xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp (revision 36b606ae6aa4b24061096ba18582e0a08ccd5dba)
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 
55 #include "llvm/Transforms/Scalar/LoopStrengthReduce.h"
56 #include "llvm/ADT/APInt.h"
57 #include "llvm/ADT/DenseMap.h"
58 #include "llvm/ADT/DenseSet.h"
59 #include "llvm/ADT/Hashing.h"
60 #include "llvm/ADT/PointerIntPair.h"
61 #include "llvm/ADT/STLExtras.h"
62 #include "llvm/ADT/SetVector.h"
63 #include "llvm/ADT/SmallBitVector.h"
64 #include "llvm/ADT/SmallPtrSet.h"
65 #include "llvm/ADT/SmallSet.h"
66 #include "llvm/ADT/SmallVector.h"
67 #include "llvm/ADT/Statistic.h"
68 #include "llvm/ADT/iterator_range.h"
69 #include "llvm/Analysis/AssumptionCache.h"
70 #include "llvm/Analysis/DomTreeUpdater.h"
71 #include "llvm/Analysis/IVUsers.h"
72 #include "llvm/Analysis/LoopAnalysisManager.h"
73 #include "llvm/Analysis/LoopInfo.h"
74 #include "llvm/Analysis/LoopPass.h"
75 #include "llvm/Analysis/MemorySSA.h"
76 #include "llvm/Analysis/MemorySSAUpdater.h"
77 #include "llvm/Analysis/ScalarEvolution.h"
78 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
79 #include "llvm/Analysis/ScalarEvolutionNormalization.h"
80 #include "llvm/Analysis/TargetLibraryInfo.h"
81 #include "llvm/Analysis/TargetTransformInfo.h"
82 #include "llvm/Analysis/ValueTracking.h"
83 #include "llvm/BinaryFormat/Dwarf.h"
84 #include "llvm/Config/llvm-config.h"
85 #include "llvm/IR/BasicBlock.h"
86 #include "llvm/IR/Constant.h"
87 #include "llvm/IR/Constants.h"
88 #include "llvm/IR/DebugInfoMetadata.h"
89 #include "llvm/IR/DerivedTypes.h"
90 #include "llvm/IR/Dominators.h"
91 #include "llvm/IR/GlobalValue.h"
92 #include "llvm/IR/IRBuilder.h"
93 #include "llvm/IR/InstrTypes.h"
94 #include "llvm/IR/Instruction.h"
95 #include "llvm/IR/Instructions.h"
96 #include "llvm/IR/IntrinsicInst.h"
97 #include "llvm/IR/Module.h"
98 #include "llvm/IR/Operator.h"
99 #include "llvm/IR/PassManager.h"
100 #include "llvm/IR/Type.h"
101 #include "llvm/IR/Use.h"
102 #include "llvm/IR/User.h"
103 #include "llvm/IR/Value.h"
104 #include "llvm/IR/ValueHandle.h"
105 #include "llvm/InitializePasses.h"
106 #include "llvm/Pass.h"
107 #include "llvm/Support/Casting.h"
108 #include "llvm/Support/CommandLine.h"
109 #include "llvm/Support/Compiler.h"
110 #include "llvm/Support/Debug.h"
111 #include "llvm/Support/ErrorHandling.h"
112 #include "llvm/Support/MathExtras.h"
113 #include "llvm/Support/raw_ostream.h"
114 #include "llvm/Transforms/Scalar.h"
115 #include "llvm/Transforms/Utils.h"
116 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
117 #include "llvm/Transforms/Utils/Local.h"
118 #include "llvm/Transforms/Utils/LoopUtils.h"
119 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
120 #include <algorithm>
121 #include <cassert>
122 #include <cstddef>
123 #include <cstdint>
124 #include <iterator>
125 #include <limits>
126 #include <map>
127 #include <numeric>
128 #include <optional>
129 #include <utility>
130 
131 using namespace llvm;
132 
133 #define DEBUG_TYPE "loop-reduce"
134 
135 /// MaxIVUsers is an arbitrary threshold that provides an early opportunity for
136 /// bail out. This threshold is far beyond the number of users that LSR can
137 /// conceivably solve, so it should not affect generated code, but catches the
138 /// worst cases before LSR burns too much compile time and stack space.
139 static const unsigned MaxIVUsers = 200;
140 
141 /// Limit the size of expression that SCEV-based salvaging will attempt to
142 /// translate into a DIExpression.
143 /// Choose a maximum size such that debuginfo is not excessively increased and
144 /// the salvaging is not too expensive for the compiler.
145 static const unsigned MaxSCEVSalvageExpressionSize = 64;
146 
147 // Cleanup congruent phis after LSR phi expansion.
148 static cl::opt<bool> EnablePhiElim(
149   "enable-lsr-phielim", cl::Hidden, cl::init(true),
150   cl::desc("Enable LSR phi elimination"));
151 
152 // The flag adds instruction count to solutions cost comparison.
153 static cl::opt<bool> InsnsCost(
154   "lsr-insns-cost", cl::Hidden, cl::init(true),
155   cl::desc("Add instruction count to a LSR cost model"));
156 
157 // Flag to choose how to narrow complex lsr solution
158 static cl::opt<bool> LSRExpNarrow(
159   "lsr-exp-narrow", cl::Hidden, cl::init(false),
160   cl::desc("Narrow LSR complex solution using"
161            " expectation of registers number"));
162 
163 // Flag to narrow search space by filtering non-optimal formulae with
164 // the same ScaledReg and Scale.
165 static cl::opt<bool> FilterSameScaledReg(
166     "lsr-filter-same-scaled-reg", cl::Hidden, cl::init(true),
167     cl::desc("Narrow LSR search space by filtering non-optimal formulae"
168              " with the same ScaledReg and Scale"));
169 
170 static cl::opt<TTI::AddressingModeKind> PreferredAddresingMode(
171   "lsr-preferred-addressing-mode", cl::Hidden, cl::init(TTI::AMK_None),
172    cl::desc("A flag that overrides the target's preferred addressing mode."),
173    cl::values(clEnumValN(TTI::AMK_None,
174                          "none",
175                          "Don't prefer any addressing mode"),
176               clEnumValN(TTI::AMK_PreIndexed,
177                          "preindexed",
178                          "Prefer pre-indexed addressing mode"),
179               clEnumValN(TTI::AMK_PostIndexed,
180                          "postindexed",
181                          "Prefer post-indexed addressing mode")));
182 
183 static cl::opt<unsigned> ComplexityLimit(
184   "lsr-complexity-limit", cl::Hidden,
185   cl::init(std::numeric_limits<uint16_t>::max()),
186   cl::desc("LSR search space complexity limit"));
187 
188 static cl::opt<unsigned> SetupCostDepthLimit(
189     "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
190     cl::desc("The limit on recursion depth for LSRs setup cost"));
191 
192 static cl::opt<cl::boolOrDefault> AllowTerminatingConditionFoldingAfterLSR(
193     "lsr-term-fold", cl::Hidden,
194     cl::desc("Attempt to replace primary IV with other IV."));
195 
196 static cl::opt<cl::boolOrDefault> AllowDropSolutionIfLessProfitable(
197     "lsr-drop-solution", cl::Hidden,
198     cl::desc("Attempt to drop solution if it is less profitable"));
199 
200 static cl::opt<bool> EnableVScaleImmediates(
201     "lsr-enable-vscale-immediates", cl::Hidden, cl::init(true),
202     cl::desc("Enable analysis of vscale-relative immediates in LSR"));
203 
204 static cl::opt<bool> DropScaledForVScale(
205     "lsr-drop-scaled-reg-for-vscale", cl::Hidden, cl::init(true),
206     cl::desc("Avoid using scaled registers with vscale-relative addressing"));
207 
208 STATISTIC(NumTermFold,
209           "Number of terminating condition fold recognized and performed");
210 
211 #ifndef NDEBUG
212 // Stress test IV chain generation.
213 static cl::opt<bool> StressIVChain(
214   "stress-ivchain", cl::Hidden, cl::init(false),
215   cl::desc("Stress test LSR IV chains"));
216 #else
217 static bool StressIVChain = false;
218 #endif
219 
220 namespace {
221 
222 struct MemAccessTy {
223   /// Used in situations where the accessed memory type is unknown.
224   static const unsigned UnknownAddressSpace =
225       std::numeric_limits<unsigned>::max();
226 
227   Type *MemTy = nullptr;
228   unsigned AddrSpace = UnknownAddressSpace;
229 
230   MemAccessTy() = default;
MemAccessTy__anonc21373340111::MemAccessTy231   MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {}
232 
operator ==__anonc21373340111::MemAccessTy233   bool operator==(MemAccessTy Other) const {
234     return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace;
235   }
236 
operator !=__anonc21373340111::MemAccessTy237   bool operator!=(MemAccessTy Other) const { return !(*this == Other); }
238 
getUnknown__anonc21373340111::MemAccessTy239   static MemAccessTy getUnknown(LLVMContext &Ctx,
240                                 unsigned AS = UnknownAddressSpace) {
241     return MemAccessTy(Type::getVoidTy(Ctx), AS);
242   }
243 
getType__anonc21373340111::MemAccessTy244   Type *getType() { return MemTy; }
245 };
246 
247 /// This class holds data which is used to order reuse candidates.
248 class RegSortData {
249 public:
250   /// This represents the set of LSRUse indices which reference
251   /// a particular register.
252   SmallBitVector UsedByIndices;
253 
254   void print(raw_ostream &OS) const;
255   void dump() const;
256 };
257 
258 // An offset from an address that is either scalable or fixed. Used for
259 // per-target optimizations of addressing modes.
260 class Immediate : public details::FixedOrScalableQuantity<Immediate, int64_t> {
Immediate(ScalarTy MinVal,bool Scalable)261   constexpr Immediate(ScalarTy MinVal, bool Scalable)
262       : FixedOrScalableQuantity(MinVal, Scalable) {}
263 
Immediate(const FixedOrScalableQuantity<Immediate,int64_t> & V)264   constexpr Immediate(const FixedOrScalableQuantity<Immediate, int64_t> &V)
265       : FixedOrScalableQuantity(V) {}
266 
267 public:
268   constexpr Immediate() = delete;
269 
getFixed(ScalarTy MinVal)270   static constexpr Immediate getFixed(ScalarTy MinVal) {
271     return {MinVal, false};
272   }
getScalable(ScalarTy MinVal)273   static constexpr Immediate getScalable(ScalarTy MinVal) {
274     return {MinVal, true};
275   }
get(ScalarTy MinVal,bool Scalable)276   static constexpr Immediate get(ScalarTy MinVal, bool Scalable) {
277     return {MinVal, Scalable};
278   }
getZero()279   static constexpr Immediate getZero() { return {0, false}; }
getFixedMin()280   static constexpr Immediate getFixedMin() {
281     return {std::numeric_limits<int64_t>::min(), false};
282   }
getFixedMax()283   static constexpr Immediate getFixedMax() {
284     return {std::numeric_limits<int64_t>::max(), false};
285   }
getScalableMin()286   static constexpr Immediate getScalableMin() {
287     return {std::numeric_limits<int64_t>::min(), true};
288   }
getScalableMax()289   static constexpr Immediate getScalableMax() {
290     return {std::numeric_limits<int64_t>::max(), true};
291   }
292 
isLessThanZero() const293   constexpr bool isLessThanZero() const { return Quantity < 0; }
294 
isGreaterThanZero() const295   constexpr bool isGreaterThanZero() const { return Quantity > 0; }
296 
isCompatibleImmediate(const Immediate & Imm) const297   constexpr bool isCompatibleImmediate(const Immediate &Imm) const {
298     return isZero() || Imm.isZero() || Imm.Scalable == Scalable;
299   }
300 
isMin() const301   constexpr bool isMin() const {
302     return Quantity == std::numeric_limits<ScalarTy>::min();
303   }
304 
isMax() const305   constexpr bool isMax() const {
306     return Quantity == std::numeric_limits<ScalarTy>::max();
307   }
308 
309   // Arithmetic 'operators' that cast to unsigned types first.
addUnsigned(const Immediate & RHS) const310   constexpr Immediate addUnsigned(const Immediate &RHS) const {
311     assert(isCompatibleImmediate(RHS) && "Incompatible Immediates");
312     ScalarTy Value = (uint64_t)Quantity + RHS.getKnownMinValue();
313     return {Value, Scalable || RHS.isScalable()};
314   }
315 
subUnsigned(const Immediate & RHS) const316   constexpr Immediate subUnsigned(const Immediate &RHS) const {
317     assert(isCompatibleImmediate(RHS) && "Incompatible Immediates");
318     ScalarTy Value = (uint64_t)Quantity - RHS.getKnownMinValue();
319     return {Value, Scalable || RHS.isScalable()};
320   }
321 
322   // Scale the quantity by a constant without caring about runtime scalability.
mulUnsigned(const ScalarTy RHS) const323   constexpr Immediate mulUnsigned(const ScalarTy RHS) const {
324     ScalarTy Value = (uint64_t)Quantity * RHS;
325     return {Value, Scalable};
326   }
327 
328   // Helpers for generating SCEVs with vscale terms where needed.
getSCEV(ScalarEvolution & SE,Type * Ty) const329   const SCEV *getSCEV(ScalarEvolution &SE, Type *Ty) const {
330     const SCEV *S = SE.getConstant(Ty, Quantity);
331     if (Scalable)
332       S = SE.getMulExpr(S, SE.getVScale(S->getType()));
333     return S;
334   }
335 
getNegativeSCEV(ScalarEvolution & SE,Type * Ty) const336   const SCEV *getNegativeSCEV(ScalarEvolution &SE, Type *Ty) const {
337     const SCEV *NegS = SE.getConstant(Ty, -(uint64_t)Quantity);
338     if (Scalable)
339       NegS = SE.getMulExpr(NegS, SE.getVScale(NegS->getType()));
340     return NegS;
341   }
342 
getUnknownSCEV(ScalarEvolution & SE,Type * Ty) const343   const SCEV *getUnknownSCEV(ScalarEvolution &SE, Type *Ty) const {
344     const SCEV *SU = SE.getUnknown(ConstantInt::getSigned(Ty, Quantity));
345     if (Scalable)
346       SU = SE.getMulExpr(SU, SE.getVScale(SU->getType()));
347     return SU;
348   }
349 };
350 
351 // This is needed for the Compare type of std::map when Immediate is used
352 // as a key. We don't need it to be fully correct against any value of vscale,
353 // just to make sure that vscale-related terms in the map are considered against
354 // each other rather than being mixed up and potentially missing opportunities.
355 struct KeyOrderTargetImmediate {
operator ()__anonc21373340111::KeyOrderTargetImmediate356   bool operator()(const Immediate &LHS, const Immediate &RHS) const {
357     if (LHS.isScalable() && !RHS.isScalable())
358       return false;
359     if (!LHS.isScalable() && RHS.isScalable())
360       return true;
361     return LHS.getKnownMinValue() < RHS.getKnownMinValue();
362   }
363 };
364 
365 // This would be nicer if we could be generic instead of directly using size_t,
366 // but there doesn't seem to be a type trait for is_orderable or
367 // is_lessthan_comparable or similar.
368 struct KeyOrderSizeTAndImmediate {
operator ()__anonc21373340111::KeyOrderSizeTAndImmediate369   bool operator()(const std::pair<size_t, Immediate> &LHS,
370                   const std::pair<size_t, Immediate> &RHS) const {
371     size_t LSize = LHS.first;
372     size_t RSize = RHS.first;
373     if (LSize != RSize)
374       return LSize < RSize;
375     return KeyOrderTargetImmediate()(LHS.second, RHS.second);
376   }
377 };
378 } // end anonymous namespace
379 
380 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print(raw_ostream & OS) const381 void RegSortData::print(raw_ostream &OS) const {
382   OS << "[NumUses=" << UsedByIndices.count() << ']';
383 }
384 
dump() const385 LLVM_DUMP_METHOD void RegSortData::dump() const {
386   print(errs()); errs() << '\n';
387 }
388 #endif
389 
390 namespace {
391 
392 /// Map register candidates to information about how they are used.
393 class RegUseTracker {
394   using RegUsesTy = DenseMap<const SCEV *, RegSortData>;
395 
396   RegUsesTy RegUsesMap;
397   SmallVector<const SCEV *, 16> RegSequence;
398 
399 public:
400   void countRegister(const SCEV *Reg, size_t LUIdx);
401   void dropRegister(const SCEV *Reg, size_t LUIdx);
402   void swapAndDropUse(size_t LUIdx, size_t LastLUIdx);
403 
404   bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
405 
406   const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
407 
408   void clear();
409 
410   using iterator = SmallVectorImpl<const SCEV *>::iterator;
411   using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator;
412 
begin()413   iterator begin() { return RegSequence.begin(); }
end()414   iterator end()   { return RegSequence.end(); }
begin() const415   const_iterator begin() const { return RegSequence.begin(); }
end() const416   const_iterator end() const   { return RegSequence.end(); }
417 };
418 
419 } // end anonymous namespace
420 
421 void
countRegister(const SCEV * Reg,size_t LUIdx)422 RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) {
423   std::pair<RegUsesTy::iterator, bool> Pair =
424     RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
425   RegSortData &RSD = Pair.first->second;
426   if (Pair.second)
427     RegSequence.push_back(Reg);
428   RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
429   RSD.UsedByIndices.set(LUIdx);
430 }
431 
432 void
dropRegister(const SCEV * Reg,size_t LUIdx)433 RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) {
434   RegUsesTy::iterator It = RegUsesMap.find(Reg);
435   assert(It != RegUsesMap.end());
436   RegSortData &RSD = It->second;
437   assert(RSD.UsedByIndices.size() > LUIdx);
438   RSD.UsedByIndices.reset(LUIdx);
439 }
440 
441 void
swapAndDropUse(size_t LUIdx,size_t LastLUIdx)442 RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
443   assert(LUIdx <= LastLUIdx);
444 
445   // Update RegUses. The data structure is not optimized for this purpose;
446   // we must iterate through it and update each of the bit vectors.
447   for (auto &Pair : RegUsesMap) {
448     SmallBitVector &UsedByIndices = Pair.second.UsedByIndices;
449     if (LUIdx < UsedByIndices.size())
450       UsedByIndices[LUIdx] =
451         LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false;
452     UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
453   }
454 }
455 
456 bool
isRegUsedByUsesOtherThan(const SCEV * Reg,size_t LUIdx) const457 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
458   RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
459   if (I == RegUsesMap.end())
460     return false;
461   const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
462   int i = UsedByIndices.find_first();
463   if (i == -1) return false;
464   if ((size_t)i != LUIdx) return true;
465   return UsedByIndices.find_next(i) != -1;
466 }
467 
getUsedByIndices(const SCEV * Reg) const468 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
469   RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
470   assert(I != RegUsesMap.end() && "Unknown register!");
471   return I->second.UsedByIndices;
472 }
473 
clear()474 void RegUseTracker::clear() {
475   RegUsesMap.clear();
476   RegSequence.clear();
477 }
478 
479 namespace {
480 
481 /// This class holds information that describes a formula for computing
482 /// satisfying a use. It may include broken-out immediates and scaled registers.
483 struct Formula {
484   /// Global base address used for complex addressing.
485   GlobalValue *BaseGV = nullptr;
486 
487   /// Base offset for complex addressing.
488   Immediate BaseOffset = Immediate::getZero();
489 
490   /// Whether any complex addressing has a base register.
491   bool HasBaseReg = false;
492 
493   /// The scale of any complex addressing.
494   int64_t Scale = 0;
495 
496   /// The list of "base" registers for this use. When this is non-empty. The
497   /// canonical representation of a formula is
498   /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and
499   /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty().
500   /// 3. The reg containing recurrent expr related with currect loop in the
501   /// formula should be put in the ScaledReg.
502   /// #1 enforces that the scaled register is always used when at least two
503   /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2.
504   /// #2 enforces that 1 * reg is reg.
505   /// #3 ensures invariant regs with respect to current loop can be combined
506   /// together in LSR codegen.
507   /// This invariant can be temporarily broken while building a formula.
508   /// However, every formula inserted into the LSRInstance must be in canonical
509   /// form.
510   SmallVector<const SCEV *, 4> BaseRegs;
511 
512   /// The 'scaled' register for this use. This should be non-null when Scale is
513   /// not zero.
514   const SCEV *ScaledReg = nullptr;
515 
516   /// An additional constant offset which added near the use. This requires a
517   /// temporary register, but the offset itself can live in an add immediate
518   /// field rather than a register.
519   Immediate UnfoldedOffset = Immediate::getZero();
520 
521   Formula() = default;
522 
523   void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
524 
525   bool isCanonical(const Loop &L) const;
526 
527   void canonicalize(const Loop &L);
528 
529   bool unscale();
530 
531   bool hasZeroEnd() const;
532 
533   size_t getNumRegs() const;
534   Type *getType() const;
535 
536   void deleteBaseReg(const SCEV *&S);
537 
538   bool referencesReg(const SCEV *S) const;
539   bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
540                                   const RegUseTracker &RegUses) const;
541 
542   void print(raw_ostream &OS) const;
543   void dump() const;
544 };
545 
546 } // end anonymous namespace
547 
548 /// Recursion helper for initialMatch.
DoInitialMatch(const SCEV * S,Loop * L,SmallVectorImpl<const SCEV * > & Good,SmallVectorImpl<const SCEV * > & Bad,ScalarEvolution & SE)549 static void DoInitialMatch(const SCEV *S, Loop *L,
550                            SmallVectorImpl<const SCEV *> &Good,
551                            SmallVectorImpl<const SCEV *> &Bad,
552                            ScalarEvolution &SE) {
553   // Collect expressions which properly dominate the loop header.
554   if (SE.properlyDominates(S, L->getHeader())) {
555     Good.push_back(S);
556     return;
557   }
558 
559   // Look at add operands.
560   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
561     for (const SCEV *S : Add->operands())
562       DoInitialMatch(S, L, Good, Bad, SE);
563     return;
564   }
565 
566   // Look at addrec operands.
567   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
568     if (!AR->getStart()->isZero() && AR->isAffine()) {
569       DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
570       DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
571                                       AR->getStepRecurrence(SE),
572                                       // FIXME: AR->getNoWrapFlags()
573                                       AR->getLoop(), SCEV::FlagAnyWrap),
574                      L, Good, Bad, SE);
575       return;
576     }
577 
578   // Handle a multiplication by -1 (negation) if it didn't fold.
579   if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
580     if (Mul->getOperand(0)->isAllOnesValue()) {
581       SmallVector<const SCEV *, 4> Ops(drop_begin(Mul->operands()));
582       const SCEV *NewMul = SE.getMulExpr(Ops);
583 
584       SmallVector<const SCEV *, 4> MyGood;
585       SmallVector<const SCEV *, 4> MyBad;
586       DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
587       const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
588         SE.getEffectiveSCEVType(NewMul->getType())));
589       for (const SCEV *S : MyGood)
590         Good.push_back(SE.getMulExpr(NegOne, S));
591       for (const SCEV *S : MyBad)
592         Bad.push_back(SE.getMulExpr(NegOne, S));
593       return;
594     }
595 
596   // Ok, we can't do anything interesting. Just stuff the whole thing into a
597   // register and hope for the best.
598   Bad.push_back(S);
599 }
600 
601 /// Incorporate loop-variant parts of S into this Formula, attempting to keep
602 /// all loop-invariant and loop-computable values in a single base register.
initialMatch(const SCEV * S,Loop * L,ScalarEvolution & SE)603 void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
604   SmallVector<const SCEV *, 4> Good;
605   SmallVector<const SCEV *, 4> Bad;
606   DoInitialMatch(S, L, Good, Bad, SE);
607   if (!Good.empty()) {
608     const SCEV *Sum = SE.getAddExpr(Good);
609     if (!Sum->isZero())
610       BaseRegs.push_back(Sum);
611     HasBaseReg = true;
612   }
613   if (!Bad.empty()) {
614     const SCEV *Sum = SE.getAddExpr(Bad);
615     if (!Sum->isZero())
616       BaseRegs.push_back(Sum);
617     HasBaseReg = true;
618   }
619   canonicalize(*L);
620 }
621 
containsAddRecDependentOnLoop(const SCEV * S,const Loop & L)622 static bool containsAddRecDependentOnLoop(const SCEV *S, const Loop &L) {
623   return SCEVExprContains(S, [&L](const SCEV *S) {
624     return isa<SCEVAddRecExpr>(S) && (cast<SCEVAddRecExpr>(S)->getLoop() == &L);
625   });
626 }
627 
628 /// Check whether or not this formula satisfies the canonical
629 /// representation.
630 /// \see Formula::BaseRegs.
isCanonical(const Loop & L) const631 bool Formula::isCanonical(const Loop &L) const {
632   if (!ScaledReg)
633     return BaseRegs.size() <= 1;
634 
635   if (Scale != 1)
636     return true;
637 
638   if (Scale == 1 && BaseRegs.empty())
639     return false;
640 
641   if (containsAddRecDependentOnLoop(ScaledReg, L))
642     return true;
643 
644   // If ScaledReg is not a recurrent expr, or it is but its loop is not current
645   // loop, meanwhile BaseRegs contains a recurrent expr reg related with current
646   // loop, we want to swap the reg in BaseRegs with ScaledReg.
647   return none_of(BaseRegs, [&L](const SCEV *S) {
648     return containsAddRecDependentOnLoop(S, L);
649   });
650 }
651 
652 /// Helper method to morph a formula into its canonical representation.
653 /// \see Formula::BaseRegs.
654 /// Every formula having more than one base register, must use the ScaledReg
655 /// field. Otherwise, we would have to do special cases everywhere in LSR
656 /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ...
657 /// On the other hand, 1*reg should be canonicalized into reg.
canonicalize(const Loop & L)658 void Formula::canonicalize(const Loop &L) {
659   if (isCanonical(L))
660     return;
661 
662   if (BaseRegs.empty()) {
663     // No base reg? Use scale reg with scale = 1 as such.
664     assert(ScaledReg && "Expected 1*reg => reg");
665     assert(Scale == 1 && "Expected 1*reg => reg");
666     BaseRegs.push_back(ScaledReg);
667     Scale = 0;
668     ScaledReg = nullptr;
669     return;
670   }
671 
672   // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg.
673   if (!ScaledReg) {
674     ScaledReg = BaseRegs.pop_back_val();
675     Scale = 1;
676   }
677 
678   // If ScaledReg is an invariant with respect to L, find the reg from
679   // BaseRegs containing the recurrent expr related with Loop L. Swap the
680   // reg with ScaledReg.
681   if (!containsAddRecDependentOnLoop(ScaledReg, L)) {
682     auto I = find_if(BaseRegs, [&L](const SCEV *S) {
683       return containsAddRecDependentOnLoop(S, L);
684     });
685     if (I != BaseRegs.end())
686       std::swap(ScaledReg, *I);
687   }
688   assert(isCanonical(L) && "Failed to canonicalize?");
689 }
690 
691 /// Get rid of the scale in the formula.
692 /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2.
693 /// \return true if it was possible to get rid of the scale, false otherwise.
694 /// \note After this operation the formula may not be in the canonical form.
unscale()695 bool Formula::unscale() {
696   if (Scale != 1)
697     return false;
698   Scale = 0;
699   BaseRegs.push_back(ScaledReg);
700   ScaledReg = nullptr;
701   return true;
702 }
703 
hasZeroEnd() const704 bool Formula::hasZeroEnd() const {
705   if (UnfoldedOffset || BaseOffset)
706     return false;
707   if (BaseRegs.size() != 1 || ScaledReg)
708     return false;
709   return true;
710 }
711 
712 /// Return the total number of register operands used by this formula. This does
713 /// not include register uses implied by non-constant addrec strides.
getNumRegs() const714 size_t Formula::getNumRegs() const {
715   return !!ScaledReg + BaseRegs.size();
716 }
717 
718 /// Return the type of this formula, if it has one, or null otherwise. This type
719 /// is meaningless except for the bit size.
getType() const720 Type *Formula::getType() const {
721   return !BaseRegs.empty() ? BaseRegs.front()->getType() :
722          ScaledReg ? ScaledReg->getType() :
723          BaseGV ? BaseGV->getType() :
724          nullptr;
725 }
726 
727 /// Delete the given base reg from the BaseRegs list.
deleteBaseReg(const SCEV * & S)728 void Formula::deleteBaseReg(const SCEV *&S) {
729   if (&S != &BaseRegs.back())
730     std::swap(S, BaseRegs.back());
731   BaseRegs.pop_back();
732 }
733 
734 /// Test if this formula references the given register.
referencesReg(const SCEV * S) const735 bool Formula::referencesReg(const SCEV *S) const {
736   return S == ScaledReg || is_contained(BaseRegs, S);
737 }
738 
739 /// Test whether this formula uses registers which are used by uses other than
740 /// the use with the given index.
hasRegsUsedByUsesOtherThan(size_t LUIdx,const RegUseTracker & RegUses) const741 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
742                                          const RegUseTracker &RegUses) const {
743   if (ScaledReg)
744     if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
745       return true;
746   for (const SCEV *BaseReg : BaseRegs)
747     if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx))
748       return true;
749   return false;
750 }
751 
752 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print(raw_ostream & OS) const753 void Formula::print(raw_ostream &OS) const {
754   bool First = true;
755   if (BaseGV) {
756     if (!First) OS << " + "; else First = false;
757     BaseGV->printAsOperand(OS, /*PrintType=*/false);
758   }
759   if (BaseOffset.isNonZero()) {
760     if (!First) OS << " + "; else First = false;
761     OS << BaseOffset;
762   }
763   for (const SCEV *BaseReg : BaseRegs) {
764     if (!First) OS << " + "; else First = false;
765     OS << "reg(" << *BaseReg << ')';
766   }
767   if (HasBaseReg && BaseRegs.empty()) {
768     if (!First) OS << " + "; else First = false;
769     OS << "**error: HasBaseReg**";
770   } else if (!HasBaseReg && !BaseRegs.empty()) {
771     if (!First) OS << " + "; else First = false;
772     OS << "**error: !HasBaseReg**";
773   }
774   if (Scale != 0) {
775     if (!First) OS << " + "; else First = false;
776     OS << Scale << "*reg(";
777     if (ScaledReg)
778       OS << *ScaledReg;
779     else
780       OS << "<unknown>";
781     OS << ')';
782   }
783   if (UnfoldedOffset.isNonZero()) {
784     if (!First) OS << " + ";
785     OS << "imm(" << UnfoldedOffset << ')';
786   }
787 }
788 
dump() const789 LLVM_DUMP_METHOD void Formula::dump() const {
790   print(errs()); errs() << '\n';
791 }
792 #endif
793 
794 /// Return true if the given addrec can be sign-extended without changing its
795 /// value.
isAddRecSExtable(const SCEVAddRecExpr * AR,ScalarEvolution & SE)796 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
797   Type *WideTy =
798     IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
799   return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
800 }
801 
802 /// Return true if the given add can be sign-extended without changing its
803 /// value.
isAddSExtable(const SCEVAddExpr * A,ScalarEvolution & SE)804 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
805   Type *WideTy =
806     IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
807   return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
808 }
809 
810 /// Return true if the given mul can be sign-extended without changing its
811 /// value.
isMulSExtable(const SCEVMulExpr * M,ScalarEvolution & SE)812 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
813   Type *WideTy =
814     IntegerType::get(SE.getContext(),
815                      SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
816   return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
817 }
818 
819 /// Return an expression for LHS /s RHS, if it can be determined and if the
820 /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits
821 /// is true, expressions like (X * Y) /s Y are simplified to X, ignoring that
822 /// the multiplication may overflow, which is useful when the result will be
823 /// used in a context where the most significant bits are ignored.
getExactSDiv(const SCEV * LHS,const SCEV * RHS,ScalarEvolution & SE,bool IgnoreSignificantBits=false)824 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
825                                 ScalarEvolution &SE,
826                                 bool IgnoreSignificantBits = false) {
827   // Handle the trivial case, which works for any SCEV type.
828   if (LHS == RHS)
829     return SE.getConstant(LHS->getType(), 1);
830 
831   // Handle a few RHS special cases.
832   const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
833   if (RC) {
834     const APInt &RA = RC->getAPInt();
835     // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
836     // some folding.
837     if (RA.isAllOnes()) {
838       if (LHS->getType()->isPointerTy())
839         return nullptr;
840       return SE.getMulExpr(LHS, RC);
841     }
842     // Handle x /s 1 as x.
843     if (RA == 1)
844       return LHS;
845   }
846 
847   // Check for a division of a constant by a constant.
848   if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
849     if (!RC)
850       return nullptr;
851     const APInt &LA = C->getAPInt();
852     const APInt &RA = RC->getAPInt();
853     if (LA.srem(RA) != 0)
854       return nullptr;
855     return SE.getConstant(LA.sdiv(RA));
856   }
857 
858   // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
859   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
860     if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) {
861       const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
862                                       IgnoreSignificantBits);
863       if (!Step) return nullptr;
864       const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
865                                        IgnoreSignificantBits);
866       if (!Start) return nullptr;
867       // FlagNW is independent of the start value, step direction, and is
868       // preserved with smaller magnitude steps.
869       // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
870       return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
871     }
872     return nullptr;
873   }
874 
875   // Distribute the sdiv over add operands, if the add doesn't overflow.
876   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
877     if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
878       SmallVector<const SCEV *, 8> Ops;
879       for (const SCEV *S : Add->operands()) {
880         const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits);
881         if (!Op) return nullptr;
882         Ops.push_back(Op);
883       }
884       return SE.getAddExpr(Ops);
885     }
886     return nullptr;
887   }
888 
889   // Check for a multiply operand that we can pull RHS out of.
890   if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
891     if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
892       // Handle special case C1*X*Y /s C2*X*Y.
893       if (const SCEVMulExpr *MulRHS = dyn_cast<SCEVMulExpr>(RHS)) {
894         if (IgnoreSignificantBits || isMulSExtable(MulRHS, SE)) {
895           const SCEVConstant *LC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
896           const SCEVConstant *RC =
897               dyn_cast<SCEVConstant>(MulRHS->getOperand(0));
898           if (LC && RC) {
899             SmallVector<const SCEV *, 4> LOps(drop_begin(Mul->operands()));
900             SmallVector<const SCEV *, 4> ROps(drop_begin(MulRHS->operands()));
901             if (LOps == ROps)
902               return getExactSDiv(LC, RC, SE, IgnoreSignificantBits);
903           }
904         }
905       }
906 
907       SmallVector<const SCEV *, 4> Ops;
908       bool Found = false;
909       for (const SCEV *S : Mul->operands()) {
910         if (!Found)
911           if (const SCEV *Q = getExactSDiv(S, RHS, SE,
912                                            IgnoreSignificantBits)) {
913             S = Q;
914             Found = true;
915           }
916         Ops.push_back(S);
917       }
918       return Found ? SE.getMulExpr(Ops) : nullptr;
919     }
920     return nullptr;
921   }
922 
923   // Otherwise we don't know.
924   return nullptr;
925 }
926 
927 /// If S involves the addition of a constant integer value, return that integer
928 /// value, and mutate S to point to a new SCEV with that value excluded.
ExtractImmediate(const SCEV * & S,ScalarEvolution & SE)929 static Immediate ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
930   if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
931     if (C->getAPInt().getSignificantBits() <= 64) {
932       S = SE.getConstant(C->getType(), 0);
933       return Immediate::getFixed(C->getValue()->getSExtValue());
934     }
935   } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
936     SmallVector<const SCEV *, 8> NewOps(Add->operands());
937     Immediate Result = ExtractImmediate(NewOps.front(), SE);
938     if (Result.isNonZero())
939       S = SE.getAddExpr(NewOps);
940     return Result;
941   } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
942     SmallVector<const SCEV *, 8> NewOps(AR->operands());
943     Immediate Result = ExtractImmediate(NewOps.front(), SE);
944     if (Result.isNonZero())
945       S = SE.getAddRecExpr(NewOps, AR->getLoop(),
946                            // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
947                            SCEV::FlagAnyWrap);
948     return Result;
949   } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
950     if (EnableVScaleImmediates && M->getNumOperands() == 2) {
951       if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
952         if (isa<SCEVVScale>(M->getOperand(1))) {
953           S = SE.getConstant(M->getType(), 0);
954           return Immediate::getScalable(C->getValue()->getSExtValue());
955         }
956     }
957   }
958   return Immediate::getZero();
959 }
960 
961 /// If S involves the addition of a GlobalValue address, return that symbol, and
962 /// mutate S to point to a new SCEV with that value excluded.
ExtractSymbol(const SCEV * & S,ScalarEvolution & SE)963 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
964   if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
965     if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
966       S = SE.getConstant(GV->getType(), 0);
967       return GV;
968     }
969   } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
970     SmallVector<const SCEV *, 8> NewOps(Add->operands());
971     GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
972     if (Result)
973       S = SE.getAddExpr(NewOps);
974     return Result;
975   } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
976     SmallVector<const SCEV *, 8> NewOps(AR->operands());
977     GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
978     if (Result)
979       S = SE.getAddRecExpr(NewOps, AR->getLoop(),
980                            // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
981                            SCEV::FlagAnyWrap);
982     return Result;
983   }
984   return nullptr;
985 }
986 
987 /// Returns true if the specified instruction is using the specified value as an
988 /// address.
isAddressUse(const TargetTransformInfo & TTI,Instruction * Inst,Value * OperandVal)989 static bool isAddressUse(const TargetTransformInfo &TTI,
990                          Instruction *Inst, Value *OperandVal) {
991   bool isAddress = isa<LoadInst>(Inst);
992   if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
993     if (SI->getPointerOperand() == OperandVal)
994       isAddress = true;
995   } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
996     // Addressing modes can also be folded into prefetches and a variety
997     // of intrinsics.
998     switch (II->getIntrinsicID()) {
999     case Intrinsic::memset:
1000     case Intrinsic::prefetch:
1001     case Intrinsic::masked_load:
1002       if (II->getArgOperand(0) == OperandVal)
1003         isAddress = true;
1004       break;
1005     case Intrinsic::masked_store:
1006       if (II->getArgOperand(1) == OperandVal)
1007         isAddress = true;
1008       break;
1009     case Intrinsic::memmove:
1010     case Intrinsic::memcpy:
1011       if (II->getArgOperand(0) == OperandVal ||
1012           II->getArgOperand(1) == OperandVal)
1013         isAddress = true;
1014       break;
1015     default: {
1016       MemIntrinsicInfo IntrInfo;
1017       if (TTI.getTgtMemIntrinsic(II, IntrInfo)) {
1018         if (IntrInfo.PtrVal == OperandVal)
1019           isAddress = true;
1020       }
1021     }
1022     }
1023   } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
1024     if (RMW->getPointerOperand() == OperandVal)
1025       isAddress = true;
1026   } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
1027     if (CmpX->getPointerOperand() == OperandVal)
1028       isAddress = true;
1029   }
1030   return isAddress;
1031 }
1032 
1033 /// Return the type of the memory being accessed.
getAccessType(const TargetTransformInfo & TTI,Instruction * Inst,Value * OperandVal)1034 static MemAccessTy getAccessType(const TargetTransformInfo &TTI,
1035                                  Instruction *Inst, Value *OperandVal) {
1036   MemAccessTy AccessTy = MemAccessTy::getUnknown(Inst->getContext());
1037 
1038   // First get the type of memory being accessed.
1039   if (Type *Ty = Inst->getAccessType())
1040     AccessTy.MemTy = Ty;
1041 
1042   // Then get the pointer address space.
