LLVM  15.0.0git
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"
69 #include "llvm/Analysis/IVUsers.h"
71 #include "llvm/Analysis/LoopInfo.h"
72 #include "llvm/Analysis/LoopPass.h"
82 #include "llvm/Config/llvm-config.h"
83 #include "llvm/IR/BasicBlock.h"
84 #include "llvm/IR/Constant.h"
85 #include "llvm/IR/Constants.h"
87 #include "llvm/IR/DerivedTypes.h"
88 #include "llvm/IR/Dominators.h"
89 #include "llvm/IR/GlobalValue.h"
90 #include "llvm/IR/IRBuilder.h"
91 #include "llvm/IR/InstrTypes.h"
92 #include "llvm/IR/Instruction.h"
93 #include "llvm/IR/Instructions.h"
94 #include "llvm/IR/IntrinsicInst.h"
95 #include "llvm/IR/Module.h"
96 #include "llvm/IR/Operator.h"
97 #include "llvm/IR/PassManager.h"
98 #include "llvm/IR/Type.h"
99 #include "llvm/IR/Use.h"
100 #include "llvm/IR/User.h"
101 #include "llvm/IR/Value.h"
102 #include "llvm/IR/ValueHandle.h"
103 #include "llvm/InitializePasses.h"
104 #include "llvm/Pass.h"
105 #include "llvm/Support/Casting.h"
107 #include "llvm/Support/Compiler.h"
108 #include "llvm/Support/Debug.h"
110 #include "llvm/Support/MathExtras.h"
112 #include "llvm/Transforms/Scalar.h"
113 #include "llvm/Transforms/Utils.h"
118 #include <algorithm>
119 #include <cassert>
120 #include <cstddef>
121 #include <cstdint>
122 #include <iterator>
123 #include <limits>
124 #include <map>
125 #include <numeric>
126 #include <utility>
127 
128 using namespace llvm;
129 
130 #define DEBUG_TYPE "loop-reduce"
131 
132 /// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
133 /// bail out. This threshold is far beyond the number of users that LSR can
134 /// conceivably solve, so it should not affect generated code, but catches the
135 /// worst cases before LSR burns too much compile time and stack space.
136 static const unsigned MaxIVUsers = 200;
137 
138 /// Limit the size of expression that SCEV-based salvaging will attempt to
139 /// translate into a DIExpression.
140 /// Choose a maximum size such that debuginfo is not excessively increased and
141 /// the salvaging is not too expensive for the compiler.
142 static const unsigned MaxSCEVSalvageExpressionSize = 64;
143 
144 // Cleanup congruent phis after LSR phi expansion.
146  "enable-lsr-phielim", cl::Hidden, cl::init(true),
147  cl::desc("Enable LSR phi elimination"));
148 
149 // The flag adds instruction count to solutions cost comparision.
150 static cl::opt<bool> InsnsCost(
151  "lsr-insns-cost", cl::Hidden, cl::init(true),
152  cl::desc("Add instruction count to a LSR cost model"));
153 
154 // Flag to choose how to narrow complex lsr solution
156  "lsr-exp-narrow", cl::Hidden, cl::init(false),
157  cl::desc("Narrow LSR complex solution using"
158  " expectation of registers number"));
159 
160 // Flag to narrow search space by filtering non-optimal formulae with
161 // the same ScaledReg and Scale.
163  "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
164  cl::desc("Narrow LSR search space by filtering non-optimal formulae"
165  " with the same ScaledReg and Scale"));
166 
168  "lsr-preferred-addressing-mode", cl::Hidden, cl::init(TTI::AMK_None),
169  cl::desc("A flag that overrides the target's preferred addressing mode."),
171  "none",
172  "Don't prefer any addressing mode"),
174  "preindexed",
175  "Prefer pre-indexed addressing mode"),
177  "postindexed",
178  "Prefer post-indexed addressing mode")));
179 
181  "lsr-complexity-limit", cl::Hidden,
183  cl::desc("LSR search space complexity limit"));
184 
186  "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
187  cl::desc("The limit on recursion depth for LSRs setup cost"));
188 
189 #ifndef NDEBUG
190 // Stress test IV chain generation.
192  "stress-ivchain", cl::Hidden, cl::init(false),
193  cl::desc("Stress test LSR IV chains"));
194 #else
195 static bool StressIVChain = false;
196 #endif
197 
198 namespace {
199 
200 struct MemAccessTy {
201  /// Used in situations where the accessed memory type is unknown.
202  static const unsigned UnknownAddressSpace =
204 
205  Type *MemTy = nullptr;
206  unsigned AddrSpace = UnknownAddressSpace;
207 
208  MemAccessTy() = default;
209  MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
210 
211  bool operator==(MemAccessTy Other) const {
212  return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
213  }
214 
215  bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
216 
217  static MemAccessTy getUnknown(LLVMContext &Ctx,
218  unsigned AS = UnknownAddressSpace) {
219  return MemAccessTy(Type::getVoidTy(Ctx), AS);
220  }
221 
222  Type *getType() { return MemTy; }
223 };
224 
225 /// This class holds data which is used to order reuse candidates.
226 class RegSortData {
227 public:
228  /// This represents the set of LSRUse indices which reference
229  /// a particular register.
230  SmallBitVector UsedByIndices;
231 
232  void print(raw_ostream &OS) const;
233  void dump() const;
234 };
235 
236 } // end anonymous namespace
237 
238 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
239 void RegSortData::print(raw_ostream &OS) const {
240  OS << "[NumUses=" << UsedByIndices.count() << ']';
241 }
242 
243 LLVM_DUMP_METHOD void RegSortData::dump() const {
244  print(errs()); errs() << '\n';
245 }
246 #endif
247 
248 namespace {
249 
250 /// Map register candidates to information about how they are used.
251 class RegUseTracker {
252  using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
253 
254  RegUsesTy RegUsesMap;
256 
257 public:
258  void countRegister(const SCEV *Reg, size_t LUIdx);
259  void dropRegister(const SCEV *Reg, size_t LUIdx);
260  void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
261 
262  bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
263 
264  const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
265 
266  void clear();
267 
269  using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
270 
271  iterator begin() { return RegSequence.begin(); }
272  iterator end() { return RegSequence.end(); }
273  const_iterator begin() const { return RegSequence.begin(); }
274  const_iterator end() const { return RegSequence.end(); }
275 };
276 
277 } // end anonymous namespace
278 
279 void
280 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
281  std::pair<RegUsesTy::iterator, bool> Pair =
282  RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
283  RegSortData &RSD = Pair.first->second;
284  if (Pair.second)
285  RegSequence.push_back(Reg);
286  RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
287  RSD.UsedByIndices.set(LUIdx);
288 }
289 
290 void
291 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
292  RegUsesTy::iterator It = RegUsesMap.find(Reg);
293  assert(It != RegUsesMap.end());
294  RegSortData &RSD = It->second;
295  assert(RSD.UsedByIndices.size() > LUIdx);
296  RSD.UsedByIndices.reset(LUIdx);
297 }
298 
299 void
300 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
301  assert(LUIdx <= LastLUIdx);
302 
303  // Update RegUses. The data structure is not optimized for this purpose;
304  // we must iterate through it and update each of the bit vectors.
305  for (auto &Pair : RegUsesMap) {
306  SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
307  if (LUIdx < UsedByIndices.size())
308  UsedByIndices[LUIdx] =
309  LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
310  UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
311  }
312 }
313 
314 bool
315 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
316  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
317  if (I == RegUsesMap.end())
318  return false;
319  const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
320  int i = UsedByIndices.find_first();
321  if (i == -1) return false;
322  if ((size_t)i != LUIdx) return true;
323  return UsedByIndices.find_next(i) != -1;
324 }
325 
326 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
327  RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
328  assert(I != RegUsesMap.end() && "Unknown register!");
329  return I->second.UsedByIndices;
330 }
331 
332 void RegUseTracker::clear() {
333  RegUsesMap.clear();
334  RegSequence.clear();
335 }
336 
337 namespace {
338 
339 /// This class holds information that describes a formula for computing
340 /// satisfying a use. It may include broken-out immediates and scaled registers.
341 struct Formula {
342  /// Global base address used for complex addressing.
343  GlobalValue *BaseGV = nullptr;
344 
345  /// Base offset for complex addressing.
346  int64_t BaseOffset = 0;
347 
348  /// Whether any complex addressing has a base register.
349  bool HasBaseReg = false;
350 
351  /// The scale of any complex addressing.
352  int64_t Scale = 0;
353 
354  /// The list of "base" registers for this use. When this is non-empty. The
355  /// canonical representation of a formula is
356  /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
357  /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
358  /// 3. The reg containing recurrent expr related with currect loop in the
359  /// formula should be put in the ScaledReg.
360  /// #1 enforces that the scaled register is always used when at least two
361  /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
362  /// #2 enforces that 1 * reg is reg.
363  /// #3 ensures invariant regs with respect to current loop can be combined
364  /// together in LSR codegen.
365  /// This invariant can be temporarily broken while building a formula.
366  /// However, every formula inserted into the LSRInstance must be in canonical
367  /// form.
369 
370  /// The 'scaled' register for this use. This should be non-null when Scale is
371  /// not zero.
372  const SCEV *ScaledReg = nullptr;
373 
374  /// An additional constant offset which added near the use. This requires a
375  /// temporary register, but the offset itself can live in an add immediate
376  /// field rather than a register.
377  int64_t UnfoldedOffset = 0;
378 
379  Formula() = default;
380 
381  void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
382 
383  bool isCanonical(const Loop &L) const;
384 
385  void canonicalize(const Loop &L);
386 
387  bool unscale();
388 
389  bool hasZeroEnd() const;
390 
391  size_t getNumRegs() const;
392  Type *getType() const;
393 
394  void deleteBaseReg(const SCEV *&S);
395 
396  bool referencesReg(const SCEV *S) const;
397  bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
398  const RegUseTracker &RegUses) const;
399 
400  void print(raw_ostream &OS) const;
401  void dump() const;
402 };
403 
404 } // end anonymous namespace
405 
406 /// Recursion helper for initialMatch.
407 static void DoInitialMatch(const SCEV *S, Loop *L,
410  ScalarEvolution &SE) {
411  // Collect expressions which properly dominate the loop header.
412  if (SE.properlyDominates(S, L->getHeader())) {
413  Good.push_back(S);
414  return;
415  }
416 
417  // Look at add operands.
418  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
419  for (const SCEV *S : Add->operands())
420  DoInitialMatch(S, L, Good, Bad, SE);
421  return;
422  }
423 
424  // Look at addrec operands.
425  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
426  if (!AR->getStart()->isZero() && AR->isAffine()) {
427  DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
428  DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
429  AR->getStepRecurrence(SE),
430  // FIXME: AR->getNoWrapFlags()
431  AR->getLoop(), SCEV::FlagAnyWrap),
432  L, Good, Bad, SE);
433  return;
434  }
435 
436  // Handle a multiplication by -1 (negation) if it didn't fold.
437  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
438  if (Mul->getOperand(0)->isAllOnesValue()) {
440  const SCEV *NewMul = SE.getMulExpr(Ops);
441 
444  DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
445  const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
446  SE.getEffectiveSCEVType(NewMul->getType())));
447  for (const SCEV *S : MyGood)
448  Good.push_back(SE.getMulExpr(NegOne, S));
449  for (const SCEV *S : MyBad)
450  Bad.push_back(SE.getMulExpr(NegOne, S));
451  return;
452  }
453 
454  // Ok, we can't do anything interesting. Just stuff the whole thing into a
455  // register and hope for the best.
456  Bad.push_back(S);
457 }
458 
459 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
460 /// all loop-invariant and loop-computable values in a single base register.
461 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
464  DoInitialMatch(S, L, Good, Bad, SE);
465  if (!Good.empty()) {
466  const SCEV *Sum = SE.getAddExpr(Good);
467  if (!Sum->isZero())
468  BaseRegs.push_back(Sum);
469  HasBaseReg = true;
470  }
471  if (!Bad.empty()) {
472  const SCEV *Sum = SE.getAddExpr(Bad);
473  if (!Sum->isZero())
474  BaseRegs.push_back(Sum);
475  HasBaseReg = true;
476  }
477  canonicalize(*L);
478 }
479 
480 static bool containsAddRecDependentOnLoop(const SCEV *S, const Loop &L) {
481  return SCEVExprContains(S, [&L](const SCEV *S) {
482  return isa<SCEVAddRecExpr>(S) && (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
483  });
484 }
485 
486 /// Check whether or not this formula satisfies the canonical
487 /// representation.
488 /// \see Formula::BaseRegs.
489 bool Formula::isCanonical(const Loop &L) const {
490  if (!ScaledReg)
491  return BaseRegs.size() <= 1;
492 
493  if (Scale != 1)
494  return true;
495 
496  if (Scale == 1 && BaseRegs.empty())
497  return false;
498 
499  if (containsAddRecDependentOnLoop(ScaledReg, L))
500  return true;
501 
502  // If ScaledReg is not a recurrent expr, or it is but its loop is not current
503  // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
504  // loop, we want to swap the reg in BaseRegs with ScaledReg.
505  return none_of(BaseRegs, [&L](const SCEV *S) {
506  return containsAddRecDependentOnLoop(S, L);
507  });
508 }
509 
510 /// Helper method to morph a formula into its canonical representation.
511 /// \see Formula::BaseRegs.
512 /// Every formula having more than one base register, must use the ScaledReg
513 /// field. Otherwise, we would have to do special cases everywhere in LSR
514 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
515 /// On the other hand, 1*reg should be canonicalized into reg.
516 void Formula::canonicalize(const Loop &L) {
517  if (isCanonical(L))
518  return;
519 
520  if (BaseRegs.empty()) {
521  // No base reg? Use scale reg with scale = 1 as such.
522  assert(ScaledReg && "Expected 1*reg => reg");
523  assert(Scale == 1 && "Expected 1*reg => reg");
524  BaseRegs.push_back(ScaledReg);
525  Scale = 0;
526  ScaledReg = nullptr;
527  return;
528  }
529 
530  // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
531  if (!ScaledReg) {
532  ScaledReg = BaseRegs.pop_back_val();
533  Scale = 1;
534  }
535 
536  // If ScaledReg is an invariant with respect to L, find the reg from
537  // BaseRegs containing the recurrent expr related with Loop L. Swap the
538  // reg with ScaledReg.
539  if (!containsAddRecDependentOnLoop(ScaledReg, L)) {
540  auto I = find_if(BaseRegs, [&L](const SCEV *S) {
541  return containsAddRecDependentOnLoop(S, L);
542  });
543  if (I != BaseRegs.end())
544  std::swap(ScaledReg, *I);
545  }
546  assert(isCanonical(L) && "Failed to canonicalize?");
547 }
548 
549 /// Get rid of the scale in the formula.
550 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
551 /// \return true if it was possible to get rid of the scale, false otherwise.
552 /// \note After this operation the formula may not be in the canonical form.
553 bool Formula::unscale() {
554  if (Scale != 1)
555  return false;
556  Scale = 0;
557  BaseRegs.push_back(ScaledReg);
558  ScaledReg = nullptr;
559  return true;
560 }
561 
562 bool Formula::hasZeroEnd() const {
563  if (UnfoldedOffset || BaseOffset)
564  return false;
565  if (BaseRegs.size() != 1 || ScaledReg)
566  return false;
567  return true;
568 }
569 
570 /// Return the total number of register operands used by this formula. This does
571 /// not include register uses implied by non-constant addrec strides.
572 size_t Formula::getNumRegs() const {
573  return !!ScaledReg + BaseRegs.size();
574 }
575 
576 /// Return the type of this formula, if it has one, or null otherwise. This type
577 /// is meaningless except for the bit size.
578 Type *Formula::getType() const {
579  return !BaseRegs.empty() ? BaseRegs.front()->getType() :
580  ScaledReg ? ScaledReg->getType() :
581  BaseGV ? BaseGV->getType() :
582  nullptr;
583 }
584 
585 /// Delete the given base reg from the BaseRegs list.
586 void Formula::deleteBaseReg(const SCEV *&S) {
587  if (&S != &BaseRegs.back())
588  std::swap(S, BaseRegs.back());
589  BaseRegs.pop_back();
590 }
591 
592 /// Test if this formula references the given register.
593 bool Formula::referencesReg(const SCEV *S) const {
594  return S == ScaledReg || is_contained(BaseRegs, S);
595 }
596 
597 /// Test whether this formula uses registers which are used by uses other than
598 /// the use with the given index.
599 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
600  const RegUseTracker &RegUses) const {
601  if (ScaledReg)
602  if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
603  return true;
604  for (const SCEV *BaseReg : BaseRegs)
605  if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
606  return true;
607  return false;
608 }
609 
610 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
611 void Formula::print(raw_ostream &OS) const {
612  bool First = true;
613  if (BaseGV) {
614  if (!First) OS << " + "; else First = false;
615  BaseGV->printAsOperand(OS, /*PrintType=*/false);
616  }
617  if (BaseOffset != 0) {
618  if (!First) OS << " + "; else First = false;
619  OS << BaseOffset;
620  }
621  for (const SCEV *BaseReg : BaseRegs) {
622  if (!First) OS << " + "; else First = false;
623  OS << "reg(" << *BaseReg << ')';
624  }
625  if (HasBaseReg && BaseRegs.empty()) {
626  if (!First) OS << " + "; else First = false;
627  OS << "**error: HasBaseReg**";
628  } else if (!HasBaseReg && !BaseRegs.empty()) {
629  if (!First) OS << " + "; else First = false;
630  OS << "**error: !HasBaseReg**";
631  }
632  if (Scale != 0) {
633  if (!First) OS << " + "; else First = false;
634  OS << Scale << "*reg(";
635  if (ScaledReg)
636  OS << *ScaledReg;
637  else
638  OS << "<unknown>";
639  OS << ')';
640  }
641  if (UnfoldedOffset != 0) {
642  if (!First) OS << " + ";
643  OS << "imm(" << UnfoldedOffset << ')';
644  }
645 }
646 
647 LLVM_DUMP_METHOD void Formula::dump() const {
648  print(errs()); errs() << '\n';
649 }
650 #endif
651 
652 /// Return true if the given addrec can be sign-extended without changing its
653 /// value.
654 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
655  Type *WideTy =
657  return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
658 }
659 
660 /// Return true if the given add can be sign-extended without changing its
661 /// value.
662 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
663  Type *WideTy =
664  IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
665  return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
666 }
667 
668 /// Return true if the given mul can be sign-extended without changing its
669 /// value.
670 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
671  Type *WideTy =
673  SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
674  return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
675 }
676 
677 /// Return an expression for LHS /s RHS, if it can be determined and if the
678 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
679 /// is true, expressions like (X * Y) /s Y are simplified to X, ignoring that
680 /// the multiplication may overflow, which is useful when the result will be
681 /// used in a context where the most significant bits are ignored.
682 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
683  ScalarEvolution &SE,
684  bool IgnoreSignificantBits = false) {
685  // Handle the trivial case, which works for any SCEV type.
686  if (LHS == RHS)
687  return SE.getConstant(LHS->getType(), 1);
688 
689  // Handle a few RHS special cases.
690  const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
691  if (RC) {
692  const APInt &RA = RC->getAPInt();
693  // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
694  // some folding.
695  if (RA.isAllOnes()) {
696  if (LHS->getType()->isPointerTy())
697  return nullptr;
698  return SE.getMulExpr(LHS, RC);
699  }
700  // Handle x /s 1 as x.
701  if (RA == 1)
702  return LHS;
703  }
704 
705  // Check for a division of a constant by a constant.
706  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
707  if (!RC)
708  return nullptr;
709  const APInt &LA = C->getAPInt();
710  const APInt &RA = RC->getAPInt();
711  if (LA.srem(RA) != 0)
712  return nullptr;
713  return SE.getConstant(LA.sdiv(RA));
714  }
715 
716  // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
717  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
718  if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
719  const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
720  IgnoreSignificantBits);
721  if (!Step) return nullptr;
722  const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
723  IgnoreSignificantBits);
724  if (!Start) return nullptr;
725  // FlagNW is independent of the start value, step direction, and is
726  // preserved with smaller magnitude steps.
727  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
728  return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
729  }
730  return nullptr;
731  }
732 
733  // Distribute the sdiv over add operands, if the add doesn't overflow.
734  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
735  if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
737  for (const SCEV *S : Add->operands()) {
738  const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
739  if (!Op) return nullptr;
740  Ops.push_back(Op);
741  }
742  return SE.getAddExpr(Ops);
743  }
744  return nullptr;
745  }
746 
747  // Check for a multiply operand that we can pull RHS out of.
748  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
749  if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
750  // Handle special case C1*X*Y /s C2*X*Y.
751  if (const SCEVMulExpr *MulRHS = dyn_cast<SCEVMulExpr>(RHS)) {
752  if (IgnoreSignificantBits || isMulSExtable(MulRHS, SE)) {
753  const SCEVConstant *LC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
754  const SCEVConstant *RC =
755  dyn_cast<SCEVConstant>(MulRHS->getOperand(0));
756  if (LC && RC) {
758  SmallVector<const SCEV *, 4> ROps(drop_begin(MulRHS->operands()));
759  if (LOps == ROps)
760  return getExactSDiv(LC, RC, SE, IgnoreSignificantBits);
761  }
762  }
763  }
764 
766  bool Found = false;
767  for (const SCEV *S : Mul->operands()) {
768  if (!Found)
769  if (const SCEV *Q = getExactSDiv(S, RHS, SE,
770  IgnoreSignificantBits)) {
771  S = Q;
772  Found = true;
773  }
774  Ops.push_back(S);
775  }
776  return Found ? SE.getMulExpr(Ops) : nullptr;
777  }
778  return nullptr;
779  }
780 
781  // Otherwise we don't know.
