, LUIdx, F);
3344  assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3345 }
3346 
3347 /// Note which registers are used by the given formula, updating RegUses.
3348 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3349  if (F.ScaledReg)
3350  RegUses.countRegister(F.ScaledReg, LUIdx);
3351  for (const SCEV *BaseReg : F.BaseRegs)
3352  RegUses.countRegister(BaseReg, LUIdx);
3353 }
3354 
3355 /// If the given formula has not yet been inserted, add it to the list, and
3356 /// return true. Return false otherwise.
3357 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3358  // Do not insert formula that we will not be able to expand.
3359  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3360  "Formula is illegal");
3361 
3362  if (!LU.InsertFormula(F, *L))
3363  return false;
3364 
3365  CountRegisters(F, LUIdx);
3366  return true;
3367 }
3368 
3369 /// Check for other uses of loop-invariant values which we're tracking. These
3370 /// other uses will pin these values in registers, making them less profitable
3371 /// for elimination.
3372 /// TODO: This currently misses non-constant addrec step registers.
3373 /// TODO: Should this give more weight to users inside the loop?
3374 void
3375 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3376  SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3378 
3379  while (!Worklist.empty()) {
3380  const SCEV *S = Worklist.pop_back_val();
3381 
3382  // Don't process the same SCEV twice
3383  if (!Visited.insert(S).second)
3384  continue;
3385 
3386  if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3387  Worklist.append(N->op_begin(), N->op_end());
3388  else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
3389  Worklist.push_back(C->getOperand());
3390  else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3391  Worklist.push_back(D->getLHS());
3392  Worklist.push_back(D->getRHS());
3393  } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3394  const Value *V = US->getValue();
3395  if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3396  // Look for instructions defined outside the loop.
3397  if (L->contains(Inst)) continue;
3398  } else if (isa<UndefValue>(V))
3399  // Undef doesn't have a live range, so it doesn't matter.
3400  continue;
3401  for (const Use &U : V->uses()) {
3402  const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3403  // Ignore non-instructions.
3404  if (!UserInst)
3405  continue;
3406  // Ignore instructions in other functions (as can happen with
3407  // Constants).
3408  if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3409  continue;
3410  // Ignore instructions not dominated by the loop.
3411  const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3412  UserInst->getParent() :
3413  cast<PHINode>(UserInst)->getIncomingBlock(
3414  PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3415  if (!DT.dominates(L->getHeader(), UseBB))
3416  continue;
3417  // Don't bother if the instruction is in a BB which ends in an EHPad.
3418  if (UseBB->getTerminator()->isEHPad())
3419  continue;
3420  // Don't bother rewriting PHIs in catchswitch blocks.
3421  if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3422  continue;
3423  // Ignore uses which are part of other SCEV expressions, to avoid
3424  // analyzing them multiple times.
3425  if (SE.isSCEVable(UserInst->getType())) {
3426  const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3427  // If the user is a no-op, look through to its uses.
3428  if (!isa<SCEVUnknown>(UserS))
3429  continue;
3430  if (UserS == US) {
3431  Worklist.push_back(
3432  SE.getUnknown(const_cast<Instruction *>(UserInst)));
3433  continue;
3434  }
3435  }
3436  // Ignore icmp instructions which are already being analyzed.
3437  if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3438  unsigned OtherIdx = !U.getOperandNo();
3439  Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3440  if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3441  continue;
3442  }
3443 
3444  std::pair<size_t, int64_t> P = getUse(
3445  S, LSRUse::Basic, MemAccessTy());
3446  size_t LUIdx = P.first;
3447  int64_t Offset = P.second;
3448  LSRUse &LU = Uses[LUIdx];
3449  LSRFixup &LF = LU.getNewFixup();
3450  LF.UserInst = const_cast<Instruction *>(UserInst);
3451  LF.OperandValToReplace = U;
3452  LF.Offset = Offset;
3453  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3454  if (!LU.WidestFixupType ||
3455  SE.getTypeSizeInBits(LU.WidestFixupType) <
3456  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3457  LU.WidestFixupType = LF.OperandValToReplace->getType();
3458  InsertSupplementalFormula(US, LU, LUIdx);
3459  CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3460  break;
3461  }
3462  }
3463  }
3464 }
3465 
3466 /// Split S into subexpressions which can be pulled out into separate
3467 /// registers. If C is non-null, multiply each subexpression by C.
3468 ///
3469 /// Return remainder expression after factoring the subexpressions captured by
3470 /// Ops. If Ops is complete, return NULL.
3471 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3473  const Loop *L,
3474  ScalarEvolution &SE,
3475  unsigned Depth = 0) {
3476  // Arbitrarily cap recursion to protect compile time.
3477  if (Depth >= 3)
3478  return S;
3479 
3480  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3481  // Break out add operands.
3482  for (const SCEV *S : Add->operands()) {
3483  const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3484  if (Remainder)
3485  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3486  }
3487  return nullptr;
3488  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3489  // Split a non-zero base out of an addrec.
3490  if (AR->getStart()->isZero() || !AR->isAffine())
3491  return S;
3492 
3493  const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3494  C, Ops, L, SE, Depth+1);
3495  // Split the non-zero AddRec unless it is part of a nested recurrence that
3496  // does not pertain to this loop.
3497  if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3498  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3499  Remainder = nullptr;
3500  }
3501  if (Remainder != AR->getStart()) {
3502  if (!Remainder)
3503  Remainder = SE.getConstant(AR->getType(), 0);
3504  return SE.getAddRecExpr(Remainder,
3505  AR->getStepRecurrence(SE),
3506  AR->getLoop(),
3507  //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3509  }
3510  } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3511  // Break (C * (a + b + c)) into C*a + C*b + C*c.
3512  if (Mul->getNumOperands() != 2)
3513  return S;
3514  if (const SCEVConstant *Op0 =
3515  dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3516  C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3517  const SCEV *Remainder =
3518  CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3519  if (Remainder)
3520  Ops.push_back(SE.getMulExpr(C, Remainder));
3521  return nullptr;
3522  }
3523  }
3524  return S;
3525 }
3526 
3527 /// Return true if the SCEV represents a value that may end up as a
3528 /// post-increment operation.
3530  LSRUse &LU, const SCEV *S, const Loop *L,
3531  ScalarEvolution &SE) {
3532  if (LU.Kind != LSRUse::Address ||
3533  !LU.AccessTy.getType()->isIntOrIntVectorTy())
3534  return false;
3535  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3536  if (!AR)
3537  return false;
3538  const SCEV *LoopStep = AR->getStepRecurrence(SE);
3539  if (!isa<SCEVConstant>(LoopStep))
3540  return false;
3541  if (LU.AccessTy.getType()->getScalarSizeInBits() !=
3542  LoopStep->getType()->getScalarSizeInBits())
3543  return false;
3544  // Check if a post-indexed load/store can be used.
3545  if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3546  TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3547  const SCEV *LoopStart = AR->getStart();
3548  if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3549  return true;
3550  }
3551  return false;
3552 }
3553 
3554 /// Helper function for LSRInstance::GenerateReassociations.
3555 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3556  const Formula &Base,
3557  unsigned Depth, size_t Idx,
3558  bool IsScaledReg) {
3559  const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3560  // Don't generate reassociations for the base register of a value that
3561  // may generate a post-increment operator. The reason is that the
3562  // reassociations cause extra base+register formula to be created,
3563  // and possibly chosen, but the post-increment is more efficient.
3564  if (TTI.shouldFavorPostInc() && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3565  return;
3567  const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3568  if (Remainder)
3569  AddOps.push_back(Remainder);
3570 
3571  if (AddOps.size() == 1)
3572  return;
3573 
3575  JE = AddOps.end();
3576  J != JE; ++J) {
3577  // Loop-variant "unknown" values are uninteresting; we won't be able to
3578  // do anything meaningful with them.
3579  if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3580  continue;
3581 
3582  // Don't pull a constant into a register if the constant could be folded
3583  // into an immediate field.
3584  if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3585  LU.AccessTy, *J, Base.getNumRegs() > 1))
3586  continue;
3587 
3588  // Collect all operands except *J.
3589  SmallVector<const SCEV *, 8> InnerAddOps(
3590  ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3591  InnerAddOps.append(std::next(J),
3592  ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3593 
3594  // Don't leave just a constant behind in a register if the constant could
3595  // be folded into an immediate field.
3596  if (InnerAddOps.size() == 1 &&
3597  isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3598  LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3599  continue;
3600 
3601  const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3602  if (InnerSum->isZero())
3603  continue;
3604  Formula F = Base;
3605 
3606  // Add the remaining pieces of the add back into the new formula.
3607  const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3608  if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3609  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3610  InnerSumSC->getValue()->getZExtValue())) {
3611  F.UnfoldedOffset =
3612  (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3613  if (IsScaledReg)
3614  F.ScaledReg = nullptr;
3615  else
3616  F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3617  } else if (IsScaledReg)
3618  F.ScaledReg = InnerSum;
3619  else
3620  F.BaseRegs[Idx] = InnerSum;
3621 
3622  // Add J as its own register, or an unfolded immediate.
3623  const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3624  if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3625  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3626  SC->getValue()->getZExtValue()))
3627  F.UnfoldedOffset =
3628  (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3629  else
3630  F.BaseRegs.push_back(*J);
3631  // We may have changed the number of register in base regs, adjust the
3632  // formula accordingly.
3633  F.canonicalize(*L);
3634 
3635  if (InsertFormula(LU, LUIdx, F))
3636  // If that formula hadn't been seen before, recurse to find more like
3637  // it.
3638  // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3639  // Because just Depth is not enough to bound compile time.
3640  // This means that every time AddOps.size() is greater 16^x we will add
3641  // x to Depth.
3642  GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3643  Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3644  }
3645 }
3646 
3647 /// Split out subexpressions from adds and the bases of addrecs.