1043   if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1044     AccessTy.AddrSpace = SI->getPointerAddressSpace();
1045   } else if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1046     AccessTy.AddrSpace = LI->getPointerAddressSpace();
1047   } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Inst)) {
1048     AccessTy.AddrSpace = RMW->getPointerAddressSpace();
1049   } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Inst)) {
1050     AccessTy.AddrSpace = CmpX->getPointerAddressSpace();
1051   } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1052     switch (II->getIntrinsicID()) {
1053     case Intrinsic::prefetch:
1054     case Intrinsic::memset:
1055       AccessTy.AddrSpace = II->getArgOperand(0)->getType()->getPointerAddressSpace();
1056       AccessTy.MemTy = OperandVal->getType();
1057       break;
1058     case Intrinsic::memmove:
1059     case Intrinsic::memcpy:
1060       AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace();
1061       AccessTy.MemTy = OperandVal->getType();
1062       break;
1063     case Intrinsic::masked_load:
1064       AccessTy.AddrSpace =
1065           II->getArgOperand(0)->getType()->getPointerAddressSpace();
1066       break;
1067     case Intrinsic::masked_store:
1068       AccessTy.AddrSpace =
1069           II->getArgOperand(1)->getType()->getPointerAddressSpace();
1070       break;
1071     default: {
1072       MemIntrinsicInfo IntrInfo;
1073       if (TTI.getTgtMemIntrinsic(II, IntrInfo) && IntrInfo.PtrVal) {
1074         AccessTy.AddrSpace
1075           = IntrInfo.PtrVal->getType()->getPointerAddressSpace();
1076       }
1077 
1078       break;
1079     }
1080     }
1081   }
1082 
1083   return AccessTy;
1084 }
1085 
1086 /// Return true if this AddRec is already a phi in its loop.
isExistingPhi(const SCEVAddRecExpr * AR,ScalarEvolution & SE)1087 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
1088   for (PHINode &PN : AR->getLoop()->getHeader()->phis()) {
1089     if (SE.isSCEVable(PN.getType()) &&
1090         (SE.getEffectiveSCEVType(PN.getType()) ==
1091          SE.getEffectiveSCEVType(AR->getType())) &&
1092         SE.getSCEV(&PN) == AR)
1093       return true;
1094   }
1095   return false;
1096 }
1097 
1098 /// Check if expanding this expression is likely to incur significant cost. This
1099 /// is tricky because SCEV doesn't track which expressions are actually computed
1100 /// by the current IR.
1101 ///
1102 /// We currently allow expansion of IV increments that involve adds,
1103 /// multiplication by constants, and AddRecs from existing phis.
1104 ///
1105 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
1106 /// obvious multiple of the UDivExpr.
isHighCostExpansion(const SCEV * S,SmallPtrSetImpl<const SCEV * > & Processed,ScalarEvolution & SE)1107 static bool isHighCostExpansion(const SCEV *S,
1108                                 SmallPtrSetImpl<const SCEV*> &Processed,
1109                                 ScalarEvolution &SE) {
1110   // Zero/One operand expressions
1111   switch (S->getSCEVType()) {
1112   case scUnknown:
1113   case scConstant:
1114   case scVScale:
1115     return false;
1116   case scTruncate:
1117     return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
1118                                Processed, SE);
1119   case scZeroExtend:
1120     return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
1121                                Processed, SE);
1122   case scSignExtend:
1123     return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
1124                                Processed, SE);
1125   default:
1126     break;
1127   }
1128 
1129   if (!Processed.insert(S).second)
1130     return false;
1131 
1132   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1133     for (const SCEV *S : Add->operands()) {
1134       if (isHighCostExpansion(S, Processed, SE))
1135         return true;
1136     }
1137     return false;
1138   }
1139 
1140   if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
1141     if (Mul->getNumOperands() == 2) {
1142       // Multiplication by a constant is ok
1143       if (isa<SCEVConstant>(Mul->getOperand(0)))
1144         return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
1145 
1146       // If we have the value of one operand, check if an existing
1147       // multiplication already generates this expression.
1148       if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
1149         Value *UVal = U->getValue();
1150         for (User *UR : UVal->users()) {
1151           // If U is a constant, it may be used by a ConstantExpr.
1152           Instruction *UI = dyn_cast<Instruction>(UR);
1153           if (UI && UI->getOpcode() == Instruction::Mul &&
1154               SE.isSCEVable(UI->getType())) {
1155             return SE.getSCEV(UI) == Mul;
1156           }
1157         }
1158       }
1159     }
1160   }
1161 
1162   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1163     if (isExistingPhi(AR, SE))
1164       return false;
1165   }
1166 
1167   // Fow now, consider any other type of expression (div/mul/min/max) high cost.
1168   return true;
1169 }
1170 
1171 namespace {
1172 
1173 class LSRUse;
1174 
1175 } // end anonymous namespace
1176 
1177 /// Check if the addressing mode defined by \p F is completely
1178 /// folded in \p LU at isel time.
1179 /// This includes address-mode folding and special icmp tricks.
1180 /// This function returns true if \p LU can accommodate what \p F
1181 /// defines and up to 1 base + 1 scaled + offset.
1182 /// In other words, if \p F has several base registers, this function may
1183 /// still return true. Therefore, users still need to account for
1184 /// additional base registers and/or unfolded offsets to derive an
1185 /// accurate cost model.
1186 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1187                                  const LSRUse &LU, const Formula &F);
1188 
1189 // Get the cost of the scaling factor used in F for LU.
1190 static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI,
1191                                             const LSRUse &LU, const Formula &F,
1192                                             const Loop &L);
1193 
1194 namespace {
1195 
1196 /// This class is used to measure and compare candidate formulae.
1197 class Cost {
1198   const Loop *L = nullptr;
1199   ScalarEvolution *SE = nullptr;
1200   const TargetTransformInfo *TTI = nullptr;
1201   TargetTransformInfo::LSRCost C;
1202   TTI::AddressingModeKind AMK = TTI::AMK_None;
1203 
1204 public:
1205   Cost() = delete;
Cost(const Loop * L,ScalarEvolution & SE,const TargetTransformInfo & TTI,TTI::AddressingModeKind AMK)1206   Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI,
1207        TTI::AddressingModeKind AMK) :
1208     L(L), SE(&SE), TTI(&TTI), AMK(AMK) {
1209     C.Insns = 0;
1210     C.NumRegs = 0;
1211     C.AddRecCost = 0;
1212     C.NumIVMuls = 0;
1213     C.NumBaseAdds = 0;
1214     C.ImmCost = 0;
1215     C.SetupCost = 0;
1216     C.ScaleCost = 0;
1217   }
1218 
1219   bool isLess(const Cost &Other) const;
1220 
1221   void Lose();
1222 
1223 #ifndef NDEBUG
1224   // Once any of the metrics loses, they must all remain losers.
isValid()1225   bool isValid() {
1226     return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds
1227              | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u)
1228       || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds
1229            & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u);
1230   }
1231 #endif
1232 
isLoser()1233   bool isLoser() {
1234     assert(isValid() && "invalid cost");
1235     return C.NumRegs == ~0u;
1236   }
1237 
1238   void RateFormula(const Formula &F,
1239                    SmallPtrSetImpl<const SCEV *> &Regs,
1240                    const DenseSet<const SCEV *> &VisitedRegs,
1241                    const LSRUse &LU,
1242                    SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr);
1243 
1244   void print(raw_ostream &OS) const;
1245   void dump() const;
1246 
1247 private:
1248   void RateRegister(const Formula &F, const SCEV *Reg,
1249                     SmallPtrSetImpl<const SCEV *> &Regs);
1250   void RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1251                            SmallPtrSetImpl<const SCEV *> &Regs,
1252                            SmallPtrSetImpl<const SCEV *> *LoserRegs);
1253 };
1254 
1255 /// An operand value in an instruction which is to be replaced with some
1256 /// equivalent, possibly strength-reduced, replacement.
1257 struct LSRFixup {
1258   /// The instruction which will be updated.
1259   Instruction *UserInst = nullptr;
1260 
1261   /// The operand of the instruction which will be replaced. The operand may be
1262   /// used more than once; every instance will be replaced.
1263   Value *OperandValToReplace = nullptr;
1264 
1265   /// If this user is to use the post-incremented value of an induction
1266   /// variable, this set is non-empty and holds the loops associated with the
1267   /// induction variable.
1268   PostIncLoopSet PostIncLoops;
1269 
1270   /// A constant offset to be added to the LSRUse expression.  This allows
1271   /// multiple fixups to share the same LSRUse with different offsets, for
1272   /// example in an unrolled loop.
1273   Immediate Offset = Immediate::getZero();
1274 
1275   LSRFixup() = default;
1276 
1277   bool isUseFullyOutsideLoop(const Loop *L) const;
1278 
1279   void print(raw_ostream &OS) const;
1280   void dump() const;
1281 };
1282 
1283 /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of sorted
1284 /// SmallVectors of const SCEV*.
1285 struct UniquifierDenseMapInfo {
getEmptyKey__anonc21373340811::UniquifierDenseMapInfo1286   static SmallVector<const SCEV *, 4> getEmptyKey() {
1287     SmallVector<const SCEV *, 4>  V;
1288     V.push_back(reinterpret_cast<const SCEV *>(-1));
1289     return V;
1290   }
1291 
getTombstoneKey__anonc21373340811::UniquifierDenseMapInfo1292   static SmallVector<const SCEV *, 4> getTombstoneKey() {
1293     SmallVector<const SCEV *, 4> V;
1294     V.push_back(reinterpret_cast<const SCEV *>(-2));
1295     return V;
1296   }
1297 
getHashValue__anonc21373340811::UniquifierDenseMapInfo1298   static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
1299     return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
1300   }
1301 
isEqual__anonc21373340811::UniquifierDenseMapInfo1302   static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
1303                       const SmallVector<const SCEV *, 4> &RHS) {
1304     return LHS == RHS;
1305   }
1306 };
1307 
1308 /// This class holds the state that LSR keeps for each use in IVUsers, as well
1309 /// as uses invented by LSR itself. It includes information about what kinds of
1310 /// things can be folded into the user, information about the user itself, and
1311 /// information about how the use may be satisfied.  TODO: Represent multiple
1312 /// users of the same expression in common?
1313 class LSRUse {
1314   DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
1315 
1316 public:
1317   /// An enum for a kind of use, indicating what types of scaled and immediate
1318   /// operands it might support.
1319   enum KindType {
1320     Basic,   ///< A normal use, with no folding.
1321     Special, ///< A special case of basic, allowing -1 scales.
1322     Address, ///< An address use; folding according to TargetLowering
1323     ICmpZero ///< An equality icmp with both operands folded into one.
1324     // TODO: Add a generic icmp too?
1325   };
1326 
1327   using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>;
1328 
1329   KindType Kind;
1330   MemAccessTy AccessTy;
1331 
1332   /// The list of operands which are to be replaced.
1333   SmallVector<LSRFixup, 8> Fixups;
1334 
1335   /// Keep track of the min and max offsets of the fixups.
1336   Immediate MinOffset = Immediate::getFixedMax();
1337   Immediate MaxOffset = Immediate::getFixedMin();
1338 
1339   /// This records whether all of the fixups using this LSRUse are outside of
1340   /// the loop, in which case some special-case heuristics may be used.
1341   bool AllFixupsOutsideLoop = true;
1342 
1343   /// RigidFormula is set to true to guarantee that this use will be associated
1344   /// with a single formula--the one that initially matched. Some SCEV
1345   /// expressions cannot be expanded. This allows LSR to consider the registers
1346   /// used by those expressions without the need to expand them later after
1347   /// changing the formula.
1348   bool RigidFormula = false;
1349 
1350   /// This records the widest use type for any fixup using this
1351   /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max
1352   /// fixup widths to be equivalent, because the narrower one may be relying on
1353   /// the implicit truncation to truncate away bogus bits.
1354   Type *WidestFixupType = nullptr;
1355 
1356   /// A list of ways to build a value that can satisfy this user.  After the
1357   /// list is populated, one of these is selected heuristically and used to
1358   /// formulate a replacement for OperandValToReplace in UserInst.
1359   SmallVector<Formula, 12> Formulae;
1360 
1361   /// The set of register candidates used by all formulae in this LSRUse.
1362   SmallPtrSet<const SCEV *, 4> Regs;
1363 
LSRUse(KindType K,MemAccessTy AT)1364   LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {}
1365 
getNewFixup()1366   LSRFixup &getNewFixup() {
1367     Fixups.push_back(LSRFixup());
1368     return Fixups.back();
1369   }
1370 
pushFixup(LSRFixup & f)1371   void pushFixup(LSRFixup &f) {
1372     Fixups.push_back(f);
1373     if (Immediate::isKnownGT(f.Offset, MaxOffset))
1374       MaxOffset = f.Offset;
1375     if (Immediate::isKnownLT(f.Offset, MinOffset))
1376       MinOffset = f.Offset;
1377   }
1378 
1379   bool HasFormulaWithSameRegs(const Formula &F) const;
1380   float getNotSelectedProbability(const SCEV *Reg) const;
1381   bool InsertFormula(const Formula &F, const Loop &L);
1382   void DeleteFormula(Formula &F);
1383   void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
1384 
1385   void print(raw_ostream &OS) const;
1386   void dump() const;
1387 };
1388 
1389 } // end anonymous namespace
1390 
1391 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1392                                  LSRUse::KindType Kind, MemAccessTy AccessTy,
1393                                  GlobalValue *BaseGV, Immediate BaseOffset,
1394                                  bool HasBaseReg, int64_t Scale,
1395                                  Instruction *Fixup = nullptr);
1396 
getSetupCost(const SCEV * Reg,unsigned Depth)1397 static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) {
1398   if (isa<SCEVUnknown>(Reg) || isa<SCEVConstant>(Reg))
1399     return 1;
1400   if (Depth == 0)
1401     return 0;
1402   if (const auto *S = dyn_cast<SCEVAddRecExpr>(Reg))
1403     return getSetupCost(S->getStart(), Depth - 1);
1404   if (auto S = dyn_cast<SCEVIntegralCastExpr>(Reg))
1405     return getSetupCost(S->getOperand(), Depth - 1);
1406   if (auto S = dyn_cast<SCEVNAryExpr>(Reg))
1407     return std::accumulate(S->operands().begin(), S->operands().end(), 0,
1408                            [&](unsigned i, const SCEV *Reg) {
1409                              return i + getSetupCost(Reg, Depth - 1);
1410                            });
1411   if (auto S = dyn_cast<SCEVUDivExpr>(Reg))
1412     return getSetupCost(S->getLHS(), Depth - 1) +
1413            getSetupCost(S->getRHS(), Depth - 1);
1414   return 0;
1415 }
1416 
1417 /// Tally up interesting quantities from the given register.
RateRegister(const Formula & F,const SCEV * Reg,SmallPtrSetImpl<const SCEV * > & Regs)1418 void Cost::RateRegister(const Formula &F, const SCEV *Reg,
1419                         SmallPtrSetImpl<const SCEV *> &Regs) {
1420   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
1421     // If this is an addrec for another loop, it should be an invariant
1422     // with respect to L since L is the innermost loop (at least
1423     // for now LSR only handles innermost loops).
1424     if (AR->getLoop() != L) {
1425       // If the AddRec exists, consider it's register free and leave it alone.
1426       if (isExistingPhi(AR, *SE) && AMK != TTI::AMK_PostIndexed)
1427         return;
1428 
1429       // It is bad to allow LSR for current loop to add induction variables
1430       // for its sibling loops.
1431       if (!AR->getLoop()->contains(L)) {
1432         Lose();
1433         return;
1434       }
1435 
1436       // Otherwise, it will be an invariant with respect to Loop L.
1437       ++C.NumRegs;
1438       return;
1439     }
1440 
1441     unsigned LoopCost = 1;
1442     if (TTI->isIndexedLoadLegal(TTI->MIM_PostInc, AR->getType()) ||
1443         TTI->isIndexedStoreLegal(TTI->MIM_PostInc, AR->getType())) {
1444 
1445       // If the step size matches the base offset, we could use pre-indexed
1446       // addressing.
1447       if (AMK == TTI::AMK_PreIndexed && F.BaseOffset.isFixed()) {
1448         if (auto *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)))
1449           if (Step->getAPInt() == F.BaseOffset.getFixedValue())
1450             LoopCost = 0;
1451       } else if (AMK == TTI::AMK_PostIndexed) {
1452         const SCEV *LoopStep = AR->getStepRecurrence(*SE);
1453         if (isa<SCEVConstant>(LoopStep)) {
1454           const SCEV *LoopStart = AR->getStart();
1455           if (!isa<SCEVConstant>(LoopStart) &&
1456               SE->isLoopInvariant(LoopStart, L))
1457             LoopCost = 0;
1458         }
1459       }
1460     }
1461     C.AddRecCost += LoopCost;
1462 
1463     // Add the step value register, if it needs one.
1464     // TODO: The non-affine case isn't precisely modeled here.
1465     if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
1466       if (!Regs.count(AR->getOperand(1))) {
1467         RateRegister(F, AR->getOperand(1), Regs);
1468         if (isLoser())
1469           return;
1470       }
1471     }
1472   }
1473   ++C.NumRegs;
1474 
1475   // Rough heuristic; favor registers which don't require extra setup
1476   // instructions in the preheader.
1477   C.SetupCost += getSetupCost(Reg, SetupCostDepthLimit);
1478   // Ensure we don't, even with the recusion limit, produce invalid costs.
1479   C.SetupCost = std::min<unsigned>(C.SetupCost, 1 << 16);
1480 
1481   C.NumIVMuls += isa<SCEVMulExpr>(Reg) &&
1482                SE->hasComputableLoopEvolution(Reg, L);
1483 }
1484 
1485 /// Record this register in the set. If we haven't seen it before, rate
1486 /// it. Optional LoserRegs provides a way to declare any formula that refers to
1487 /// one of those regs an instant loser.
RatePrimaryRegister(const Formula & F,const SCEV * Reg,SmallPtrSetImpl<const SCEV * > & Regs,SmallPtrSetImpl<const SCEV * > * LoserRegs)1488 void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg,
1489                                SmallPtrSetImpl<const SCEV *> &Regs,
1490                                SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1491   if (LoserRegs && LoserRegs->count(Reg)) {
1492     Lose();
1493     return;
1494   }
1495   if (Regs.insert(Reg).second) {
1496     RateRegister(F, Reg, Regs);
1497     if (LoserRegs && isLoser())
1498       LoserRegs->insert(Reg);
1499   }
1500 }
1501 
RateFormula(const Formula & F,SmallPtrSetImpl<const SCEV * > & Regs,const DenseSet<const SCEV * > & VisitedRegs,const LSRUse & LU,SmallPtrSetImpl<const SCEV * > * LoserRegs)1502 void Cost::RateFormula(const Formula &F,
1503                        SmallPtrSetImpl<const SCEV *> &Regs,
1504                        const DenseSet<const SCEV *> &VisitedRegs,
1505                        const LSRUse &LU,
1506                        SmallPtrSetImpl<const SCEV *> *LoserRegs) {
1507   if (isLoser())
1508     return;
1509   assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula");
1510   // Tally up the registers.
1511   unsigned PrevAddRecCost = C.AddRecCost;
1512   unsigned PrevNumRegs = C.NumRegs;
1513   unsigned PrevNumBaseAdds = C.NumBaseAdds;
1514   if (const SCEV *ScaledReg = F.ScaledReg) {
1515     if (VisitedRegs.count(ScaledReg)) {
1516       Lose();
1517       return;
1518     }
1519     RatePrimaryRegister(F, ScaledReg, Regs, LoserRegs);
1520     if (isLoser())
1521       return;
1522   }
1523   for (const SCEV *BaseReg : F.BaseRegs) {
1524     if (VisitedRegs.count(BaseReg)) {
1525       Lose();
1526       return;
1527     }
1528     RatePrimaryRegister(F, BaseReg, Regs, LoserRegs);
1529     if (isLoser())
1530       return;
1531   }
1532 
1533   // Determine how many (unfolded) adds we'll need inside the loop.
1534   size_t NumBaseParts = F.getNumRegs();
1535   if (NumBaseParts > 1)
1536     // Do not count the base and a possible second register if the target
1537     // allows to fold 2 registers.
1538     C.NumBaseAdds +=
1539         NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(*TTI, LU, F)));
1540   C.NumBaseAdds += (F.UnfoldedOffset.isNonZero());
1541 
1542   // Accumulate non-free scaling amounts.
1543   C.ScaleCost += *getScalingFactorCost(*TTI, LU, F, *L).getValue();
1544 
1545   // Tally up the non-zero immediates.
1546   for (const LSRFixup &Fixup : LU.Fixups) {
1547     if (Fixup.Offset.isCompatibleImmediate(F.BaseOffset)) {
1548       Immediate Offset = Fixup.Offset.addUnsigned(F.BaseOffset);
1549       if (F.BaseGV)
1550         C.ImmCost += 64; // Handle symbolic values conservatively.
1551                          // TODO: This should probably be the pointer size.
1552       else if (Offset.isNonZero())
1553         C.ImmCost +=
1554             APInt(64, Offset.getKnownMinValue(), true).getSignificantBits();
1555 
1556       // Check with target if this offset with this instruction is
1557       // specifically not supported.
1558       if (LU.Kind == LSRUse::Address && Offset.isNonZero() &&
1559           !isAMCompletelyFolded(*TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1560                                 Offset, F.HasBaseReg, F.Scale, Fixup.UserInst))
1561         C.NumBaseAdds++;
1562     } else {
1563       // Incompatible immediate type, increase cost to avoid using
1564       C.ImmCost += 2048;
1565     }
1566   }
1567 
1568   // If we don't count instruction cost exit here.
1569   if (!InsnsCost) {
1570     assert(isValid() && "invalid cost");
1571     return;
1572   }
1573 
1574   // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as
1575   // additional instruction (at least fill).
1576   // TODO: Need distinguish register class?
1577   unsigned TTIRegNum = TTI->getNumberOfRegisters(
1578                        TTI->getRegisterClassForType(false, F.getType())) - 1;
1579   if (C.NumRegs > TTIRegNum) {
1580     // Cost already exceeded TTIRegNum, then only newly added register can add
1581     // new instructions.
1582     if (PrevNumRegs > TTIRegNum)
1583       C.Insns += (C.NumRegs - PrevNumRegs);
1584     else
1585       C.Insns += (C.NumRegs - TTIRegNum);
1586   }
1587 
1588   // If ICmpZero formula ends with not 0, it could not be replaced by
1589   // just add or sub. We'll need to compare final result of AddRec.
1590   // That means we'll need an additional instruction. But if the target can
1591   // macro-fuse a compare with a branch, don't count this extra instruction.
1592   // For -10 + {0, +, 1}:
1593   // i = i + 1;
1594   // cmp i, 10
1595   //
1596   // For {-10, +, 1}:
1597   // i = i + 1;
1598   if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() &&
1599       !TTI->canMacroFuseCmp())
1600     C.Insns++;
1601   // Each new AddRec adds 1 instruction to calculation.
1602   C.Insns += (C.AddRecCost - PrevAddRecCost);
1603 
1604   // BaseAdds adds instructions for unfolded registers.
1605   if (LU.Kind != LSRUse::ICmpZero)
1606     C.Insns += C.NumBaseAdds - PrevNumBaseAdds;
1607   assert(isValid() && "invalid cost");
1608 }
1609 
1610 /// Set this cost to a losing value.
Lose()1611 void Cost::Lose() {
1612   C.Insns = std::numeric_limits<unsigned>::max();
1613   C.NumRegs = std::numeric_limits<unsigned>::max();
1614   C.AddRecCost = std::numeric_limits<unsigned>::max();
1615   C.NumIVMuls = std::numeric_limits<unsigned>::max();
1616   C.NumBaseAdds = std::numeric_limits<unsigned>::max();
1617   C.ImmCost = std::numeric_limits<unsigned>::max();
1618   C.SetupCost = std::numeric_limits<unsigned>::max();
1619   C.ScaleCost = std::numeric_limits<unsigned>::max();
1620 }
1621 
1622 /// Choose the lower cost.
isLess(const Cost & Other) const1623 bool Cost::isLess(const Cost &Other) const {
1624   if (InsnsCost.getNumOccurrences() > 0 && InsnsCost &&
1625       C.Insns != Other.C.Insns)
1626     return C.Insns < Other.C.Insns;
1627   return TTI->isLSRCostLess(C, Other.C);
1628 }
1629 
1630 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print(raw_ostream & OS) const1631 void Cost::print(raw_ostream &OS) const {
1632   if (InsnsCost)
1633     OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s ");
1634   OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s");
1635   if (C.AddRecCost != 0)
1636     OS << ", with addrec cost " << C.AddRecCost;
1637   if (C.NumIVMuls != 0)
1638     OS << ", plus " << C.NumIVMuls << " IV mul"
1639        << (C.NumIVMuls == 1 ? "" : "s");
1640   if (C.NumBaseAdds != 0)
1641     OS << ", plus " << C.NumBaseAdds << " base add"
1642        << (C.NumBaseAdds == 1 ? "" : "s");
1643   if (C.ScaleCost != 0)
1644     OS << ", plus " << C.ScaleCost << " scale cost";
1645   if (C.ImmCost != 0)
1646     OS << ", plus " << C.ImmCost << " imm cost";
1647   if (C.SetupCost != 0)
1648     OS << ", plus " << C.SetupCost << " setup cost";
1649 }
1650 
dump() const1651 LLVM_DUMP_METHOD void Cost::dump() const {
1652   print(errs()); errs() << '\n';
1653 }
1654 #endif
1655 
1656 /// Test whether this fixup always uses its value outside of the given loop.
isUseFullyOutsideLoop(const Loop * L) const1657 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
1658   // PHI nodes use their value in their incoming blocks.
1659   if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
1660     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1661       if (PN->getIncomingValue(i) == OperandValToReplace &&
1662           L->contains(PN->getIncomingBlock(i)))
1663         return false;
1664     return true;
1665   }
1666 
1667   return !L->contains(UserInst);
1668 }
1669 
1670 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print(raw_ostream & OS) const1671 void LSRFixup::print(raw_ostream &OS) const {
1672   OS << "UserInst=";
1673   // Store is common and interesting enough to be worth special-casing.
1674   if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
1675     OS << "store ";
1676     Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
1677   } else if (UserInst->getType()->isVoidTy())
1678     OS << UserInst->getOpcodeName();
1679   else
1680     UserInst->printAsOperand(OS, /*PrintType=*/false);
1681 
1682   OS << ", OperandValToReplace=";
1683   OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
1684 
1685   for (const Loop *PIL : PostIncLoops) {
1686     OS << ", PostIncLoop=";
1687     PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false);
1688   }
1689 
1690   if (Offset.isNonZero())
1691     OS << ", Offset=" << Offset;
1692 }
1693 
dump() const1694 LLVM_DUMP_METHOD void LSRFixup::dump() const {
1695   print(errs()); errs() << '\n';
1696 }
1697 #endif
1698 
1699 /// Test whether this use as a formula which has the same registers as the given
1700 /// formula.
HasFormulaWithSameRegs(const Formula & F) const1701 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
1702   SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1703   if (F.ScaledReg) Key.push_back(F.ScaledReg);
1704   // Unstable sort by host order ok, because this is only used for uniquifying.
1705   llvm::sort(Key);
1706   return Uniquifier.count(Key);
1707 }
1708 
1709 /// The function returns a probability of selecting formula without Reg.
getNotSelectedProbability(const SCEV * Reg) const1710 float LSRUse::getNotSelectedProbability(const SCEV *Reg) const {
1711   unsigned FNum = 0;
1712   for (const Formula &F : Formulae)
1713     if (F.referencesReg(Reg))
1714       FNum++;
1715   return ((float)(Formulae.size() - FNum)) / Formulae.size();
1716 }
1717 
1718 /// If the given formula has not yet been inserted, add it to the list, and
1719 /// return true. Return false otherwise.  The formula must be in canonical form.
InsertFormula(const Formula & F,const Loop & L)1720 bool LSRUse::InsertFormula(const Formula &F, const Loop &L) {
1721   assert(F.isCanonical(L) && "Invalid canonical representation");
1722 
1723   if (!Formulae.empty() && RigidFormula)
1724     return false;
1725 
1726   SmallVector<const SCEV *, 4> Key = F.BaseRegs;
1727   if (F.ScaledReg) Key.push_back(F.ScaledReg);
1728   // Unstable sort by host order ok, because this is only used for uniquifying.
1729   llvm::sort(Key);
1730 
1731   if (!Uniquifier.insert(Key).second)
1732     return false;
1733 
1734   // Using a register to hold the value of 0 is not profitable.
1735   assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
1736          "Zero allocated in a scaled register!");
1737 #ifndef NDEBUG
1738   for (const SCEV *BaseReg : F.BaseRegs)
1739     assert(!BaseReg->isZero() && "Zero allocated in a base register!");
1740 #endif
1741 
1742   // Add the formula to the list.
1743   Formulae.push_back(F);
1744 
1745   // Record registers now being used by this use.
1746   Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1747   if (F.ScaledReg)
1748     Regs.insert(F.ScaledReg);
1749 
1750   return true;
1751 }
1752 
1753 /// Remove the given formula from this use's list.
DeleteFormula(Formula & F)1754 void LSRUse::DeleteFormula(Formula &F) {
1755   if (&F != &Formulae.back())
1756     std::swap(F, Formulae.back());
1757   Formulae.pop_back();
1758 }
1759 
1760 /// Recompute the Regs field, and update RegUses.
RecomputeRegs(size_t LUIdx,RegUseTracker & RegUses)1761 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
1762   // Now that we've filtered out some formulae, recompute the Regs set.
1763   SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs);
1764   Regs.clear();
1765   for (const Formula &F : Formulae) {
1766     if (F.ScaledReg) Regs.insert(F.ScaledReg);
1767     Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
1768   }
1769 
1770   // Update the RegTracker.
1771   for (const SCEV *S : OldRegs)
1772     if (!Regs.count(S))
1773       RegUses.dropRegister(S, LUIdx);
1774 }
1775 
1776 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print(raw_ostream & OS) const1777 void LSRUse::print(raw_ostream &OS) const {
1778   OS << "LSR Use: Kind=";
1779   switch (Kind) {
1780   case Basic:    OS << "Basic"; break;
1781   case Special:  OS << "Special"; break;
1782   case ICmpZero: OS << "ICmpZero"; break;
1783   case Address:
1784     OS << "Address of ";
1785     if (AccessTy.MemTy->isPointerTy())
1786       OS << "pointer"; // the full pointer type could be really verbose
1787     else {
1788       OS << *AccessTy.MemTy;
1789     }
1790 
1791     OS << " in addrspace(" << AccessTy.AddrSpace << ')';
1792   }
1793 
1794   OS << ", Offsets={";
1795   bool NeedComma = false;
1796   for (const LSRFixup &Fixup : Fixups) {
1797     if (NeedComma) OS << ',';
1798     OS << Fixup.Offset;
1799     NeedComma = true;
1800   }
1801   OS << '}';
1802 
1803   if (AllFixupsOutsideLoop)
1804     OS << ", all-fixups-outside-loop";
1805 
1806   if (WidestFixupType)
1807     OS << ", widest fixup type: " << *WidestFixupType;
1808 }
1809 
dump() const1810 LLVM_DUMP_METHOD void LSRUse::dump() const {
1811   print(errs()); errs() << '\n';
1812 }
1813 #endif
1814 
isAMCompletelyFolded(const TargetTransformInfo & TTI,LSRUse::KindType Kind,MemAccessTy AccessTy,GlobalValue * BaseGV,Immediate BaseOffset,bool HasBaseReg,int64_t Scale,Instruction * Fixup)1815 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1816                                  LSRUse::KindType Kind, MemAccessTy AccessTy,
1817                                  GlobalValue *BaseGV, Immediate BaseOffset,
1818                                  bool HasBaseReg, int64_t Scale,
1819                                  Instruction *Fixup /* = nullptr */) {
1820   switch (Kind) {
1821   case LSRUse::Address: {
1822     int64_t FixedOffset =
1823         BaseOffset.isScalable() ? 0 : BaseOffset.getFixedValue();
1824     int64_t ScalableOffset =
1825         BaseOffset.isScalable() ? BaseOffset.getKnownMinValue() : 0;
1826     return TTI.isLegalAddressingMode(AccessTy.MemTy, BaseGV, FixedOffset,
1827                                      HasBaseReg, Scale, AccessTy.AddrSpace,
1828                                      Fixup, ScalableOffset);
1829   }
1830   case LSRUse::ICmpZero:
1831     // There's not even a target hook for querying whether it would be legal to
1832     // fold a GV into an ICmp.
1833     if (BaseGV)
1834       return false;
1835 
1836     // ICmp only has two operands; don't allow more than two non-trivial parts.
1837     if (Scale != 0 && HasBaseReg && BaseOffset.isNonZero())
1838       return false;
1839 
1840     // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1841     // putting the scaled register in the other operand of the icmp.
1842     if (Scale != 0 && Scale != -1)
1843       return false;
1844 
1845     // If we have low-level target information, ask the target if it can fold an
1846     // integer immediate on an icmp.
1847     if (BaseOffset.isNonZero()) {
1848       // We don't have an interface to query whether the target supports
1849       // icmpzero against scalable quantities yet.
1850       if (BaseOffset.isScalable())
1851         return false;
1852 
1853       // We have one of:
1854       // ICmpZero     BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
1855       // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
1856       // Offs is the ICmp immediate.
1857       if (Scale == 0)
1858         // The cast does the right thing with
1859         // std::numeric_limits<int64_t>::min().
1860         BaseOffset = BaseOffset.getFixed(-(uint64_t)BaseOffset.getFixedValue());
1861       return TTI.isLegalICmpImmediate(BaseOffset.getFixedValue());
1862     }
1863 
1864     // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
1865     return true;
1866 
1867   case LSRUse::Basic:
1868     // Only handle single-register values.
1869     return !BaseGV && Scale == 0 && BaseOffset.isZero();
1870 
1871   case LSRUse::Special:
1872     // Special case Basic to handle -1 scales.
1873     return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset.isZero();
1874   }
1875 
1876   llvm_unreachable("Invalid LSRUse Kind!");
1877 }
1878 
isAMCompletelyFolded(const TargetTransformInfo & TTI,Immediate MinOffset,Immediate MaxOffset,LSRUse::KindType Kind,MemAccessTy AccessTy,GlobalValue * BaseGV,Immediate BaseOffset,bool HasBaseReg,int64_t Scale)1879 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1880                                  Immediate MinOffset, Immediate MaxOffset,
1881                                  LSRUse::KindType Kind, MemAccessTy AccessTy,
1882                                  GlobalValue *BaseGV, Immediate BaseOffset,
1883                                  bool HasBaseReg, int64_t Scale) {
1884   if (BaseOffset.isNonZero() &&
1885       (BaseOffset.isScalable() != MinOffset.isScalable() ||
1886        BaseOffset.isScalable() != MaxOffset.isScalable()))
1887     return false;
1888   // Check for overflow.
1889   int64_t Base = BaseOffset.getKnownMinValue();
1890   int64_t Min = MinOffset.getKnownMinValue();
1891   int64_t Max = MaxOffset.getKnownMinValue();
1892   if (((int64_t)((uint64_t)Base + Min) > Base) != (Min > 0))
1893     return false;
1894   MinOffset = Immediate::get((uint64_t)Base + Min, MinOffset.isScalable());
1895   if (((int64_t)((uint64_t)Base + Max) > Base) != (Max > 0))
1896     return false;
1897   MaxOffset = Immediate::get((uint64_t)Base + Max, MaxOffset.isScalable());
1898 
1899   return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset,
1900                               HasBaseReg, Scale) &&
1901          isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset,
1902                               HasBaseReg, Scale);
1903 }
1904 
isAMCompletelyFolded(const TargetTransformInfo & TTI,Immediate MinOffset,Immediate MaxOffset,LSRUse::KindType Kind,MemAccessTy AccessTy,const Formula & F,const Loop & L)1905 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1906                                  Immediate MinOffset, Immediate MaxOffset,
1907                                  LSRUse::KindType Kind, MemAccessTy AccessTy,
1908                                  const Formula &F, const Loop &L) {
1909   // For the purpose of isAMCompletelyFolded either having a canonical formula
1910   // or a scale not equal to zero is correct.
1911   // Problems may arise from non canonical formulae having a scale == 0.
1912   // Strictly speaking it would best to just rely on canonical formulae.
1913   // However, when we generate the scaled formulae, we first check that the
1914   // scaling factor is profitable before computing the actual ScaledReg for
1915   // compile time sake.
1916   assert((F.isCanonical(L) || F.Scale != 0));
1917   return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1918                               F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale);
1919 }
1920 
1921 /// Test whether we know how to expand the current formula.
isLegalUse(const TargetTransformInfo & TTI,Immediate MinOffset,Immediate MaxOffset,LSRUse::KindType Kind,MemAccessTy AccessTy,GlobalValue * BaseGV,Immediate BaseOffset,bool HasBaseReg,int64_t Scale)1922 static bool isLegalUse(const TargetTransformInfo &TTI, Immediate MinOffset,
1923                        Immediate MaxOffset, LSRUse::KindType Kind,
1924                        MemAccessTy AccessTy, GlobalValue *BaseGV,
1925                        Immediate BaseOffset, bool HasBaseReg, int64_t Scale) {
1926   // We know how to expand completely foldable formulae.
1927   return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
1928                               BaseOffset, HasBaseReg, Scale) ||
1929          // Or formulae that use a base register produced by a sum of base
1930          // registers.
1931          (Scale == 1 &&
1932           isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy,
1933                                BaseGV, BaseOffset, true, 0));
1934 }
1935 
isLegalUse(const TargetTransformInfo & TTI,Immediate MinOffset,Immediate MaxOffset,LSRUse::KindType Kind,MemAccessTy AccessTy,const Formula & F)1936 static bool isLegalUse(const TargetTransformInfo &TTI, Immediate MinOffset,
1937                        Immediate MaxOffset, LSRUse::KindType Kind,
1938                        MemAccessTy AccessTy, const Formula &F) {
1939   return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
1940                     F.BaseOffset, F.HasBaseReg, F.Scale);
1941 }
1942 
isLegalAddImmediate(const TargetTransformInfo & TTI,Immediate Offset)1943 static bool isLegalAddImmediate(const TargetTransformInfo &TTI,
1944                                 Immediate Offset) {
1945   if (Offset.isScalable())
1946     return TTI.isLegalAddScalableImmediate(Offset.getKnownMinValue());
1947 
1948   return TTI.isLegalAddImmediate(Offset.getFixedValue());
1949 }
1950 
isAMCompletelyFolded(const TargetTransformInfo & TTI,const LSRUse & LU,const Formula & F)1951 static bool isAMCompletelyFolded(const TargetTransformInfo &TTI,
1952                                  const LSRUse &LU, const Formula &F) {
1953   // Target may want to look at the user instructions.
1954   if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) {
1955     for (const LSRFixup &Fixup : LU.Fixups)
1956       if (!isAMCompletelyFolded(TTI, LSRUse::Address, LU.AccessTy, F.BaseGV,
1957                                 (F.BaseOffset + Fixup.Offset), F.HasBaseReg,
1958                                 F.Scale, Fixup.UserInst))
1959         return false;
1960     return true;
1961   }
1962 
1963   return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1964                               LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg,
1965                               F.Scale);
1966 }
1967 
getScalingFactorCost(const TargetTransformInfo & TTI,const LSRUse & LU,const Formula & F,const Loop & L)1968 static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI,
1969                                             const LSRUse &LU, const Formula &F,
1970                                             const Loop &L) {
1971   if (!F.Scale)
1972     return 0;
1973 
1974   // If the use is not completely folded in that instruction, we will have to
1975   // pay an extra cost only for scale != 1.