782  return nullptr;
783 }
784 
785 /// If S involves the addition of a constant integer value, return that integer
786 /// value, and mutate S to point to a new SCEV with that value excluded.
787 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
788  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
789  if (C->getAPInt().getMinSignedBits() <= 64) {
790  S = SE.getConstant(C->getType(), 0);
791  return C->getValue()->getSExtValue();
792  }
793  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
794  SmallVector<const SCEV *, 8> NewOps(Add->operands());
795  int64_t Result = ExtractImmediate(NewOps.front(), SE);
796  if (Result != 0)
797  S = SE.getAddExpr(NewOps);
798  return Result;
799  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
800  SmallVector<const SCEV *, 8> NewOps(AR->operands());
801  int64_t Result = ExtractImmediate(NewOps.front(), SE);
802  if (Result != 0)
803  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
804  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
806  return Result;
807  }
808  return 0;
809 }
810 
811 /// If S involves the addition of a GlobalValue address, return that symbol, and
812 /// mutate S to point to a new SCEV with that value excluded.
814  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
815  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
816  S = SE.getConstant(GV->getType(), 0);
817  return GV;
818  }
819  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
820  SmallVector<const SCEV *, 8> NewOps(Add->operands());
821  GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
822  if (Result)
823  S = SE.getAddExpr(NewOps);
824  return Result;
825  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
826  SmallVector<const SCEV *, 8> NewOps(AR->operands());
827  GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
828  if (Result)
829  S = SE.getAddRecExpr(NewOps, AR->getLoop(),
830  // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
832  return Result;
833  }
834  return nullptr;
835 }
836 
837 /// Returns true if the specified instruction is using the specified value as an
838 /// address.
840  Instruction *Inst, Value *OperandVal) {
841  bool isAddress = isa<LoadInst>(Inst);
842  if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
843  if (SI->getPointerOperand() == OperandVal)
844  isAddress = true;
845  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
846  // Addressing modes can also be folded into prefetches and a variety
847  // of intrinsics.
848  switch (II->getIntrinsicID()) {
849  case Intrinsic::memset:
850  case Intrinsic::prefetch:
851  case Intrinsic::masked_load:
852  if (II->getArgOperand(0) == OperandVal)
853  isAddress = true;
854  break;
855  case Intrinsic::masked_store:
856  if (II->getArgOperand(1) == OperandVal)
857  isAddress = true;
858  break;
859  case Intrinsic::memmove:
860  case Intrinsic::memcpy:
861  if (II->getArgOperand(0) == OperandVal ||
862  II->getArgOperand(1) == OperandVal)
863  isAddress = true;
864  break;
865  default: {
866  MemIntrinsicInfo IntrInfo;
867  if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
868  if (IntrInfo.PtrVal == OperandVal)
869  isAddress = true;
870  }
871  }
872  }
873  } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
874  if (RMW->getPointerOperand() == OperandVal)
875  isAddress = true;
876  } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
877  if (CmpX->getPointerOperand() == OperandVal)
878  isAddress = true;
879  }
880  return isAddress;
881 }
882 
883 /// Return the type of the memory being accessed.
884 static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
885  Instruction *Inst, Value *OperandVal) {
886  MemAccessTy AccessTy(Inst->getType(), MemAccessTy::UnknownAddressSpace);
887  if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
888  AccessTy.MemTy = SI->getOperand(0)->getType();
889  AccessTy.AddrSpace = SI->getPointerAddressSpace();
890  } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
891  AccessTy.AddrSpace = LI->getPointerAddressSpace();
892  } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
893  AccessTy.AddrSpace = RMW->getPointerAddressSpace();
894  } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
895  AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
896  } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
897  switch (II->getIntrinsicID()) {
898  case Intrinsic::prefetch:
899  case Intrinsic::memset:
900  AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
901  AccessTy.MemTy = OperandVal->getType();
902  break;
903  case Intrinsic::memmove:
904  case Intrinsic::memcpy:
905  AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
906  AccessTy.MemTy = OperandVal->getType();
907  break;
908  case Intrinsic::masked_load:
909  AccessTy.AddrSpace =
910  II->getArgOperand(0)->getType()->getPointerAddressSpace();
911  break;
912  case Intrinsic::masked_store:
913  AccessTy.MemTy = II->getOperand(0)->getType();
914  AccessTy.AddrSpace =
915  II->getArgOperand(1)->getType()->getPointerAddressSpace();
916  break;
917  default: {
918  MemIntrinsicInfo IntrInfo;
919  if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
920  AccessTy.AddrSpace
921  = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
922  }
923 
924  break;
925  }
926  }
927  }
928 
929  // All pointers have the same requirements, so canonicalize them to an
930  // arbitrary pointer type to minimize variation.
931  if (PointerType *PTy = dyn_cast<PointerType>(AccessTy.MemTy))
932  AccessTy.MemTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
933  PTy->getAddressSpace());
934 
935  return AccessTy;
936 }
937 
938 /// Return true if this AddRec is already a phi in its loop.
939 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
940  for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
941  if (SE.isSCEVable(PN.getType()) &&
942  (SE.getEffectiveSCEVType(PN.getType()) ==
943  SE.getEffectiveSCEVType(AR->getType())) &&
944  SE.getSCEV(&PN) == AR)
945  return true;
946  }
947  return false;
948 }
949 
950 /// Check if expanding this expression is likely to incur significant cost. This
951 /// is tricky because SCEV doesn't track which expressions are actually computed
952 /// by the current IR.
953 ///
954 /// We currently allow expansion of IV increments that involve adds,
955 /// multiplication by constants, and AddRecs from existing phis.
956 ///
957 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
958 /// obvious multiple of the UDivExpr.
959 static bool isHighCostExpansion(const SCEV *S,
960  SmallPtrSetImpl<const SCEV*> &Processed,
961  ScalarEvolution &SE) {
962  // Zero/One operand expressions
963  switch (S->getSCEVType()) {
964  case scUnknown:
965  case scConstant:
966  return false;
967  case scTruncate:
968  return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
969  Processed, SE);
970  case scZeroExtend:
971  return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
972  Processed, SE);
973  case scSignExtend:
974  return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
975  Processed, SE);
976  default:
977  break;
978  }
979 
980  if (!Processed.insert(S).second)
981  return false;
982 
983  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
984  for (const SCEV *S : Add->operands()) {
985  if (isHighCostExpansion(S, Processed, SE))
986  return true;
987  }
988  return false;
989  }
990 
991  if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
992  if (Mul->getNumOperands() == 2) {
993  // Multiplication by a constant is ok
994  if (isa<SCEVConstant>(Mul->getOperand(0)))
995  return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
996 
997  // If we have the value of one operand, check if an existing
998  // multiplication already generates this expression.
999  if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
1000  Value *UVal = U->getValue();
1001  for (User *UR : UVal->users()) {
1002  // If U is a constant, it may be used by a ConstantExpr.
1003  Instruction *UI = dyn_cast<Instruction>(UR);
1004  if (UI && UI->getOpcode() == Instruction::Mul &&
1005  SE.isSCEVable(UI->getType())) {
1006  return SE.getSCEV(UI) == Mul;
1007  }
1008  }
1009  }
1010  }
1011  }
1012 
1013  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1014  if (isExistingPhi(AR, SE))
1015  return false;
1016  }
1017 
1018  // Fow now, consider any other type of expression (div/mul/min/max) high cost.
1019  return true;
1020 }
1021 
1022 namespace {
1023 
1024 class LSRUse;
1025 
1026 } // end anonymous namespace
1027 
1028 /// Check if the addressing mode defined by \p F is completely
1029 /// folded in \p LU at isel time.
1030 /// This includes address-mode folding and special icmp tricks.
1031 /// This function returns true if \p LU can accommodate what \p F
1032 /// defines and up to 1 base + 1 scaled + offset.
1033 /// In other words, if \p F has several base registers, this function may
1034 /// still return true. Therefore, users still need to account for
1035 /// additional base registers and/or unfolded offsets to derive an
1036 /// accurate cost model.
1037 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1038  const LSRUse &LU, const Formula &F);
1039 
1040 // Get the cost of the scaling factor used in F for LU.
1042  const LSRUse &LU, const Formula &F,
1043  const Loop &L);
1044 
1045 namespace {
1046 
1047 /// This class is used to measure and compare candidate formulae.
1048 class Cost {
1049  const Loop *L = nullptr;
1050  ScalarEvolution *SE = nullptr;
1051  const TargetTransformInfo *TTI = nullptr;
1054 
1055 public:
1056  Cost() = delete;
1057  Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI,
1059  L(L), SE(&SE), TTI(&TTI), AMK(AMK) {
1060  C.Insns = 0;
1061  C.NumRegs = 0;
1062  C.AddRecCost = 0;
1063  C.NumIVMuls = 0;
1064  C.NumBaseAdds = 0;
1065  C.ImmCost = 0;
1066  C.SetupCost = 0;
1067  C.ScaleCost = 0;
1068  }
1069 
1070  bool isLess(const Cost &Other);
1071 
1072  void Lose();
1073 
1074 #ifndef NDEBUG
1075  // Once any of the metrics loses, they must all remain losers.
1076  bool isValid() {
1077  return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1078  | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1079  || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1080  & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1081  }
1082 #endif
1083 
1084  bool isLoser() {
1085  assert(isValid() && "invalid cost");
1086  return C.NumRegs == ~0u;
1087  }
1088 
1089  void RateFormula(const Formula &F,
1091  const DenseSet<const SCEV *> &VisitedRegs,
1092  const LSRUse &LU,
1093  SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1094 
1095  void print(raw_ostream &OS) const;
1096  void dump() const;
1097 
1098 private:
1099  void RateRegister(const Formula &F, const SCEV *Reg,
1101  void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1103  SmallPtrSetImpl<const SCEV *> *LoserRegs);
1104 };
1105 
1106 /// An operand value in an instruction which is to be replaced with some
1107 /// equivalent, possibly strength-reduced, replacement.
1108 struct LSRFixup {
1109  /// The instruction which will be updated.
1110  Instruction *UserInst = nullptr;
1111 
1112  /// The operand of the instruction which will be replaced. The operand may be
1113  /// used more than once; every instance will be replaced.
1114  Value *OperandValToReplace = nullptr;
1115 
1116  /// If this user is to use the post-incremented value of an induction
1117  /// variable, this set is non-empty and holds the loops associated with the
1118  /// induction variable.
1119  PostIncLoopSet PostIncLoops;
1120 
1121  /// A constant offset to be added to the LSRUse expression. This allows
1122  /// multiple fixups to share the same LSRUse with different offsets, for
1123  /// example in an unrolled loop.
1124  int64_t Offset = 0;
1125 
1126  LSRFixup() = default;
1127 
1128  bool isUseFullyOutsideLoop(const Loop *L) const;
1129 
1130  void print(raw_ostream &OS) const;
1131  void dump() const;
1132 };
1133 
1134 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1135 /// SmallVectors of const SCEV*.
1136 struct UniquifierDenseMapInfo {
1137  static SmallVector<const SCEV *, 4> getEmptyKey() {
1139  V.push_back(reinterpret_cast<const SCEV *>(-1));
1140  return V;
1141  }
1142 
1143  static SmallVector<const SCEV *, 4> getTombstoneKey() {
1145  V.push_back(reinterpret_cast<const SCEV *>(-2));
1146  return V;
1147  }
1148 
1149  static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1150  return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1151  }
1152 
1153  static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1155  return LHS == RHS;
1156  }
1157 };
1158 
1159 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1160 /// as uses invented by LSR itself. It includes information about what kinds of
1161 /// things can be folded into the user, information about the user itself, and
1162 /// information about how the use may be satisfied. TODO: Represent multiple
1163 /// users of the same expression in common?
1164 class LSRUse {
1165  DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1166 
1167 public:
1168  /// An enum for a kind of use, indicating what types of scaled and immediate
1169  /// operands it might support.
1170  enum KindType {
1171  Basic, ///< A normal use, with no folding.
1172  Special, ///< A special case of basic, allowing -1 scales.
1173  Address, ///< An address use; folding according to TargetLowering
1174  ICmpZero ///< An equality icmp with both operands folded into one.
1175  // TODO: Add a generic icmp too?
1176  };
1177 
1178  using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1179 
1180  KindType Kind;
1181  MemAccessTy AccessTy;
1182 
1183  /// The list of operands which are to be replaced.
1185 
1186  /// Keep track of the min and max offsets of the fixups.
1187  int64_t MinOffset = std::numeric_limits<int64_t>::max();
1188  int64_t MaxOffset = std::numeric_limits<int64_t>::min();
1189 
1190  /// This records whether all of the fixups using this LSRUse are outside of
1191  /// the loop, in which case some special-case heuristics may be used.
1192  bool AllFixupsOutsideLoop = true;
1193 
1194  /// RigidFormula is set to true to guarantee that this use will be associated
1195  /// with a single formula--the one that initially matched. Some SCEV
1196  /// expressions cannot be expanded. This allows LSR to consider the registers
1197  /// used by those expressions without the need to expand them later after
1198  /// changing the formula.
1199  bool RigidFormula = false;
1200 
1201  /// This records the widest use type for any fixup using this
1202  /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1203  /// fixup widths to be equivalent, because the narrower one may be relying on
1204  /// the implicit truncation to truncate away bogus bits.
1205  Type *WidestFixupType = nullptr;
1206 
1207  /// A list of ways to build a value that can satisfy this user. After the
1208  /// list is populated, one of these is selected heuristically and used to
1209  /// formulate a replacement for OperandValToReplace in UserInst.
1210  SmallVector<Formula, 12> Formulae;
1211 
1212  /// The set of register candidates used by all formulae in this LSRUse.
1214 
1215  LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1216 
1217  LSRFixup &getNewFixup() {
1218  Fixups.push_back(LSRFixup());
1219  return Fixups.back();
1220  }
1221 
1222  void pushFixup(LSRFixup &f) {
1223  Fixups.push_back(f);
1224  if (f.Offset > MaxOffset)
1225  MaxOffset = f.Offset;
1226  if (f.Offset < MinOffset)
1227  MinOffset = f.Offset;
1228  }
1229 
1230  bool HasFormulaWithSameRegs(const Formula &F) const;
1231  float getNotSelectedProbability(const SCEV *Reg) const;
1232  bool InsertFormula(const Formula &F, const Loop &L);
1233  void DeleteFormula(Formula &F);
1234  void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1235 
1236  void print(raw_ostream &OS) const;
1237  void dump() const;
1238 };
1239 
1240 } // end anonymous namespace
1241 
1242 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1243  LSRUse::KindType Kind, MemAccessTy AccessTy,
1244  GlobalValue *BaseGV, int64_t BaseOffset,
1245  bool HasBaseReg, int64_t Scale,
1246  Instruction *Fixup = nullptr);
1247 
1248 static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) {
1249  if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1250  return 1;
1251  if (Depth == 0)
1252  return 0;
1253  if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1254  return getSetupCost(S->getStart(), Depth - 1);
1255  if (auto S = dyn_cast<SCEVIntegralCastExpr>(Reg))
1256  return getSetupCost(S->getOperand(), Depth - 1);
1257  if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
1258  return std::accumulate(S->op_begin(), S->op_end(), 0,
1259  [&](unsigned i, const SCEV *Reg) {
1260  return i + getSetupCost(Reg, Depth - 1);
1261  });
1262  if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1263  return getSetupCost(S->getLHS(), Depth - 1) +
1264  getSetupCost(S->getRHS(), Depth - 1);
1265  return 0;
1266 }
1267 
1268 /// Tally up interesting quantities from the given register.
1269 void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1271  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1272  // If this is an addrec for another loop, it should be an invariant
1273  // with respect to L since L is the innermost loop (at least
1274  // for now LSR only handles innermost loops).
1275  if (AR->getLoop() != L) {
1276  // If the AddRec exists, consider it's register free and leave it alone.
1277  if (isExistingPhi(AR, *SE) && AMK != TTI::AMK_PostIndexed)
1278  return;
1279 
1280  // It is bad to allow LSR for current loop to add induction variables
1281  // for its sibling loops.
1282  if (!AR->getLoop()->contains(L)) {
1283  Lose();
1284  return;
1285  }
1286 
1287  // Otherwise, it will be an invariant with respect to Loop L.
1288  ++C.NumRegs;
1289  return;
1290  }
1291 
1292  unsigned LoopCost = 1;
1293  if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1294  TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1295 
1296  // If the step size matches the base offset, we could use pre-indexed
1297  // addressing.
1298  if (AMK == TTI::AMK_PreIndexed) {
1299  if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1300  if (Step->getAPInt() == F.BaseOffset)
1301  LoopCost = 0;
1302  } else if (AMK == TTI::AMK_PostIndexed) {
1303  const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1304  if (isa<SCEVConstant>(LoopStep)) {
1305  const SCEV *LoopStart = AR->getStart();
1306  if (!isa<SCEVConstant>(LoopStart) &&
1307  SE->isLoopInvariant(LoopStart, L))
1308  LoopCost = 0;
1309  }
1310  }
1311  }
1312  C.AddRecCost += LoopCost;
1313 
1314  // Add the step value register, if it needs one.
1315  // TODO: The non-affine case isn't precisely modeled here.
1316  if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1317  if (!Regs.count(AR->getOperand(1))) {
1318  RateRegister(F, AR->getOperand(1), Regs);
1319  if (isLoser())
1320  return;
1321  }
1322  }
1323  }
1324  ++C.NumRegs;
1325 
1326  // Rough heuristic; favor registers which don't require extra setup
1327  // instructions in the preheader.
1328  C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
1329  // Ensure we don't, even with the recusion limit, produce invalid costs.
1330  C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);
1331 
1332  C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1334 }
1335 
1336 /// Record this register in the set. If we haven't seen it before, rate
1337 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1338 /// one of those regs an instant loser.
1339 void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1341  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1342  if (LoserRegs && LoserRegs->count(Reg)) {
1343  Lose();
1344  return;
1345  }
1346  if (Regs.insert(Reg).second) {
1347  RateRegister(F, Reg, Regs);
1348  if (LoserRegs && isLoser())
1349  LoserRegs->insert(Reg);
1350  }
1351 }
1352 
1353 void Cost::RateFormula(const Formula &F,
1355  const DenseSet<const SCEV *> &VisitedRegs,
1356  const LSRUse &LU,
1357  SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1358  if (isLoser())
1359  return;
1360  assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1361  // Tally up the registers.
1362  unsigned PrevAddRecCost = C.AddRecCost;
1363  unsigned PrevNumRegs = C.NumRegs;
1364  unsigned PrevNumBaseAdds = C.NumBaseAdds;
1365  if (const SCEV *ScaledReg = F.ScaledReg) {
1366  if (VisitedRegs.count(ScaledReg)) {
1367  Lose();
1368  return;
1369  }
1370  RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1371  if (isLoser())
1372  return;
1373  }
1374  for (const SCEV *BaseReg : F.BaseRegs) {
1375  if (VisitedRegs.count(BaseReg)) {
1376  Lose();
1377  return;
1378  }
1379  RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1380  if (isLoser())
1381  return;
1382  }
1383 
1384  // Determine how many (unfolded) adds we'll need inside the loop.
1385  size_t NumBaseParts = F.getNumRegs();
1386  if (NumBaseParts > 1)
1387  // Do not count the base and a possible second register if the target
1388  // allows to fold 2 registers.
1389  C.NumBaseAdds +=
1390  NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1391  C.NumBaseAdds += (F.UnfoldedOffset != 0);
1392 
1393  // Accumulate non-free scaling amounts.
1394  C.ScaleCost += *getScalingFactorCost(*TTI, LU, F, *L).getValue();
1395 
1396  // Tally up the non-zero immediates.
1397  for (const LSRFixup &Fixup : LU.Fixups) {
1398  int64_t O = Fixup.Offset;
1399  int64_t Offset = (uint64_t)O + F.BaseOffset;
1400  if (F.BaseGV)
1401  C.ImmCost += 64; // Handle symbolic values conservatively.
1402  // TODO: This should probably be the pointer size.
1403  else if (Offset != 0)
1404  C.ImmCost += APInt(64, Offset, true).getMinSignedBits();
1405 
1406  // Check with target if this offset with this instruction is
1407  // specifically not supported.
1408  if (LU.Kind == LSRUse::Address && Offset != 0 &&
1409  !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1410  Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1411  C.NumBaseAdds++;
1412  }
1413 
1414  // If we don't count instruction cost exit here.
1415  if (!InsnsCost) {
1416  assert(isValid() && "invalid cost");
1417  return;
1418  }
1419 
1420  // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1421  // additional instruction (at least fill).
1422  // TODO: Need distinguish register class?
1423  unsigned TTIRegNum = TTI->getNumberOfRegisters(
1424  TTI->getRegisterClassForType(false, F.getType())) - 1;
1425  if (C.NumRegs > TTIRegNum) {
1426  // Cost already exceeded TTIRegNum, then only newly added register can add
1427  // new instructions.
1428  if (PrevNumRegs > TTIRegNum)
1429  C.Insns += (C.NumRegs - PrevNumRegs);
1430  else
1431  C.Insns += (C.NumRegs - TTIRegNum);
1432  }
1433 
1434  // If ICmpZero formula ends with not 0, it could not be replaced by
1435  // just add or sub. We'll need to compare final result of AddRec.
1436  // That means we'll need an additional instruction. But if the target can
1437  // macro-fuse a compare with a branch, don't count this extra instruction.
1438  // For -10 + {0, +, 1}:
1439  // i = i + 1;
1440  // cmp i, 10
1441  //
1442  // For {-10, +, 1}:
1443  // i = i + 1;
1444  if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1445  !TTI->canMacroFuseCmp())
1446  C.Insns++;
1447  // Each new AddRec adds 1 instruction to calculation.
