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/PointerIntPair.h" |
60 | #include "llvm/ADT/STLExtras.h" |
61 | #include "llvm/ADT/SetVector.h" |
62 | #include "llvm/ADT/SmallBitVector.h" |
63 | #include "llvm/ADT/SmallPtrSet.h" |
64 | #include "llvm/ADT/SmallSet.h" |
65 | #include "llvm/ADT/SmallVector.h" |
66 | #include "llvm/ADT/Statistic.h" |
67 | #include "llvm/ADT/iterator_range.h" |
68 | #include "llvm/Analysis/AssumptionCache.h" |
69 | #include "llvm/Analysis/DomTreeUpdater.h" |
70 | #include "llvm/Analysis/IVUsers.h" |
71 | #include "llvm/Analysis/LoopAnalysisManager.h" |
72 | #include "llvm/Analysis/LoopInfo.h" |
73 | #include "llvm/Analysis/LoopPass.h" |
74 | #include "llvm/Analysis/MemorySSA.h" |
75 | #include "llvm/Analysis/MemorySSAUpdater.h" |
76 | #include "llvm/Analysis/ScalarEvolution.h" |
77 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
78 | #include "llvm/Analysis/ScalarEvolutionNormalization.h" |
79 | #include "llvm/Analysis/ScalarEvolutionPatternMatch.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/IR/BasicBlock.h" |
85 | #include "llvm/IR/Constant.h" |
86 | #include "llvm/IR/Constants.h" |
87 | #include "llvm/IR/DebugInfoMetadata.h" |
88 | #include "llvm/IR/DerivedTypes.h" |
89 | #include "llvm/IR/Dominators.h" |
90 | #include "llvm/IR/GlobalValue.h" |
91 | #include "llvm/IR/IRBuilder.h" |
92 | #include "llvm/IR/InstrTypes.h" |
93 | #include "llvm/IR/Instruction.h" |
94 | #include "llvm/IR/Instructions.h" |
95 | #include "llvm/IR/IntrinsicInst.h" |
96 | #include "llvm/IR/Module.h" |
97 | #include "llvm/IR/Operator.h" |
98 | #include "llvm/IR/Type.h" |
99 | #include "llvm/IR/Use.h" |
100 | #include "llvm/IR/User.h" |
101 | #include "llvm/IR/Value.h" |
102 | #include "llvm/IR/ValueHandle.h" |
103 | #include "llvm/InitializePasses.h" |
104 | #include "llvm/Pass.h" |
105 | #include "llvm/Support/Casting.h" |
106 | #include "llvm/Support/CommandLine.h" |
107 | #include "llvm/Support/Compiler.h" |
108 | #include "llvm/Support/Debug.h" |
109 | #include "llvm/Support/ErrorHandling.h" |
110 | #include "llvm/Support/MathExtras.h" |
111 | #include "llvm/Support/raw_ostream.h" |
112 | #include "llvm/Transforms/Scalar.h" |
113 | #include "llvm/Transforms/Utils.h" |
114 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
115 | #include "llvm/Transforms/Utils/Local.h" |
116 | #include "llvm/Transforms/Utils/LoopUtils.h" |
117 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" |
118 | #include <algorithm> |
119 | #include <cassert> |
120 | #include <cstddef> |
121 | #include <cstdint> |
122 | #include <iterator> |
123 | #include <limits> |
124 | #include <map> |
125 | #include <numeric> |
126 | #include <optional> |
127 | #include <utility> |
128 | |
129 | using namespace llvm; |
130 | using namespace SCEVPatternMatch; |
131 | |
132 | #define DEBUG_TYPE "loop-reduce" |
133 | |
134 | /// MaxIVUsers is an arbitrary threshold that provides an early opportunity for |
135 | /// bail out. This threshold is far beyond the number of users that LSR can |
136 | /// conceivably solve, so it should not affect generated code, but catches the |
137 | /// worst cases before LSR burns too much compile time and stack space. |
138 | static const unsigned MaxIVUsers = 200; |
139 | |
140 | /// Limit the size of expression that SCEV-based salvaging will attempt to |
141 | /// translate into a DIExpression. |
142 | /// Choose a maximum size such that debuginfo is not excessively increased and |
143 | /// the salvaging is not too expensive for the compiler. |
144 | static const unsigned MaxSCEVSalvageExpressionSize = 64; |
145 | |
146 | // Cleanup congruent phis after LSR phi expansion. |
147 | static cl::opt<bool> EnablePhiElim( |
148 | "enable-lsr-phielim" , cl::Hidden, cl::init(Val: true), |
149 | cl::desc("Enable LSR phi elimination" )); |
150 | |
151 | // The flag adds instruction count to solutions cost comparison. |
152 | static cl::opt<bool> InsnsCost( |
153 | "lsr-insns-cost" , cl::Hidden, cl::init(Val: true), |
154 | cl::desc("Add instruction count to a LSR cost model" )); |
155 | |
156 | // Flag to choose how to narrow complex lsr solution |
157 | static cl::opt<bool> LSRExpNarrow( |
158 | "lsr-exp-narrow" , cl::Hidden, cl::init(Val: false), |
159 | cl::desc("Narrow LSR complex solution using" |
160 | " expectation of registers number" )); |
161 | |
162 | // Flag to narrow search space by filtering non-optimal formulae with |
163 | // the same ScaledReg and Scale. |
164 | static cl::opt<bool> FilterSameScaledReg( |
165 | "lsr-filter-same-scaled-reg" , cl::Hidden, cl::init(Val: true), |
166 | cl::desc("Narrow LSR search space by filtering non-optimal formulae" |
167 | " with the same ScaledReg and Scale" )); |
168 | |
169 | static cl::opt<TTI::AddressingModeKind> PreferredAddresingMode( |
170 | "lsr-preferred-addressing-mode" , cl::Hidden, cl::init(Val: TTI::AMK_None), |
171 | cl::desc("A flag that overrides the target's preferred addressing mode." ), |
172 | cl::values(clEnumValN(TTI::AMK_None, |
173 | "none" , |
174 | "Don't prefer any addressing mode" ), |
175 | clEnumValN(TTI::AMK_PreIndexed, |
176 | "preindexed" , |
177 | "Prefer pre-indexed addressing mode" ), |
178 | clEnumValN(TTI::AMK_PostIndexed, |
179 | "postindexed" , |
180 | "Prefer post-indexed addressing mode" ))); |
181 | |
182 | static cl::opt<unsigned> ComplexityLimit( |
183 | "lsr-complexity-limit" , cl::Hidden, |
184 | cl::init(Val: std::numeric_limits<uint16_t>::max()), |
185 | cl::desc("LSR search space complexity limit" )); |
186 | |
187 | static cl::opt<unsigned> SetupCostDepthLimit( |
188 | "lsr-setupcost-depth-limit" , cl::Hidden, cl::init(Val: 7), |
189 | cl::desc("The limit on recursion depth for LSRs setup cost" )); |
190 | |
191 | static cl::opt<cl::boolOrDefault> AllowDropSolutionIfLessProfitable( |
192 | "lsr-drop-solution" , cl::Hidden, |
193 | cl::desc("Attempt to drop solution if it is less profitable" )); |
194 | |
195 | static cl::opt<bool> EnableVScaleImmediates( |
196 | "lsr-enable-vscale-immediates" , cl::Hidden, cl::init(Val: true), |
197 | cl::desc("Enable analysis of vscale-relative immediates in LSR" )); |
198 | |
199 | static cl::opt<bool> DropScaledForVScale( |
200 | "lsr-drop-scaled-reg-for-vscale" , cl::Hidden, cl::init(Val: true), |
201 | cl::desc("Avoid using scaled registers with vscale-relative addressing" )); |
202 | |
203 | #ifndef NDEBUG |
204 | // Stress test IV chain generation. |
205 | static cl::opt<bool> StressIVChain( |
206 | "stress-ivchain" , cl::Hidden, cl::init(false), |
207 | cl::desc("Stress test LSR IV chains" )); |
208 | #else |
209 | static bool StressIVChain = false; |
210 | #endif |
211 | |
212 | namespace { |
213 | |
214 | struct MemAccessTy { |
215 | /// Used in situations where the accessed memory type is unknown. |
216 | static const unsigned UnknownAddressSpace = |
217 | std::numeric_limits<unsigned>::max(); |
218 | |
219 | Type *MemTy = nullptr; |
220 | unsigned AddrSpace = UnknownAddressSpace; |
221 | |
222 | MemAccessTy() = default; |
223 | MemAccessTy(Type *Ty, unsigned AS) : MemTy(Ty), AddrSpace(AS) {} |
224 | |
225 | bool operator==(MemAccessTy Other) const { |
226 | return MemTy == Other.MemTy && AddrSpace == Other.AddrSpace; |
227 | } |
228 | |
229 | bool operator!=(MemAccessTy Other) const { return !(*this == Other); } |
230 | |
231 | static MemAccessTy getUnknown(LLVMContext &Ctx, |
232 | unsigned AS = UnknownAddressSpace) { |
233 | return MemAccessTy(Type::getVoidTy(C&: Ctx), AS); |
234 | } |
235 | |
236 | Type *getType() { return MemTy; } |
237 | }; |
238 | |
239 | /// This class holds data which is used to order reuse candidates. |
240 | class RegSortData { |
241 | public: |
242 | /// This represents the set of LSRUse indices which reference |
243 | /// a particular register. |
244 | SmallBitVector UsedByIndices; |
245 | |
246 | void print(raw_ostream &OS) const; |
247 | void dump() const; |
248 | }; |
249 | |
250 | // An offset from an address that is either scalable or fixed. Used for |
251 | // per-target optimizations of addressing modes. |
252 | class Immediate : public details::FixedOrScalableQuantity<Immediate, int64_t> { |
253 | constexpr Immediate(ScalarTy MinVal, bool Scalable) |
254 | : FixedOrScalableQuantity(MinVal, Scalable) {} |
255 | |
256 | constexpr Immediate(const FixedOrScalableQuantity<Immediate, int64_t> &V) |
257 | : FixedOrScalableQuantity(V) {} |
258 | |
259 | public: |
260 | constexpr Immediate() = delete; |
261 | |
262 | static constexpr Immediate getFixed(ScalarTy MinVal) { |
263 | return {MinVal, false}; |
264 | } |
265 | static constexpr Immediate getScalable(ScalarTy MinVal) { |
266 | return {MinVal, true}; |
267 | } |
268 | static constexpr Immediate get(ScalarTy MinVal, bool Scalable) { |
269 | return {MinVal, Scalable}; |
270 | } |
271 | static constexpr Immediate getZero() { return {0, false}; } |
272 | static constexpr Immediate getFixedMin() { |
273 | return {std::numeric_limits<int64_t>::min(), false}; |
274 | } |
275 | static constexpr Immediate getFixedMax() { |
276 | return {std::numeric_limits<int64_t>::max(), false}; |
277 | } |
278 | static constexpr Immediate getScalableMin() { |
279 | return {std::numeric_limits<int64_t>::min(), true}; |
280 | } |
281 | static constexpr Immediate getScalableMax() { |
282 | return {std::numeric_limits<int64_t>::max(), true}; |
283 | } |
284 | |
285 | constexpr bool isLessThanZero() const { return Quantity < 0; } |
286 | |
287 | constexpr bool isGreaterThanZero() const { return Quantity > 0; } |
288 | |
289 | constexpr bool isCompatibleImmediate(const Immediate &Imm) const { |
290 | return isZero() || Imm.isZero() || Imm.Scalable == Scalable; |
291 | } |
292 | |
293 | constexpr bool isMin() const { |
294 | return Quantity == std::numeric_limits<ScalarTy>::min(); |
295 | } |
296 | |
297 | constexpr bool isMax() const { |
298 | return Quantity == std::numeric_limits<ScalarTy>::max(); |
299 | } |
300 | |
301 | // Arithmetic 'operators' that cast to unsigned types first. |
302 | constexpr Immediate addUnsigned(const Immediate &RHS) const { |
303 | assert(isCompatibleImmediate(RHS) && "Incompatible Immediates" ); |
304 | ScalarTy Value = (uint64_t)Quantity + RHS.getKnownMinValue(); |
305 | return {Value, Scalable || RHS.isScalable()}; |
306 | } |
307 | |
308 | constexpr Immediate subUnsigned(const Immediate &RHS) const { |
309 | assert(isCompatibleImmediate(RHS) && "Incompatible Immediates" ); |
310 | ScalarTy Value = (uint64_t)Quantity - RHS.getKnownMinValue(); |
311 | return {Value, Scalable || RHS.isScalable()}; |
312 | } |
313 | |
314 | // Scale the quantity by a constant without caring about runtime scalability. |
315 | constexpr Immediate mulUnsigned(const ScalarTy RHS) const { |
316 | ScalarTy Value = (uint64_t)Quantity * RHS; |
317 | return {Value, Scalable}; |
318 | } |
319 | |
320 | // Helpers for generating SCEVs with vscale terms where needed. |
321 | const SCEV *getSCEV(ScalarEvolution &SE, Type *Ty) const { |
322 | const SCEV *S = SE.getConstant(Ty, V: Quantity); |
323 | if (Scalable) |
324 | S = SE.getMulExpr(LHS: S, RHS: SE.getVScale(Ty: S->getType())); |
325 | return S; |
326 | } |
327 | |
328 | const SCEV *getNegativeSCEV(ScalarEvolution &SE, Type *Ty) const { |
329 | const SCEV *NegS = SE.getConstant(Ty, V: -(uint64_t)Quantity); |
330 | if (Scalable) |
331 | NegS = SE.getMulExpr(LHS: NegS, RHS: SE.getVScale(Ty: NegS->getType())); |
332 | return NegS; |
333 | } |
334 | |
335 | const SCEV *getUnknownSCEV(ScalarEvolution &SE, Type *Ty) const { |
336 | const SCEV *SU = SE.getUnknown(V: ConstantInt::getSigned(Ty, V: Quantity)); |
337 | if (Scalable) |
338 | SU = SE.getMulExpr(LHS: SU, RHS: SE.getVScale(Ty: SU->getType())); |
339 | return SU; |
340 | } |
341 | }; |
342 | |
343 | // This is needed for the Compare type of std::map when Immediate is used |
344 | // as a key. We don't need it to be fully correct against any value of vscale, |
345 | // just to make sure that vscale-related terms in the map are considered against |
346 | // each other rather than being mixed up and potentially missing opportunities. |
347 | struct KeyOrderTargetImmediate { |
348 | bool operator()(const Immediate &LHS, const Immediate &RHS) const { |
349 | if (LHS.isScalable() && !RHS.isScalable()) |
350 | return false; |
351 | if (!LHS.isScalable() && RHS.isScalable()) |
352 | return true; |
353 | return LHS.getKnownMinValue() < RHS.getKnownMinValue(); |
354 | } |
355 | }; |
356 | |
357 | // This would be nicer if we could be generic instead of directly using size_t, |
358 | // but there doesn't seem to be a type trait for is_orderable or |
359 | // is_lessthan_comparable or similar. |
360 | struct KeyOrderSizeTAndImmediate { |
361 | bool operator()(const std::pair<size_t, Immediate> &LHS, |
362 | const std::pair<size_t, Immediate> &RHS) const { |
363 | size_t LSize = LHS.first; |
364 | size_t RSize = RHS.first; |
365 | if (LSize != RSize) |
366 | return LSize < RSize; |
367 | return KeyOrderTargetImmediate()(LHS.second, RHS.second); |
368 | } |
369 | }; |
370 | } // end anonymous namespace |
371 | |
372 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
373 | void RegSortData::print(raw_ostream &OS) const { |
374 | OS << "[NumUses=" << UsedByIndices.count() << ']'; |
375 | } |
376 | |
377 | LLVM_DUMP_METHOD void RegSortData::dump() const { |
378 | print(errs()); errs() << '\n'; |
379 | } |
380 | #endif |
381 | |
382 | namespace { |
383 | |
384 | /// Map register candidates to information about how they are used. |
385 | class RegUseTracker { |
386 | using RegUsesTy = DenseMap<const SCEV *, RegSortData>; |
387 | |
388 | RegUsesTy RegUsesMap; |
389 | SmallVector<const SCEV *, 16> RegSequence; |
390 | |
391 | public: |
392 | void countRegister(const SCEV *Reg, size_t LUIdx); |
393 | void dropRegister(const SCEV *Reg, size_t LUIdx); |
394 | void swapAndDropUse(size_t LUIdx, size_t LastLUIdx); |
395 | |
396 | bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const; |
397 | |
398 | const SmallBitVector &getUsedByIndices(const SCEV *Reg) const; |
399 | |
400 | void clear(); |
401 | |
402 | using iterator = SmallVectorImpl<const SCEV *>::iterator; |
403 | using const_iterator = SmallVectorImpl<const SCEV *>::const_iterator; |
404 | |
405 | iterator begin() { return RegSequence.begin(); } |
406 | iterator end() { return RegSequence.end(); } |
407 | const_iterator begin() const { return RegSequence.begin(); } |
408 | const_iterator end() const { return RegSequence.end(); } |
409 | }; |
410 | |
411 | } // end anonymous namespace |
412 | |
413 | void |
414 | RegUseTracker::countRegister(const SCEV *Reg, size_t LUIdx) { |
415 | std::pair<RegUsesTy::iterator, bool> Pair = RegUsesMap.try_emplace(Key: Reg); |
416 | RegSortData &RSD = Pair.first->second; |