1976   if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
1977                             LU.AccessTy, F, L))
1978     return F.Scale != 1;
1979 
1980   switch (LU.Kind) {
1981   case LSRUse::Address: {
1982     // Check the scaling factor cost with both the min and max offsets.
1983     int64_t ScalableMin = 0, ScalableMax = 0, FixedMin = 0, FixedMax = 0;
1984     if (F.BaseOffset.isScalable()) {
1985       ScalableMin = (F.BaseOffset + LU.MinOffset).getKnownMinValue();
1986       ScalableMax = (F.BaseOffset + LU.MaxOffset).getKnownMinValue();
1987     } else {
1988       FixedMin = (F.BaseOffset + LU.MinOffset).getFixedValue();
1989       FixedMax = (F.BaseOffset + LU.MaxOffset).getFixedValue();
1990     }
1991     InstructionCost ScaleCostMinOffset = TTI.getScalingFactorCost(
1992         LU.AccessTy.MemTy, F.BaseGV, StackOffset::get(FixedMin, ScalableMin),
1993         F.HasBaseReg, F.Scale, LU.AccessTy.AddrSpace);
1994     InstructionCost ScaleCostMaxOffset = TTI.getScalingFactorCost(
1995         LU.AccessTy.MemTy, F.BaseGV, StackOffset::get(FixedMax, ScalableMax),
1996         F.HasBaseReg, F.Scale, LU.AccessTy.AddrSpace);
1997 
1998     assert(ScaleCostMinOffset.isValid() && ScaleCostMaxOffset.isValid() &&
1999            "Legal addressing mode has an illegal cost!");
2000     return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
2001   }
2002   case LSRUse::ICmpZero:
2003   case LSRUse::Basic:
2004   case LSRUse::Special:
2005     // The use is completely folded, i.e., everything is folded into the
2006     // instruction.
2007     return 0;
2008   }
2009 
2010   llvm_unreachable("Invalid LSRUse Kind!");
2011 }
2012 
isAlwaysFoldable(const TargetTransformInfo & TTI,LSRUse::KindType Kind,MemAccessTy AccessTy,GlobalValue * BaseGV,Immediate BaseOffset,bool HasBaseReg)2013 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
2014                              LSRUse::KindType Kind, MemAccessTy AccessTy,
2015                              GlobalValue *BaseGV, Immediate BaseOffset,
2016                              bool HasBaseReg) {
2017   // Fast-path: zero is always foldable.
2018   if (BaseOffset.isZero() && !BaseGV)
2019     return true;
2020 
2021   // Conservatively, create an address with an immediate and a
2022   // base and a scale.
2023   int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
2024 
2025   // Canonicalize a scale of 1 to a base register if the formula doesn't
2026   // already have a base register.
2027   if (!HasBaseReg && Scale == 1) {
2028     Scale = 0;
2029     HasBaseReg = true;
2030   }
2031 
2032   // FIXME: Try with + without a scale? Maybe based on TTI?
2033   // I think basereg + scaledreg + immediateoffset isn't a good 'conservative'
2034   // default for many architectures, not just AArch64 SVE. More investigation
2035   // needed later to determine if this should be used more widely than just
2036   // on scalable types.
2037   if (HasBaseReg && BaseOffset.isNonZero() && Kind != LSRUse::ICmpZero &&
2038       AccessTy.MemTy && AccessTy.MemTy->isScalableTy() && DropScaledForVScale)
2039     Scale = 0;
2040 
2041   return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset,
2042                               HasBaseReg, Scale);
2043 }
2044 
isAlwaysFoldable(const TargetTransformInfo & TTI,ScalarEvolution & SE,Immediate MinOffset,Immediate MaxOffset,LSRUse::KindType Kind,MemAccessTy AccessTy,const SCEV * S,bool HasBaseReg)2045 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
2046                              ScalarEvolution &SE, Immediate MinOffset,
2047                              Immediate MaxOffset, LSRUse::KindType Kind,
2048                              MemAccessTy AccessTy, const SCEV *S,
2049                              bool HasBaseReg) {
2050   // Fast-path: zero is always foldable.
2051   if (S->isZero()) return true;
2052 
2053   // Conservatively, create an address with an immediate and a
2054   // base and a scale.
2055   Immediate BaseOffset = ExtractImmediate(S, SE);
2056   GlobalValue *BaseGV = ExtractSymbol(S, SE);
2057 
2058   // If there's anything else involved, it's not foldable.
2059   if (!S->isZero()) return false;
2060 
2061   // Fast-path: zero is always foldable.
2062   if (BaseOffset.isZero() && !BaseGV)
2063     return true;
2064 
2065   if (BaseOffset.isScalable())
2066     return false;
2067 
2068   // Conservatively, create an address with an immediate and a
2069   // base and a scale.
2070   int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
2071 
2072   return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
2073                               BaseOffset, HasBaseReg, Scale);
2074 }
2075 
2076 namespace {
2077 
2078 /// An individual increment in a Chain of IV increments.  Relate an IV user to
2079 /// an expression that computes the IV it uses from the IV used by the previous
2080 /// link in the Chain.
2081 ///
2082 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
2083 /// original IVOperand. The head of the chain's IVOperand is only valid during
2084 /// chain collection, before LSR replaces IV users. During chain generation,
2085 /// IncExpr can be used to find the new IVOperand that computes the same
2086 /// expression.
2087 struct IVInc {
2088   Instruction *UserInst;
2089   Value* IVOperand;
2090   const SCEV *IncExpr;
2091 
IVInc__anonc21373340a11::IVInc2092   IVInc(Instruction *U, Value *O, const SCEV *E)
2093       : UserInst(U), IVOperand(O), IncExpr(E) {}
2094 };
2095 
2096 // The list of IV increments in program order.  We typically add the head of a
2097 // chain without finding subsequent links.
2098 struct IVChain {
2099   SmallVector<IVInc, 1> Incs;
2100   const SCEV *ExprBase = nullptr;
2101 
2102   IVChain() = default;
IVChain__anonc21373340a11::IVChain2103   IVChain(const IVInc &Head, const SCEV *Base)
2104       : Incs(1, Head), ExprBase(Base) {}
2105 
2106   using const_iterator = SmallVectorImpl<IVInc>::const_iterator;
2107 
2108   // Return the first increment in the chain.
begin__anonc21373340a11::IVChain2109   const_iterator begin() const {
2110     assert(!Incs.empty());
2111     return std::next(Incs.begin());
2112   }
end__anonc21373340a11::IVChain2113   const_iterator end() const {
2114     return Incs.end();
2115   }
2116 
2117   // Returns true if this chain contains any increments.
hasIncs__anonc21373340a11::IVChain2118   bool hasIncs() const { return Incs.size() >= 2; }
2119 
2120   // Add an IVInc to the end of this chain.
add__anonc21373340a11::IVChain2121   void add(const IVInc &X) { Incs.push_back(X); }
2122 
2123   // Returns the last UserInst in the chain.
tailUserInst__anonc21373340a11::IVChain2124   Instruction *tailUserInst() const { return Incs.back().UserInst; }
2125 
2126   // Returns true if IncExpr can be profitably added to this chain.
2127   bool isProfitableIncrement(const SCEV *OperExpr,
2128                              const SCEV *IncExpr,
2129                              ScalarEvolution&);
2130 };
2131 
2132 /// Helper for CollectChains to track multiple IV increment uses.  Distinguish
2133 /// between FarUsers that definitely cross IV increments and NearUsers that may
2134 /// be used between IV increments.
2135 struct ChainUsers {
2136   SmallPtrSet<Instruction*, 4> FarUsers;
2137   SmallPtrSet<Instruction*, 4> NearUsers;
2138 };
2139 
2140 /// This class holds state for the main loop strength reduction logic.
2141 class LSRInstance {
2142   IVUsers &IU;
2143   ScalarEvolution &SE;
2144   DominatorTree &DT;
2145   LoopInfo &LI;
2146   AssumptionCache &AC;
2147   TargetLibraryInfo &TLI;
2148   const TargetTransformInfo &TTI;
2149   Loop *const L;
2150   MemorySSAUpdater *MSSAU;
2151   TTI::AddressingModeKind AMK;
2152   mutable SCEVExpander Rewriter;
2153   bool Changed = false;
2154 
2155   /// This is the insert position that the current loop's induction variable
2156   /// increment should be placed. In simple loops, this is the latch block's
2157   /// terminator. But in more complicated cases, this is a position which will
2158   /// dominate all the in-loop post-increment users.
2159   Instruction *IVIncInsertPos = nullptr;
2160 
2161   /// Interesting factors between use strides.
2162   ///
2163   /// We explicitly use a SetVector which contains a SmallSet, instead of the
2164   /// default, a SmallDenseSet, because we need to use the full range of
2165   /// int64_ts, and there's currently no good way of doing that with
2166   /// SmallDenseSet.
2167   SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors;
2168 
2169   /// The cost of the current SCEV, the best solution by LSR will be dropped if
2170   /// the solution is not profitable.
2171   Cost BaselineCost;
2172 
2173   /// Interesting use types, to facilitate truncation reuse.
2174   SmallSetVector<Type *, 4> Types;
2175 
2176   /// The list of interesting uses.
2177   mutable SmallVector<LSRUse, 16> Uses;
2178 
2179   /// Track which uses use which register candidates.
2180   RegUseTracker RegUses;
2181 
2182   // Limit the number of chains to avoid quadratic behavior. We don't expect to
2183   // have more than a few IV increment chains in a loop. Missing a Chain falls
2184   // back to normal LSR behavior for those uses.
2185   static const unsigned MaxChains = 8;
2186 
2187   /// IV users can form a chain of IV increments.
2188   SmallVector<IVChain, MaxChains> IVChainVec;
2189 
2190   /// IV users that belong to profitable IVChains.
2191   SmallPtrSet<Use*, MaxChains> IVIncSet;
2192 
2193   /// Induction variables that were generated and inserted by the SCEV Expander.
2194   SmallVector<llvm::WeakVH, 2> ScalarEvolutionIVs;
2195 
2196   void OptimizeShadowIV();
2197   bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
2198   ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
2199   void OptimizeLoopTermCond();
2200 
2201   void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
2202                         SmallVectorImpl<ChainUsers> &ChainUsersVec);
2203   void FinalizeChain(IVChain &Chain);
2204   void CollectChains();
2205   void GenerateIVChain(const IVChain &Chain,
2206                        SmallVectorImpl<WeakTrackingVH> &DeadInsts);
2207 
2208   void CollectInterestingTypesAndFactors();
2209   void CollectFixupsAndInitialFormulae();
2210 
2211   // Support for sharing of LSRUses between LSRFixups.
2212   using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>;
2213   UseMapTy UseMap;
2214 
2215   bool reconcileNewOffset(LSRUse &LU, Immediate NewOffset, bool HasBaseReg,
2216                           LSRUse::KindType Kind, MemAccessTy AccessTy);
2217 
2218   std::pair<size_t, Immediate> getUse(const SCEV *&Expr, LSRUse::KindType Kind,
2219                                       MemAccessTy AccessTy);
2220 
2221   void DeleteUse(LSRUse &LU, size_t LUIdx);
2222 
2223   LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
2224 
2225   void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2226   void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
2227   void CountRegisters(const Formula &F, size_t LUIdx);
2228   bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
2229 
2230   void CollectLoopInvariantFixupsAndFormulae();
2231 
2232   void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
2233                               unsigned Depth = 0);
2234 
2235   void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
2236                                   const Formula &Base, unsigned Depth,
2237                                   size_t Idx, bool IsScaledReg = false);
2238   void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
2239   void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2240                                    const Formula &Base, size_t Idx,
2241                                    bool IsScaledReg = false);
2242   void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2243   void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx,
2244                                    const Formula &Base,
2245                                    const SmallVectorImpl<Immediate> &Worklist,
2246                                    size_t Idx, bool IsScaledReg = false);
2247   void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
2248   void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2249   void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
2250   void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
2251   void GenerateCrossUseConstantOffsets();
2252   void GenerateAllReuseFormulae();
2253 
2254   void FilterOutUndesirableDedicatedRegisters();
2255 
2256   size_t EstimateSearchSpaceComplexity() const;
2257   void NarrowSearchSpaceByDetectingSupersets();
2258   void NarrowSearchSpaceByCollapsingUnrolledCode();
2259   void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
2260   void NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
2261   void NarrowSearchSpaceByFilterPostInc();
2262   void NarrowSearchSpaceByDeletingCostlyFormulas();
2263   void NarrowSearchSpaceByPickingWinnerRegs();
2264   void NarrowSearchSpaceUsingHeuristics();
2265 
2266   void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2267                     Cost &SolutionCost,
2268                     SmallVectorImpl<const Formula *> &Workspace,
2269                     const Cost &CurCost,
2270                     const SmallPtrSet<const SCEV *, 16> &CurRegs,
2271                     DenseSet<const SCEV *> &VisitedRegs) const;
2272   void Solve(SmallVectorImpl<const Formula *> &Solution) const;
2273 
2274   BasicBlock::iterator
2275   HoistInsertPosition(BasicBlock::iterator IP,
2276                       const SmallVectorImpl<Instruction *> &Inputs) const;
2277   BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2278                                                      const LSRFixup &LF,
2279                                                      const LSRUse &LU) const;
2280 
2281   Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2282                 BasicBlock::iterator IP,
2283                 SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2284   void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF,
2285                      const Formula &F,
2286                      SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2287   void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F,
2288                SmallVectorImpl<WeakTrackingVH> &DeadInsts) const;
2289   void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution);
2290 
2291 public:
2292   LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT,
2293               LoopInfo &LI, const TargetTransformInfo &TTI, AssumptionCache &AC,
2294               TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU);
2295 
getChanged() const2296   bool getChanged() const { return Changed; }
getScalarEvolutionIVs() const2297   const SmallVectorImpl<WeakVH> &getScalarEvolutionIVs() const {
2298     return ScalarEvolutionIVs;
2299   }
2300 
2301   void print_factors_and_types(raw_ostream &OS) const;
2302   void print_fixups(raw_ostream &OS) const;
2303   void print_uses(raw_ostream &OS) const;
2304   void print(raw_ostream &OS) const;
2305   void dump() const;
2306 };
2307 
2308 } // end anonymous namespace
2309 
2310 /// If IV is used in a int-to-float cast inside the loop then try to eliminate
2311 /// the cast operation.
OptimizeShadowIV()2312 void LSRInstance::OptimizeShadowIV() {
2313   const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2314   if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2315     return;
2316 
2317   for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
2318        UI != E; /* empty */) {
2319     IVUsers::const_iterator CandidateUI = UI;
2320     ++UI;
2321     Instruction *ShadowUse = CandidateUI->getUser();
2322     Type *DestTy = nullptr;
2323     bool IsSigned = false;
2324 
2325     /* If shadow use is a int->float cast then insert a second IV
2326        to eliminate this cast.
2327 
2328          for (unsigned i = 0; i < n; ++i)
2329            foo((double)i);
2330 
2331        is transformed into
2332 
2333          double d = 0.0;
2334          for (unsigned i = 0; i < n; ++i, ++d)
2335            foo(d);
2336     */
2337     if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
2338       IsSigned = false;
2339       DestTy = UCast->getDestTy();
2340     }
2341     else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
2342       IsSigned = true;
2343       DestTy = SCast->getDestTy();
2344     }
2345     if (!DestTy) continue;
2346 
2347     // If target does not support DestTy natively then do not apply
2348     // this transformation.
2349     if (!TTI.isTypeLegal(DestTy)) continue;
2350 
2351     PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2352     if (!PH) continue;
2353     if (PH->getNumIncomingValues() != 2) continue;
2354 
2355     // If the calculation in integers overflows, the result in FP type will
2356     // differ. So we only can do this transformation if we are guaranteed to not
2357     // deal with overflowing values
2358     const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PH));
2359     if (!AR) continue;
2360     if (IsSigned && !AR->hasNoSignedWrap()) continue;
2361     if (!IsSigned && !AR->hasNoUnsignedWrap()) continue;
2362 
2363     Type *SrcTy = PH->getType();
2364     int Mantissa = DestTy->getFPMantissaWidth();
2365     if (Mantissa == -1) continue;
2366     if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
2367       continue;
2368 
2369     unsigned Entry, Latch;
2370     if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2371       Entry = 0;
2372       Latch = 1;
2373     } else {
2374       Entry = 1;
2375       Latch = 0;
2376     }
2377 
2378     ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2379     if (!Init) continue;
2380     Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
2381                                         (double)Init->getSExtValue() :
2382                                         (double)Init->getZExtValue());
2383 
2384     BinaryOperator *Incr =
2385       dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2386     if (!Incr) continue;
2387     if (Incr->getOpcode() != Instruction::Add
2388         && Incr->getOpcode() != Instruction::Sub)
2389       continue;
2390 
2391     /* Initialize new IV, double d = 0.0 in above example. */
2392     ConstantInt *C = nullptr;
2393     if (Incr->getOperand(0) == PH)
2394       C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2395     else if (Incr->getOperand(1) == PH)
2396       C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2397     else
2398       continue;
2399 
2400     if (!C) continue;
2401 
2402     // Ignore negative constants, as the code below doesn't handle them
2403     // correctly. TODO: Remove this restriction.
2404     if (!C->getValue().isStrictlyPositive())
2405       continue;
2406 
2407     /* Add new PHINode. */
2408     PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH->getIterator());
2409     NewPH->setDebugLoc(PH->getDebugLoc());
2410 
2411     /* create new increment. '++d' in above example. */
2412     Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2413     BinaryOperator *NewIncr = BinaryOperator::Create(
2414         Incr->getOpcode() == Instruction::Add ? Instruction::FAdd
2415                                               : Instruction::FSub,
2416         NewPH, CFP, "IV.S.next.", Incr->getIterator());
2417     NewIncr->setDebugLoc(Incr->getDebugLoc());
2418 
2419     NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2420     NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2421 
2422     /* Remove cast operation */
2423     ShadowUse->replaceAllUsesWith(NewPH);
2424     ShadowUse->eraseFromParent();
2425     Changed = true;
2426     break;
2427   }
2428 }
2429 
2430 /// If Cond has an operand that is an expression of an IV, set the IV user and
2431 /// stride information and return true, otherwise return false.
FindIVUserForCond(ICmpInst * Cond,IVStrideUse * & CondUse)2432 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
2433   for (IVStrideUse &U : IU)
2434     if (U.getUser() == Cond) {
2435       // NOTE: we could handle setcc instructions with multiple uses here, but
2436       // InstCombine does it as well for simple uses, it's not clear that it
2437       // occurs enough in real life to handle.
2438       CondUse = &U;
2439       return true;
2440     }
2441   return false;
2442 }
2443 
2444 /// Rewrite the loop's terminating condition if it uses a max computation.
2445 ///
2446 /// This is a narrow solution to a specific, but acute, problem. For loops
2447 /// like this:
2448 ///
2449 ///   i = 0;
2450 ///   do {
2451 ///     p[i] = 0.0;
2452 ///   } while (++i < n);
2453 ///
2454 /// the trip count isn't just 'n', because 'n' might not be positive. And
2455 /// unfortunately this can come up even for loops where the user didn't use
2456 /// a C do-while loop. For example, seemingly well-behaved top-test loops
2457 /// will commonly be lowered like this:
2458 ///
2459 ///   if (n > 0) {
2460 ///     i = 0;
2461 ///     do {
2462 ///       p[i] = 0.0;
2463 ///     } while (++i < n);
2464 ///   }
2465 ///
2466 /// and then it's possible for subsequent optimization to obscure the if
2467 /// test in such a way that indvars can't find it.
2468 ///
2469 /// When indvars can't find the if test in loops like this, it creates a
2470 /// max expression, which allows it to give the loop a canonical
2471 /// induction variable:
2472 ///
2473 ///   i = 0;
2474 ///   max = n < 1 ? 1 : n;
2475 ///   do {
2476 ///     p[i] = 0.0;
2477 ///   } while (++i != max);
2478 ///
2479 /// Canonical induction variables are necessary because the loop passes
2480 /// are designed around them. The most obvious example of this is the
2481 /// LoopInfo analysis, which doesn't remember trip count values. It
2482 /// expects to be able to rediscover the trip count each time it is
2483 /// needed, and it does this using a simple analysis that only succeeds if
2484 /// the loop has a canonical induction variable.
2485 ///
2486 /// However, when it comes time to generate code, the maximum operation
2487 /// can be quite costly, especially if it's inside of an outer loop.
2488 ///
2489 /// This function solves this problem by detecting this type of loop and
2490 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2491 /// the instructions for the maximum computation.
OptimizeMax(ICmpInst * Cond,IVStrideUse * & CondUse)2492 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
2493   // Check that the loop matches the pattern we're looking for.
2494   if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2495       Cond->getPredicate() != CmpInst::ICMP_NE)
2496     return Cond;
2497 
2498   SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2499   if (!Sel || !Sel->hasOneUse()) return Cond;
2500 
2501   const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
2502   if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2503     return Cond;
2504   const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
2505 
2506   // Add one to the backedge-taken count to get the trip count.
2507   const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
2508   if (IterationCount != SE.getSCEV(Sel)) return Cond;
2509 
2510   // Check for a max calculation that matches the pattern. There's no check
2511   // for ICMP_ULE here because the comparison would be with zero, which
2512   // isn't interesting.
2513   CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
2514   const SCEVNAryExpr *Max = nullptr;
2515   if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
2516     Pred = ICmpInst::ICMP_SLE;
2517     Max = S;
2518   } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
2519     Pred = ICmpInst::ICMP_SLT;
2520     Max = S;
2521   } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
2522     Pred = ICmpInst::ICMP_ULT;
2523     Max = U;
2524   } else {
2525     // No match; bail.
2526     return Cond;
2527   }
2528 
2529   // To handle a max with more than two operands, this optimization would
2530   // require additional checking and setup.
2531   if (Max->getNumOperands() != 2)
2532     return Cond;
2533 
2534   const SCEV *MaxLHS = Max->getOperand(0);
2535   const SCEV *MaxRHS = Max->getOperand(1);
2536 
2537   // ScalarEvolution canonicalizes constants to the left. For < and >, look
2538   // for a comparison with 1. For <= and >=, a comparison with zero.
2539   if (!MaxLHS ||
2540       (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
2541     return Cond;
2542 
2543   // Check the relevant induction variable for conformance to
2544   // the pattern.
2545   const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
2546   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2547   if (!AR || !AR->isAffine() ||
2548       AR->getStart() != One ||
2549       AR->getStepRecurrence(SE) != One)
2550     return Cond;
2551 
2552   assert(AR->getLoop() == L &&
2553          "Loop condition operand is an addrec in a different loop!");
2554 
2555   // Check the right operand of the select, and remember it, as it will
2556   // be used in the new comparison instruction.
2557   Value *NewRHS = nullptr;
2558   if (ICmpInst::isTrueWhenEqual(Pred)) {
2559     // Look for n+1, and grab n.
2560     if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
2561       if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2562          if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2563            NewRHS = BO->getOperand(0);
2564     if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
2565       if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
2566         if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
2567           NewRHS = BO->getOperand(0);
2568     if (!NewRHS)
2569       return Cond;
2570   } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
2571     NewRHS = Sel->getOperand(1);
2572   else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
2573     NewRHS = Sel->getOperand(2);
2574   else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
2575     NewRHS = SU->getValue();
2576   else
2577     // Max doesn't match expected pattern.
2578     return Cond;
2579 
2580   // Determine the new comparison opcode. It may be signed or unsigned,
2581   // and the original comparison may be either equality or inequality.
2582   if (Cond->getPredicate() == CmpInst::ICMP_EQ)
2583     Pred = CmpInst::getInversePredicate(Pred);
2584 
2585   // Ok, everything looks ok to change the condition into an SLT or SGE and
2586   // delete the max calculation.
2587   ICmpInst *NewCond = new ICmpInst(Cond->getIterator(), Pred,
2588                                    Cond->getOperand(0), NewRHS, "scmp");
2589 
2590   // Delete the max calculation instructions.
2591   NewCond->setDebugLoc(Cond->getDebugLoc());
2592   Cond->replaceAllUsesWith(NewCond);
2593   CondUse->setUser(NewCond);
2594   Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2595   Cond->eraseFromParent();
2596   Sel->eraseFromParent();
2597   if (Cmp->use_empty())
2598     Cmp->eraseFromParent();
2599   return NewCond;
2600 }
2601 
2602 /// Change loop terminating condition to use the postinc iv when possible.
2603 void
OptimizeLoopTermCond()2604 LSRInstance::OptimizeLoopTermCond() {
2605   SmallPtrSet<Instruction *, 4> PostIncs;
2606 
2607   // We need a different set of heuristics for rotated and non-rotated loops.
2608   // If a loop is rotated then the latch is also the backedge, so inserting
2609   // post-inc expressions just before the latch is ideal. To reduce live ranges
2610   // it also makes sense to rewrite terminating conditions to use post-inc
2611   // expressions.
2612   //
2613   // If the loop is not rotated then the latch is not a backedge; the latch
2614   // check is done in the loop head. Adding post-inc expressions before the
2615   // latch will cause overlapping live-ranges of pre-inc and post-inc expressions
2616   // in the loop body. In this case we do *not* want to use post-inc expressions
2617   // in the latch check, and we want to insert post-inc expressions before
2618   // the backedge.
2619   BasicBlock *LatchBlock = L->getLoopLatch();
2620   SmallVector<BasicBlock*, 8> ExitingBlocks;
2621   L->getExitingBlocks(ExitingBlocks);
2622   if (!llvm::is_contained(ExitingBlocks, LatchBlock)) {
2623     // The backedge doesn't exit the loop; treat this as a head-tested loop.
2624     IVIncInsertPos = LatchBlock->getTerminator();
2625     return;
2626   }
2627 
2628   // Otherwise treat this as a rotated loop.
2629   for (BasicBlock *ExitingBlock : ExitingBlocks) {
2630     // Get the terminating condition for the loop if possible.  If we
2631     // can, we want to change it to use a post-incremented version of its
2632     // induction variable, to allow coalescing the live ranges for the IV into
2633     // one register value.
2634 
2635     BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
2636     if (!TermBr)
2637       continue;
2638     // FIXME: Overly conservative, termination condition could be an 'or' etc..
2639     if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
2640       continue;
2641 
2642     // Search IVUsesByStride to find Cond's IVUse if there is one.
2643     IVStrideUse *CondUse = nullptr;
2644     ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2645     if (!FindIVUserForCond(Cond, CondUse))
2646       continue;
2647 
2648     // If the trip count is computed in terms of a max (due to ScalarEvolution
2649     // being unable to find a sufficient guard, for example), change the loop
2650     // comparison to use SLT or ULT instead of NE.
2651     // One consequence of doing this now is that it disrupts the count-down
2652     // optimization. That's not always a bad thing though, because in such
2653     // cases it may still be worthwhile to avoid a max.
2654     Cond = OptimizeMax(Cond, CondUse);
2655 
2656     // If this exiting block dominates the latch block, it may also use
2657     // the post-inc value if it won't be shared with other uses.
2658     // Check for dominance.
2659     if (!DT.dominates(ExitingBlock, LatchBlock))
2660       continue;
2661 
2662     // Conservatively avoid trying to use the post-inc value in non-latch
2663     // exits if there may be pre-inc users in intervening blocks.
2664     if (LatchBlock != ExitingBlock)
2665       for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
2666         // Test if the use is reachable from the exiting block. This dominator
2667         // query is a conservative approximation of reachability.
2668         if (&*UI != CondUse &&
2669             !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
2670           // Conservatively assume there may be reuse if the quotient of their
2671           // strides could be a legal scale.
2672           const SCEV *A = IU.getStride(*CondUse, L);
2673           const SCEV *B = IU.getStride(*UI, L);
2674           if (!A || !B) continue;
2675           if (SE.getTypeSizeInBits(A->getType()) !=
2676               SE.getTypeSizeInBits(B->getType())) {
2677             if (SE.getTypeSizeInBits(A->getType()) >
2678                 SE.getTypeSizeInBits(B->getType()))
2679               B = SE.getSignExtendExpr(B, A->getType());
2680             else
2681               A = SE.getSignExtendExpr(A, B->getType());
2682           }
2683           if (const SCEVConstant *D =
2684                 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
2685             const ConstantInt *C = D->getValue();
2686             // Stride of one or negative one can have reuse with non-addresses.
2687             if (C->isOne() || C->isMinusOne())
2688               goto decline_post_inc;
2689             // Avoid weird situations.
2690             if (C->getValue().getSignificantBits() >= 64 ||
2691                 C->getValue().isMinSignedValue())
2692               goto decline_post_inc;
2693             // Check for possible scaled-address reuse.
2694             if (isAddressUse(TTI, UI->getUser(), UI->getOperandValToReplace())) {
2695               MemAccessTy AccessTy = getAccessType(
2696                   TTI, UI->getUser(), UI->getOperandValToReplace());
2697               int64_t Scale = C->getSExtValue();
2698               if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2699                                             /*BaseOffset=*/0,
2700                                             /*HasBaseReg=*/true, Scale,
2701                                             AccessTy.AddrSpace))
2702                 goto decline_post_inc;
2703               Scale = -Scale;
2704               if (TTI.isLegalAddressingMode(AccessTy.MemTy, /*BaseGV=*/nullptr,
2705                                             /*BaseOffset=*/0,
2706                                             /*HasBaseReg=*/true, Scale,
2707                                             AccessTy.AddrSpace))
2708                 goto decline_post_inc;
2709             }
2710           }
2711         }
2712 
2713     LLVM_DEBUG(dbgs() << "  Change loop exiting icmp to use postinc iv: "
2714                       << *Cond << '\n');
2715 
2716     // It's possible for the setcc instruction to be anywhere in the loop, and
2717     // possible for it to have multiple users.  If it is not immediately before
2718     // the exiting block branch, move it.
2719     if (Cond->getNextNonDebugInstruction() != TermBr) {
2720       if (Cond->hasOneUse()) {
2721         Cond->moveBefore(TermBr);
2722       } else {
2723         // Clone the terminating condition and insert into the loopend.
2724         ICmpInst *OldCond = Cond;
2725         Cond = cast<ICmpInst>(Cond->clone());
2726         Cond->setName(L->getHeader()->getName() + ".termcond");
2727         Cond->insertInto(ExitingBlock, TermBr->getIterator());
2728 
2729         // Clone the IVUse, as the old use still exists!
2730         CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
2731         TermBr->replaceUsesOfWith(OldCond, Cond);
2732       }
2733     }
2734 
2735     // If we get to here, we know that we can transform the setcc instruction to
2736     // use the post-incremented version of the IV, allowing us to coalesce the
2737     // live ranges for the IV correctly.
2738     CondUse->transformToPostInc(L);
2739     Changed = true;
2740 
2741     PostIncs.insert(Cond);
2742   decline_post_inc:;
2743   }
2744 
2745   // Determine an insertion point for the loop induction variable increment. It
2746   // must dominate all the post-inc comparisons we just set up, and it must
2747   // dominate the loop latch edge.
2748   IVIncInsertPos = L->getLoopLatch()->getTerminator();
2749   for (Instruction *Inst : PostIncs)
2750     IVIncInsertPos = DT.findNearestCommonDominator(IVIncInsertPos, Inst);
2751 }
2752 
2753 /// Determine if the given use can accommodate a fixup at the given offset and
2754 /// other details. If so, update the use and return true.
reconcileNewOffset(LSRUse & LU,Immediate NewOffset,bool HasBaseReg,LSRUse::KindType Kind,MemAccessTy AccessTy)2755 bool LSRInstance::reconcileNewOffset(LSRUse &LU, Immediate NewOffset,
2756                                      bool HasBaseReg, LSRUse::KindType Kind,
2757                                      MemAccessTy AccessTy) {
2758   Immediate NewMinOffset = LU.MinOffset;
2759   Immediate NewMaxOffset = LU.MaxOffset;
2760   MemAccessTy NewAccessTy = AccessTy;
2761 
2762   // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
2763   // something conservative, however this can pessimize in the case that one of
2764   // the uses will have all its uses outside the loop, for example.
2765   if (LU.Kind != Kind)
2766     return false;
2767 
2768   // Check for a mismatched access type, and fall back conservatively as needed.
2769   // TODO: Be less conservative when the type is similar and can use the same
2770   // addressing modes.
2771   if (Kind == LSRUse::Address) {
2772     if (AccessTy.MemTy != LU.AccessTy.MemTy) {
2773       NewAccessTy = MemAccessTy::getUnknown(AccessTy.MemTy->getContext(),
2774                                             AccessTy.AddrSpace);
2775     }
2776   }
2777 
2778   // Conservatively assume HasBaseReg is true for now.
2779   if (Immediate::isKnownLT(NewOffset, LU.MinOffset)) {
2780     if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2781                           LU.MaxOffset - NewOffset, HasBaseReg))
2782       return false;
2783     NewMinOffset = NewOffset;
2784   } else if (Immediate::isKnownGT(NewOffset, LU.MaxOffset)) {
2785     if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr,
2786                           NewOffset - LU.MinOffset, HasBaseReg))
2787       return false;
2788     NewMaxOffset = NewOffset;
2789   }
2790 
2791   // FIXME: We should be able to handle some level of scalable offset support
2792   // for 'void', but in order to get basic support up and running this is
2793   // being left out.
2794   if (NewAccessTy.MemTy && NewAccessTy.MemTy->isVoidTy() &&
2795       (NewMinOffset.isScalable() || NewMaxOffset.isScalable()))
2796     return false;
2797 
2798   // Update the use.
2799   LU.MinOffset = NewMinOffset;
2800   LU.MaxOffset = NewMaxOffset;
2801   LU.AccessTy = NewAccessTy;
2802   return true;
2803 }
2804 
2805 /// Return an LSRUse index and an offset value for a fixup which needs the given
2806 /// expression, with the given kind and optional access type.  Either reuse an
2807 /// existing use or create a new one, as needed.
getUse(const SCEV * & Expr,LSRUse::KindType Kind,MemAccessTy AccessTy)2808 std::pair<size_t, Immediate> LSRInstance::getUse(const SCEV *&Expr,
2809                                                  LSRUse::KindType Kind,
2810                                                  MemAccessTy AccessTy) {
2811   const SCEV *Copy = Expr;
2812   Immediate Offset = ExtractImmediate(Expr, SE);
2813 
2814   // Basic uses can't accept any offset, for example.
2815   if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr,
2816                         Offset, /*HasBaseReg=*/ true)) {
2817     Expr = Copy;
2818     Offset = Immediate::getFixed(0);
2819   }
2820 
2821   std::pair<UseMapTy::iterator, bool> P =
2822     UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
2823   if (!P.second) {
2824     // A use already existed with this base.
2825     size_t LUIdx = P.first->second;
2826     LSRUse &LU = Uses[LUIdx];
2827     if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
2828       // Reuse this use.
2829       return std::make_pair(LUIdx, Offset);
2830   }
2831 
2832   // Create a new use.
2833   size_t LUIdx = Uses.size();
2834   P.first->second = LUIdx;
2835   Uses.push_back(LSRUse(Kind, AccessTy));
2836   LSRUse &LU = Uses[LUIdx];
2837 
2838   LU.MinOffset = Offset;
2839   LU.MaxOffset = Offset;
2840   return std::make_pair(LUIdx, Offset);
2841 }
2842 
2843 /// Delete the given use from the Uses list.
DeleteUse(LSRUse & LU,size_t LUIdx)2844 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
2845   if (&LU != &Uses.back())
2846     std::swap(LU, Uses.back());
2847   Uses.pop_back();
2848 
2849   // Update RegUses.
2850   RegUses.swapAndDropUse(LUIdx, Uses.size());
2851 }
2852 
2853 /// Look for a use distinct from OrigLU which is has a formula that has the same
2854 /// registers as the given formula.
2855 LSRUse *
FindUseWithSimilarFormula(const Formula & OrigF,const LSRUse & OrigLU)2856 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
2857                                        const LSRUse &OrigLU) {
2858   // Search all uses for the formula. This could be more clever.
2859   for (LSRUse &LU : Uses) {
2860     // Check whether this use is close enough to OrigLU, to see whether it's
2861     // worthwhile looking through its formulae.
2862     // Ignore ICmpZero uses because they may contain formulae generated by
2863     // GenerateICmpZeroScales, in which case adding fixup offsets may
2864     // be invalid.
2865     if (&LU != &OrigLU &&
2866         LU.Kind != LSRUse::ICmpZero &&
2867         LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
2868         LU.WidestFixupType == OrigLU.WidestFixupType &&
2869         LU.HasFormulaWithSameRegs(OrigF)) {
2870       // Scan through this use's formulae.
2871       for (const Formula &F : LU.Formulae) {
2872         // Check to see if this formula has the same registers and symbols
2873         // as OrigF.
2874         if (F.BaseRegs == OrigF.BaseRegs &&
2875             F.ScaledReg == OrigF.ScaledReg &&
2876             F.BaseGV == OrigF.BaseGV &&
2877             F.Scale == OrigF.Scale &&
2878             F.UnfoldedOffset == OrigF.UnfoldedOffset) {
2879           if (F.BaseOffset.isZero())
2880             return &LU;
2881           // This is the formula where all the registers and symbols matched;
2882           // there aren't going to be any others. Since we declined it, we
2883           // can skip the rest of the formulae and proceed to the next LSRUse.
2884           break;
2885         }
2886       }
2887     }
2888   }
2889 
2890   // Nothing looked good.
2891   return nullptr;
2892 }
2893 
CollectInterestingTypesAndFactors()2894 void LSRInstance::CollectInterestingTypesAndFactors() {
2895   SmallSetVector<const SCEV *, 4> Strides;
2896 
2897   // Collect interesting types and strides.
2898   SmallVector<const SCEV *, 4> Worklist;
2899   for (const IVStrideUse &U : IU) {
2900     const SCEV *Expr = IU.getExpr(U);
2901     if (!Expr)
2902       continue;
2903 
2904     // Collect interesting types.
2905     Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
2906 
2907     // Add strides for mentioned loops.
2908     Worklist.push_back(Expr);
2909     do {
2910       const SCEV *S = Worklist.pop_back_val();
2911       if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2912         if (AR->getLoop() == L)
2913           Strides.insert(AR->getStepRecurrence(SE));
2914         Worklist.push_back(AR->getStart());
2915       } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
2916         append_range(Worklist, Add->operands());
2917       }
2918     } while (!Worklist.empty());
2919   }
2920 
2921   // Compute interesting factors from the set of interesting strides.