1448  C.Insns += (C.AddRecCost - PrevAddRecCost);
1449 
1450  // BaseAdds adds instructions for unfolded registers.
1451  if (LU.Kind != LSRUse::ICmpZero)
1452  C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1453  assert(isValid() && "invalid cost");
1454 }
1455 
1456 /// Set this cost to a losing value.
1457 void Cost::Lose() {
1460  C.AddRecCost = std::numeric_limits<unsigned>::max();
1461  C.NumIVMuls = std::numeric_limits<unsigned>::max();
1462  C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1464  C.SetupCost = std::numeric_limits<unsigned>::max();
1465  C.ScaleCost = std::numeric_limits<unsigned>::max();
1466 }
1467 
1468 /// Choose the lower cost.
1469 bool Cost::isLess(const Cost &Other) {
1470  if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1471  C.Insns != Other.C.Insns)
1472  return C.Insns < Other.C.Insns;
1473  return TTI->isLSRCostLess(C, Other.C);
1474 }
1475 
1476 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1477 void Cost::print(raw_ostream &OS) const {
1478  if (InsnsCost)
1479  OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1480  OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1481  if (C.AddRecCost != 0)
1482  OS << ", with addrec cost " << C.AddRecCost;
1483  if (C.NumIVMuls != 0)
1484  OS << ", plus " << C.NumIVMuls << " IV mul"
1485  << (C.NumIVMuls == 1 ? "" : "s");
1486  if (C.NumBaseAdds != 0)
1487  OS << ", plus " << C.NumBaseAdds << " base add"
1488  << (C.NumBaseAdds == 1 ? "" : "s");
1489  if (C.ScaleCost != 0)
1490  OS << ", plus " << C.ScaleCost << " scale cost";
1491  if (C.ImmCost != 0)
1492  OS << ", plus " << C.ImmCost << " imm cost";
1493  if (C.SetupCost != 0)
1494  OS << ", plus " << C.SetupCost << " setup cost";
1495 }
1496 
1497 LLVM_DUMP_METHOD void Cost::dump() const {
1498  print(errs()); errs() << '\n';
1499 }
1500 #endif
1501 
1502 /// Test whether this fixup always uses its value outside of the given loop.
1503 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1504  // PHI nodes use their value in their incoming blocks.
1505  if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1506  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1507  if (PN->getIncomingValue(i) == OperandValToReplace &&
1508  L->contains(PN->getIncomingBlock(i)))
1509  return false;
1510  return true;
1511  }
1512 
1513  return !L->contains(UserInst);
1514 }
1515 
1516 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1517 void LSRFixup::print(raw_ostream &OS) const {
1518  OS << "UserInst=";
1519  // Store is common and interesting enough to be worth special-casing.
1520  if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1521  OS << "store ";
1522  Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1523  } else if (UserInst->getType()->isVoidTy())
1524  OS << UserInst->getOpcodeName();
1525  else
1526  UserInst->printAsOperand(OS, /*PrintType=*/false);
1527 
1528  OS << ", OperandValToReplace=";
1529  OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1530 
1531  for (const Loop *PIL : PostIncLoops) {
1532  OS << ", PostIncLoop=";
1533  PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1534  }
1535 
1536  if (Offset != 0)
1537  OS << ", Offset=" << Offset;
1538 }
1539 
1540 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1541  print(errs()); errs() << '\n';
1542 }
1543 #endif
1544 
1545 /// Test whether this use as a formula which has the same registers as the given
1546 /// formula.
1547 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1548  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1549  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1550  // Unstable sort by host order ok, because this is only used for uniquifying.
1551  llvm::sort(Key);
1552  return Uniquifier.count(Key);
1553 }
1554 
1555 /// The function returns a probability of selecting formula without Reg.
1556 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1557  unsigned FNum = 0;
1558  for (const Formula &F : Formulae)
1559  if (F.referencesReg(Reg))
1560  FNum++;
1561  return ((float)(Formulae.size() - FNum)) / Formulae.size();
1562 }
1563 
1564 /// If the given formula has not yet been inserted, add it to the list, and
1565 /// return true. Return false otherwise. The formula must be in canonical form.
1566 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1567  assert(F.isCanonical(L) && "Invalid canonical representation");
1568 
1569  if (!Formulae.empty() && RigidFormula)
1570  return false;
1571 
1572  SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1573  if (F.ScaledReg) Key.push_back(F.ScaledReg);
1574  // Unstable sort by host order ok, because this is only used for uniquifying.
1575  llvm::sort(Key);
1576 
1577  if (!Uniquifier.insert(Key).second)
1578  return false;
1579 
1580  // Using a register to hold the value of 0 is not profitable.
1581  assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1582  "Zero allocated in a scaled register!");
1583 #ifndef NDEBUG
1584  for (const SCEV *BaseReg : F.BaseRegs)
1585  assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1586 #endif
1587 
1588  // Add the formula to the list.
1589  Formulae.push_back(F);
1590 
1591  // Record registers now being used by this use.
1592  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1593  if (F.ScaledReg)
1594  Regs.insert(F.ScaledReg);
1595 
1596  return true;
1597 }
1598 
1599 /// Remove the given formula from this use's list.
1600 void LSRUse::DeleteFormula(Formula &F) {
1601  if (&F != &Formulae.back())
1602  std::swap(F, Formulae.back());
1603  Formulae.pop_back();
1604 }
1605 
1606 /// Recompute the Regs field, and update RegUses.
1607 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1608  // Now that we've filtered out some formulae, recompute the Regs set.
1609  SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1610  Regs.clear();
1611  for (const Formula &F : Formulae) {
1612  if (F.ScaledReg) Regs.insert(F.ScaledReg);
1613  Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1614  }
1615 
1616  // Update the RegTracker.
1617  for (const SCEV *S : OldRegs)
1618  if (!Regs.count(S))
1619  RegUses.dropRegister(S, LUIdx);
1620 }
1621 
1622 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1623 void LSRUse::print(raw_ostream &OS) const {
1624  OS << "LSR Use: Kind=";
1625  switch (Kind) {
1626  case Basic: OS << "Basic"; break;
1627  case Special: OS << "Special"; break;
1628  case ICmpZero: OS << "ICmpZero"; break;
1629  case Address:
1630  OS << "Address of ";
1631  if (AccessTy.MemTy->isPointerTy())
1632  OS << "pointer"; // the full pointer type could be really verbose
1633  else {
1634  OS << *AccessTy.MemTy;
1635  }
1636 
1637  OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1638  }
1639 
1640  OS << ", Offsets={";
1641  bool NeedComma = false;
1642  for (const LSRFixup &Fixup : Fixups) {
1643  if (NeedComma) OS << ',';
1644  OS << Fixup.Offset;
1645  NeedComma = true;
1646  }
1647  OS << '}';
1648 
1649  if (AllFixupsOutsideLoop)
1650  OS << ", all-fixups-outside-loop";
1651 
1652  if (WidestFixupType)
1653  OS << ", widest fixup type: " << *WidestFixupType;
1654 }
1655 
1656 LLVM_DUMP_METHOD void LSRUse::dump() const {
1657  print(errs()); errs() << '\n';
1658 }
1659 #endif
1660 
1662  LSRUse::KindType Kind, MemAccessTy AccessTy,
1663  GlobalValue *BaseGV, int64_t BaseOffset,
1664  bool HasBaseReg, int64_t Scale,
1665  Instruction *Fixup/*= nullptr*/) {
1666  switch (Kind) {
1667  case LSRUse::Address:
1668  return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, BaseOffset,
1669  HasBaseReg, Scale, AccessTy.AddrSpace, Fixup);
1670 
1671  case LSRUse::ICmpZero:
1672  // There's not even a target hook for querying whether it would be legal to
1673  // fold a GV into an ICmp.
1674  if (BaseGV)
1675  return false;
1676 
1677  // ICmp only has two operands; don't allow more than two non-trivial parts.
1678  if (Scale != 0 && HasBaseReg && BaseOffset != 0)
1679  return false;
1680 
1681  // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1682  // putting the scaled register in the other operand of the icmp.
1683  if (Scale != 0 && Scale != -1)
1684  return false;
1685 
1686  // If we have low-level target information, ask the target if it can fold an
1687  // integer immediate on an icmp.
1688  if (BaseOffset != 0) {
1689  // We have one of:
1690  // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1691  // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1692  // Offs is the ICmp immediate.
1693  if (Scale == 0)
1694  // The cast does the right thing with
1695  // std::numeric_limits<int64_t>::min().
1696  BaseOffset = -(uint64_t)BaseOffset;
1697  return TTI.isLegalICmpImmediate(BaseOffset);
1698  }
1699 
1700  // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1701  return true;
1702 
1703  case LSRUse::Basic:
1704  // Only handle single-register values.
1705  return !BaseGV && Scale == 0 && BaseOffset == 0;
1706 
1707  case LSRUse::Special:
1708  // Special case Basic to handle -1 scales.
1709  return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
1710  }
1711 
1712  llvm_unreachable("Invalid LSRUse Kind!");
1713 }
1714 
1716  int64_t MinOffset, int64_t MaxOffset,
1717  LSRUse::KindType Kind, MemAccessTy AccessTy,
1718  GlobalValue *BaseGV, int64_t BaseOffset,
1719  bool HasBaseReg, int64_t Scale) {
1720  // Check for overflow.
1721  if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
1722  (MinOffset > 0))
1723  return false;
1724  MinOffset = (uint64_t)BaseOffset + MinOffset;
1725  if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
1726  (MaxOffset > 0))
1727  return false;
1728  MaxOffset = (uint64_t)BaseOffset + MaxOffset;
1729 
1730  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1731  HasBaseReg, Scale) &&
1732  isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1733  HasBaseReg, Scale);
1734 }
1735 
1737  int64_t MinOffset, int64_t MaxOffset,
1738  LSRUse::KindType Kind, MemAccessTy AccessTy,
1739  const Formula &F, const Loop &L) {
1740  // For the purpose of isAMCompletelyFolded either having a canonical formula
1741  // or a scale not equal to zero is correct.
1742  // Problems may arise from non canonical formulae having a scale == 0.
1743  // Strictly speaking it would best to just rely on canonical formulae.
1744  // However, when we generate the scaled formulae, we first check that the
1745  // scaling factor is profitable before computing the actual ScaledReg for
1746  // compile time sake.
1747  assert((F.isCanonical(L) || F.Scale != 0));
1748  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1749  F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1750 }
1751 
1752 /// Test whether we know how to expand the current formula.
1753 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1754  int64_t MaxOffset, LSRUse::KindType Kind,
1755  MemAccessTy AccessTy, GlobalValue *BaseGV,
1756  int64_t BaseOffset, bool HasBaseReg, int64_t Scale) {
1757  // We know how to expand completely foldable formulae.
1758  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1759  BaseOffset, HasBaseReg, Scale) ||
1760  // Or formulae that use a base register produced by a sum of base
1761  // registers.
1762  (Scale == 1 &&
1763  isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1764  BaseGV, BaseOffset, true, 0));
1765 }
1766 
1767 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
1768  int64_t MaxOffset, LSRUse::KindType Kind,
1769  MemAccessTy AccessTy, const Formula &F) {
1770  return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1771  F.BaseOffset, F.HasBaseReg, F.Scale);
1772 }
1773 
1775  const LSRUse &LU, const Formula &F) {
1776  // Target may want to look at the user instructions.
1777  if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1778  for (const LSRFixup &Fixup : LU.Fixups)
1779  if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1780  (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1781  F.Scale, Fixup.UserInst))
1782  return false;
1783  return true;
1784  }
1785 
1786  return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1787  LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1788  F.Scale);
1789 }
1790 
1792  const LSRUse &LU, const Formula &F,
1793  const Loop &L) {
1794  if (!F.Scale)
1795  return 0;
1796 
1797  // If the use is not completely folded in that instruction, we will have to
1798  // pay an extra cost only for scale != 1.
1799  if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1800  LU.AccessTy, F, L))
1801  return F.Scale != 1;
1802 
1803  switch (LU.Kind) {
1804  case LSRUse::Address: {
1805  // Check the scaling factor cost with both the min and max offsets.
1806  InstructionCost ScaleCostMinOffset = TTI.getScalingFactorCost(
1807  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MinOffset, F.HasBaseReg,
1808  F.Scale, LU.AccessTy.AddrSpace);
1809  InstructionCost ScaleCostMaxOffset = TTI.getScalingFactorCost(
1810  LU.AccessTy.MemTy, F.BaseGV, F.BaseOffset + LU.MaxOffset, F.HasBaseReg,
1811  F.Scale, LU.AccessTy.AddrSpace);
1812 
1813  assert(ScaleCostMinOffset.isValid() && ScaleCostMaxOffset.isValid() &&
1814  "Legal addressing mode has an illegal cost!");
1815  return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
1816  }
1817  case LSRUse::ICmpZero:
1818  case LSRUse::Basic:
1819  case LSRUse::Special:
1820  // The use is completely folded, i.e., everything is folded into the
1821  // instruction.
1822  return 0;
1823  }
1824 
1825  llvm_unreachable("Invalid LSRUse Kind!");
1826 }
1827 
1829  LSRUse::KindType Kind, MemAccessTy AccessTy,
1830  GlobalValue *BaseGV, int64_t BaseOffset,
1831  bool HasBaseReg) {
1832  // Fast-path: zero is always foldable.
1833  if (BaseOffset == 0 && !BaseGV) return true;
1834 
1835  // Conservatively, create an address with an immediate and a
1836  // base and a scale.
1837  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1838 
1839  // Canonicalize a scale of 1 to a base register if the formula doesn't
1840  // already have a base register.
1841  if (!HasBaseReg && Scale == 1) {
1842  Scale = 0;
1843  HasBaseReg = true;
1844  }
1845 
1846  return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
1847  HasBaseReg, Scale);
1848 }
1849 
1851  ScalarEvolution &SE, int64_t MinOffset,
1852  int64_t MaxOffset, LSRUse::KindType Kind,
1853  MemAccessTy AccessTy, const SCEV *S,
1854  bool HasBaseReg) {
1855  // Fast-path: zero is always foldable.
1856  if (S->isZero()) return true;
1857 
1858  // Conservatively, create an address with an immediate and a
1859  // base and a scale.
1860  int64_t BaseOffset = ExtractImmediate(S, SE);
1861  GlobalValue *BaseGV = ExtractSymbol(S, SE);
1862 
1863  // If there's anything else involved, it's not foldable.
1864  if (!S->isZero()) return false;
1865 
1866  // Fast-path: zero is always foldable.
1867  if (BaseOffset == 0 && !BaseGV) return true;
1868 
1869  // Conservatively, create an address with an immediate and a
1870  // base and a scale.
1871  int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1872 
1873  return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1874  BaseOffset, HasBaseReg, Scale);
1875 }
1876 
1877 namespace {
1878 
1879 /// An individual increment in a Chain of IV increments. Relate an IV user to
1880 /// an expression that computes the IV it uses from the IV used by the previous
1881 /// link in the Chain.
1882 ///
1883 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
1884 /// original IVOperand. The head of the chain's IVOperand is only valid during
1885 /// chain collection, before LSR replaces IV users. During chain generation,
1886 /// IncExpr can be used to find the new IVOperand that computes the same
1887 /// expression.
1888 struct IVInc {
1889  Instruction *UserInst;
1890  Value* IVOperand;
1891  const SCEV *IncExpr;
1892 
1893  IVInc(Instruction *U, Value *O, const SCEV *E)
1894  : UserInst(U), IVOperand(O), IncExpr(E) {}
1895 };
1896 
1897 // The list of IV increments in program order. We typically add the head of a
1898 // chain without finding subsequent links.
1899 struct IVChain {
1900  SmallVector<IVInc, 1> Incs;
1901  const SCEV *ExprBase = nullptr;
1902 
1903  IVChain() = default;
1904  IVChain(const IVInc &Head, const SCEV *Base)
1905  : Incs(1, Head), ExprBase(Base) {}
1906 
1907  using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
1908 
1909  // Return the first increment in the chain.
1910  const_iterator begin() const {
1911  assert(!Incs.empty());
1912  return std::next(Incs.begin());
1913  }
1914  const_iterator end() const {
1915  return Incs.end();
1916  }
1917 
1918  // Returns true if this chain contains any increments.
1919  bool hasIncs() const { return Incs.size() >= 2; }
1920 
1921  // Add an IVInc to the end of this chain.
1922  void add(const IVInc &X) { Incs.push_back(X); }
1923 
1924  // Returns the last UserInst in the chain.
1925  Instruction *tailUserInst() const { return Incs.back().UserInst; }
1926 
1927  // Returns true if IncExpr can be profitably added to this chain.
1928  bool isProfitableIncrement(const SCEV *OperExpr,
1929  const SCEV *IncExpr,
1930  ScalarEvolution&);
1931 };
1932 
1933 /// Helper for CollectChains to track multiple IV increment uses. Distinguish
1934 /// between FarUsers that definitely cross IV increments and NearUsers that may
1935 /// be used between IV increments.
1936 struct ChainUsers {
1938  SmallPtrSet<Instruction*, 4> NearUsers;
1939 };
1940 
1941 /// This class holds state for the main loop strength reduction logic.
1942 class LSRInstance {
1943  IVUsers &IU;
1944  ScalarEvolution &SE;
1945  DominatorTree &DT;
1946  LoopInfo &LI;
1947  AssumptionCache &AC;
1948  TargetLibraryInfo &TLI;
1949  const TargetTransformInfo &TTI;
1950  Loop *const L;
1951  MemorySSAUpdater *MSSAU;
1953  bool Changed = false;
1954 
1955  /// This is the insert position that the current loop's induction variable
1956  /// increment should be placed. In simple loops, this is the latch block's
1957  /// terminator. But in more complicated cases, this is a position which will
1958  /// dominate all the in-loop post-increment users.
1959  Instruction *IVIncInsertPos = nullptr;
1960 
1961  /// Interesting factors between use strides.
1962  ///
1963  /// We explicitly use a SetVector which contains a SmallSet, instead of the
1964  /// default, a SmallDenseSet, because we need to use the full range of
1965  /// int64_ts, and there's currently no good way of doing that with
1966  /// SmallDenseSet.
1968 
1969  /// Interesting use types, to facilitate truncation reuse.
1971 
1972  /// The list of interesting uses.
1973  mutable SmallVector<LSRUse, 16> Uses;
1974 
1975  /// Track which uses use which register candidates.
1976  RegUseTracker RegUses;
1977 
1978  // Limit the number of chains to avoid quadratic behavior. We don't expect to
1979  // have more than a few IV increment chains in a loop. Missing a Chain falls
1980  // back to normal LSR behavior for those uses.
1981  static const unsigned MaxChains = 8;
1982 
1983  /// IV users can form a chain of IV increments.
1985 
1986  /// IV users that belong to profitable IVChains.
1988 
1989  /// Induction variables that were generated and inserted by the SCEV Expander.
1990  SmallVector<llvm::WeakVH, 2> ScalarEvolutionIVs;
1991 
1992  void OptimizeShadowIV();
1993  bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1994  ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1995  void OptimizeLoopTermCond();
1996 
1997  void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
1998  SmallVectorImpl<ChainUsers> &ChainUsersVec);
1999  void FinalizeChain(IVChain &Chain);
2000  void CollectChains();
2001  void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
2002  SmallVectorImpl<WeakTrackingVH> &DeadInsts);
2003 
2004  void CollectInterestingTypesAndFactors();
2005  void CollectFixupsAndInitialFormulae();
2006 
2007  // Support for sharing of LSRUses between LSRFixups.
2009  UseMapTy UseMap;
2010 
2011  bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
2012  LSRUse::KindType Kind, MemAccessTy AccessTy);
2013 
2014  std::pair<size_t, int64_t> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
2015  MemAccessTy AccessTy);
2016 
2017  void DeleteUse(LSRUse &LU, size_t LUIdx);
2018 
2019  LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
2020 
2021  void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2022  void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2023  void CountRegisters(const Formula &F, size_t LUIdx);
2024  bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
2025 
2026  void CollectLoopInvariantFixupsAndFormulae();
2027 
2028  void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
2029  unsigned Depth = 0);
2030 
2031  void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
2032  const Formula &Base, unsigned Depth,
2033  size_t Idx, bool IsScaledReg = false);
2034  void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
2035  void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2036  const Formula &Base, size_t Idx,
2037  bool IsScaledReg = false);
2038  void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2039  void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2040  const Formula &Base,
2041  const SmallVectorImpl<int64_t> &Worklist,
2042  size_t Idx, bool IsScaledReg = false);
2043  void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2044  void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2045  void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2046  void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2047  void GenerateCrossUseConstantOffsets();
2048  void GenerateAllReuseFormulae();
2049 
2050  void FilterOutUndesirableDedicatedRegisters();
2051 
2052  size_t EstimateSearchSpaceComplexity() const;
2053  void NarrowSearchSpaceByDetectingSupersets();
2054  void NarrowSearchSpaceByCollapsingUnrolledCode();
2055  void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2056  void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2057  void NarrowSearchSpaceByFilterPostInc();
2058  void NarrowSearchSpaceByDeletingCostlyFormulas();
2059  void NarrowSearchSpaceByPickingWinnerRegs();
2060  void NarrowSearchSpaceUsingHeuristics();
2061 
2062  void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2063  Cost &SolutionCost,
2065  const Cost &CurCost,
2066  const SmallPtrSet<const SCEV *, 16> &CurRegs,
2067  DenseSet<const SCEV *> &VisitedRegs) const;
2068  void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2069 
2071  HoistInsertPosition(BasicBlock::iterator IP,
2072  const SmallVectorImpl<Instruction *> &Inputs) const;
2074  AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2075  const LSRFixup &LF,
2076  const LSRUse &LU,
2077  SCEVExpander &Rewriter) const;
2078 
2079  Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2081  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2082  void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2083  const Formula &F, SCEVExpander &Rewriter,
2084  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2085  void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2087  SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2088  void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2089 
2090 public:
2091  LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2093  TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU);
2094 
2095  bool getChanged() const { return Changed; }
2096  const SmallVectorImpl<WeakVH> &getScalarEvolutionIVs() const {
2097  return ScalarEvolutionIVs;
2098  }
2099 
2100  void print_factors_and_types(raw_ostream &OS) const;
2101  void print_fixups(raw_ostream &OS) const;
2102  void print_uses(raw_ostream &OS) const;
2103  void print(raw_ostream &OS) const;
2104  void dump() const;
2105 };
2106 
2107 } // end anonymous namespace
2108 
2109 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
2110 /// the cast operation.