417 | if (Pair.second) |
418 | RegSequence.push_back(Elt: Reg); |
419 | RSD.UsedByIndices.resize(N: std::max(a: RSD.UsedByIndices.size(), b: LUIdx + 1)); |
420 | RSD.UsedByIndices.set(LUIdx); |
421 | } |
422 | |
423 | void |
424 | RegUseTracker::dropRegister(const SCEV *Reg, size_t LUIdx) { |
425 | RegUsesTy::iterator It = RegUsesMap.find(Val: Reg); |
426 | assert(It != RegUsesMap.end()); |
427 | RegSortData &RSD = It->second; |
428 | assert(RSD.UsedByIndices.size() > LUIdx); |
429 | RSD.UsedByIndices.reset(Idx: LUIdx); |
430 | } |
431 | |
432 | void |
433 | RegUseTracker::swapAndDropUse(size_t LUIdx, size_t LastLUIdx) { |
434 | assert(LUIdx <= LastLUIdx); |
435 | |
436 | // Update RegUses. The data structure is not optimized for this purpose; |
437 | // we must iterate through it and update each of the bit vectors. |
438 | for (auto &Pair : RegUsesMap) { |
439 | SmallBitVector &UsedByIndices = Pair.second.UsedByIndices; |
440 | if (LUIdx < UsedByIndices.size()) |
441 | UsedByIndices[LUIdx] = |
442 | LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : false; |
443 | UsedByIndices.resize(N: std::min(a: UsedByIndices.size(), b: LastLUIdx)); |
444 | } |
445 | } |
446 | |
447 | bool |
448 | RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const { |
449 | RegUsesTy::const_iterator I = RegUsesMap.find(Val: Reg); |
450 | if (I == RegUsesMap.end()) |
451 | return false; |
452 | const SmallBitVector &UsedByIndices = I->second.UsedByIndices; |
453 | int i = UsedByIndices.find_first(); |
454 | if (i == -1) return false; |
455 | if ((size_t)i != LUIdx) return true; |
456 | return UsedByIndices.find_next(Prev: i) != -1; |
457 | } |
458 | |
459 | const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const { |
460 | RegUsesTy::const_iterator I = RegUsesMap.find(Val: Reg); |
461 | assert(I != RegUsesMap.end() && "Unknown register!" ); |
462 | return I->second.UsedByIndices; |
463 | } |
464 | |
465 | void RegUseTracker::clear() { |
466 | RegUsesMap.clear(); |
467 | RegSequence.clear(); |
468 | } |
469 | |
470 | namespace { |
471 | |
472 | /// This class holds information that describes a formula for computing |
473 | /// satisfying a use. It may include broken-out immediates and scaled registers. |
474 | struct Formula { |
475 | /// Global base address used for complex addressing. |
476 | GlobalValue *BaseGV = nullptr; |
477 | |
478 | /// Base offset for complex addressing. |
479 | Immediate BaseOffset = Immediate::getZero(); |
480 | |
481 | /// Whether any complex addressing has a base register. |
482 | bool HasBaseReg = false; |
483 | |
484 | /// The scale of any complex addressing. |
485 | int64_t Scale = 0; |
486 | |
487 | /// The list of "base" registers for this use. When this is non-empty. The |
488 | /// canonical representation of a formula is |
489 | /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and |
490 | /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty(). |
491 | /// 3. The reg containing recurrent expr related with currect loop in the |
492 | /// formula should be put in the ScaledReg. |
493 | /// #1 enforces that the scaled register is always used when at least two |
494 | /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2. |
495 | /// #2 enforces that 1 * reg is reg. |
496 | /// #3 ensures invariant regs with respect to current loop can be combined |
497 | /// together in LSR codegen. |
498 | /// This invariant can be temporarily broken while building a formula. |
499 | /// However, every formula inserted into the LSRInstance must be in canonical |
500 | /// form. |
501 | SmallVector<const SCEV *, 4> BaseRegs; |
502 | |
503 | /// The 'scaled' register for this use. This should be non-null when Scale is |
504 | /// not zero. |
505 | const SCEV *ScaledReg = nullptr; |
506 | |
507 | /// An additional constant offset which added near the use. This requires a |
508 | /// temporary register, but the offset itself can live in an add immediate |
509 | /// field rather than a register. |
510 | Immediate UnfoldedOffset = Immediate::getZero(); |
511 | |
512 | Formula() = default; |
513 | |
514 | void initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE); |
515 | |
516 | bool isCanonical(const Loop &L) const; |
517 | |
518 | void canonicalize(const Loop &L); |
519 | |
520 | bool unscale(); |
521 | |
522 | bool hasZeroEnd() const; |
523 | |
524 | size_t getNumRegs() const; |
525 | Type *getType() const; |
526 | |
527 | void deleteBaseReg(const SCEV *&S); |
528 | |
529 | bool referencesReg(const SCEV *S) const; |
530 | bool hasRegsUsedByUsesOtherThan(size_t LUIdx, |
531 | const RegUseTracker &RegUses) const; |
532 | |
533 | void print(raw_ostream &OS) const; |
534 | void dump() const; |
535 | }; |
536 | |
537 | } // end anonymous namespace |
538 | |
539 | /// Recursion helper for initialMatch. |
540 | static void DoInitialMatch(const SCEV *S, Loop *L, |
541 | SmallVectorImpl<const SCEV *> &Good, |
542 | SmallVectorImpl<const SCEV *> &Bad, |
543 | ScalarEvolution &SE) { |
544 | // Collect expressions which properly dominate the loop header. |
545 | if (SE.properlyDominates(S, BB: L->getHeader())) { |
546 | Good.push_back(Elt: S); |
547 | return; |
548 | } |
549 | |
550 | // Look at add operands. |
551 | if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Val: S)) { |
552 | for (const SCEV *S : Add->operands()) |
553 | DoInitialMatch(S, L, Good, Bad, SE); |
554 | return; |
555 | } |
556 | |
557 | // Look at addrec operands. |
558 | const SCEV *Start, *Step; |
559 | const Loop *ARLoop; |
560 | if (match(S, |
561 | P: m_scev_AffineAddRec(Op0: m_SCEV(V&: Start), Op1: m_SCEV(V&: Step), L: m_Loop(L&: ARLoop))) && |
562 | !Start->isZero()) { |
563 | DoInitialMatch(S: Start, L, Good, Bad, SE); |
564 | DoInitialMatch(S: SE.getAddRecExpr(Start: SE.getConstant(Ty: S->getType(), V: 0), Step, |
565 | // FIXME: AR->getNoWrapFlags() |
566 | L: ARLoop, Flags: SCEV::FlagAnyWrap), |
567 | L, Good, Bad, SE); |
568 | return; |
569 | } |
570 | |
571 | // Handle a multiplication by -1 (negation) if it didn't fold. |
572 | if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Val: S)) |
573 | if (Mul->getOperand(i: 0)->isAllOnesValue()) { |
574 | SmallVector<const SCEV *, 4> Ops(drop_begin(RangeOrContainer: Mul->operands())); |
575 | const SCEV *NewMul = SE.getMulExpr(Ops); |
576 | |
577 | SmallVector<const SCEV *, 4> MyGood; |
578 | SmallVector<const SCEV *, 4> MyBad; |
579 | DoInitialMatch(S: NewMul, L, Good&: MyGood, Bad&: MyBad, SE); |
580 | const SCEV *NegOne = SE.getSCEV(V: ConstantInt::getAllOnesValue( |
581 | Ty: SE.getEffectiveSCEVType(Ty: NewMul->getType()))); |
582 | for (const SCEV *S : MyGood) |
583 | Good.push_back(Elt: SE.getMulExpr(LHS: NegOne, RHS: S)); |
584 | for (const SCEV *S : MyBad) |
585 | Bad.push_back(Elt: SE.getMulExpr(LHS: NegOne, RHS: S)); |
586 | return; |
587 | } |
588 | |
589 | // Ok, we can't do anything interesting. Just stuff the whole thing into a |
590 | // register and hope for the best. |
591 | Bad.push_back(Elt: S); |
592 | } |
593 | |
594 | /// Incorporate loop-variant parts of S into this Formula, attempting to keep |
595 | /// all loop-invariant and loop-computable values in a single base register. |
596 | void Formula::initialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) { |
597 | SmallVector<const SCEV *, 4> Good; |
598 | SmallVector<const SCEV *, 4> Bad; |
599 | DoInitialMatch(S, L, Good, Bad, SE); |
600 | if (!Good.empty()) { |
601 | const SCEV *Sum = SE.getAddExpr(Ops&: Good); |
602 | if (!Sum->isZero()) |
603 | BaseRegs.push_back(Elt: Sum); |
604 | HasBaseReg = true; |
605 | } |
606 | if (!Bad.empty()) { |
607 | const SCEV *Sum = SE.getAddExpr(Ops&: Bad); |
608 | if (!Sum->isZero()) |
609 | BaseRegs.push_back(Elt: Sum); |
610 | HasBaseReg = true; |
611 | } |
612 | canonicalize(L: *L); |
613 | } |
614 | |
615 | static bool containsAddRecDependentOnLoop(const SCEV *S, const Loop &L) { |
616 | return SCEVExprContains(Root: S, Pred: [&L](const SCEV *S) { |
617 | return isa<SCEVAddRecExpr>(Val: S) && (cast<SCEVAddRecExpr>(Val: S)->getLoop() == &L); |
618 | }); |
619 | } |
620 | |
621 | /// Check whether or not this formula satisfies the canonical |
622 | /// representation. |
623 | /// \see Formula::BaseRegs. |
624 | bool Formula::isCanonical(const Loop &L) const { |
625 | assert((Scale == 0 || ScaledReg) && |
626 | "ScaledReg must be non-null if Scale is non-zero" ); |
627 | |
628 | if (!ScaledReg) |
629 | return BaseRegs.size() <= 1; |
630 | |
631 | if (Scale != 1) |
632 | return true; |
633 | |
634 | if (Scale == 1 && BaseRegs.empty()) |
635 | return false; |
636 | |
637 | if (containsAddRecDependentOnLoop(S: ScaledReg, L)) |
638 | return true; |
639 | |
640 | // If ScaledReg is not a recurrent expr, or it is but its loop is not current |
641 | // loop, meanwhile BaseRegs contains a recurrent expr reg related with current |
642 | // loop, we want to swap the reg in BaseRegs with ScaledReg. |
643 | return none_of(Range: BaseRegs, P: [&L](const SCEV *S) { |
644 | return containsAddRecDependentOnLoop(S, L); |
645 | }); |
646 | } |
647 | |
648 | /// Helper method to morph a formula into its canonical representation. |
649 | /// \see Formula::BaseRegs. |
650 | /// Every formula having more than one base register, must use the ScaledReg |
651 | /// field. Otherwise, we would have to do special cases everywhere in LSR |
652 | /// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ... |
653 | /// On the other hand, 1*reg should be canonicalized into reg. |
654 | void Formula::canonicalize(const Loop &L) { |
655 | if (isCanonical(L)) |
656 | return; |
657 | |
658 | if (BaseRegs.empty()) { |
659 | // No base reg? Use scale reg with scale = 1 as such. |
660 | assert(ScaledReg && "Expected 1*reg => reg" ); |
661 | assert(Scale == 1 && "Expected 1*reg => reg" ); |
662 | BaseRegs.push_back(Elt: ScaledReg); |
663 | Scale = 0; |
664 | ScaledReg = nullptr; |
665 | return; |
666 | } |
667 | |
668 | // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg. |
669 | if (!ScaledReg) { |
670 | ScaledReg = BaseRegs.pop_back_val(); |
671 | Scale = 1; |
672 | } |
673 | |
674 | // If ScaledReg is an invariant with respect to L, find the reg from |
675 | // BaseRegs containing the recurrent expr related with Loop L. Swap the |
676 | // reg with ScaledReg. |
677 | if (!containsAddRecDependentOnLoop(S: ScaledReg, L)) { |
678 | auto I = find_if(Range&: BaseRegs, P: [&L](const SCEV *S) { |
679 | return containsAddRecDependentOnLoop(S, L); |
680 | }); |
681 | if (I != BaseRegs.end()) |
682 | std::swap(a&: ScaledReg, b&: *I); |
683 | } |
684 | assert(isCanonical(L) && "Failed to canonicalize?" ); |
685 | } |
686 | |
687 | /// Get rid of the scale in the formula. |
688 | /// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2. |
689 | /// \return true if it was possible to get rid of the scale, false otherwise. |
690 | /// \note After this operation the formula may not be in the canonical form. |
691 | bool Formula::unscale() { |
692 | if (Scale != 1) |
693 | return false; |
694 | Scale = 0; |
695 | BaseRegs.push_back(Elt: ScaledReg); |
696 | ScaledReg = nullptr; |
697 | return true; |
698 | } |
699 | |
700 | bool Formula::hasZeroEnd() const { |
701 | if (UnfoldedOffset || BaseOffset) |
702 | return false; |
703 | if (BaseRegs.size() != 1 || ScaledReg) |
704 | return false; |
705 | return true; |
706 | } |
707 | |
708 | /// Return the total number of register operands used by this formula. This does |
709 | /// not include register uses implied by non-constant addrec strides. |
710 | size_t Formula::getNumRegs() const { |
711 | return !!ScaledReg + BaseRegs.size(); |
712 | } |
713 | |
714 | /// Return the type of this formula, if it has one, or null otherwise. This type |
715 | /// is meaningless except for the bit size. |
716 | Type *Formula::getType() const { |
717 | return !BaseRegs.empty() ? BaseRegs.front()->getType() : |
718 | ScaledReg ? ScaledReg->getType() : |
719 | BaseGV ? BaseGV->getType() : |
720 | nullptr; |
721 | } |
722 | |
723 | /// Delete the given base reg from the BaseRegs list. |
724 | void Formula::deleteBaseReg(const SCEV *&S) { |
725 | if (&S != &BaseRegs.back()) |
726 | std::swap(a&: S, b&: BaseRegs.back()); |
727 | BaseRegs.pop_back(); |
728 | } |
729 | |
730 | /// Test if this formula references the given register. |
731 | bool Formula::referencesReg(const SCEV *S) const { |
732 | return S == ScaledReg || is_contained(Range: BaseRegs, Element: S); |
733 | } |
734 | |
735 | /// Test whether this formula uses registers which are used by uses other than |
736 | /// the use with the given index. |
737 | bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx, |
738 | const RegUseTracker &RegUses) const { |
739 | if (ScaledReg) |
740 | if (RegUses.isRegUsedByUsesOtherThan(Reg: ScaledReg, LUIdx)) |
741 | return true; |
742 | for (const SCEV *BaseReg : BaseRegs) |
743 | if (RegUses.isRegUsedByUsesOtherThan(Reg: BaseReg, LUIdx)) |
744 | return true; |
745 | return false; |
746 | } |
747 | |
748 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
749 | void Formula::print(raw_ostream &OS) const { |
750 | bool First = true; |
751 | if (BaseGV) { |
752 | if (!First) OS << " + " ; else First = false; |
753 | BaseGV->printAsOperand(OS, /*PrintType=*/false); |
754 | } |
755 | if (BaseOffset.isNonZero()) { |
756 | if (!First) OS << " + " ; else First = false; |
757 | OS << BaseOffset; |
758 | } |
759 | for (const SCEV *BaseReg : BaseRegs) { |
760 | if (!First) OS << " + " ; else First = false; |
761 | OS << "reg(" << *BaseReg << ')'; |
762 | } |
763 | if (HasBaseReg && BaseRegs.empty()) { |
764 | if (!First) OS << " + " ; else First = false; |
765 | OS << "**error: HasBaseReg**" ; |
766 | } else if (!HasBaseReg && !BaseRegs.empty()) { |
767 | if (!First) OS << " + " ; else First = false; |
768 | OS << "**error: !HasBaseReg**" ; |
769 | } |
770 | if (Scale != 0) { |
771 | if (!First) OS << " + " ; else First = false; |
772 | OS << Scale << "*reg(" ; |
773 | if (ScaledReg) |
774 | OS << *ScaledReg; |
775 | else |
776 | OS << "<unknown>" ; |
777 | OS << ')'; |
778 | } |
779 | if (UnfoldedOffset.isNonZero()) { |
780 | if (!First) OS << " + " ; |
781 | OS << "imm(" << UnfoldedOffset << ')'; |
782 | } |
783 | } |
784 | |
785 | LLVM_DUMP_METHOD void Formula::dump() const { |
786 | print(errs()); errs() << '\n'; |
787 | } |
788 | #endif |
789 | |
790 | /// Return true if the given addrec can be sign-extended without changing its |
791 | /// value. |
792 | static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { |
793 | Type *WideTy = |
794 | IntegerType::get(C&: SE.getContext(), NumBits: SE.getTypeSizeInBits(Ty: AR->getType()) + 1); |
795 | return isa<SCEVAddRecExpr>(Val: SE.getSignExtendExpr(Op: AR, Ty: WideTy)); |