3648 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3649  Formula Base, unsigned Depth) {
3650  assert(Base.isCanonical(*L) && "Input must be in the canonical form");
3651  // Arbitrarily cap recursion to protect compile time.
3652  if (Depth >= 3)
3653  return;
3654 
3655  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3656  GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3657 
3658  if (Base.Scale == 1)
3659  GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3660  /* Idx */ -1, /* IsScaledReg */ true);
3661 }
3662 
3663 /// Generate a formula consisting of all of the loop-dominating registers added
3664 /// into a single register.
3665 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3666  Formula Base) {
3667  // This method is only interesting on a plurality of registers.
3668  if (Base.BaseRegs.size() + (Base.Scale == 1) +
3669  (Base.UnfoldedOffset != 0) <= 1)
3670  return;
3671 
3672  // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3673  // processing the formula.
3674  Base.unscale();
3676  Formula NewBase = Base;
3677  NewBase.BaseRegs.clear();
3678  Type *CombinedIntegerType = nullptr;
3679  for (const SCEV *BaseReg : Base.BaseRegs) {
3680  if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3681  !SE.hasComputableLoopEvolution(BaseReg, L)) {
3682  if (!CombinedIntegerType)
3683  CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3684  Ops.push_back(BaseReg);
3685  }
3686  else
3687  NewBase.BaseRegs.push_back(BaseReg);
3688  }
3689 
3690  // If no register is relevant, we're done.
3691  if (Ops.size() == 0)
3692  return;
3693 
3694  // Utility function for generating the required variants of the combined
3695  // registers.
3696  auto GenerateFormula = [&](const SCEV *Sum) {
3697  Formula F = NewBase;
3698 
3699  // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3700  // opportunity to fold something. For now, just ignore such cases
3701  // rather than proceed with zero in a register.
3702  if (Sum->isZero())
3703  return;
3704 
3705  F.BaseRegs.push_back(Sum);
3706  F.canonicalize(*L);
3707  (void)InsertFormula(LU, LUIdx, F);
3708  };
3709 
3710  // If we collected at least two registers, generate a formula combining them.
3711  if (Ops.size() > 1) {
3712  SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3713  GenerateFormula(SE.getAddExpr(OpsCopy));
3714  }
3715 
3716  // If we have an unfolded offset, generate a formula combining it with the
3717  // registers collected.
3718  if (NewBase.UnfoldedOffset) {
3719  assert(CombinedIntegerType && "Missing a type for the unfolded offset");
3720  Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3721  true));
3722  NewBase.UnfoldedOffset = 0;
3723  GenerateFormula(SE.getAddExpr(Ops));
3724  }
3725 }
3726 
3727 /// Helper function for LSRInstance::GenerateSymbolicOffsets.
3728 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3729  const Formula &Base, size_t Idx,
3730  bool IsScaledReg) {
3731  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3732  GlobalValue *GV = ExtractSymbol(G, SE);
3733  if (G->isZero() || !GV)
3734  return;
3735  Formula F = Base;
3736  F.BaseGV = GV;
3737  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3738  return;
3739  if (IsScaledReg)
3740  F.ScaledReg = G;
3741  else
3742  F.BaseRegs[Idx] = G;
3743  (void)InsertFormula(LU, LUIdx, F);
3744 }
3745 
3746 /// Generate reuse formulae using symbolic offsets.
3747 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3748  Formula Base) {
3749  // We can't add a symbolic offset if the address already contains one.
3750  if (Base.BaseGV) return;
3751 
3752  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3753  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3754  if (Base.Scale == 1)
3755  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3756  /* IsScaledReg */ true);
3757 }
3758 
3759 /// Helper function for LSRInstance::GenerateConstantOffsets.
3760 void LSRInstance::GenerateConstantOffsetsImpl(
3761  LSRUse &LU, unsigned LUIdx, const Formula &Base,
3762  const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3763 
3764  auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3765  Formula F = Base;
3766  F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3767 
3768  if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind,
3769  LU.AccessTy, F)) {
3770  // Add the offset to the base register.
3771  const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3772  // If it cancelled out, drop the base register, otherwise update it.
3773  if (NewG->isZero()) {
3774  if (IsScaledReg) {
3775  F.Scale = 0;
3776  F.ScaledReg = nullptr;
3777  } else
3778  F.deleteBaseReg(F.BaseRegs[Idx]);
3779  F.canonicalize(*L);
3780  } else if (IsScaledReg)
3781  F.ScaledReg = NewG;
3782  else
3783  F.BaseRegs[Idx] = NewG;
3784 
3785  (void)InsertFormula(LU, LUIdx, F);
3786  }
3787  };
3788 
3789  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3790 
3791  // With constant offsets and constant steps, we can generate pre-inc
3792  // accesses by having the offset equal the step. So, for access #0 with a
3793  // step of 8, we generate a G - 8 base which would require the first access
3794  // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3795  // for itself and hopefully becomes the base for other accesses. This means
3796  // means that a single pre-indexed access can be generated to become the new
3797  // base pointer for each iteration of the loop, resulting in no extra add/sub
3798  // instructions for pointer updating.
3799  if (FavorBackedgeIndex && LU.Kind == LSRUse::Address) {
3800  if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3801  if (auto *StepRec =
3802  dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3803  const APInt &StepInt = StepRec->getAPInt();
3804  int64_t Step = StepInt.isNegative() ?
3805  StepInt.getSExtValue() : StepInt.getZExtValue();
3806 
3807  for (int64_t Offset : Worklist) {
3808  Offset -= Step;
3809  GenerateOffset(G, Offset);
3810  }
3811  }
3812  }
3813  }
3814  for (int64_t Offset : Worklist)
3815  GenerateOffset(G, Offset);
3816 
3817  int64_t Imm = ExtractImmediate(G, SE);
3818  if (G->isZero() || Imm == 0)
3819  return;
3820  Formula F = Base;
3821  F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3822  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3823  return;
3824  if (IsScaledReg)
3825  F.ScaledReg = G;
3826  else
3827  F.BaseRegs[Idx] = G;
3828  (void)InsertFormula(LU, LUIdx, F);
3829 }
3830 
3831 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3832 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3833  Formula Base) {
3834  // TODO: For now, just add the min and max offset, because it usually isn't
3835  // worthwhile looking at everything inbetween.
3836  SmallVector<int64_t, 2> Worklist;
3837  Worklist.push_back(LU.MinOffset);
3838  if (LU.MaxOffset != LU.MinOffset)
3839  Worklist.push_back(LU.MaxOffset);
3840 
3841  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3842  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3843  if (Base.Scale == 1)
3844  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3845  /* IsScaledReg */ true);
3846 }
3847 
3848 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3849 /// == y -> x*c == y*c.
3850 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3851  Formula Base) {
3852  if (LU.Kind != LSRUse::ICmpZero) return;
3853 
3854  // Determine the integer type for the base formula.
3855  Type *IntTy = Base.getType();
3856  if (!IntTy) return;
3857  if (SE.getTypeSizeInBits(IntTy) > 64) return;
3858 
3859  // Don't do this if there is more than one offset.
3860  if (LU.MinOffset != LU.MaxOffset) return;
3861 
3862  // Check if transformation is valid. It is illegal to multiply pointer.
3863  if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3864  return;
3865  for (const SCEV *BaseReg : Base.BaseRegs)
3866  if (BaseReg->getType()->isPointerTy())
3867  return;
3868  assert(!Base.BaseGV && "ICmpZero use is not legal!");
3869 
3870  // Check each interesting stride.
3871  for (int64_t Factor : Factors) {
3872  // Check that the multiplication doesn't overflow.
3873  if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3874  continue;
3875  int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3876  if (NewBaseOffset / Factor != Base.BaseOffset)
3877  continue;
3878  // If the offset will be truncated at this use, check that it is in bounds.
3879  if (!IntTy->isPointerTy() &&
3880  !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3881  continue;
3882 
3883  // Check that multiplying with the use offset doesn't overflow.
3884  int64_t Offset = LU.MinOffset;
3885  if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3886  continue;
3887  Offset = (uint64_t)Offset * Factor;
3888  if (Offset / Factor != LU.MinOffset)
3889  continue;
3890  // If the offset will be truncated at this use, check that it is in bounds.
3891  if (!IntTy->isPointerTy() &&
3892  !ConstantInt::isValueValidForType(IntTy, Offset))
3893  continue;
3894 
3895  Formula F = Base;
3896  F.BaseOffset = NewBaseOffset;
3897 
3898  // Check that this scale is legal.
3899  if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3900  continue;
3901 
3902  // Compensate for the use having MinOffset built into it.
3903  F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
3904 
3905  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3906 
3907  // Check that multiplying with each base register doesn't overflow.
3908  for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
3909  F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
3910  if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
3911  goto next;
3912  }
3913 
3914  // Check that multiplying with the scaled register doesn't overflow.
3915  if (F.ScaledReg) {
3916  F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
3917  if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
3918  continue;
3919  }
3920 
3921  // Check that multiplying with the unfolded offset doesn't overflow.
3922  if (F.UnfoldedOffset != 0) {
3923  if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
3924  Factor == -1)
3925  continue;
3926  F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
3927  if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
3928  continue;
3929  // If the offset will be truncated, check that it is in bounds.
3930  if (!IntTy->isPointerTy() &&
3931  !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
3932  continue;
3933  }
3934 
3935  // If we make it here and it's legal, add it.
3936  (void)InsertFormula(LU, LUIdx, F);
3937  next:;
3938  }
3939 }
3940 
3941 /// Generate stride factor reuse formulae by making use of scaled-offset address
3942 /// modes, for example.
3943 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
3944  // Determine the integer type for the base formula.
3945  Type *IntTy = Base.getType();
3946  if (!IntTy) return;
3947 
3948  // If this Formula already has a scaled register, we can't add another one.
3949  // Try to unscale the formula to generate a better scale.