2111 void LSRInstance::OptimizeShadowIV() {
2112  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2113  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2114  return;
2115 
2116  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2117  UI != E; /* empty */) {
2118  IVUsers::const_iterator CandidateUI = UI;
2119  ++UI;
2120  Instruction *ShadowUse = CandidateUI->getUser();
2121  Type *DestTy = nullptr;
2122  bool IsSigned = false;
2123 
2124  /* If shadow use is a int->float cast then insert a second IV
2125  to eliminate this cast.
2126 
2127  for (unsigned i = 0; i < n; ++i)
2128  foo((double)i);
2129 
2130  is transformed into
2131 
2132  double d = 0.0;
2133  for (unsigned i = 0; i < n; ++i, ++d)
2134  foo(d);
2135  */
2136  if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2137  IsSigned = false;
2138  DestTy = UCast->getDestTy();
2139  }
2140  else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2141  IsSigned = true;
2142  DestTy = SCast->getDestTy();
2143  }
2144  if (!DestTy) continue;
2145 
2146  // If target does not support DestTy natively then do not apply
2147  // this transformation.
2148  if (!TTI.isTypeLegal(DestTy)) continue;
2149 
2150  PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2151  if (!PH) continue;
2152  if (PH->getNumIncomingValues() != 2) continue;
2153 
2154  // If the calculation in integers overflows, the result in FP type will
2155  // differ. So we only can do this transformation if we are guaranteed to not
2156  // deal with overflowing values
2157  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2158  if (!AR) continue;
2159  if (IsSigned && !AR->hasNoSignedWrap()) continue;
2160  if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2161 
2162  Type *SrcTy = PH->getType();
2163  int Mantissa = DestTy->getFPMantissaWidth();
2164  if (Mantissa == -1) continue;
2165  if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2166  continue;
2167 
2168  unsigned Entry, Latch;
2169  if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2170  Entry = 0;
2171  Latch = 1;
2172  } else {
2173  Entry = 1;
2174  Latch = 0;
2175  }
2176 
2177  ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2178  if (!Init) continue;
2179  Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2180  (double)Init->getSExtValue() :
2181  (double)Init->getZExtValue());
2182 
2183  BinaryOperator *Incr =
2184  dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2185  if (!Incr) continue;
2186  if (Incr->getOpcode() != Instruction::Add
2187  && Incr->getOpcode() != Instruction::Sub)
2188  continue;
2189 
2190  /* Initialize new IV, double d = 0.0 in above example. */
2191  ConstantInt *C = nullptr;
2192  if (Incr->getOperand(0) == PH)
2193  C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2194  else if (Incr->getOperand(1) == PH)
2195  C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2196  else
2197  continue;
2198 
2199  if (!C) continue;
2200 
2201  // Ignore negative constants, as the code below doesn't handle them
2202  // correctly. TODO: Remove this restriction.
2203  if (!C->getValue().isStrictlyPositive()) continue;
2204 
2205  /* Add new PHINode. */
2206  PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
2207 
2208  /* create new increment. '++d' in above example. */
2209  Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2210  BinaryOperator *NewIncr =
2212  Instruction::FAdd : Instruction::FSub,
2213  NewPH, CFP, "IV.S.next.", Incr);
2214 
2215  NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2216  NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2217 
2218  /* Remove cast operation */
2219  ShadowUse->replaceAllUsesWith(NewPH);
2220  ShadowUse->eraseFromParent();
2221  Changed = true;
2222  break;
2223  }
2224 }
2225 
2226 /// If Cond has an operand that is an expression of an IV, set the IV user and
2227 /// stride information and return true, otherwise return false.
2228 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2229  for (IVStrideUse &U : IU)
2230  if (U.getUser() == Cond) {
2231  // NOTE: we could handle setcc instructions with multiple uses here, but
2232  // InstCombine does it as well for simple uses, it's not clear that it
2233  // occurs enough in real life to handle.
2234  CondUse = &U;
2235  return true;
2236  }
2237  return false;
2238 }
2239 
2240 /// Rewrite the loop's terminating condition if it uses a max computation.
2241 ///
2242 /// This is a narrow solution to a specific, but acute, problem. For loops
2243 /// like this:
2244 ///
2245 /// i = 0;
2246 /// do {
2247 /// p[i] = 0.0;
2248 /// } while (++i < n);
2249 ///
2250 /// the trip count isn't just 'n', because 'n' might not be positive. And
2251 /// unfortunately this can come up even for loops where the user didn't use
2252 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2253 /// will commonly be lowered like this:
2254 ///
2255 /// if (n > 0) {
2256 /// i = 0;
2257 /// do {
2258 /// p[i] = 0.0;
2259 /// } while (++i < n);
2260 /// }
2261 ///
2262 /// and then it's possible for subsequent optimization to obscure the if
2263 /// test in such a way that indvars can't find it.
2264 ///
2265 /// When indvars can't find the if test in loops like this, it creates a
2266 /// max expression, which allows it to give the loop a canonical
2267 /// induction variable:
2268 ///
2269 /// i = 0;
2270 /// max = n < 1 ? 1 : n;
2271 /// do {
2272 /// p[i] = 0.0;
2273 /// } while (++i != max);
2274 ///
2275 /// Canonical induction variables are necessary because the loop passes
2276 /// are designed around them. The most obvious example of this is the
2277 /// LoopInfo analysis, which doesn't remember trip count values. It
2278 /// expects to be able to rediscover the trip count each time it is
2279 /// needed, and it does this using a simple analysis that only succeeds if
2280 /// the loop has a canonical induction variable.
2281 ///
2282 /// However, when it comes time to generate code, the maximum operation
2283 /// can be quite costly, especially if it's inside of an outer loop.
2284 ///
2285 /// This function solves this problem by detecting this type of loop and
2286 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2287 /// the instructions for the maximum computation.
2288 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2289  // Check that the loop matches the pattern we're looking for.
2290  if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2291  Cond->getPredicate() != CmpInst::ICMP_NE)
2292  return Cond;
2293 
2294  SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2295  if (!Sel || !Sel->hasOneUse()) return Cond;
2296 
2297  const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2298  if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2299  return Cond;
2300  const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2301 
2302  // Add one to the backedge-taken count to get the trip count.
2303  const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2304  if (IterationCount != SE.getSCEV(Sel)) return Cond;
2305 
2306  // Check for a max calculation that matches the pattern. There's no check
2307  // for ICMP_ULE here because the comparison would be with zero, which
2308  // isn't interesting.
2310  const SCEVNAryExpr *Max = nullptr;
2311  if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2312  Pred = ICmpInst::ICMP_SLE;
2313  Max = S;
2314  } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2315  Pred = ICmpInst::ICMP_SLT;
2316  Max = S;
2317  } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2318  Pred = ICmpInst::ICMP_ULT;
2319  Max = U;
2320  } else {
2321  // No match; bail.
2322  return Cond;
2323  }
2324 
2325  // To handle a max with more than two operands, this optimization would
2326  // require additional checking and setup.
2327  if (Max->getNumOperands() != 2)
2328  return Cond;
2329 
2330  const SCEV *MaxLHS = Max->getOperand(0);
2331  const SCEV *MaxRHS = Max->getOperand(1);
2332 
2333  // ScalarEvolution canonicalizes constants to the left. For < and >, look
2334  // for a comparison with 1. For <= and >=, a comparison with zero.
2335  if (!MaxLHS ||
2336  (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2337  return Cond;
2338 
2339  // Check the relevant induction variable for conformance to
2340  // the pattern.
2341  const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2342  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2343  if (!AR || !AR->isAffine() ||
2344  AR->getStart() != One ||
2345  AR->getStepRecurrence(SE) != One)
2346  return Cond;
2347 
2348  assert(AR->getLoop() == L &&
2349  "Loop condition operand is an addrec in a different loop!");
2350 
2351  // Check the right operand of the select, and remember it, as it will
2352  // be used in the new comparison instruction.
2353  Value *NewRHS = nullptr;
2354  if (ICmpInst::isTrueWhenEqual(Pred)) {
2355  // Look for n+1, and grab n.
2356  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2357  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2358  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2359  NewRHS = BO->getOperand(0);
2360  if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2361  if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2362  if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2363  NewRHS = BO->getOperand(0);
2364  if (!NewRHS)
2365  return Cond;
2366  } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2367  NewRHS = Sel->getOperand(1);
2368  else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2369  NewRHS = Sel->getOperand(2);
2370  else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2371  NewRHS = SU->getValue();
2372  else
2373  // Max doesn't match expected pattern.
2374  return Cond;
2375 
2376  // Determine the new comparison opcode. It may be signed or unsigned,
2377  // and the original comparison may be either equality or inequality.
2378  if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2379  Pred = CmpInst::getInversePredicate(Pred);
2380 
2381  // Ok, everything looks ok to change the condition into an SLT or SGE and
2382  // delete the max calculation.
2383  ICmpInst *NewCond =
2384  new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
2385 
2386  // Delete the max calculation instructions.
2387  NewCond->setDebugLoc(Cond->getDebugLoc());
2388  Cond->replaceAllUsesWith(NewCond);
2389  CondUse->setUser(NewCond);
2390  Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2391  Cond->eraseFromParent();
2392  Sel->eraseFromParent();
2393  if (Cmp->use_empty())
2394  Cmp->eraseFromParent();
2395  return NewCond;
2396 }
2397 
2398 /// Change loop terminating condition to use the postinc iv when possible.
2399 void
2400 LSRInstance::OptimizeLoopTermCond() {
2402 
2403  // We need a different set of heuristics for rotated and non-rotated loops.
2404  // If a loop is rotated then the latch is also the backedge, so inserting
2405  // post-inc expressions just before the latch is ideal. To reduce live ranges
2406  // it also makes sense to rewrite terminating conditions to use post-inc
2407  // expressions.
2408  //
2409  // If the loop is not rotated then the latch is not a backedge; the latch
2410  // check is done in the loop head. Adding post-inc expressions before the
2411  // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2412  // in the loop body. In this case we do *not* want to use post-inc expressions
2413  // in the latch check, and we want to insert post-inc expressions before
2414  // the backedge.
2415  BasicBlock *LatchBlock = L->getLoopLatch();
2416  SmallVector<BasicBlock*, 8> ExitingBlocks;
2417  L->getExitingBlocks(ExitingBlocks);
2418  if (llvm::all_of(ExitingBlocks, [&LatchBlock](const BasicBlock *BB) {
2419  return LatchBlock != BB;
2420  })) {
2421  // The backedge doesn't exit the loop; treat this as a head-tested loop.
2422  IVIncInsertPos = LatchBlock->getTerminator();
2423  return;
2424  }
2425 
2426  // Otherwise treat this as a rotated loop.
2427  for (BasicBlock *ExitingBlock : ExitingBlocks) {
2428  // Get the terminating condition for the loop if possible. If we
2429  // can, we want to change it to use a post-incremented version of its
2430  // induction variable, to allow coalescing the live ranges for the IV into
2431  // one register value.
2432 
2433  BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2434  if (!TermBr)
2435  continue;
2436  // FIXME: Overly conservative, termination condition could be an 'or' etc..
2437  if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2438  continue;
2439 
2440  // Search IVUsesByStride to find Cond's IVUse if there is one.
2441  IVStrideUse *CondUse = nullptr;
2442  ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2443  if (!FindIVUserForCond(Cond, CondUse))
2444  continue;
2445 
2446  // If the trip count is computed in terms of a max (due to ScalarEvolution
2447  // being unable to find a sufficient guard, for example), change the loop
2448  // comparison to use SLT or ULT instead of NE.
2449  // One consequence of doing this now is that it disrupts the count-down
2450  // optimization. That's not always a bad thing though, because in such
2451  // cases it may still be worthwhile to avoid a max.
2452  Cond = OptimizeMax(Cond, CondUse);
2453 
2454  // If this exiting block dominates the latch block, it may also use
2455  // the post-inc value if it won't be shared with other uses.
2456  // Check for dominance.
2457  if (!DT.dominates(ExitingBlock, LatchBlock))
2458  continue;
2459 
2460  // Conservatively avoid trying to use the post-inc value in non-latch
2461  // exits if there may be pre-inc users in intervening blocks.
2462  if (LatchBlock != ExitingBlock)
2463  for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2464  // Test if the use is reachable from the exiting block. This dominator
2465  // query is a conservative approximation of reachability.
2466  if (&*UI != CondUse &&
2467  !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2468  // Conservatively assume there may be reuse if the quotient of their
2469  // strides could be a legal scale.
2470  const SCEV *A = IU.getStride(*CondUse, L);
2471  const SCEV *B = IU.getStride(*UI, L);
2472  if (!A || !B) continue;
2473  if (SE.getTypeSizeInBits(A->getType()) !=
2474  SE.getTypeSizeInBits(B->getType())) {
2475  if (SE.getTypeSizeInBits(A->getType()) >
2476  SE.getTypeSizeInBits(B->getType()))
2477  B = SE.getSignExtendExpr(B, A->getType());
2478  else
2479  A = SE.getSignExtendExpr(A, B->getType());
2480  }
2481  if (const SCEVConstant *D =
2482  dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2483  const ConstantInt *C = D->getValue();
2484  // Stride of one or negative one can have reuse with non-addresses.
2485  if (C->isOne() || C->isMinusOne())
2486  goto decline_post_inc;
2487  // Avoid weird situations.
2488  if (C->getValue().getMinSignedBits() >= 64 ||
2489  C->getValue().isMinSignedValue())
2490  goto decline_post_inc;
2491  // Check for possible scaled-address reuse.
2492  if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2493  MemAccessTy AccessTy = getAccessType(
2494  TTI, UI->getUser(), UI->getOperandValToReplace());
2495  int64_t Scale = C->getSExtValue();
2496  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2497  /*BaseOffset=*/0,
2498  /*HasBaseReg=*/false, Scale,
2499  AccessTy.AddrSpace))
2500  goto decline_post_inc;
2501  Scale = -Scale;
2502  if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2503  /*BaseOffset=*/0,
2504  /*HasBaseReg=*/false, Scale,
2505  AccessTy.AddrSpace))
2506  goto decline_post_inc;
2507  }
2508  }
2509  }
2510 
2511  LLVM_DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
2512  << *Cond << '\n');
2513 
2514  // It's possible for the setcc instruction to be anywhere in the loop, and
2515  // possible for it to have multiple users. If it is not immediately before
2516  // the exiting block branch, move it.
2517  if (Cond->getNextNonDebugInstruction() != TermBr) {
2518  if (Cond->hasOneUse()) {
2519  Cond->moveBefore(TermBr);
2520  } else {
2521  // Clone the terminating condition and insert into the loopend.
2522  ICmpInst *OldCond = Cond;
2523  Cond = cast<ICmpInst>(Cond->clone());
2524  Cond->setName(L->getHeader()->getName() + ".termcond");
2525  ExitingBlock->getInstList().insert(TermBr->getIterator(), Cond);
2526 
2527  // Clone the IVUse, as the old use still exists!
2528  CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2529  TermBr->replaceUsesOfWith(OldCond, Cond);
2530  }
2531  }
2532 
2533  // If we get to here, we know that we can transform the setcc instruction to
2534  // use the post-incremented version of the IV, allowing us to coalesce the
2535  // live ranges for the IV correctly.
2536  CondUse->transformToPostInc(L);
2537  Changed = true;
2538 
2539  PostIncs.insert(Cond);
2540  decline_post_inc:;
2541  }
2542 
2543  // Determine an insertion point for the loop induction variable increment. It
2544  // must dominate all the post-inc comparisons we just set up, and it must
2545  // dominate the loop latch edge.
2546  IVIncInsertPos = L->getLoopLatch()->getTerminator();
2547  for (Instruction *Inst : PostIncs) {
2548  BasicBlock *BB =
2549  DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
2550  Inst->getParent());
2551  if (BB == Inst->getParent())
2552  IVIncInsertPos = Inst;
2553  else if (BB != IVIncInsertPos->getParent())
2554  IVIncInsertPos = BB->getTerminator();
2555  }
2556 }
2557 
2558 /// Determine if the given use can accommodate a fixup at the given offset and
2559 /// other details. If so, update the use and return true.
2560 bool LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
2561  bool HasBaseReg, LSRUse::KindType Kind,
2562  MemAccessTy AccessTy) {
2563  int64_t NewMinOffset = LU.MinOffset;
2564  int64_t NewMaxOffset = LU.MaxOffset;
2565  MemAccessTy NewAccessTy = AccessTy;
2566 
2567  // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2568  // something conservative, however this can pessimize in the case that one of
2569  // the uses will have all its uses outside the loop, for example.
2570  if (LU.Kind != Kind)
2571  return false;
2572 
2573  // Check for a mismatched access type, and fall back conservatively as needed.
2574  // TODO: Be less conservative when the type is similar and can use the same
2575  // addressing modes.
2576  if (Kind == LSRUse::Address) {
2577  if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2578  NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2579  AccessTy.AddrSpace);
2580  }
2581  }
2582 
2583  // Conservatively assume HasBaseReg is true for now.
2584  if (NewOffset < LU.MinOffset) {
2585  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2586  LU.MaxOffset - NewOffset, HasBaseReg))
2587  return false;
2588  NewMinOffset = NewOffset;
2589  } else if (NewOffset > LU.MaxOffset) {
2590  if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2591  NewOffset - LU.MinOffset, HasBaseReg))
2592  return false;
2593  NewMaxOffset = NewOffset;
2594  }
2595 
2596  // Update the use.
2597  LU.MinOffset = NewMinOffset;
2598  LU.MaxOffset = NewMaxOffset;
2599  LU.AccessTy = NewAccessTy;
2600  return true;
2601 }
2602 
2603 /// Return an LSRUse index and an offset value for a fixup which needs the given
2604 /// expression, with the given kind and optional access type. Either reuse an
2605 /// existing use or create a new one, as needed.
2606 std::pair<size_t, int64_t> LSRInstance::getUse(const SCEV *&Expr,
2607  LSRUse::KindType Kind,
2608  MemAccessTy AccessTy) {
2609  const SCEV *Copy = Expr;
2610  int64_t Offset = ExtractImmediate(Expr, SE);
2611 
2612  // Basic uses can't accept any offset, for example.
2613  if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2614  Offset, /*HasBaseReg=*/ true)) {
2615  Expr = Copy;
2616  Offset = 0;
2617  }
2618 
2619  std::pair<UseMapTy::iterator, bool> P =
2620  UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2621  if (!P.second) {
2622  // A use already existed with this base.
2623  size_t LUIdx = P.first->second;
2624  LSRUse &LU = Uses[LUIdx];
2625  if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2626  // Reuse this use.
2627  return std::make_pair(LUIdx, Offset);
2628  }
2629 
2630  // Create a new use.
2631  size_t LUIdx = Uses.size();
2632  P.first->second = LUIdx;
2633  Uses.push_back(LSRUse(Kind, AccessTy));
2634  LSRUse &LU = Uses[LUIdx];
2635 
2636  LU.MinOffset = Offset;
2637  LU.MaxOffset = Offset;
2638  return std::make_pair(LUIdx, Offset);
2639 }
2640 
2641 /// Delete the given use from the Uses list.
2642 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2643  if (&LU != &Uses.back())
2644  std::swap(LU, Uses.back());
2645  Uses.pop_back();
2646 
2647  // Update RegUses.
2648  RegUses.swapAndDropUse(LUIdx, Uses.size());
2649 }
2650 
2651 /// Look for a use distinct from OrigLU which is has a formula that has the same
2652 /// registers as the given formula.
2653 LSRUse *
2654 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2655  const LSRUse &OrigLU) {
2656  // Search all uses for the formula. This could be more clever.
2657  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2658  LSRUse &LU = Uses[LUIdx];
2659  // Check whether this use is close enough to OrigLU, to see whether it's
2660  // worthwhile looking through its formulae.
2661  // Ignore ICmpZero uses because they may contain formulae generated by
2662  // GenerateICmpZeroScales, in which case adding fixup offsets may
2663  // be invalid.
2664  if (&LU != &OrigLU &&
2665  LU.Kind != LSRUse::ICmpZero &&
2666  LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2667  LU.WidestFixupType == OrigLU.WidestFixupType &&
2668  LU.HasFormulaWithSameRegs(OrigF)) {
2669  // Scan through this use's formulae.
2670  for (const Formula &F : LU.Formulae) {
2671  // Check to see if this formula has the same registers and symbols
2672  // as OrigF.
2673  if (F.BaseRegs == OrigF.BaseRegs &&
2674  F.ScaledReg == OrigF.ScaledReg &&
2675  F.BaseGV == OrigF.BaseGV &&
2676  F.Scale == OrigF.Scale &&
2677  F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2678  if (F.BaseOffset == 0)
2679  return &LU;
2680  // This is the formula where all the registers and symbols matched;
2681  // there aren't going to be any others. Since we declined it, we
2682  // can skip the rest of the formulae and proceed to the next LSRUse.