796 | } |
797 | |
798 | /// Return true if the given add can be sign-extended without changing its |
799 | /// value. |
800 | static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) { |
801 | Type *WideTy = |
802 | IntegerType::get(C&: SE.getContext(), NumBits: SE.getTypeSizeInBits(Ty: A->getType()) + 1); |
803 | return isa<SCEVAddExpr>(Val: SE.getSignExtendExpr(Op: A, Ty: WideTy)); |
804 | } |
805 | |
806 | /// Return true if the given mul can be sign-extended without changing its |
807 | /// value. |
808 | static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) { |
809 | Type *WideTy = |
810 | IntegerType::get(C&: SE.getContext(), |
811 | NumBits: SE.getTypeSizeInBits(Ty: M->getType()) * M->getNumOperands()); |
812 | return isa<SCEVMulExpr>(Val: SE.getSignExtendExpr(Op: M, Ty: WideTy)); |
813 | } |
814 | |
815 | /// Return an expression for LHS /s RHS, if it can be determined and if the |
816 | /// remainder is known to be zero, or null otherwise. If IgnoreSignificantBits |
817 | /// is true, expressions like (X * Y) /s Y are simplified to X, ignoring that |
818 | /// the multiplication may overflow, which is useful when the result will be |
819 | /// used in a context where the most significant bits are ignored. |
820 | static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS, |
821 | ScalarEvolution &SE, |
822 | bool IgnoreSignificantBits = false) { |
823 | // Handle the trivial case, which works for any SCEV type. |
824 | if (LHS == RHS) |
825 | return SE.getConstant(Ty: LHS->getType(), V: 1); |
826 | |
827 | // Handle a few RHS special cases. |
828 | const SCEVConstant *RC = dyn_cast<SCEVConstant>(Val: RHS); |
829 | if (RC) { |
830 | const APInt &RA = RC->getAPInt(); |
831 | // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do |
832 | // some folding. |
833 | if (RA.isAllOnes()) { |
834 | if (LHS->getType()->isPointerTy()) |
835 | return nullptr; |
836 | return SE.getMulExpr(LHS, RHS: RC); |
837 | } |
838 | // Handle x /s 1 as x. |
839 | if (RA == 1) |
840 | return LHS; |
841 | } |
842 | |
843 | // Check for a division of a constant by a constant. |
844 | if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Val: LHS)) { |
845 | if (!RC) |
846 | return nullptr; |
847 | const APInt &LA = C->getAPInt(); |
848 | const APInt &RA = RC->getAPInt(); |
849 | if (LA.srem(RHS: RA) != 0) |
850 | return nullptr; |
851 | return SE.getConstant(Val: LA.sdiv(RHS: RA)); |
852 | } |
853 | |
854 | // Distribute the sdiv over addrec operands, if the addrec doesn't overflow. |
855 | if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: LHS)) { |
856 | if ((IgnoreSignificantBits || isAddRecSExtable(AR, SE)) && AR->isAffine()) { |
857 | const SCEV *Step = getExactSDiv(LHS: AR->getStepRecurrence(SE), RHS, SE, |
858 | IgnoreSignificantBits); |
859 | if (!Step) return nullptr; |
860 | const SCEV *Start = getExactSDiv(LHS: AR->getStart(), RHS, SE, |
861 | IgnoreSignificantBits); |
862 | if (!Start) return nullptr; |
863 | // FlagNW is independent of the start value, step direction, and is |
864 | // preserved with smaller magnitude steps. |
865 | // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) |
866 | return SE.getAddRecExpr(Start, Step, L: AR->getLoop(), Flags: SCEV::FlagAnyWrap); |
867 | } |
868 | return nullptr; |
869 | } |
870 | |
871 | // Distribute the sdiv over add operands, if the add doesn't overflow. |
872 | if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Val: LHS)) { |
873 | if (IgnoreSignificantBits || isAddSExtable(A: Add, SE)) { |
874 | SmallVector<const SCEV *, 8> Ops; |
875 | for (const SCEV *S : Add->operands()) { |
876 | const SCEV *Op = getExactSDiv(LHS: S, RHS, SE, IgnoreSignificantBits); |
877 | if (!Op) return nullptr; |
878 | Ops.push_back(Elt: Op); |
879 | } |
880 | return SE.getAddExpr(Ops); |
881 | } |
882 | return nullptr; |
883 | } |
884 | |
885 | // Check for a multiply operand that we can pull RHS out of. |
886 | if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Val: LHS)) { |
887 | if (IgnoreSignificantBits || isMulSExtable(M: Mul, SE)) { |
888 | // Handle special case C1*X*Y /s C2*X*Y. |
889 | if (const SCEVMulExpr *MulRHS = dyn_cast<SCEVMulExpr>(Val: RHS)) { |
890 | if (IgnoreSignificantBits || isMulSExtable(M: MulRHS, SE)) { |
891 | const SCEVConstant *LC = dyn_cast<SCEVConstant>(Val: Mul->getOperand(i: 0)); |
892 | const SCEVConstant *RC = |
893 | dyn_cast<SCEVConstant>(Val: MulRHS->getOperand(i: 0)); |
894 | if (LC && RC) { |
895 | SmallVector<const SCEV *, 4> LOps(drop_begin(RangeOrContainer: Mul->operands())); |
896 | SmallVector<const SCEV *, 4> ROps(drop_begin(RangeOrContainer: MulRHS->operands())); |
897 | if (LOps == ROps) |
898 | return getExactSDiv(LHS: LC, RHS: RC, SE, IgnoreSignificantBits); |
899 | } |
900 | } |
901 | } |
902 | |
903 | SmallVector<const SCEV *, 4> Ops; |
904 | bool Found = false; |
905 | for (const SCEV *S : Mul->operands()) { |
906 | if (!Found) |
907 | if (const SCEV *Q = getExactSDiv(LHS: S, RHS, SE, |
908 | IgnoreSignificantBits)) { |
909 | S = Q; |
910 | Found = true; |
911 | } |
912 | Ops.push_back(Elt: S); |
913 | } |
914 | return Found ? SE.getMulExpr(Ops) : nullptr; |
915 | } |
916 | return nullptr; |
917 | } |
918 | |
919 | // Otherwise we don't know. |
920 | return nullptr; |
921 | } |
922 | |
923 | /// If S involves the addition of a constant integer value, return that integer |
924 | /// value, and mutate S to point to a new SCEV with that value excluded. |
925 | static Immediate (const SCEV *&S, ScalarEvolution &SE) { |
926 | const APInt *C; |
927 | if (match(S, P: m_scev_APInt(C))) { |
928 | if (C->getSignificantBits() <= 64) { |
929 | S = SE.getConstant(Ty: S->getType(), V: 0); |
930 | return Immediate::getFixed(MinVal: C->getSExtValue()); |
931 | } |
932 | } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Val: S)) { |
933 | SmallVector<const SCEV *, 8> NewOps(Add->operands()); |
934 | Immediate Result = ExtractImmediate(S&: NewOps.front(), SE); |
935 | if (Result.isNonZero()) |
936 | S = SE.getAddExpr(Ops&: NewOps); |
937 | return Result; |
938 | } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: S)) { |
939 | SmallVector<const SCEV *, 8> NewOps(AR->operands()); |
940 | Immediate Result = ExtractImmediate(S&: NewOps.front(), SE); |
941 | if (Result.isNonZero()) |
942 | S = SE.getAddRecExpr(Operands&: NewOps, L: AR->getLoop(), |
943 | // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) |
944 | Flags: SCEV::FlagAnyWrap); |
945 | return Result; |
946 | } else if (EnableVScaleImmediates && |
947 | match(S, P: m_scev_Mul(Op0: m_scev_APInt(C), Op1: m_SCEVVScale()))) { |
948 | S = SE.getConstant(Ty: S->getType(), V: 0); |
949 | return Immediate::getScalable(MinVal: C->getSExtValue()); |
950 | } |
951 | return Immediate::getZero(); |
952 | } |
953 | |
954 | /// If S involves the addition of a GlobalValue address, return that symbol, and |
955 | /// mutate S to point to a new SCEV with that value excluded. |
956 | static GlobalValue *(const SCEV *&S, ScalarEvolution &SE) { |
957 | if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Val: S)) { |
958 | if (GlobalValue *GV = dyn_cast<GlobalValue>(Val: U->getValue())) { |
959 | S = SE.getConstant(Ty: GV->getType(), V: 0); |
960 | return GV; |
961 | } |
962 | } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Val: S)) { |
963 | SmallVector<const SCEV *, 8> NewOps(Add->operands()); |
964 | GlobalValue *Result = ExtractSymbol(S&: NewOps.back(), SE); |
965 | if (Result) |
966 | S = SE.getAddExpr(Ops&: NewOps); |
967 | return Result; |
968 | } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: S)) { |
969 | SmallVector<const SCEV *, 8> NewOps(AR->operands()); |
970 | GlobalValue *Result = ExtractSymbol(S&: NewOps.front(), SE); |
971 | if (Result) |
972 | S = SE.getAddRecExpr(Operands&: NewOps, L: AR->getLoop(), |
973 | // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) |
974 | Flags: SCEV::FlagAnyWrap); |
975 | return Result; |
976 | } |
977 | return nullptr; |
978 | } |
979 | |
980 | /// Returns true if the specified instruction is using the specified value as an |
981 | /// address. |
982 | static bool isAddressUse(const TargetTransformInfo &TTI, |
983 | Instruction *Inst, Value *OperandVal) { |
984 | bool isAddress = isa<LoadInst>(Val: Inst); |
985 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: Inst)) { |
986 | if (SI->getPointerOperand() == OperandVal) |
987 | isAddress = true; |
988 | } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: Inst)) { |
989 | // Addressing modes can also be folded into prefetches and a variety |
990 | // of intrinsics. |
991 | switch (II->getIntrinsicID()) { |
992 | case Intrinsic::memset: |
993 | case Intrinsic::prefetch: |
994 | case Intrinsic::masked_load: |
995 | if (II->getArgOperand(i: 0) == OperandVal) |
996 | isAddress = true; |
997 | break; |
998 | case Intrinsic::masked_store: |
999 | if (II->getArgOperand(i: 1) == OperandVal) |
1000 | isAddress = true; |
1001 | break; |
1002 | case Intrinsic::memmove: |
1003 | case Intrinsic::memcpy: |
1004 | if (II->getArgOperand(i: 0) == OperandVal || |
1005 | II->getArgOperand(i: 1) == OperandVal) |
1006 | isAddress = true; |
1007 | break; |
1008 | default: { |
1009 | MemIntrinsicInfo IntrInfo; |
1010 | if (TTI.getTgtMemIntrinsic(Inst: II, Info&: IntrInfo)) { |
1011 | if (IntrInfo.PtrVal == OperandVal) |
1012 | isAddress = true; |
1013 | } |
1014 | } |
1015 | } |
1016 | } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: Inst)) { |
1017 | if (RMW->getPointerOperand() == OperandVal) |
1018 | isAddress = true; |
1019 | } else if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: Inst)) { |
1020 | if (CmpX->getPointerOperand() == OperandVal) |
1021 | isAddress = true; |
1022 | } |
1023 | return isAddress; |
1024 | } |
1025 | |
1026 | /// Return the type of the memory being accessed. |
1027 | static MemAccessTy getAccessType(const TargetTransformInfo &TTI, |
1028 | Instruction *Inst, Value *OperandVal) { |
1029 | MemAccessTy AccessTy = MemAccessTy::getUnknown(Ctx&: Inst->getContext()); |
1030 | |
1031 | // First get the type of memory being accessed. |
1032 | if (Type *Ty = Inst->getAccessType()) |
1033 | AccessTy.MemTy = Ty; |
1034 | |
1035 | // Then get the pointer address space. |
1036 | if (const StoreInst *SI = dyn_cast<StoreInst>(Val: Inst)) { |
1037 | AccessTy.AddrSpace = SI->getPointerAddressSpace(); |
1038 | } else if (const LoadInst *LI = dyn_cast<LoadInst>(Val: Inst)) { |
1039 | AccessTy.AddrSpace = LI->getPointerAddressSpace(); |
1040 | } else if (const AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: Inst)) { |
1041 | AccessTy.AddrSpace = RMW->getPointerAddressSpace(); |
1042 | } else if (const AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: Inst)) { |
1043 | AccessTy.AddrSpace = CmpX->getPointerAddressSpace(); |
1044 | } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: Inst)) { |
1045 | switch (II->getIntrinsicID()) { |
1046 | case Intrinsic::prefetch: |
1047 | case Intrinsic::memset: |
1048 | AccessTy.AddrSpace = II->getArgOperand(i: 0)->getType()->getPointerAddressSpace(); |
1049 | AccessTy.MemTy = OperandVal->getType(); |
1050 | break; |
1051 | case Intrinsic::memmove: |
1052 | case Intrinsic::memcpy: |
1053 | AccessTy.AddrSpace = OperandVal->getType()->getPointerAddressSpace(); |
1054 | AccessTy.MemTy = OperandVal->getType(); |
1055 | break; |
1056 | case Intrinsic::masked_load: |
1057 | AccessTy.AddrSpace = |
1058 | II->getArgOperand(i: 0)->getType()->getPointerAddressSpace(); |
1059 | break; |
1060 | case Intrinsic::masked_store: |
1061 | AccessTy.AddrSpace = |
1062 | II->getArgOperand(i: 1)->getType()->getPointerAddressSpace(); |
1063 | break; |
1064 | default: { |
1065 | MemIntrinsicInfo IntrInfo; |
1066 | if (TTI.getTgtMemIntrinsic(Inst: II, Info&: IntrInfo) && IntrInfo.PtrVal) { |
1067 | AccessTy.AddrSpace |
1068 | = IntrInfo.PtrVal->getType()->getPointerAddressSpace(); |
1069 | } |
1070 | |
1071 | break; |
1072 | } |
1073 | } |
1074 | } |
1075 | |
1076 | return AccessTy; |
1077 | } |
1078 | |
1079 | /// Return true if this AddRec is already a phi in its loop. |
1080 | static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { |
1081 | for (PHINode &PN : AR->getLoop()->getHeader()->phis()) { |
1082 | if (SE.isSCEVable(Ty: PN.getType()) && |
1083 | (SE.getEffectiveSCEVType(Ty: PN.getType()) == |
1084 | SE.getEffectiveSCEVType(Ty: AR->getType())) && |
1085 | SE.getSCEV(V: &PN) == AR) |
1086 | return true; |
1087 | } |
1088 | return false; |
1089 | } |
1090 | |
1091 | /// Check if expanding this expression is likely to incur significant cost. This |
1092 | /// is tricky because SCEV doesn't track which expressions are actually computed |
1093 | /// by the current IR. |
1094 | /// |
1095 | /// We currently allow expansion of IV increments that involve adds, |
1096 | /// multiplication by constants, and AddRecs from existing phis. |
1097 | /// |
1098 | /// TODO: Allow UDivExpr if we can find an existing IV increment that is an |
1099 | /// obvious multiple of the UDivExpr. |
1100 | static bool isHighCostExpansion(const SCEV *S, |
1101 | SmallPtrSetImpl<const SCEV*> &Processed, |
1102 | ScalarEvolution &SE) { |
1103 | // Zero/One operand expressions |
1104 | switch (S->getSCEVType()) { |
1105 | case scUnknown: |
1106 | case scConstant: |
1107 | case scVScale: |
1108 | return false; |
1109 | case scTruncate: |
1110 | return isHighCostExpansion(S: cast<SCEVTruncateExpr>(Val: S)->getOperand(), |
1111 | Processed, SE); |
1112 | case scZeroExtend: |
1113 | return isHighCostExpansion(S: cast<SCEVZeroExtendExpr>(Val: S)->getOperand(), |
1114 | Processed, SE); |
1115 | case scSignExtend: |
1116 | return isHighCostExpansion(S: cast<SCEVSignExtendExpr>(Val: S)->getOperand(), |
1117 | Processed, SE); |
1118 | default: |
1119 | break; |
1120 | } |
1121 | |
1122 | if (!Processed.insert(Ptr: S).second) |
1123 | return false; |
1124 | |
1125 | if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Val: S)) { |
1126 | for (const SCEV *S : Add->operands()) { |
1127 | if (isHighCostExpansion(S, Processed, SE)) |
1128 | return true; |
1129 | } |
1130 | return false; |
1131 | } |
1132 | |
1133 | const SCEV *Op0, *Op1; |
1134 | if (match(S, P: m_scev_Mul(Op0: m_SCEV(V&: Op0), Op1: m_SCEV(V&: Op1)))) { |
1135 | // Multiplication by a constant is ok |
1136 | if (isa<SCEVConstant>(Val: Op0)) |
1137 | return isHighCostExpansion(S: Op1, Processed, SE); |
1138 | |
1139 | // If we have the value of one operand, check if an existing |
1140 | // multiplication already generates this expression. |
1141 | if (const auto *U = dyn_cast<SCEVUnknown>(Val: Op1)) { |
1142 | Value *UVal = U->getValue(); |
1143 | for (User *UR : UVal->users()) { |
1144 | // If U is a constant, it may be used by a ConstantExpr. |