3950  if (Base.Scale != 0 && !Base.unscale())
3951  return;
3952 
3953  assert(Base.Scale == 0 && "unscale did not did its job!");
3954 
3955  // Check each interesting stride.
3956  for (int64_t Factor : Factors) {
3957  Base.Scale = Factor;
3958  Base.HasBaseReg = Base.BaseRegs.size() > 1;
3959  // Check whether this scale is going to be legal.
3960  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
3961  Base)) {
3962  // As a special-case, handle special out-of-loop Basic users specially.
3963  // TODO: Reconsider this special case.
3964  if (LU.Kind == LSRUse::Basic &&
3965  isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
3966  LU.AccessTy, Base) &&
3967  LU.AllFixupsOutsideLoop)
3968  LU.Kind = LSRUse::Special;
3969  else
3970  continue;
3971  }
3972  // For an ICmpZero, negating a solitary base register won't lead to
3973  // new solutions.
3974  if (LU.Kind == LSRUse::ICmpZero &&
3975  !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
3976  continue;
3977  // For each addrec base reg, if its loop is current loop, apply the scale.
3978  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
3979  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
3980  if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
3981  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
3982  if (FactorS->isZero())
3983  continue;
3984  // Divide out the factor, ignoring high bits, since we'll be
3985  // scaling the value back up in the end.
3986  if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
3987  // TODO: This could be optimized to avoid all the copying.
3988  Formula F = Base;
3989  F.ScaledReg = Quotient;
3990  F.deleteBaseReg(F.BaseRegs[i]);
3991  // The canonical representation of 1*reg is reg, which is already in
3992  // Base. In that case, do not try to insert the formula, it will be
3993  // rejected anyway.
3994  if (F.Scale == 1 && (F.BaseRegs.empty() ||
3995  (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
3996  continue;
3997  // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
3998  // non canonical Formula with ScaledReg's loop not being L.
3999  if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4000  F.canonicalize(*L);
4001  (void)InsertFormula(LU, LUIdx, F);
4002  }
4003  }
4004  }
4005  }
4006 }
4007 
4008 /// Generate reuse formulae from different IV types.
4009 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4010  // Don't bother truncating symbolic values.
4011  if (Base.BaseGV) return;
4012 
4013  // Determine the integer type for the base formula.
4014  Type *DstTy = Base.getType();
4015  if (!DstTy) return;
4016  DstTy = SE.getEffectiveSCEVType(DstTy);
4017 
4018  for (Type *SrcTy : Types) {
4019  if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4020  Formula F = Base;
4021 
4022  // Sometimes SCEV is able to prove zero during ext transform. It may
4023  // happen if SCEV did not do all possible transforms while creating the
4024  // initial node (maybe due to depth limitations), but it can do them while
4025  // taking ext.
4026  if (F.ScaledReg) {
4027  const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4028  if (NewScaledReg->isZero())
4029  continue;
4030  F.ScaledReg = NewScaledReg;
4031  }
4032  bool HasZeroBaseReg = false;
4033  for (const SCEV *&BaseReg : F.BaseRegs) {
4034  const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4035  if (NewBaseReg->isZero()) {
4036  HasZeroBaseReg = true;
4037  break;
4038  }
4039  BaseReg = NewBaseReg;
4040  }
4041  if (HasZeroBaseReg)
4042  continue;
4043 
4044  // TODO: This assumes we've done basic processing on all uses and
4045  // have an idea what the register usage is.
4046  if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4047  continue;
4048 
4049  F.canonicalize(*L);
4050  (void)InsertFormula(LU, LUIdx, F);
4051  }
4052  }
4053 }
4054 
4055 namespace {
4056 
4057 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4058 /// modifications so that the search phase doesn't have to worry about the data
4059 /// structures moving underneath it.
4060 struct WorkItem {
4061  size_t LUIdx;
4062  int64_t Imm;
4063  const SCEV *OrigReg;
4064 
4065  WorkItem(size_t LI, int64_t I, const SCEV *R)
4066  : LUIdx(LI), Imm(I), OrigReg(R) {}
4067 
4068  void print(raw_ostream &OS) const;
4069  void dump() const;
4070 };
4071 
4072 } // end anonymous namespace
4073 
4074 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4075 void WorkItem::print(raw_ostream &OS) const {
4076  OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4077  << " , add offset " << Imm;
4078 }
4079 
4080 LLVM_DUMP_METHOD void WorkItem::dump() const {
4081  print(errs()); errs() << '\n';
4082 }
4083 #endif
4084 
4085 /// Look for registers which are a constant distance apart and try to form reuse
4086 /// opportunities between them.
4087 void LSRInstance::GenerateCrossUseConstantOffsets() {
4088  // Group the registers by their value without any added constant offset.
4089  using ImmMapTy = std::map<int64_t, const SCEV *>;
4090 
4092  DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4094  for (const SCEV *Use : RegUses) {
4095  const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4096  int64_t Imm = ExtractImmediate(Reg, SE);
4097  auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4098  if (Pair.second)
4099  Sequence.push_back(Reg);
4100  Pair.first->second.insert(std::make_pair(Imm, Use));
4101  UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4102  }
4103 
4104  // Now examine each set of registers with the same base value. Build up
4105  // a list of work to do and do the work in a separate step so that we're
4106  // not adding formulae and register counts while we're searching.
4107  SmallVector<WorkItem, 32> WorkItems;
4108  SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4109  for (const SCEV *Reg : Sequence) {
4110  const ImmMapTy &Imms = Map.find(Reg)->second;
4111 
4112  // It's not worthwhile looking for reuse if there's only one offset.
4113  if (Imms.size() == 1)
4114  continue;
4115 
4116  LLVM_DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
4117  for (const auto &Entry
4118  : Imms) dbgs()
4119  << ' ' << Entry.first;
4120  dbgs() << '\n');
4121 
4122  // Examine each offset.
4123  for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4124  J != JE; ++J) {
4125  const SCEV *OrigReg = J->second;
4126 
4127  int64_t JImm = J->first;
4128  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4129 
4130  if (!isa<SCEVConstant>(OrigReg) &&
4131  UsedByIndicesMap[Reg].count() == 1) {
4132  LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg
4133  << '\n');
4134  continue;
4135  }
4136 
4137  // Conservatively examine offsets between this orig reg a few selected
4138  // other orig regs.
4139  int64_t First = Imms.begin()->first;
4140  int64_t Last = std::prev(Imms.end())->first;
4141  // Compute (First + Last) / 2 without overflow using the fact that
4142  // First + Last = 2 * (First + Last) + (First ^ Last).
4143  int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4144  // If the result is negative and First is odd and Last even (or vice versa),
4145  // we rounded towards -inf. Add 1 in that case, to round towards 0.
4146  Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4147  ImmMapTy::const_iterator OtherImms[] = {
4148  Imms.begin(), std::prev(Imms.end()),
4149  Imms.lower_bound(Avg)};
4150  for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4151  ImmMapTy::const_iterator M = OtherImms[i];
4152  if (M == J || M == JE) continue;
4153 
4154  // Compute the difference between the two.
4155  int64_t Imm = (uint64_t)JImm - M->first;
4156  for (unsigned LUIdx : UsedByIndices.set_bits())
4157  // Make a memo of this use, offset, and register tuple.
4158  if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4159  WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4160  }
4161  }
4162  }
4163 
4164  Map.clear();
4165  Sequence.clear();
4166  UsedByIndicesMap.clear();
4167  UniqueItems.clear();
4168 
4169  // Now iterate through the worklist and add new formulae.
4170  for (const WorkItem &WI : WorkItems) {
4171  size_t LUIdx = WI.LUIdx;
4172  LSRUse &LU = Uses[LUIdx];
4173  int64_t Imm = WI.Imm;
4174  const SCEV *OrigReg = WI.OrigReg;
4175 
4176  Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4177  const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4178  unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4179 
4180  // TODO: Use a more targeted data structure.
4181  for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4182  Formula F = LU.Formulae[L];
4183  // FIXME: The code for the scaled and unscaled registers looks
4184  // very similar but slightly different. Investigate if they
4185  // could be merged. That way, we would not have to unscale the
4186  // Formula.
4187  F.unscale();
4188  // Use the immediate in the scaled register.
4189  if (F.ScaledReg == OrigReg) {
4190  int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4191  // Don't create 50 + reg(-50).
4192  if (F.referencesReg(SE.getSCEV(
4193  ConstantInt::get(IntTy, -(uint64_t)Offset))))
4194  continue;
4195  Formula NewF = F;
4196  NewF.BaseOffset = Offset;
4197  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4198  NewF))
4199  continue;
4200  NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4201 
4202  // If the new scale is a constant in a register, and adding the constant
4203  // value to the immediate would produce a value closer to zero than the
4204  // immediate itself, then the formula isn't worthwhile.
4205  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4206  if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4207  (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4208  .ule(std::abs(NewF.BaseOffset)))
4209  continue;
4210 
4211  // OK, looks good.
4212  NewF.canonicalize(*this->L);
4213  (void)InsertFormula(LU, LUIdx, NewF);
4214  } else {
4215  // Use the immediate in a base register.
4216  for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4217  const SCEV *BaseReg = F.BaseRegs[N];
4218  if (BaseReg != OrigReg)
4219  continue;
4220  Formula NewF = F;
4221  NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4222  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4223  LU.Kind, LU.AccessTy, NewF)) {
4224  if (TTI.shouldFavorPostInc() &&
4225  mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4226  continue;
4227  if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4228  continue;
4229  NewF = F;
4230  NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4231  }
4232  NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4233 
4234  // If the new formula has a constant in a register, and adding the
4235  // constant value to the immediate would produce a value closer to
4236  // zero than the immediate itself, then the formula isn't worthwhile.
4237  for (const SCEV *NewReg : NewF.BaseRegs)
4238  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4239  if ((C->getAPInt() + NewF.BaseOffset)
4240  .abs()
4241  .slt(std::abs(NewF.BaseOffset)) &&
4242  (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4243  countTrailingZeros<uint64_t>(NewF.BaseOffset))
4244  goto skip_formula;
4245 
4246  // Ok, looks good.