2683  break;
2684  }
2685  }
2686  }
2687  }
2688 
2689  // Nothing looked good.
2690  return nullptr;
2691 }
2692 
2693 void LSRInstance::CollectInterestingTypesAndFactors() {
2695 
2696  // Collect interesting types and strides.
2698  for (const IVStrideUse &U : IU) {
2699  const SCEV *Expr = IU.getExpr(U);
2700 
2701  // Collect interesting types.
2702  Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2703 
2704  // Add strides for mentioned loops.
2705  Worklist.push_back(Expr);
2706  do {
2707  const SCEV *S = Worklist.pop_back_val();
2708  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2709  if (AR->getLoop() == L)
2710  Strides.insert(AR->getStepRecurrence(SE));
2711  Worklist.push_back(AR->getStart());
2712  } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2713  Worklist.append(Add->op_begin(), Add->op_end());
2714  }
2715  } while (!Worklist.empty());
2716  }
2717 
2718  // Compute interesting factors from the set of interesting strides.
2720  I = Strides.begin(), E = Strides.end(); I != E; ++I)
2722  std::next(I); NewStrideIter != E; ++NewStrideIter) {
2723  const SCEV *OldStride = *I;
2724  const SCEV *NewStride = *NewStrideIter;
2725 
2726  if (SE.getTypeSizeInBits(OldStride->getType()) !=
2727  SE.getTypeSizeInBits(NewStride->getType())) {
2728  if (SE.getTypeSizeInBits(OldStride->getType()) >
2729  SE.getTypeSizeInBits(NewStride->getType()))
2730  NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2731  else
2732  OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2733  }
2734  if (const SCEVConstant *Factor =
2735  dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2736  SE, true))) {
2737  if (Factor->getAPInt().getMinSignedBits() <= 64 && !Factor->isZero())
2738  Factors.insert(Factor->getAPInt().getSExtValue());
2739  } else if (const SCEVConstant *Factor =
2740  dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2741  NewStride,
2742  SE, true))) {
2743  if (Factor->getAPInt().getMinSignedBits() <= 64 && !Factor->isZero())
2744  Factors.insert(Factor->getAPInt().getSExtValue());
2745  }
2746  }
2747 
2748  // If all uses use the same type, don't bother looking for truncation-based
2749  // reuse.
2750  if (Types.size() == 1)
2751  Types.clear();
2752 
2753  LLVM_DEBUG(print_factors_and_types(dbgs()));
2754 }
2755 
2756 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2757 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2758 /// IVStrideUses, we could partially skip this.
2759 static User::op_iterator
2761  Loop *L, ScalarEvolution &SE) {
2762  for(; OI != OE; ++OI) {
2763  if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2764  if (!SE.isSCEVable(Oper->getType()))
2765  continue;
2766 
2767  if (const SCEVAddRecExpr *AR =
2768  dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2769  if (AR->getLoop() == L)
2770  break;
2771  }
2772  }
2773  }
2774  return OI;
2775 }
2776 
2777 /// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2778 /// a convenient helper.
2779 static Value *getWideOperand(Value *Oper) {
2780  if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2781  return Trunc->getOperand(0);
2782  return Oper;
2783 }
2784 
2785 /// Return true if we allow an IV chain to include both types.
2786 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
2787  Type *LType = LVal->getType();
2788  Type *RType = RVal->getType();
2789  return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy() &&
2790  // Different address spaces means (possibly)
2791  // different types of the pointer implementation,
2792  // e.g. i16 vs i32 so disallow that.
2793  (LType->getPointerAddressSpace() ==
2794  RType->getPointerAddressSpace()));
2795 }
2796 
2797 /// Return an approximation of this SCEV expression's "base", or NULL for any
2798 /// constant. Returning the expression itself is conservative. Returning a
2799 /// deeper subexpression is more precise and valid as long as it isn't less
2800 /// complex than another subexpression. For expressions involving multiple
2801 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2802 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2803 /// IVInc==b-a.
2804 ///
2805 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2806 /// SCEVUnknown, we simply return the rightmost SCEV operand.
2807 static const SCEV *getExprBase(const SCEV *S) {
2808  switch (S->getSCEVType()) {
2809  default: // uncluding scUnknown.
2810  return S;
2811  case scConstant:
2812  return nullptr;
2813  case scTruncate:
2814  return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
2815  case scZeroExtend:
2816  return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
2817  case scSignExtend:
2818  return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
2819  case scAddExpr: {
2820  // Skip over scaled operands (scMulExpr) to follow add operands as long as
2821  // there's nothing more complex.
2822  // FIXME: not sure if we want to recognize negation.
2823  const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
2824  for (const SCEV *SubExpr : reverse(Add->operands())) {
2825  if (SubExpr->getSCEVType() == scAddExpr)
2826  return getExprBase(SubExpr);
2827 
2828  if (SubExpr->getSCEVType() != scMulExpr)
2829  return SubExpr;
2830  }
2831  return S; // all operands are scaled, be conservative.
2832  }
2833  case scAddRecExpr:
2834  return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
2835  }
2836  llvm_unreachable("Unknown SCEV kind!");
2837 }
2838 
2839 /// Return true if the chain increment is profitable to expand into a loop
2840 /// invariant value, which may require its own register. A profitable chain
2841 /// increment will be an offset relative to the same base. We allow such offsets
2842 /// to potentially be used as chain increment as long as it's not obviously
2843 /// expensive to expand using real instructions.
2844 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
2845  const SCEV *IncExpr,
2846  ScalarEvolution &SE) {
2847  // Aggressively form chains when -stress-ivchain.
2848  if (StressIVChain)
2849  return true;
2850 
2851  // Do not replace a constant offset from IV head with a nonconstant IV
2852  // increment.
2853  if (!isa<SCEVConstant>(IncExpr)) {
2854  const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
2855  if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
2856  return false;
2857  }
2858 
2859  SmallPtrSet<const SCEV*, 8> Processed;
2860  return !isHighCostExpansion(IncExpr, Processed, SE);
2861 }
2862 
2863 /// Return true if the number of registers needed for the chain is estimated to
2864 /// be less than the number required for the individual IV users. First prohibit
2865 /// any IV users that keep the IV live across increments (the Users set should
2866 /// be empty). Next count the number and type of increments in the chain.
2867 ///
2868 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
2869 /// effectively use postinc addressing modes. Only consider it profitable it the
2870 /// increments can be computed in fewer registers when chained.
2871 ///
2872 /// TODO: Consider IVInc free if it's already used in another chains.
2873 static bool isProfitableChain(IVChain &Chain,
2875  ScalarEvolution &SE,
2876  const TargetTransformInfo &TTI) {
2877  if (StressIVChain)
2878  return true;
2879 
2880  if (!Chain.hasIncs())
2881  return false;
2882 
2883  if (!Users.empty()) {
2884  LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
2885  for (Instruction *Inst
2886  : Users) { dbgs() << " " << *Inst << "\n"; });
2887  return false;
2888  }
2889  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
2890 
2891  // The chain itself may require a register, so intialize cost to 1.
2892  int cost = 1;
2893 
2894  // A complete chain likely eliminates the need for keeping the original IV in
2895  // a register. LSR does not currently know how to form a complete chain unless
2896  // the header phi already exists.
2897  if (isa<PHINode>(Chain.tailUserInst())
2898  && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
2899  --cost;
2900  }
2901  const SCEV *LastIncExpr = nullptr;
2902  unsigned NumConstIncrements = 0;
2903  unsigned NumVarIncrements = 0;
2904  unsigned NumReusedIncrements = 0;
2905 
2906  if (TTI.isProfitableLSRChainElement(Chain.Incs[0].UserInst))
2907  return true;
2908 
2909  for (const IVInc &Inc : Chain) {
2910  if (TTI.isProfitableLSRChainElement(Inc.UserInst))
2911  return true;
2912  if (Inc.IncExpr->isZero())
2913  continue;
2914 
2915  // Incrementing by zero or some constant is neutral. We assume constants can
2916  // be folded into an addressing mode or an add's immediate operand.
2917  if (isa<SCEVConstant>(Inc.IncExpr)) {
2918  ++NumConstIncrements;
2919  continue;
2920  }
2921 
2922  if (Inc.IncExpr == LastIncExpr)
2923  ++NumReusedIncrements;
2924  else
2925  ++NumVarIncrements;
2926 
2927  LastIncExpr = Inc.IncExpr;
2928  }
2929  // An IV chain with a single increment is handled by LSR's postinc
2930  // uses. However, a chain with multiple increments requires keeping the IV's
2931  // value live longer than it needs to be if chained.
2932  if (NumConstIncrements > 1)
2933  --cost;
2934 
2935  // Materializing increment expressions in the preheader that didn't exist in
2936  // the original code may cost a register. For example, sign-extended array
2937  // indices can produce ridiculous increments like this:
2938  // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
2939  cost += NumVarIncrements;
2940 
2941  // Reusing variable increments likely saves a register to hold the multiple of
2942  // the stride.
2943  cost -= NumReusedIncrements;
2944 
2945  LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
2946  << "\n");
2947 
2948  return cost < 0;
2949 }
2950 
2951 /// Add this IV user to an existing chain or make it the head of a new chain.
2952 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2953  SmallVectorImpl<ChainUsers> &ChainUsersVec) {
2954  // When IVs are used as types of varying widths, they are generally converted
2955  // to a wider type with some uses remaining narrow under a (free) trunc.
2956  Value *const NextIV = getWideOperand(IVOper);
2957  const SCEV *const OperExpr = SE.getSCEV(NextIV);
2958  const SCEV *const OperExprBase = getExprBase(OperExpr);
2959 
2960  // Visit all existing chains. Check if its IVOper can be computed as a
2961  // profitable loop invariant increment from the last link in the Chain.
2962  unsigned ChainIdx = 0, NChains = IVChainVec.size();
2963  const SCEV *LastIncExpr = nullptr;
2964  for (; ChainIdx < NChains; ++ChainIdx) {
2965  IVChain &Chain = IVChainVec[ChainIdx];
2966 
2967  // Prune the solution space aggressively by checking that both IV operands
2968  // are expressions that operate on the same unscaled SCEVUnknown. This
2969  // "base" will be canceled by the subsequent getMinusSCEV call. Checking
2970  // first avoids creating extra SCEV expressions.
2971  if (!StressIVChain && Chain.ExprBase != OperExprBase)
2972  continue;
2973 
2974  Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
2975  if (!isCompatibleIVType(PrevIV, NextIV))
2976  continue;
2977 
2978  // A phi node terminates a chain.
2979  if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
2980  continue;
2981 
2982  // The increment must be loop-invariant so it can be kept in a register.
2983  const SCEV *PrevExpr = SE.getSCEV(PrevIV);
2984  const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
2985  if (isa<SCEVCouldNotCompute>(IncExpr) || !SE.isLoopInvariant(IncExpr, L))
2986  continue;
2987 
2988  if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
2989  LastIncExpr = IncExpr;
2990  break;
2991  }
2992  }
2993  // If we haven't found a chain, create a new one, unless we hit the max. Don't
2994  // bother for phi nodes, because they must be last in the chain.
2995  if (ChainIdx == NChains) {
2996  if (isa<PHINode>(UserInst))
2997  return;
2998  if (NChains >= MaxChains && !StressIVChain) {
2999  LLVM_DEBUG(dbgs() << "IV Chain Limit\n");
3000  return;
3001  }
3002  LastIncExpr = OperExpr;
3003  // IVUsers may have skipped over sign/zero extensions. We don't currently
3004  // attempt to form chains involving extensions unless they can be hoisted
3005  // into this loop's AddRec.
3006  if (!isa<SCEVAddRecExpr>(LastIncExpr))
3007  return;
3008  ++NChains;
3009  IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
3010  OperExprBase));
3011  ChainUsersVec.resize(NChains);
3012  LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
3013  << ") IV=" << *LastIncExpr << "\n");
3014  } else {
3015  LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst
3016  << ") IV+" << *LastIncExpr << "\n");
3017  // Add this IV user to the end of the chain.
3018  IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
3019  }
3020  IVChain &Chain = IVChainVec[ChainIdx];
3021 
3022  SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
3023  // This chain's NearUsers become FarUsers.
3024  if (!LastIncExpr->isZero()) {
3025  ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
3026  NearUsers.end());
3027  NearUsers.clear();
3028  }
3029 
3030  // All other uses of IVOperand become near uses of the chain.
3031  // We currently ignore intermediate values within SCEV expressions, assuming
3032  // they will eventually be used be the current chain, or can be computed
3033  // from one of the chain increments. To be more precise we could
3034  // transitively follow its user and only add leaf IV users to the set.
3035  for (User *U : IVOper->users()) {
3036  Instruction *OtherUse = dyn_cast<Instruction>(U);
3037  if (!OtherUse)
3038  continue;
3039  // Uses in the chain will no longer be uses if the chain is formed.
3040  // Include the head of the chain in this iteration (not Chain.begin()).
3041  IVChain::const_iterator IncIter = Chain.Incs.begin();
3042  IVChain::const_iterator IncEnd = Chain.Incs.end();
3043  for( ; IncIter != IncEnd; ++IncIter) {
3044  if (IncIter->UserInst == OtherUse)
3045  break;
3046  }
3047  if (IncIter != IncEnd)
3048  continue;
3049 
3050  if (SE.isSCEVable(OtherUse->getType())
3051  && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
3052  && IU.isIVUserOrOperand(OtherUse)) {
3053  continue;
3054  }
3055  NearUsers.insert(OtherUse);
3056  }
3057 
3058  // Since this user is part of the chain, it's no longer considered a use
3059  // of the chain.
3060  ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3061 }
3062 
3063 /// Populate the vector of Chains.
3064 ///
3065 /// This decreases ILP at the architecture level. Targets with ample registers,
3066 /// multiple memory ports, and no register renaming probably don't want
3067 /// this. However, such targets should probably disable LSR altogether.
3068 ///
3069 /// The job of LSR is to make a reasonable choice of induction variables across
3070 /// the loop. Subsequent passes can easily "unchain" computation exposing more
3071 /// ILP *within the loop* if the target wants it.
3072 ///
3073 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
3074 /// will not reorder memory operations, it will recognize this as a chain, but
3075 /// will generate redundant IV increments. Ideally this would be corrected later
3076 /// by a smart scheduler:
3077 /// = A[i]
3078 /// = A[i+x]
3079 /// A[i] =
3080 /// A[i+x] =
3081 ///
3082 /// TODO: Walk the entire domtree within this loop, not just the path to the
3083 /// loop latch. This will discover chains on side paths, but requires
3084 /// maintaining multiple copies of the Chains state.
3085 void LSRInstance::CollectChains() {
3086  LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n");
3087  SmallVector<ChainUsers, 8> ChainUsersVec;
3088 
3089  SmallVector<BasicBlock *,8> LatchPath;
3090  BasicBlock *LoopHeader = L->getHeader();
3091  for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3092  Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3093  LatchPath.push_back(Rung->getBlock());
3094  }
3095  LatchPath.push_back(LoopHeader);
3096 
3097  // Walk the instruction stream from the loop header to the loop latch.
3098  for (BasicBlock *BB : reverse(LatchPath)) {
3099  for (Instruction &I : *BB) {
3100  // Skip instructions that weren't seen by IVUsers analysis.
3101  if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3102  continue;
3103 
3104  // Ignore users that are part of a SCEV expression. This way we only
3105  // consider leaf IV Users. This effectively rediscovers a portion of
3106  // IVUsers analysis but in program order this time.
3107  if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3108  continue;
3109 
3110  // Remove this instruction from any NearUsers set it may be in.
3111  for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3112  ChainIdx < NChains; ++ChainIdx) {
3113  ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3114  }
3115  // Search for operands that can be chained.
3116  SmallPtrSet<Instruction*, 4> UniqueOperands;
3117  User::op_iterator IVOpEnd = I.op_end();
3118  User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3119  while (IVOpIter != IVOpEnd) {
3120  Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3121  if (UniqueOperands.insert(IVOpInst).second)
3122  ChainInstruction(&I, IVOpInst, ChainUsersVec);
3123  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3124  }
3125  } // Continue walking down the instructions.
3126  } // Continue walking down the domtree.
3127  // Visit phi backedges to determine if the chain can generate the IV postinc.
3128  for (PHINode &PN : L->getHeader()->phis()) {
3129  if (!SE.isSCEVable(PN.getType()))
3130  continue;
3131 
3132  Instruction *IncV =
3133  dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3134  if (IncV)
3135  ChainInstruction(&PN, IncV, ChainUsersVec);
3136  }
3137  // Remove any unprofitable chains.
3138  unsigned ChainIdx = 0;
3139  for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3140  UsersIdx < NChains; ++UsersIdx) {
3141  if (!isProfitableChain(IVChainVec[UsersIdx],
3142  ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3143  continue;
3144  // Preserve the chain at UsesIdx.
3145  if (ChainIdx != UsersIdx)
3146  IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3147  FinalizeChain(IVChainVec[ChainIdx]);
3148  ++ChainIdx;
3149  }
3150  IVChainVec.resize(ChainIdx);
3151 }
3152 
3153 void LSRInstance::FinalizeChain(IVChain &Chain) {
3154  assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3155  LLVM_DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
3156 
3157  for (const IVInc &Inc : Chain) {
3158  LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n");
3159  auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3160  assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
3161  IVIncSet.insert(UseI);
3162  }
3163 }
3164 
3165 /// Return true if the IVInc can be folded into an addressing mode.
3166 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3167  Value *Operand, const TargetTransformInfo &TTI) {
3168  const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3169  if (!IncConst || !isAddressUse(TTI, UserInst, Operand))
3170  return false;
3171 
3172  if (IncConst->getAPInt().getMinSignedBits() > 64)
3173  return false;
3174 
3175  MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3176  int64_t IncOffset = IncConst->getValue()->getSExtValue();
3177  if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3178  IncOffset, /*HasBaseReg=*/false))
3179  return false;
3180 
3181  return true;
3182 }
3183 
3184 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3185 /// user's operand from the previous IV user's operand.
3186 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
3187  SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3188  // Find the new IVOperand for the head of the chain. It may have been replaced
3189  // by LSR.
3190  const IVInc &Head = Chain.Incs[0];
3191  User::op_iterator IVOpEnd = Head.UserInst->op_end();
3192  // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3193  User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3194  IVOpEnd, L, SE);
3195  Value *IVSrc = nullptr;
3196  while (IVOpIter != IVOpEnd) {
3197  IVSrc = getWideOperand(*IVOpIter);
3198 
3199  // If this operand computes the expression that the chain needs, we may use
3200  // it. (Check this after setting IVSrc which is used below.)
3201  //
3202  // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3203  // narrow for the chain, so we can no longer use it. We do allow using a
3204  // wider phi, assuming the LSR checked for free truncation. In that case we
3205  // should already have a truncate on this operand such that
3206  // getSCEV(IVSrc) == IncExpr.
3207  if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3208  || SE.getSCEV(IVSrc) == Head.IncExpr) {
3209  break;
3210  }
3211  IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3212  }
3213  if (IVOpIter == IVOpEnd) {
3214  // Gracefully give up on this chain.
3215  LLVM_DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3216  return;
3217  }
3218  assert(IVSrc && "Failed to find IV chain source");
3219 
3220  LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3221  Type *IVTy = IVSrc->getType();
3222  Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3223  const SCEV *LeftOverExpr = nullptr;
3224  for (const IVInc &Inc : Chain) {
3225  Instruction *InsertPt = Inc.UserInst;
3226  if (isa<PHINode>(InsertPt))
3227  InsertPt = L->getLoopLatch()->getTerminator();
3228 
3229  // IVOper will replace the current IV User's operand. IVSrc is the IV
3230  // value currently held in a register.
3231  Value *IVOper = IVSrc;
3232  if (!Inc.IncExpr->isZero()) {
3233  // IncExpr was the result of subtraction of two narrow values, so must
3234  // be signed.
3235  const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3236  LeftOverExpr = LeftOverExpr ?
3237  SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3238  }
3239  if (LeftOverExpr && !LeftOverExpr->isZero()) {
3240  // Expand the IV increment.
3241  Rewriter.clearPostInc();
3242  Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3243  const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3244  SE.getUnknown(IncV));
3245  IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3246 
3247  // If an IV increment can't be folded, use it as the next IV value.
3248  if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3249  assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3250  IVSrc = IVOper;
3251  LeftOverExpr = nullptr;
3252  }
3253  }
3254  Type *OperTy = Inc.IVOperand->getType();
3255  if (IVTy != OperTy) {
3256  assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3257  "cannot extend a chained IV");
3258  IRBuilder<> Builder(InsertPt);
3259  IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3260  }
3261  Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3262  if (auto *OperandIsInstr = dyn_cast<Instruction>(Inc.IVOperand))
3263  DeadInsts.emplace_back(OperandIsInstr);
3264  }
3265  // If LSR created a new, wider phi, we may also replace its postinc. We only
3266  // do this if we also found a wide value for the head of the chain.
3267  if (isa<PHINode>(Chain.tailUserInst())) {
3268  for (PHINode &Phi : L->getHeader()->phis()) {
3269  if (!isCompatibleIVType(&Phi, IVSrc))
3270  continue;
3271  Instruction *PostIncV = dyn_cast<Instruction>(
3272  Phi.getIncomingValueForBlock(L->getLoopLatch()));
3273  if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3274  continue;
3275  Value *IVOper = IVSrc;
3276  Type *PostIncTy = PostIncV->getType();
3277  if (IVTy != PostIncTy) {
3278  assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3280  Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3281  IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3282  }
3283  Phi.replaceUsesOfWith(PostIncV, IVOper);
3284  DeadInsts.emplace_back(PostIncV);
3285  }
3286  }
3287 }
3288 
3289 void LSRInstance::CollectFixupsAndInitialFormulae() {
3290  BranchInst *ExitBranch = nullptr;
3291  bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &TLI);
3292 
3293  for (const IVStrideUse &U : IU) {
3294  Instruction *UserInst = U.getUser();
3295  // Skip IV users that are part of profitable IV Chains.