1145 | Instruction *UI = dyn_cast<Instruction>(Val: UR); |
1146 | if (UI && UI->getOpcode() == Instruction::Mul && |
1147 | SE.isSCEVable(Ty: UI->getType())) { |
1148 | return SE.getSCEV(V: UI) == S; |
1149 | } |
1150 | } |
1151 | } |
1152 | } |
1153 | |
1154 | if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: S)) { |
1155 | if (isExistingPhi(AR, SE)) |
1156 | return false; |
1157 | } |
1158 | |
1159 | // Fow now, consider any other type of expression (div/mul/min/max) high cost. |
1160 | return true; |
1161 | } |
1162 | |
1163 | namespace { |
1164 | |
1165 | class LSRUse; |
1166 | |
1167 | } // end anonymous namespace |
1168 | |
1169 | /// Check if the addressing mode defined by \p F is completely |
1170 | /// folded in \p LU at isel time. |
1171 | /// This includes address-mode folding and special icmp tricks. |
1172 | /// This function returns true if \p LU can accommodate what \p F |
1173 | /// defines and up to 1 base + 1 scaled + offset. |
1174 | /// In other words, if \p F has several base registers, this function may |
1175 | /// still return true. Therefore, users still need to account for |
1176 | /// additional base registers and/or unfolded offsets to derive an |
1177 | /// accurate cost model. |
1178 | static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, |
1179 | const LSRUse &LU, const Formula &F); |
1180 | |
1181 | // Get the cost of the scaling factor used in F for LU. |
1182 | static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI, |
1183 | const LSRUse &LU, const Formula &F, |
1184 | const Loop &L); |
1185 | |
1186 | namespace { |
1187 | |
1188 | /// This class is used to measure and compare candidate formulae. |
1189 | class Cost { |
1190 | const Loop *L = nullptr; |
1191 | ScalarEvolution *SE = nullptr; |
1192 | const TargetTransformInfo *TTI = nullptr; |
1193 | TargetTransformInfo::LSRCost C; |
1194 | TTI::AddressingModeKind AMK = TTI::AMK_None; |
1195 | |
1196 | public: |
1197 | Cost() = delete; |
1198 | Cost(const Loop *L, ScalarEvolution &SE, const TargetTransformInfo &TTI, |
1199 | TTI::AddressingModeKind AMK) : |
1200 | L(L), SE(&SE), TTI(&TTI), AMK(AMK) { |
1201 | C.Insns = 0; |
1202 | C.NumRegs = 0; |
1203 | C.AddRecCost = 0; |
1204 | C.NumIVMuls = 0; |
1205 | C.NumBaseAdds = 0; |
1206 | C.ImmCost = 0; |
1207 | C.SetupCost = 0; |
1208 | C.ScaleCost = 0; |
1209 | } |
1210 | |
1211 | bool isLess(const Cost &Other) const; |
1212 | |
1213 | void Lose(); |
1214 | |
1215 | #ifndef NDEBUG |
1216 | // Once any of the metrics loses, they must all remain losers. |
1217 | bool isValid() { |
1218 | return ((C.Insns | C.NumRegs | C.AddRecCost | C.NumIVMuls | C.NumBaseAdds |
1219 | | C.ImmCost | C.SetupCost | C.ScaleCost) != ~0u) |
1220 | || ((C.Insns & C.NumRegs & C.AddRecCost & C.NumIVMuls & C.NumBaseAdds |
1221 | & C.ImmCost & C.SetupCost & C.ScaleCost) == ~0u); |
1222 | } |
1223 | #endif |
1224 | |
1225 | bool isLoser() { |
1226 | assert(isValid() && "invalid cost" ); |
1227 | return C.NumRegs == ~0u; |
1228 | } |
1229 | |
1230 | void RateFormula(const Formula &F, |
1231 | SmallPtrSetImpl<const SCEV *> &Regs, |
1232 | const DenseSet<const SCEV *> &VisitedRegs, |
1233 | const LSRUse &LU, |
1234 | SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr); |
1235 | |
1236 | void print(raw_ostream &OS) const; |
1237 | void dump() const; |
1238 | |
1239 | private: |
1240 | void RateRegister(const Formula &F, const SCEV *Reg, |
1241 | SmallPtrSetImpl<const SCEV *> &Regs); |
1242 | void RatePrimaryRegister(const Formula &F, const SCEV *Reg, |
1243 | SmallPtrSetImpl<const SCEV *> &Regs, |
1244 | SmallPtrSetImpl<const SCEV *> *LoserRegs); |
1245 | }; |
1246 | |
1247 | /// An operand value in an instruction which is to be replaced with some |
1248 | /// equivalent, possibly strength-reduced, replacement. |
1249 | struct LSRFixup { |
1250 | /// The instruction which will be updated. |
1251 | Instruction *UserInst = nullptr; |
1252 | |
1253 | /// The operand of the instruction which will be replaced. The operand may be |
1254 | /// used more than once; every instance will be replaced. |
1255 | Value *OperandValToReplace = nullptr; |
1256 | |
1257 | /// If this user is to use the post-incremented value of an induction |
1258 | /// variable, this set is non-empty and holds the loops associated with the |
1259 | /// induction variable. |
1260 | PostIncLoopSet PostIncLoops; |
1261 | |
1262 | /// A constant offset to be added to the LSRUse expression. This allows |
1263 | /// multiple fixups to share the same LSRUse with different offsets, for |
1264 | /// example in an unrolled loop. |
1265 | Immediate Offset = Immediate::getZero(); |
1266 | |
1267 | LSRFixup() = default; |
1268 | |
1269 | bool isUseFullyOutsideLoop(const Loop *L) const; |
1270 | |
1271 | void print(raw_ostream &OS) const; |
1272 | void dump() const; |
1273 | }; |
1274 | |
1275 | /// This class holds the state that LSR keeps for each use in IVUsers, as well |
1276 | /// as uses invented by LSR itself. It includes information about what kinds of |
1277 | /// things can be folded into the user, information about the user itself, and |
1278 | /// information about how the use may be satisfied. TODO: Represent multiple |
1279 | /// users of the same expression in common? |
1280 | class LSRUse { |
1281 | DenseSet<SmallVector<const SCEV *, 4>> Uniquifier; |
1282 | |
1283 | public: |
1284 | /// An enum for a kind of use, indicating what types of scaled and immediate |
1285 | /// operands it might support. |
1286 | enum KindType { |
1287 | Basic, ///< A normal use, with no folding. |
1288 | Special, ///< A special case of basic, allowing -1 scales. |
1289 | Address, ///< An address use; folding according to TargetLowering |
1290 | ICmpZero ///< An equality icmp with both operands folded into one. |
1291 | // TODO: Add a generic icmp too? |
1292 | }; |
1293 | |
1294 | using SCEVUseKindPair = PointerIntPair<const SCEV *, 2, KindType>; |
1295 | |
1296 | KindType Kind; |
1297 | MemAccessTy AccessTy; |
1298 | |
1299 | /// The list of operands which are to be replaced. |
1300 | SmallVector<LSRFixup, 8> Fixups; |
1301 | |
1302 | /// Keep track of the min and max offsets of the fixups. |
1303 | Immediate MinOffset = Immediate::getFixedMax(); |
1304 | Immediate MaxOffset = Immediate::getFixedMin(); |
1305 | |
1306 | /// This records whether all of the fixups using this LSRUse are outside of |
1307 | /// the loop, in which case some special-case heuristics may be used. |
1308 | bool AllFixupsOutsideLoop = true; |
1309 | |
1310 | /// RigidFormula is set to true to guarantee that this use will be associated |
1311 | /// with a single formula--the one that initially matched. Some SCEV |
1312 | /// expressions cannot be expanded. This allows LSR to consider the registers |
1313 | /// used by those expressions without the need to expand them later after |
1314 | /// changing the formula. |
1315 | bool RigidFormula = false; |
1316 | |
1317 | /// This records the widest use type for any fixup using this |
1318 | /// LSRUse. FindUseWithSimilarFormula can't consider uses with different max |
1319 | /// fixup widths to be equivalent, because the narrower one may be relying on |
1320 | /// the implicit truncation to truncate away bogus bits. |
1321 | Type *WidestFixupType = nullptr; |
1322 | |
1323 | /// A list of ways to build a value that can satisfy this user. After the |
1324 | /// list is populated, one of these is selected heuristically and used to |
1325 | /// formulate a replacement for OperandValToReplace in UserInst. |
1326 | SmallVector<Formula, 12> Formulae; |
1327 | |
1328 | /// The set of register candidates used by all formulae in this LSRUse. |
1329 | SmallPtrSet<const SCEV *, 4> Regs; |
1330 | |
1331 | LSRUse(KindType K, MemAccessTy AT) : Kind(K), AccessTy(AT) {} |
1332 | |
1333 | LSRFixup &getNewFixup() { |
1334 | Fixups.push_back(Elt: LSRFixup()); |
1335 | return Fixups.back(); |
1336 | } |
1337 | |
1338 | void pushFixup(LSRFixup &f) { |
1339 | Fixups.push_back(Elt: f); |
1340 | if (Immediate::isKnownGT(LHS: f.Offset, RHS: MaxOffset)) |
1341 | MaxOffset = f.Offset; |
1342 | if (Immediate::isKnownLT(LHS: f.Offset, RHS: MinOffset)) |
1343 | MinOffset = f.Offset; |
1344 | } |
1345 | |
1346 | bool HasFormulaWithSameRegs(const Formula &F) const; |
1347 | float getNotSelectedProbability(const SCEV *Reg) const; |
1348 | bool InsertFormula(const Formula &F, const Loop &L); |
1349 | void DeleteFormula(Formula &F); |
1350 | void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses); |
1351 | |
1352 | void print(raw_ostream &OS) const; |
1353 | void dump() const; |
1354 | }; |
1355 | |
1356 | } // end anonymous namespace |
1357 | |
1358 | static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, |
1359 | LSRUse::KindType Kind, MemAccessTy AccessTy, |
1360 | GlobalValue *BaseGV, Immediate BaseOffset, |
1361 | bool HasBaseReg, int64_t Scale, |
1362 | Instruction *Fixup = nullptr); |
1363 | |
1364 | static unsigned getSetupCost(const SCEV *Reg, unsigned Depth) { |
1365 | if (isa<SCEVUnknown>(Val: Reg) || isa<SCEVConstant>(Val: Reg)) |
1366 | return 1; |
1367 | if (Depth == 0) |
1368 | return 0; |
1369 | if (const auto *S = dyn_cast<SCEVAddRecExpr>(Val: Reg)) |
1370 | return getSetupCost(Reg: S->getStart(), Depth: Depth - 1); |
1371 | if (auto S = dyn_cast<SCEVIntegralCastExpr>(Val: Reg)) |
1372 | return getSetupCost(Reg: S->getOperand(), Depth: Depth - 1); |
1373 | if (auto S = dyn_cast<SCEVNAryExpr>(Val: Reg)) |
1374 | return std::accumulate(first: S->operands().begin(), last: S->operands().end(), init: 0, |
1375 | binary_op: [&](unsigned i, const SCEV *Reg) { |
1376 | return i + getSetupCost(Reg, Depth: Depth - 1); |
1377 | }); |
1378 | if (auto S = dyn_cast<SCEVUDivExpr>(Val: Reg)) |
1379 | return getSetupCost(Reg: S->getLHS(), Depth: Depth - 1) + |
1380 | getSetupCost(Reg: S->getRHS(), Depth: Depth - 1); |
1381 | return 0; |
1382 | } |
1383 | |
1384 | /// Tally up interesting quantities from the given register. |
1385 | void Cost::RateRegister(const Formula &F, const SCEV *Reg, |
1386 | SmallPtrSetImpl<const SCEV *> &Regs) { |
1387 | if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: Reg)) { |
1388 | // If this is an addrec for another loop, it should be an invariant |
1389 | // with respect to L since L is the innermost loop (at least |
1390 | // for now LSR only handles innermost loops). |
1391 | if (AR->getLoop() != L) { |
1392 | // If the AddRec exists, consider it's register free and leave it alone. |
1393 | if (isExistingPhi(AR, SE&: *SE) && AMK != TTI::AMK_PostIndexed) |
1394 | return; |
1395 | |
1396 | // It is bad to allow LSR for current loop to add induction variables |
1397 | // for its sibling loops. |
1398 | if (!AR->getLoop()->contains(L)) { |
1399 | Lose(); |
1400 | return; |
1401 | } |
1402 | |
1403 | // Otherwise, it will be an invariant with respect to Loop L. |
1404 | ++C.NumRegs; |
1405 | return; |
1406 | } |
1407 | |
1408 | unsigned LoopCost = 1; |
1409 | if (TTI->isIndexedLoadLegal(Mode: TTI->MIM_PostInc, Ty: AR->getType()) || |
1410 | TTI->isIndexedStoreLegal(Mode: TTI->MIM_PostInc, Ty: AR->getType())) { |
1411 | const SCEV *Start; |
1412 | const SCEVConstant *Step; |
1413 | if (match(S: AR, P: m_scev_AffineAddRec(Op0: m_SCEV(V&: Start), Op1: m_SCEVConstant(V&: Step)))) |
1414 | // If the step size matches the base offset, we could use pre-indexed |
1415 | // addressing. |
1416 | if ((AMK == TTI::AMK_PreIndexed && F.BaseOffset.isFixed() && |
1417 | Step->getAPInt() == F.BaseOffset.getFixedValue()) || |
1418 | (AMK == TTI::AMK_PostIndexed && !isa<SCEVConstant>(Val: Start) && |
1419 | SE->isLoopInvariant(S: Start, L))) |
1420 | LoopCost = 0; |
1421 | } |
1422 | C.AddRecCost += LoopCost; |
1423 | |
1424 | // Add the step value register, if it needs one. |
1425 | // TODO: The non-affine case isn't precisely modeled here. |
1426 | if (!AR->isAffine() || !isa<SCEVConstant>(Val: AR->getOperand(i: 1))) { |
1427 | if (!Regs.count(Ptr: AR->getOperand(i: 1))) { |
1428 | RateRegister(F, Reg: AR->getOperand(i: 1), Regs); |
1429 | if (isLoser()) |
1430 | return; |
1431 | } |
1432 | } |
1433 | } |
1434 | ++C.NumRegs; |
1435 | |
1436 | // Rough heuristic; favor registers which don't require extra setup |
1437 | // instructions in the preheader. |
1438 | C.SetupCost += getSetupCost(Reg, Depth: SetupCostDepthLimit); |
1439 | // Ensure we don't, even with the recusion limit, produce invalid costs. |
1440 | C.SetupCost = std::min<unsigned>(a: C.SetupCost, b: 1 << 16); |
1441 | |
1442 | C.NumIVMuls += isa<SCEVMulExpr>(Val: Reg) && |
1443 | SE->hasComputableLoopEvolution(S: Reg, L); |
1444 | } |
1445 | |
1446 | /// Record this register in the set. If we haven't seen it before, rate |
1447 | /// it. Optional LoserRegs provides a way to declare any formula that refers to |
1448 | /// one of those regs an instant loser. |
1449 | void Cost::RatePrimaryRegister(const Formula &F, const SCEV *Reg, |
1450 | SmallPtrSetImpl<const SCEV *> &Regs, |
1451 | SmallPtrSetImpl<const SCEV *> *LoserRegs) { |
1452 | if (LoserRegs && LoserRegs->count(Ptr: Reg)) { |
1453 | Lose(); |
1454 | return; |
1455 | } |
1456 | if (Regs.insert(Ptr: Reg).second) { |
1457 | RateRegister(F, Reg, Regs); |
1458 | if (LoserRegs && isLoser()) |
1459 | LoserRegs->insert(Ptr: Reg); |
1460 | } |
1461 | } |
1462 | |
1463 | void Cost::RateFormula(const Formula &F, |
1464 | SmallPtrSetImpl<const SCEV *> &Regs, |
1465 | const DenseSet<const SCEV *> &VisitedRegs, |
1466 | const LSRUse &LU, |
1467 | SmallPtrSetImpl<const SCEV *> *LoserRegs) { |
1468 | if (isLoser()) |
1469 | return; |
1470 | assert(F.isCanonical(*L) && "Cost is accurate only for canonical formula" ); |
1471 | // Tally up the registers. |
1472 | unsigned PrevAddRecCost = C.AddRecCost; |
1473 | unsigned PrevNumRegs = C.NumRegs; |
1474 | unsigned PrevNumBaseAdds = C.NumBaseAdds; |
1475 | if (const SCEV *ScaledReg = F.ScaledReg) { |
1476 | if (VisitedRegs.count(V: ScaledReg)) { |
1477 | Lose(); |
1478 | return; |
1479 | } |
1480 | RatePrimaryRegister(F, Reg: ScaledReg, Regs, LoserRegs); |
1481 | if (isLoser()) |
1482 | return; |
1483 | } |
1484 | for (const SCEV *BaseReg : F.BaseRegs) { |
1485 | if (VisitedRegs.count(V: BaseReg)) { |
1486 | Lose(); |
1487 | return; |
1488 | } |
1489 | RatePrimaryRegister(F, Reg: BaseReg, Regs, LoserRegs); |
1490 | if (isLoser()) |
1491 | return; |
1492 | } |
1493 | |
1494 | // Determine how many (unfolded) adds we'll need inside the loop. |
1495 | size_t NumBaseParts = F.getNumRegs(); |
1496 | if (NumBaseParts > 1) |
1497 | // Do not count the base and a possible second register if the target |
1498 | // allows to fold 2 registers. |
1499 | C.NumBaseAdds += |
1500 | NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI: *TTI, LU, F))); |