2922   for (SmallSetVector<const SCEV *, 4>::const_iterator
2923        I = Strides.begin(), E = Strides.end(); I != E; ++I)
2924     for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
2925          std::next(I); NewStrideIter != E; ++NewStrideIter) {
2926       const SCEV *OldStride = *I;
2927       const SCEV *NewStride = *NewStrideIter;
2928 
2929       if (SE.getTypeSizeInBits(OldStride->getType()) !=
2930           SE.getTypeSizeInBits(NewStride->getType())) {
2931         if (SE.getTypeSizeInBits(OldStride->getType()) >
2932             SE.getTypeSizeInBits(NewStride->getType()))
2933           NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
2934         else
2935           OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
2936       }
2937       if (const SCEVConstant *Factor =
2938             dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
2939                                                         SE, true))) {
2940         if (Factor->getAPInt().getSignificantBits() <= 64 && !Factor->isZero())
2941           Factors.insert(Factor->getAPInt().getSExtValue());
2942       } else if (const SCEVConstant *Factor =
2943                    dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
2944                                                                NewStride,
2945                                                                SE, true))) {
2946         if (Factor->getAPInt().getSignificantBits() <= 64 && !Factor->isZero())
2947           Factors.insert(Factor->getAPInt().getSExtValue());
2948       }
2949     }
2950 
2951   // If all uses use the same type, don't bother looking for truncation-based
2952   // reuse.
2953   if (Types.size() == 1)
2954     Types.clear();
2955 
2956   LLVM_DEBUG(print_factors_and_types(dbgs()));
2957 }
2958 
2959 /// Helper for CollectChains that finds an IV operand (computed by an AddRec in
2960 /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to
2961 /// IVStrideUses, we could partially skip this.
2962 static User::op_iterator
findIVOperand(User::op_iterator OI,User::op_iterator OE,Loop * L,ScalarEvolution & SE)2963 findIVOperand(User::op_iterator OI, User::op_iterator OE,
2964               Loop *L, ScalarEvolution &SE) {
2965   for(; OI != OE; ++OI) {
2966     if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
2967       if (!SE.isSCEVable(Oper->getType()))
2968         continue;
2969 
2970       if (const SCEVAddRecExpr *AR =
2971           dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
2972         if (AR->getLoop() == L)
2973           break;
2974       }
2975     }
2976   }
2977   return OI;
2978 }
2979 
2980 /// IVChain logic must consistently peek base TruncInst operands, so wrap it in
2981 /// a convenient helper.
getWideOperand(Value * Oper)2982 static Value *getWideOperand(Value *Oper) {
2983   if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
2984     return Trunc->getOperand(0);
2985   return Oper;
2986 }
2987 
2988 /// Return an approximation of this SCEV expression's "base", or NULL for any
2989 /// constant. Returning the expression itself is conservative. Returning a
2990 /// deeper subexpression is more precise and valid as long as it isn't less
2991 /// complex than another subexpression. For expressions involving multiple
2992 /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids
2993 /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i],
2994 /// IVInc==b-a.
2995 ///
2996 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
2997 /// SCEVUnknown, we simply return the rightmost SCEV operand.
getExprBase(const SCEV * S)2998 static const SCEV *getExprBase(const SCEV *S) {
2999   switch (S->getSCEVType()) {
3000   default: // including scUnknown.
3001     return S;
3002   case scConstant:
3003   case scVScale:
3004     return nullptr;
3005   case scTruncate:
3006     return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
3007   case scZeroExtend:
3008     return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
3009   case scSignExtend:
3010     return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
3011   case scAddExpr: {
3012     // Skip over scaled operands (scMulExpr) to follow add operands as long as
3013     // there's nothing more complex.
3014     // FIXME: not sure if we want to recognize negation.
3015     const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
3016     for (const SCEV *SubExpr : reverse(Add->operands())) {
3017       if (SubExpr->getSCEVType() == scAddExpr)
3018         return getExprBase(SubExpr);
3019 
3020       if (SubExpr->getSCEVType() != scMulExpr)
3021         return SubExpr;
3022     }
3023     return S; // all operands are scaled, be conservative.
3024   }
3025   case scAddRecExpr:
3026     return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
3027   }
3028   llvm_unreachable("Unknown SCEV kind!");
3029 }
3030 
3031 /// Return true if the chain increment is profitable to expand into a loop
3032 /// invariant value, which may require its own register. A profitable chain
3033 /// increment will be an offset relative to the same base. We allow such offsets
3034 /// to potentially be used as chain increment as long as it's not obviously
3035 /// expensive to expand using real instructions.
isProfitableIncrement(const SCEV * OperExpr,const SCEV * IncExpr,ScalarEvolution & SE)3036 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
3037                                     const SCEV *IncExpr,
3038                                     ScalarEvolution &SE) {
3039   // Aggressively form chains when -stress-ivchain.
3040   if (StressIVChain)
3041     return true;
3042 
3043   // Do not replace a constant offset from IV head with a nonconstant IV
3044   // increment.
3045   if (!isa<SCEVConstant>(IncExpr)) {
3046     const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
3047     if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
3048       return false;
3049   }
3050 
3051   SmallPtrSet<const SCEV*, 8> Processed;
3052   return !isHighCostExpansion(IncExpr, Processed, SE);
3053 }
3054 
3055 /// Return true if the number of registers needed for the chain is estimated to
3056 /// be less than the number required for the individual IV users. First prohibit
3057 /// any IV users that keep the IV live across increments (the Users set should
3058 /// be empty). Next count the number and type of increments in the chain.
3059 ///
3060 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
3061 /// effectively use postinc addressing modes. Only consider it profitable it the
3062 /// increments can be computed in fewer registers when chained.
3063 ///
3064 /// TODO: Consider IVInc free if it's already used in another chains.
isProfitableChain(IVChain & Chain,SmallPtrSetImpl<Instruction * > & Users,ScalarEvolution & SE,const TargetTransformInfo & TTI)3065 static bool isProfitableChain(IVChain &Chain,
3066                               SmallPtrSetImpl<Instruction *> &Users,
3067                               ScalarEvolution &SE,
3068                               const TargetTransformInfo &TTI) {
3069   if (StressIVChain)
3070     return true;
3071 
3072   if (!Chain.hasIncs())
3073     return false;
3074 
3075   if (!Users.empty()) {
3076     LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
3077                for (Instruction *Inst
3078                     : Users) { dbgs() << "  " << *Inst << "\n"; });
3079     return false;
3080   }
3081   assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3082 
3083   // The chain itself may require a register, so intialize cost to 1.
3084   int cost = 1;
3085 
3086   // A complete chain likely eliminates the need for keeping the original IV in
3087   // a register. LSR does not currently know how to form a complete chain unless
3088   // the header phi already exists.
3089   if (isa<PHINode>(Chain.tailUserInst())
3090       && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
3091     --cost;
3092   }
3093   const SCEV *LastIncExpr = nullptr;
3094   unsigned NumConstIncrements = 0;
3095   unsigned NumVarIncrements = 0;
3096   unsigned NumReusedIncrements = 0;
3097 
3098   if (TTI.isProfitableLSRChainElement(Chain.Incs[0].UserInst))
3099     return true;
3100 
3101   for (const IVInc &Inc : Chain) {
3102     if (TTI.isProfitableLSRChainElement(Inc.UserInst))
3103       return true;
3104     if (Inc.IncExpr->isZero())
3105       continue;
3106 
3107     // Incrementing by zero or some constant is neutral. We assume constants can
3108     // be folded into an addressing mode or an add's immediate operand.
3109     if (isa<SCEVConstant>(Inc.IncExpr)) {
3110       ++NumConstIncrements;
3111       continue;
3112     }
3113 
3114     if (Inc.IncExpr == LastIncExpr)
3115       ++NumReusedIncrements;
3116     else
3117       ++NumVarIncrements;
3118 
3119     LastIncExpr = Inc.IncExpr;
3120   }
3121   // An IV chain with a single increment is handled by LSR's postinc
3122   // uses. However, a chain with multiple increments requires keeping the IV's
3123   // value live longer than it needs to be if chained.
3124   if (NumConstIncrements > 1)
3125     --cost;
3126 
3127   // Materializing increment expressions in the preheader that didn't exist in
3128   // the original code may cost a register. For example, sign-extended array
3129   // indices can produce ridiculous increments like this:
3130   // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
3131   cost += NumVarIncrements;
3132 
3133   // Reusing variable increments likely saves a register to hold the multiple of
3134   // the stride.
3135   cost -= NumReusedIncrements;
3136 
3137   LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
3138                     << "\n");
3139 
3140   return cost < 0;
3141 }
3142 
3143 /// Add this IV user to an existing chain or make it the head of a new chain.
ChainInstruction(Instruction * UserInst,Instruction * IVOper,SmallVectorImpl<ChainUsers> & ChainUsersVec)3144 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
3145                                    SmallVectorImpl<ChainUsers> &ChainUsersVec) {
3146   // When IVs are used as types of varying widths, they are generally converted
3147   // to a wider type with some uses remaining narrow under a (free) trunc.
3148   Value *const NextIV = getWideOperand(IVOper);
3149   const SCEV *const OperExpr = SE.getSCEV(NextIV);
3150   const SCEV *const OperExprBase = getExprBase(OperExpr);
3151 
3152   // Visit all existing chains. Check if its IVOper can be computed as a
3153   // profitable loop invariant increment from the last link in the Chain.
3154   unsigned ChainIdx = 0, NChains = IVChainVec.size();
3155   const SCEV *LastIncExpr = nullptr;
3156   for (; ChainIdx < NChains; ++ChainIdx) {
3157     IVChain &Chain = IVChainVec[ChainIdx];
3158 
3159     // Prune the solution space aggressively by checking that both IV operands
3160     // are expressions that operate on the same unscaled SCEVUnknown. This
3161     // "base" will be canceled by the subsequent getMinusSCEV call. Checking
3162     // first avoids creating extra SCEV expressions.
3163     if (!StressIVChain && Chain.ExprBase != OperExprBase)
3164       continue;
3165 
3166     Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
3167     if (PrevIV->getType() != NextIV->getType())
3168       continue;
3169 
3170     // A phi node terminates a chain.
3171     if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
3172       continue;
3173 
3174     // The increment must be loop-invariant so it can be kept in a register.
3175     const SCEV *PrevExpr = SE.getSCEV(PrevIV);
3176     const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
3177     if (isa<SCEVCouldNotCompute>(IncExpr) || !SE.isLoopInvariant(IncExpr, L))
3178       continue;
3179 
3180     if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
3181       LastIncExpr = IncExpr;
3182       break;
3183     }
3184   }
3185   // If we haven't found a chain, create a new one, unless we hit the max. Don't
3186   // bother for phi nodes, because they must be last in the chain.
3187   if (ChainIdx == NChains) {
3188     if (isa<PHINode>(UserInst))
3189       return;
3190     if (NChains >= MaxChains && !StressIVChain) {
3191       LLVM_DEBUG(dbgs() << "IV Chain Limit\n");
3192       return;
3193     }
3194     LastIncExpr = OperExpr;
3195     // IVUsers may have skipped over sign/zero extensions. We don't currently
3196     // attempt to form chains involving extensions unless they can be hoisted
3197     // into this loop's AddRec.
3198     if (!isa<SCEVAddRecExpr>(LastIncExpr))
3199       return;
3200     ++NChains;
3201     IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
3202                                  OperExprBase));
3203     ChainUsersVec.resize(NChains);
3204     LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
3205                       << ") IV=" << *LastIncExpr << "\n");
3206   } else {
3207     LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << "  Inc: (" << *UserInst
3208                       << ") IV+" << *LastIncExpr << "\n");
3209     // Add this IV user to the end of the chain.
3210     IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
3211   }
3212   IVChain &Chain = IVChainVec[ChainIdx];
3213 
3214   SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
3215   // This chain's NearUsers become FarUsers.
3216   if (!LastIncExpr->isZero()) {
3217     ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
3218                                             NearUsers.end());
3219     NearUsers.clear();
3220   }
3221 
3222   // All other uses of IVOperand become near uses of the chain.
3223   // We currently ignore intermediate values within SCEV expressions, assuming
3224   // they will eventually be used be the current chain, or can be computed
3225   // from one of the chain increments. To be more precise we could
3226   // transitively follow its user and only add leaf IV users to the set.
3227   for (User *U : IVOper->users()) {
3228     Instruction *OtherUse = dyn_cast<Instruction>(U);
3229     if (!OtherUse)
3230       continue;
3231     // Uses in the chain will no longer be uses if the chain is formed.
3232     // Include the head of the chain in this iteration (not Chain.begin()).
3233     IVChain::const_iterator IncIter = Chain.Incs.begin();
3234     IVChain::const_iterator IncEnd = Chain.Incs.end();
3235     for( ; IncIter != IncEnd; ++IncIter) {
3236       if (IncIter->UserInst == OtherUse)
3237         break;
3238     }
3239     if (IncIter != IncEnd)
3240       continue;
3241 
3242     if (SE.isSCEVable(OtherUse->getType())
3243         && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
3244         && IU.isIVUserOrOperand(OtherUse)) {
3245       continue;
3246     }
3247     NearUsers.insert(OtherUse);
3248   }
3249 
3250   // Since this user is part of the chain, it's no longer considered a use
3251   // of the chain.
3252   ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
3253 }
3254 
3255 /// Populate the vector of Chains.
3256 ///
3257 /// This decreases ILP at the architecture level. Targets with ample registers,
3258 /// multiple memory ports, and no register renaming probably don't want
3259 /// this. However, such targets should probably disable LSR altogether.
3260 ///
3261 /// The job of LSR is to make a reasonable choice of induction variables across
3262 /// the loop. Subsequent passes can easily "unchain" computation exposing more
3263 /// ILP *within the loop* if the target wants it.
3264 ///
3265 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
3266 /// will not reorder memory operations, it will recognize this as a chain, but
3267 /// will generate redundant IV increments. Ideally this would be corrected later
3268 /// by a smart scheduler:
3269 ///        = A[i]
3270 ///        = A[i+x]
3271 /// A[i]   =
3272 /// A[i+x] =
3273 ///
3274 /// TODO: Walk the entire domtree within this loop, not just the path to the
3275 /// loop latch. This will discover chains on side paths, but requires
3276 /// maintaining multiple copies of the Chains state.
CollectChains()3277 void LSRInstance::CollectChains() {
3278   LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n");
3279   SmallVector<ChainUsers, 8> ChainUsersVec;
3280 
3281   SmallVector<BasicBlock *,8> LatchPath;
3282   BasicBlock *LoopHeader = L->getHeader();
3283   for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
3284        Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
3285     LatchPath.push_back(Rung->getBlock());
3286   }
3287   LatchPath.push_back(LoopHeader);
3288 
3289   // Walk the instruction stream from the loop header to the loop latch.
3290   for (BasicBlock *BB : reverse(LatchPath)) {
3291     for (Instruction &I : *BB) {
3292       // Skip instructions that weren't seen by IVUsers analysis.
3293       if (isa<PHINode>(I) || !IU.isIVUserOrOperand(&I))
3294         continue;
3295 
3296       // Ignore users that are part of a SCEV expression. This way we only
3297       // consider leaf IV Users. This effectively rediscovers a portion of
3298       // IVUsers analysis but in program order this time.
3299       if (SE.isSCEVable(I.getType()) && !isa<SCEVUnknown>(SE.getSCEV(&I)))
3300           continue;
3301 
3302       // Remove this instruction from any NearUsers set it may be in.
3303       for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
3304            ChainIdx < NChains; ++ChainIdx) {
3305         ChainUsersVec[ChainIdx].NearUsers.erase(&I);
3306       }
3307       // Search for operands that can be chained.
3308       SmallPtrSet<Instruction*, 4> UniqueOperands;
3309       User::op_iterator IVOpEnd = I.op_end();
3310       User::op_iterator IVOpIter = findIVOperand(I.op_begin(), IVOpEnd, L, SE);
3311       while (IVOpIter != IVOpEnd) {
3312         Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
3313         if (UniqueOperands.insert(IVOpInst).second)
3314           ChainInstruction(&I, IVOpInst, ChainUsersVec);
3315         IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3316       }
3317     } // Continue walking down the instructions.
3318   } // Continue walking down the domtree.
3319   // Visit phi backedges to determine if the chain can generate the IV postinc.
3320   for (PHINode &PN : L->getHeader()->phis()) {
3321     if (!SE.isSCEVable(PN.getType()))
3322       continue;
3323 
3324     Instruction *IncV =
3325         dyn_cast<Instruction>(PN.getIncomingValueForBlock(L->getLoopLatch()));
3326     if (IncV)
3327       ChainInstruction(&PN, IncV, ChainUsersVec);
3328   }
3329   // Remove any unprofitable chains.
3330   unsigned ChainIdx = 0;
3331   for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
3332        UsersIdx < NChains; ++UsersIdx) {
3333     if (!isProfitableChain(IVChainVec[UsersIdx],
3334                            ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
3335       continue;
3336     // Preserve the chain at UsesIdx.
3337     if (ChainIdx != UsersIdx)
3338       IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
3339     FinalizeChain(IVChainVec[ChainIdx]);
3340     ++ChainIdx;
3341   }
3342   IVChainVec.resize(ChainIdx);
3343 }
3344 
FinalizeChain(IVChain & Chain)3345 void LSRInstance::FinalizeChain(IVChain &Chain) {
3346   assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
3347   LLVM_DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
3348 
3349   for (const IVInc &Inc : Chain) {
3350     LLVM_DEBUG(dbgs() << "        Inc: " << *Inc.UserInst << "\n");
3351     auto UseI = find(Inc.UserInst->operands(), Inc.IVOperand);
3352     assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand");
3353     IVIncSet.insert(UseI);
3354   }
3355 }
3356 
3357 /// Return true if the IVInc can be folded into an addressing mode.
canFoldIVIncExpr(const SCEV * IncExpr,Instruction * UserInst,Value * Operand,const TargetTransformInfo & TTI)3358 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
3359                              Value *Operand, const TargetTransformInfo &TTI) {
3360   const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
3361   Immediate IncOffset = Immediate::getZero();
3362   if (IncConst) {
3363     if (IncConst && IncConst->getAPInt().getSignificantBits() > 64)
3364       return false;
3365     IncOffset = Immediate::getFixed(IncConst->getValue()->getSExtValue());
3366   } else {
3367     // Look for mul(vscale, constant), to detect a scalable offset.
3368     auto *IncVScale = dyn_cast<SCEVMulExpr>(IncExpr);
3369     if (!IncVScale || IncVScale->getNumOperands() != 2 ||
3370         !isa<SCEVVScale>(IncVScale->getOperand(1)))
3371       return false;
3372     auto *Scale = dyn_cast<SCEVConstant>(IncVScale->getOperand(0));
3373     if (!Scale || Scale->getType()->getScalarSizeInBits() > 64)
3374       return false;
3375     IncOffset = Immediate::getScalable(Scale->getValue()->getSExtValue());
3376   }
3377 
3378   if (!isAddressUse(TTI, UserInst, Operand))
3379     return false;
3380 
3381   MemAccessTy AccessTy = getAccessType(TTI, UserInst, Operand);
3382   if (!isAlwaysFoldable(TTI, LSRUse::Address, AccessTy, /*BaseGV=*/nullptr,
3383                         IncOffset, /*HasBaseReg=*/false))
3384     return false;
3385 
3386   return true;
3387 }
3388 
3389 /// Generate an add or subtract for each IVInc in a chain to materialize the IV
3390 /// user's operand from the previous IV user's operand.
GenerateIVChain(const IVChain & Chain,SmallVectorImpl<WeakTrackingVH> & DeadInsts)3391 void LSRInstance::GenerateIVChain(const IVChain &Chain,
3392                                   SmallVectorImpl<WeakTrackingVH> &DeadInsts) {
3393   // Find the new IVOperand for the head of the chain. It may have been replaced
3394   // by LSR.
3395   const IVInc &Head = Chain.Incs[0];
3396   User::op_iterator IVOpEnd = Head.UserInst->op_end();
3397   // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
3398   User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
3399                                              IVOpEnd, L, SE);
3400   Value *IVSrc = nullptr;
3401   while (IVOpIter != IVOpEnd) {
3402     IVSrc = getWideOperand(*IVOpIter);
3403 
3404     // If this operand computes the expression that the chain needs, we may use
3405     // it. (Check this after setting IVSrc which is used below.)
3406     //
3407     // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
3408     // narrow for the chain, so we can no longer use it. We do allow using a
3409     // wider phi, assuming the LSR checked for free truncation. In that case we
3410     // should already have a truncate on this operand such that
3411     // getSCEV(IVSrc) == IncExpr.
3412     if (SE.getSCEV(*IVOpIter) == Head.IncExpr
3413         || SE.getSCEV(IVSrc) == Head.IncExpr) {
3414       break;
3415     }
3416     IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
3417   }
3418   if (IVOpIter == IVOpEnd) {
3419     // Gracefully give up on this chain.
3420     LLVM_DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
3421     return;
3422   }
3423   assert(IVSrc && "Failed to find IV chain source");
3424 
3425   LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
3426   Type *IVTy = IVSrc->getType();
3427   Type *IntTy = SE.getEffectiveSCEVType(IVTy);
3428   const SCEV *LeftOverExpr = nullptr;
3429   const SCEV *Accum = SE.getZero(IntTy);
3430   SmallVector<std::pair<const SCEV *, Value *>> Bases;
3431   Bases.emplace_back(Accum, IVSrc);
3432 
3433   for (const IVInc &Inc : Chain) {
3434     Instruction *InsertPt = Inc.UserInst;
3435     if (isa<PHINode>(InsertPt))
3436       InsertPt = L->getLoopLatch()->getTerminator();
3437 
3438     // IVOper will replace the current IV User's operand. IVSrc is the IV
3439     // value currently held in a register.
3440     Value *IVOper = IVSrc;
3441     if (!Inc.IncExpr->isZero()) {
3442       // IncExpr was the result of subtraction of two narrow values, so must
3443       // be signed.
3444       const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy);
3445       Accum = SE.getAddExpr(Accum, IncExpr);
3446       LeftOverExpr = LeftOverExpr ?
3447         SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
3448     }
3449 
3450     // Look through each base to see if any can produce a nice addressing mode.
3451     bool FoundBase = false;
3452     for (auto [MapScev, MapIVOper] : reverse(Bases)) {
3453       const SCEV *Remainder = SE.getMinusSCEV(Accum, MapScev);
3454       if (canFoldIVIncExpr(Remainder, Inc.UserInst, Inc.IVOperand, TTI)) {
3455         if (!Remainder->isZero()) {
3456           Rewriter.clearPostInc();
3457           Value *IncV = Rewriter.expandCodeFor(Remainder, IntTy, InsertPt);
3458           const SCEV *IVOperExpr =
3459               SE.getAddExpr(SE.getUnknown(MapIVOper), SE.getUnknown(IncV));
3460           IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3461         } else {
3462           IVOper = MapIVOper;
3463         }
3464 
3465         FoundBase = true;
3466         break;
3467       }
3468     }
3469     if (!FoundBase && LeftOverExpr && !LeftOverExpr->isZero()) {
3470       // Expand the IV increment.
3471       Rewriter.clearPostInc();
3472       Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
3473       const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
3474                                              SE.getUnknown(IncV));
3475       IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
3476 
3477       // If an IV increment can't be folded, use it as the next IV value.
3478       if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) {
3479         assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
3480         Bases.emplace_back(Accum, IVOper);
3481         IVSrc = IVOper;
3482         LeftOverExpr = nullptr;
3483       }
3484     }
3485     Type *OperTy = Inc.IVOperand->getType();
3486     if (IVTy != OperTy) {
3487       assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
3488              "cannot extend a chained IV");
3489       IRBuilder<> Builder(InsertPt);
3490       IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
3491     }
3492     Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper);
3493     if (auto *OperandIsInstr = dyn_cast<Instruction>(Inc.IVOperand))
3494       DeadInsts.emplace_back(OperandIsInstr);
3495   }
3496   // If LSR created a new, wider phi, we may also replace its postinc. We only
3497   // do this if we also found a wide value for the head of the chain.
3498   if (isa<PHINode>(Chain.tailUserInst())) {
3499     for (PHINode &Phi : L->getHeader()->phis()) {
3500       if (Phi.getType() != IVSrc->getType())
3501         continue;
3502       Instruction *PostIncV = dyn_cast<Instruction>(
3503           Phi.getIncomingValueForBlock(L->getLoopLatch()));
3504       if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
3505         continue;
3506       Value *IVOper = IVSrc;
3507       Type *PostIncTy = PostIncV->getType();
3508       if (IVTy != PostIncTy) {
3509         assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
3510         IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
3511         Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
3512         IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
3513       }
3514       Phi.replaceUsesOfWith(PostIncV, IVOper);
3515       DeadInsts.emplace_back(PostIncV);
3516     }
3517   }
3518 }
3519 
CollectFixupsAndInitialFormulae()3520 void LSRInstance::CollectFixupsAndInitialFormulae() {
3521   BranchInst *ExitBranch = nullptr;
3522   bool SaveCmp = TTI.canSaveCmp(L, &ExitBranch, &SE, &LI, &DT, &AC, &TLI);
3523 
3524   // For calculating baseline cost
3525   SmallPtrSet<const SCEV *, 16> Regs;
3526   DenseSet<const SCEV *> VisitedRegs;
3527   DenseSet<size_t> VisitedLSRUse;
3528 
3529   for (const IVStrideUse &U : IU) {
3530     Instruction *UserInst = U.getUser();
3531     // Skip IV users that are part of profitable IV Chains.
3532     User::op_iterator UseI =
3533         find(UserInst->operands(), U.getOperandValToReplace());
3534     assert(UseI != UserInst->op_end() && "cannot find IV operand");
3535     if (IVIncSet.count(UseI)) {
3536       LLVM_DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n');
3537       continue;
3538     }
3539 
3540     LSRUse::KindType Kind = LSRUse::Basic;
3541     MemAccessTy AccessTy;
3542     if (isAddressUse(TTI, UserInst, U.getOperandValToReplace())) {
3543       Kind = LSRUse::Address;
3544       AccessTy = getAccessType(TTI, UserInst, U.getOperandValToReplace());
3545     }
3546 
3547     const SCEV *S = IU.getExpr(U);
3548     if (!S)
3549       continue;
3550     PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops();
3551 
3552     // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
3553     // (N - i == 0), and this allows (N - i) to be the expression that we work
3554     // with rather than just N or i, so we can consider the register
3555     // requirements for both N and i at the same time. Limiting this code to
3556     // equality icmps is not a problem because all interesting loops use
3557     // equality icmps, thanks to IndVarSimplify.
3558     if (ICmpInst *CI = dyn_cast<ICmpInst>(UserInst)) {
3559       // If CI can be saved in some target, like replaced inside hardware loop
3560       // in PowerPC, no need to generate initial formulae for it.
3561       if (SaveCmp && CI == dyn_cast<ICmpInst>(ExitBranch->getCondition()))
3562         continue;
3563       if (CI->isEquality()) {
3564         // Swap the operands if needed to put the OperandValToReplace on the
3565         // left, for consistency.
3566         Value *NV = CI->getOperand(1);
3567         if (NV == U.getOperandValToReplace()) {
3568           CI->setOperand(1, CI->getOperand(0));
3569           CI->setOperand(0, NV);
3570           NV = CI->getOperand(1);
3571           Changed = true;
3572         }
3573 
3574         // x == y  -->  x - y == 0
3575         const SCEV *N = SE.getSCEV(NV);
3576         if (SE.isLoopInvariant(N, L) && Rewriter.isSafeToExpand(N) &&
3577             (!NV->getType()->isPointerTy() ||
3578              SE.getPointerBase(N) == SE.getPointerBase(S))) {
3579           // S is normalized, so normalize N before folding it into S
3580           // to keep the result normalized.
3581           N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3582           if (!N)
3583             continue;
3584           Kind = LSRUse::ICmpZero;
3585           S = SE.getMinusSCEV(N, S);
3586         } else if (L->isLoopInvariant(NV) &&
3587                    (!isa<Instruction>(NV) ||
3588                     DT.dominates(cast<Instruction>(NV), L->getHeader())) &&
3589                    !NV->getType()->isPointerTy()) {
3590           // If we can't generally expand the expression (e.g. it contains
3591           // a divide), but it is already at a loop invariant point before the
3592           // loop, wrap it in an unknown (to prevent the expander from trying
3593           // to re-expand in a potentially unsafe way.)  The restriction to
3594           // integer types is required because the unknown hides the base, and
3595           // SCEV can't compute the difference of two unknown pointers.
3596           N = SE.getUnknown(NV);
3597           N = normalizeForPostIncUse(N, TmpPostIncLoops, SE);
3598           if (!N)
3599             continue;
3600           Kind = LSRUse::ICmpZero;
3601           S = SE.getMinusSCEV(N, S);
3602           assert(!isa<SCEVCouldNotCompute>(S));
3603         }
3604 
3605         // -1 and the negations of all interesting strides (except the negation
3606         // of -1) are now also interesting.
3607         for (size_t i = 0, e = Factors.size(); i != e; ++i)
3608           if (Factors[i] != -1)
3609             Factors.insert(-(uint64_t)Factors[i]);
3610         Factors.insert(-1);
3611       }
3612     }
3613 
3614     // Get or create an LSRUse.
3615     std::pair<size_t, Immediate> P = getUse(S, Kind, AccessTy);
3616     size_t LUIdx = P.first;
3617     Immediate Offset = P.second;
3618     LSRUse &LU = Uses[LUIdx];
3619 
3620     // Record the fixup.
3621     LSRFixup &LF = LU.getNewFixup();
3622     LF.UserInst = UserInst;
3623     LF.OperandValToReplace = U.getOperandValToReplace();
3624     LF.PostIncLoops = TmpPostIncLoops;
3625     LF.Offset = Offset;
3626     LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3627 
3628     // Create SCEV as Formula for calculating baseline cost
3629     if (!VisitedLSRUse.count(LUIdx) && !LF.isUseFullyOutsideLoop(L)) {
3630       Formula F;
3631       F.initialMatch(S, L, SE);
3632       BaselineCost.RateFormula(F, Regs, VisitedRegs, LU);
3633       VisitedLSRUse.insert(LUIdx);
3634     }
3635 
3636     if (!LU.WidestFixupType ||
3637         SE.getTypeSizeInBits(LU.WidestFixupType) <
3638         SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3639       LU.WidestFixupType = LF.OperandValToReplace->getType();
3640 
3641     // If this is the first use of this LSRUse, give it a formula.
3642     if (LU.Formulae.empty()) {
3643       InsertInitialFormula(S, LU, LUIdx);
3644       CountRegisters(LU.Formulae.back(), LUIdx);
3645     }
3646   }
3647 
3648   LLVM_DEBUG(print_fixups(dbgs()));
3649 }
3650 
3651 /// Insert a formula for the given expression into the given use, separating out
3652 /// loop-variant portions from loop-invariant and loop-computable portions.
InsertInitialFormula(const SCEV * S,LSRUse & LU,size_t LUIdx)3653 void LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU,
3654                                        size_t LUIdx) {
3655   // Mark uses whose expressions cannot be expanded.
3656   if (!Rewriter.isSafeToExpand(S))
3657     LU.RigidFormula = true;
3658 
3659   Formula F;
3660   F.initialMatch(S, L, SE);
3661   bool Inserted = InsertFormula(LU, LUIdx, F);
3662   assert(Inserted && "Initial formula already exists!"); (void)Inserted;
3663 }
3664 
3665 /// Insert a simple single-register formula for the given expression into the
3666 /// given use.
3667 void
InsertSupplementalFormula(const SCEV * S,LSRUse & LU,size_t LUIdx)3668 LSRInstance::InsertSupplementalFormula(const SCEV *S,
3669                                        LSRUse &LU, size_t LUIdx) {
3670   Formula F;
3671   F.BaseRegs.push_back(S);
3672   F.HasBaseReg = true;
3673   bool Inserted = InsertFormula(LU, LUIdx, F);
3674   assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
3675 }
3676 
3677 /// Note which registers are used by the given formula, updating RegUses.
CountRegisters(const Formula & F,size_t LUIdx)3678 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
3679   if (F.ScaledReg)
3680     RegUses.countRegister(F.ScaledReg, LUIdx);
3681   for (const SCEV *BaseReg : F.BaseRegs)
3682     RegUses.countRegister(BaseReg, LUIdx);
3683 }
3684 
3685 /// If the given formula has not yet been inserted, add it to the list, and
3686 /// return true. Return false otherwise.
InsertFormula(LSRUse & LU,unsigned LUIdx,const Formula & F)3687 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
3688   // Do not insert formula that we will not be able to expand.
3689   assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) &&
3690          "Formula is illegal");
3691 
3692   if (!LU.InsertFormula(F, *L))
3693     return false;
3694 
3695   CountRegisters(F, LUIdx);
3696   return true;
3697 }
3698 
3699 /// Check for other uses of loop-invariant values which we're tracking. These
3700 /// other uses will pin these values in registers, making them less profitable
3701 /// for elimination.
3702 /// TODO: This currently misses non-constant addrec step registers.
3703 /// TODO: Should this give more weight to users inside the loop?
3704 void
CollectLoopInvariantFixupsAndFormulae()3705 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
3706   SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
3707   SmallPtrSet<const SCEV *, 32> Visited;
3708 
3709   // Don't collect outside uses if we are favoring postinc - the instructions in
3710   // the loop are more important than the ones outside of it.
3711   if (AMK == TTI::AMK_PostIndexed)
3712     return;
3713 
3714   while (!Worklist.empty()) {
3715     const SCEV *S = Worklist.pop_back_val();
3716 
3717     // Don't process the same SCEV twice
3718     if (!Visited.insert(S).second)
3719       continue;
3720 
3721     if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
3722       append_range(Worklist, N->operands());
3723     else if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(S))
3724       Worklist.push_back(C->getOperand());
3725     else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
3726       Worklist.push_back(D->getLHS());
3727       Worklist.push_back(D->getRHS());
3728     } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
3729       const Value *V = US->getValue();
3730       if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
3731         // Look for instructions defined outside the loop.
3732         if (L->contains(Inst)) continue;
3733       } else if (isa<Constant>(V))
3734         // Constants can be re-materialized.
3735         continue;
3736       for (const Use &U : V->uses()) {
3737         const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
3738         // Ignore non-instructions.
3739         if (!UserInst)
3740           continue;
3741         // Don't bother if the instruction is an EHPad.
3742         if (UserInst->isEHPad())
3743           continue;
3744         // Ignore instructions in other functions (as can happen with
3745         // Constants).
3746         if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
3747           continue;
3748         // Ignore instructions not dominated by the loop.
3749         const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
3750           UserInst->getParent() :
3751           cast<PHINode>(UserInst)->getIncomingBlock(
3752             PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
3753         if (!DT.dominates(L->getHeader(), UseBB))
3754           continue;
3755         // Don't bother if the instruction is in a BB which ends in an EHPad.
3756         if (UseBB->getTerminator()->isEHPad())
3757           continue;
3758 
3759         // Ignore cases in which the currently-examined value could come from
3760         // a basic block terminated with an EHPad. This checks all incoming
3761         // blocks of the phi node since it is possible that the same incoming
3762         // value comes from multiple basic blocks, only some of which may end
3763         // in an EHPad. If any of them do, a subsequent rewrite attempt by this
3764         // pass would try to insert instructions into an EHPad, hitting an
3765         // assertion.
3766         if (isa<PHINode>(UserInst)) {
3767           const auto *PhiNode = cast<PHINode>(UserInst);
3768           bool HasIncompatibleEHPTerminatedBlock = false;
3769           llvm::Value *ExpectedValue = U;
3770           for (unsigned int I = 0; I < PhiNode->getNumIncomingValues(); I++) {
3771             if (PhiNode->getIncomingValue(I) == ExpectedValue) {
3772               if (PhiNode->getIncomingBlock(I)->getTerminator()->isEHPad()) {
3773                 HasIncompatibleEHPTerminatedBlock = true;
3774                 break;
3775               }
3776             }
3777           }
3778           if (HasIncompatibleEHPTerminatedBlock) {
3779             continue;
3780           }
3781         }
3782 
3783         // Don't bother rewriting PHIs in catchswitch blocks.
3784         if (isa<CatchSwitchInst>(UserInst->getParent()->getTerminator()))
3785           continue;
3786         // Ignore uses which are part of other SCEV expressions, to avoid
3787         // analyzing them multiple times.
3788         if (SE.isSCEVable(UserInst->getType())) {
3789           const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
3790           // If the user is a no-op, look through to its uses.
3791           if (!isa<SCEVUnknown>(UserS))
3792             continue;
3793           if (UserS == US) {
3794             Worklist.push_back(
3795               SE.getUnknown(const_cast<Instruction *>(UserInst)));
3796             continue;
3797           }
3798         }
3799         // Ignore icmp instructions which are already being analyzed.
3800         if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
3801           unsigned OtherIdx = !U.getOperandNo();
3802           Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
3803           if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
3804             continue;
3805         }
3806 
3807         std::pair<size_t, Immediate> P =
3808             getUse(S, LSRUse::Basic, MemAccessTy());
3809         size_t LUIdx = P.first;
3810         Immediate Offset = P.second;
3811         LSRUse &LU = Uses[LUIdx];
3812         LSRFixup &LF = LU.getNewFixup();
3813         LF.UserInst = const_cast<Instruction *>(UserInst);
3814         LF.OperandValToReplace = U;
3815         LF.Offset = Offset;
3816         LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
3817         if (!LU.WidestFixupType ||
3818             SE.getTypeSizeInBits(LU.WidestFixupType) <
3819             SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
3820           LU.WidestFixupType = LF.OperandValToReplace->getType();
3821         InsertSupplementalFormula(US, LU, LUIdx);
3822         CountRegisters(LU.Formulae.back(), Uses.size() - 1);
3823         break;
3824       }
3825     }
3826   }
3827 }
3828 
3829 /// Split S into subexpressions which can be pulled out into separate
3830 /// registers. If C is non-null, multiply each subexpression by C.
3831 ///
3832 /// Return remainder expression after factoring the subexpressions captured by
3833 /// Ops. If Ops is complete, return NULL.
CollectSubexprs(const SCEV * S,const SCEVConstant * C,SmallVectorImpl<const SCEV * > & Ops,const Loop * L,ScalarEvolution & SE,unsigned Depth=0)3834 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
3835                                    SmallVectorImpl<const SCEV *> &Ops,
3836                                    const Loop *L,
3837                                    ScalarEvolution &SE,
3838                                    unsigned Depth = 0) {
3839   // Arbitrarily cap recursion to protect compile time.
3840   if (Depth >= 3)
3841     return S;
3842 
3843   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
3844     // Break out add operands.
3845     for (const SCEV *S : Add->operands()) {
3846       const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1);
3847       if (Remainder)
3848         Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3849     }
3850     return nullptr;
3851   } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
3852     // Split a non-zero base out of an addrec.
3853     if (AR->getStart()->isZero() || !AR->isAffine())
3854       return S;
3855 
3856     const SCEV *Remainder = CollectSubexprs(AR->getStart(),
3857                                             C, Ops, L, SE, Depth+1);
3858     // Split the non-zero AddRec unless it is part of a nested recurrence that
3859     // does not pertain to this loop.
3860     if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
3861       Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
3862       Remainder = nullptr;
3863     }
3864     if (Remainder != AR->getStart()) {
3865       if (!Remainder)
3866         Remainder = SE.getConstant(AR->getType(), 0);
3867       return SE.getAddRecExpr(Remainder,
3868                               AR->getStepRecurrence(SE),
3869                               AR->getLoop(),
3870                               //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
3871                               SCEV::FlagAnyWrap);
3872     }
3873   } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
3874     // Break (C * (a + b + c)) into C*a + C*b + C*c.