3296  User::op_iterator UseI =
3297  find(UserInst->operands(), U.getOperandValToReplace());
3298  assert(UseI != UserInst->op_end() && "cannot find IV operand");
3299  if (IVIncSet.count(UseI)) {
3300  LLVM_DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3301  continue;
3302  }
3303 
3304  LSRUse::KindType Kind = LSRUse::Basic;
3305  MemAccessTy AccessTy;
3306  if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3307  Kind = LSRUse::Address;
3308  AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3309  }
3310 
3311  const SCEV *S = IU.getExpr(U);
3312  PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3313 
3314  // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3315  // (N - i == 0), and this allows (N - i) to be the expression that we work
3316  // with rather than just N or i, so we can consider the register
3317  // requirements for both N and i at the same time. Limiting this code to
3318  // equality icmps is not a problem because all interesting loops use
3319  // equality icmps, thanks to IndVarSimplify.
3320  if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
3321  // If CI can be saved in some target, like replaced inside hardware loop
3322  // in PowerPC, no need to generate initial formulae for it.
3323  if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
3324  continue;
3325  if (CI->isEquality()) {
3326  // Swap the operands if needed to put the OperandValToReplace on the
3327  // left, for consistency.
3328  Value *NV = CI->getOperand(1);
3329  if (NV == U.getOperandValToReplace()) {
3330  CI->setOperand(1, CI->getOperand(0));
3331  CI->setOperand(0, NV);
3332  NV = CI->getOperand(1);
3333  Changed = true;
3334  }
3335 
3336  // x == y --> x - y == 0
3337  const SCEV *N = SE.getSCEV(NV);
3338  if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE) &&
3339  (!NV->getType()->isPointerTy() ||
3340  SE.getPointerBase(N) == SE.getPointerBase(S))) {
3341  // S is normalized, so normalize N before folding it into S
3342  // to keep the result normalized.
3343  N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3344  Kind = LSRUse::ICmpZero;
3345  S = SE.getMinusSCEV(N, S);
3346  }
3347 
3348  // -1 and the negations of all interesting strides (except the negation
3349  // of -1) are now also interesting.
3350  for (size_t i = 0, e = Factors.size(); i != e; ++i)
3351  if (Factors[i] != -1)
3352  Factors.insert(-(uint64_t)Factors[i]);
3353  Factors.insert(-1);
3354  }
3355  }
3356 
3357  // Get or create an LSRUse.
3358  std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
3359  size_t LUIdx = P.first;
3360  int64_t Offset = P.second;
3361  LSRUse &LU = Uses[LUIdx];
3362 
3363  // Record the fixup.
3364  LSRFixup &LF = LU.getNewFixup();
3365  LF.UserInst = UserInst;
3366  LF.OperandValToReplace = U.getOperandValToReplace();
3367  LF.PostIncLoops = TmpPostIncLoops;
3368  LF.Offset = Offset;
3369  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3370 
3371  if (!LU.WidestFixupType ||
3372  SE.getTypeSizeInBits(LU.WidestFixupType) <
3373  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3374  LU.WidestFixupType = LF.OperandValToReplace->getType();
3375 
3376  // If this is the first use of this LSRUse, give it a formula.
3377  if (LU.Formulae.empty()) {
3378  InsertInitialFormula(S, LU, LUIdx);
3379  CountRegisters(LU.Formulae.back(), LUIdx);
3380  }
3381  }
3382 
3383  LLVM_DEBUG(print_fixups(dbgs()));
3384 }
3385 
3386 /// Insert a formula for the given expression into the given use, separating out
3387 /// loop-variant portions from loop-invariant and loop-computable portions.
3388 void
3389 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
3390  // Mark uses whose expressions cannot be expanded.
3391  if (!isSafeToExpand(S, SE, /*CanonicalMode*/ false))
3392  LU.RigidFormula = true;
3393 
3394  Formula F;
3395  F.initialMatch(S, L, SE);
3396  bool Inserted = InsertFormula(LU, LUIdx, F);
3397  assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3398 }
3399 
3400 /// Insert a simple single-register formula for the given expression into the
3401 /// given use.
3402 void
3403 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3404  LSRUse &LU, size_t LUIdx) {
3405  Formula F;
3406  F.BaseRegs.push_back(S);
3407  F.HasBaseReg = true;
3408  bool Inserted = InsertFormula(LU, LUIdx, F);
3409  assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3410 }
3411 
3412 /// Note which registers are used by the given formula, updating RegUses.
3413 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3414  if (F.ScaledReg)
3415  RegUses.countRegister(F.ScaledReg, LUIdx);
3416  for (const SCEV *BaseReg : F.BaseRegs)
3417  RegUses.countRegister(BaseReg, LUIdx);
3418 }
3419 
3420 /// If the given formula has not yet been inserted, add it to the list, and
3421 /// return true. Return false otherwise.
3422 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3423  // Do not insert formula that we will not be able to expand.
3424  assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3425  "Formula is illegal");
3426 
3427  if (!LU.InsertFormula(F, *L))
3428  return false;
3429 
3430  CountRegisters(F, LUIdx);
3431  return true;
3432 }
3433 
3434 /// Check for other uses of loop-invariant values which we're tracking. These
3435 /// other uses will pin these values in registers, making them less profitable
3436 /// for elimination.
3437 /// TODO: This currently misses non-constant addrec step registers.
3438 /// TODO: Should this give more weight to users inside the loop?
3439 void
3440 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3441  SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3443 
3444  while (!Worklist.empty()) {
3445  const SCEV *S = Worklist.pop_back_val();
3446 
3447  // Don't process the same SCEV twice
3448  if (!Visited.insert(S).second)
3449  continue;
3450 
3451  if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3452  Worklist.append(N->op_begin(), N->op_end());
3453  else if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(S))
3454  Worklist.push_back(C->getOperand());
3455  else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3456  Worklist.push_back(D->getLHS());
3457  Worklist.push_back(D->getRHS());
3458  } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3459  const Value *V = US->getValue();
3460  if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3461  // Look for instructions defined outside the loop.
3462  if (L->contains(Inst)) continue;
3463  } else if (isa<UndefValue>(V))
3464  // Undef doesn't have a live range, so it doesn't matter.
3465  continue;
3466  for (const Use &U : V->uses()) {
3467  const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3468  // Ignore non-instructions.
3469  if (!UserInst)
3470  continue;
3471  // Don't bother if the instruction is an EHPad.
3472  if (UserInst->isEHPad())
3473  continue;
3474  // Ignore instructions in other functions (as can happen with
3475  // Constants).
3476  if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3477  continue;
3478  // Ignore instructions not dominated by the loop.
3479  const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3480  UserInst->getParent() :
3481  cast<PHINode>(UserInst)->getIncomingBlock(
3482  PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3483  if (!DT.dominates(L->getHeader(), UseBB))
3484  continue;
3485  // Don't bother if the instruction is in a BB which ends in an EHPad.
3486  if (UseBB->getTerminator()->isEHPad())
3487  continue;
3488 
3489  // Ignore cases in which the currently-examined value could come from
3490  // a basic block terminated with an EHPad. This checks all incoming
3491  // blocks of the phi node since it is possible that the same incoming
3492  // value comes from multiple basic blocks, only some of which may end
3493  // in an EHPad. If any of them do, a subsequent rewrite attempt by this
3494  // pass would try to insert instructions into an EHPad, hitting an
3495  // assertion.
3496  if (isa<PHINode>(UserInst)) {
3497  const auto *PhiNode = cast<PHINode>(UserInst);
3498  bool HasIncompatibleEHPTerminatedBlock = false;
3499  llvm::Value *ExpectedValue = U;
3500  for (unsigned int I = 0; I < PhiNode->getNumIncomingValues(); I++) {
3501  if (PhiNode->getIncomingValue(I) == ExpectedValue) {
3502  if (PhiNode->getIncomingBlock(I)->getTerminator()->isEHPad()) {
3503  HasIncompatibleEHPTerminatedBlock = true;
3504  break;
3505  }
3506  }
3507  }
3508  if (HasIncompatibleEHPTerminatedBlock) {
3509  continue;
3510  }
3511  }
3512 
3513  // Don't bother rewriting PHIs in catchswitch blocks.
3514  if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3515  continue;
3516  // Ignore uses which are part of other SCEV expressions, to avoid
3517  // analyzing them multiple times.
3518  if (SE.isSCEVable(UserInst->getType())) {
3519  const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3520  // If the user is a no-op, look through to its uses.
3521  if (!isa<SCEVUnknown>(UserS))
3522  continue;
3523  if (UserS == US) {
3524  Worklist.push_back(
3525  SE.getUnknown(const_cast<Instruction *>(UserInst)));
3526  continue;
3527  }
3528  }
3529  // Ignore icmp instructions which are already being analyzed.
3530  if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3531  unsigned OtherIdx = !U.getOperandNo();
3532  Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3533  if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3534  continue;
3535  }
3536 
3537  std::pair<size_t, int64_t> P = getUse(
3538  S, LSRUse::Basic, MemAccessTy());
3539  size_t LUIdx = P.first;
3540  int64_t Offset = P.second;
3541  LSRUse &LU = Uses[LUIdx];
3542  LSRFixup &LF = LU.getNewFixup();
3543  LF.UserInst = const_cast<Instruction *>(UserInst);
3544  LF.OperandValToReplace = U;
3545  LF.Offset = Offset;
3546  LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3547  if (!LU.WidestFixupType ||
3548  SE.getTypeSizeInBits(LU.WidestFixupType) <
3549  SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3550  LU.WidestFixupType = LF.OperandValToReplace->getType();
3551  InsertSupplementalFormula(US, LU, LUIdx);
3552  CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3553  break;
3554  }
3555  }
3556  }
3557 }
3558 
3559 /// Split S into subexpressions which can be pulled out into separate
3560 /// registers. If C is non-null, multiply each subexpression by C.
3561 ///
3562 /// Return remainder expression after factoring the subexpressions captured by
3563 /// Ops. If Ops is complete, return NULL.
3564 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3566  const Loop *L,
3567  ScalarEvolution &SE,
3568  unsigned Depth = 0) {
3569  // Arbitrarily cap recursion to protect compile time.
3570  if (Depth >= 3)
3571  return S;
3572 
3573  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3574  // Break out add operands.
3575  for (const SCEV *S : Add->operands()) {
3576  const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3577  if (Remainder)
3578  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3579  }
3580  return nullptr;
3581  } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3582  // Split a non-zero base out of an addrec.
3583  if (AR->getStart()->isZero() || !AR->isAffine())
3584  return S;
3585 
3586  const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3587  C, Ops, L, SE, Depth+1);
3588  // Split the non-zero AddRec unless it is part of a nested recurrence that
3589  // does not pertain to this loop.
3590  if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3591  Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3592  Remainder = nullptr;
3593  }
3594  if (Remainder != AR->getStart()) {
3595  if (!Remainder)
3596  Remainder = SE.getConstant(AR->getType(), 0);
3597  return SE.getAddRecExpr(Remainder,
3598  AR->getStepRecurrence(SE),
3599  AR->getLoop(),
3600  //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3602  }
3603  } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3604  // Break (C * (a + b + c)) into C*a + C*b + C*c.
3605  if (Mul->getNumOperands() != 2)
3606  return S;
3607  if (const SCEVConstant *Op0 =
3608  dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3609  C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3610  const SCEV *Remainder =
3611  CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3612  if (Remainder)
3613  Ops.push_back(SE.getMulExpr(C, Remainder));
3614  return nullptr;
3615  }
3616  }
3617  return S;
3618 }
3619 
3620 /// Return true if the SCEV represents a value that may end up as a
3621 /// post-increment operation.
3623  LSRUse &LU, const SCEV *S, const Loop *L,
3624  ScalarEvolution &SE) {
3625  if (LU.Kind != LSRUse::Address ||
3626  !LU.AccessTy.getType()->isIntOrIntVectorTy())
3627  return false;
3628  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3629  if (!AR)
3630  return false;
3631  const SCEV *LoopStep = AR->getStepRecurrence(SE);
3632  if (!isa<SCEVConstant>(LoopStep))
3633  return false;
3634  // Check if a post-indexed load/store can be used.
3637  const SCEV *LoopStart = AR->getStart();
3638  if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3639  return true;
3640  }
3641  return false;
3642 }
3643 
3644 /// Helper function for LSRInstance::GenerateReassociations.
3645 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3646  const Formula &Base,
3647  unsigned Depth, size_t Idx,
3648  bool IsScaledReg) {
3649  const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3650  // Don't generate reassociations for the base register of a value that
3651  // may generate a post-increment operator. The reason is that the
3652  // reassociations cause extra base+register formula to be created,
3653  // and possibly chosen, but the post-increment is more efficient.
3654  if (AMK == TTI::AMK_PostIndexed && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3655  return;
3657  const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3658  if (Remainder)
3659  AddOps.push_back(Remainder);
3660 
3661  if (AddOps.size() == 1)
3662  return;
3663 
3664  for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3665  JE = AddOps.end();
3666  J != JE; ++J) {
3667  // Loop-variant "unknown" values are uninteresting; we won't be able to
3668  // do anything meaningful with them.
3669  if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3670  continue;
3671 
3672  // Don't pull a constant into a register if the constant could be folded
3673  // into an immediate field.
3674  if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3675  LU.AccessTy, *J, Base.getNumRegs() > 1))
3676  continue;
3677 
3678  // Collect all operands except *J.
3679  SmallVector<const SCEV *, 8> InnerAddOps(
3680  ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3681  InnerAddOps.append(std::next(J),
3682  ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3683 
3684  // Don't leave just a constant behind in a register if the constant could
3685  // be folded into an immediate field.
3686  if (InnerAddOps.size() == 1 &&
3687  isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3688  LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3689  continue;
3690 
3691  const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3692  if (InnerSum->isZero())
3693  continue;
3694  Formula F = Base;
3695 
3696  // Add the remaining pieces of the add back into the new formula.
3697  const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3698  if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3699  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3700  InnerSumSC->getValue()->getZExtValue())) {
3701  F.UnfoldedOffset =
3702  (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue();
3703  if (IsScaledReg)
3704  F.ScaledReg = nullptr;
3705  else
3706  F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3707  } else if (IsScaledReg)
3708  F.ScaledReg = InnerSum;
3709  else
3710  F.BaseRegs[Idx] = InnerSum;
3711 
3712  // Add J as its own register, or an unfolded immediate.
3713  const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3714  if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3715  TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
3716  SC->getValue()->getZExtValue()))
3717  F.UnfoldedOffset =
3718  (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue();
3719  else
3720  F.BaseRegs.push_back(*J);
3721  // We may have changed the number of register in base regs, adjust the
3722  // formula accordingly.
3723  F.canonicalize(*L);
3724 
3725  if (InsertFormula(LU, LUIdx, F))
3726  // If that formula hadn't been seen before, recurse to find more like
3727  // it.
3728  // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
3729  // Because just Depth is not enough to bound compile time.
3730  // This means that every time AddOps.size() is greater 16^x we will add
3731  // x to Depth.
3732  GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
3733  Depth + 1 + (Log2_32(AddOps.size()) >> 2));
3734  }
3735 }
3736 
3737 /// Split out subexpressions from adds and the bases of addrecs.
3738 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
3739  Formula Base, unsigned Depth) {
3740  assert(Base.isCanonical(*L) && "Input must be in the canonical form");
3741  // Arbitrarily cap recursion to protect compile time.
3742  if (Depth >= 3)
3743  return;
3744 
3745  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3746  GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
3747 
3748  if (Base.Scale == 1)
3749  GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
3750  /* Idx */ -1, /* IsScaledReg */ true);
3751 }
3752 
3753 /// Generate a formula consisting of all of the loop-dominating registers added
3754 /// into a single register.
3755 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
3756  Formula Base) {
3757  // This method is only interesting on a plurality of registers.
3758  if (Base.BaseRegs.size() + (Base.Scale == 1) +
3759  (Base.UnfoldedOffset != 0) <= 1)
3760  return;
3761 
3762  // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
3763  // processing the formula.
3764  Base.unscale();
3766  Formula NewBase = Base;
3767  NewBase.BaseRegs.clear();
3768  Type *CombinedIntegerType = nullptr;
3769  for (const SCEV *BaseReg : Base.BaseRegs) {
3770  if (SE.properlyDominates(BaseReg, L->getHeader()) &&
3771  !SE.hasComputableLoopEvolution(BaseReg, L)) {
3772  if (!CombinedIntegerType)
3773  CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
3774  Ops.push_back(BaseReg);
3775  }
3776  else
3777  NewBase.BaseRegs.push_back(BaseReg);
3778  }
3779 
3780  // If no register is relevant, we're done.
3781  if (Ops.size() == 0)
3782  return;
3783 
3784  // Utility function for generating the required variants of the combined
3785  // registers.
3786  auto GenerateFormula = [&](const SCEV *Sum) {
3787  Formula F = NewBase;
3788 
3789  // TODO: If Sum is zero, it probably means ScalarEvolution missed an
3790  // opportunity to fold something. For now, just ignore such cases
3791  // rather than proceed with zero in a register.
3792  if (Sum->isZero())
3793  return;
3794 
3795  F.BaseRegs.push_back(Sum);
3796  F.canonicalize(*L);
3797  (void)InsertFormula(LU, LUIdx, F);
3798  };
3799 
3800  // If we collected at least two registers, generate a formula combining them.
3801  if (Ops.size() > 1) {
3802  SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
3803  GenerateFormula(SE.getAddExpr(OpsCopy));
3804  }
3805 
3806  // If we have an unfolded offset, generate a formula combining it with the
3807  // registers collected.
3808  if (NewBase.UnfoldedOffset) {
3809  assert(CombinedIntegerType && "Missing a type for the unfolded offset");
3810  Ops.push_back(SE.getConstant(CombinedIntegerType, NewBase.UnfoldedOffset,
3811  true));
3812  NewBase.UnfoldedOffset = 0;
3813  GenerateFormula(SE.getAddExpr(Ops));
3814  }
3815 }
3816 
3817 /// Helper function for LSRInstance::GenerateSymbolicOffsets.
3818 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
3819  const Formula &Base, size_t Idx,
3820  bool IsScaledReg) {
3821  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3822  GlobalValue *GV = ExtractSymbol(G, SE);
3823  if (G->isZero() || !GV)
3824  return;
3825  Formula F = Base;
3826  F.BaseGV = GV;
3827  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3828  return;
3829  if (IsScaledReg)
3830  F.ScaledReg = G;
3831  else
3832  F.BaseRegs[Idx] = G;
3833  (void)InsertFormula(LU, LUIdx, F);
3834 }
3835 
3836 /// Generate reuse formulae using symbolic offsets.
3837 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
3838  Formula Base) {
3839  // We can't add a symbolic offset if the address already contains one.
3840  if (Base.BaseGV) return;
3841 
3842  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3843  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
3844  if (Base.Scale == 1)
3845  GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
3846  /* IsScaledReg */ true);
3847 }
3848 
3849 /// Helper function for LSRInstance::GenerateConstantOffsets.
3850 void LSRInstance::GenerateConstantOffsetsImpl(
3851  LSRUse &LU, unsigned LUIdx, const Formula &Base,
3852  const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) {
3853 
3854  auto GenerateOffset = [&](const SCEV *G, int64_t Offset) {
3855  Formula F = Base;
3856  F.BaseOffset = (uint64_t)Base.BaseOffset - Offset;
3857 
3858  if (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) {
3859  // Add the offset to the base register.
3860  const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G);
3861  // If it cancelled out, drop the base register, otherwise update it.
3862  if (NewG->isZero()) {
3863  if (IsScaledReg) {
3864  F.Scale = 0;
3865  F.ScaledReg = nullptr;
3866  } else
3867  F.deleteBaseReg(F.BaseRegs[Idx]);
3868  F.canonicalize(*L);
3869  } else if (IsScaledReg)
3870  F.ScaledReg = NewG;
3871  else
3872  F.BaseRegs[Idx] = NewG;
3873 
3874  (void)InsertFormula(LU, LUIdx, F);
3875  }
3876  };
3877 
3878  const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3879 
3880  // With constant offsets and constant steps, we can generate pre-inc
3881  // accesses by having the offset equal the step. So, for access #0 with a
3882  // step of 8, we generate a G - 8 base which would require the first access
3883  // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
3884  // for itself and hopefully becomes the base for other accesses. This means
3885  // means that a single pre-indexed access can be generated to become the new
3886  // base pointer for each iteration of the loop, resulting in no extra add/sub
3887  // instructions for pointer updating.
3888  if (AMK == TTI::AMK_PreIndexed && LU.Kind == LSRUse::Address) {
3889  if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
3890  if (auto *StepRec =
3891  dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
3892  const APInt &StepInt = StepRec->getAPInt();
3893  int64_t Step = StepInt.isNegative() ?
3894  StepInt.getSExtValue() : StepInt.getZExtValue();
3895 
3896  for (int64_t Offset : Worklist) {
3897  Offset -= Step;
3898  GenerateOffset(G, Offset);
3899  }
3900  }
3901  }
3902  }
3903  for (int64_t Offset : Worklist)
3904  GenerateOffset(G, Offset);
3905 
3906  int64_t Imm = ExtractImmediate(G, SE);
3907  if (G->isZero() || Imm == 0)
3908  return;
3909  Formula F = Base;
3910  F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
3911  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
3912  return;
3913  if (IsScaledReg) {
3914  F.ScaledReg = G;
3915  } else {
3916  F.BaseRegs[Idx] = G;
3917  // We may generate non canonical Formula if G is a recurrent expr reg
3918  // related with current loop while F.ScaledReg is not.