1501 | C.NumBaseAdds += (F.UnfoldedOffset.isNonZero()); |
1502 | |
1503 | // Accumulate non-free scaling amounts. |
1504 | C.ScaleCost += getScalingFactorCost(TTI: *TTI, LU, F, L: *L).getValue(); |
1505 | |
1506 | // Tally up the non-zero immediates. |
1507 | for (const LSRFixup &Fixup : LU.Fixups) { |
1508 | if (Fixup.Offset.isCompatibleImmediate(Imm: F.BaseOffset)) { |
1509 | Immediate Offset = Fixup.Offset.addUnsigned(RHS: F.BaseOffset); |
1510 | if (F.BaseGV) |
1511 | C.ImmCost += 64; // Handle symbolic values conservatively. |
1512 | // TODO: This should probably be the pointer size. |
1513 | else if (Offset.isNonZero()) |
1514 | C.ImmCost += |
1515 | APInt(64, Offset.getKnownMinValue(), true).getSignificantBits(); |
1516 | |
1517 | // Check with target if this offset with this instruction is |
1518 | // specifically not supported. |
1519 | if (LU.Kind == LSRUse::Address && Offset.isNonZero() && |
1520 | !isAMCompletelyFolded(TTI: *TTI, Kind: LSRUse::Address, AccessTy: LU.AccessTy, BaseGV: F.BaseGV, |
1521 | BaseOffset: Offset, HasBaseReg: F.HasBaseReg, Scale: F.Scale, Fixup: Fixup.UserInst)) |
1522 | C.NumBaseAdds++; |
1523 | } else { |
1524 | // Incompatible immediate type, increase cost to avoid using |
1525 | C.ImmCost += 2048; |
1526 | } |
1527 | } |
1528 | |
1529 | // If we don't count instruction cost exit here. |
1530 | if (!InsnsCost) { |
1531 | assert(isValid() && "invalid cost" ); |
1532 | return; |
1533 | } |
1534 | |
1535 | // Treat every new register that exceeds TTI.getNumberOfRegisters() - 1 as |
1536 | // additional instruction (at least fill). |
1537 | // TODO: Need distinguish register class? |
1538 | unsigned TTIRegNum = TTI->getNumberOfRegisters( |
1539 | ClassID: TTI->getRegisterClassForType(Vector: false, Ty: F.getType())) - 1; |
1540 | if (C.NumRegs > TTIRegNum) { |
1541 | // Cost already exceeded TTIRegNum, then only newly added register can add |
1542 | // new instructions. |
1543 | if (PrevNumRegs > TTIRegNum) |
1544 | C.Insns += (C.NumRegs - PrevNumRegs); |
1545 | else |
1546 | C.Insns += (C.NumRegs - TTIRegNum); |
1547 | } |
1548 | |
1549 | // If ICmpZero formula ends with not 0, it could not be replaced by |
1550 | // just add or sub. We'll need to compare final result of AddRec. |
1551 | // That means we'll need an additional instruction. But if the target can |
1552 | // macro-fuse a compare with a branch, don't count this extra instruction. |
1553 | // For -10 + {0, +, 1}: |
1554 | // i = i + 1; |
1555 | // cmp i, 10 |
1556 | // |
1557 | // For {-10, +, 1}: |
1558 | // i = i + 1; |
1559 | if (LU.Kind == LSRUse::ICmpZero && !F.hasZeroEnd() && |
1560 | !TTI->canMacroFuseCmp()) |
1561 | C.Insns++; |
1562 | // Each new AddRec adds 1 instruction to calculation. |
1563 | C.Insns += (C.AddRecCost - PrevAddRecCost); |
1564 | |
1565 | // BaseAdds adds instructions for unfolded registers. |
1566 | if (LU.Kind != LSRUse::ICmpZero) |
1567 | C.Insns += C.NumBaseAdds - PrevNumBaseAdds; |
1568 | assert(isValid() && "invalid cost" ); |
1569 | } |
1570 | |
1571 | /// Set this cost to a losing value. |
1572 | void Cost::Lose() { |
1573 | C.Insns = std::numeric_limits<unsigned>::max(); |
1574 | C.NumRegs = std::numeric_limits<unsigned>::max(); |
1575 | C.AddRecCost = std::numeric_limits<unsigned>::max(); |
1576 | C.NumIVMuls = std::numeric_limits<unsigned>::max(); |
1577 | C.NumBaseAdds = std::numeric_limits<unsigned>::max(); |
1578 | C.ImmCost = std::numeric_limits<unsigned>::max(); |
1579 | C.SetupCost = std::numeric_limits<unsigned>::max(); |
1580 | C.ScaleCost = std::numeric_limits<unsigned>::max(); |
1581 | } |
1582 | |
1583 | /// Choose the lower cost. |
1584 | bool Cost::isLess(const Cost &Other) const { |
1585 | if (InsnsCost.getNumOccurrences() > 0 && InsnsCost && |
1586 | C.Insns != Other.C.Insns) |
1587 | return C.Insns < Other.C.Insns; |
1588 | return TTI->isLSRCostLess(C1: C, C2: Other.C); |
1589 | } |
1590 | |
1591 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1592 | void Cost::print(raw_ostream &OS) const { |
1593 | if (InsnsCost) |
1594 | OS << C.Insns << " instruction" << (C.Insns == 1 ? " " : "s " ); |
1595 | OS << C.NumRegs << " reg" << (C.NumRegs == 1 ? "" : "s" ); |
1596 | if (C.AddRecCost != 0) |
1597 | OS << ", with addrec cost " << C.AddRecCost; |
1598 | if (C.NumIVMuls != 0) |
1599 | OS << ", plus " << C.NumIVMuls << " IV mul" |
1600 | << (C.NumIVMuls == 1 ? "" : "s" ); |
1601 | if (C.NumBaseAdds != 0) |
1602 | OS << ", plus " << C.NumBaseAdds << " base add" |
1603 | << (C.NumBaseAdds == 1 ? "" : "s" ); |
1604 | if (C.ScaleCost != 0) |
1605 | OS << ", plus " << C.ScaleCost << " scale cost" ; |
1606 | if (C.ImmCost != 0) |
1607 | OS << ", plus " << C.ImmCost << " imm cost" ; |
1608 | if (C.SetupCost != 0) |
1609 | OS << ", plus " << C.SetupCost << " setup cost" ; |
1610 | } |
1611 | |
1612 | LLVM_DUMP_METHOD void Cost::dump() const { |
1613 | print(errs()); errs() << '\n'; |
1614 | } |
1615 | #endif |
1616 | |
1617 | /// Test whether this fixup always uses its value outside of the given loop. |
1618 | bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const { |
1619 | // PHI nodes use their value in their incoming blocks. |
1620 | if (const PHINode *PN = dyn_cast<PHINode>(Val: UserInst)) { |
1621 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) |
1622 | if (PN->getIncomingValue(i) == OperandValToReplace && |
1623 | L->contains(BB: PN->getIncomingBlock(i))) |
1624 | return false; |
1625 | return true; |
1626 | } |
1627 | |
1628 | return !L->contains(Inst: UserInst); |
1629 | } |
1630 | |
1631 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1632 | void LSRFixup::print(raw_ostream &OS) const { |
1633 | OS << "UserInst=" ; |
1634 | // Store is common and interesting enough to be worth special-casing. |
1635 | if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) { |
1636 | OS << "store " ; |
1637 | Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false); |
1638 | } else if (UserInst->getType()->isVoidTy()) |
1639 | OS << UserInst->getOpcodeName(); |
1640 | else |
1641 | UserInst->printAsOperand(OS, /*PrintType=*/false); |
1642 | |
1643 | OS << ", OperandValToReplace=" ; |
1644 | OperandValToReplace->printAsOperand(OS, /*PrintType=*/false); |
1645 | |
1646 | for (const Loop *PIL : PostIncLoops) { |
1647 | OS << ", PostIncLoop=" ; |
1648 | PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false); |
1649 | } |
1650 | |
1651 | if (Offset.isNonZero()) |
1652 | OS << ", Offset=" << Offset; |
1653 | } |
1654 | |
1655 | LLVM_DUMP_METHOD void LSRFixup::dump() const { |
1656 | print(errs()); errs() << '\n'; |
1657 | } |
1658 | #endif |
1659 | |
1660 | /// Test whether this use as a formula which has the same registers as the given |
1661 | /// formula. |
1662 | bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const { |
1663 | SmallVector<const SCEV *, 4> Key = F.BaseRegs; |
1664 | if (F.ScaledReg) Key.push_back(Elt: F.ScaledReg); |
1665 | // Unstable sort by host order ok, because this is only used for uniquifying. |
1666 | llvm::sort(C&: Key); |
1667 | return Uniquifier.count(V: Key); |
1668 | } |
1669 | |
1670 | /// The function returns a probability of selecting formula without Reg. |
1671 | float LSRUse::getNotSelectedProbability(const SCEV *Reg) const { |
1672 | unsigned FNum = 0; |
1673 | for (const Formula &F : Formulae) |
1674 | if (F.referencesReg(S: Reg)) |
1675 | FNum++; |
1676 | return ((float)(Formulae.size() - FNum)) / Formulae.size(); |
1677 | } |
1678 | |
1679 | /// If the given formula has not yet been inserted, add it to the list, and |
1680 | /// return true. Return false otherwise. The formula must be in canonical form. |
1681 | bool LSRUse::InsertFormula(const Formula &F, const Loop &L) { |
1682 | assert(F.isCanonical(L) && "Invalid canonical representation" ); |
1683 | |
1684 | if (!Formulae.empty() && RigidFormula) |
1685 | return false; |
1686 | |
1687 | SmallVector<const SCEV *, 4> Key = F.BaseRegs; |
1688 | if (F.ScaledReg) Key.push_back(Elt: F.ScaledReg); |
1689 | // Unstable sort by host order ok, because this is only used for uniquifying. |
1690 | llvm::sort(C&: Key); |
1691 | |
1692 | if (!Uniquifier.insert(V: Key).second) |
1693 | return false; |
1694 | |
1695 | // Using a register to hold the value of 0 is not profitable. |
1696 | assert((!F.ScaledReg || !F.ScaledReg->isZero()) && |
1697 | "Zero allocated in a scaled register!" ); |
1698 | #ifndef NDEBUG |
1699 | for (const SCEV *BaseReg : F.BaseRegs) |
1700 | assert(!BaseReg->isZero() && "Zero allocated in a base register!" ); |
1701 | #endif |
1702 | |
1703 | // Add the formula to the list. |
1704 | Formulae.push_back(Elt: F); |
1705 | |
1706 | // Record registers now being used by this use. |
1707 | Regs.insert_range(R: F.BaseRegs); |
1708 | if (F.ScaledReg) |
1709 | Regs.insert(Ptr: F.ScaledReg); |
1710 | |
1711 | return true; |
1712 | } |
1713 | |
1714 | /// Remove the given formula from this use's list. |
1715 | void LSRUse::DeleteFormula(Formula &F) { |
1716 | if (&F != &Formulae.back()) |
1717 | std::swap(a&: F, b&: Formulae.back()); |
1718 | Formulae.pop_back(); |
1719 | } |
1720 | |
1721 | /// Recompute the Regs field, and update RegUses. |
1722 | void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) { |
1723 | // Now that we've filtered out some formulae, recompute the Regs set. |
1724 | SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs); |
1725 | Regs.clear(); |
1726 | for (const Formula &F : Formulae) { |
1727 | if (F.ScaledReg) Regs.insert(Ptr: F.ScaledReg); |
1728 | Regs.insert_range(R: F.BaseRegs); |
1729 | } |
1730 | |
1731 | // Update the RegTracker. |
1732 | for (const SCEV *S : OldRegs) |
1733 | if (!Regs.count(Ptr: S)) |
1734 | RegUses.dropRegister(Reg: S, LUIdx); |
1735 | } |
1736 | |
1737 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1738 | void LSRUse::print(raw_ostream &OS) const { |
1739 | OS << "LSR Use: Kind=" ; |
1740 | switch (Kind) { |
1741 | case Basic: OS << "Basic" ; break; |
1742 | case Special: OS << "Special" ; break; |
1743 | case ICmpZero: OS << "ICmpZero" ; break; |
1744 | case Address: |
1745 | OS << "Address of " ; |
1746 | if (AccessTy.MemTy->isPointerTy()) |
1747 | OS << "pointer" ; // the full pointer type could be really verbose |
1748 | else { |
1749 | OS << *AccessTy.MemTy; |
1750 | } |
1751 | |
1752 | OS << " in addrspace(" << AccessTy.AddrSpace << ')'; |
1753 | } |
1754 | |
1755 | OS << ", Offsets={" ; |
1756 | bool NeedComma = false; |
1757 | for (const LSRFixup &Fixup : Fixups) { |
1758 | if (NeedComma) OS << ','; |
1759 | OS << Fixup.Offset; |
1760 | NeedComma = true; |
1761 | } |
1762 | OS << '}'; |
1763 | |
1764 | if (AllFixupsOutsideLoop) |
1765 | OS << ", all-fixups-outside-loop" ; |
1766 | |
1767 | if (WidestFixupType) |
1768 | OS << ", widest fixup type: " << *WidestFixupType; |
1769 | } |
1770 | |
1771 | LLVM_DUMP_METHOD void LSRUse::dump() const { |
1772 | print(errs()); errs() << '\n'; |
1773 | } |
1774 | #endif |
1775 | |
1776 | static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, |
1777 | LSRUse::KindType Kind, MemAccessTy AccessTy, |
1778 | GlobalValue *BaseGV, Immediate BaseOffset, |
1779 | bool HasBaseReg, int64_t Scale, |
1780 | Instruction *Fixup /* = nullptr */) { |
1781 | switch (Kind) { |
1782 | case LSRUse::Address: { |
1783 | int64_t FixedOffset = |
1784 | BaseOffset.isScalable() ? 0 : BaseOffset.getFixedValue(); |
1785 | int64_t ScalableOffset = |
1786 | BaseOffset.isScalable() ? BaseOffset.getKnownMinValue() : 0; |
1787 | return TTI.isLegalAddressingMode(Ty: AccessTy.MemTy, BaseGV, BaseOffset: FixedOffset, |
1788 | HasBaseReg, Scale, AddrSpace: AccessTy.AddrSpace, |
1789 | I: Fixup, ScalableOffset); |
1790 | } |
1791 | case LSRUse::ICmpZero: |
1792 | // There's not even a target hook for querying whether it would be legal to |
1793 | // fold a GV into an ICmp. |
1794 | if (BaseGV) |
1795 | return false; |
1796 | |
1797 | // ICmp only has two operands; don't allow more than two non-trivial parts. |
1798 | if (Scale != 0 && HasBaseReg && BaseOffset.isNonZero()) |
1799 | return false; |
1800 | |
1801 | // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by |
1802 | // putting the scaled register in the other operand of the icmp. |
1803 | if (Scale != 0 && Scale != -1) |
1804 | return false; |
1805 | |
1806 | // If we have low-level target information, ask the target if it can fold an |
1807 | // integer immediate on an icmp. |
1808 | if (BaseOffset.isNonZero()) { |
1809 | // We don't have an interface to query whether the target supports |
1810 | // icmpzero against scalable quantities yet. |
1811 | if (BaseOffset.isScalable()) |
1812 | return false; |
1813 | |
1814 | // We have one of: |
1815 | // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset |
1816 | // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset |
1817 | // Offs is the ICmp immediate. |
1818 | if (Scale == 0) |
1819 | // The cast does the right thing with |
1820 | // std::numeric_limits<int64_t>::min(). |
1821 | BaseOffset = BaseOffset.getFixed(MinVal: -(uint64_t)BaseOffset.getFixedValue()); |
1822 | return TTI.isLegalICmpImmediate(Imm: BaseOffset.getFixedValue()); |
1823 | } |
1824 | |
1825 | // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg |
1826 | return true; |
1827 | |
1828 | case LSRUse::Basic: |
1829 | // Only handle single-register values. |
1830 | return !BaseGV && Scale == 0 && BaseOffset.isZero(); |
1831 | |
1832 | case LSRUse::Special: |
1833 | // Special case Basic to handle -1 scales. |
1834 | return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset.isZero(); |
1835 | } |
1836 | |
1837 | llvm_unreachable("Invalid LSRUse Kind!" ); |
1838 | } |
1839 | |
1840 | static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, |
1841 | Immediate MinOffset, Immediate MaxOffset, |
1842 | LSRUse::KindType Kind, MemAccessTy AccessTy, |
1843 | GlobalValue *BaseGV, Immediate BaseOffset, |
1844 | bool HasBaseReg, int64_t Scale) { |
1845 | if (BaseOffset.isNonZero() && |
1846 | (BaseOffset.isScalable() != MinOffset.isScalable() || |
1847 | BaseOffset.isScalable() != MaxOffset.isScalable())) |
1848 | return false; |
1849 | // Check for overflow. |
1850 | int64_t Base = BaseOffset.getKnownMinValue(); |
1851 | int64_t Min = MinOffset.getKnownMinValue(); |
1852 | int64_t Max = MaxOffset.getKnownMinValue(); |
1853 | if (((int64_t)((uint64_t)Base + Min) > Base) != (Min > 0)) |
1854 | return false; |
1855 | MinOffset = Immediate::get(MinVal: (uint64_t)Base + Min, Scalable: MinOffset.isScalable()); |
1856 | if (((int64_t)((uint64_t)Base + Max) > Base) != (Max > 0)) |
1857 | return false; |
1858 | MaxOffset = Immediate::get(MinVal: (uint64_t)Base + Max, Scalable: MaxOffset.isScalable()); |
1859 | |