4247  NewF.canonicalize(*this->L);
4248  (void)InsertFormula(LU, LUIdx, NewF);
4249  break;
4250  skip_formula:;
4251  }
4252  }
4253  }
4254  }
4255 }
4256 
4257 /// Generate formulae for each use.
4258 void
4259 LSRInstance::GenerateAllReuseFormulae() {
4260  // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4261  // queries are more precise.
4262  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4263  LSRUse &LU = Uses[LUIdx];
4264  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4265  GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4266  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4267  GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4268  }
4269  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4270  LSRUse &LU = Uses[LUIdx];
4271  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4272  GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4273  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4274  GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4275  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4276  GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4277  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4278  GenerateScales(LU, LUIdx, LU.Formulae[i]);
4279  }
4280  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4281  LSRUse &LU = Uses[LUIdx];
4282  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4283  GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4284  }
4285 
4286  GenerateCrossUseConstantOffsets();
4287 
4288  LLVM_DEBUG(dbgs() << "\n"
4289  "After generating reuse formulae:\n";
4290  print_uses(dbgs()));
4291 }
4292 
4293 /// If there are multiple formulae with the same set of registers used
4294 /// by other uses, pick the best one and delete the others.
4295 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4296  DenseSet<const SCEV *> VisitedRegs;
4299 #ifndef NDEBUG
4300  bool ChangedFormulae = false;
4301 #endif
4302 
4303  // Collect the best formula for each unique set of shared registers. This
4304  // is reset for each use.
4305  using BestFormulaeTy =
4306  DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4307 
4308  BestFormulaeTy BestFormulae;
4309 
4310  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4311  LSRUse &LU = Uses[LUIdx];
4312  LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
4313  dbgs() << '\n');
4314 
4315  bool Any = false;
4316  for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4317  FIdx != NumForms; ++FIdx) {
4318  Formula &F = LU.Formulae[FIdx];
4319 
4320  // Some formulas are instant losers. For example, they may depend on
4321  // nonexistent AddRecs from other loops. These need to be filtered
4322  // immediately, otherwise heuristics could choose them over others leading
4323  // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4324  // avoids the need to recompute this information across formulae using the
4325  // same bad AddRec. Passing LoserRegs is also essential unless we remove
4326  // the corresponding bad register from the Regs set.
4327  Cost CostF(L, SE, TTI);
4328  Regs.clear();
4329  CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4330  if (CostF.isLoser()) {
4331  // During initial formula generation, undesirable formulae are generated
4332  // by uses within other loops that have some non-trivial address mode or
4333  // use the postinc form of the IV. LSR needs to provide these formulae
4334  // as the basis of rediscovering the desired formula that uses an AddRec
4335  // corresponding to the existing phi. Once all formulae have been
4336  // generated, these initial losers may be pruned.
4337  LLVM_DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
4338  dbgs() << "\n");
4339  }
4340  else {
4342  for (const SCEV *Reg : F.BaseRegs) {
4343  if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4344  Key.push_back(Reg);
4345  }
4346  if (F.ScaledReg &&
4347  RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4348  Key.push_back(F.ScaledReg);
4349  // Unstable sort by host order ok, because this is only used for
4350  // uniquifying.
4351  llvm::sort(Key);
4352 
4353  std::pair<BestFormulaeTy::const_iterator, bool> P =
4354  BestFormulae.insert(std::make_pair(Key, FIdx));
4355  if (P.second)
4356  continue;
4357 
4358  Formula &Best = LU.Formulae[P.first->second];
4359 
4360  Cost CostBest(L, SE, TTI);
4361  Regs.clear();
4362  CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4363  if (CostF.isLess(CostBest))
4364  std::swap(F, Best);
4365  LLVM_DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4366  dbgs() << "\n"
4367  " in favor of formula ";
4368  Best.print(dbgs()); dbgs() << '\n');
4369  }
4370 #ifndef NDEBUG
4371  ChangedFormulae = true;
4372 #endif
4373  LU.DeleteFormula(F);
4374  --FIdx;
4375  --NumForms;
4376  Any = true;
4377  }
4378 
4379  // Now that we've filtered out some formulae, recompute the Regs set.
4380  if (Any)
4381  LU.RecomputeRegs(LUIdx, RegUses);
4382 
4383  // Reset this to prepare for the next use.
4384  BestFormulae.clear();
4385  }
4386 
4387  LLVM_DEBUG(if (ChangedFormulae) {
4388  dbgs() << "\n"
4389  "After filtering out undesirable candidates:\n";
4390  print_uses(dbgs());
4391  });
4392 }
4393 
4394 /// Estimate the worst-case number of solutions the solver might have to
4395 /// consider. It almost never considers this many solutions because it prune the
4396 /// search space, but the pruning isn't always sufficient.
4397 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4398  size_t Power = 1;
4399  for (const LSRUse &LU : Uses) {
4400  size_t FSize = LU.Formulae.size();
4401  if (FSize >= ComplexityLimit) {
4402  Power = ComplexityLimit;
4403  break;
4404  }
4405  Power *= FSize;
4406  if (Power >= ComplexityLimit)
4407  break;
4408  }
4409  return Power;
4410 }
4411 
4412 /// When one formula uses a superset of the registers of another formula, it
4413 /// won't help reduce register pressure (though it may not necessarily hurt
4414 /// register pressure); remove it to simplify the system.
4415 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4416  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4417  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4418 
4419  LLVM_DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4420  "which use a superset of registers used by other "
4421  "formulae.\n");
4422 
4423  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4424  LSRUse &LU = Uses[LUIdx];
4425  bool Any = false;
4426  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4427  Formula &F = LU.Formulae[i];
4428  // Look for a formula with a constant or GV in a register. If the use
4429  // also has a formula with that same value in an immediate field,
4430  // delete the one that uses a register.
4432  I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4433  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4434  Formula NewF = F;
4435  //FIXME: Formulas should store bitwidth to do wrapping properly.
4436  // See PR41034.
4437  NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4438  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4439  (I - F.BaseRegs.begin()));
4440  if (LU.HasFormulaWithSameRegs(NewF)) {
4441  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4442  dbgs() << '\n');
4443  LU.DeleteFormula(F);
4444  --i;
4445  --e;
4446  Any = true;
4447  break;
4448  }
4449  } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4450  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4451  if (!F.BaseGV) {
4452  Formula NewF = F;
4453  NewF.BaseGV = GV;
4454  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4455  (I - F.BaseRegs.begin()));
4456  if (LU.HasFormulaWithSameRegs(NewF)) {
4457  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4458  dbgs() << '\n');
4459  LU.DeleteFormula(F);
4460  --i;
4461  --e;
4462  Any = true;
4463  break;
4464  }
4465  }
4466  }
4467  }
4468  }
4469  if (Any)
4470  LU.RecomputeRegs(LUIdx, RegUses);
4471  }
4472 
4473  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4474  }
4475 }
4476 
4477 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4478 /// allocate a single register for them.
4479 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4480  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4481  return;
4482 
4483  LLVM_DEBUG(
4484  dbgs() << "The search space is too complex.\n"
4485  "Narrowing the search space by assuming that uses separated "
4486  "by a constant offset will use the same registers.\n");
4487 
4488  // This is especially useful for unrolled loops.
4489 
4490  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4491  LSRUse &LU = Uses[LUIdx];
4492  for (const Formula &F : LU.Formulae) {
4493  if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4494  continue;
4495 
4496  LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4497  if (!LUThatHas)
4498  continue;
4499 
4500  if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4501  LU.Kind, LU.AccessTy))
4502  continue;
4503 
4504  LLVM_DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4505 
4506  LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4507 
4508  // Transfer the fixups of LU to LUThatHas.
4509  for (LSRFixup &Fixup : LU.Fixups) {
4510  Fixup.Offset += F.BaseOffset;
4511  LUThatHas->pushFixup(Fixup);
4512  LLVM_DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4513  }
4514 
4515  // Delete formulae from the new use which are no longer legal.
4516  bool Any = false;
4517  for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4518  Formula &F = LUThatHas->Formulae[i];
4519  if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4520  LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4521  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4522  LUThatHas->DeleteFormula(F);
4523  --i;
4524  --e;
4525  Any = true;
4526  }
4527  }
4528 
4529  if (Any)
4530  LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4531 
4532  // Delete the old use.
4533  DeleteUse(LU, LUIdx);
4534  --LUIdx;
4535  --NumUses;
4536  break;
4537  }
4538  }
4539 
4540  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4541 }
4542 
4543 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4544 /// we've done more filtering, as it may be able to find more formulae to
4545 /// eliminate.
4546 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4547  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4548  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4549 
4550  LLVM_DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4551  "undesirable dedicated registers.\n");
4552 
4553  FilterOutUndesirableDedicatedRegisters();
4554 
4555  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4556  }
4557 }
4558 
4559 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4560 /// Pick the best one and delete the others.
4561 /// This narrowing heuristic is to keep as many formulae with different
4562 /// Scale and ScaledReg pair as possible while narrowing the search space.
4563 /// The benefit is that it is more likely to find out a better solution
4564 /// from a formulae set with more Scale and ScaledReg variations than
4565 /// a formulae set with the same Scale and ScaledReg. The picking winner
4566 /// reg heuristic will often keep the formulae with the same Scale and
4567 /// ScaledReg and filter others, and we want to avoid that if possible.
4568 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4569  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4570  return;
4571 
4572  LLVM_DEBUG(
4573  dbgs() << "The search space is too complex.\n"
4574  "Narrowing the search space by choosing the best Formula "
4575  "from the Formulae with the same Scale and ScaledReg.\n");
4576 
4577  // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4578  using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4579 
4580  BestFormulaeTy BestFormulae;
4581 #ifndef NDEBUG
4582  bool ChangedFormulae = false;
4583 #endif
4584  DenseSet<const SCEV *> VisitedRegs;
4586 
4587  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4588  LSRUse &LU = Uses[LUIdx];
4589  LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
4590  dbgs() << '\n');
4591 
4592  // Return true if Formula FA is better than Formula FB.