3919  F.canonicalize(*L);
3920  }
3921  (void)InsertFormula(LU, LUIdx, F);
3922 }
3923 
3924 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
3925 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
3926  Formula Base) {
3927  // TODO: For now, just add the min and max offset, because it usually isn't
3928  // worthwhile looking at everything inbetween.
3929  SmallVector<int64_t, 2> Worklist;
3930  Worklist.push_back(LU.MinOffset);
3931  if (LU.MaxOffset != LU.MinOffset)
3932  Worklist.push_back(LU.MaxOffset);
3933 
3934  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
3935  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
3936  if (Base.Scale == 1)
3937  GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
3938  /* IsScaledReg */ true);
3939 }
3940 
3941 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
3942 /// == y -> x*c == y*c.
3943 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
3944  Formula Base) {
3945  if (LU.Kind != LSRUse::ICmpZero) return;
3946 
3947  // Determine the integer type for the base formula.
3948  Type *IntTy = Base.getType();
3949  if (!IntTy) return;
3950  if (SE.getTypeSizeInBits(IntTy) > 64) return;
3951 
3952  // Don't do this if there is more than one offset.
3953  if (LU.MinOffset != LU.MaxOffset) return;
3954 
3955  // Check if transformation is valid. It is illegal to multiply pointer.
3956  if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
3957  return;
3958  for (const SCEV *BaseReg : Base.BaseRegs)
3959  if (BaseReg->getType()->isPointerTy())
3960  return;
3961  assert(!Base.BaseGV && "ICmpZero use is not legal!");
3962 
3963  // Check each interesting stride.
3964  for (int64_t Factor : Factors) {
3965  // Check that Factor can be represented by IntTy
3966  if (!ConstantInt::isValueValidForType(IntTy, Factor))
3967  continue;
3968  // Check that the multiplication doesn't overflow.
3969  if (Base.BaseOffset == std::numeric_limits<int64_t>::min() && Factor == -1)
3970  continue;
3971  int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
3972  assert(Factor != 0 && "Zero factor not expected!");
3973  if (NewBaseOffset / Factor != Base.BaseOffset)
3974  continue;
3975  // If the offset will be truncated at this use, check that it is in bounds.
3976  if (!IntTy->isPointerTy() &&
3977  !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
3978  continue;
3979 
3980  // Check that multiplying with the use offset doesn't overflow.
3981  int64_t Offset = LU.MinOffset;
3982  if (Offset == std::numeric_limits<int64_t>::min() && Factor == -1)
3983  continue;
3984  Offset = (uint64_t)Offset * Factor;
3985  if (Offset / Factor != LU.MinOffset)
3986  continue;
3987  // If the offset will be truncated at this use, check that it is in bounds.
3988  if (!IntTy->isPointerTy() &&
3989  !ConstantInt::isValueValidForType(IntTy, Offset))
3990  continue;
3991 
3992  Formula F = Base;
3993  F.BaseOffset = NewBaseOffset;
3994 
3995  // Check that this scale is legal.
3996  if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
3997  continue;
3998 
3999  // Compensate for the use having MinOffset built into it.
4000  F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
4001 
4002  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4003 
4004  // Check that multiplying with each base register doesn't overflow.
4005  for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
4006  F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
4007  if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
4008  goto next;
4009  }
4010 
4011  // Check that multiplying with the scaled register doesn't overflow.
4012  if (F.ScaledReg) {
4013  F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
4014  if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
4015  continue;
4016  }
4017 
4018  // Check that multiplying with the unfolded offset doesn't overflow.
4019  if (F.UnfoldedOffset != 0) {
4020  if (F.UnfoldedOffset == std::numeric_limits<int64_t>::min() &&
4021  Factor == -1)
4022  continue;
4023  F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
4024  if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
4025  continue;
4026  // If the offset will be truncated, check that it is in bounds.
4027  if (!IntTy->isPointerTy() &&
4028  !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
4029  continue;
4030  }
4031 
4032  // If we make it here and it's legal, add it.
4033  (void)InsertFormula(LU, LUIdx, F);
4034  next:;
4035  }
4036 }
4037 
4038 /// Generate stride factor reuse formulae by making use of scaled-offset address
4039 /// modes, for example.
4040 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
4041  // Determine the integer type for the base formula.
4042  Type *IntTy = Base.getType();
4043  if (!IntTy) return;
4044 
4045  // If this Formula already has a scaled register, we can't add another one.
4046  // Try to unscale the formula to generate a better scale.
4047  if (Base.Scale != 0 && !Base.unscale())
4048  return;
4049 
4050  assert(Base.Scale == 0 && "unscale did not did its job!");
4051 
4052  // Check each interesting stride.
4053  for (int64_t Factor : Factors) {
4054  Base.Scale = Factor;
4055  Base.HasBaseReg = Base.BaseRegs.size() > 1;
4056  // Check whether this scale is going to be legal.
4057  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4058  Base)) {
4059  // As a special-case, handle special out-of-loop Basic users specially.
4060  // TODO: Reconsider this special case.
4061  if (LU.Kind == LSRUse::Basic &&
4062  isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
4063  LU.AccessTy, Base) &&
4064  LU.AllFixupsOutsideLoop)
4065  LU.Kind = LSRUse::Special;
4066  else
4067  continue;
4068  }
4069  // For an ICmpZero, negating a solitary base register won't lead to
4070  // new solutions.
4071  if (LU.Kind == LSRUse::ICmpZero &&
4072  !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
4073  continue;
4074  // For each addrec base reg, if its loop is current loop, apply the scale.
4075  for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
4076  const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
4077  if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
4078  const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4079  if (FactorS->isZero())
4080  continue;
4081  // Divide out the factor, ignoring high bits, since we'll be
4082  // scaling the value back up in the end.
4083  if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true))
4084  if (!Quotient->isZero()) {
4085  // TODO: This could be optimized to avoid all the copying.
4086  Formula F = Base;
4087  F.ScaledReg = Quotient;
4088  F.deleteBaseReg(F.BaseRegs[i]);
4089  // The canonical representation of 1*reg is reg, which is already in
4090  // Base. In that case, do not try to insert the formula, it will be
4091  // rejected anyway.
4092  if (F.Scale == 1 && (F.BaseRegs.empty() ||
4093  (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
4094  continue;
4095  // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
4096  // non canonical Formula with ScaledReg's loop not being L.
4097  if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4098  F.canonicalize(*L);
4099  (void)InsertFormula(LU, LUIdx, F);
4100  }
4101  }
4102  }
4103  }
4104 }
4105 
4106 /// Generate reuse formulae from different IV types.
4107 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4108  // Don't bother truncating symbolic values.
4109  if (Base.BaseGV) return;
4110 
4111  // Determine the integer type for the base formula.
4112  Type *DstTy = Base.getType();
4113  if (!DstTy) return;
4114  if (DstTy->isPointerTy())
4115  return;
4116 
4117  // It is invalid to extend a pointer type so exit early if ScaledReg or
4118  // any of the BaseRegs are pointers.
4119  if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
4120  return;
4121  if (any_of(Base.BaseRegs,
4122  [](const SCEV *S) { return S->getType()->isPointerTy(); }))
4123  return;
4124 
4125  for (Type *SrcTy : Types) {
4126  if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4127  Formula F = Base;
4128 
4129  // Sometimes SCEV is able to prove zero during ext transform. It may
4130  // happen if SCEV did not do all possible transforms while creating the
4131  // initial node (maybe due to depth limitations), but it can do them while
4132  // taking ext.
4133  if (F.ScaledReg) {
4134  const SCEV *NewScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy);
4135  if (NewScaledReg->isZero())
4136  continue;
4137  F.ScaledReg = NewScaledReg;
4138  }
4139  bool HasZeroBaseReg = false;
4140  for (const SCEV *&BaseReg : F.BaseRegs) {
4141  const SCEV *NewBaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy);
4142  if (NewBaseReg->isZero()) {
4143  HasZeroBaseReg = true;
4144  break;
4145  }
4146  BaseReg = NewBaseReg;
4147  }
4148  if (HasZeroBaseReg)
4149  continue;
4150 
4151  // TODO: This assumes we've done basic processing on all uses and
4152  // have an idea what the register usage is.
4153  if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4154  continue;
4155 
4156  F.canonicalize(*L);
4157  (void)InsertFormula(LU, LUIdx, F);
4158  }
4159  }
4160 }
4161 
4162 namespace {
4163 
4164 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4165 /// modifications so that the search phase doesn't have to worry about the data
4166 /// structures moving underneath it.
4167 struct WorkItem {
4168  size_t LUIdx;
4169  int64_t Imm;
4170  const SCEV *OrigReg;
4171 
4172  WorkItem(size_t LI, int64_t I, const SCEV *R)
4173  : LUIdx(LI), Imm(I), OrigReg(R) {}
4174 
4175  void print(raw_ostream &OS) const;
4176  void dump() const;
4177 };
4178 
4179 } // end anonymous namespace
4180 
4181 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
4182 void WorkItem::print(raw_ostream &OS) const {
4183  OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4184  << " , add offset " << Imm;
4185 }
4186 
4187 LLVM_DUMP_METHOD void WorkItem::dump() const {
4188  print(errs()); errs() << '\n';
4189 }
4190 #endif
4191 
4192 /// Look for registers which are a constant distance apart and try to form reuse
4193 /// opportunities between them.
4194 void LSRInstance::GenerateCrossUseConstantOffsets() {
4195  // Group the registers by their value without any added constant offset.
4196  using ImmMapTy = std::map<int64_t, const SCEV *>;
4197 
4199  DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4201  for (const SCEV *Use : RegUses) {
4202  const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4203  int64_t Imm = ExtractImmediate(Reg, SE);
4204  auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4205  if (Pair.second)
4206  Sequence.push_back(Reg);
4207  Pair.first->second.insert(std::make_pair(Imm, Use));
4208  UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4209  }
4210 
4211  // Now examine each set of registers with the same base value. Build up
4212  // a list of work to do and do the work in a separate step so that we're
4213  // not adding formulae and register counts while we're searching.
4214  SmallVector<WorkItem, 32> WorkItems;
4215  SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
4216  for (const SCEV *Reg : Sequence) {
4217  const ImmMapTy &Imms = Map.find(Reg)->second;
4218 
4219  // It's not worthwhile looking for reuse if there's only one offset.
4220  if (Imms.size() == 1)
4221  continue;
4222 
4223  LLVM_DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
4224  for (const auto &Entry
4225  : Imms) dbgs()
4226  << ' ' << Entry.first;
4227  dbgs() << '\n');
4228 
4229  // Examine each offset.
4230  for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4231  J != JE; ++J) {
4232  const SCEV *OrigReg = J->second;
4233 
4234  int64_t JImm = J->first;
4235  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4236 
4237  if (!isa<SCEVConstant>(OrigReg) &&
4238  UsedByIndicesMap[Reg].count() == 1) {
4239  LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg
4240  << '\n');
4241  continue;
4242  }
4243 
4244  // Conservatively examine offsets between this orig reg a few selected
4245  // other orig regs.
4246  int64_t First = Imms.begin()->first;
4247  int64_t Last = std::prev(Imms.end())->first;
4248  // Compute (First + Last) / 2 without overflow using the fact that
4249  // First + Last = 2 * (First + Last) + (First ^ Last).
4250  int64_t Avg = (First & Last) + ((First ^ Last) >> 1);
4251  // If the result is negative and First is odd and Last even (or vice versa),
4252  // we rounded towards -inf. Add 1 in that case, to round towards 0.
4253  Avg = Avg + ((First ^ Last) & ((uint64_t)Avg >> 63));
4254  ImmMapTy::const_iterator OtherImms[] = {
4255  Imms.begin(), std::prev(Imms.end()),
4256  Imms.lower_bound(Avg)};
4257  for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
4258  ImmMapTy::const_iterator M = OtherImms[i];
4259  if (M == J || M == JE) continue;
4260 
4261  // Compute the difference between the two.
4262  int64_t Imm = (uint64_t)JImm - M->first;
4263  for (unsigned LUIdx : UsedByIndices.set_bits())
4264  // Make a memo of this use, offset, and register tuple.
4265  if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4266  WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4267  }
4268  }
4269  }
4270 
4271  Map.clear();
4272  Sequence.clear();
4273  UsedByIndicesMap.clear();
4274  UniqueItems.clear();
4275 
4276  // Now iterate through the worklist and add new formulae.
4277  for (const WorkItem &WI : WorkItems) {
4278  size_t LUIdx = WI.LUIdx;
4279  LSRUse &LU = Uses[LUIdx];
4280  int64_t Imm = WI.Imm;
4281  const SCEV *OrigReg = WI.OrigReg;
4282 
4283  Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4284  const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
4285  unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4286 
4287  // TODO: Use a more targeted data structure.
4288  for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4289  Formula F = LU.Formulae[L];
4290  // FIXME: The code for the scaled and unscaled registers looks
4291  // very similar but slightly different. Investigate if they
4292  // could be merged. That way, we would not have to unscale the
4293  // Formula.
4294  F.unscale();
4295  // Use the immediate in the scaled register.
4296  if (F.ScaledReg == OrigReg) {
4297  int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
4298  // Don't create 50 + reg(-50).
4299  if (F.referencesReg(SE.getSCEV(
4300  ConstantInt::get(IntTy, -(uint64_t)Offset))))
4301  continue;
4302  Formula NewF = F;
4303  NewF.BaseOffset = Offset;
4304  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4305  NewF))
4306  continue;
4307  NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4308 
4309  // If the new scale is a constant in a register, and adding the constant
4310  // value to the immediate would produce a value closer to zero than the
4311  // immediate itself, then the formula isn't worthwhile.
4312  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
4313  if (C->getValue()->isNegative() != (NewF.BaseOffset < 0) &&
4314  (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4315  .ule(std::abs(NewF.BaseOffset)))
4316  continue;
4317 
4318  // OK, looks good.
4319  NewF.canonicalize(*this->L);
4320  (void)InsertFormula(LU, LUIdx, NewF);
4321  } else {
4322  // Use the immediate in a base register.
4323  for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4324  const SCEV *BaseReg = F.BaseRegs[N];
4325  if (BaseReg != OrigReg)
4326  continue;
4327  Formula NewF = F;
4328  NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
4329  if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4330  LU.Kind, LU.AccessTy, NewF)) {
4331  if (AMK == TTI::AMK_PostIndexed &&
4332  mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4333  continue;
4334  if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
4335  continue;
4336  NewF = F;
4337  NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
4338  }
4339  NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4340 
4341  // If the new formula has a constant in a register, and adding the
4342  // constant value to the immediate would produce a value closer to
4343  // zero than the immediate itself, then the formula isn't worthwhile.
4344  for (const SCEV *NewReg : NewF.BaseRegs)
4345  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg))
4346  if ((C->getAPInt() + NewF.BaseOffset)
4347  .abs()
4348  .slt(std::abs(NewF.BaseOffset)) &&
4349  (C->getAPInt() + NewF.BaseOffset).countTrailingZeros() >=
4350  countTrailingZeros<uint64_t>(NewF.BaseOffset))
4351  goto skip_formula;
4352 
4353  // Ok, looks good.
4354  NewF.canonicalize(*this->L);
4355  (void)InsertFormula(LU, LUIdx, NewF);
4356  break;
4357  skip_formula:;
4358  }
4359  }
4360  }
4361  }
4362 }
4363 
4364 /// Generate formulae for each use.
4365 void
4366 LSRInstance::GenerateAllReuseFormulae() {
4367  // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4368  // queries are more precise.
4369  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4370  LSRUse &LU = Uses[LUIdx];
4371  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4372  GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4373  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4374  GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4375  }
4376  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4377  LSRUse &LU = Uses[LUIdx];
4378  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4379  GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4380  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4381  GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4382  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4383  GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4384  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4385  GenerateScales(LU, LUIdx, LU.Formulae[i]);
4386  }
4387  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4388  LSRUse &LU = Uses[LUIdx];
4389  for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4390  GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4391  }
4392 
4393  GenerateCrossUseConstantOffsets();
4394 
4395  LLVM_DEBUG(dbgs() << "\n"
4396  "After generating reuse formulae:\n";
4397  print_uses(dbgs()));
4398 }
4399 
4400 /// If there are multiple formulae with the same set of registers used
4401 /// by other uses, pick the best one and delete the others.
4402 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4403  DenseSet<const SCEV *> VisitedRegs;
4406 #ifndef NDEBUG
4407  bool ChangedFormulae = false;
4408 #endif
4409 
4410  // Collect the best formula for each unique set of shared registers. This
4411  // is reset for each use.
4412  using BestFormulaeTy =
4413  DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4414 
4415  BestFormulaeTy BestFormulae;
4416 
4417  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4418  LSRUse &LU = Uses[LUIdx];
4419  LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
4420  dbgs() << '\n');
4421 
4422  bool Any = false;
4423  for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4424  FIdx != NumForms; ++FIdx) {
4425  Formula &F = LU.Formulae[FIdx];
4426 
4427  // Some formulas are instant losers. For example, they may depend on
4428  // nonexistent AddRecs from other loops. These need to be filtered
4429  // immediately, otherwise heuristics could choose them over others leading
4430  // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4431  // avoids the need to recompute this information across formulae using the
4432  // same bad AddRec. Passing LoserRegs is also essential unless we remove
4433  // the corresponding bad register from the Regs set.
4434  Cost CostF(L, SE, TTI, AMK);
4435  Regs.clear();
4436  CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4437  if (CostF.isLoser()) {
4438  // During initial formula generation, undesirable formulae are generated
4439  // by uses within other loops that have some non-trivial address mode or
4440  // use the postinc form of the IV. LSR needs to provide these formulae
4441  // as the basis of rediscovering the desired formula that uses an AddRec
4442  // corresponding to the existing phi. Once all formulae have been
4443  // generated, these initial losers may be pruned.
4444  LLVM_DEBUG(dbgs() << " Filtering loser "; F.print(dbgs());
4445  dbgs() << "\n");
4446  }
4447  else {
4449  for (const SCEV *Reg : F.BaseRegs) {
4450  if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4451  Key.push_back(Reg);
4452  }
4453  if (F.ScaledReg &&
4454  RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4455  Key.push_back(F.ScaledReg);
4456  // Unstable sort by host order ok, because this is only used for
4457  // uniquifying.
4458  llvm::sort(Key);
4459 
4460  std::pair<BestFormulaeTy::const_iterator, bool> P =
4461  BestFormulae.insert(std::make_pair(Key, FIdx));
4462  if (P.second)
4463  continue;
4464 
4465  Formula &Best = LU.Formulae[P.first->second];
4466 
4467  Cost CostBest(L, SE, TTI, AMK);
4468  Regs.clear();
4469  CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4470  if (CostF.isLess(CostBest))
4471  std::swap(F, Best);
4472  LLVM_DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4473  dbgs() << "\n"
4474  " in favor of formula ";
4475  Best.print(dbgs()); dbgs() << '\n');
4476  }
4477 #ifndef NDEBUG
4478  ChangedFormulae = true;
4479 #endif
4480  LU.DeleteFormula(F);
4481  --FIdx;
4482  --NumForms;
4483  Any = true;
4484  }
4485 
4486  // Now that we've filtered out some formulae, recompute the Regs set.
4487  if (Any)
4488  LU.RecomputeRegs(LUIdx, RegUses);
4489 
4490  // Reset this to prepare for the next use.
4491  BestFormulae.clear();
4492  }
4493 
4494  LLVM_DEBUG(if (ChangedFormulae) {
4495  dbgs() << "\n"
4496  "After filtering out undesirable candidates:\n";
4497  print_uses(dbgs());
4498  });
4499 }
4500 
4501 /// Estimate the worst-case number of solutions the solver might have to
4502 /// consider. It almost never considers this many solutions because it prune the
4503 /// search space, but the pruning isn't always sufficient.
4504 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4505  size_t Power = 1;
4506  for (const LSRUse &LU : Uses) {
4507  size_t FSize = LU.Formulae.size();
4508  if (FSize >= ComplexityLimit) {
4509  Power = ComplexityLimit;
4510  break;
4511  }
4512  Power *= FSize;
4513  if (Power >= ComplexityLimit)
4514  break;
4515  }
4516  return Power;
4517 }
4518 
4519 /// When one formula uses a superset of the registers of another formula, it
4520 /// won't help reduce register pressure (though it may not necessarily hurt
4521 /// register pressure); remove it to simplify the system.
4522 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4523  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4524  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4525 
4526  LLVM_DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4527  "which use a superset of registers used by other "
4528  "formulae.\n");
4529 
4530  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4531  LSRUse &LU = Uses[LUIdx];
4532  bool Any = false;
4533  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4534  Formula &F = LU.Formulae[i];
4535  // Look for a formula with a constant or GV in a register. If the use
4536  // also has a formula with that same value in an immediate field,
4537  // delete the one that uses a register.
4539  I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4540  if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4541  Formula NewF = F;
4542  //FIXME: Formulas should store bitwidth to do wrapping properly.
4543  // See PR41034.