1860 | return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset: MinOffset, |
1861 | HasBaseReg, Scale) && |
1862 | isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset: MaxOffset, |
1863 | HasBaseReg, Scale); |
1864 | } |
1865 | |
1866 | static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, |
1867 | Immediate MinOffset, Immediate MaxOffset, |
1868 | LSRUse::KindType Kind, MemAccessTy AccessTy, |
1869 | const Formula &F, const Loop &L) { |
1870 | // For the purpose of isAMCompletelyFolded either having a canonical formula |
1871 | // or a scale not equal to zero is correct. |
1872 | // Problems may arise from non canonical formulae having a scale == 0. |
1873 | // Strictly speaking it would best to just rely on canonical formulae. |
1874 | // However, when we generate the scaled formulae, we first check that the |
1875 | // scaling factor is profitable before computing the actual ScaledReg for |
1876 | // compile time sake. |
1877 | assert((F.isCanonical(L) || F.Scale != 0)); |
1878 | return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, |
1879 | BaseGV: F.BaseGV, BaseOffset: F.BaseOffset, HasBaseReg: F.HasBaseReg, Scale: F.Scale); |
1880 | } |
1881 | |
1882 | /// Test whether we know how to expand the current formula. |
1883 | static bool isLegalUse(const TargetTransformInfo &TTI, Immediate MinOffset, |
1884 | Immediate MaxOffset, LSRUse::KindType Kind, |
1885 | MemAccessTy AccessTy, GlobalValue *BaseGV, |
1886 | Immediate BaseOffset, bool HasBaseReg, int64_t Scale) { |
1887 | // We know how to expand completely foldable formulae. |
1888 | return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV, |
1889 | BaseOffset, HasBaseReg, Scale) || |
1890 | // Or formulae that use a base register produced by a sum of base |
1891 | // registers. |
1892 | (Scale == 1 && |
1893 | isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, |
1894 | BaseGV, BaseOffset, HasBaseReg: true, Scale: 0)); |
1895 | } |
1896 | |
1897 | static bool isLegalUse(const TargetTransformInfo &TTI, Immediate MinOffset, |
1898 | Immediate MaxOffset, LSRUse::KindType Kind, |
1899 | MemAccessTy AccessTy, const Formula &F) { |
1900 | return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV: F.BaseGV, |
1901 | BaseOffset: F.BaseOffset, HasBaseReg: F.HasBaseReg, Scale: F.Scale); |
1902 | } |
1903 | |
1904 | static bool isLegalAddImmediate(const TargetTransformInfo &TTI, |
1905 | Immediate Offset) { |
1906 | if (Offset.isScalable()) |
1907 | return TTI.isLegalAddScalableImmediate(Imm: Offset.getKnownMinValue()); |
1908 | |
1909 | return TTI.isLegalAddImmediate(Imm: Offset.getFixedValue()); |
1910 | } |
1911 | |
1912 | static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, |
1913 | const LSRUse &LU, const Formula &F) { |
1914 | // Target may want to look at the user instructions. |
1915 | if (LU.Kind == LSRUse::Address && TTI.LSRWithInstrQueries()) { |
1916 | for (const LSRFixup &Fixup : LU.Fixups) |
1917 | if (!isAMCompletelyFolded(TTI, Kind: LSRUse::Address, AccessTy: LU.AccessTy, BaseGV: F.BaseGV, |
1918 | BaseOffset: (F.BaseOffset + Fixup.Offset), HasBaseReg: F.HasBaseReg, |
1919 | Scale: F.Scale, Fixup: Fixup.UserInst)) |
1920 | return false; |
1921 | return true; |
1922 | } |
1923 | |
1924 | return isAMCompletelyFolded(TTI, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LU.Kind, |
1925 | AccessTy: LU.AccessTy, BaseGV: F.BaseGV, BaseOffset: F.BaseOffset, HasBaseReg: F.HasBaseReg, |
1926 | Scale: F.Scale); |
1927 | } |
1928 | |
1929 | static InstructionCost getScalingFactorCost(const TargetTransformInfo &TTI, |
1930 | const LSRUse &LU, const Formula &F, |
1931 | const Loop &L) { |
1932 | if (!F.Scale) |
1933 | return 0; |
1934 | |
1935 | // If the use is not completely folded in that instruction, we will have to |
1936 | // pay an extra cost only for scale != 1. |
1937 | if (!isAMCompletelyFolded(TTI, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LU.Kind, |
1938 | AccessTy: LU.AccessTy, F, L)) |
1939 | return F.Scale != 1; |
1940 | |
1941 | switch (LU.Kind) { |
1942 | case LSRUse::Address: { |
1943 | // Check the scaling factor cost with both the min and max offsets. |
1944 | int64_t ScalableMin = 0, ScalableMax = 0, FixedMin = 0, FixedMax = 0; |
1945 | if (F.BaseOffset.isScalable()) { |
1946 | ScalableMin = (F.BaseOffset + LU.MinOffset).getKnownMinValue(); |
1947 | ScalableMax = (F.BaseOffset + LU.MaxOffset).getKnownMinValue(); |
1948 | } else { |
1949 | FixedMin = (F.BaseOffset + LU.MinOffset).getFixedValue(); |
1950 | FixedMax = (F.BaseOffset + LU.MaxOffset).getFixedValue(); |
1951 | } |
1952 | InstructionCost ScaleCostMinOffset = TTI.getScalingFactorCost( |
1953 | Ty: LU.AccessTy.MemTy, BaseGV: F.BaseGV, BaseOffset: StackOffset::get(Fixed: FixedMin, Scalable: ScalableMin), |
1954 | HasBaseReg: F.HasBaseReg, Scale: F.Scale, AddrSpace: LU.AccessTy.AddrSpace); |
1955 | InstructionCost ScaleCostMaxOffset = TTI.getScalingFactorCost( |
1956 | Ty: LU.AccessTy.MemTy, BaseGV: F.BaseGV, BaseOffset: StackOffset::get(Fixed: FixedMax, Scalable: ScalableMax), |
1957 | HasBaseReg: F.HasBaseReg, Scale: F.Scale, AddrSpace: LU.AccessTy.AddrSpace); |
1958 | |
1959 | assert(ScaleCostMinOffset.isValid() && ScaleCostMaxOffset.isValid() && |
1960 | "Legal addressing mode has an illegal cost!" ); |
1961 | return std::max(a: ScaleCostMinOffset, b: ScaleCostMaxOffset); |
1962 | } |
1963 | case LSRUse::ICmpZero: |
1964 | case LSRUse::Basic: |
1965 | case LSRUse::Special: |
1966 | // The use is completely folded, i.e., everything is folded into the |
1967 | // instruction. |
1968 | return 0; |
1969 | } |
1970 | |
1971 | llvm_unreachable("Invalid LSRUse Kind!" ); |
1972 | } |
1973 | |
1974 | static bool isAlwaysFoldable(const TargetTransformInfo &TTI, |
1975 | LSRUse::KindType Kind, MemAccessTy AccessTy, |
1976 | GlobalValue *BaseGV, Immediate BaseOffset, |
1977 | bool HasBaseReg) { |
1978 | // Fast-path: zero is always foldable. |
1979 | if (BaseOffset.isZero() && !BaseGV) |
1980 | return true; |
1981 | |
1982 | // Conservatively, create an address with an immediate and a |
1983 | // base and a scale. |
1984 | int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1; |
1985 | |
1986 | // Canonicalize a scale of 1 to a base register if the formula doesn't |
1987 | // already have a base register. |
1988 | if (!HasBaseReg && Scale == 1) { |
1989 | Scale = 0; |
1990 | HasBaseReg = true; |
1991 | } |
1992 | |
1993 | // FIXME: Try with + without a scale? Maybe based on TTI? |
1994 | // I think basereg + scaledreg + immediateoffset isn't a good 'conservative' |
1995 | // default for many architectures, not just AArch64 SVE. More investigation |
1996 | // needed later to determine if this should be used more widely than just |
1997 | // on scalable types. |
1998 | if (HasBaseReg && BaseOffset.isNonZero() && Kind != LSRUse::ICmpZero && |
1999 | AccessTy.MemTy && AccessTy.MemTy->isScalableTy() && DropScaledForVScale) |
2000 | Scale = 0; |
2001 | |
2002 | return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset, |
2003 | HasBaseReg, Scale); |
2004 | } |
2005 | |
2006 | static bool isAlwaysFoldable(const TargetTransformInfo &TTI, |
2007 | ScalarEvolution &SE, Immediate MinOffset, |
2008 | Immediate MaxOffset, LSRUse::KindType Kind, |
2009 | MemAccessTy AccessTy, const SCEV *S, |
2010 | bool HasBaseReg) { |
2011 | // Fast-path: zero is always foldable. |
2012 | if (S->isZero()) return true; |
2013 | |
2014 | // Conservatively, create an address with an immediate and a |
2015 | // base and a scale. |
2016 | Immediate BaseOffset = ExtractImmediate(S, SE); |
2017 | GlobalValue *BaseGV = ExtractSymbol(S, SE); |
2018 | |
2019 | // If there's anything else involved, it's not foldable. |
2020 | if (!S->isZero()) return false; |
2021 | |
2022 | // Fast-path: zero is always foldable. |
2023 | if (BaseOffset.isZero() && !BaseGV) |
2024 | return true; |
2025 | |
2026 | if (BaseOffset.isScalable()) |
2027 | return false; |
2028 | |
2029 | // Conservatively, create an address with an immediate and a |
2030 | // base and a scale. |
2031 | int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1; |
2032 | |
2033 | return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV, |
2034 | BaseOffset, HasBaseReg, Scale); |
2035 | } |
2036 | |
2037 | namespace { |
2038 | |
2039 | /// An individual increment in a Chain of IV increments. Relate an IV user to |
2040 | /// an expression that computes the IV it uses from the IV used by the previous |
2041 | /// link in the Chain. |
2042 | /// |
2043 | /// For the head of a chain, IncExpr holds the absolute SCEV expression for the |
2044 | /// original IVOperand. The head of the chain's IVOperand is only valid during |
2045 | /// chain collection, before LSR replaces IV users. During chain generation, |
2046 | /// IncExpr can be used to find the new IVOperand that computes the same |
2047 | /// expression. |
2048 | struct IVInc { |
2049 | Instruction *UserInst; |
2050 | Value* IVOperand; |
2051 | const SCEV *IncExpr; |
2052 | |
2053 | IVInc(Instruction *U, Value *O, const SCEV *E) |
2054 | : UserInst(U), IVOperand(O), IncExpr(E) {} |
2055 | }; |
2056 | |
2057 | // The list of IV increments in program order. We typically add the head of a |
2058 | // chain without finding subsequent links. |
2059 | struct IVChain { |
2060 | SmallVector<IVInc, 1> Incs; |
2061 | const SCEV *ExprBase = nullptr; |
2062 | |
2063 | IVChain() = default; |
2064 | IVChain(const IVInc &Head, const SCEV *Base) |
2065 | : Incs(1, Head), ExprBase(Base) {} |
2066 | |
2067 | using const_iterator = SmallVectorImpl<IVInc>::const_iterator; |
2068 | |
2069 | // Return the first increment in the chain. |
2070 | const_iterator begin() const { |
2071 | assert(!Incs.empty()); |
2072 | return std::next(x: Incs.begin()); |
2073 | } |
2074 | const_iterator end() const { |
2075 | return Incs.end(); |
2076 | } |
2077 | |
2078 | // Returns true if this chain contains any increments. |
2079 | bool hasIncs() const { return Incs.size() >= 2; } |
2080 | |
2081 | // Add an IVInc to the end of this chain. |
2082 | void add(const IVInc &X) { Incs.push_back(Elt: X); } |
2083 | |
2084 | // Returns the last UserInst in the chain. |
2085 | Instruction *tailUserInst() const { return Incs.back().UserInst; } |
2086 | |
2087 | // Returns true if IncExpr can be profitably added to this chain. |
2088 | bool isProfitableIncrement(const SCEV *OperExpr, |
2089 | const SCEV *IncExpr, |
2090 | ScalarEvolution&); |
2091 | }; |
2092 | |
2093 | /// Helper for CollectChains to track multiple IV increment uses. Distinguish |
2094 | /// between FarUsers that definitely cross IV increments and NearUsers that may |
2095 | /// be used between IV increments. |
2096 | struct ChainUsers { |
2097 | SmallPtrSet<Instruction*, 4> FarUsers; |
2098 | SmallPtrSet<Instruction*, 4> NearUsers; |
2099 | }; |
2100 | |
2101 | /// This class holds state for the main loop strength reduction logic. |
2102 | class LSRInstance { |
2103 | IVUsers &IU; |
2104 | ScalarEvolution &SE; |
2105 | DominatorTree &DT; |
2106 | LoopInfo &LI; |
2107 | AssumptionCache &AC; |
2108 | TargetLibraryInfo &TLI; |
2109 | const TargetTransformInfo &TTI; |
2110 | Loop *const L; |
2111 | MemorySSAUpdater *MSSAU; |
2112 | TTI::AddressingModeKind AMK; |
2113 | mutable SCEVExpander Rewriter; |
2114 | bool Changed = false; |
2115 | |
2116 | /// This is the insert position that the current loop's induction variable |
2117 | /// increment should be placed. In simple loops, this is the latch block's |
2118 | /// terminator. But in more complicated cases, this is a position which will |
2119 | /// dominate all the in-loop post-increment users. |
2120 | Instruction *IVIncInsertPos = nullptr; |
2121 | |
2122 | /// Interesting factors between use strides. |
2123 | /// |
2124 | /// We explicitly use a SetVector which contains a SmallSet, instead of the |
2125 | /// default, a SmallDenseSet, because we need to use the full range of |
2126 | /// int64_ts, and there's currently no good way of doing that with |
2127 | /// SmallDenseSet. |
2128 | SetVector<int64_t, SmallVector<int64_t, 8>, SmallSet<int64_t, 8>> Factors; |
2129 | |
2130 | /// The cost of the current SCEV, the best solution by LSR will be dropped if |
2131 | /// the solution is not profitable. |
2132 | Cost BaselineCost; |
2133 | |
2134 | /// Interesting use types, to facilitate truncation reuse. |
2135 | SmallSetVector<Type *, 4> Types; |
2136 | |
2137 | /// The list of interesting uses. |
2138 | mutable SmallVector<LSRUse, 16> Uses; |
2139 | |
2140 | /// Track which uses use which register candidates. |
2141 | RegUseTracker RegUses; |
2142 | |
2143 | // Limit the number of chains to avoid quadratic behavior. We don't expect to |
2144 | // have more than a few IV increment chains in a loop. Missing a Chain falls |
2145 | // back to normal LSR behavior for those uses. |
2146 | static const unsigned MaxChains = 8; |
2147 | |
2148 | /// IV users can form a chain of IV increments. |
2149 | SmallVector<IVChain, MaxChains> IVChainVec; |
2150 | |
2151 | /// IV users that belong to profitable IVChains. |
2152 | SmallPtrSet<Use*, MaxChains> IVIncSet; |
2153 | |
2154 | /// Induction variables that were generated and inserted by the SCEV Expander. |
2155 | SmallVector<llvm::WeakVH, 2> ScalarEvolutionIVs; |
2156 | |
2157 | // Inserting instructions in the loop and using them as PHI's input could |
2158 | // break LCSSA in case if PHI's parent block is not a loop exit (i.e. the |
2159 | // corresponding incoming block is not loop exiting). So collect all such |
2160 | // instructions to form LCSSA for them later. |
2161 | SmallSetVector<Instruction *, 4> InsertedNonLCSSAInsts; |
2162 | |
2163 | void OptimizeShadowIV(); |
2164 | bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse); |
2165 | ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse); |
2166 | void OptimizeLoopTermCond(); |
2167 | |
2168 | void ChainInstruction(Instruction *UserInst, Instruction *IVOper, |
2169 | SmallVectorImpl<ChainUsers> &ChainUsersVec); |
2170 | void FinalizeChain(IVChain &Chain); |
2171 | void CollectChains(); |
2172 | void GenerateIVChain(const IVChain &Chain, |
2173 | SmallVectorImpl<WeakTrackingVH> &DeadInsts); |
2174 | |
2175 | void CollectInterestingTypesAndFactors(); |
2176 | void CollectFixupsAndInitialFormulae(); |
2177 | |
2178 | // Support for sharing of LSRUses between LSRFixups. |
2179 | using UseMapTy = DenseMap<LSRUse::SCEVUseKindPair, size_t>; |
2180 | UseMapTy UseMap; |
2181 | |
2182 | bool reconcileNewOffset(LSRUse &LU, Immediate NewOffset, bool HasBaseReg, |
2183 | LSRUse::KindType Kind, MemAccessTy AccessTy); |
2184 | |
2185 | std::pair<size_t, Immediate> getUse(const SCEV *&Expr, LSRUse::KindType Kind, |
2186 | MemAccessTy AccessTy); |