3875     if (Mul->getNumOperands() != 2)
3876       return S;
3877     if (const SCEVConstant *Op0 =
3878         dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
3879       C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
3880       const SCEV *Remainder =
3881         CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
3882       if (Remainder)
3883         Ops.push_back(SE.getMulExpr(C, Remainder));
3884       return nullptr;
3885     }
3886   }
3887   return S;
3888 }
3889 
3890 /// Return true if the SCEV represents a value that may end up as a
3891 /// post-increment operation.
mayUsePostIncMode(const TargetTransformInfo & TTI,LSRUse & LU,const SCEV * S,const Loop * L,ScalarEvolution & SE)3892 static bool mayUsePostIncMode(const TargetTransformInfo &TTI,
3893                               LSRUse &LU, const SCEV *S, const Loop *L,
3894                               ScalarEvolution &SE) {
3895   if (LU.Kind != LSRUse::Address ||
3896       !LU.AccessTy.getType()->isIntOrIntVectorTy())
3897     return false;
3898   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
3899   if (!AR)
3900     return false;
3901   const SCEV *LoopStep = AR->getStepRecurrence(SE);
3902   if (!isa<SCEVConstant>(LoopStep))
3903     return false;
3904   // Check if a post-indexed load/store can be used.
3905   if (TTI.isIndexedLoadLegal(TTI.MIM_PostInc, AR->getType()) ||
3906       TTI.isIndexedStoreLegal(TTI.MIM_PostInc, AR->getType())) {
3907     const SCEV *LoopStart = AR->getStart();
3908     if (!isa<SCEVConstant>(LoopStart) && SE.isLoopInvariant(LoopStart, L))
3909       return true;
3910   }
3911   return false;
3912 }
3913 
3914 /// Helper function for LSRInstance::GenerateReassociations.
GenerateReassociationsImpl(LSRUse & LU,unsigned LUIdx,const Formula & Base,unsigned Depth,size_t Idx,bool IsScaledReg)3915 void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx,
3916                                              const Formula &Base,
3917                                              unsigned Depth, size_t Idx,
3918                                              bool IsScaledReg) {
3919   const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
3920   // Don't generate reassociations for the base register of a value that
3921   // may generate a post-increment operator. The reason is that the
3922   // reassociations cause extra base+register formula to be created,
3923   // and possibly chosen, but the post-increment is more efficient.
3924   if (AMK == TTI::AMK_PostIndexed && mayUsePostIncMode(TTI, LU, BaseReg, L, SE))
3925     return;
3926   SmallVector<const SCEV *, 8> AddOps;
3927   const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE);
3928   if (Remainder)
3929     AddOps.push_back(Remainder);
3930 
3931   if (AddOps.size() == 1)
3932     return;
3933 
3934   for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
3935                                                      JE = AddOps.end();
3936        J != JE; ++J) {
3937     // Loop-variant "unknown" values are uninteresting; we won't be able to
3938     // do anything meaningful with them.
3939     if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
3940       continue;
3941 
3942     // Don't pull a constant into a register if the constant could be folded
3943     // into an immediate field.
3944     if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3945                          LU.AccessTy, *J, Base.getNumRegs() > 1))
3946       continue;
3947 
3948     // Collect all operands except *J.
3949     SmallVector<const SCEV *, 8> InnerAddOps(
3950         ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
3951     InnerAddOps.append(std::next(J),
3952                        ((const SmallVector<const SCEV *, 8> &)AddOps).end());
3953 
3954     // Don't leave just a constant behind in a register if the constant could
3955     // be folded into an immediate field.
3956     if (InnerAddOps.size() == 1 &&
3957         isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
3958                          LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
3959       continue;
3960 
3961     const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
3962     if (InnerSum->isZero())
3963       continue;
3964     Formula F = Base;
3965 
3966     if (F.UnfoldedOffset.isNonZero() && F.UnfoldedOffset.isScalable())
3967       continue;
3968 
3969     // Add the remaining pieces of the add back into the new formula.
3970     const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
3971     if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
3972         TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset.getFixedValue() +
3973                                 InnerSumSC->getValue()->getZExtValue())) {
3974       F.UnfoldedOffset =
3975           Immediate::getFixed((uint64_t)F.UnfoldedOffset.getFixedValue() +
3976                               InnerSumSC->getValue()->getZExtValue());
3977       if (IsScaledReg)
3978         F.ScaledReg = nullptr;
3979       else
3980         F.BaseRegs.erase(F.BaseRegs.begin() + Idx);
3981     } else if (IsScaledReg)
3982       F.ScaledReg = InnerSum;
3983     else
3984       F.BaseRegs[Idx] = InnerSum;
3985 
3986     // Add J as its own register, or an unfolded immediate.
3987     const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
3988     if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
3989         TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset.getFixedValue() +
3990                                 SC->getValue()->getZExtValue()))
3991       F.UnfoldedOffset =
3992           Immediate::getFixed((uint64_t)F.UnfoldedOffset.getFixedValue() +
3993                               SC->getValue()->getZExtValue());
3994     else
3995       F.BaseRegs.push_back(*J);
3996     // We may have changed the number of register in base regs, adjust the
3997     // formula accordingly.
3998     F.canonicalize(*L);
3999 
4000     if (InsertFormula(LU, LUIdx, F))
4001       // If that formula hadn't been seen before, recurse to find more like
4002       // it.
4003       // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2)
4004       // Because just Depth is not enough to bound compile time.
4005       // This means that every time AddOps.size() is greater 16^x we will add
4006       // x to Depth.
4007       GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
4008                              Depth + 1 + (Log2_32(AddOps.size()) >> 2));
4009   }
4010 }
4011 
4012 /// Split out subexpressions from adds and the bases of addrecs.
GenerateReassociations(LSRUse & LU,unsigned LUIdx,Formula Base,unsigned Depth)4013 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
4014                                          Formula Base, unsigned Depth) {
4015   assert(Base.isCanonical(*L) && "Input must be in the canonical form");
4016   // Arbitrarily cap recursion to protect compile time.
4017   if (Depth >= 3)
4018     return;
4019 
4020   for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
4021     GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i);
4022 
4023   if (Base.Scale == 1)
4024     GenerateReassociationsImpl(LU, LUIdx, Base, Depth,
4025                                /* Idx */ -1, /* IsScaledReg */ true);
4026 }
4027 
4028 ///  Generate a formula consisting of all of the loop-dominating registers added
4029 /// into a single register.
GenerateCombinations(LSRUse & LU,unsigned LUIdx,Formula Base)4030 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
4031                                        Formula Base) {
4032   // This method is only interesting on a plurality of registers.
4033   if (Base.BaseRegs.size() + (Base.Scale == 1) +
4034           (Base.UnfoldedOffset.isNonZero()) <=
4035       1)
4036     return;
4037 
4038   // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before
4039   // processing the formula.
4040   Base.unscale();
4041   SmallVector<const SCEV *, 4> Ops;
4042   Formula NewBase = Base;
4043   NewBase.BaseRegs.clear();
4044   Type *CombinedIntegerType = nullptr;
4045   for (const SCEV *BaseReg : Base.BaseRegs) {
4046     if (SE.properlyDominates(BaseReg, L->getHeader()) &&
4047         !SE.hasComputableLoopEvolution(BaseReg, L)) {
4048       if (!CombinedIntegerType)
4049         CombinedIntegerType = SE.getEffectiveSCEVType(BaseReg->getType());
4050       Ops.push_back(BaseReg);
4051     }
4052     else
4053       NewBase.BaseRegs.push_back(BaseReg);
4054   }
4055 
4056   // If no register is relevant, we're done.
4057   if (Ops.size() == 0)
4058     return;
4059 
4060   // Utility function for generating the required variants of the combined
4061   // registers.
4062   auto GenerateFormula = [&](const SCEV *Sum) {
4063     Formula F = NewBase;
4064 
4065     // TODO: If Sum is zero, it probably means ScalarEvolution missed an
4066     // opportunity to fold something. For now, just ignore such cases
4067     // rather than proceed with zero in a register.
4068     if (Sum->isZero())
4069       return;
4070 
4071     F.BaseRegs.push_back(Sum);
4072     F.canonicalize(*L);
4073     (void)InsertFormula(LU, LUIdx, F);
4074   };
4075 
4076   // If we collected at least two registers, generate a formula combining them.
4077   if (Ops.size() > 1) {
4078     SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops.
4079     GenerateFormula(SE.getAddExpr(OpsCopy));
4080   }
4081 
4082   // If we have an unfolded offset, generate a formula combining it with the
4083   // registers collected.
4084   if (NewBase.UnfoldedOffset.isNonZero() && NewBase.UnfoldedOffset.isFixed()) {
4085     assert(CombinedIntegerType && "Missing a type for the unfolded offset");
4086     Ops.push_back(SE.getConstant(CombinedIntegerType,
4087                                  NewBase.UnfoldedOffset.getFixedValue(), true));
4088     NewBase.UnfoldedOffset = Immediate::getFixed(0);
4089     GenerateFormula(SE.getAddExpr(Ops));
4090   }
4091 }
4092 
4093 /// Helper function for LSRInstance::GenerateSymbolicOffsets.
GenerateSymbolicOffsetsImpl(LSRUse & LU,unsigned LUIdx,const Formula & Base,size_t Idx,bool IsScaledReg)4094 void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx,
4095                                               const Formula &Base, size_t Idx,
4096                                               bool IsScaledReg) {
4097   const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
4098   GlobalValue *GV = ExtractSymbol(G, SE);
4099   if (G->isZero() || !GV)
4100     return;
4101   Formula F = Base;
4102   F.BaseGV = GV;
4103   if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
4104     return;
4105   if (IsScaledReg)
4106     F.ScaledReg = G;
4107   else
4108     F.BaseRegs[Idx] = G;
4109   (void)InsertFormula(LU, LUIdx, F);
4110 }
4111 
4112 /// Generate reuse formulae using symbolic offsets.
GenerateSymbolicOffsets(LSRUse & LU,unsigned LUIdx,Formula Base)4113 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
4114                                           Formula Base) {
4115   // We can't add a symbolic offset if the address already contains one.
4116   if (Base.BaseGV) return;
4117 
4118   for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
4119     GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i);
4120   if (Base.Scale == 1)
4121     GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1,
4122                                 /* IsScaledReg */ true);
4123 }
4124 
4125 /// Helper function for LSRInstance::GenerateConstantOffsets.
GenerateConstantOffsetsImpl(LSRUse & LU,unsigned LUIdx,const Formula & Base,const SmallVectorImpl<Immediate> & Worklist,size_t Idx,bool IsScaledReg)4126 void LSRInstance::GenerateConstantOffsetsImpl(
4127     LSRUse &LU, unsigned LUIdx, const Formula &Base,
4128     const SmallVectorImpl<Immediate> &Worklist, size_t Idx, bool IsScaledReg) {
4129 
4130   auto GenerateOffset = [&](const SCEV *G, Immediate Offset) {
4131     Formula F = Base;
4132     if (!Base.BaseOffset.isCompatibleImmediate(Offset))
4133       return;
4134     F.BaseOffset = Base.BaseOffset.subUnsigned(Offset);
4135 
4136     if (isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) {
4137       // Add the offset to the base register.
4138       const SCEV *NewOffset = Offset.getSCEV(SE, G->getType());
4139       const SCEV *NewG = SE.getAddExpr(NewOffset, G);
4140       // If it cancelled out, drop the base register, otherwise update it.
4141       if (NewG->isZero()) {
4142         if (IsScaledReg) {
4143           F.Scale = 0;
4144           F.ScaledReg = nullptr;
4145         } else
4146           F.deleteBaseReg(F.BaseRegs[Idx]);
4147         F.canonicalize(*L);
4148       } else if (IsScaledReg)
4149         F.ScaledReg = NewG;
4150       else
4151         F.BaseRegs[Idx] = NewG;
4152 
4153       (void)InsertFormula(LU, LUIdx, F);
4154     }
4155   };
4156 
4157   const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx];
4158 
4159   // With constant offsets and constant steps, we can generate pre-inc
4160   // accesses by having the offset equal the step. So, for access #0 with a
4161   // step of 8, we generate a G - 8 base which would require the first access
4162   // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer
4163   // for itself and hopefully becomes the base for other accesses. This means
4164   // means that a single pre-indexed access can be generated to become the new
4165   // base pointer for each iteration of the loop, resulting in no extra add/sub
4166   // instructions for pointer updating.
4167   if (AMK == TTI::AMK_PreIndexed && LU.Kind == LSRUse::Address) {
4168     if (auto *GAR = dyn_cast<SCEVAddRecExpr>(G)) {
4169       if (auto *StepRec =
4170           dyn_cast<SCEVConstant>(GAR->getStepRecurrence(SE))) {
4171         const APInt &StepInt = StepRec->getAPInt();
4172         int64_t Step = StepInt.isNegative() ?
4173           StepInt.getSExtValue() : StepInt.getZExtValue();
4174 
4175         for (Immediate Offset : Worklist) {
4176           if (Offset.isFixed()) {
4177             Offset = Immediate::getFixed(Offset.getFixedValue() - Step);
4178             GenerateOffset(G, Offset);
4179           }
4180         }
4181       }
4182     }
4183   }
4184   for (Immediate Offset : Worklist)
4185     GenerateOffset(G, Offset);
4186 
4187   Immediate Imm = ExtractImmediate(G, SE);
4188   if (G->isZero() || Imm.isZero() ||
4189       !Base.BaseOffset.isCompatibleImmediate(Imm))
4190     return;
4191   Formula F = Base;
4192   F.BaseOffset = F.BaseOffset.addUnsigned(Imm);
4193   if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
4194     return;
4195   if (IsScaledReg) {
4196     F.ScaledReg = G;
4197   } else {
4198     F.BaseRegs[Idx] = G;
4199     // We may generate non canonical Formula if G is a recurrent expr reg
4200     // related with current loop while F.ScaledReg is not.
4201     F.canonicalize(*L);
4202   }
4203   (void)InsertFormula(LU, LUIdx, F);
4204 }
4205 
4206 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
GenerateConstantOffsets(LSRUse & LU,unsigned LUIdx,Formula Base)4207 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
4208                                           Formula Base) {
4209   // TODO: For now, just add the min and max offset, because it usually isn't
4210   // worthwhile looking at everything inbetween.
4211   SmallVector<Immediate, 2> Worklist;
4212   Worklist.push_back(LU.MinOffset);
4213   if (LU.MaxOffset != LU.MinOffset)
4214     Worklist.push_back(LU.MaxOffset);
4215 
4216   for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
4217     GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i);
4218   if (Base.Scale == 1)
4219     GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1,
4220                                 /* IsScaledReg */ true);
4221 }
4222 
4223 /// For ICmpZero, check to see if we can scale up the comparison. For example, x
4224 /// == y -> x*c == y*c.
GenerateICmpZeroScales(LSRUse & LU,unsigned LUIdx,Formula Base)4225 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
4226                                          Formula Base) {
4227   if (LU.Kind != LSRUse::ICmpZero) return;
4228 
4229   // Determine the integer type for the base formula.
4230   Type *IntTy = Base.getType();
4231   if (!IntTy) return;
4232   if (SE.getTypeSizeInBits(IntTy) > 64) return;
4233 
4234   // Don't do this if there is more than one offset.
4235   if (LU.MinOffset != LU.MaxOffset) return;
4236 
4237   // Check if transformation is valid. It is illegal to multiply pointer.
4238   if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
4239     return;
4240   for (const SCEV *BaseReg : Base.BaseRegs)
4241     if (BaseReg->getType()->isPointerTy())
4242       return;
4243   assert(!Base.BaseGV && "ICmpZero use is not legal!");
4244 
4245   // Check each interesting stride.
4246   for (int64_t Factor : Factors) {
4247     // Check that Factor can be represented by IntTy
4248     if (!ConstantInt::isValueValidForType(IntTy, Factor))
4249       continue;
4250     // Check that the multiplication doesn't overflow.
4251     if (Base.BaseOffset.isMin() && Factor == -1)
4252       continue;
4253     // Not supporting scalable immediates.
4254     if (Base.BaseOffset.isNonZero() && Base.BaseOffset.isScalable())
4255       continue;
4256     Immediate NewBaseOffset = Base.BaseOffset.mulUnsigned(Factor);
4257     assert(Factor != 0 && "Zero factor not expected!");
4258     if (NewBaseOffset.getFixedValue() / Factor !=
4259         Base.BaseOffset.getFixedValue())
4260       continue;
4261     // If the offset will be truncated at this use, check that it is in bounds.
4262     if (!IntTy->isPointerTy() &&
4263         !ConstantInt::isValueValidForType(IntTy, NewBaseOffset.getFixedValue()))
4264       continue;
4265 
4266     // Check that multiplying with the use offset doesn't overflow.
4267     Immediate Offset = LU.MinOffset;
4268     if (Offset.isMin() && Factor == -1)
4269       continue;
4270     Offset = Offset.mulUnsigned(Factor);
4271     if (Offset.getFixedValue() / Factor != LU.MinOffset.getFixedValue())
4272       continue;
4273     // If the offset will be truncated at this use, check that it is in bounds.
4274     if (!IntTy->isPointerTy() &&
4275         !ConstantInt::isValueValidForType(IntTy, Offset.getFixedValue()))
4276       continue;
4277 
4278     Formula F = Base;
4279     F.BaseOffset = NewBaseOffset;
4280 
4281     // Check that this scale is legal.
4282     if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
4283       continue;
4284 
4285     // Compensate for the use having MinOffset built into it.
4286     F.BaseOffset = F.BaseOffset.addUnsigned(Offset).subUnsigned(LU.MinOffset);
4287 
4288     const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4289 
4290     // Check that multiplying with each base register doesn't overflow.
4291     for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
4292       F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
4293       if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
4294         goto next;
4295     }
4296 
4297     // Check that multiplying with the scaled register doesn't overflow.
4298     if (F.ScaledReg) {
4299       F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
4300       if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
4301         continue;
4302     }
4303 
4304     // Check that multiplying with the unfolded offset doesn't overflow.
4305     if (F.UnfoldedOffset.isNonZero()) {
4306       if (F.UnfoldedOffset.isMin() && Factor == -1)
4307         continue;
4308       F.UnfoldedOffset = F.UnfoldedOffset.mulUnsigned(Factor);
4309       if (F.UnfoldedOffset.getFixedValue() / Factor !=
4310           Base.UnfoldedOffset.getFixedValue())
4311         continue;
4312       // If the offset will be truncated, check that it is in bounds.
4313       if (!IntTy->isPointerTy() && !ConstantInt::isValueValidForType(
4314                                        IntTy, F.UnfoldedOffset.getFixedValue()))
4315         continue;
4316     }
4317 
4318     // If we make it here and it's legal, add it.
4319     (void)InsertFormula(LU, LUIdx, F);
4320   next:;
4321   }
4322 }
4323 
4324 /// Generate stride factor reuse formulae by making use of scaled-offset address
4325 /// modes, for example.
GenerateScales(LSRUse & LU,unsigned LUIdx,Formula Base)4326 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
4327   // Determine the integer type for the base formula.
4328   Type *IntTy = Base.getType();
4329   if (!IntTy) return;
4330 
4331   // If this Formula already has a scaled register, we can't add another one.
4332   // Try to unscale the formula to generate a better scale.
4333   if (Base.Scale != 0 && !Base.unscale())
4334     return;
4335 
4336   assert(Base.Scale == 0 && "unscale did not did its job!");
4337 
4338   // Check each interesting stride.
4339   for (int64_t Factor : Factors) {
4340     Base.Scale = Factor;
4341     Base.HasBaseReg = Base.BaseRegs.size() > 1;
4342     // Check whether this scale is going to be legal.
4343     if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4344                     Base)) {
4345       // As a special-case, handle special out-of-loop Basic users specially.
4346       // TODO: Reconsider this special case.
4347       if (LU.Kind == LSRUse::Basic &&
4348           isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
4349                      LU.AccessTy, Base) &&
4350           LU.AllFixupsOutsideLoop)
4351         LU.Kind = LSRUse::Special;
4352       else
4353         continue;
4354     }
4355     // For an ICmpZero, negating a solitary base register won't lead to
4356     // new solutions.
4357     if (LU.Kind == LSRUse::ICmpZero && !Base.HasBaseReg &&
4358         Base.BaseOffset.isZero() && !Base.BaseGV)
4359       continue;
4360     // For each addrec base reg, if its loop is current loop, apply the scale.
4361     for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
4362       const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
4363       if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
4364         const SCEV *FactorS = SE.getConstant(IntTy, Factor);
4365         if (FactorS->isZero())
4366           continue;
4367         // Divide out the factor, ignoring high bits, since we'll be
4368         // scaling the value back up in the end.
4369         if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true))
4370           if (!Quotient->isZero()) {
4371             // TODO: This could be optimized to avoid all the copying.
4372             Formula F = Base;
4373             F.ScaledReg = Quotient;
4374             F.deleteBaseReg(F.BaseRegs[i]);
4375             // The canonical representation of 1*reg is reg, which is already in
4376             // Base. In that case, do not try to insert the formula, it will be
4377             // rejected anyway.
4378             if (F.Scale == 1 && (F.BaseRegs.empty() ||
4379                                  (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
4380               continue;
4381             // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
4382             // non canonical Formula with ScaledReg's loop not being L.
4383             if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
4384               F.canonicalize(*L);
4385             (void)InsertFormula(LU, LUIdx, F);
4386           }
4387       }
4388     }
4389   }
4390 }
4391 
4392 /// Extend/Truncate \p Expr to \p ToTy considering post-inc uses in \p Loops.
4393 /// For all PostIncLoopSets in \p Loops, first de-normalize \p Expr, then
4394 /// perform the extension/truncate and normalize again, as the normalized form
4395 /// can result in folds that are not valid in the post-inc use contexts. The
4396 /// expressions for all PostIncLoopSets must match, otherwise return nullptr.
4397 static const SCEV *
getAnyExtendConsideringPostIncUses(ArrayRef<PostIncLoopSet> Loops,const SCEV * Expr,Type * ToTy,ScalarEvolution & SE)4398 getAnyExtendConsideringPostIncUses(ArrayRef<PostIncLoopSet> Loops,
4399                                    const SCEV *Expr, Type *ToTy,
4400                                    ScalarEvolution &SE) {
4401   const SCEV *Result = nullptr;
4402   for (auto &L : Loops) {
4403     auto *DenormExpr = denormalizeForPostIncUse(Expr, L, SE);
4404     const SCEV *NewDenormExpr = SE.getAnyExtendExpr(DenormExpr, ToTy);
4405     const SCEV *New = normalizeForPostIncUse(NewDenormExpr, L, SE);
4406     if (!New || (Result && New != Result))
4407       return nullptr;
4408     Result = New;
4409   }
4410 
4411   assert(Result && "failed to create expression");
4412   return Result;
4413 }
4414 
4415 /// Generate reuse formulae from different IV types.
GenerateTruncates(LSRUse & LU,unsigned LUIdx,Formula Base)4416 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
4417   // Don't bother truncating symbolic values.
4418   if (Base.BaseGV) return;
4419 
4420   // Determine the integer type for the base formula.
4421   Type *DstTy = Base.getType();
4422   if (!DstTy) return;
4423   if (DstTy->isPointerTy())
4424     return;
4425 
4426   // It is invalid to extend a pointer type so exit early if ScaledReg or
4427   // any of the BaseRegs are pointers.
4428   if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy())
4429     return;
4430   if (any_of(Base.BaseRegs,
4431              [](const SCEV *S) { return S->getType()->isPointerTy(); }))
4432     return;
4433 
4434   SmallVector<PostIncLoopSet> Loops;
4435   for (auto &LF : LU.Fixups)
4436     Loops.push_back(LF.PostIncLoops);
4437 
4438   for (Type *SrcTy : Types) {
4439     if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
4440       Formula F = Base;
4441 
4442       // Sometimes SCEV is able to prove zero during ext transform. It may
4443       // happen if SCEV did not do all possible transforms while creating the
4444       // initial node (maybe due to depth limitations), but it can do them while
4445       // taking ext.
4446       if (F.ScaledReg) {
4447         const SCEV *NewScaledReg =
4448             getAnyExtendConsideringPostIncUses(Loops, F.ScaledReg, SrcTy, SE);
4449         if (!NewScaledReg || NewScaledReg->isZero())
4450           continue;
4451         F.ScaledReg = NewScaledReg;
4452       }
4453       bool HasZeroBaseReg = false;
4454       for (const SCEV *&BaseReg : F.BaseRegs) {
4455         const SCEV *NewBaseReg =
4456             getAnyExtendConsideringPostIncUses(Loops, BaseReg, SrcTy, SE);
4457         if (!NewBaseReg || NewBaseReg->isZero()) {
4458           HasZeroBaseReg = true;
4459           break;
4460         }
4461         BaseReg = NewBaseReg;
4462       }
4463       if (HasZeroBaseReg)
4464         continue;
4465 
4466       // TODO: This assumes we've done basic processing on all uses and
4467       // have an idea what the register usage is.
4468       if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
4469         continue;
4470 
4471       F.canonicalize(*L);
4472       (void)InsertFormula(LU, LUIdx, F);
4473     }
4474   }
4475 }
4476 
4477 namespace {
4478 
4479 /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer
4480 /// modifications so that the search phase doesn't have to worry about the data
4481 /// structures moving underneath it.
4482 struct WorkItem {
4483   size_t LUIdx;
4484   Immediate Imm;
4485   const SCEV *OrigReg;
4486 
WorkItem__anonc21373340e11::WorkItem4487   WorkItem(size_t LI, Immediate I, const SCEV *R)
4488       : LUIdx(LI), Imm(I), OrigReg(R) {}
4489 
4490   void print(raw_ostream &OS) const;
4491   void dump() const;
4492 };
4493 
4494 } // end anonymous namespace
4495 
4496 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print(raw_ostream & OS) const4497 void WorkItem::print(raw_ostream &OS) const {
4498   OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
4499      << " , add offset " << Imm;
4500 }
4501 
dump() const4502 LLVM_DUMP_METHOD void WorkItem::dump() const {
4503   print(errs()); errs() << '\n';
4504 }
4505 #endif
4506 
4507 /// Look for registers which are a constant distance apart and try to form reuse
4508 /// opportunities between them.
GenerateCrossUseConstantOffsets()4509 void LSRInstance::GenerateCrossUseConstantOffsets() {
4510   // Group the registers by their value without any added constant offset.
4511   using ImmMapTy = std::map<Immediate, const SCEV *, KeyOrderTargetImmediate>;
4512 
4513   DenseMap<const SCEV *, ImmMapTy> Map;
4514   DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
4515   SmallVector<const SCEV *, 8> Sequence;
4516   for (const SCEV *Use : RegUses) {
4517     const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify.
4518     Immediate Imm = ExtractImmediate(Reg, SE);
4519     auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy()));
4520     if (Pair.second)
4521       Sequence.push_back(Reg);
4522     Pair.first->second.insert(std::make_pair(Imm, Use));
4523     UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use);
4524   }
4525 
4526   // Now examine each set of registers with the same base value. Build up
4527   // a list of work to do and do the work in a separate step so that we're
4528   // not adding formulae and register counts while we're searching.
4529   SmallVector<WorkItem, 32> WorkItems;
4530   SmallSet<std::pair<size_t, Immediate>, 32, KeyOrderSizeTAndImmediate>
4531       UniqueItems;
4532   for (const SCEV *Reg : Sequence) {
4533     const ImmMapTy &Imms = Map.find(Reg)->second;
4534 
4535     // It's not worthwhile looking for reuse if there's only one offset.
4536     if (Imms.size() == 1)
4537       continue;
4538 
4539     LLVM_DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
4540                for (const auto &Entry
4541                     : Imms) dbgs()
4542                << ' ' << Entry.first;
4543                dbgs() << '\n');
4544 
4545     // Examine each offset.
4546     for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
4547          J != JE; ++J) {
4548       const SCEV *OrigReg = J->second;
4549 
4550       Immediate JImm = J->first;
4551       const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
4552 
4553       if (!isa<SCEVConstant>(OrigReg) &&
4554           UsedByIndicesMap[Reg].count() == 1) {
4555         LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg
4556                           << '\n');
4557         continue;
4558       }
4559 
4560       // Conservatively examine offsets between this orig reg a few selected
4561       // other orig regs.
4562       Immediate First = Imms.begin()->first;
4563       Immediate Last = std::prev(Imms.end())->first;
4564       if (!First.isCompatibleImmediate(Last)) {
4565         LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg
4566                           << "\n");
4567         continue;
4568       }
4569       // Only scalable if both terms are scalable, or if one is scalable and
4570       // the other is 0.
4571       bool Scalable = First.isScalable() || Last.isScalable();
4572       int64_t FI = First.getKnownMinValue();
4573       int64_t LI = Last.getKnownMinValue();
4574       // Compute (First + Last)  / 2 without overflow using the fact that
4575       // First + Last = 2 * (First + Last) + (First ^ Last).
4576       int64_t Avg = (FI & LI) + ((FI ^ LI) >> 1);
4577       // If the result is negative and FI is odd and LI even (or vice versa),
4578       // we rounded towards -inf. Add 1 in that case, to round towards 0.
4579       Avg = Avg + ((FI ^ LI) & ((uint64_t)Avg >> 63));
4580       ImmMapTy::const_iterator OtherImms[] = {
4581           Imms.begin(), std::prev(Imms.end()),
4582           Imms.lower_bound(Immediate::get(Avg, Scalable))};
4583       for (const auto &M : OtherImms) {
4584         if (M == J || M == JE) continue;
4585         if (!JImm.isCompatibleImmediate(M->first))
4586           continue;
4587 
4588         // Compute the difference between the two.
4589         Immediate Imm = JImm.subUnsigned(M->first);
4590         for (unsigned LUIdx : UsedByIndices.set_bits())
4591           // Make a memo of this use, offset, and register tuple.
4592           if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second)
4593             WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
4594       }
4595     }
4596   }
4597 
4598   Map.clear();
4599   Sequence.clear();
4600   UsedByIndicesMap.clear();
4601   UniqueItems.clear();
4602 
4603   // Now iterate through the worklist and add new formulae.
4604   for (const WorkItem &WI : WorkItems) {
4605     size_t LUIdx = WI.LUIdx;
4606     LSRUse &LU = Uses[LUIdx];
4607     Immediate Imm = WI.Imm;
4608     const SCEV *OrigReg = WI.OrigReg;
4609 
4610     Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
4611     const SCEV *NegImmS = Imm.getNegativeSCEV(SE, IntTy);
4612     unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
4613 
4614     // TODO: Use a more targeted data structure.
4615     for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
4616       Formula F = LU.Formulae[L];
4617       // FIXME: The code for the scaled and unscaled registers looks
4618       // very similar but slightly different. Investigate if they
4619       // could be merged. That way, we would not have to unscale the
4620       // Formula.
4621       F.unscale();
4622       // Use the immediate in the scaled register.
4623       if (F.ScaledReg == OrigReg) {
4624         if (!F.BaseOffset.isCompatibleImmediate(Imm))
4625           continue;
4626         Immediate Offset = F.BaseOffset.addUnsigned(Imm.mulUnsigned(F.Scale));
4627         // Don't create 50 + reg(-50).
4628         const SCEV *S = Offset.getNegativeSCEV(SE, IntTy);
4629         if (F.referencesReg(S))
4630           continue;
4631         Formula NewF = F;
4632         NewF.BaseOffset = Offset;
4633         if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
4634                         NewF))
4635           continue;
4636         NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
4637 
4638         // If the new scale is a constant in a register, and adding the constant
4639         // value to the immediate would produce a value closer to zero than the
4640         // immediate itself, then the formula isn't worthwhile.
4641         if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg)) {
4642           // FIXME: Do we need to do something for scalable immediates here?
4643           //        A scalable SCEV won't be constant, but we might still have
4644           //        something in the offset? Bail out for now to be safe.
4645           if (NewF.BaseOffset.isNonZero() && NewF.BaseOffset.isScalable())
4646             continue;
4647           if (C->getValue()->isNegative() !=
4648                   (NewF.BaseOffset.isLessThanZero()) &&
4649               (C->getAPInt().abs() * APInt(BitWidth, F.Scale))
4650                   .ule(std::abs(NewF.BaseOffset.getFixedValue())))
4651             continue;
4652         }
4653 
4654         // OK, looks good.
4655         NewF.canonicalize(*this->L);
4656         (void)InsertFormula(LU, LUIdx, NewF);
4657       } else {
4658         // Use the immediate in a base register.
4659         for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
4660           const SCEV *BaseReg = F.BaseRegs[N];
4661           if (BaseReg != OrigReg)
4662             continue;
4663           Formula NewF = F;
4664           if (!NewF.BaseOffset.isCompatibleImmediate(Imm) ||
4665               !NewF.UnfoldedOffset.isCompatibleImmediate(Imm) ||
4666               !NewF.BaseOffset.isCompatibleImmediate(NewF.UnfoldedOffset))
4667             continue;
4668           NewF.BaseOffset = NewF.BaseOffset.addUnsigned(Imm);
4669           if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
4670                           LU.Kind, LU.AccessTy, NewF)) {
4671             if (AMK == TTI::AMK_PostIndexed &&
4672                 mayUsePostIncMode(TTI, LU, OrigReg, this->L, SE))
4673               continue;
4674             Immediate NewUnfoldedOffset = NewF.UnfoldedOffset.addUnsigned(Imm);
4675             if (!isLegalAddImmediate(TTI, NewUnfoldedOffset))
4676               continue;
4677             NewF = F;
4678             NewF.UnfoldedOffset = NewUnfoldedOffset;
4679           }
4680           NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
4681 
4682           // If the new formula has a constant in a register, and adding the
4683           // constant value to the immediate would produce a value closer to
4684           // zero than the immediate itself, then the formula isn't worthwhile.
4685           for (const SCEV *NewReg : NewF.BaseRegs)
4686             if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg)) {
4687               if (NewF.BaseOffset.isNonZero() && NewF.BaseOffset.isScalable())
4688                 goto skip_formula;
4689               if ((C->getAPInt() + NewF.BaseOffset.getFixedValue())
4690                       .abs()
4691                       .slt(std::abs(NewF.BaseOffset.getFixedValue())) &&
4692                   (C->getAPInt() + NewF.BaseOffset.getFixedValue())
4693                           .countr_zero() >=
4694                       (unsigned)llvm::countr_zero<uint64_t>(
4695                           NewF.BaseOffset.getFixedValue()))
4696                 goto skip_formula;
4697             }
4698 
4699           // Ok, looks good.
4700           NewF.canonicalize(*this->L);
4701           (void)InsertFormula(LU, LUIdx, NewF);
4702           break;
4703         skip_formula:;
4704         }
4705       }
4706     }
4707   }
4708 }
4709 
4710 /// Generate formulae for each use.
4711 void
GenerateAllReuseFormulae()4712 LSRInstance::GenerateAllReuseFormulae() {
4713   // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
4714   // queries are more precise.
4715   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4716     LSRUse &LU = Uses[LUIdx];
4717     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4718       GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
4719     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4720       GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
4721   }
4722   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4723     LSRUse &LU = Uses[LUIdx];
4724     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4725       GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
4726     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4727       GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
4728     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4729       GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
4730     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4731       GenerateScales(LU, LUIdx, LU.Formulae[i]);
4732   }
4733   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4734     LSRUse &LU = Uses[LUIdx];
4735     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
4736       GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
4737   }
4738 
4739   GenerateCrossUseConstantOffsets();
4740 
4741   LLVM_DEBUG(dbgs() << "\n"
4742                        "After generating reuse formulae:\n";
4743              print_uses(dbgs()));
4744 }
4745 
4746 /// If there are multiple formulae with the same set of registers used
4747 /// by other uses, pick the best one and delete the others.
FilterOutUndesirableDedicatedRegisters()4748 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
4749   DenseSet<const SCEV *> VisitedRegs;
4750   SmallPtrSet<const SCEV *, 16> Regs;
4751   SmallPtrSet<const SCEV *, 16> LoserRegs;
4752 #ifndef NDEBUG
4753   bool ChangedFormulae = false;
4754 #endif
4755 
4756   // Collect the best formula for each unique set of shared registers. This
4757   // is reset for each use.
4758   using BestFormulaeTy =
4759       DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>;
4760 
4761   BestFormulaeTy BestFormulae;
4762 
4763   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4764     LSRUse &LU = Uses[LUIdx];
4765     LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
4766                dbgs() << '\n');
4767 
4768     bool Any = false;
4769     for (size_t FIdx = 0, NumForms = LU.Formulae.size();
4770          FIdx != NumForms; ++FIdx) {
4771       Formula &F = LU.Formulae[FIdx];
4772 
4773       // Some formulas are instant losers. For example, they may depend on
4774       // nonexistent AddRecs from other loops. These need to be filtered
4775       // immediately, otherwise heuristics could choose them over others leading
4776       // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
4777       // avoids the need to recompute this information across formulae using the
4778       // same bad AddRec. Passing LoserRegs is also essential unless we remove
4779       // the corresponding bad register from the Regs set.
4780       Cost CostF(L, SE, TTI, AMK);
4781       Regs.clear();
4782       CostF.RateFormula(F, Regs, VisitedRegs, LU, &LoserRegs);
4783       if (CostF.isLoser()) {
4784         // During initial formula generation, undesirable formulae are generated
4785         // by uses within other loops that have some non-trivial address mode or
4786         // use the postinc form of the IV. LSR needs to provide these formulae
4787         // as the basis of rediscovering the desired formula that uses an AddRec
4788         // corresponding to the existing phi. Once all formulae have been
4789         // generated, these initial losers may be pruned.
4790         LLVM_DEBUG(dbgs() << "  Filtering loser "; F.print(dbgs());
4791                    dbgs() << "\n");
4792       }
4793       else {
4794         SmallVector<const SCEV *, 4> Key;
4795         for (const SCEV *Reg : F.BaseRegs) {
4796           if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
4797             Key.push_back(Reg);
4798         }
4799         if (F.ScaledReg &&
4800             RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
4801           Key.push_back(F.ScaledReg);
4802         // Unstable sort by host order ok, because this is only used for
4803         // uniquifying.
4804         llvm::sort(Key);
4805 
4806         std::pair<BestFormulaeTy::const_iterator, bool> P =
4807           BestFormulae.insert(std::make_pair(Key, FIdx));
4808         if (P.second)
4809           continue;
4810 
4811         Formula &Best = LU.Formulae[P.first->second];
4812 
4813         Cost CostBest(L, SE, TTI, AMK);
4814         Regs.clear();
4815         CostBest.RateFormula(Best, Regs, VisitedRegs, LU);
4816         if (CostF.isLess(CostBest))
4817           std::swap(F, Best);
4818         LLVM_DEBUG(dbgs() << "  Filtering out formula "; F.print(dbgs());
4819                    dbgs() << "\n"
4820                              "    in favor of formula ";
4821                    Best.print(dbgs()); dbgs() << '\n');
4822       }
4823 #ifndef NDEBUG
4824       ChangedFormulae = true;
4825 #endif
4826       LU.DeleteFormula(F);
4827       --FIdx;
4828       --NumForms;
4829       Any = true;
4830     }
4831 
4832     // Now that we've filtered out some formulae, recompute the Regs set.
4833     if (Any)
4834       LU.RecomputeRegs(LUIdx, RegUses);
4835 
4836     // Reset this to prepare for the next use.
4837     BestFormulae.clear();
4838   }
4839 
4840   LLVM_DEBUG(if (ChangedFormulae) {
4841     dbgs() << "\n"
4842               "After filtering out undesirable candidates:\n";
4843     print_uses(dbgs());
4844   });
4845 }
4846 
4847 /// Estimate the worst-case number of solutions the solver might have to
4848 /// consider. It almost never considers this many solutions because it prune the
4849 /// search space, but the pruning isn't always sufficient.