4593  auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4594  // First we will try to choose the Formula with fewer new registers.
4595  // For a register used by current Formula, the more the register is
4596  // shared among LSRUses, the less we increase the register number
4597  // counter of the formula.
4598  size_t FARegNum = 0;
4599  for (const SCEV *Reg : FA.BaseRegs) {
4600  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4601  FARegNum += (NumUses - UsedByIndices.count() + 1);
4602  }
4603  size_t FBRegNum = 0;
4604  for (const SCEV *Reg : FB.BaseRegs) {
4605  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4606  FBRegNum += (NumUses - UsedByIndices.count() + 1);
4607  }
4608  if (FARegNum != FBRegNum)
4609  return FARegNum < FBRegNum;
4610 
4611  // If the new register numbers are the same, choose the Formula with
4612  // less Cost.
4613  Cost CostFA(L, SE, TTI);
4614  Cost CostFB(L, SE, TTI);
4615  Regs.clear();
4616  CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4617  Regs.clear();
4618  CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4619  return CostFA.isLess(CostFB);
4620  };
4621 
4622  bool Any = false;
4623  for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4624  ++FIdx) {
4625  Formula &F = LU.Formulae[FIdx];
4626  if (!F.ScaledReg)
4627  continue;
4628  auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4629  if (P.second)
4630  continue;
4631 
4632  Formula &Best = LU.Formulae[P.first->second];
4633  if (IsBetterThan(F, Best))
4634  std::swap(F, Best);
4635  LLVM_DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4636  dbgs() << "\n"
4637  " in favor of formula ";
4638  Best.print(dbgs()); dbgs() << '\n');
4639 #ifndef NDEBUG
4640  ChangedFormulae = true;
4641 #endif
4642  LU.DeleteFormula(F);
4643  --FIdx;
4644  --NumForms;
4645  Any = true;
4646  }
4647  if (Any)
4648  LU.RecomputeRegs(LUIdx, RegUses);
4649 
4650  // Reset this to prepare for the next use.
4651  BestFormulae.clear();
4652  }
4653 
4654  LLVM_DEBUG(if (ChangedFormulae) {
4655  dbgs() << "\n"
4656  "After filtering out undesirable candidates:\n";
4657  print_uses(dbgs());
4658  });
4659 }
4660 
4661 /// The function delete formulas with high registers number expectation.
4662 /// Assuming we don't know the value of each formula (already delete
4663 /// all inefficient), generate probability of not selecting for each
4664 /// register.
4665 /// For example,
4666 /// Use1:
4667 /// reg(a) + reg({0,+,1})
4668 /// reg(a) + reg({-1,+,1}) + 1
4669 /// reg({a,+,1})
4670 /// Use2:
4671 /// reg(b) + reg({0,+,1})
4672 /// reg(b) + reg({-1,+,1}) + 1
4673 /// reg({b,+,1})
4674 /// Use3:
4675 /// reg(c) + reg(b) + reg({0,+,1})
4676 /// reg(c) + reg({b,+,1})
4677 ///
4678 /// Probability of not selecting
4679 /// Use1 Use2 Use3
4680 /// reg(a) (1/3) * 1 * 1
4681 /// reg(b) 1 * (1/3) * (1/2)
4682 /// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4683 /// reg({-1,+,1}) (2/3) * (2/3) * 1
4684 /// reg({a,+,1}) (2/3) * 1 * 1
4685 /// reg({b,+,1}) 1 * (2/3) * (2/3)
4686 /// reg(c) 1 * 1 * 0
4687 ///
4688 /// Now count registers number mathematical expectation for each formula:
4689 /// Note that for each use we exclude probability if not selecting for the use.
4690 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4691 /// probabilty 1/3 of not selecting for Use1).
4692 /// Use1:
4693 /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4694 /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4695 /// reg({a,+,1}) 1
4696 /// Use2:
4697 /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4698 /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4699 /// reg({b,+,1}) 2/3
4700 /// Use3:
4701 /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4702 /// reg(c) + reg({b,+,1}) 1 + 2/3
4703 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4704  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4705  return;
4706  // Ok, we have too many of formulae on our hands to conveniently handle.
4707  // Use a rough heuristic to thin out the list.
4708 
4709  // Set of Regs wich will be 100% used in final solution.
4710  // Used in each formula of a solution (in example above this is reg(c)).
4711  // We can skip them in calculations.
4713  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4714 
4715  // Map each register to probability of not selecting
4717  for (const SCEV *Reg : RegUses) {
4718  if (UniqRegs.count(Reg))
4719  continue;
4720  float PNotSel = 1;
4721  for (const LSRUse &LU : Uses) {
4722  if (!LU.Regs.count(Reg))
4723  continue;
4724  float P = LU.getNotSelectedProbability(Reg);
4725  if (P != 0.0)
4726  PNotSel *= P;
4727  else
4728  UniqRegs.insert(Reg);
4729  }
4730  RegNumMap.insert(std::make_pair(Reg, PNotSel));
4731  }
4732 
4733  LLVM_DEBUG(
4734  dbgs() << "Narrowing the search space by deleting costly formulas\n");
4735 
4736  // Delete formulas where registers number expectation is high.
4737  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4738  LSRUse &LU = Uses[LUIdx];
4739  // If nothing to delete - continue.
4740  if (LU.Formulae.size() < 2)
4741  continue;
4742  // This is temporary solution to test performance. Float should be
4743  // replaced with round independent type (based on integers) to avoid
4744  // different results for different target builds.
4745  float FMinRegNum = LU.Formulae[0].getNumRegs();
4746  float FMinARegNum = LU.Formulae[0].getNumRegs();
4747  size_t MinIdx = 0;
4748  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4749  Formula &F = LU.Formulae[i];
4750  float FRegNum = 0;
4751  float FARegNum = 0;
4752  for (const SCEV *BaseReg : F.BaseRegs) {
4753  if (UniqRegs.count(BaseReg))
4754  continue;
4755  FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4756  if (isa<SCEVAddRecExpr>(BaseReg))
4757  FARegNum +=
4758  RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4759  }
4760  if (const SCEV *ScaledReg = F.ScaledReg) {
4761  if (!UniqRegs.count(ScaledReg)) {
4762  FRegNum +=
4763  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4764  if (isa<SCEVAddRecExpr>(ScaledReg))
4765  FARegNum +=
4766  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4767  }
4768  }
4769  if (FMinRegNum > FRegNum ||
4770  (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4771  FMinRegNum = FRegNum;
4772  FMinARegNum = FARegNum;
4773  MinIdx = i;
4774  }
4775  }
4776  LLVM_DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());
4777  dbgs() << " with min reg num " << FMinRegNum << '\n');
4778  if (MinIdx != 0)
4779  std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4780  while (LU.Formulae.size() != 1) {
4781  LLVM_DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());
4782  dbgs() << '\n');
4783  LU.Formulae.pop_back();
4784  }
4785  LU.RecomputeRegs(LUIdx, RegUses);
4786  assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
4787  Formula &F = LU.Formulae[0];
4788  LLVM_DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n');
4789  // When we choose the formula, the regs become unique.
4790  UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4791  if (F.ScaledReg)
4792  UniqRegs.insert(F.ScaledReg);
4793  }
4794  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4795 }
4796 
4797 /// Pick a register which seems likely to be profitable, and then in any use
4798 /// which has any reference to that register, delete all formulae which do not
4799 /// reference that register.
4800 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4801  // With all other options exhausted, loop until the system is simple
4802  // enough to handle.
4804  while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4805  // Ok, we have too many of formulae on our hands to conveniently handle.
4806  // Use a rough heuristic to thin out the list.
4807  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4808 
4809  // Pick the register which is used by the most LSRUses, which is likely
4810  // to be a good reuse register candidate.
4811  const SCEV *Best = nullptr;
4812  unsigned BestNum = 0;
4813  for (const SCEV *Reg : RegUses) {
4814  if (Taken.count(Reg))
4815  continue;
4816  if (!Best) {
4817  Best = Reg;
4818  BestNum = RegUses.getUsedByIndices(Reg).count();
4819  } else {
4820  unsigned Count = RegUses.getUsedByIndices(Reg).count();
4821  if (Count > BestNum) {
4822  Best = Reg;
4823  BestNum = Count;
4824  }
4825  }
4826  }
4827 
4828  LLVM_DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4829  << " will yield profitable reuse.\n");
4830  Taken.insert(Best);
4831 
4832  // In any use with formulae which references this register, delete formulae
4833  // which don't reference it.
4834  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4835  LSRUse &LU = Uses[LUIdx];
4836  if (!LU.Regs.count(Best)) continue;
4837 
4838  bool Any = false;
4839  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4840  Formula &F = LU.Formulae[i];
4841  if (!F.referencesReg(Best)) {
4842  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4843  LU.DeleteFormula(F);
4844  --e;
4845  --i;
4846  Any = true;
4847  assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
4848  continue;
4849  }
4850  }
4851 
4852  if (Any)
4853  LU.RecomputeRegs(LUIdx, RegUses);
4854  }
4855 
4856  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4857  }
4858 }
4859 
4860 /// If there are an extraordinary number of formulae to choose from, use some
4861 /// rough heuristics to prune down the number of formulae. This keeps the main
4862 /// solver from taking an extraordinary amount of time in some worst-case
4863 /// scenarios.