2187 | |
2188 | void DeleteUse(LSRUse &LU, size_t LUIdx); |
2189 | |
2190 | LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU); |
2191 | |
2192 | void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); |
2193 | void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); |
2194 | void CountRegisters(const Formula &F, size_t LUIdx); |
2195 | bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F); |
2196 | |
2197 | void CollectLoopInvariantFixupsAndFormulae(); |
2198 | |
2199 | void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base, |
2200 | unsigned Depth = 0); |
2201 | |
2202 | void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx, |
2203 | const Formula &Base, unsigned Depth, |
2204 | size_t Idx, bool IsScaledReg = false); |
2205 | void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base); |
2206 | void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx, |
2207 | const Formula &Base, size_t Idx, |
2208 | bool IsScaledReg = false); |
2209 | void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); |
2210 | void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx, |
2211 | const Formula &Base, |
2212 | const SmallVectorImpl<Immediate> &Worklist, |
2213 | size_t Idx, bool IsScaledReg = false); |
2214 | void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); |
2215 | void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base); |
2216 | void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base); |
2217 | void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base); |
2218 | void GenerateCrossUseConstantOffsets(); |
2219 | void GenerateAllReuseFormulae(); |
2220 | |
2221 | void FilterOutUndesirableDedicatedRegisters(); |
2222 | |
2223 | size_t EstimateSearchSpaceComplexity() const; |
2224 | void NarrowSearchSpaceByDetectingSupersets(); |
2225 | void NarrowSearchSpaceByCollapsingUnrolledCode(); |
2226 | void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); |
2227 | void NarrowSearchSpaceByFilterFormulaWithSameScaledReg(); |
2228 | void NarrowSearchSpaceByFilterPostInc(); |
2229 | void NarrowSearchSpaceByDeletingCostlyFormulas(); |
2230 | void NarrowSearchSpaceByPickingWinnerRegs(); |
2231 | void NarrowSearchSpaceUsingHeuristics(); |
2232 | |
2233 | void SolveRecurse(SmallVectorImpl<const Formula *> &Solution, |
2234 | Cost &SolutionCost, |
2235 | SmallVectorImpl<const Formula *> &Workspace, |
2236 | const Cost &CurCost, |
2237 | const SmallPtrSet<const SCEV *, 16> &CurRegs, |
2238 | DenseSet<const SCEV *> &VisitedRegs) const; |
2239 | void Solve(SmallVectorImpl<const Formula *> &Solution) const; |
2240 | |
2241 | BasicBlock::iterator |
2242 | HoistInsertPosition(BasicBlock::iterator IP, |
2243 | const SmallVectorImpl<Instruction *> &Inputs) const; |
2244 | BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP, |
2245 | const LSRFixup &LF, |
2246 | const LSRUse &LU) const; |
2247 | |
2248 | Value *Expand(const LSRUse &LU, const LSRFixup &LF, const Formula &F, |
2249 | BasicBlock::iterator IP, |
2250 | SmallVectorImpl<WeakTrackingVH> &DeadInsts) const; |
2251 | void RewriteForPHI(PHINode *PN, const LSRUse &LU, const LSRFixup &LF, |
2252 | const Formula &F, |
2253 | SmallVectorImpl<WeakTrackingVH> &DeadInsts); |
2254 | void Rewrite(const LSRUse &LU, const LSRFixup &LF, const Formula &F, |
2255 | SmallVectorImpl<WeakTrackingVH> &DeadInsts); |
2256 | void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution); |
2257 | |
2258 | public: |
2259 | LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, DominatorTree &DT, |
2260 | LoopInfo &LI, const TargetTransformInfo &TTI, AssumptionCache &AC, |
2261 | TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU); |
2262 | |
2263 | bool getChanged() const { return Changed; } |
2264 | const SmallVectorImpl<WeakVH> &getScalarEvolutionIVs() const { |
2265 | return ScalarEvolutionIVs; |
2266 | } |
2267 | |
2268 | void print_factors_and_types(raw_ostream &OS) const; |
2269 | void print_fixups(raw_ostream &OS) const; |
2270 | void print_uses(raw_ostream &OS) const; |
2271 | void print(raw_ostream &OS) const; |
2272 | void dump() const; |
2273 | }; |
2274 | |
2275 | } // end anonymous namespace |
2276 | |
2277 | /// If IV is used in a int-to-float cast inside the loop then try to eliminate |
2278 | /// the cast operation. |
2279 | void LSRInstance::OptimizeShadowIV() { |
2280 | const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); |
2281 | if (isa<SCEVCouldNotCompute>(Val: BackedgeTakenCount)) |
2282 | return; |
2283 | |
2284 | for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); |
2285 | UI != E; /* empty */) { |
2286 | IVUsers::const_iterator CandidateUI = UI; |
2287 | ++UI; |
2288 | Instruction *ShadowUse = CandidateUI->getUser(); |
2289 | Type *DestTy = nullptr; |
2290 | bool IsSigned = false; |
2291 | |
2292 | /* If shadow use is a int->float cast then insert a second IV |
2293 | to eliminate this cast. |
2294 | |
2295 | for (unsigned i = 0; i < n; ++i) |
2296 | foo((double)i); |
2297 | |
2298 | is transformed into |
2299 | |
2300 | double d = 0.0; |
2301 | for (unsigned i = 0; i < n; ++i, ++d) |
2302 | foo(d); |
2303 | */ |
2304 | if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(Val: CandidateUI->getUser())) { |
2305 | IsSigned = false; |
2306 | DestTy = UCast->getDestTy(); |
2307 | } |
2308 | else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(Val: CandidateUI->getUser())) { |
2309 | IsSigned = true; |
2310 | DestTy = SCast->getDestTy(); |
2311 | } |
2312 | if (!DestTy) continue; |
2313 | |
2314 | // If target does not support DestTy natively then do not apply |
2315 | // this transformation. |
2316 | if (!TTI.isTypeLegal(Ty: DestTy)) continue; |
2317 | |
2318 | PHINode *PH = dyn_cast<PHINode>(Val: ShadowUse->getOperand(i: 0)); |
2319 | if (!PH) continue; |
2320 | if (PH->getNumIncomingValues() != 2) continue; |
2321 | |
2322 | // If the calculation in integers overflows, the result in FP type will |
2323 | // differ. So we only can do this transformation if we are guaranteed to not |
2324 | // deal with overflowing values |
2325 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: SE.getSCEV(V: PH)); |
2326 | if (!AR) continue; |
2327 | if (IsSigned && !AR->hasNoSignedWrap()) continue; |
2328 | if (!IsSigned && !AR->hasNoUnsignedWrap()) continue; |
2329 | |
2330 | Type *SrcTy = PH->getType(); |
2331 | int Mantissa = DestTy->getFPMantissaWidth(); |
2332 | if (Mantissa == -1) continue; |
2333 | if ((int)SE.getTypeSizeInBits(Ty: SrcTy) > Mantissa) |
2334 | continue; |
2335 | |
2336 | unsigned Entry, Latch; |
2337 | if (PH->getIncomingBlock(i: 0) == L->getLoopPreheader()) { |
2338 | Entry = 0; |
2339 | Latch = 1; |
2340 | } else { |
2341 | Entry = 1; |
2342 | Latch = 0; |
2343 | } |
2344 | |
2345 | ConstantInt *Init = dyn_cast<ConstantInt>(Val: PH->getIncomingValue(i: Entry)); |
2346 | if (!Init) continue; |
2347 | Constant *NewInit = ConstantFP::get(Ty: DestTy, V: IsSigned ? |
2348 | (double)Init->getSExtValue() : |
2349 | (double)Init->getZExtValue()); |
2350 | |
2351 | BinaryOperator *Incr = |
2352 | dyn_cast<BinaryOperator>(Val: PH->getIncomingValue(i: Latch)); |
2353 | if (!Incr) continue; |
2354 | if (Incr->getOpcode() != Instruction::Add |
2355 | && Incr->getOpcode() != Instruction::Sub) |
2356 | continue; |
2357 | |
2358 | /* Initialize new IV, double d = 0.0 in above example. */ |
2359 | ConstantInt *C = nullptr; |
2360 | if (Incr->getOperand(i_nocapture: 0) == PH) |
2361 | C = dyn_cast<ConstantInt>(Val: Incr->getOperand(i_nocapture: 1)); |
2362 | else if (Incr->getOperand(i_nocapture: 1) == PH) |
2363 | C = dyn_cast<ConstantInt>(Val: Incr->getOperand(i_nocapture: 0)); |
2364 | else |
2365 | continue; |
2366 | |
2367 | if (!C) continue; |
2368 | |
2369 | // Ignore negative constants, as the code below doesn't handle them |
2370 | // correctly. TODO: Remove this restriction. |
2371 | if (!C->getValue().isStrictlyPositive()) |
2372 | continue; |
2373 | |
2374 | /* Add new PHINode. */ |
2375 | PHINode *NewPH = PHINode::Create(Ty: DestTy, NumReservedValues: 2, NameStr: "IV.S." , InsertBefore: PH->getIterator()); |
2376 | NewPH->setDebugLoc(PH->getDebugLoc()); |
2377 | |
2378 | /* create new increment. '++d' in above example. */ |
2379 | Constant *CFP = ConstantFP::get(Ty: DestTy, V: C->getZExtValue()); |
2380 | BinaryOperator *NewIncr = BinaryOperator::Create( |
2381 | Op: Incr->getOpcode() == Instruction::Add ? Instruction::FAdd |
2382 | : Instruction::FSub, |
2383 | S1: NewPH, S2: CFP, Name: "IV.S.next." , InsertBefore: Incr->getIterator()); |
2384 | NewIncr->setDebugLoc(Incr->getDebugLoc()); |
2385 | |
2386 | NewPH->addIncoming(V: NewInit, BB: PH->getIncomingBlock(i: Entry)); |
2387 | NewPH->addIncoming(V: NewIncr, BB: PH->getIncomingBlock(i: Latch)); |
2388 | |
2389 | /* Remove cast operation */ |
2390 | ShadowUse->replaceAllUsesWith(V: NewPH); |
2391 | ShadowUse->eraseFromParent(); |
2392 | Changed = true; |
2393 | break; |
2394 | } |
2395 | } |
2396 | |
2397 | /// If Cond has an operand that is an expression of an IV, set the IV user and |
2398 | /// stride information and return true, otherwise return false. |
2399 | bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) { |
2400 | for (IVStrideUse &U : IU) |
2401 | if (U.getUser() == Cond) { |
2402 | // NOTE: we could handle setcc instructions with multiple uses here, but |
2403 | // InstCombine does it as well for simple uses, it's not clear that it |
2404 | // occurs enough in real life to handle. |
2405 | CondUse = &U; |
2406 | return true; |
2407 | } |
2408 | return false; |
2409 | } |
2410 | |
2411 | /// Rewrite the loop's terminating condition if it uses a max computation. |
2412 | /// |
2413 | /// This is a narrow solution to a specific, but acute, problem. For loops |
2414 | /// like this: |
2415 | /// |
2416 | /// i = 0; |
2417 | /// do { |
2418 | /// p[i] = 0.0; |
2419 | /// } while (++i < n); |
2420 | /// |
2421 | /// the trip count isn't just 'n', because 'n' might not be positive. And |
2422 | /// unfortunately this can come up even for loops where the user didn't use |
2423 | /// a C do-while loop. For example, seemingly well-behaved top-test loops |
2424 | /// will commonly be lowered like this: |
2425 | /// |
2426 | /// if (n > 0) { |
2427 | /// i = 0; |
2428 | /// do { |
2429 | /// p[i] = 0.0; |
2430 | /// } while (++i < n); |
2431 | /// } |
2432 | /// |
2433 | /// and then it's possible for subsequent optimization to obscure the if |
2434 | /// test in such a way that indvars can't find it. |
2435 | /// |
2436 | /// When indvars can't find the if test in loops like this, it creates a |
2437 | /// max expression, which allows it to give the loop a canonical |
2438 | /// induction variable: |
2439 | /// |
2440 | /// i = 0; |
2441 | /// max = n < 1 ? 1 : n; |
2442 | /// do { |
2443 | /// p[i] = 0.0; |
2444 | /// } while (++i != max); |
2445 | /// |
2446 | /// Canonical induction variables are necessary because the loop passes |
2447 | /// are designed around them. The most obvious example of this is the |
2448 | /// LoopInfo analysis, which doesn't remember trip count values. It |
2449 | /// expects to be able to rediscover the trip count each time it is |
2450 | /// needed, and it does this using a simple analysis that only succeeds if |
2451 | /// the loop has a canonical induction variable. |
2452 | /// |
2453 | /// However, when it comes time to generate code, the maximum operation |
2454 | /// can be quite costly, especially if it's inside of an outer loop. |
2455 | /// |
2456 | /// This function solves this problem by detecting this type of loop and |
2457 | /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting |
2458 | /// the instructions for the maximum computation. |
2459 | ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) { |
2460 | // Check that the loop matches the pattern we're looking for. |
2461 | if (Cond->getPredicate() != CmpInst::ICMP_EQ && |
2462 | Cond->getPredicate() != CmpInst::ICMP_NE) |
2463 | return Cond; |
2464 | |
2465 | SelectInst *Sel = dyn_cast<SelectInst>(Val: Cond->getOperand(i_nocapture: 1)); |
2466 | if (!Sel || !Sel->hasOneUse()) return Cond; |
2467 | |
2468 | const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); |
2469 | if (isa<SCEVCouldNotCompute>(Val: BackedgeTakenCount)) |
2470 | return Cond; |
2471 | const SCEV *One = SE.getConstant(Ty: BackedgeTakenCount->getType(), V: 1); |
2472 | |
2473 | // Add one to the backedge-taken count to get the trip count. |
2474 | const SCEV *IterationCount = SE.getAddExpr(LHS: One, RHS: BackedgeTakenCount); |
2475 | if (IterationCount != SE.getSCEV(V: Sel)) return Cond; |
2476 | |
2477 | // Check for a max calculation that matches the pattern. There's no check |
2478 | // for ICMP_ULE here because the comparison would be with zero, which |
2479 | // isn't interesting. |
2480 | CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; |
2481 | const SCEVNAryExpr *Max = nullptr; |
2482 | if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(Val: BackedgeTakenCount)) { |
2483 | Pred = ICmpInst::ICMP_SLE; |
2484 | Max = S; |
2485 | } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(Val: IterationCount)) { |
2486 | Pred = ICmpInst::ICMP_SLT; |
2487 | Max = S; |
2488 | } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(Val: IterationCount)) { |
2489 | Pred = ICmpInst::ICMP_ULT; |
2490 | Max = U; |
2491 | } else { |
2492 | // No match; bail. |
2493 | return Cond; |
2494 | } |
2495 | |
2496 | // To handle a max with more than two operands, this optimization would |
2497 | // require additional checking and setup. |
2498 | if (Max->getNumOperands() != 2) |
2499 | return Cond; |
2500 | |
2501 | const SCEV *MaxLHS = Max->getOperand(i: 0); |
2502 | const SCEV *MaxRHS = Max->getOperand(i: 1); |
2503 | |
2504 | // ScalarEvolution canonicalizes constants to the left. For < and >, look |
2505 | // for a comparison with 1. For <= and >=, a comparison with zero. |
2506 | if (!MaxLHS || |
2507 | (ICmpInst::isTrueWhenEqual(predicate: Pred) ? !MaxLHS->isZero() : (MaxLHS != One))) |
2508 | return Cond; |
2509 | |
2510 | // Check the relevant induction variable for conformance to |
2511 | // the pattern. |
2512 | const SCEV *IV = SE.getSCEV(V: Cond->getOperand(i_nocapture: 0)); |
2513 | if (!match(S: IV, |
2514 | P: m_scev_AffineAddRec(Op0: m_scev_SpecificInt(V: 1), Op1: m_scev_SpecificInt(V: 1)))) |
2515 | return Cond; |
2516 | |
2517 | assert(cast<SCEVAddRecExpr>(IV)->getLoop() == L && |
2518 | "Loop condition operand is an addrec in a different loop!" ); |
2519 | |
2520 | // Check the right operand of the select, and remember it, as it will |
2521 | // be used in the new comparison instruction. |
2522 | Value *NewRHS = nullptr; |
2523 | if (ICmpInst::isTrueWhenEqual(predicate: Pred)) { |