EstimateSearchSpaceComplexity() const4850 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
4851   size_t Power = 1;
4852   for (const LSRUse &LU : Uses) {
4853     size_t FSize = LU.Formulae.size();
4854     if (FSize >= ComplexityLimit) {
4855       Power = ComplexityLimit;
4856       break;
4857     }
4858     Power *= FSize;
4859     if (Power >= ComplexityLimit)
4860       break;
4861   }
4862   return Power;
4863 }
4864 
4865 /// When one formula uses a superset of the registers of another formula, it
4866 /// won't help reduce register pressure (though it may not necessarily hurt
4867 /// register pressure); remove it to simplify the system.
NarrowSearchSpaceByDetectingSupersets()4868 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
4869   if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
4870     LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
4871 
4872     LLVM_DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
4873                          "which use a superset of registers used by other "
4874                          "formulae.\n");
4875 
4876     for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4877       LSRUse &LU = Uses[LUIdx];
4878       bool Any = false;
4879       for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
4880         Formula &F = LU.Formulae[i];
4881         if (F.BaseOffset.isNonZero() && F.BaseOffset.isScalable())
4882           continue;
4883         // Look for a formula with a constant or GV in a register. If the use
4884         // also has a formula with that same value in an immediate field,
4885         // delete the one that uses a register.
4886         for (SmallVectorImpl<const SCEV *>::const_iterator
4887              I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
4888           if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
4889             Formula NewF = F;
4890             //FIXME: Formulas should store bitwidth to do wrapping properly.
4891             //       See PR41034.
4892             NewF.BaseOffset =
4893                 Immediate::getFixed(NewF.BaseOffset.getFixedValue() +
4894                                     (uint64_t)C->getValue()->getSExtValue());
4895             NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4896                                 (I - F.BaseRegs.begin()));
4897             if (LU.HasFormulaWithSameRegs(NewF)) {
4898               LLVM_DEBUG(dbgs() << "  Deleting "; F.print(dbgs());
4899                          dbgs() << '\n');
4900               LU.DeleteFormula(F);
4901               --i;
4902               --e;
4903               Any = true;
4904               break;
4905             }
4906           } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
4907             if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
4908               if (!F.BaseGV) {
4909                 Formula NewF = F;
4910                 NewF.BaseGV = GV;
4911                 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
4912                                     (I - F.BaseRegs.begin()));
4913                 if (LU.HasFormulaWithSameRegs(NewF)) {
4914                   LLVM_DEBUG(dbgs() << "  Deleting "; F.print(dbgs());
4915                              dbgs() << '\n');
4916                   LU.DeleteFormula(F);
4917                   --i;
4918                   --e;
4919                   Any = true;
4920                   break;
4921                 }
4922               }
4923           }
4924         }
4925       }
4926       if (Any)
4927         LU.RecomputeRegs(LUIdx, RegUses);
4928     }
4929 
4930     LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4931   }
4932 }
4933 
4934 /// When there are many registers for expressions like A, A+1, A+2, etc.,
4935 /// allocate a single register for them.
NarrowSearchSpaceByCollapsingUnrolledCode()4936 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
4937   if (EstimateSearchSpaceComplexity() < ComplexityLimit)
4938     return;
4939 
4940   LLVM_DEBUG(
4941       dbgs() << "The search space is too complex.\n"
4942                 "Narrowing the search space by assuming that uses separated "
4943                 "by a constant offset will use the same registers.\n");
4944 
4945   // This is especially useful for unrolled loops.
4946 
4947   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
4948     LSRUse &LU = Uses[LUIdx];
4949     for (const Formula &F : LU.Formulae) {
4950       if (F.BaseOffset.isZero() || (F.Scale != 0 && F.Scale != 1))
4951         continue;
4952 
4953       LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
4954       if (!LUThatHas)
4955         continue;
4956 
4957       if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
4958                               LU.Kind, LU.AccessTy))
4959         continue;
4960 
4961       LLVM_DEBUG(dbgs() << "  Deleting use "; LU.print(dbgs()); dbgs() << '\n');
4962 
4963       LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
4964 
4965       // Transfer the fixups of LU to LUThatHas.
4966       for (LSRFixup &Fixup : LU.Fixups) {
4967         Fixup.Offset += F.BaseOffset;
4968         LUThatHas->pushFixup(Fixup);
4969         LLVM_DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
4970       }
4971 
4972       // Delete formulae from the new use which are no longer legal.
4973       bool Any = false;
4974       for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
4975         Formula &F = LUThatHas->Formulae[i];
4976         if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
4977                         LUThatHas->Kind, LUThatHas->AccessTy, F)) {
4978           LLVM_DEBUG(dbgs() << "  Deleting "; F.print(dbgs()); dbgs() << '\n');
4979           LUThatHas->DeleteFormula(F);
4980           --i;
4981           --e;
4982           Any = true;
4983         }
4984       }
4985 
4986       if (Any)
4987         LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
4988 
4989       // Delete the old use.
4990       DeleteUse(LU, LUIdx);
4991       --LUIdx;
4992       --NumUses;
4993       break;
4994     }
4995   }
4996 
4997   LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
4998 }
4999 
5000 /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that
5001 /// we've done more filtering, as it may be able to find more formulae to
5002 /// eliminate.
NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters()5003 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
5004   if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
5005     LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
5006 
5007     LLVM_DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
5008                          "undesirable dedicated registers.\n");
5009 
5010     FilterOutUndesirableDedicatedRegisters();
5011 
5012     LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
5013   }
5014 }
5015 
5016 /// If a LSRUse has multiple formulae with the same ScaledReg and Scale.
5017 /// Pick the best one and delete the others.
5018 /// This narrowing heuristic is to keep as many formulae with different
5019 /// Scale and ScaledReg pair as possible while narrowing the search space.
5020 /// The benefit is that it is more likely to find out a better solution
5021 /// from a formulae set with more Scale and ScaledReg variations than
5022 /// a formulae set with the same Scale and ScaledReg. The picking winner
5023 /// reg heuristic will often keep the formulae with the same Scale and
5024 /// ScaledReg and filter others, and we want to avoid that if possible.
NarrowSearchSpaceByFilterFormulaWithSameScaledReg()5025 void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() {
5026   if (EstimateSearchSpaceComplexity() < ComplexityLimit)
5027     return;
5028 
5029   LLVM_DEBUG(
5030       dbgs() << "The search space is too complex.\n"
5031                 "Narrowing the search space by choosing the best Formula "
5032                 "from the Formulae with the same Scale and ScaledReg.\n");
5033 
5034   // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse.
5035   using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>;
5036 
5037   BestFormulaeTy BestFormulae;
5038 #ifndef NDEBUG
5039   bool ChangedFormulae = false;
5040 #endif
5041   DenseSet<const SCEV *> VisitedRegs;
5042   SmallPtrSet<const SCEV *, 16> Regs;
5043 
5044   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
5045     LSRUse &LU = Uses[LUIdx];
5046     LLVM_DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs());
5047                dbgs() << '\n');
5048 
5049     // Return true if Formula FA is better than Formula FB.
5050     auto IsBetterThan = [&](Formula &FA, Formula &FB) {
5051       // First we will try to choose the Formula with fewer new registers.
5052       // For a register used by current Formula, the more the register is
5053       // shared among LSRUses, the less we increase the register number
5054       // counter of the formula.
5055       size_t FARegNum = 0;
5056       for (const SCEV *Reg : FA.BaseRegs) {
5057         const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
5058         FARegNum += (NumUses - UsedByIndices.count() + 1);
5059       }
5060       size_t FBRegNum = 0;
5061       for (const SCEV *Reg : FB.BaseRegs) {
5062         const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg);
5063         FBRegNum += (NumUses - UsedByIndices.count() + 1);
5064       }
5065       if (FARegNum != FBRegNum)
5066         return FARegNum < FBRegNum;
5067 
5068       // If the new register numbers are the same, choose the Formula with
5069       // less Cost.
5070       Cost CostFA(L, SE, TTI, AMK);
5071       Cost CostFB(L, SE, TTI, AMK);
5072       Regs.clear();
5073       CostFA.RateFormula(FA, Regs, VisitedRegs, LU);
5074       Regs.clear();
5075       CostFB.RateFormula(FB, Regs, VisitedRegs, LU);
5076       return CostFA.isLess(CostFB);
5077     };
5078 
5079     bool Any = false;
5080     for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
5081          ++FIdx) {
5082       Formula &F = LU.Formulae[FIdx];
5083       if (!F.ScaledReg)
5084         continue;
5085       auto P = BestFormulae.insert({{F.ScaledReg, F.Scale}, FIdx});
5086       if (P.second)
5087         continue;
5088 
5089       Formula &Best = LU.Formulae[P.first->second];
5090       if (IsBetterThan(F, Best))
5091         std::swap(F, Best);
5092       LLVM_DEBUG(dbgs() << "  Filtering out formula "; F.print(dbgs());
5093                  dbgs() << "\n"
5094                            "    in favor of formula ";
5095                  Best.print(dbgs()); dbgs() << '\n');
5096 #ifndef NDEBUG
5097       ChangedFormulae = true;
5098 #endif
5099       LU.DeleteFormula(F);
5100       --FIdx;
5101       --NumForms;
5102       Any = true;
5103     }
5104     if (Any)
5105       LU.RecomputeRegs(LUIdx, RegUses);
5106 
5107     // Reset this to prepare for the next use.
5108     BestFormulae.clear();
5109   }
5110 
5111   LLVM_DEBUG(if (ChangedFormulae) {
5112     dbgs() << "\n"
5113               "After filtering out undesirable candidates:\n";
5114     print_uses(dbgs());
5115   });
5116 }
5117 
5118 /// If we are over the complexity limit, filter out any post-inc prefering
5119 /// variables to only post-inc values.
NarrowSearchSpaceByFilterPostInc()5120 void LSRInstance::NarrowSearchSpaceByFilterPostInc() {
5121   if (AMK != TTI::AMK_PostIndexed)
5122     return;
5123   if (EstimateSearchSpaceComplexity() < ComplexityLimit)
5124     return;
5125 
5126   LLVM_DEBUG(dbgs() << "The search space is too complex.\n"
5127                        "Narrowing the search space by choosing the lowest "
5128                        "register Formula for PostInc Uses.\n");
5129 
5130   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
5131     LSRUse &LU = Uses[LUIdx];
5132 
5133     if (LU.Kind != LSRUse::Address)
5134       continue;
5135     if (!TTI.isIndexedLoadLegal(TTI.MIM_PostInc, LU.AccessTy.getType()) &&
5136         !TTI.isIndexedStoreLegal(TTI.MIM_PostInc, LU.AccessTy.getType()))
5137       continue;
5138 
5139     size_t MinRegs = std::numeric_limits<size_t>::max();
5140     for (const Formula &F : LU.Formulae)
5141       MinRegs = std::min(F.getNumRegs(), MinRegs);
5142 
5143     bool Any = false;
5144     for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms;
5145          ++FIdx) {
5146       Formula &F = LU.Formulae[FIdx];
5147       if (F.getNumRegs() > MinRegs) {
5148         LLVM_DEBUG(dbgs() << "  Filtering out formula "; F.print(dbgs());
5149                    dbgs() << "\n");
5150         LU.DeleteFormula(F);
5151         --FIdx;
5152         --NumForms;
5153         Any = true;
5154       }
5155     }
5156     if (Any)
5157       LU.RecomputeRegs(LUIdx, RegUses);
5158 
5159     if (EstimateSearchSpaceComplexity() < ComplexityLimit)
5160       break;
5161   }
5162 
5163   LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
5164 }
5165 
5166 /// The function delete formulas with high registers number expectation.
5167 /// Assuming we don't know the value of each formula (already delete
5168 /// all inefficient), generate probability of not selecting for each
5169 /// register.
5170 /// For example,
5171 /// Use1:
5172 ///  reg(a) + reg({0,+,1})
5173 ///  reg(a) + reg({-1,+,1}) + 1
5174 ///  reg({a,+,1})
5175 /// Use2:
5176 ///  reg(b) + reg({0,+,1})
5177 ///  reg(b) + reg({-1,+,1}) + 1
5178 ///  reg({b,+,1})
5179 /// Use3:
5180 ///  reg(c) + reg(b) + reg({0,+,1})
5181 ///  reg(c) + reg({b,+,1})
5182 ///
5183 /// Probability of not selecting
5184 ///                 Use1   Use2    Use3
5185 /// reg(a)         (1/3) *   1   *   1
5186 /// reg(b)           1   * (1/3) * (1/2)
5187 /// reg({0,+,1})   (2/3) * (2/3) * (1/2)
5188 /// reg({-1,+,1})  (2/3) * (2/3) *   1
5189 /// reg({a,+,1})   (2/3) *   1   *   1
5190 /// reg({b,+,1})     1   * (2/3) * (2/3)
5191 /// reg(c)           1   *   1   *   0
5192 ///
5193 /// Now count registers number mathematical expectation for each formula:
5194 /// Note that for each use we exclude probability if not selecting for the use.
5195 /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding
5196 /// probabilty 1/3 of not selecting for Use1).
5197 /// Use1:
5198 ///  reg(a) + reg({0,+,1})          1 + 1/3       -- to be deleted
5199 ///  reg(a) + reg({-1,+,1}) + 1     1 + 4/9       -- to be deleted
5200 ///  reg({a,+,1})                   1
5201 /// Use2:
5202 ///  reg(b) + reg({0,+,1})          1/2 + 1/3     -- to be deleted
5203 ///  reg(b) + reg({-1,+,1}) + 1     1/2 + 2/3     -- to be deleted
5204 ///  reg({b,+,1})                   2/3
5205 /// Use3:
5206 ///  reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted
5207 ///  reg(c) + reg({b,+,1})          1 + 2/3
NarrowSearchSpaceByDeletingCostlyFormulas()5208 void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() {
5209   if (EstimateSearchSpaceComplexity() < ComplexityLimit)
5210     return;
5211   // Ok, we have too many of formulae on our hands to conveniently handle.
5212   // Use a rough heuristic to thin out the list.
5213 
5214   // Set of Regs wich will be 100% used in final solution.
5215   // Used in each formula of a solution (in example above this is reg(c)).
5216   // We can skip them in calculations.
5217   SmallPtrSet<const SCEV *, 4> UniqRegs;
5218   LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
5219 
5220   // Map each register to probability of not selecting
5221   DenseMap <const SCEV *, float> RegNumMap;
5222   for (const SCEV *Reg : RegUses) {
5223     if (UniqRegs.count(Reg))
5224       continue;
5225     float PNotSel = 1;
5226     for (const LSRUse &LU : Uses) {
5227       if (!LU.Regs.count(Reg))
5228         continue;
5229       float P = LU.getNotSelectedProbability(Reg);
5230       if (P != 0.0)
5231         PNotSel *= P;
5232       else
5233         UniqRegs.insert(Reg);
5234     }
5235     RegNumMap.insert(std::make_pair(Reg, PNotSel));
5236   }
5237 
5238   LLVM_DEBUG(
5239       dbgs() << "Narrowing the search space by deleting costly formulas\n");
5240 
5241   // Delete formulas where registers number expectation is high.
5242   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
5243     LSRUse &LU = Uses[LUIdx];
5244     // If nothing to delete - continue.
5245     if (LU.Formulae.size() < 2)
5246       continue;
5247     // This is temporary solution to test performance. Float should be
5248     // replaced with round independent type (based on integers) to avoid
5249     // different results for different target builds.
5250     float FMinRegNum = LU.Formulae[0].getNumRegs();
5251     float FMinARegNum = LU.Formulae[0].getNumRegs();
5252     size_t MinIdx = 0;
5253     for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
5254       Formula &F = LU.Formulae[i];
5255       float FRegNum = 0;
5256       float FARegNum = 0;
5257       for (const SCEV *BaseReg : F.BaseRegs) {
5258         if (UniqRegs.count(BaseReg))
5259           continue;
5260         FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
5261         if (isa<SCEVAddRecExpr>(BaseReg))
5262           FARegNum +=
5263               RegNumMap[BaseReg] / LU.getNotSelectedProbability(BaseReg);
5264       }
5265       if (const SCEV *ScaledReg = F.ScaledReg) {
5266         if (!UniqRegs.count(ScaledReg)) {
5267           FRegNum +=
5268               RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
5269           if (isa<SCEVAddRecExpr>(ScaledReg))
5270             FARegNum +=
5271                 RegNumMap[ScaledReg] / LU.getNotSelectedProbability(ScaledReg);
5272         }
5273       }
5274       if (FMinRegNum > FRegNum ||
5275           (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) {
5276         FMinRegNum = FRegNum;
5277         FMinARegNum = FARegNum;
5278         MinIdx = i;
5279       }
5280     }
5281     LLVM_DEBUG(dbgs() << "  The formula "; LU.Formulae[MinIdx].print(dbgs());
5282                dbgs() << " with min reg num " << FMinRegNum << '\n');
5283     if (MinIdx != 0)
5284       std::swap(LU.Formulae[MinIdx], LU.Formulae[0]);
5285     while (LU.Formulae.size() != 1) {
5286       LLVM_DEBUG(dbgs() << "  Deleting "; LU.Formulae.back().print(dbgs());
5287                  dbgs() << '\n');
5288       LU.Formulae.pop_back();
5289     }
5290     LU.RecomputeRegs(LUIdx, RegUses);
5291     assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula");
5292     Formula &F = LU.Formulae[0];
5293     LLVM_DEBUG(dbgs() << "  Leaving only "; F.print(dbgs()); dbgs() << '\n');
5294     // When we choose the formula, the regs become unique.
5295     UniqRegs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
5296     if (F.ScaledReg)
5297       UniqRegs.insert(F.ScaledReg);
5298   }
5299   LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
5300 }
5301 
5302 // Check if Best and Reg are SCEVs separated by a constant amount C, and if so
5303 // would the addressing offset +C would be legal where the negative offset -C is
5304 // not.
IsSimplerBaseSCEVForTarget(const TargetTransformInfo & TTI,ScalarEvolution & SE,const SCEV * Best,const SCEV * Reg,MemAccessTy AccessType)5305 static bool IsSimplerBaseSCEVForTarget(const TargetTransformInfo &TTI,
5306                                        ScalarEvolution &SE, const SCEV *Best,
5307                                        const SCEV *Reg,
5308                                        MemAccessTy AccessType) {
5309   if (Best->getType() != Reg->getType() ||
5310       (isa<SCEVAddRecExpr>(Best) && isa<SCEVAddRecExpr>(Reg) &&
5311        cast<SCEVAddRecExpr>(Best)->getLoop() !=
5312            cast<SCEVAddRecExpr>(Reg)->getLoop()))
5313     return false;
5314   const auto *Diff = dyn_cast<SCEVConstant>(SE.getMinusSCEV(Best, Reg));
5315   if (!Diff)
5316     return false;
5317 
5318   return TTI.isLegalAddressingMode(
5319              AccessType.MemTy, /*BaseGV=*/nullptr,
5320              /*BaseOffset=*/Diff->getAPInt().getSExtValue(),
5321              /*HasBaseReg=*/true, /*Scale=*/0, AccessType.AddrSpace) &&
5322          !TTI.isLegalAddressingMode(
5323              AccessType.MemTy, /*BaseGV=*/nullptr,
5324              /*BaseOffset=*/-Diff->getAPInt().getSExtValue(),
5325              /*HasBaseReg=*/true, /*Scale=*/0, AccessType.AddrSpace);
5326 }
5327 
5328 /// Pick a register which seems likely to be profitable, and then in any use
5329 /// which has any reference to that register, delete all formulae which do not
5330 /// reference that register.
NarrowSearchSpaceByPickingWinnerRegs()5331 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
5332   // With all other options exhausted, loop until the system is simple
5333   // enough to handle.
5334   SmallPtrSet<const SCEV *, 4> Taken;
5335   while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
5336     // Ok, we have too many of formulae on our hands to conveniently handle.
5337     // Use a rough heuristic to thin out the list.
5338     LLVM_DEBUG(dbgs() << "The search space is too complex.\n");
5339 
5340     // Pick the register which is used by the most LSRUses, which is likely
5341     // to be a good reuse register candidate.
5342     const SCEV *Best = nullptr;
5343     unsigned BestNum = 0;
5344     for (const SCEV *Reg : RegUses) {
5345       if (Taken.count(Reg))
5346         continue;
5347       if (!Best) {
5348         Best = Reg;
5349         BestNum = RegUses.getUsedByIndices(Reg).count();
5350       } else {
5351         unsigned Count = RegUses.getUsedByIndices(Reg).count();
5352         if (Count > BestNum) {
5353           Best = Reg;
5354           BestNum = Count;
5355         }
5356 
5357         // If the scores are the same, but the Reg is simpler for the target
5358         // (for example {x,+,1} as opposed to {x+C,+,1}, where the target can
5359         // handle +C but not -C), opt for the simpler formula.
5360         if (Count == BestNum) {
5361           int LUIdx = RegUses.getUsedByIndices(Reg).find_first();
5362           if (LUIdx >= 0 && Uses[LUIdx].Kind == LSRUse::Address &&
5363               IsSimplerBaseSCEVForTarget(TTI, SE, Best, Reg,
5364                                          Uses[LUIdx].AccessTy)) {
5365             Best = Reg;
5366             BestNum = Count;
5367           }
5368         }
5369       }
5370     }
5371     assert(Best && "Failed to find best LSRUse candidate");
5372 
5373     LLVM_DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
5374                       << " will yield profitable reuse.\n");
5375     Taken.insert(Best);
5376 
5377     // In any use with formulae which references this register, delete formulae
5378     // which don't reference it.
5379     for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
5380       LSRUse &LU = Uses[LUIdx];
5381       if (!LU.Regs.count(Best)) continue;
5382 
5383       bool Any = false;
5384       for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
5385         Formula &F = LU.Formulae[i];
5386         if (!F.referencesReg(Best)) {
5387           LLVM_DEBUG(dbgs() << "  Deleting "; F.print(dbgs()); dbgs() << '\n');
5388           LU.DeleteFormula(F);
5389           --e;
5390           --i;
5391           Any = true;
5392           assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
5393           continue;
5394         }
5395       }
5396 
5397       if (Any)
5398         LU.RecomputeRegs(LUIdx, RegUses);
5399     }
5400 
5401     LLVM_DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
5402   }
5403 }
5404 
5405 /// If there are an extraordinary number of formulae to choose from, use some
5406 /// rough heuristics to prune down the number of formulae. This keeps the main
5407 /// solver from taking an extraordinary amount of time in some worst-case
5408 /// scenarios.
NarrowSearchSpaceUsingHeuristics()5409 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
5410   NarrowSearchSpaceByDetectingSupersets();
5411   NarrowSearchSpaceByCollapsingUnrolledCode();
5412   NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
5413   if (FilterSameScaledReg)
5414     NarrowSearchSpaceByFilterFormulaWithSameScaledReg();
5415   NarrowSearchSpaceByFilterPostInc();
5416   if (LSRExpNarrow)
5417     NarrowSearchSpaceByDeletingCostlyFormulas();
5418   else
5419     NarrowSearchSpaceByPickingWinnerRegs();
5420 }
5421 
5422 /// This is the recursive solver.
SolveRecurse(SmallVectorImpl<const Formula * > & Solution,Cost & SolutionCost,SmallVectorImpl<const Formula * > & Workspace,const Cost & CurCost,const SmallPtrSet<const SCEV *,16> & CurRegs,DenseSet<const SCEV * > & VisitedRegs) const5423 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
5424                                Cost &SolutionCost,
5425                                SmallVectorImpl<const Formula *> &Workspace,
5426                                const Cost &CurCost,
5427                                const SmallPtrSet<const SCEV *, 16> &CurRegs,
5428                                DenseSet<const SCEV *> &VisitedRegs) const {
5429   // Some ideas:
5430   //  - prune more:
5431   //    - use more aggressive filtering
5432   //    - sort the formula so that the most profitable solutions are found first
5433   //    - sort the uses too
5434   //  - search faster:
5435   //    - don't compute a cost, and then compare. compare while computing a cost
5436   //      and bail early.
5437   //    - track register sets with SmallBitVector
5438 
5439   const LSRUse &LU = Uses[Workspace.size()];
5440 
5441   // If this use references any register that's already a part of the
5442   // in-progress solution, consider it a requirement that a formula must
5443   // reference that register in order to be considered. This prunes out
5444   // unprofitable searching.
5445   SmallSetVector<const SCEV *, 4> ReqRegs;
5446   for (const SCEV *S : CurRegs)
5447     if (LU.Regs.count(S))
5448       ReqRegs.insert(S);
5449 
5450   SmallPtrSet<const SCEV *, 16> NewRegs;
5451   Cost NewCost(L, SE, TTI, AMK);
5452   for (const Formula &F : LU.Formulae) {
5453     // Ignore formulae which may not be ideal in terms of register reuse of
5454     // ReqRegs.  The formula should use all required registers before
5455     // introducing new ones.
5456     // This can sometimes (notably when trying to favour postinc) lead to
5457     // sub-optimial decisions. There it is best left to the cost modelling to
5458     // get correct.
5459     if (AMK != TTI::AMK_PostIndexed || LU.Kind != LSRUse::Address) {
5460       int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size());
5461       for (const SCEV *Reg : ReqRegs) {
5462         if ((F.ScaledReg && F.ScaledReg == Reg) ||
5463             is_contained(F.BaseRegs, Reg)) {
5464           --NumReqRegsToFind;
5465           if (NumReqRegsToFind == 0)
5466             break;
5467         }
5468       }
5469       if (NumReqRegsToFind != 0) {
5470         // If none of the formulae satisfied the required registers, then we could
5471         // clear ReqRegs and try again. Currently, we simply give up in this case.
5472         continue;
5473       }
5474     }
5475 
5476     // Evaluate the cost of the current formula. If it's already worse than
5477     // the current best, prune the search at that point.
5478     NewCost = CurCost;
5479     NewRegs = CurRegs;
5480     NewCost.RateFormula(F, NewRegs, VisitedRegs, LU);
5481     if (NewCost.isLess(SolutionCost)) {
5482       Workspace.push_back(&F);
5483       if (Workspace.size() != Uses.size()) {
5484         SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
5485                      NewRegs, VisitedRegs);
5486         if (F.getNumRegs() == 1 && Workspace.size() == 1)
5487           VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
5488       } else {
5489         LLVM_DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
5490                    dbgs() << ".\nRegs:\n";
5491                    for (const SCEV *S : NewRegs) dbgs()
5492                       << "- " << *S << "\n";
5493                    dbgs() << '\n');
5494 
5495         SolutionCost = NewCost;
5496         Solution = Workspace;
5497       }
5498       Workspace.pop_back();
5499     }
5500   }
5501 }
5502 
5503 /// Choose one formula from each use. Return the results in the given Solution
5504 /// vector.
Solve(SmallVectorImpl<const Formula * > & Solution) const5505 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
5506   SmallVector<const Formula *, 8> Workspace;
5507   Cost SolutionCost(L, SE, TTI, AMK);
5508   SolutionCost.Lose();
5509   Cost CurCost(L, SE, TTI, AMK);
5510   SmallPtrSet<const SCEV *, 16> CurRegs;
5511   DenseSet<const SCEV *> VisitedRegs;
5512   Workspace.reserve(Uses.size());
5513 
5514   // SolveRecurse does all the work.
5515   SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
5516                CurRegs, VisitedRegs);
5517   if (Solution.empty()) {
5518     LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
5519     return;
5520   }
5521 
5522   // Ok, we've now made all our decisions.
5523   LLVM_DEBUG(dbgs() << "\n"
5524                        "The chosen solution requires ";
5525              SolutionCost.print(dbgs()); dbgs() << ":\n";
5526              for (size_t i = 0, e = Uses.size(); i != e; ++i) {
5527                dbgs() << "  ";
5528                Uses[i].print(dbgs());
5529                dbgs() << "\n"
5530                          "    ";
5531                Solution[i]->print(dbgs());
5532                dbgs() << '\n';
5533              });
5534 
5535   assert(Solution.size() == Uses.size() && "Malformed solution!");
5536 
5537   const bool EnableDropUnprofitableSolution = [&] {
5538     switch (AllowDropSolutionIfLessProfitable) {
5539     case cl::BOU_TRUE:
5540       return true;
5541     case cl::BOU_FALSE:
5542       return false;
5543     case cl::BOU_UNSET:
5544       return TTI.shouldDropLSRSolutionIfLessProfitable();
5545     }
5546     llvm_unreachable("Unhandled cl::boolOrDefault enum");
5547   }();
5548 
5549   if (BaselineCost.isLess(SolutionCost)) {
5550     if (!EnableDropUnprofitableSolution)
5551       LLVM_DEBUG(
5552           dbgs() << "Baseline is more profitable than chosen solution, "
5553                     "add option 'lsr-drop-solution' to drop LSR solution.\n");
5554     else {
5555       LLVM_DEBUG(dbgs() << "Baseline is more profitable than chosen "
5556                            "solution, dropping LSR solution.\n";);
5557       Solution.clear();
5558     }
5559   }
5560 }
5561 
5562 /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as
5563 /// we can go while still being dominated by the input positions. This helps
5564 /// canonicalize the insert position, which encourages sharing.
5565 BasicBlock::iterator
HoistInsertPosition(BasicBlock::iterator IP,const SmallVectorImpl<Instruction * > & Inputs) const5566 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
5567                                  const SmallVectorImpl<Instruction *> &Inputs)
5568                                                                          const {
5569   Instruction *Tentative = &*IP;
5570   while (true) {
5571     bool AllDominate = true;
5572     Instruction *BetterPos = nullptr;
5573     // Don't bother attempting to insert before a catchswitch, their basic block
5574     // cannot have other non-PHI instructions.
5575     if (isa<CatchSwitchInst>(Tentative))
5576       return IP;
5577 
5578     for (Instruction *Inst : Inputs) {
5579       if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
5580         AllDominate = false;
5581         break;
5582       }
5583       // Attempt to find an insert position in the middle of the block,
5584       // instead of at the end, so that it can be used for other expansions.
5585       if (Tentative->getParent() == Inst->getParent() &&
5586           (!BetterPos || !DT.dominates(Inst, BetterPos)))
5587         BetterPos = &*std::next(BasicBlock::iterator(Inst));
5588     }
5589     if (!AllDominate)
5590       break;
5591     if (BetterPos)
5592       IP = BetterPos->getIterator();
5593     else
5594       IP = Tentative->getIterator();
5595 
5596     const Loop *IPLoop = LI.getLoopFor(IP->getParent());
5597     unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
5598 
5599     BasicBlock *IDom;
5600     for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
5601       if (!Rung) return IP;
5602       Rung = Rung->getIDom();
5603       if (!Rung) return IP;
5604       IDom = Rung->getBlock();
5605 
5606       // Don't climb into a loop though.
5607       const Loop *IDomLoop = LI.getLoopFor(IDom);
5608       unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
5609       if (IDomDepth <= IPLoopDepth &&
5610           (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
5611         break;
5612     }
5613 
5614     Tentative = IDom->getTerminator();
5615   }
5616 
5617   return IP;
5618 }
5619 
5620 /// Determine an input position which will be dominated by the operands and
5621 /// which will dominate the result.
AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,const LSRFixup & LF,const LSRUse & LU) const5622 BasicBlock::iterator LSRInstance::AdjustInsertPositionForExpand(
5623     BasicBlock::iterator LowestIP, const LSRFixup &LF, const LSRUse &LU) const {
5624   // Collect some instructions which must be dominated by the
5625   // expanding replacement. These must be dominated by any operands that
5626   // will be required in the expansion.
5627   SmallVector<Instruction *, 4> Inputs;
5628   if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
5629     Inputs.push_back(I);
5630   if (LU.Kind == LSRUse::ICmpZero)
5631     if (Instruction *I =
5632           dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
5633       Inputs.push_back(I);
5634   if (LF.PostIncLoops.count(L)) {
5635     if (LF.isUseFullyOutsideLoop(L))
5636       Inputs.push_back(L->getLoopLatch()->getTerminator());
5637     else
5638       Inputs.push_back(IVIncInsertPos);
5639   }
5640   // The expansion must also be dominated by the increment positions of any
5641   // loops it for which it is using post-inc mode.
5642   for (const Loop *PIL : LF.PostIncLoops) {
5643     if (PIL == L) continue;
5644 
5645     // Be dominated by the loop exit.
5646     SmallVector<BasicBlock *, 4> ExitingBlocks;
5647     PIL->getExitingBlocks(ExitingBlocks);
5648     if (!ExitingBlocks.empty()) {
5649       BasicBlock *BB = ExitingBlocks[0];
5650       for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
5651         BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
5652       Inputs.push_back(BB->getTerminator());
5653     }
5654   }
5655 
5656   assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad()
5657          && !isa<DbgInfoIntrinsic>(LowestIP) &&
5658          "Insertion point must be a normal instruction");
5659 
5660   // Then, climb up the immediate dominator tree as far as we can go while
5661   // still being dominated by the input positions.
5662   BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
5663 
5664   // Don't insert instructions before PHI nodes.
5665   while (isa<PHINode>(IP)) ++IP;
5666 
5667   // Ignore landingpad instructions.
5668   while (IP->isEHPad()) ++IP;
5669 
5670   // Ignore debug intrinsics.
5671   while (isa<DbgInfoIntrinsic>(IP)) ++IP;
5672 
5673   // Set IP below instructions recently inserted by SCEVExpander. This keeps the
5674   // IP consistent across expansions and allows the previously inserted
5675   // instructions to be reused by subsequent expansion.
5676   while (Rewriter.isInsertedInstruction(&*IP) && IP != LowestIP)
5677     ++IP;
5678 
5679   return IP;
5680 }
5681 
5682 /// Emit instructions for the leading candidate expression for this LSRUse (this
5683 /// is called "expanding").
Expand(const LSRUse & LU,const LSRFixup & LF,const Formula & F,BasicBlock::iterator IP,SmallVectorImpl<WeakTrackingVH> & DeadInsts) const5684 Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF,
5685                            const Formula &F, BasicBlock::iterator IP,
5686                            SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5687   if (LU.RigidFormula)
5688     return LF.OperandValToReplace;
5689 
5690   // Determine an input position which will be dominated by the operands and
5691   // which will dominate the result.
5692   IP = AdjustInsertPositionForExpand(IP, LF, LU);
5693   Rewriter.setInsertPoint(&*IP);
5694 
5695   // Inform the Rewriter if we have a post-increment use, so that it can
5696   // perform an advantageous expansion.
5697   Rewriter.setPostInc(LF.PostIncLoops);
5698 
5699   // This is the type that the user actually needs.
5700   Type *OpTy = LF.OperandValToReplace->getType();
5701   // This will be the type that we'll initially expand to.
5702   Type *Ty = F.getType();
5703   if (!Ty)
5704     // No type known; just expand directly to the ultimate type.
5705     Ty = OpTy;
5706   else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
5707     // Expand directly to the ultimate type if it's the right size.
5708     Ty = OpTy;
5709   // This is the type to do integer arithmetic in.
5710   Type *IntTy = SE.getEffectiveSCEVType(Ty);
5711 
5712   // Build up a list of operands to add together to form the full base.
5713   SmallVector<const SCEV *, 8> Ops;
5714 
5715   // Expand the BaseRegs portion.
5716   for (const SCEV *Reg : F.BaseRegs) {
5717     assert(!Reg->isZero() && "Zero allocated in a base register!");
5718 
5719     // If we're expanding for a post-inc user, make the post-inc adjustment.
5720     Reg = denormalizeForPostIncUse(Reg, LF.PostIncLoops, SE);
5721     Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr)));
5722   }
5723 
5724   // Expand the ScaledReg portion.
5725   Value *ICmpScaledV = nullptr;
5726   if (F.Scale != 0) {
5727     const SCEV *ScaledS = F.ScaledReg;
5728 
5729     // If we're expanding for a post-inc user, make the post-inc adjustment.
5730     PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
5731     ScaledS = denormalizeForPostIncUse(ScaledS, Loops, SE);
5732 
5733     if (LU.Kind == LSRUse::ICmpZero) {
5734       // Expand ScaleReg as if it was part of the base regs.
5735       if (F.Scale == 1)
5736         Ops.push_back(
5737             SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr)));
5738       else {
5739         // An interesting way of "folding" with an icmp is to use a negated
5740         // scale, which we'll implement by inserting it into the other operand
5741         // of the icmp.
5742         assert(F.Scale == -1 &&
5743                "The only scale supported by ICmpZero uses is -1!");
5744         ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr);
5745       }
5746     } else {
5747       // Otherwise just expand the scaled register and an explicit scale,
5748       // which is expected to be matched as part of the address.
5749 
5750       // Flush the operand list to suppress SCEVExpander hoisting address modes.
5751       // Unless the addressing mode will not be folded.
5752       if (!Ops.empty() && LU.Kind == LSRUse::Address &&
5753           isAMCompletelyFolded(TTI, LU, F)) {
5754         Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), nullptr);
5755         Ops.clear();
5756         Ops.push_back(SE.getUnknown(FullV));
5757       }
5758       ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr));
5759       if (F.Scale != 1)
5760         ScaledS =
5761             SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale));
5762       Ops.push_back(ScaledS);
5763     }
5764   }
5765 
5766   // Expand the GV portion.
5767   if (F.BaseGV) {
5768     // Flush the operand list to suppress SCEVExpander hoisting.
5769     if (!Ops.empty()) {
5770       Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), IntTy);
5771       Ops.clear();
5772       Ops.push_back(SE.getUnknown(FullV));
5773     }
5774     Ops.push_back(SE.getUnknown(F.BaseGV));
5775   }
5776 
5777   // Flush the operand list to suppress SCEVExpander hoisting of both folded and
5778   // unfolded offsets. LSR assumes they both live next to their uses.
5779   if (!Ops.empty()) {
5780     Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty);
5781     Ops.clear();
5782     Ops.push_back(SE.getUnknown(FullV));
5783   }
5784 
5785   // FIXME: Are we sure we won't get a mismatch here? Is there a way to bail
5786   // out at this point, or should we generate a SCEV adding together mixed
5787   // offsets?
5788   assert(F.BaseOffset.isCompatibleImmediate(LF.Offset) &&
5789          "Expanding mismatched offsets\n");
5790   // Expand the immediate portion.
5791   Immediate Offset = F.BaseOffset.addUnsigned(LF.Offset);
5792   if (Offset.isNonZero()) {
5793     if (LU.Kind == LSRUse::ICmpZero) {
5794       // The other interesting way of "folding" with an ICmpZero is to use a
5795       // negated immediate.
5796       if (!ICmpScaledV)
5797         ICmpScaledV =
5798             ConstantInt::get(IntTy, -(uint64_t)Offset.getFixedValue());
5799       else {
5800         Ops.push_back(SE.getUnknown(ICmpScaledV));
5801         ICmpScaledV = ConstantInt::get(IntTy, Offset.getFixedValue());
5802       }
5803     } else {
5804       // Just add the immediate values. These again are expected to be matched
5805       // as part of the address.
5806       Ops.push_back(Offset.getUnknownSCEV(SE, IntTy));
5807     }
5808   }
5809 
5810   // Expand the unfolded offset portion.
5811   Immediate UnfoldedOffset = F.UnfoldedOffset;
5812   if (UnfoldedOffset.isNonZero()) {
5813     // Just add the immediate values.
5814     Ops.push_back(UnfoldedOffset.getUnknownSCEV(SE, IntTy));
5815   }
5816 
5817   // Emit instructions summing all the operands.