4864 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
4865  NarrowSearchSpaceByDetectingSupersets();
4866  NarrowSearchSpaceByCollapsingUnrolledCode();
4867  NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
4868  if (FilterSameScaledReg)
4869  NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
4870  if (LSRExpNarrow)
4871  NarrowSearchSpaceByDeletingCostlyFormulas();
4872  else
4873  NarrowSearchSpaceByPickingWinnerRegs();
4874 }
4875 
4876 /// This is the recursive solver.
4877 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
4878  Cost &SolutionCost,
4880  const Cost &CurCost,
4881  const SmallPtrSet<const SCEV *, 16> &CurRegs,
4882  DenseSet<const SCEV *> &VisitedRegs) const {
4883  // Some ideas:
4884  // - prune more:
4885  // - use more aggressive filtering
4886  // - sort the formula so that the most profitable solutions are found first
4887  // - sort the uses too
4888  // - search faster:
4889  // - don't compute a cost, and then compare. compare while computing a cost
4890  // and bail early.
4891  // - track register sets with SmallBitVector
4892 
4893  const LSRUse &LU = Uses[Workspace.size()];
4894 
4895  // If this use references any register that's already a part of the
4896  // in-progress solution, consider it a requirement that a formula must
4897  // reference that register in order to be considered. This prunes out
4898  // unprofitable searching.
4900  for (const SCEV *S : CurRegs)
4901  if (LU.Regs.count(S))
4902  ReqRegs.insert(S);
4903 
4905  Cost NewCost(L, SE, TTI);
4906  for (const Formula &F : LU.Formulae) {
4907  // Ignore formulae which may not be ideal in terms of register reuse of
4908  // ReqRegs. The formula should use all required registers before
4909  // introducing new ones.
4910  int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
4911  for (const SCEV *Reg : ReqRegs) {
4912  if ((F.ScaledReg && F.ScaledReg == Reg) ||
4913  is_contained(F.BaseRegs, Reg)) {
4914  --NumReqRegsToFind;
4915  if (NumReqRegsToFind == 0)
4916  break;
4917  }
4918  }
4919  if (NumReqRegsToFind != 0) {
4920  // If none of the formulae satisfied the required registers, then we could
4921  // clear ReqRegs and try again. Currently, we simply give up in this case.
4922  continue;
4923  }
4924 
4925  // Evaluate the cost of the current formula. If it's already worse than
4926  // the current best, prune the search at that point.
4927  NewCost = CurCost;
4928  NewRegs = CurRegs;
4929  NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
4930  if (NewCost.isLess(SolutionCost)) {
4931  Workspace.push_back(&F);
4932  if (Workspace.size() != Uses.size()) {
4933  SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
4934  NewRegs, VisitedRegs);
4935  if (F.getNumRegs() == 1 && Workspace.size() == 1)
4936  VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
4937  } else {
4938  LLVM_DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
4939  dbgs() << ".\nRegs:\n";
4940  for (const SCEV *S : NewRegs) dbgs()
4941  << "- " << *S << "\n";
4942  dbgs() << '\n');
4943 
4944  SolutionCost = NewCost;
4945  Solution = Workspace;
4946  }
4947  Workspace.pop_back();
4948  }
4949  }
4950 }
4951 
4952 /// Choose one formula from each use. Return the results in the given Solution
4953 /// vector.
4954 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
4956  Cost SolutionCost(L, SE, TTI);
4957  SolutionCost.Lose();
4958  Cost CurCost(L, SE, TTI);
4960  DenseSet<const SCEV *> VisitedRegs;
4961  Workspace.reserve(Uses.size());
4962 
4963  // SolveRecurse does all the work.
4964  SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
4965  CurRegs, VisitedRegs);
4966  if (Solution.empty()) {
4967  LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
4968  return;
4969  }
4970 
4971  // Ok, we've now made all our decisions.
4972  LLVM_DEBUG(dbgs() << "\n"
4973  "The chosen solution requires ";
4974  SolutionCost.print(dbgs()); dbgs() << ":\n";
4975  for (size_t i = 0, e = Uses.size(); i != e; ++i) {
4976  dbgs() << " ";
4977  Uses[i].print(dbgs());
4978  dbgs() << "\n"
4979  " ";
4980  Solution[i]->print(dbgs());
4981  dbgs() << '\n';
4982  });
4983 
4984  assert(Solution.size() == Uses.size() && "Malformed solution!");
4985 }
4986 
4987 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
4988 /// we can go while still being dominated by the input positions. This helps
4989 /// canonicalize the insert position, which encourages sharing.
4991 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
4992  const SmallVectorImpl<Instruction *> &Inputs)
4993  const {
4994  Instruction *Tentative = &*IP;
4995  while (true) {
4996  bool AllDominate = true;
4997  Instruction *BetterPos = nullptr;
4998  // Don't bother attempting to insert before a catchswitch, their basic block
4999  // cannot have other non-PHI instructions.
5000  if (isa<CatchSwitchInst>(Tentative))
5001  return IP;
5002 
5003  for (Instruction *Inst : Inputs) {
5004  if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5005  AllDominate = false;
5006  break;
5007  }
5008  // Attempt to find an insert position in the middle of the block,
5009  // instead of at the end, so that it can be used for other expansions.
5010  if (Tentative->getParent() == Inst->getParent() &&
5011  (!BetterPos || !DT.dominates(Inst, BetterPos)))
5012  BetterPos = &*std::next(BasicBlock::iterator(Inst));
5013  }
5014  if (!AllDominate)
5015  break;
5016  if (BetterPos)
5017  IP = BetterPos->getIterator();
5018  else
5019  IP = Tentative->getIterator();
5020 
5021  const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5022  unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5023 
5024  BasicBlock *IDom;
5025  for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5026  if (!Rung) return IP;
5027  Rung = Rung->getIDom();
5028  if (!Rung) return IP;
5029  IDom = Rung->getBlock();
5030 
5031  // Don't climb into a loop though.
5032  const Loop *IDomLoop = LI.getLoopFor(IDom);
5033  unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5034  if (IDomDepth <= IPLoopDepth &&
5035  (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5036  break;
5037  }
5038 
5039  Tentative = IDom->getTerminator();
5040  }
5041 
5042  return IP;
5043 }
5044 
5045 /// Determine an input position which will be dominated by the operands and
5046 /// which will dominate the result.
5048 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
5049  const LSRFixup &LF,
5050  const LSRUse &LU,
5051  SCEVExpander &Rewriter) const {
5052  // Collect some instructions which must be dominated by the
5053  // expanding replacement. These must be dominated by any operands that
5054  // will be required in the expansion.
5056  if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5057  Inputs.push_back(I);
5058  if (LU.Kind == LSRUse::ICmpZero)
5059  if (Instruction *I =
5060  dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5061  Inputs.push_back(I);
5062  if (LF.PostIncLoops.count(L)) {
5063  if (LF.isUseFullyOutsideLoop(L))
5064  Inputs.push_back(L->getLoopLatch()->getTerminator());
5065  else
5066  Inputs.push_back(IVIncInsertPos);
5067  }
5068  // The expansion must also be dominated by the increment positions of any
5069  // loops it for which it is using post-inc mode.
5070  for (const Loop *PIL : LF.PostIncLoops) {
5071  if (PIL == L) continue;
5072 
5073  // Be dominated by the loop exit.
5074  SmallVector<BasicBlock *, 4> ExitingBlocks;
5075  PIL->getExitingBlocks(ExitingBlocks);
5076  if (!ExitingBlocks.empty()) {
5077  BasicBlock *BB = ExitingBlocks[0];
5078  for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5079  BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5080  Inputs.push_back(BB->getTerminator());
5081  }
5082  }
5083 
5084  assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
5085  && !isa<DbgInfoIntrinsic>(LowestIP) &&
5086  "Insertion point must be a normal instruction");
5087 
5088  // Then, climb up the immediate dominator tree as far as we can go while
5089  // still being dominated by the input positions.
5090  BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5091 
5092  // Don't insert instructions before PHI nodes.
5093  while (isa<PHINode>(IP)) ++IP;
5094 
5095  // Ignore landingpad instructions.
5096  while (IP->isEHPad()) ++IP;
5097 
5098  // Ignore debug intrinsics.
5099  while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5100 
5101  // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5102  // IP consistent across expansions and allows the previously inserted
5103  // instructions to be reused by subsequent expansion.
5104  while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5105  ++IP;
5106 
5107  return IP;
5108 }
5109 
5110 /// Emit instructions for the leading candidate expression for this LSRUse (this
5111 /// is called "expanding").
5112 Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5113  const Formula &F, BasicBlock::iterator IP,
5114  SCEVExpander &Rewriter,
5115  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5116  if (LU.RigidFormula)
5117  return LF.OperandValToReplace;
5118 
5119  // Determine an input position which will be dominated by the operands and
5120  // which will dominate the result.
5121  IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
5122  Rewriter.setInsertPoint(&*IP);
5123 
5124  // Inform the Rewriter if we have a post-increment use, so that it can
5125  // perform an advantageous expansion.
5126  Rewriter.setPostInc(LF.PostIncLoops);
5127 
5128  // This is the type that the user actually needs.
5129  Type *OpTy = LF.OperandValToReplace->getType();
5130  // This will be the type that we'll initially expand to.
5131  Type *Ty = F.getType();
5132  if (!Ty)
5133  // No type known; just expand directly to the ultimate type.
5134  Ty = OpTy;
5135  else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5136  // Expand directly to the ultimate type if it's the right size.
5137  Ty = OpTy;
5138  // This is the type to do integer arithmetic in.
5139  Type *IntTy = SE.getEffectiveSCEVType(Ty);
5140 
5141  // Build up a list of operands to add together to form the full base.
5143 
5144  // Expand the BaseRegs portion.
5145  for (const SCEV *Reg : F.BaseRegs) {
5146  assert(!Reg->isZero() && "Zero allocated in a base register!");
5147 
5148  // If we're expanding for a post-inc user, make the post-inc adjustment.
5149  Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5150  Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5151  }
5152 
5153  // Expand the ScaledReg portion.