4544  NewF.BaseOffset += (uint64_t)C->getValue()->getSExtValue();
4545  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4546  (I - F.BaseRegs.begin()));
4547  if (LU.HasFormulaWithSameRegs(NewF)) {
4548  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4549  dbgs() << '\n');
4550  LU.DeleteFormula(F);
4551  --i;
4552  --e;
4553  Any = true;
4554  break;
4555  }
4556  } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4557  if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4558  if (!F.BaseGV) {
4559  Formula NewF = F;
4560  NewF.BaseGV = GV;
4561  NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4562  (I - F.BaseRegs.begin()));
4563  if (LU.HasFormulaWithSameRegs(NewF)) {
4564  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs());
4565  dbgs() << '\n');
4566  LU.DeleteFormula(F);
4567  --i;
4568  --e;
4569  Any = true;
4570  break;
4571  }
4572  }
4573  }
4574  }
4575  }
4576  if (Any)
4577  LU.RecomputeRegs(LUIdx, RegUses);
4578  }
4579 
4580  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4581  }
4582 }
4583 
4584 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4585 /// allocate a single register for them.
4586 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4587  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4588  return;
4589 
4590  LLVM_DEBUG(
4591  dbgs() << "The search space is too complex.\n"
4592  "Narrowing the search space by assuming that uses separated "
4593  "by a constant offset will use the same registers.\n");
4594 
4595  // This is especially useful for unrolled loops.
4596 
4597  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4598  LSRUse &LU = Uses[LUIdx];
4599  for (const Formula &F : LU.Formulae) {
4600  if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1))
4601  continue;
4602 
4603  LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4604  if (!LUThatHas)
4605  continue;
4606 
4607  if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4608  LU.Kind, LU.AccessTy))
4609  continue;
4610 
4611  LLVM_DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4612 
4613  LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4614 
4615  // Transfer the fixups of LU to LUThatHas.
4616  for (LSRFixup &Fixup : LU.Fixups) {
4617  Fixup.Offset += F.BaseOffset;
4618  LUThatHas->pushFixup(Fixup);
4619  LLVM_DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4620  }
4621 
4622  // Delete formulae from the new use which are no longer legal.
4623  bool Any = false;
4624  for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4625  Formula &F = LUThatHas->Formulae[i];
4626  if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4627  LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4628  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4629  LUThatHas->DeleteFormula(F);
4630  --i;
4631  --e;
4632  Any = true;
4633  }
4634  }
4635 
4636  if (Any)
4637  LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4638 
4639  // Delete the old use.
4640  DeleteUse(LU, LUIdx);
4641  --LUIdx;
4642  --NumUses;
4643  break;
4644  }
4645  }
4646 
4647  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4648 }
4649 
4650 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
4651 /// we've done more filtering, as it may be able to find more formulae to
4652 /// eliminate.
4653 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
4654  if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4655  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4656 
4657  LLVM_DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
4658  "undesirable dedicated registers.\n");
4659 
4660  FilterOutUndesirableDedicatedRegisters();
4661 
4662  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4663  }
4664 }
4665 
4666 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
4667 /// Pick the best one and delete the others.
4668 /// This narrowing heuristic is to keep as many formulae with different
4669 /// Scale and ScaledReg pair as possible while narrowing the search space.
4670 /// The benefit is that it is more likely to find out a better solution
4671 /// from a formulae set with more Scale and ScaledReg variations than
4672 /// a formulae set with the same Scale and ScaledReg. The picking winner
4673 /// reg heuristic will often keep the formulae with the same Scale and
4674 /// ScaledReg and filter others, and we want to avoid that if possible.
4675 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
4676  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4677  return;
4678 
4679  LLVM_DEBUG(
4680  dbgs() << "The search space is too complex.\n"
4681  "Narrowing the search space by choosing the best Formula "
4682  "from the Formulae with the same Scale and ScaledReg.\n");
4683 
4684  // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
4685  using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
4686 
4687  BestFormulaeTy BestFormulae;
4688 #ifndef NDEBUG
4689  bool ChangedFormulae = false;
4690 #endif
4691  DenseSet<const SCEV *> VisitedRegs;
4693 
4694  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4695  LSRUse &LU = Uses[LUIdx];
4696  LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
4697  dbgs() << '\n');
4698 
4699  // Return true if Formula FA is better than Formula FB.
4700  auto IsBetterThan = [&](Formula &FA, Formula &FB) {
4701  // First we will try to choose the Formula with fewer new registers.
4702  // For a register used by current Formula, the more the register is
4703  // shared among LSRUses, the less we increase the register number
4704  // counter of the formula.
4705  size_t FARegNum = 0;
4706  for (const SCEV *Reg : FA.BaseRegs) {
4707  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4708  FARegNum += (NumUses - UsedByIndices.count() + 1);
4709  }
4710  size_t FBRegNum = 0;
4711  for (const SCEV *Reg : FB.BaseRegs) {
4712  const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
4713  FBRegNum += (NumUses - UsedByIndices.count() + 1);
4714  }
4715  if (FARegNum != FBRegNum)
4716  return FARegNum < FBRegNum;
4717 
4718  // If the new register numbers are the same, choose the Formula with
4719  // less Cost.
4720  Cost CostFA(L, SE, TTI, AMK);
4721  Cost CostFB(L, SE, TTI, AMK);
4722  Regs.clear();
4723  CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
4724  Regs.clear();
4725  CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
4726  return CostFA.isLess(CostFB);
4727  };
4728 
4729  bool Any = false;
4730  for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4731  ++FIdx) {
4732  Formula &F = LU.Formulae[FIdx];
4733  if (!F.ScaledReg)
4734  continue;
4735  auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
4736  if (P.second)
4737  continue;
4738 
4739  Formula &Best = LU.Formulae[P.first->second];
4740  if (IsBetterThan(F, Best))
4741  std::swap(F, Best);
4742  LLVM_DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4743  dbgs() << "\n"
4744  " in favor of formula ";
4745  Best.print(dbgs()); dbgs() << '\n');
4746 #ifndef NDEBUG
4747  ChangedFormulae = true;
4748 #endif
4749  LU.DeleteFormula(F);
4750  --FIdx;
4751  --NumForms;
4752  Any = true;
4753  }
4754  if (Any)
4755  LU.RecomputeRegs(LUIdx, RegUses);
4756 
4757  // Reset this to prepare for the next use.
4758  BestFormulae.clear();
4759  }
4760 
4761  LLVM_DEBUG(if (ChangedFormulae) {
4762  dbgs() << "\n"
4763  "After filtering out undesirable candidates:\n";
4764  print_uses(dbgs());
4765  });
4766 }
4767 
4768 /// If we are over the complexity limit, filter out any post-inc prefering
4769 /// variables to only post-inc values.
4770 void LSRInstance::NarrowSearchSpaceByFilterPostInc() {
4771  if (AMK != TTI::AMK_PostIndexed)
4772  return;
4773  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4774  return;
4775 
4776  LLVM_DEBUG(dbgs() << "The search space is too complex.\n"
4777  "Narrowing the search space by choosing the lowest "
4778  "register Formula for PostInc Uses.\n");
4779 
4780  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4781  LSRUse &LU = Uses[LUIdx];
4782 
4783  if (LU.Kind != LSRUse::Address)
4784  continue;
4785  if (!TTI.isIndexedLoadLegal(TTI.MIM_PostInc, LU.AccessTy.getType()) &&
4786  !TTI.isIndexedStoreLegal(TTI.MIM_PostInc, LU.AccessTy.getType()))
4787  continue;
4788 
4789  size_t MinRegs = std::numeric_limits<size_t>::max();
4790  for (const Formula &F : LU.Formulae)
4791  MinRegs = std::min(F.getNumRegs(), MinRegs);
4792 
4793  bool Any = false;
4794  for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
4795  ++FIdx) {
4796  Formula &F = LU.Formulae[FIdx];
4797  if (F.getNumRegs() > MinRegs) {
4798  LLVM_DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs());
4799  dbgs() << "\n");
4800  LU.DeleteFormula(F);
4801  --FIdx;
4802  --NumForms;
4803  Any = true;
4804  }
4805  }
4806  if (Any)
4807  LU.RecomputeRegs(LUIdx, RegUses);
4808 
4809  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4810  break;
4811  }
4812 
4813  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4814 }
4815 
4816 /// The function delete formulas with high registers number expectation.
4817 /// Assuming we don't know the value of each formula (already delete
4818 /// all inefficient), generate probability of not selecting for each
4819 /// register.
4820 /// For example,
4821 /// Use1:
4822 /// reg(a) + reg({0,+,1})
4823 /// reg(a) + reg({-1,+,1}) + 1
4824 /// reg({a,+,1})
4825 /// Use2:
4826 /// reg(b) + reg({0,+,1})
4827 /// reg(b) + reg({-1,+,1}) + 1
4828 /// reg({b,+,1})
4829 /// Use3:
4830 /// reg(c) + reg(b) + reg({0,+,1})
4831 /// reg(c) + reg({b,+,1})
4832 ///
4833 /// Probability of not selecting
4834 /// Use1 Use2 Use3
4835 /// reg(a) (1/3) * 1 * 1
4836 /// reg(b) 1 * (1/3) * (1/2)
4837 /// reg({0,+,1}) (2/3) * (2/3) * (1/2)
4838 /// reg({-1,+,1}) (2/3) * (2/3) * 1
4839 /// reg({a,+,1}) (2/3) * 1 * 1
4840 /// reg({b,+,1}) 1 * (2/3) * (2/3)
4841 /// reg(c) 1 * 1 * 0
4842 ///
4843 /// Now count registers number mathematical expectation for each formula:
4844 /// Note that for each use we exclude probability if not selecting for the use.
4845 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
4846 /// probabilty 1/3 of not selecting for Use1).
4847 /// Use1:
4848 /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted
4849 /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted
4850 /// reg({a,+,1}) 1
4851 /// Use2:
4852 /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted
4853 /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted
4854 /// reg({b,+,1}) 2/3
4855 /// Use3:
4856 /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
4857 /// reg(c) + reg({b,+,1}) 1 + 2/3
4858 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
4859  if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4860  return;
4861  // Ok, we have too many of formulae on our hands to conveniently handle.
4862  // Use a rough heuristic to thin out the list.
4863 
4864  // Set of Regs wich will be 100% used in final solution.
4865  // Used in each formula of a solution (in example above this is reg(c)).
4866  // We can skip them in calculations.
4868  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4869 
4870  // Map each register to probability of not selecting
4872  for (const SCEV *Reg : RegUses) {
4873  if (UniqRegs.count(Reg))
4874  continue;
4875  float PNotSel = 1;
4876  for (const LSRUse &LU : Uses) {
4877  if (!LU.Regs.count(Reg))
4878  continue;
4879  float P = LU.getNotSelectedProbability(Reg);
4880  if (P != 0.0)
4881  PNotSel *= P;
4882  else
4883  UniqRegs.insert(Reg);
4884  }
4885  RegNumMap.insert(std::make_pair(Reg, PNotSel));
4886  }
4887 
4888  LLVM_DEBUG(
4889  dbgs() << "Narrowing the search space by deleting costly formulas\n");
4890 
4891  // Delete formulas where registers number expectation is high.
4892  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4893  LSRUse &LU = Uses[LUIdx];
4894  // If nothing to delete - continue.
4895  if (LU.Formulae.size() < 2)
4896  continue;
4897  // This is temporary solution to test performance. Float should be
4898  // replaced with round independent type (based on integers) to avoid
4899  // different results for different target builds.
4900  float FMinRegNum = LU.Formulae[0].getNumRegs();
4901  float FMinARegNum = LU.Formulae[0].getNumRegs();
4902  size_t MinIdx = 0;
4903  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4904  Formula &F = LU.Formulae[i];
4905  float FRegNum = 0;
4906  float FARegNum = 0;
4907  for (const SCEV *BaseReg : F.BaseRegs) {
4908  if (UniqRegs.count(BaseReg))
4909  continue;
4910  FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4911  if (isa<SCEVAddRecExpr>(BaseReg))
4912  FARegNum +=
4913  RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
4914  }
4915  if (const SCEV *ScaledReg = F.ScaledReg) {
4916  if (!UniqRegs.count(ScaledReg)) {
4917  FRegNum +=
4918  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4919  if (isa<SCEVAddRecExpr>(ScaledReg))
4920  FARegNum +=
4921  RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
4922  }
4923  }
4924  if (FMinRegNum > FRegNum ||
4925  (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
4926  FMinRegNum = FRegNum;
4927  FMinARegNum = FARegNum;
4928  MinIdx = i;
4929  }
4930  }
4931  LLVM_DEBUG(dbgs() << " The formula "; LU.Formulae[MinIdx].print(dbgs());
4932  dbgs() << " with min reg num " << FMinRegNum << '\n');
4933  if (MinIdx != 0)
4934  std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
4935  while (LU.Formulae.size() != 1) {
4936  LLVM_DEBUG(dbgs() << " Deleting "; LU.Formulae.back().print(dbgs());
4937  dbgs() << '\n');
4938  LU.Formulae.pop_back();
4939  }
4940  LU.RecomputeRegs(LUIdx, RegUses);
4941  assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
4942  Formula &F = LU.Formulae[0];
4943  LLVM_DEBUG(dbgs() << " Leaving only "; F.print(dbgs()); dbgs() << '\n');
4944  // When we choose the formula, the regs become unique.
4945  UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
4946  if (F.ScaledReg)
4947  UniqRegs.insert(F.ScaledReg);
4948  }
4949  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4950 }
4951 
4952 /// Pick a register which seems likely to be profitable, and then in any use
4953 /// which has any reference to that register, delete all formulae which do not
4954 /// reference that register.
4955 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
4956  // With all other options exhausted, loop until the system is simple
4957  // enough to handle.
4959  while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4960  // Ok, we have too many of formulae on our hands to conveniently handle.
4961  // Use a rough heuristic to thin out the list.
4962  LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4963 
4964  // Pick the register which is used by the most LSRUses, which is likely
4965  // to be a good reuse register candidate.
4966  const SCEV *Best = nullptr;
4967  unsigned BestNum = 0;
4968  for (const SCEV *Reg : RegUses) {
4969  if (Taken.count(Reg))
4970  continue;
4971  if (!Best) {
4972  Best = Reg;
4973  BestNum = RegUses.getUsedByIndices(Reg).count();
4974  } else {
4975  unsigned Count = RegUses.getUsedByIndices(Reg).count();
4976  if (Count > BestNum) {
4977  Best = Reg;
4978  BestNum = Count;
4979  }
4980  }
4981  }
4982  assert(Best && "Failed to find best LSRUse candidate");
4983 
4984  LLVM_DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
4985  << " will yield profitable reuse.\n");
4986  Taken.insert(Best);
4987 
4988  // In any use with formulae which references this register, delete formulae
4989  // which don't reference it.
4990  for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4991  LSRUse &LU = Uses[LUIdx];
4992  if (!LU.Regs.count(Best)) continue;
4993 
4994  bool Any = false;
4995  for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4996  Formula &F = LU.Formulae[i];
4997  if (!F.referencesReg(Best)) {
4998  LLVM_DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
4999  LU.DeleteFormula(F);
5000  --e;
5001  --i;
5002  Any = true;
5003  assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
5004  continue;
5005  }
5006  }
5007 
5008  if (Any)
5009  LU.RecomputeRegs(LUIdx, RegUses);
5010  }
5011 
5012  LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
5013  }
5014 }
5015 
5016 /// If there are an extraordinary number of formulae to choose from, use some
5017 /// rough heuristics to prune down the number of formulae. This keeps the main
5018 /// solver from taking an extraordinary amount of time in some worst-case
5019 /// scenarios.
5020 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
5021  NarrowSearchSpaceByDetectingSupersets();
5022  NarrowSearchSpaceByCollapsingUnrolledCode();
5023  NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
5024  if (FilterSameScaledReg)
5025  NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
5026  NarrowSearchSpaceByFilterPostInc();
5027  if (LSRExpNarrow)
5028  NarrowSearchSpaceByDeletingCostlyFormulas();
5029  else
5030  NarrowSearchSpaceByPickingWinnerRegs();
5031 }
5032 
5033 /// This is the recursive solver.
5034 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
5035  Cost &SolutionCost,
5037  const Cost &CurCost,
5038  const SmallPtrSet<const SCEV *, 16> &CurRegs,
5039  DenseSet<const SCEV *> &VisitedRegs) const {
5040  // Some ideas:
5041  // - prune more:
5042  // - use more aggressive filtering
5043  // - sort the formula so that the most profitable solutions are found first
5044  // - sort the uses too
5045  // - search faster:
5046  // - don't compute a cost, and then compare. compare while computing a cost
5047  // and bail early.
5048  // - track register sets with SmallBitVector
5049 
5050  const LSRUse &LU = Uses[Workspace.size()];
5051 
5052  // If this use references any register that's already a part of the
5053  // in-progress solution, consider it a requirement that a formula must
5054  // reference that register in order to be considered. This prunes out
5055  // unprofitable searching.
5057  for (const SCEV *S : CurRegs)
5058  if (LU.Regs.count(S))
5059  ReqRegs.insert(S);
5060 
5062  Cost NewCost(L, SE, TTI, AMK);
5063  for (const Formula &F : LU.Formulae) {
5064  // Ignore formulae which may not be ideal in terms of register reuse of
5065  // ReqRegs. The formula should use all required registers before
5066  // introducing new ones.
5067  // This can sometimes (notably when trying to favour postinc) lead to
5068  // sub-optimial decisions. There it is best left to the cost modelling to
5069  // get correct.
5070  if (AMK != TTI::AMK_PostIndexed || LU.Kind != LSRUse::Address) {
5071  int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
5072  for (const SCEV *Reg : ReqRegs) {
5073  if ((F.ScaledReg && F.ScaledReg == Reg) ||
5074  is_contained(F.BaseRegs, Reg)) {
5075  --NumReqRegsToFind;
5076  if (NumReqRegsToFind == 0)
5077  break;
5078  }
5079  }
5080  if (NumReqRegsToFind != 0) {
5081  // If none of the formulae satisfied the required registers, then we could
5082  // clear ReqRegs and try again. Currently, we simply give up in this case.
5083  continue;
5084  }
5085  }
5086 
5087  // Evaluate the cost of the current formula. If it's already worse than
5088  // the current best, prune the search at that point.
5089  NewCost = CurCost;
5090  NewRegs = CurRegs;
5091  NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
5092  if (NewCost.isLess(SolutionCost)) {
5093  Workspace.push_back(&F);
5094  if (Workspace.size() != Uses.size()) {
5095  SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
5096  NewRegs, VisitedRegs);
5097  if (F.getNumRegs() == 1 && Workspace.size() == 1)
5098  VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
5099  } else {
5100  LLVM_DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
5101  dbgs() << ".\nRegs:\n";
5102  for (const SCEV *S : NewRegs) dbgs()
5103  << "- " << *S << "\n";
5104  dbgs() << '\n');
5105 
5106  SolutionCost = NewCost;
5107  Solution = Workspace;
5108  }
5109  Workspace.pop_back();
5110  }
5111  }
5112 }
5113 
5114 /// Choose one formula from each use. Return the results in the given Solution
5115 /// vector.
5116 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
5118  Cost SolutionCost(L, SE, TTI, AMK);
5119  SolutionCost.Lose();
5120  Cost CurCost(L, SE, TTI, AMK);
5122  DenseSet<const SCEV *> VisitedRegs;
5123  Workspace.reserve(Uses.size());
5124 
5125  // SolveRecurse does all the work.
5126  SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
5127  CurRegs, VisitedRegs);
5128  if (Solution.empty()) {
5129  LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
5130  return;
5131  }
5132 
5133  // Ok, we've now made all our decisions.
5134  LLVM_DEBUG(dbgs() << "\n"
5135  "The chosen solution requires ";
5136  SolutionCost.print(dbgs()); dbgs() << ":\n";
5137  for (size_t i = 0, e = Uses.size(); i != e; ++i) {
5138  dbgs() << " ";
5139  Uses[i].print(dbgs());
5140  dbgs() << "\n"
5141  " ";
5142  Solution[i]->print(dbgs());
5143  dbgs() << '\n';
5144  });
5145 
5146  assert(Solution.size() == Uses.size() && "Malformed solution!");
5147 }
5148 
5149 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
5150 /// we can go while still being dominated by the input positions. This helps
5151 /// canonicalize the insert position, which encourages sharing.
5153 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
5154  const SmallVectorImpl<Instruction *> &Inputs)
5155  const {
5156  Instruction *Tentative = &*IP;
5157  while (true) {
5158  bool AllDominate = true;
5159  Instruction *BetterPos = nullptr;
5160  // Don't bother attempting to insert before a catchswitch, their basic block
5161  // cannot have other non-PHI instructions.
5162  if (isa<CatchSwitchInst>(Tentative))
5163  return IP;
5164 
5165  for (Instruction *Inst : Inputs) {
5166  if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5167  AllDominate = false;
5168  break;
5169  }
5170  // Attempt to find an insert position in the middle of the block,
5171  // instead of at the end, so that it can be used for other expansions.
5172  if (Tentative->getParent() == Inst->getParent() &&
5173  (!BetterPos || !DT.dominates(Inst, BetterPos)))
5174  BetterPos = &*std::next(BasicBlock::iterator(Inst));
5175  }
5176  if (!AllDominate)
5177  break;
5178  if (BetterPos)
5179  IP = BetterPos->getIterator();
5180  else
5181  IP = Tentative->getIterator();
5182 
5183  const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5184  unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5185 
5186  BasicBlock *IDom;
5187  for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5188  if (!Rung) return IP;
5189  Rung = Rung->getIDom();
5190  if (!Rung) return IP;
5191  IDom = Rung->getBlock();
5192 
5193  // Don't climb into a loop though.
5194  const Loop *IDomLoop = LI.getLoopFor(IDom);
5195  unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5196  if (IDomDepth <= IPLoopDepth &&
5197  (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5198  break;
5199  }
5200 
5201  Tentative = IDom->getTerminator();
5202  }
5203 
5204  return IP;
5205 }
5206 
5207 /// Determine an input position which will be dominated by the operands and
5208 /// which will dominate the result.
5210 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
5211  const LSRFixup &LF,
5212