5818   const SCEV *FullS = Ops.empty() ?
5819                       SE.getConstant(IntTy, 0) :
5820                       SE.getAddExpr(Ops);
5821   Value *FullV = Rewriter.expandCodeFor(FullS, Ty);
5822 
5823   // We're done expanding now, so reset the rewriter.
5824   Rewriter.clearPostInc();
5825 
5826   // An ICmpZero Formula represents an ICmp which we're handling as a
5827   // comparison against zero. Now that we've expanded an expression for that
5828   // form, update the ICmp's other operand.
5829   if (LU.Kind == LSRUse::ICmpZero) {
5830     ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
5831     if (auto *OperandIsInstr = dyn_cast<Instruction>(CI->getOperand(1)))
5832       DeadInsts.emplace_back(OperandIsInstr);
5833     assert(!F.BaseGV && "ICmp does not support folding a global value and "
5834                            "a scale at the same time!");
5835     if (F.Scale == -1) {
5836       if (ICmpScaledV->getType() != OpTy) {
5837         Instruction *Cast = CastInst::Create(
5838             CastInst::getCastOpcode(ICmpScaledV, false, OpTy, false),
5839             ICmpScaledV, OpTy, "tmp", CI->getIterator());
5840         ICmpScaledV = Cast;
5841       }
5842       CI->setOperand(1, ICmpScaledV);
5843     } else {
5844       // A scale of 1 means that the scale has been expanded as part of the
5845       // base regs.
5846       assert((F.Scale == 0 || F.Scale == 1) &&
5847              "ICmp does not support folding a global value and "
5848              "a scale at the same time!");
5849       Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
5850                                            -(uint64_t)Offset.getFixedValue());
5851       if (C->getType() != OpTy) {
5852         C = ConstantFoldCastOperand(
5853             CastInst::getCastOpcode(C, false, OpTy, false), C, OpTy,
5854             CI->getDataLayout());
5855         assert(C && "Cast of ConstantInt should have folded");
5856       }
5857 
5858       CI->setOperand(1, C);
5859     }
5860   }
5861 
5862   return FullV;
5863 }
5864 
5865 /// Helper for Rewrite. PHI nodes are special because the use of their operands
5866 /// effectively happens in their predecessor blocks, so the expression may need
5867 /// to be expanded in multiple places.
RewriteForPHI(PHINode * PN,const LSRUse & LU,const LSRFixup & LF,const Formula & F,SmallVectorImpl<WeakTrackingVH> & DeadInsts) const5868 void LSRInstance::RewriteForPHI(
5869     PHINode *PN, const LSRUse &LU, const LSRFixup &LF, const Formula &F,
5870     SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
5871   DenseMap<BasicBlock *, Value *> Inserted;
5872 
5873   // Inserting instructions in the loop and using them as PHI's input could
5874   // break LCSSA in case if PHI's parent block is not a loop exit (i.e. the
5875   // corresponding incoming block is not loop exiting). So collect all such
5876   // instructions to form LCSSA for them later.
5877   SmallVector<Instruction *, 4> InsertedNonLCSSAInsts;
5878 
5879   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
5880     if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
5881       bool needUpdateFixups = false;
5882       BasicBlock *BB = PN->getIncomingBlock(i);
5883 
5884       // If this is a critical edge, split the edge so that we do not insert
5885       // the code on all predecessor/successor paths.  We do this unless this
5886       // is the canonical backedge for this loop, which complicates post-inc
5887       // users.
5888       if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
5889           !isa<IndirectBrInst>(BB->getTerminator()) &&
5890           !isa<CatchSwitchInst>(BB->getTerminator())) {
5891         BasicBlock *Parent = PN->getParent();
5892         Loop *PNLoop = LI.getLoopFor(Parent);
5893         if (!PNLoop || Parent != PNLoop->getHeader()) {
5894           // Split the critical edge.
5895           BasicBlock *NewBB = nullptr;
5896           if (!Parent->isLandingPad()) {
5897             NewBB =
5898                 SplitCriticalEdge(BB, Parent,
5899                                   CriticalEdgeSplittingOptions(&DT, &LI, MSSAU)
5900                                       .setMergeIdenticalEdges()
5901                                       .setKeepOneInputPHIs());
5902           } else {
5903             SmallVector<BasicBlock*, 2> NewBBs;
5904             DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
5905             SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, &DTU, &LI);
5906             NewBB = NewBBs[0];
5907           }
5908           // If NewBB==NULL, then SplitCriticalEdge refused to split because all
5909           // phi predecessors are identical. The simple thing to do is skip
5910           // splitting in this case rather than complicate the API.
5911           if (NewBB) {
5912             // If PN is outside of the loop and BB is in the loop, we want to
5913             // move the block to be immediately before the PHI block, not
5914             // immediately after BB.
5915             if (L->contains(BB) && !L->contains(PN))
5916               NewBB->moveBefore(PN->getParent());
5917 
5918             // Splitting the edge can reduce the number of PHI entries we have.
5919             e = PN->getNumIncomingValues();
5920             BB = NewBB;
5921             i = PN->getBasicBlockIndex(BB);
5922 
5923             needUpdateFixups = true;
5924           }
5925         }
5926       }
5927 
5928       std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
5929         Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr)));
5930       if (!Pair.second)
5931         PN->setIncomingValue(i, Pair.first->second);
5932       else {
5933         Value *FullV =
5934             Expand(LU, LF, F, BB->getTerminator()->getIterator(), DeadInsts);
5935 
5936         // If this is reuse-by-noop-cast, insert the noop cast.
5937         Type *OpTy = LF.OperandValToReplace->getType();
5938         if (FullV->getType() != OpTy)
5939           FullV = CastInst::Create(
5940               CastInst::getCastOpcode(FullV, false, OpTy, false), FullV,
5941               LF.OperandValToReplace->getType(), "tmp",
5942               BB->getTerminator()->getIterator());
5943 
5944         // If the incoming block for this value is not in the loop, it means the
5945         // current PHI is not in a loop exit, so we must create a LCSSA PHI for
5946         // the inserted value.
5947         if (auto *I = dyn_cast<Instruction>(FullV))
5948           if (L->contains(I) && !L->contains(BB))
5949             InsertedNonLCSSAInsts.push_back(I);
5950 
5951         PN->setIncomingValue(i, FullV);
5952         Pair.first->second = FullV;
5953       }
5954 
5955       // If LSR splits critical edge and phi node has other pending
5956       // fixup operands, we need to update those pending fixups. Otherwise
5957       // formulae will not be implemented completely and some instructions
5958       // will not be eliminated.
5959       if (needUpdateFixups) {
5960         for (LSRUse &LU : Uses)
5961           for (LSRFixup &Fixup : LU.Fixups)
5962             // If fixup is supposed to rewrite some operand in the phi
5963             // that was just updated, it may be already moved to
5964             // another phi node. Such fixup requires update.
5965             if (Fixup.UserInst == PN) {
5966               // Check if the operand we try to replace still exists in the
5967               // original phi.
5968               bool foundInOriginalPHI = false;
5969               for (const auto &val : PN->incoming_values())
5970                 if (val == Fixup.OperandValToReplace) {
5971                   foundInOriginalPHI = true;
5972                   break;
5973                 }
5974 
5975               // If fixup operand found in original PHI - nothing to do.
5976               if (foundInOriginalPHI)
5977                 continue;
5978 
5979               // Otherwise it might be moved to another PHI and requires update.
5980               // If fixup operand not found in any of the incoming blocks that
5981               // means we have already rewritten it - nothing to do.
5982               for (const auto &Block : PN->blocks())
5983                 for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(I);
5984                      ++I) {
5985                   PHINode *NewPN = cast<PHINode>(I);
5986                   for (const auto &val : NewPN->incoming_values())
5987                     if (val == Fixup.OperandValToReplace)
5988                       Fixup.UserInst = NewPN;
5989                 }
5990             }
5991       }
5992     }
5993 
5994   formLCSSAForInstructions(InsertedNonLCSSAInsts, DT, LI, &SE);
5995 }
5996 
5997 /// Emit instructions for the leading candidate expression for this LSRUse (this
5998 /// is called "expanding"), and update the UserInst to reference the newly
5999 /// expanded value.
Rewrite(const LSRUse & LU,const LSRFixup & LF,const Formula & F,SmallVectorImpl<WeakTrackingVH> & DeadInsts) const6000 void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF,
6001                           const Formula &F,
6002                           SmallVectorImpl<WeakTrackingVH> &DeadInsts) const {
6003   // First, find an insertion point that dominates UserInst. For PHI nodes,
6004   // find the nearest block which dominates all the relevant uses.
6005   if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
6006     RewriteForPHI(PN, LU, LF, F, DeadInsts);
6007   } else {
6008     Value *FullV = Expand(LU, LF, F, LF.UserInst->getIterator(), DeadInsts);
6009 
6010     // If this is reuse-by-noop-cast, insert the noop cast.
6011     Type *OpTy = LF.OperandValToReplace->getType();
6012     if (FullV->getType() != OpTy) {
6013       Instruction *Cast =
6014           CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
6015                            FullV, OpTy, "tmp", LF.UserInst->getIterator());
6016       FullV = Cast;
6017     }
6018 
6019     // Update the user. ICmpZero is handled specially here (for now) because
6020     // Expand may have updated one of the operands of the icmp already, and
6021     // its new value may happen to be equal to LF.OperandValToReplace, in
6022     // which case doing replaceUsesOfWith leads to replacing both operands
6023     // with the same value. TODO: Reorganize this.
6024     if (LU.Kind == LSRUse::ICmpZero)
6025       LF.UserInst->setOperand(0, FullV);
6026     else
6027       LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
6028   }
6029 
6030   if (auto *OperandIsInstr = dyn_cast<Instruction>(LF.OperandValToReplace))
6031     DeadInsts.emplace_back(OperandIsInstr);
6032 }
6033 
6034 // Trying to hoist the IVInc to loop header if all IVInc users are in
6035 // the loop header. It will help backend to generate post index load/store
6036 // when the latch block is different from loop header block.
canHoistIVInc(const TargetTransformInfo & TTI,const LSRFixup & Fixup,const LSRUse & LU,Instruction * IVIncInsertPos,Loop * L)6037 static bool canHoistIVInc(const TargetTransformInfo &TTI, const LSRFixup &Fixup,
6038                           const LSRUse &LU, Instruction *IVIncInsertPos,
6039                           Loop *L) {
6040   if (LU.Kind != LSRUse::Address)
6041     return false;
6042 
6043   // For now this code do the conservative optimization, only work for
6044   // the header block. Later we can hoist the IVInc to the block post
6045   // dominate all users.
6046   BasicBlock *LHeader = L->getHeader();
6047   if (IVIncInsertPos->getParent() == LHeader)
6048     return false;
6049 
6050   if (!Fixup.OperandValToReplace ||
6051       any_of(Fixup.OperandValToReplace->users(), [&LHeader](User *U) {
6052         Instruction *UI = cast<Instruction>(U);
6053         return UI->getParent() != LHeader;
6054       }))
6055     return false;
6056 
6057   Instruction *I = Fixup.UserInst;
6058   Type *Ty = I->getType();
6059   return Ty->isIntegerTy() &&
6060          ((isa<LoadInst>(I) && TTI.isIndexedLoadLegal(TTI.MIM_PostInc, Ty)) ||
6061           (isa<StoreInst>(I) && TTI.isIndexedStoreLegal(TTI.MIM_PostInc, Ty)));
6062 }
6063 
6064 /// Rewrite all the fixup locations with new values, following the chosen
6065 /// solution.
ImplementSolution(const SmallVectorImpl<const Formula * > & Solution)6066 void LSRInstance::ImplementSolution(
6067     const SmallVectorImpl<const Formula *> &Solution) {
6068   // Keep track of instructions we may have made dead, so that
6069   // we can remove them after we are done working.
6070   SmallVector<WeakTrackingVH, 16> DeadInsts;
6071 
6072   // Mark phi nodes that terminate chains so the expander tries to reuse them.
6073   for (const IVChain &Chain : IVChainVec) {
6074     if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst()))
6075       Rewriter.setChainedPhi(PN);
6076   }
6077 
6078   // Expand the new value definitions and update the users.
6079   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx)
6080     for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) {
6081       Instruction *InsertPos =
6082           canHoistIVInc(TTI, Fixup, Uses[LUIdx], IVIncInsertPos, L)
6083               ? L->getHeader()->getTerminator()
6084               : IVIncInsertPos;
6085       Rewriter.setIVIncInsertPos(L, InsertPos);
6086       Rewrite(Uses[LUIdx], Fixup, *Solution[LUIdx], DeadInsts);
6087       Changed = true;
6088     }
6089 
6090   for (const IVChain &Chain : IVChainVec) {
6091     GenerateIVChain(Chain, DeadInsts);
6092     Changed = true;
6093   }
6094 
6095   for (const WeakVH &IV : Rewriter.getInsertedIVs())
6096     if (IV && dyn_cast<Instruction>(&*IV)->getParent())
6097       ScalarEvolutionIVs.push_back(IV);
6098 
6099   // Clean up after ourselves. This must be done before deleting any
6100   // instructions.
6101   Rewriter.clear();
6102 
6103   Changed |= RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts,
6104                                                                   &TLI, MSSAU);
6105 
6106   // In our cost analysis above, we assume that each addrec consumes exactly
6107   // one register, and arrange to have increments inserted just before the
6108   // latch to maximimize the chance this is true.  However, if we reused
6109   // existing IVs, we now need to move the increments to match our
6110   // expectations.  Otherwise, our cost modeling results in us having a
6111   // chosen a non-optimal result for the actual schedule.  (And yes, this
6112   // scheduling decision does impact later codegen.)
6113   for (PHINode &PN : L->getHeader()->phis()) {
6114     BinaryOperator *BO = nullptr;
6115     Value *Start = nullptr, *Step = nullptr;
6116     if (!matchSimpleRecurrence(&PN, BO, Start, Step))
6117       continue;
6118 
6119     switch (BO->getOpcode()) {
6120     case Instruction::Sub:
6121       if (BO->getOperand(0) != &PN)
6122         // sub is non-commutative - match handling elsewhere in LSR
6123         continue;
6124       break;
6125     case Instruction::Add:
6126       break;
6127     default:
6128       continue;
6129     };
6130 
6131     if (!isa<Constant>(Step))
6132       // If not a constant step, might increase register pressure
6133       // (We assume constants have been canonicalized to RHS)
6134       continue;
6135 
6136     if (BO->getParent() == IVIncInsertPos->getParent())
6137       // Only bother moving across blocks.  Isel can handle block local case.
6138       continue;
6139 
6140     // Can we legally schedule inc at the desired point?
6141     if (!llvm::all_of(BO->uses(),
6142                       [&](Use &U) {return DT.dominates(IVIncInsertPos, U);}))
6143       continue;
6144     BO->moveBefore(IVIncInsertPos);
6145     Changed = true;
6146   }
6147 
6148 
6149 }
6150 
LSRInstance(Loop * L,IVUsers & IU,ScalarEvolution & SE,DominatorTree & DT,LoopInfo & LI,const TargetTransformInfo & TTI,AssumptionCache & AC,TargetLibraryInfo & TLI,MemorySSAUpdater * MSSAU)6151 LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE,
6152                          DominatorTree &DT, LoopInfo &LI,
6153                          const TargetTransformInfo &TTI, AssumptionCache &AC,
6154                          TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU)
6155     : IU(IU), SE(SE), DT(DT), LI(LI), AC(AC), TLI(TLI), TTI(TTI), L(L),
6156       MSSAU(MSSAU), AMK(PreferredAddresingMode.getNumOccurrences() > 0
6157                             ? PreferredAddresingMode
6158                             : TTI.getPreferredAddressingMode(L, &SE)),
6159       Rewriter(SE, L->getHeader()->getDataLayout(), "lsr", false),
6160       BaselineCost(L, SE, TTI, AMK) {
6161   // If LoopSimplify form is not available, stay out of trouble.
6162   if (!L->isLoopSimplifyForm())
6163     return;
6164 
6165   // If there's no interesting work to be done, bail early.
6166   if (IU.empty()) return;
6167 
6168   // If there's too much analysis to be done, bail early. We won't be able to
6169   // model the problem anyway.
6170   unsigned NumUsers = 0;
6171   for (const IVStrideUse &U : IU) {
6172     if (++NumUsers > MaxIVUsers) {
6173       (void)U;
6174       LLVM_DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U
6175                         << "\n");
6176       return;
6177     }
6178     // Bail out if we have a PHI on an EHPad that gets a value from a
6179     // CatchSwitchInst.  Because the CatchSwitchInst cannot be split, there is
6180     // no good place to stick any instructions.
6181     if (auto *PN = dyn_cast<PHINode>(U.getUser())) {
6182        auto *FirstNonPHI = PN->getParent()->getFirstNonPHI();
6183        if (isa<FuncletPadInst>(FirstNonPHI) ||
6184            isa<CatchSwitchInst>(FirstNonPHI))
6185          for (BasicBlock *PredBB : PN->blocks())
6186            if (isa<CatchSwitchInst>(PredBB->getFirstNonPHI()))
6187              return;
6188     }
6189   }
6190 
6191   LLVM_DEBUG(dbgs() << "\nLSR on loop ";
6192              L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
6193              dbgs() << ":\n");
6194 
6195   // Configure SCEVExpander already now, so the correct mode is used for
6196   // isSafeToExpand() checks.
6197 #ifndef NDEBUG
6198   Rewriter.setDebugType(DEBUG_TYPE);
6199 #endif
6200   Rewriter.disableCanonicalMode();
6201   Rewriter.enableLSRMode();
6202 
6203   // First, perform some low-level loop optimizations.
6204   OptimizeShadowIV();
6205   OptimizeLoopTermCond();
6206 
6207   // If loop preparation eliminates all interesting IV users, bail.
6208   if (IU.empty()) return;
6209 
6210   // Skip nested loops until we can model them better with formulae.
6211   if (!L->isInnermost()) {
6212     LLVM_DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
6213     return;
6214   }
6215 
6216   // Start collecting data and preparing for the solver.
6217   // If number of registers is not the major cost, we cannot benefit from the
6218   // current profitable chain optimization which is based on number of
6219   // registers.
6220   // FIXME: add profitable chain optimization for other kinds major cost, for
6221   // example number of instructions.
6222   if (TTI.isNumRegsMajorCostOfLSR() || StressIVChain)
6223     CollectChains();
6224   CollectInterestingTypesAndFactors();
6225   CollectFixupsAndInitialFormulae();
6226   CollectLoopInvariantFixupsAndFormulae();
6227 
6228   if (Uses.empty())
6229     return;
6230 
6231   LLVM_DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
6232              print_uses(dbgs()));
6233   LLVM_DEBUG(dbgs() << "The baseline solution requires ";
6234              BaselineCost.print(dbgs()); dbgs() << "\n");
6235 
6236   // Now use the reuse data to generate a bunch of interesting ways
6237   // to formulate the values needed for the uses.
6238   GenerateAllReuseFormulae();
6239 
6240   FilterOutUndesirableDedicatedRegisters();
6241   NarrowSearchSpaceUsingHeuristics();
6242 
6243   SmallVector<const Formula *, 8> Solution;
6244   Solve(Solution);
6245 
6246   // Release memory that is no longer needed.
6247   Factors.clear();
6248   Types.clear();
6249   RegUses.clear();
6250 
6251   if (Solution.empty())
6252     return;
6253 
6254 #ifndef NDEBUG
6255   // Formulae should be legal.
6256   for (const LSRUse &LU : Uses) {
6257     for (const Formula &F : LU.Formulae)
6258       assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
6259                         F) && "Illegal formula generated!");
6260   };
6261 #endif
6262 
6263   // Now that we've decided what we want, make it so.
6264   ImplementSolution(Solution);
6265 }
6266 
6267 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
print_factors_and_types(raw_ostream & OS) const6268 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
6269   if (Factors.empty() && Types.empty()) return;
6270 
6271   OS << "LSR has identified the following interesting factors and types: ";
6272   bool First = true;
6273 
6274   for (int64_t Factor : Factors) {
6275     if (!First) OS << ", ";
6276     First = false;
6277     OS << '*' << Factor;
6278   }
6279 
6280   for (Type *Ty : Types) {
6281     if (!First) OS << ", ";
6282     First = false;
6283     OS << '(' << *Ty << ')';
6284   }
6285   OS << '\n';
6286 }
6287 
print_fixups(raw_ostream & OS) const6288 void LSRInstance::print_fixups(raw_ostream &OS) const {
6289   OS << "LSR is examining the following fixup sites:\n";
6290   for (const LSRUse &LU : Uses)
6291     for (const LSRFixup &LF : LU.Fixups) {
6292       dbgs() << "  ";
6293       LF.print(OS);
6294       OS << '\n';
6295     }
6296 }
6297 
print_uses(raw_ostream & OS) const6298 void LSRInstance::print_uses(raw_ostream &OS) const {
6299   OS << "LSR is examining the following uses:\n";
6300   for (const LSRUse &LU : Uses) {
6301     dbgs() << "  ";
6302     LU.print(OS);
6303     OS << '\n';
6304     for (const Formula &F : LU.Formulae) {
6305       OS << "    ";
6306       F.print(OS);
6307       OS << '\n';
6308     }
6309   }
6310 }
6311 
print(raw_ostream & OS) const6312 void LSRInstance::print(raw_ostream &OS) const {
6313   print_factors_and_types(OS);
6314   print_fixups(OS);
6315   print_uses(OS);
6316 }
6317 
dump() const6318 LLVM_DUMP_METHOD void LSRInstance::dump() const {
6319   print(errs()); errs() << '\n';
6320 }
6321 #endif
6322 
6323 namespace {
6324 
6325 class LoopStrengthReduce : public LoopPass {
6326 public:
6327   static char ID; // Pass ID, replacement for typeid
6328 
6329   LoopStrengthReduce();
6330 
6331 private:
6332   bool runOnLoop(Loop *L, LPPassManager &LPM) override;
6333   void getAnalysisUsage(AnalysisUsage &AU) const override;
6334 };
6335 
6336 } // end anonymous namespace
6337 
LoopStrengthReduce()6338 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
6339   initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
6340 }
6341 
getAnalysisUsage(AnalysisUsage & AU) const6342 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
6343   // We split critical edges, so we change the CFG.  However, we do update
6344   // many analyses if they are around.
6345   AU.addPreservedID(LoopSimplifyID);
6346 
6347   AU.addRequired<LoopInfoWrapperPass>();
6348   AU.addPreserved<LoopInfoWrapperPass>();
6349   AU.addRequiredID(LoopSimplifyID);
6350   AU.addRequired<DominatorTreeWrapperPass>();
6351   AU.addPreserved<DominatorTreeWrapperPass>();
6352   AU.addRequired<ScalarEvolutionWrapperPass>();
6353   AU.addPreserved<ScalarEvolutionWrapperPass>();
6354   AU.addRequired<AssumptionCacheTracker>();
6355   AU.addRequired<TargetLibraryInfoWrapperPass>();
6356   // Requiring LoopSimplify a second time here prevents IVUsers from running
6357   // twice, since LoopSimplify was invalidated by running ScalarEvolution.
6358   AU.addRequiredID(LoopSimplifyID);
6359   AU.addRequired<IVUsersWrapperPass>();
6360   AU.addPreserved<IVUsersWrapperPass>();
6361   AU.addRequired<TargetTransformInfoWrapperPass>();
6362   AU.addPreserved<MemorySSAWrapperPass>();
6363 }
6364 
6365 namespace {
6366 
6367 /// Enables more convenient iteration over a DWARF expression vector.
6368 static iterator_range<llvm::DIExpression::expr_op_iterator>
ToDwarfOpIter(SmallVectorImpl<uint64_t> & Expr)6369 ToDwarfOpIter(SmallVectorImpl<uint64_t> &Expr) {
6370   llvm::DIExpression::expr_op_iterator Begin =
6371       llvm::DIExpression::expr_op_iterator(Expr.begin());
6372   llvm::DIExpression::expr_op_iterator End =
6373       llvm::DIExpression::expr_op_iterator(Expr.end());
6374   return {Begin, End};
6375 }
6376 
6377 struct SCEVDbgValueBuilder {
6378   SCEVDbgValueBuilder() = default;
SCEVDbgValueBuilder__anonc21373341411::SCEVDbgValueBuilder6379   SCEVDbgValueBuilder(const SCEVDbgValueBuilder &Base) { clone(Base); }
6380 
clone__anonc21373341411::SCEVDbgValueBuilder6381   void clone(const SCEVDbgValueBuilder &Base) {
6382     LocationOps = Base.LocationOps;
6383     Expr = Base.Expr;
6384   }
6385 
clear__anonc21373341411::SCEVDbgValueBuilder6386   void clear() {
6387     LocationOps.clear();
6388     Expr.clear();
6389   }
6390 
6391   /// The DIExpression as we translate the SCEV.
6392   SmallVector<uint64_t, 6> Expr;
6393   /// The location ops of the DIExpression.
6394   SmallVector<Value *, 2> LocationOps;
6395 
pushOperator__anonc21373341411::SCEVDbgValueBuilder6396   void pushOperator(uint64_t Op) { Expr.push_back(Op); }
pushUInt__anonc21373341411::SCEVDbgValueBuilder6397   void pushUInt(uint64_t Operand) { Expr.push_back(Operand); }
6398 
6399   /// Add a DW_OP_LLVM_arg to the expression, followed by the index of the value
6400   /// in the set of values referenced by the expression.
pushLocation__anonc21373341411::SCEVDbgValueBuilder6401   void pushLocation(llvm::Value *V) {
6402     Expr.push_back(llvm::dwarf::DW_OP_LLVM_arg);
6403     auto *It = llvm::find(LocationOps, V);
6404     unsigned ArgIndex = 0;
6405     if (It != LocationOps.end()) {
6406       ArgIndex = std::distance(LocationOps.begin(), It);
6407     } else {
6408       ArgIndex = LocationOps.size();
6409       LocationOps.push_back(V);
6410     }
6411     Expr.push_back(ArgIndex);
6412   }
6413 
pushValue__anonc21373341411::SCEVDbgValueBuilder6414   void pushValue(const SCEVUnknown *U) {
6415     llvm::Value *V = cast<SCEVUnknown>(U)->getValue();
6416     pushLocation(V);
6417   }
6418 
pushConst__anonc21373341411::SCEVDbgValueBuilder6419   bool pushConst(const SCEVConstant *C) {
6420     if (C->getAPInt().getSignificantBits() > 64)
6421       return false;
6422     Expr.push_back(llvm::dwarf::DW_OP_consts);
6423     Expr.push_back(C->getAPInt().getSExtValue());
6424     return true;
6425   }
6426 
6427   // Iterating the expression as DWARF ops is convenient when updating
6428   // DWARF_OP_LLVM_args.
expr_ops__anonc21373341411::SCEVDbgValueBuilder6429   iterator_range<llvm::DIExpression::expr_op_iterator> expr_ops() {
6430     return ToDwarfOpIter(Expr);
6431   }
6432 
6433   /// Several SCEV types are sequences of the same arithmetic operator applied
6434   /// to constants and values that may be extended or truncated.
pushArithmeticExpr__anonc21373341411::SCEVDbgValueBuilder6435   bool pushArithmeticExpr(const llvm::SCEVCommutativeExpr *CommExpr,
6436                           uint64_t DwarfOp) {
6437     assert((isa<llvm::SCEVAddExpr>(CommExpr) || isa<SCEVMulExpr>(CommExpr)) &&
6438            "Expected arithmetic SCEV type");
6439     bool Success = true;
6440     unsigned EmitOperator = 0;
6441     for (const auto &Op : CommExpr->operands()) {
6442       Success &= pushSCEV(Op);
6443 
6444       if (EmitOperator >= 1)
6445         pushOperator(DwarfOp);
6446       ++EmitOperator;
6447     }
6448     return Success;
6449   }
6450 
6451   // TODO: Identify and omit noop casts.
pushCast__anonc21373341411::SCEVDbgValueBuilder6452   bool pushCast(const llvm::SCEVCastExpr *C, bool IsSigned) {
6453     const llvm::SCEV *Inner = C->getOperand(0);
6454     const llvm::Type *Type = C->getType();
6455     uint64_t ToWidth = Type->getIntegerBitWidth();
6456     bool Success = pushSCEV(Inner);
6457     uint64_t CastOps[] = {dwarf::DW_OP_LLVM_convert, ToWidth,
6458                           IsSigned ? llvm::dwarf::DW_ATE_signed
6459                                    : llvm::dwarf::DW_ATE_unsigned};
6460     for (const auto &Op : CastOps)
6461       pushOperator(Op);
6462     return Success;
6463   }
6464 
6465   // TODO: MinMax - although these haven't been encountered in the test suite.
pushSCEV__anonc21373341411::SCEVDbgValueBuilder6466   bool pushSCEV(const llvm::SCEV *S) {
6467     bool Success = true;
6468     if (const SCEVConstant *StartInt = dyn_cast<SCEVConstant>(S)) {
6469       Success &= pushConst(StartInt);
6470 
6471     } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
6472       if (!U->getValue())
6473         return false;
6474       pushLocation(U->getValue());
6475 
6476     } else if (const SCEVMulExpr *MulRec = dyn_cast<SCEVMulExpr>(S)) {
6477       Success &= pushArithmeticExpr(MulRec, llvm::dwarf::DW_OP_mul);
6478 
6479     } else if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
6480       Success &= pushSCEV(UDiv->getLHS());
6481       Success &= pushSCEV(UDiv->getRHS());
6482       pushOperator(llvm::dwarf::DW_OP_div);
6483 
6484     } else if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(S)) {
6485       // Assert if a new and unknown SCEVCastEXpr type is encountered.
6486       assert((isa<SCEVZeroExtendExpr>(Cast) || isa<SCEVTruncateExpr>(Cast) ||
6487               isa<SCEVPtrToIntExpr>(Cast) || isa<SCEVSignExtendExpr>(Cast)) &&
6488              "Unexpected cast type in SCEV.");
6489       Success &= pushCast(Cast, (isa<SCEVSignExtendExpr>(Cast)));
6490 
6491     } else if (const SCEVAddExpr *AddExpr = dyn_cast<SCEVAddExpr>(S)) {
6492       Success &= pushArithmeticExpr(AddExpr, llvm::dwarf::DW_OP_plus);
6493 
6494     } else if (isa<SCEVAddRecExpr>(S)) {
6495       // Nested SCEVAddRecExpr are generated by nested loops and are currently
6496       // unsupported.
6497       return false;
6498 
6499     } else {
6500       return false;
6501     }
6502     return Success;
6503   }
6504 
6505   /// Return true if the combination of arithmetic operator and underlying
6506   /// SCEV constant value is an identity function.
isIdentityFunction__anonc21373341411::SCEVDbgValueBuilder6507   bool isIdentityFunction(uint64_t Op, const SCEV *S) {
6508     if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
6509       if (C->getAPInt().getSignificantBits() > 64)
6510         return false;
6511       int64_t I = C->getAPInt().getSExtValue();
6512       switch (Op) {
6513       case llvm::dwarf::DW_OP_plus:
6514       case llvm::dwarf::DW_OP_minus:
6515         return I == 0;
6516       case llvm::dwarf::DW_OP_mul:
6517       case llvm::dwarf::DW_OP_div:
6518         return I == 1;
6519       }
6520     }
6521     return false;
6522   }
6523 
6524   /// Convert a SCEV of a value to a DIExpression that is pushed onto the
6525   /// builder's expression stack. The stack should already contain an
6526   /// expression for the iteration count, so that it can be multiplied by
6527   /// the stride and added to the start.
6528   /// Components of the expression are omitted if they are an identity function.
6529   /// Chain (non-affine) SCEVs are not supported.
SCEVToValueExpr__anonc21373341411::SCEVDbgValueBuilder6530   bool SCEVToValueExpr(const llvm::SCEVAddRecExpr &SAR, ScalarEvolution &SE) {
6531     assert(SAR.isAffine() && "Expected affine SCEV");
6532     // TODO: Is this check needed?
6533     if (isa<SCEVAddRecExpr>(SAR.getStart()))
6534       return false;
6535 
6536     const SCEV *Start = SAR.getStart();
6537     const SCEV *Stride = SAR.getStepRecurrence(SE);
6538 
6539     // Skip pushing arithmetic noops.
6540     if (!isIdentityFunction(llvm::dwarf::DW_OP_mul, Stride)) {
6541       if (!pushSCEV(Stride))
6542         return false;
6543       pushOperator(llvm::dwarf::DW_OP_mul);
6544     }
6545     if (!isIdentityFunction(llvm::dwarf::DW_OP_plus, Start)) {
6546       if (!pushSCEV(Start))
6547         return false;
6548       pushOperator(llvm::dwarf::DW_OP_plus);
6549     }
6550     return true;
6551   }
6552 
6553   /// Create an expression that is an offset from a value (usually the IV).
createOffsetExpr__anonc21373341411::SCEVDbgValueBuilder6554   void createOffsetExpr(int64_t Offset, Value *OffsetValue) {
6555     pushLocation(OffsetValue);
6556     DIExpression::appendOffset(Expr, Offset);
6557     LLVM_DEBUG(
6558         dbgs() << "scev-salvage: Generated IV offset expression. Offset: "
6559                << std::to_string(Offset) << "\n");
6560   }
6561 
6562   /// Combine a translation of the SCEV and the IV to create an expression that
6563   /// recovers a location's value.
6564   /// returns true if an expression was created.
createIterCountExpr__anonc21373341411::SCEVDbgValueBuilder6565   bool createIterCountExpr(const SCEV *S,
6566                            const SCEVDbgValueBuilder &IterationCount,
6567                            ScalarEvolution &SE) {
6568     // SCEVs for SSA values are most frquently of the form
6569     // {start,+,stride}, but sometimes they are ({start,+,stride} + %a + ..).
6570     // This is because %a is a PHI node that is not the IV. However, these
6571     // SCEVs have not been observed to result in debuginfo-lossy optimisations,
6572     // so its not expected this point will be reached.
6573     if (!isa<SCEVAddRecExpr>(S))
6574       return false;
6575 
6576     LLVM_DEBUG(dbgs() << "scev-salvage: Location to salvage SCEV: " << *S
6577                       << '\n');
6578 
6579     const auto *Rec = cast<SCEVAddRecExpr>(S);
6580     if (!Rec->isAffine())
6581       return false;
6582 
6583     if (S->getExpressionSize() > MaxSCEVSalvageExpressionSize)
6584       return false;
6585 
6586     // Initialise a new builder with the iteration count expression. In
6587     // combination with the value's SCEV this enables recovery.
6588     clone(IterationCount);
6589     if (!SCEVToValueExpr(*Rec, SE))
6590       return false;
6591 
6592     return true;
6593   }
6594 
6595   /// Convert a SCEV of a value to a DIExpression that is pushed onto the
6596   /// builder's expression stack. The stack should already contain an
6597   /// expression for the iteration count, so that it can be multiplied by
6598   /// the stride and added to the start.
6599   /// Components of the expression are omitted if they are an identity function.
SCEVToIterCountExpr__anonc21373341411::SCEVDbgValueBuilder6600   bool SCEVToIterCountExpr(const llvm::SCEVAddRecExpr &SAR,
6601                            ScalarEvolution &SE) {
6602     assert(SAR.isAffine() && "Expected affine SCEV");
6603     if (isa<SCEVAddRecExpr>(SAR.getStart())) {
6604       LLVM_DEBUG(dbgs() << "scev-salvage: IV SCEV. Unsupported nested AddRec: "
6605                         << SAR << '\n');
6606       return false;
6607     }
6608     const SCEV *Start = SAR.getStart();
6609     const SCEV *Stride = SAR.getStepRecurrence(SE);
6610 
6611     // Skip pushing arithmetic noops.
6612     if (!isIdentityFunction(llvm::dwarf::DW_OP_minus, Start)) {
6613       if (!pushSCEV(Start))
6614         return false;
6615       pushOperator(llvm::dwarf::DW_OP_minus);
6616     }
6617     if (!isIdentityFunction(llvm::dwarf::DW_OP_div, Stride)) {
6618       if (!pushSCEV(Stride))
6619         return false;
6620       pushOperator(llvm::dwarf::DW_OP_div);
6621     }
6622     return true;
6623   }
6624 
6625   // Append the current expression and locations to a location list and an
6626   // expression list. Modify the DW_OP_LLVM_arg indexes to account for
6627   // the locations already present in the destination list.
appendToVectors__anonc21373341411::SCEVDbgValueBuilder6628   void appendToVectors(SmallVectorImpl<uint64_t> &DestExpr,
6629                        SmallVectorImpl<Value *> &DestLocations) {
6630     assert(!DestLocations.empty() &&
6631            "Expected the locations vector to contain the IV");
6632     // The DWARF_OP_LLVM_arg arguments of the expression being appended must be
6633     // modified to account for the locations already in the destination vector.
6634     // All builders contain the IV as the first location op.
6635     assert(!LocationOps.empty() &&
6636            "Expected the location ops to contain the IV.");
6637     // DestIndexMap[n] contains the index in DestLocations for the nth
6638     // location in this SCEVDbgValueBuilder.
6639     SmallVector<uint64_t, 2> DestIndexMap;
6640     for (const auto &Op : LocationOps) {
6641       auto It = find(DestLocations, Op);
6642       if (It != DestLocations.end()) {
6643         // Location already exists in DestLocations, reuse existing ArgIndex.
6644         DestIndexMap.push_back(std::distance(DestLocations.begin(), It));
6645         continue;
6646       }
6647       // Location is not in DestLocations, add it.
6648       DestIndexMap.push_back(DestLocations.size());
6649       DestLocations.push_back(Op);
6650     }
6651 
6652     for (const auto &Op : expr_ops()) {
6653       if (Op.getOp() != dwarf::DW_OP_LLVM_arg) {
6654         Op.appendToVector(DestExpr);
6655         continue;
6656       }
6657 
6658       DestExpr.push_back(dwarf::DW_OP_LLVM_arg);
6659       // `DW_OP_LLVM_arg n` represents the nth LocationOp in this SCEV,
6660       // DestIndexMap[n] contains its new index in DestLocations.
6661       uint64_t NewIndex = DestIndexMap[Op.getArg(0)];
6662       DestExpr.push_back(NewIndex);
6663     }
6664   }
6665 };
6666 
6667 /// Holds all the required data to salvage a dbg.value using the pre-LSR SCEVs
6668 /// and DIExpression.
6669 struct DVIRecoveryRec {
DVIRecoveryRec__anonc21373341411::DVIRecoveryRec6670   DVIRecoveryRec(DbgValueInst *DbgValue)
6671       : DbgRef(DbgValue), Expr(DbgValue->getExpression()),
6672         HadLocationArgList(false) {}
DVIRecoveryRec__anonc21373341411::DVIRecoveryRec6673   DVIRecoveryRec(DbgVariableRecord *DVR)
6674       : DbgRef(DVR), Expr(DVR->getExpression()), HadLocationArgList(false) {}
6675 
6676   PointerUnion<DbgValueInst *, DbgVariableRecord *> DbgRef;
6677   DIExpression *Expr;
6678   bool HadLocationArgList;
6679   SmallVector<WeakVH, 2> LocationOps;
6680   SmallVector<const llvm::SCEV *, 2> SCEVs;
6681   SmallVector<std::unique_ptr<SCEVDbgValueBuilder>, 2> RecoveryExprs;
6682 
clear__anonc21373341411::DVIRecoveryRec6683   void clear() {
6684     for (auto &RE : RecoveryExprs)
6685       RE.reset();
6686     RecoveryExprs.clear();
6687   }
6688 
~DVIRecoveryRec__anonc21373341411::DVIRecoveryRec6689   ~DVIRecoveryRec() { clear(); }
6690 };
6691 } // namespace
6692 
6693 /// Returns the total number of DW_OP_llvm_arg operands in the expression.