5154  Value *ICmpScaledV = nullptr;
5155  if (F.Scale != 0) {
5156  const SCEV *ScaledS = F.ScaledReg;
5157 
5158  // If we're expanding for a post-inc user, make the post-inc adjustment.
5159  PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5160  ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5161 
5162  if (LU.Kind == LSRUse::ICmpZero) {
5163  // Expand ScaleReg as if it was part of the base regs.
5164  if (F.Scale == 1)
5165  Ops.push_back(
5166  SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5167  else {
5168  // An interesting way of "folding" with an icmp is to use a negated
5169  // scale, which we'll implement by inserting it into the other operand
5170  // of the icmp.
5171  assert(F.Scale == -1 &&
5172  "The only scale supported by ICmpZero uses is -1!");
5173  ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5174  }
5175  } else {
5176  // Otherwise just expand the scaled register and an explicit scale,
5177  // which is expected to be matched as part of the address.
5178 
5179  // Flush the operand list to suppress SCEVExpander hoisting address modes.
5180  // Unless the addressing mode will not be folded.
5181  if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5182  isAMCompletelyFolded(TTI, LU, F)) {
5183  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5184  Ops.clear();
5185  Ops.push_back(SE.getUnknown(FullV));
5186  }
5187  ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5188  if (F.Scale != 1)
5189  ScaledS =
5190  SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5191  Ops.push_back(ScaledS);
5192  }
5193  }
5194 
5195  // Expand the GV portion.
5196  if (F.BaseGV) {
5197  // Flush the operand list to suppress SCEVExpander hoisting.
5198  if (!Ops.empty()) {
5199  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5200  Ops.clear();
5201  Ops.push_back(SE.getUnknown(FullV));
5202  }
5203  Ops.push_back(SE.getUnknown(F.BaseGV));
5204  }
5205 
5206  // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5207  // unfolded offsets. LSR assumes they both live next to their uses.
5208  if (!Ops.empty()) {
5209  Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5210  Ops.clear();
5211  Ops.push_back(SE.getUnknown(FullV));
5212  }
5213 
5214  // Expand the immediate portion.
5215  int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
5216  if (Offset != 0) {
5217  if (LU.Kind == LSRUse::ICmpZero) {
5218  // The other interesting way of "folding" with an ICmpZero is to use a
5219  // negated immediate.
5220  if (!ICmpScaledV)
5221  ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
5222  else {
5223  Ops.push_back(SE.getUnknown(ICmpScaledV));
5224  ICmpScaledV = ConstantInt::get(IntTy, Offset);
5225  }
5226  } else {
5227  // Just add the immediate values. These again are expected to be matched
5228  // as part of the address.
5229  Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
5230  }
5231  }
5232 
5233  // Expand the unfolded offset portion.
5234  int64_t UnfoldedOffset = F.UnfoldedOffset;
5235  if (UnfoldedOffset != 0) {
5236  // Just add the immediate values.
5238  UnfoldedOffset)));
5239  }
5240 
5241  // Emit instructions summing all the operands.
5242  const SCEV *FullS = Ops.empty() ?
5243  SE.getConstant(IntTy, 0) :
5244  SE.getAddExpr(Ops);
5245  Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5246 
5247  // We're done expanding now, so reset the rewriter.
5248  Rewriter.clearPostInc();
5249 
5250  // An ICmpZero Formula represents an ICmp which we're handling as a
5251  // comparison against zero. Now that we've expanded an expression for that
5252  // form, update the ICmp's other operand.
5253  if (LU.Kind == LSRUse::ICmpZero) {
5254  ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5255  DeadInsts.emplace_back(CI->getOperand(1));
5256  assert(!F.BaseGV && "ICmp does not support folding a global value and "
5257  "a scale at the same time!");
5258  if (F.Scale == -1) {
5259  if (ICmpScaledV->getType() != OpTy) {
5260  Instruction *Cast =
5261  CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
5262  OpTy, false),
5263  ICmpScaledV, OpTy, "tmp", CI);
5264  ICmpScaledV = Cast;
5265  }
5266  CI->setOperand(1, ICmpScaledV);
5267  } else {
5268  // A scale of 1 means that the scale has been expanded as part of the
5269  // base regs.
5270  assert((F.Scale == 0 || F.Scale == 1) &&
5271  "ICmp does not support folding a global value and "
5272  "a scale at the same time!");
5274  -(uint64_t)Offset);
5275  if (C->getType() != OpTy)
5277  OpTy, false),
5278  C, OpTy);
5279 
5280  CI->setOperand(1, C);
5281  }
5282  }
5283 
5284  return FullV;
5285 }
5286 
5287 /// Helper for Rewrite. PHI nodes are special because the use of their operands
5288 /// effectively happens in their predecessor blocks, so the expression may need
5289 /// to be expanded in multiple places.
5290 void LSRInstance::RewriteForPHI(
5291  PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5292  SCEVExpander &Rewriter, SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5294  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5295  if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5296  bool needUpdateFixups = false;
5297  BasicBlock *BB = PN->getIncomingBlock(i);
5298 
5299  // If this is a critical edge, split the edge so that we do not insert
5300  // the code on all predecessor/successor paths. We do this unless this
5301  // is the canonical backedge for this loop, which complicates post-inc
5302  // users.
5303  if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5304  !isa<IndirectBrInst>(BB->getTerminator()) &&
5305  !isa<CatchSwitchInst>(BB->getTerminator())) {
5306  BasicBlock *Parent = PN->getParent();
5307  Loop *PNLoop = LI.getLoopFor(Parent);
5308  if (!PNLoop || Parent != PNLoop->getHeader()) {
5309  // Split the critical edge.
5310  BasicBlock *NewBB = nullptr;
5311  if (!Parent->isLandingPad()) {
5312  NewBB = SplitCriticalEdge(BB, Parent,
5314  .setMergeIdenticalEdges()
5315  .setKeepOneInputPHIs());
5316  } else {
5318  SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DT, &LI);
5319  NewBB = NewBBs[0];
5320  }
5321  // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5322  // phi predecessors are identical. The simple thing to do is skip
5323  // splitting in this case rather than complicate the API.
5324  if (NewBB) {
5325  // If PN is outside of the loop and BB is in the loop, we want to
5326  // move the block to be immediately before the PHI block, not
5327  // immediately after BB.
5328  if (L->contains(BB) && !L->contains(PN))
5329  NewBB->moveBefore(PN->getParent());
5330 
5331  // Splitting the edge can reduce the number of PHI entries we have.
5332  e = PN->getNumIncomingValues();
5333  BB = NewBB;
5334  i = PN->getBasicBlockIndex(BB);
5335 
5336  needUpdateFixups = true;
5337  }
5338  }
5339  }
5340 
5341  std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5342  Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5343  if (!Pair.second)
5344  PN->setIncomingValue(i, Pair.first->second);
5345  else {
5346  Value *FullV = Expand(LU, LF, F, BB->getTerminator()->getIterator(),
5347  Rewriter, DeadInsts);
5348 
5349  // If this is reuse-by-noop-cast, insert the noop cast.
5350  Type *OpTy = LF.OperandValToReplace->getType();
5351  if (FullV->getType() != OpTy)
5352  FullV =
5354  OpTy, false),
5355  FullV, LF.OperandValToReplace->getType(),
5356  "tmp", BB->getTerminator());
5357 
5358  PN->setIncomingValue(i, FullV);
5359  Pair.first->second = FullV;
5360  }
5361 
5362  // If LSR splits critical edge and phi node has other pending
5363  // fixup operands, we need to update those pending fixups. Otherwise
5364  // formulae will not be implemented completely and some instructions
5365  // will not be eliminated.
5366  if (needUpdateFixups) {
5367  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5368  for (LSRFixup &Fixup : Uses[LUIdx].Fixups)
5369  // If fixup is supposed to rewrite some operand in the phi
5370  // that was just updated, it may be already moved to
5371  // another phi node. Such fixup requires update.
5372  if (Fixup.UserInst == PN) {
5373  // Check if the operand we try to replace still exists in the
5374  // original phi.
5375  bool foundInOriginalPHI = false;
5376  for (const auto &val : PN->incoming_values())
5377  if (val == Fixup.OperandValToReplace) {
5378  foundInOriginalPHI = true;
5379  break;
5380  }
5381 
5382  // If fixup operand found in original PHI - nothing to do.
5383  if (foundInOriginalPHI)
5384  continue;
5385 
5386  // Otherwise it might be moved to another PHI and requires update.
5387  // If fixup operand not found in any of the incoming blocks that
5388  // means we have already rewritten it - nothing to do.
5389  for (const auto &Block : PN->blocks())
5390  for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5391  ++I) {
5392  PHINode *NewPN = cast<PHINode>(I);
5393  for (const auto &val : NewPN->incoming_values())
5394  if (val == Fixup.OperandValToReplace)
5395  Fixup.UserInst = NewPN;
5396  }
5397  }
5398  }
5399  }
5400 }
5401 
5402 /// Emit instructions for the leading candidate expression for this LSRUse (this
5403 /// is called "expanding"), and update the UserInst to reference the newly
5404 /// expanded value.
5405 void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
5406  const Formula &F, SCEVExpander &Rewriter,
5407  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5408  // First, find an insertion point that dominates UserInst. For PHI nodes,
5409  // find the nearest block which dominates all the relevant uses.
5410  if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
5411  RewriteForPHI(PN, LU, LF, F, Rewriter, DeadInsts);
5412  } else {
5413  Value *FullV =
5414  Expand(LU, LF, F, LF.UserInst->getIterator(), Rewriter, DeadInsts);
5415 
5416  // If this is reuse-by-noop-cast, insert the noop cast.
5417  Type *OpTy = LF.OperandValToReplace->getType();
5418  if (FullV->getType() != OpTy) {
5419  Instruction *Cast =
5420  CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
5421  FullV, OpTy, "tmp", LF.UserInst);
5422  FullV = Cast;
5423  }
5424 
5425  // Update the user. ICmpZero is handled specially here (for now) because
5426  // Expand may have updated one of the operands of the icmp already, and
5427  // its new value may happen to be equal to LF.OperandValToReplace, in
5428  // which case doing replaceUsesOfWith leads to replacing both operands
5429  // with the same value. TODO: Reorganize this.
5430  if (LU.Kind == LSRUse::ICmpZero)
5431  LF.UserInst->setOperand(0, FullV);
5432  else
5433  LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
5434  }
5435 
5436  DeadInsts.emplace_back(LF.OperandValToReplace);
5437 }
5438 
5439 /// Rewrite all the fixup locations with new values, following the chosen
5440 /// solution.
5441 void LSRInstance::ImplementSolution(
5442  const SmallVectorImpl<const Formula *> &Solution) {
5443  // Keep track of instructions we may have made dead, so that
5444  // we can remove them after we are done working.
5446 
5448  "lsr");
5449 #ifndef NDEBUG
5450  Rewriter.setDebugType(DEBUG_TYPE);
5451 #endif
5452  Rewriter.disableCanonicalMode();
5453  Rewriter.enableLSRMode();
5454  Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
5455 
5456  // Mark phi nodes that terminate chains so the expander tries to reuse them.
5457  for (const IVChain &Chain : IVChainVec) {
5458  if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
5459  Rewriter.setChainedPhi(PN);
5460  }
5461 
5462  // Expand the new value definitions and update the users.
5463  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
5464  for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
5465  Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], Rewriter, DeadInsts);
5466  Changed = true;
5467  }
5468 
5469  for (const IVChain &Chain : IVChainVec) {
5470  GenerateIVChain(Chain, Rewriter, DeadInsts);
5471  Changed = true;
5472  }
5473  // Clean up after ourselves. This must be done before deleting any
5474  // instructions.
5475  Rewriter.clear();
5476 
5477  Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
5478 }
5479 
5480 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
5481  DominatorTree &DT, LoopInfo &LI,
5482  const TargetTransformInfo &TTI)
5483  : IU(IU), SE(SE), DT(DT), LI(LI), TTI(TTI), L(L),
5484  FavorBackedgeIndex(EnableBackedgeIndexing &&
5485  TTI.shouldFavorBackedgeIndex(L)) {
5486  // If LoopSimplify form is not available, stay out of trouble.
5487  if (!L->isLoopSimplifyForm())
5488  return;
5489 
5490  // If there's no interesting work to be done, bail early.
5491  if (IU.empty()) return;
5492 
5493  // If there's too much analysis to be done, bail early. We won't be able to
5494  // model the problem anyway.
5495  unsigned NumUsers = 0;
5496  for (const IVStrideUse &U : IU) {
5497  if (++NumUsers > MaxIVUsers) {
5498  (void)U;
5499  LLVM_DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U
5500  << "\n");
5501  return;
5502  }
5503  // Bail out if we have a PHI on an EHPad that gets a value from a
5504  // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is
5505  // no good place to stick any instructions.
5506  if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
5507  auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
5508  if (isa<FuncletPadInst>(FirstNonPHI) ||
5509  isa<CatchSwitchInst>(FirstNonPHI))
5510  for (BasicBlock *PredBB : PN->blocks())
5511  if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
5512  return;
5513  }
5514  }
5515 
5516 #ifndef NDEBUG
5517  // All dominating loops must have preheaders, or SCEVExpander may not be able
5518  // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
5519  //
5520  // IVUsers analysis should only create users that are dominated by simple loop
5521  // headers. Since this loop should dominate all of its users, its user list
5522  // should be empty if this loop itself is not within a simple loop nest.
5523  for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
5524  Rung; Rung = Rung->getIDom()) {
5525  BasicBlock *BB = Rung->getBlock();
5526  const Loop *DomLoop = LI.getLoopFor(BB);
5527  if (DomLoop && DomLoop->getHeader() == BB) {
5528  assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
5529  }
5530  }
5531 #endif // DEBUG
5532 
5533  LLVM_DEBUG(dbgs() << "\nLSR on loop ";
5534  L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
5535  dbgs() << ":\n");
5536 
5537  // First, perform some low-level loop optimizations.
5538  OptimizeShadowIV();
5539  OptimizeLoopTermCond();
5540 
5541  // If loop preparation eliminates all interesting IV users, bail.
5542  if (IU.empty()) return;
5543 
5544  // Skip nested loops until we can model them better with formulae.
5545  if (!L->empty()) {
5546  LLVM_DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
5547  return;
5548  }
5549 
5550  // Start collecting data and preparing for the solver.
5551  CollectChains();
5552  CollectInterestingTypesAndFactors();
5553  CollectFixupsAndInitialFormulae();
5554  CollectLoopInvariantFixupsAndFormulae();
5555 
5556  if (Uses.empty())
5557  return;
5558 
5559  LLVM_DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
5560  print_uses(dbgs()));
5561 
5562  // Now use the reuse data to generate a bunch of interesting ways
5563  // to formulate the values needed for the uses.
5564  GenerateAllReuseFormulae();
5565 
5566  FilterOutUndesirableDedicatedRegisters();
5567  NarrowSearchSpaceUsingHeuristics();
5568 
5570  Solve(Solution);
5571 
5572  // Release memory that is no longer needed.
5573  Factors.clear();
5574  Types.clear();
5575  RegUses.clear();
5576 
5577  if (Solution.empty())
5578  return;
5579 
5580 #ifndef NDEBUG
5581  // Formulae should be legal.
5582  for (const LSRUse &LU : Uses) {
5583  for (const Formula &F : LU.Formulae)
5584  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
5585  F) && "Illegal formula generated!");
5586  };
5587 #endif
5588 
5589  // Now that we've decided what we want, make it so.
5590  ImplementSolution(Solution);
5591 }
5592 
5593 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
5594 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
5595  if (Factors.empty() && Types.empty()) return;
5596 
5597  OS << "LSR has identified the following interesting factors and types: ";
5598  bool First = true;
5599 
5600  for (int64_t Factor : Factors) {
5601  if (!First) OS << ", ";
5602  First = false;
5603  OS << '*' << Factor;
5604  }
5605 
5606  for (Type *Ty : Types) {
5607  if (!First) OS << ", ";
5608  First = false;
5609  OS << '(' << *Ty << ')';
5610  }
5611  OS << '\n';
5612 }
5613 
5614 void LSRInstance::print_fixups(raw_ostream &OS) const {
5615  OS << "LSR is examining the following fixup sites:\n";
5616  for (const LSRUse &LU : Uses)
5617  for (const LSRFixup &LF : LU.Fixups) {
5618  dbgs() << " ";
5619  LF.print(OS);
5620  OS << '\n';
5621  }
5622 }
5623 
5624 void LSRInstance::print_uses(raw_ostream &OS) const {
5625  OS << "LSR is examining the following uses:\n";
5626  for (const LSRUse &LU : Uses) {
5627  dbgs() << " ";
5628  LU.print(OS);
5629  OS << '\n';
5630  for (const Formula &F : LU.Formulae) {
5631  OS << " ";
5632  F.print(OS);
5633  OS << '\n';
5634  }
5635  }
5636 }
5637 
5638 void LSRInstance::print(raw_ostream &OS) const {
5639  print_factors_and_types(OS);
5640  print_fixups(OS);
5641  print_uses(OS);
5642 }
5643 
5644 LLVM_DUMP_METHOD void LSRInstance::dump() const {
5645  print(errs()); errs() << '\n';
5646 }
5647 #endif
5648 
5649 namespace {
5650 
5651 class LoopStrengthReduce : public LoopPass {
5652 public:
5653  static char ID; // Pass ID, replacement for typeid
5654 
5655  LoopStrengthReduce();
5656 
5657 private:
5658  bool runOnLoop(Loop *L, LPPassManager &LPM) override;
5659  void getAnalysisUsage(AnalysisUsage &AU) const override;
5660 };
5661 
5662 } // end anonymous namespace
5663 
5664 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
5666 }
5667 
5668 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
5669  // We split critical edges, so we change the CFG. However, we do update
5670  // many analyses if they are around.
5672 
5680  // Requiring LoopSimplify a second time here prevents IVUsers from running
5681  // twice, since LoopSimplify was invalidated by running ScalarEvolution.
5686 }
5687 
5689  DominatorTree &DT, LoopInfo &LI,
5690  const TargetTransformInfo &TTI) {
5691  bool Changed = false;
5692 
5693  // Run the main LSR transformation.
5694  Changed |= LSRInstance(L, IU, SE, DT, LI, TTI).getChanged();
5695 
5696  // Remove any extra phis created by processing inner loops.
5697  Changed |= DeleteDeadPHIs(L->getHeader());
5698  if (EnablePhiElim && L->isLoopSimplifyForm()) {
5700  const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
5701  SCEVExpander Rewriter(SE, DL, "lsr");
5702 #ifndef NDEBUG
5703  Rewriter.setDebugType(DEBUG_TYPE);
5704 #endif
5705  unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
5706  if (numFolded) {
5707  Changed = true;
5709  DeleteDeadPHIs(L->getHeader());
5710  }
5711  }
5712  return Changed;
5713 }
5714 
5715 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
5716  if (skipLoop(L))
5717  return false;
5718 
5719  auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
5720  auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
5721  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
5722  auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
5723  const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
5724  *L->getHeader()->getParent());
5725  return ReduceLoopStrength(L, IU, SE, DT, LI, TTI);
5726 }
5727 
5730  LPMUpdater &) {
5731  if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
5732  AR.DT, AR.LI, AR.TTI))
5733  return PreservedAnalyses::all();
5734 
5736 }
5737 
5738 char LoopStrengthReduce::ID = 0;
5739 
5740 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
5741  "Loop Strength Reduction", false, false)
5747 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5748 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
5749  "Loop Strength Reduction", false, false)
5750 
5751 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:641