6694 /// This helps in determining if a DIArglist is necessary or can be omitted from
6695 /// the dbg.value.
numLLVMArgOps(SmallVectorImpl<uint64_t> & Expr)6696 static unsigned numLLVMArgOps(SmallVectorImpl<uint64_t> &Expr) {
6697   auto expr_ops = ToDwarfOpIter(Expr);
6698   unsigned Count = 0;
6699   for (auto Op : expr_ops)
6700     if (Op.getOp() == dwarf::DW_OP_LLVM_arg)
6701       Count++;
6702   return Count;
6703 }
6704 
6705 /// Overwrites DVI with the location and Ops as the DIExpression. This will
6706 /// create an invalid expression if Ops has any dwarf::DW_OP_llvm_arg operands,
6707 /// because a DIArglist is not created for the first argument of the dbg.value.
6708 template <typename T>
updateDVIWithLocation(T & DbgVal,Value * Location,SmallVectorImpl<uint64_t> & Ops)6709 static void updateDVIWithLocation(T &DbgVal, Value *Location,
6710                                   SmallVectorImpl<uint64_t> &Ops) {
6711   assert(numLLVMArgOps(Ops) == 0 && "Expected expression that does not "
6712                                     "contain any DW_OP_llvm_arg operands.");
6713   DbgVal.setRawLocation(ValueAsMetadata::get(Location));
6714   DbgVal.setExpression(DIExpression::get(DbgVal.getContext(), Ops));
6715   DbgVal.setExpression(DIExpression::get(DbgVal.getContext(), Ops));
6716 }
6717 
6718 /// Overwrite DVI with locations placed into a DIArglist.
6719 template <typename T>
updateDVIWithLocations(T & DbgVal,SmallVectorImpl<Value * > & Locations,SmallVectorImpl<uint64_t> & Ops)6720 static void updateDVIWithLocations(T &DbgVal,
6721                                    SmallVectorImpl<Value *> &Locations,
6722                                    SmallVectorImpl<uint64_t> &Ops) {
6723   assert(numLLVMArgOps(Ops) != 0 &&
6724          "Expected expression that references DIArglist locations using "
6725          "DW_OP_llvm_arg operands.");
6726   SmallVector<ValueAsMetadata *, 3> MetadataLocs;
6727   for (Value *V : Locations)
6728     MetadataLocs.push_back(ValueAsMetadata::get(V));
6729   auto ValArrayRef = llvm::ArrayRef<llvm::ValueAsMetadata *>(MetadataLocs);
6730   DbgVal.setRawLocation(llvm::DIArgList::get(DbgVal.getContext(), ValArrayRef));
6731   DbgVal.setExpression(DIExpression::get(DbgVal.getContext(), Ops));
6732 }
6733 
6734 /// Write the new expression and new location ops for the dbg.value. If possible
6735 /// reduce the szie of the dbg.value intrinsic by omitting DIArglist. This
6736 /// can be omitted if:
6737 /// 1. There is only a single location, refenced by a single DW_OP_llvm_arg.
6738 /// 2. The DW_OP_LLVM_arg is the first operand in the expression.
UpdateDbgValueInst(DVIRecoveryRec & DVIRec,SmallVectorImpl<Value * > & NewLocationOps,SmallVectorImpl<uint64_t> & NewExpr)6739 static void UpdateDbgValueInst(DVIRecoveryRec &DVIRec,
6740                                SmallVectorImpl<Value *> &NewLocationOps,
6741                                SmallVectorImpl<uint64_t> &NewExpr) {
6742   auto UpdateDbgValueInstImpl = [&](auto *DbgVal) {
6743     unsigned NumLLVMArgs = numLLVMArgOps(NewExpr);
6744     if (NumLLVMArgs == 0) {
6745       // Location assumed to be on the stack.
6746       updateDVIWithLocation(*DbgVal, NewLocationOps[0], NewExpr);
6747     } else if (NumLLVMArgs == 1 && NewExpr[0] == dwarf::DW_OP_LLVM_arg) {
6748       // There is only a single DW_OP_llvm_arg at the start of the expression,
6749       // so it can be omitted along with DIArglist.
6750       assert(NewExpr[1] == 0 &&
6751              "Lone LLVM_arg in a DIExpression should refer to location-op 0.");
6752       llvm::SmallVector<uint64_t, 6> ShortenedOps(llvm::drop_begin(NewExpr, 2));
6753       updateDVIWithLocation(*DbgVal, NewLocationOps[0], ShortenedOps);
6754     } else {
6755       // Multiple DW_OP_llvm_arg, so DIArgList is strictly necessary.
6756       updateDVIWithLocations(*DbgVal, NewLocationOps, NewExpr);
6757     }
6758 
6759     // If the DIExpression was previously empty then add the stack terminator.
6760     // Non-empty expressions have only had elements inserted into them and so
6761     // the terminator should already be present e.g. stack_value or fragment.
6762     DIExpression *SalvageExpr = DbgVal->getExpression();
6763     if (!DVIRec.Expr->isComplex() && SalvageExpr->isComplex()) {
6764       SalvageExpr =
6765           DIExpression::append(SalvageExpr, {dwarf::DW_OP_stack_value});
6766       DbgVal->setExpression(SalvageExpr);
6767     }
6768   };
6769   if (isa<DbgValueInst *>(DVIRec.DbgRef))
6770     UpdateDbgValueInstImpl(cast<DbgValueInst *>(DVIRec.DbgRef));
6771   else
6772     UpdateDbgValueInstImpl(cast<DbgVariableRecord *>(DVIRec.DbgRef));
6773 }
6774 
6775 /// Cached location ops may be erased during LSR, in which case a poison is
6776 /// required when restoring from the cache. The type of that location is no
6777 /// longer available, so just use int8. The poison will be replaced by one or
6778 /// more locations later when a SCEVDbgValueBuilder selects alternative
6779 /// locations to use for the salvage.
getValueOrPoison(WeakVH & VH,LLVMContext & C)6780 static Value *getValueOrPoison(WeakVH &VH, LLVMContext &C) {
6781   return (VH) ? VH : PoisonValue::get(llvm::Type::getInt8Ty(C));
6782 }
6783 
6784 /// Restore the DVI's pre-LSR arguments. Substitute undef for any erased values.
restorePreTransformState(DVIRecoveryRec & DVIRec)6785 static void restorePreTransformState(DVIRecoveryRec &DVIRec) {
6786   auto RestorePreTransformStateImpl = [&](auto *DbgVal) {
6787     LLVM_DEBUG(dbgs() << "scev-salvage: restore dbg.value to pre-LSR state\n"
6788                       << "scev-salvage: post-LSR: " << *DbgVal << '\n');
6789     assert(DVIRec.Expr && "Expected an expression");
6790     DbgVal->setExpression(DVIRec.Expr);
6791 
6792     // Even a single location-op may be inside a DIArgList and referenced with
6793     // DW_OP_LLVM_arg, which is valid only with a DIArgList.
6794     if (!DVIRec.HadLocationArgList) {
6795       assert(DVIRec.LocationOps.size() == 1 &&
6796              "Unexpected number of location ops.");
6797       // LSR's unsuccessful salvage attempt may have added DIArgList, which in
6798       // this case was not present before, so force the location back to a
6799       // single uncontained Value.
6800       Value *CachedValue =
6801           getValueOrPoison(DVIRec.LocationOps[0], DbgVal->getContext());
6802       DbgVal->setRawLocation(ValueAsMetadata::get(CachedValue));
6803     } else {
6804       SmallVector<ValueAsMetadata *, 3> MetadataLocs;
6805       for (WeakVH VH : DVIRec.LocationOps) {
6806         Value *CachedValue = getValueOrPoison(VH, DbgVal->getContext());
6807         MetadataLocs.push_back(ValueAsMetadata::get(CachedValue));
6808       }
6809       auto ValArrayRef = llvm::ArrayRef<llvm::ValueAsMetadata *>(MetadataLocs);
6810       DbgVal->setRawLocation(
6811           llvm::DIArgList::get(DbgVal->getContext(), ValArrayRef));
6812     }
6813     LLVM_DEBUG(dbgs() << "scev-salvage: pre-LSR: " << *DbgVal << '\n');
6814   };
6815   if (isa<DbgValueInst *>(DVIRec.DbgRef))
6816     RestorePreTransformStateImpl(cast<DbgValueInst *>(DVIRec.DbgRef));
6817   else
6818     RestorePreTransformStateImpl(cast<DbgVariableRecord *>(DVIRec.DbgRef));
6819 }
6820 
SalvageDVI(llvm::Loop * L,ScalarEvolution & SE,llvm::PHINode * LSRInductionVar,DVIRecoveryRec & DVIRec,const SCEV * SCEVInductionVar,SCEVDbgValueBuilder IterCountExpr)6821 static bool SalvageDVI(llvm::Loop *L, ScalarEvolution &SE,
6822                        llvm::PHINode *LSRInductionVar, DVIRecoveryRec &DVIRec,
6823                        const SCEV *SCEVInductionVar,
6824                        SCEVDbgValueBuilder IterCountExpr) {
6825 
6826   if (isa<DbgValueInst *>(DVIRec.DbgRef)
6827           ? !cast<DbgValueInst *>(DVIRec.DbgRef)->isKillLocation()
6828           : !cast<DbgVariableRecord *>(DVIRec.DbgRef)->isKillLocation())
6829     return false;
6830 
6831   // LSR may have caused several changes to the dbg.value in the failed salvage
6832   // attempt. So restore the DIExpression, the location ops and also the
6833   // location ops format, which is always DIArglist for multiple ops, but only
6834   // sometimes for a single op.
6835   restorePreTransformState(DVIRec);
6836 
6837   // LocationOpIndexMap[i] will store the post-LSR location index of
6838   // the non-optimised out location at pre-LSR index i.
6839   SmallVector<int64_t, 2> LocationOpIndexMap;
6840   LocationOpIndexMap.assign(DVIRec.LocationOps.size(), -1);
6841   SmallVector<Value *, 2> NewLocationOps;
6842   NewLocationOps.push_back(LSRInductionVar);
6843 
6844   for (unsigned i = 0; i < DVIRec.LocationOps.size(); i++) {
6845     WeakVH VH = DVIRec.LocationOps[i];
6846     // Place the locations not optimised out in the list first, avoiding
6847     // inserts later. The map is used to update the DIExpression's
6848     // DW_OP_LLVM_arg arguments as the expression is updated.
6849     if (VH && !isa<UndefValue>(VH)) {
6850       NewLocationOps.push_back(VH);
6851       LocationOpIndexMap[i] = NewLocationOps.size() - 1;
6852       LLVM_DEBUG(dbgs() << "scev-salvage: Location index " << i
6853                         << " now at index " << LocationOpIndexMap[i] << "\n");
6854       continue;
6855     }
6856 
6857     // It's possible that a value referred to in the SCEV may have been
6858     // optimised out by LSR.
6859     if (SE.containsErasedValue(DVIRec.SCEVs[i]) ||
6860         SE.containsUndefs(DVIRec.SCEVs[i])) {
6861       LLVM_DEBUG(dbgs() << "scev-salvage: SCEV for location at index: " << i
6862                         << " refers to a location that is now undef or erased. "
6863                            "Salvage abandoned.\n");
6864       return false;
6865     }
6866 
6867     LLVM_DEBUG(dbgs() << "scev-salvage: salvaging location at index " << i
6868                       << " with SCEV: " << *DVIRec.SCEVs[i] << "\n");
6869 
6870     DVIRec.RecoveryExprs[i] = std::make_unique<SCEVDbgValueBuilder>();
6871     SCEVDbgValueBuilder *SalvageExpr = DVIRec.RecoveryExprs[i].get();
6872 
6873     // Create an offset-based salvage expression if possible, as it requires
6874     // less DWARF ops than an iteration count-based expression.
6875     if (std::optional<APInt> Offset =
6876             SE.computeConstantDifference(DVIRec.SCEVs[i], SCEVInductionVar)) {
6877       if (Offset->getSignificantBits() <= 64)
6878         SalvageExpr->createOffsetExpr(Offset->getSExtValue(), LSRInductionVar);
6879     } else if (!SalvageExpr->createIterCountExpr(DVIRec.SCEVs[i], IterCountExpr,
6880                                                  SE))
6881       return false;
6882   }
6883 
6884   // Merge the DbgValueBuilder generated expressions and the original
6885   // DIExpression, place the result into an new vector.
6886   SmallVector<uint64_t, 3> NewExpr;
6887   if (DVIRec.Expr->getNumElements() == 0) {
6888     assert(DVIRec.RecoveryExprs.size() == 1 &&
6889            "Expected only a single recovery expression for an empty "
6890            "DIExpression.");
6891     assert(DVIRec.RecoveryExprs[0] &&
6892            "Expected a SCEVDbgSalvageBuilder for location 0");
6893     SCEVDbgValueBuilder *B = DVIRec.RecoveryExprs[0].get();
6894     B->appendToVectors(NewExpr, NewLocationOps);
6895   }
6896   for (const auto &Op : DVIRec.Expr->expr_ops()) {
6897     // Most Ops needn't be updated.
6898     if (Op.getOp() != dwarf::DW_OP_LLVM_arg) {
6899       Op.appendToVector(NewExpr);
6900       continue;
6901     }
6902 
6903     uint64_t LocationArgIndex = Op.getArg(0);
6904     SCEVDbgValueBuilder *DbgBuilder =
6905         DVIRec.RecoveryExprs[LocationArgIndex].get();
6906     // The location doesn't have s SCEVDbgValueBuilder, so LSR did not
6907     // optimise it away. So just translate the argument to the updated
6908     // location index.
6909     if (!DbgBuilder) {
6910       NewExpr.push_back(dwarf::DW_OP_LLVM_arg);
6911       assert(LocationOpIndexMap[Op.getArg(0)] != -1 &&
6912              "Expected a positive index for the location-op position.");
6913       NewExpr.push_back(LocationOpIndexMap[Op.getArg(0)]);
6914       continue;
6915     }
6916     // The location has a recovery expression.
6917     DbgBuilder->appendToVectors(NewExpr, NewLocationOps);
6918   }
6919 
6920   UpdateDbgValueInst(DVIRec, NewLocationOps, NewExpr);
6921   if (isa<DbgValueInst *>(DVIRec.DbgRef))
6922     LLVM_DEBUG(dbgs() << "scev-salvage: Updated DVI: "
6923                       << *cast<DbgValueInst *>(DVIRec.DbgRef) << "\n");
6924   else
6925     LLVM_DEBUG(dbgs() << "scev-salvage: Updated DVI: "
6926                       << *cast<DbgVariableRecord *>(DVIRec.DbgRef) << "\n");
6927   return true;
6928 }
6929 
6930 /// Obtain an expression for the iteration count, then attempt to salvage the
6931 /// dbg.value intrinsics.
DbgRewriteSalvageableDVIs(llvm::Loop * L,ScalarEvolution & SE,llvm::PHINode * LSRInductionVar,SmallVector<std::unique_ptr<DVIRecoveryRec>,2> & DVIToUpdate)6932 static void DbgRewriteSalvageableDVIs(
6933     llvm::Loop *L, ScalarEvolution &SE, llvm::PHINode *LSRInductionVar,
6934     SmallVector<std::unique_ptr<DVIRecoveryRec>, 2> &DVIToUpdate) {
6935   if (DVIToUpdate.empty())
6936     return;
6937 
6938   const llvm::SCEV *SCEVInductionVar = SE.getSCEV(LSRInductionVar);
6939   assert(SCEVInductionVar &&
6940          "Anticipated a SCEV for the post-LSR induction variable");
6941 
6942   if (const SCEVAddRecExpr *IVAddRec =
6943           dyn_cast<SCEVAddRecExpr>(SCEVInductionVar)) {
6944     if (!IVAddRec->isAffine())
6945       return;
6946 
6947     // Prevent translation using excessive resources.
6948     if (IVAddRec->getExpressionSize() > MaxSCEVSalvageExpressionSize)
6949       return;
6950 
6951     // The iteration count is required to recover location values.
6952     SCEVDbgValueBuilder IterCountExpr;
6953     IterCountExpr.pushLocation(LSRInductionVar);
6954     if (!IterCountExpr.SCEVToIterCountExpr(*IVAddRec, SE))
6955       return;
6956 
6957     LLVM_DEBUG(dbgs() << "scev-salvage: IV SCEV: " << *SCEVInductionVar
6958                       << '\n');
6959 
6960     for (auto &DVIRec : DVIToUpdate) {
6961       SalvageDVI(L, SE, LSRInductionVar, *DVIRec, SCEVInductionVar,
6962                  IterCountExpr);
6963     }
6964   }
6965 }
6966 
6967 /// Identify and cache salvageable DVI locations and expressions along with the
6968 /// corresponding SCEV(s). Also ensure that the DVI is not deleted between
6969 /// cacheing and salvaging.
DbgGatherSalvagableDVI(Loop * L,ScalarEvolution & SE,SmallVector<std::unique_ptr<DVIRecoveryRec>,2> & SalvageableDVISCEVs,SmallSet<AssertingVH<DbgValueInst>,2> & DVIHandles)6970 static void DbgGatherSalvagableDVI(
6971     Loop *L, ScalarEvolution &SE,
6972     SmallVector<std::unique_ptr<DVIRecoveryRec>, 2> &SalvageableDVISCEVs,
6973     SmallSet<AssertingVH<DbgValueInst>, 2> &DVIHandles) {
6974   for (const auto &B : L->getBlocks()) {
6975     for (auto &I : *B) {
6976       auto ProcessDbgValue = [&](auto *DbgVal) -> bool {
6977         // Ensure that if any location op is undef that the dbg.vlue is not
6978         // cached.
6979         if (DbgVal->isKillLocation())
6980           return false;
6981 
6982         // Check that the location op SCEVs are suitable for translation to
6983         // DIExpression.
6984         const auto &HasTranslatableLocationOps =
6985             [&](const auto *DbgValToTranslate) -> bool {
6986           for (const auto LocOp : DbgValToTranslate->location_ops()) {
6987             if (!LocOp)
6988               return false;
6989 
6990             if (!SE.isSCEVable(LocOp->getType()))
6991               return false;
6992 
6993             const SCEV *S = SE.getSCEV(LocOp);
6994             if (SE.containsUndefs(S))
6995               return false;
6996           }
6997           return true;
6998         };
6999 
7000         if (!HasTranslatableLocationOps(DbgVal))
7001           return false;
7002 
7003         std::unique_ptr<DVIRecoveryRec> NewRec =
7004             std::make_unique<DVIRecoveryRec>(DbgVal);
7005         // Each location Op may need a SCEVDbgValueBuilder in order to recover
7006         // it. Pre-allocating a vector will enable quick lookups of the builder
7007         // later during the salvage.
7008         NewRec->RecoveryExprs.resize(DbgVal->getNumVariableLocationOps());
7009         for (const auto LocOp : DbgVal->location_ops()) {
7010           NewRec->SCEVs.push_back(SE.getSCEV(LocOp));
7011           NewRec->LocationOps.push_back(LocOp);
7012           NewRec->HadLocationArgList = DbgVal->hasArgList();
7013         }
7014         SalvageableDVISCEVs.push_back(std::move(NewRec));
7015         return true;
7016       };
7017       for (DbgVariableRecord &DVR : filterDbgVars(I.getDbgRecordRange())) {
7018         if (DVR.isDbgValue() || DVR.isDbgAssign())
7019           ProcessDbgValue(&DVR);
7020       }
7021       auto DVI = dyn_cast<DbgValueInst>(&I);
7022       if (!DVI)
7023         continue;
7024       if (ProcessDbgValue(DVI))
7025         DVIHandles.insert(DVI);
7026     }
7027   }
7028 }
7029 
7030 /// Ideally pick the PHI IV inserted by ScalarEvolutionExpander. As a fallback
7031 /// any PHi from the loop header is usable, but may have less chance of
7032 /// surviving subsequent transforms.
GetInductionVariable(const Loop & L,ScalarEvolution & SE,const LSRInstance & LSR)7033 static llvm::PHINode *GetInductionVariable(const Loop &L, ScalarEvolution &SE,
7034                                            const LSRInstance &LSR) {
7035 
7036   auto IsSuitableIV = [&](PHINode *P) {
7037     if (!SE.isSCEVable(P->getType()))
7038       return false;
7039     if (const SCEVAddRecExpr *Rec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(P)))
7040       return Rec->isAffine() && !SE.containsUndefs(SE.getSCEV(P));
7041     return false;
7042   };
7043 
7044   // For now, just pick the first IV that was generated and inserted by
7045   // ScalarEvolution. Ideally pick an IV that is unlikely to be optimised away
7046   // by subsequent transforms.
7047   for (const WeakVH &IV : LSR.getScalarEvolutionIVs()) {
7048     if (!IV)
7049       continue;
7050 
7051     // There should only be PHI node IVs.
7052     PHINode *P = cast<PHINode>(&*IV);
7053 
7054     if (IsSuitableIV(P))
7055       return P;
7056   }
7057 
7058   for (PHINode &P : L.getHeader()->phis()) {
7059     if (IsSuitableIV(&P))
7060       return &P;
7061   }
7062   return nullptr;
7063 }
7064 
7065 static std::optional<std::tuple<PHINode *, PHINode *, const SCEV *, bool>>
canFoldTermCondOfLoop(Loop * L,ScalarEvolution & SE,DominatorTree & DT,const LoopInfo & LI,const TargetTransformInfo & TTI)7066 canFoldTermCondOfLoop(Loop *L, ScalarEvolution &SE, DominatorTree &DT,
7067                       const LoopInfo &LI, const TargetTransformInfo &TTI) {
7068   if (!L->isInnermost()) {
7069     LLVM_DEBUG(dbgs() << "Cannot fold on non-innermost loop\n");
7070     return std::nullopt;
7071   }
7072   // Only inspect on simple loop structure
7073   if (!L->isLoopSimplifyForm()) {
7074     LLVM_DEBUG(dbgs() << "Cannot fold on non-simple loop\n");
7075     return std::nullopt;
7076   }
7077 
7078   if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
7079     LLVM_DEBUG(dbgs() << "Cannot fold on backedge that is loop variant\n");
7080     return std::nullopt;
7081   }
7082 
7083   BasicBlock *LoopLatch = L->getLoopLatch();
7084   BranchInst *BI = dyn_cast<BranchInst>(LoopLatch->getTerminator());
7085   if (!BI || BI->isUnconditional())
7086     return std::nullopt;
7087   auto *TermCond = dyn_cast<ICmpInst>(BI->getCondition());
7088   if (!TermCond) {
7089     LLVM_DEBUG(
7090         dbgs() << "Cannot fold on branching condition that is not an ICmpInst");
7091     return std::nullopt;
7092   }
7093   if (!TermCond->hasOneUse()) {
7094     LLVM_DEBUG(
7095         dbgs()
7096         << "Cannot replace terminating condition with more than one use\n");
7097     return std::nullopt;
7098   }
7099 
7100   BinaryOperator *LHS = dyn_cast<BinaryOperator>(TermCond->getOperand(0));
7101   Value *RHS = TermCond->getOperand(1);
7102   if (!LHS || !L->isLoopInvariant(RHS))
7103     // We could pattern match the inverse form of the icmp, but that is
7104     // non-canonical, and this pass is running *very* late in the pipeline.
7105     return std::nullopt;
7106 
7107   // Find the IV used by the current exit condition.
7108   PHINode *ToFold;
7109   Value *ToFoldStart, *ToFoldStep;
7110   if (!matchSimpleRecurrence(LHS, ToFold, ToFoldStart, ToFoldStep))
7111     return std::nullopt;
7112 
7113   // Ensure the simple recurrence is a part of the current loop.
7114   if (ToFold->getParent() != L->getHeader())
7115     return std::nullopt;
7116 
7117   // If that IV isn't dead after we rewrite the exit condition in terms of
7118   // another IV, there's no point in doing the transform.
7119   if (!isAlmostDeadIV(ToFold, LoopLatch, TermCond))
7120     return std::nullopt;
7121 
7122   // Inserting instructions in the preheader has a runtime cost, scale
7123   // the allowed cost with the loops trip count as best we can.
7124   const unsigned ExpansionBudget = [&]() {
7125     unsigned Budget = 2 * SCEVCheapExpansionBudget;
7126     if (unsigned SmallTC = SE.getSmallConstantMaxTripCount(L))
7127       return std::min(Budget, SmallTC);
7128     if (std::optional<unsigned> SmallTC = getLoopEstimatedTripCount(L))
7129       return std::min(Budget, *SmallTC);
7130     // Unknown trip count, assume long running by default.
7131     return Budget;
7132   }();
7133 
7134   const SCEV *BECount = SE.getBackedgeTakenCount(L);
7135   const DataLayout &DL = L->getHeader()->getDataLayout();
7136   SCEVExpander Expander(SE, DL, "lsr_fold_term_cond");
7137 
7138   PHINode *ToHelpFold = nullptr;
7139   const SCEV *TermValueS = nullptr;
7140   bool MustDropPoison = false;
7141   auto InsertPt = L->getLoopPreheader()->getTerminator();
7142   for (PHINode &PN : L->getHeader()->phis()) {
7143     if (ToFold == &PN)
7144       continue;
7145 
7146     if (!SE.isSCEVable(PN.getType())) {
7147       LLVM_DEBUG(dbgs() << "IV of phi '" << PN
7148                         << "' is not SCEV-able, not qualified for the "
7149                            "terminating condition folding.\n");
7150       continue;
7151     }
7152     const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(&PN));
7153     // Only speculate on affine AddRec
7154     if (!AddRec || !AddRec->isAffine()) {
7155       LLVM_DEBUG(dbgs() << "SCEV of phi '" << PN
7156                         << "' is not an affine add recursion, not qualified "
7157                            "for the terminating condition folding.\n");
7158       continue;
7159     }
7160 
7161     // Check that we can compute the value of AddRec on the exiting iteration
7162     // without soundness problems.  evaluateAtIteration internally needs
7163     // to multiply the stride of the iteration number - which may wrap around.
7164     // The issue here is subtle because computing the result accounting for
7165     // wrap is insufficient. In order to use the result in an exit test, we
7166     // must also know that AddRec doesn't take the same value on any previous
7167     // iteration. The simplest case to consider is a candidate IV which is
7168     // narrower than the trip count (and thus original IV), but this can
7169     // also happen due to non-unit strides on the candidate IVs.
7170     if (!AddRec->hasNoSelfWrap() ||
7171         !SE.isKnownNonZero(AddRec->getStepRecurrence(SE)))
7172       continue;
7173 
7174     const SCEVAddRecExpr *PostInc = AddRec->getPostIncExpr(SE);
7175     const SCEV *TermValueSLocal = PostInc->evaluateAtIteration(BECount, SE);
7176     if (!Expander.isSafeToExpand(TermValueSLocal)) {
7177       LLVM_DEBUG(
7178           dbgs() << "Is not safe to expand terminating value for phi node" << PN
7179                  << "\n");
7180       continue;
7181     }
7182 
7183     if (Expander.isHighCostExpansion(TermValueSLocal, L, ExpansionBudget,
7184                                      &TTI, InsertPt)) {
7185       LLVM_DEBUG(
7186           dbgs() << "Is too expensive to expand terminating value for phi node"
7187                  << PN << "\n");
7188       continue;
7189     }
7190 
7191     // The candidate IV may have been otherwise dead and poison from the
7192     // very first iteration.  If we can't disprove that, we can't use the IV.
7193     if (!mustExecuteUBIfPoisonOnPathTo(&PN, LoopLatch->getTerminator(), &DT)) {
7194       LLVM_DEBUG(dbgs() << "Can not prove poison safety for IV "
7195                         << PN << "\n");
7196       continue;
7197     }
7198 
7199     // The candidate IV may become poison on the last iteration.  If this
7200     // value is not branched on, this is a well defined program.  We're
7201     // about to add a new use to this IV, and we have to ensure we don't
7202     // insert UB which didn't previously exist.
7203     bool MustDropPoisonLocal = false;
7204     Instruction *PostIncV =
7205       cast<Instruction>(PN.getIncomingValueForBlock(LoopLatch));
7206     if (!mustExecuteUBIfPoisonOnPathTo(PostIncV, LoopLatch->getTerminator(),
7207                                        &DT)) {
7208       LLVM_DEBUG(dbgs() << "Can not prove poison safety to insert use"
7209                         << PN << "\n");
7210 
7211       // If this is a complex recurrance with multiple instructions computing
7212       // the backedge value, we might need to strip poison flags from all of
7213       // them.
7214       if (PostIncV->getOperand(0) != &PN)
7215         continue;
7216 
7217       // In order to perform the transform, we need to drop the poison generating
7218       // flags on this instruction (if any).
7219       MustDropPoisonLocal = PostIncV->hasPoisonGeneratingFlags();
7220     }
7221 
7222     // We pick the last legal alternate IV.  We could expore choosing an optimal
7223     // alternate IV if we had a decent heuristic to do so.
7224     ToHelpFold = &PN;
7225     TermValueS = TermValueSLocal;
7226     MustDropPoison = MustDropPoisonLocal;
7227   }
7228 
7229   LLVM_DEBUG(if (ToFold && !ToHelpFold) dbgs()
7230                  << "Cannot find other AddRec IV to help folding\n";);
7231 
7232   LLVM_DEBUG(if (ToFold && ToHelpFold) dbgs()
7233              << "\nFound loop that can fold terminating condition\n"
7234              << "  BECount (SCEV): " << *SE.getBackedgeTakenCount(L) << "\n"
7235              << "  TermCond: " << *TermCond << "\n"
7236              << "  BrandInst: " << *BI << "\n"
7237              << "  ToFold: " << *ToFold << "\n"
7238              << "  ToHelpFold: " << *ToHelpFold << "\n");
7239 
7240   if (!ToFold || !ToHelpFold)
7241     return std::nullopt;
7242   return std::make_tuple(ToFold, ToHelpFold, TermValueS, MustDropPoison);
7243 }
7244 
ReduceLoopStrength(Loop * L,IVUsers & IU,ScalarEvolution & SE,DominatorTree & DT,LoopInfo & LI,const TargetTransformInfo & TTI,AssumptionCache & AC,TargetLibraryInfo & TLI,MemorySSA * MSSA)7245 static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
7246                                DominatorTree &DT, LoopInfo &LI,
7247                                const TargetTransformInfo &TTI,
7248                                AssumptionCache &AC, TargetLibraryInfo &TLI,
7249                                MemorySSA *MSSA) {
7250 
7251   // Debug preservation - before we start removing anything identify which DVI
7252   // meet the salvageable criteria and store their DIExpression and SCEVs.
7253   SmallVector<std::unique_ptr<DVIRecoveryRec>, 2> SalvageableDVIRecords;
7254   SmallSet<AssertingVH<DbgValueInst>, 2> DVIHandles;
7255   DbgGatherSalvagableDVI(L, SE, SalvageableDVIRecords, DVIHandles);
7256 
7257   bool Changed = false;
7258   std::unique_ptr<MemorySSAUpdater> MSSAU;
7259   if (MSSA)
7260     MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
7261 
7262   // Run the main LSR transformation.
7263   const LSRInstance &Reducer =
7264       LSRInstance(L, IU, SE, DT, LI, TTI, AC, TLI, MSSAU.get());
7265   Changed |= Reducer.getChanged();
7266 
7267   // Remove any extra phis created by processing inner loops.
7268   Changed |= DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
7269   if (EnablePhiElim && L->isLoopSimplifyForm()) {
7270     SmallVector<WeakTrackingVH, 16> DeadInsts;
7271     const DataLayout &DL = L->getHeader()->getDataLayout();
7272     SCEVExpander Rewriter(SE, DL, "lsr", false);
7273 #ifndef NDEBUG
7274     Rewriter.setDebugType(DEBUG_TYPE);
7275 #endif
7276     unsigned numFolded = Rewriter.replaceCongruentIVs(L, &DT, DeadInsts, &TTI);
7277     Rewriter.clear();
7278     if (numFolded) {
7279       Changed = true;
7280       RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts, &TLI,
7281                                                            MSSAU.get());
7282       DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
7283     }
7284   }
7285   // LSR may at times remove all uses of an induction variable from a loop.
7286   // The only remaining use is the PHI in the exit block.
7287   // When this is the case, if the exit value of the IV can be calculated using
7288   // SCEV, we can replace the exit block PHI with the final value of the IV and
7289   // skip the updates in each loop iteration.
7290   if (L->isRecursivelyLCSSAForm(DT, LI) && L->getExitBlock()) {
7291     SmallVector<WeakTrackingVH, 16> DeadInsts;
7292     const DataLayout &DL = L->getHeader()->getDataLayout();
7293     SCEVExpander Rewriter(SE, DL, "lsr", true);
7294     int Rewrites = rewriteLoopExitValues(L, &LI, &TLI, &SE, &TTI, Rewriter, &DT,
7295                                          UnusedIndVarInLoop, DeadInsts);
7296     Rewriter.clear();
7297     if (Rewrites) {
7298       Changed = true;
7299       RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts, &TLI,
7300                                                            MSSAU.get());
7301       DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
7302     }
7303   }
7304 
7305   const bool EnableFormTerm = [&] {
7306     switch (AllowTerminatingConditionFoldingAfterLSR) {
7307     case cl::BOU_TRUE:
7308       return true;
7309     case cl::BOU_FALSE:
7310       return false;
7311     case cl::BOU_UNSET:
7312       return TTI.shouldFoldTerminatingConditionAfterLSR();
7313     }
7314     llvm_unreachable("Unhandled cl::boolOrDefault enum");
7315   }();
7316 
7317   if (EnableFormTerm) {
7318     if (auto Opt = canFoldTermCondOfLoop(L, SE, DT, LI, TTI)) {
7319       auto [ToFold, ToHelpFold, TermValueS, MustDrop] = *Opt;
7320 
7321       Changed = true;
7322       NumTermFold++;
7323 
7324       BasicBlock *LoopPreheader = L->getLoopPreheader();
7325       BasicBlock *LoopLatch = L->getLoopLatch();
7326 
7327       (void)ToFold;
7328       LLVM_DEBUG(dbgs() << "To fold phi-node:\n"
7329                         << *ToFold << "\n"
7330                         << "New term-cond phi-node:\n"
7331                         << *ToHelpFold << "\n");
7332 
7333       Value *StartValue = ToHelpFold->getIncomingValueForBlock(LoopPreheader);
7334       (void)StartValue;
7335       Value *LoopValue = ToHelpFold->getIncomingValueForBlock(LoopLatch);
7336 
7337       // See comment in canFoldTermCondOfLoop on why this is sufficient.
7338       if (MustDrop)
7339         cast<Instruction>(LoopValue)->dropPoisonGeneratingFlags();
7340 
7341       // SCEVExpander for both use in preheader and latch
7342       const DataLayout &DL = L->getHeader()->getDataLayout();
7343       SCEVExpander Expander(SE, DL, "lsr_fold_term_cond");
7344 
7345       assert(Expander.isSafeToExpand(TermValueS) &&
7346              "Terminating value was checked safe in canFoldTerminatingCondition");
7347 
7348       // Create new terminating value at loop preheader
7349       Value *TermValue = Expander.expandCodeFor(TermValueS, ToHelpFold->getType(),
7350                                                 LoopPreheader->getTerminator());
7351 
7352       LLVM_DEBUG(dbgs() << "Start value of new term-cond phi-node:\n"
7353                         << *StartValue << "\n"
7354                         << "Terminating value of new term-cond phi-node:\n"
7355                         << *TermValue << "\n");
7356 
7357       // Create new terminating condition at loop latch
7358       BranchInst *BI = cast<BranchInst>(LoopLatch->getTerminator());
7359       ICmpInst *OldTermCond = cast<ICmpInst>(BI->getCondition());
7360       IRBuilder<> LatchBuilder(LoopLatch->getTerminator());
7361       Value *NewTermCond =
7362           LatchBuilder.CreateICmp(CmpInst::ICMP_EQ, LoopValue, TermValue,
7363                                   "lsr_fold_term_cond.replaced_term_cond");
7364       // Swap successors to exit loop body if IV equals to new TermValue
7365       if (BI->getSuccessor(0) == L->getHeader())
7366         BI->swapSuccessors();
7367 
7368       LLVM_DEBUG(dbgs() << "Old term-cond:\n"
7369                         << *OldTermCond << "\n"
7370                         << "New term-cond:\n" << *NewTermCond << "\n");
7371 
7372       BI->setCondition(NewTermCond);
7373 
7374       Expander.clear();
7375       OldTermCond->eraseFromParent();
7376       DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
7377     }
7378   }
7379 
7380   if (SalvageableDVIRecords.empty())
7381     return Changed;
7382 
7383   // Obtain relevant IVs and attempt to rewrite the salvageable DVIs with
7384   // expressions composed using the derived iteration count.
7385   // TODO: Allow for multiple IV references for nested AddRecSCEVs
7386   for (const auto &L : LI) {
7387     if (llvm::PHINode *IV = GetInductionVariable(*L, SE, Reducer))
7388       DbgRewriteSalvageableDVIs(L, SE, IV, SalvageableDVIRecords);
7389     else {
7390       LLVM_DEBUG(dbgs() << "scev-salvage: SCEV salvaging not possible. An IV "
7391                            "could not be identified.\n");
7392     }
7393   }
7394 
7395   for (auto &Rec : SalvageableDVIRecords)
7396     Rec->clear();
7397   SalvageableDVIRecords.clear();
7398   DVIHandles.clear();
7399   return Changed;
7400 }
7401 
runOnLoop(Loop * L,LPPassManager &)7402 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
7403   if (skipLoop(L))
7404     return false;
7405 
7406   auto &IU = getAnalysis<IVUsersWrapperPass>().getIU();
7407   auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
7408   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
7409   auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
7410   const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
7411       *L->getHeader()->getParent());
7412   auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
7413       *L->getHeader()->getParent());
7414   auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
7415       *L->getHeader()->getParent());
7416   auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>();
7417   MemorySSA *MSSA = nullptr;
7418   if (MSSAAnalysis)
7419     MSSA = &MSSAAnalysis->getMSSA();
7420   return ReduceLoopStrength(L, IU, SE, DT, LI, TTI, AC, TLI, MSSA);
7421 }
7422 
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater &)7423 PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM,
7424                                               LoopStandardAnalysisResults &AR,
7425                                               LPMUpdater &) {
7426   if (!ReduceLoopStrength(&L, AM.getResult<IVUsersAnalysis>(L, AR), AR.SE,
7427                           AR.DT, AR.LI, AR.TTI, AR.AC, AR.TLI, AR.MSSA))
7428     return PreservedAnalyses::all();
7429 
7430   auto PA = getLoopPassPreservedAnalyses();
7431   if (AR.MSSA)
7432     PA.preserve<MemorySSAAnalysis>();
7433   return PA;
7434 }
7435 
7436 char LoopStrengthReduce::ID = 0;
7437 
7438 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
7439                       "Loop Strength Reduction", false, false)
INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)7440 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
7441 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
7442 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
7443 INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass)
7444 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
7445 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
7446 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
7447                     "Loop Strength Reduction", false, false)
7448 
7449 Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); }
7450