2524 | // Look for n+1, and grab n. |
2525 | if (AddOperator *BO = dyn_cast<AddOperator>(Val: Sel->getOperand(i_nocapture: 1))) |
2526 | if (ConstantInt *BO1 = dyn_cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))) |
2527 | if (BO1->isOne() && SE.getSCEV(V: BO->getOperand(i_nocapture: 0)) == MaxRHS) |
2528 | NewRHS = BO->getOperand(i_nocapture: 0); |
2529 | if (AddOperator *BO = dyn_cast<AddOperator>(Val: Sel->getOperand(i_nocapture: 2))) |
2530 | if (ConstantInt *BO1 = dyn_cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))) |
2531 | if (BO1->isOne() && SE.getSCEV(V: BO->getOperand(i_nocapture: 0)) == MaxRHS) |
2532 | NewRHS = BO->getOperand(i_nocapture: 0); |
2533 | if (!NewRHS) |
2534 | return Cond; |
2535 | } else if (SE.getSCEV(V: Sel->getOperand(i_nocapture: 1)) == MaxRHS) |
2536 | NewRHS = Sel->getOperand(i_nocapture: 1); |
2537 | else if (SE.getSCEV(V: Sel->getOperand(i_nocapture: 2)) == MaxRHS) |
2538 | NewRHS = Sel->getOperand(i_nocapture: 2); |
2539 | else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Val: MaxRHS)) |
2540 | NewRHS = SU->getValue(); |
2541 | else |
2542 | // Max doesn't match expected pattern. |
2543 | return Cond; |
2544 | |
2545 | // Determine the new comparison opcode. It may be signed or unsigned, |
2546 | // and the original comparison may be either equality or inequality. |
2547 | if (Cond->getPredicate() == CmpInst::ICMP_EQ) |
2548 | Pred = CmpInst::getInversePredicate(pred: Pred); |
2549 | |
2550 | // Ok, everything looks ok to change the condition into an SLT or SGE and |
2551 | // delete the max calculation. |
2552 | ICmpInst *NewCond = new ICmpInst(Cond->getIterator(), Pred, |
2553 | Cond->getOperand(i_nocapture: 0), NewRHS, "scmp" ); |
2554 | |
2555 | // Delete the max calculation instructions. |
2556 | NewCond->setDebugLoc(Cond->getDebugLoc()); |
2557 | Cond->replaceAllUsesWith(V: NewCond); |
2558 | CondUse->setUser(NewCond); |
2559 | Instruction *Cmp = cast<Instruction>(Val: Sel->getOperand(i_nocapture: 0)); |
2560 | Cond->eraseFromParent(); |
2561 | Sel->eraseFromParent(); |
2562 | if (Cmp->use_empty()) |
2563 | Cmp->eraseFromParent(); |
2564 | return NewCond; |
2565 | } |
2566 | |
2567 | /// Change loop terminating condition to use the postinc iv when possible. |
2568 | void |
2569 | LSRInstance::OptimizeLoopTermCond() { |
2570 | SmallPtrSet<Instruction *, 4> PostIncs; |
2571 | |
2572 | // We need a different set of heuristics for rotated and non-rotated loops. |
2573 | // If a loop is rotated then the latch is also the backedge, so inserting |
2574 | // post-inc expressions just before the latch is ideal. To reduce live ranges |
2575 | // it also makes sense to rewrite terminating conditions to use post-inc |
2576 | // expressions. |
2577 | // |
2578 | // If the loop is not rotated then the latch is not a backedge; the latch |
2579 | // check is done in the loop head. Adding post-inc expressions before the |
2580 | // latch will cause overlapping live-ranges of pre-inc and post-inc expressions |
2581 | // in the loop body. In this case we do *not* want to use post-inc expressions |
2582 | // in the latch check, and we want to insert post-inc expressions before |
2583 | // the backedge. |
2584 | BasicBlock *LatchBlock = L->getLoopLatch(); |
2585 | SmallVector<BasicBlock*, 8> ExitingBlocks; |
2586 | L->getExitingBlocks(ExitingBlocks); |
2587 | if (!llvm::is_contained(Range&: ExitingBlocks, Element: LatchBlock)) { |
2588 | // The backedge doesn't exit the loop; treat this as a head-tested loop. |
2589 | IVIncInsertPos = LatchBlock->getTerminator(); |
2590 | return; |
2591 | } |
2592 | |
2593 | // Otherwise treat this as a rotated loop. |
2594 | for (BasicBlock *ExitingBlock : ExitingBlocks) { |
2595 | // Get the terminating condition for the loop if possible. If we |
2596 | // can, we want to change it to use a post-incremented version of its |
2597 | // induction variable, to allow coalescing the live ranges for the IV into |
2598 | // one register value. |
2599 | |
2600 | BranchInst *TermBr = dyn_cast<BranchInst>(Val: ExitingBlock->getTerminator()); |
2601 | if (!TermBr) |
2602 | continue; |
2603 | // FIXME: Overly conservative, termination condition could be an 'or' etc.. |
2604 | if (TermBr->isUnconditional() || !isa<ICmpInst>(Val: TermBr->getCondition())) |
2605 | continue; |
2606 | |
2607 | // Search IVUsesByStride to find Cond's IVUse if there is one. |
2608 | IVStrideUse *CondUse = nullptr; |
2609 | ICmpInst *Cond = cast<ICmpInst>(Val: TermBr->getCondition()); |
2610 | if (!FindIVUserForCond(Cond, CondUse)) |
2611 | continue; |
2612 | |
2613 | // If the trip count is computed in terms of a max (due to ScalarEvolution |
2614 | // being unable to find a sufficient guard, for example), change the loop |
2615 | // comparison to use SLT or ULT instead of NE. |
2616 | // One consequence of doing this now is that it disrupts the count-down |
2617 | // optimization. That's not always a bad thing though, because in such |
2618 | // cases it may still be worthwhile to avoid a max. |
2619 | Cond = OptimizeMax(Cond, CondUse); |
2620 | |
2621 | // If this exiting block dominates the latch block, it may also use |
2622 | // the post-inc value if it won't be shared with other uses. |
2623 | // Check for dominance. |
2624 | if (!DT.dominates(A: ExitingBlock, B: LatchBlock)) |
2625 | continue; |
2626 | |
2627 | // Conservatively avoid trying to use the post-inc value in non-latch |
2628 | // exits if there may be pre-inc users in intervening blocks. |
2629 | if (LatchBlock != ExitingBlock) |
2630 | for (const IVStrideUse &UI : IU) |
2631 | // Test if the use is reachable from the exiting block. This dominator |
2632 | // query is a conservative approximation of reachability. |
2633 | if (&UI != CondUse && |
2634 | !DT.properlyDominates(A: UI.getUser()->getParent(), B: ExitingBlock)) { |
2635 | // Conservatively assume there may be reuse if the quotient of their |
2636 | // strides could be a legal scale. |
2637 | const SCEV *A = IU.getStride(IU: *CondUse, L); |
2638 | const SCEV *B = IU.getStride(IU: UI, L); |
2639 | if (!A || !B) continue; |
2640 | if (SE.getTypeSizeInBits(Ty: A->getType()) != |
2641 | SE.getTypeSizeInBits(Ty: B->getType())) { |
2642 | if (SE.getTypeSizeInBits(Ty: A->getType()) > |
2643 | SE.getTypeSizeInBits(Ty: B->getType())) |
2644 | B = SE.getSignExtendExpr(Op: B, Ty: A->getType()); |
2645 | else |
2646 | A = SE.getSignExtendExpr(Op: A, Ty: B->getType()); |
2647 | } |
2648 | if (const SCEVConstant *D = |
2649 | dyn_cast_or_null<SCEVConstant>(Val: getExactSDiv(LHS: B, RHS: A, SE))) { |
2650 | const ConstantInt *C = D->getValue(); |
2651 | // Stride of one or negative one can have reuse with non-addresses. |
2652 | if (C->isOne() || C->isMinusOne()) |
2653 | goto decline_post_inc; |
2654 | // Avoid weird situations. |
2655 | if (C->getValue().getSignificantBits() >= 64 || |
2656 | C->getValue().isMinSignedValue()) |
2657 | goto decline_post_inc; |
2658 | // Check for possible scaled-address reuse. |
2659 | if (isAddressUse(TTI, Inst: UI.getUser(), OperandVal: UI.getOperandValToReplace())) { |
2660 | MemAccessTy AccessTy = |
2661 | getAccessType(TTI, Inst: UI.getUser(), OperandVal: UI.getOperandValToReplace()); |
2662 | int64_t Scale = C->getSExtValue(); |
2663 | if (TTI.isLegalAddressingMode(Ty: AccessTy.MemTy, /*BaseGV=*/nullptr, |
2664 | /*BaseOffset=*/0, |
2665 | /*HasBaseReg=*/true, Scale, |
2666 | AddrSpace: AccessTy.AddrSpace)) |
2667 | goto decline_post_inc; |
2668 | Scale = -Scale; |
2669 | if (TTI.isLegalAddressingMode(Ty: AccessTy.MemTy, /*BaseGV=*/nullptr, |
2670 | /*BaseOffset=*/0, |
2671 | /*HasBaseReg=*/true, Scale, |
2672 | AddrSpace: AccessTy.AddrSpace)) |
2673 | goto decline_post_inc; |
2674 | } |
2675 | } |
2676 | } |
2677 | |
2678 | LLVM_DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " |
2679 | << *Cond << '\n'); |
2680 | |
2681 | // It's possible for the setcc instruction to be anywhere in the loop, and |
2682 | // possible for it to have multiple users. If it is not immediately before |
2683 | // the exiting block branch, move it. |
2684 | if (Cond->getNextNonDebugInstruction() != TermBr) { |
2685 | if (Cond->hasOneUse()) { |
2686 | Cond->moveBefore(InsertPos: TermBr->getIterator()); |
2687 | } else { |
2688 | // Clone the terminating condition and insert into the loopend. |
2689 | ICmpInst *OldCond = Cond; |
2690 | Cond = cast<ICmpInst>(Val: Cond->clone()); |
2691 | Cond->setName(L->getHeader()->getName() + ".termcond" ); |
2692 | Cond->insertInto(ParentBB: ExitingBlock, It: TermBr->getIterator()); |
2693 | |
2694 | // Clone the IVUse, as the old use still exists! |
2695 | CondUse = &IU.AddUser(User: Cond, Operand: CondUse->getOperandValToReplace()); |
2696 | TermBr->replaceUsesOfWith(From: OldCond, To: Cond); |
2697 | } |
2698 | } |
2699 | |
2700 | // If we get to here, we know that we can transform the setcc instruction to |
2701 | // use the post-incremented version of the IV, allowing us to coalesce the |
2702 | // live ranges for the IV correctly. |
2703 | CondUse->transformToPostInc(L); |
2704 | Changed = true; |
2705 | |
2706 | PostIncs.insert(Ptr: Cond); |
2707 | decline_post_inc:; |
2708 | } |
2709 | |
2710 | // Determine an insertion point for the loop induction variable increment. It |
2711 | // must dominate all the post-inc comparisons we just set up, and it must |
2712 | // dominate the loop latch edge. |
2713 | IVIncInsertPos = L->getLoopLatch()->getTerminator(); |
2714 | for (Instruction *Inst : PostIncs) |
2715 | IVIncInsertPos = DT.findNearestCommonDominator(I1: IVIncInsertPos, I2: Inst); |
2716 | } |
2717 | |
2718 | /// Determine if the given use can accommodate a fixup at the given offset and |
2719 | /// other details. If so, update the use and return true. |
2720 | bool LSRInstance::reconcileNewOffset(LSRUse &LU, Immediate NewOffset, |
2721 | bool HasBaseReg, LSRUse::KindType Kind, |
2722 | MemAccessTy AccessTy) { |
2723 | Immediate NewMinOffset = LU.MinOffset; |
2724 | Immediate NewMaxOffset = LU.MaxOffset; |
2725 | MemAccessTy NewAccessTy = AccessTy; |
2726 | |
2727 | // Check for a mismatched kind. It's tempting to collapse mismatched kinds to |
2728 | // something conservative, however this can pessimize in the case that one of |
2729 | // the uses will have all its uses outside the loop, for example. |
2730 | if (LU.Kind != Kind) |
2731 | return false; |
2732 | |
2733 | // Check for a mismatched access type, and fall back conservatively as needed. |
2734 | // TODO: Be less conservative when the type is similar and can use the same |
2735 | // addressing modes. |
2736 | if (Kind == LSRUse::Address) { |
2737 | if (AccessTy.MemTy != LU.AccessTy.MemTy) { |
2738 | NewAccessTy = MemAccessTy::getUnknown(Ctx&: AccessTy.MemTy->getContext(), |
2739 | AS: AccessTy.AddrSpace); |
2740 | } |
2741 | } |
2742 | |
2743 | // Conservatively assume HasBaseReg is true for now. |
2744 | if (Immediate::isKnownLT(LHS: NewOffset, RHS: LU.MinOffset)) { |
2745 | if (!isAlwaysFoldable(TTI, Kind, AccessTy: NewAccessTy, /*BaseGV=*/nullptr, |
2746 | BaseOffset: LU.MaxOffset - NewOffset, HasBaseReg)) |
2747 | return false; |
2748 | NewMinOffset = NewOffset; |
2749 | } else if (Immediate::isKnownGT(LHS: NewOffset, RHS: LU.MaxOffset)) { |
2750 | if (!isAlwaysFoldable(TTI, Kind, AccessTy: NewAccessTy, /*BaseGV=*/nullptr, |
2751 | BaseOffset: NewOffset - LU.MinOffset, HasBaseReg)) |
2752 | return false; |
2753 | NewMaxOffset = NewOffset; |
2754 | } |
2755 | |
2756 | // FIXME: We should be able to handle some level of scalable offset support |
2757 | // for 'void', but in order to get basic support up and running this is |
2758 | // being left out. |
2759 | if (NewAccessTy.MemTy && NewAccessTy.MemTy->isVoidTy() && |
2760 | (NewMinOffset.isScalable() || NewMaxOffset.isScalable())) |
2761 | return false; |
2762 | |
2763 | // Update the use. |
2764 | LU.MinOffset = NewMinOffset; |
2765 | LU.MaxOffset = NewMaxOffset; |
2766 | LU.AccessTy = NewAccessTy; |
2767 | return true; |
2768 | } |
2769 | |
2770 | /// Return an LSRUse index and an offset value for a fixup which needs the given |
2771 | /// expression, with the given kind and optional access type. Either reuse an |
2772 | /// existing use or create a new one, as needed. |
2773 | std::pair<size_t, Immediate> LSRInstance::getUse(const SCEV *&Expr, |
2774 | LSRUse::KindType Kind, |
2775 | MemAccessTy AccessTy) { |
2776 | const SCEV *Copy = Expr; |
2777 | Immediate Offset = ExtractImmediate(S&: Expr, SE); |
2778 | |
2779 | // Basic uses can't accept any offset, for example. |
2780 | if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr, |
2781 | BaseOffset: Offset, /*HasBaseReg=*/ true)) { |
2782 | Expr = Copy; |
2783 | Offset = Immediate::getFixed(MinVal: 0); |
2784 | } |
2785 | |
2786 | std::pair<UseMapTy::iterator, bool> P = |
2787 | UseMap.try_emplace(Key: LSRUse::SCEVUseKindPair(Expr, Kind)); |
2788 | if (!P.second) { |
2789 | // A use already existed with this base. |
2790 | size_t LUIdx = P.first->second; |
2791 | LSRUse &LU = Uses[LUIdx]; |
2792 | if (reconcileNewOffset(LU, NewOffset: Offset, /*HasBaseReg=*/true, Kind, AccessTy)) |
2793 | // Reuse this use. |
2794 | return std::make_pair(x&: LUIdx, y&: Offset); |
2795 | } |
2796 | |
2797 | // Create a new use. |
2798 | size_t LUIdx = Uses.size(); |
2799 | P.first->second = LUIdx; |
2800 | Uses.push_back(Elt: LSRUse(Kind, AccessTy)); |
2801 | LSRUse &LU = Uses[LUIdx]; |
2802 | |
2803 | LU.MinOffset = Offset; |
2804 | LU.MaxOffset = Offset; |
2805 | return std::make_pair(x&: LUIdx, y&: Offset); |
2806 | } |
2807 | |
2808 | /// Delete the given use from the Uses list. |
2809 | void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) { |
2810 | if (&LU != &Uses.back()) |
2811 | std::swap(a&: LU, b&: Uses.back()); |
2812 | Uses.pop_back(); |
2813 | |
2814 | // Update RegUses. |
2815 | RegUses.swapAndDropUse(LUIdx, LastLUIdx: Uses.size()); |
2816 | } |
2817 | |
2818 | /// Look for a use distinct from OrigLU which is has a formula that has the same |
2819 | /// registers as the given formula. |
2820 | LSRUse * |
2821 | LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF, |
2822 | const LSRUse &OrigLU) { |
2823 | // Search all uses for the formula. This could be more clever. |
2824 | for (LSRUse &LU : Uses) { |
2825 | // Check whether this use is close enough to OrigLU, to see whether it's |
2826 | // worthwhile looking through its formulae. |
2827 | // Ignore ICmpZero uses because they may contain formulae generated by |
2828 | // GenerateICmpZeroScales, in which case adding fixup offsets may |
2829 | // be invalid. |
2830 | if (&LU != &OrigLU && |
2831 | LU.Kind != LSRUse::ICmpZero && |
2832 | LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy && |
2833 | LU.WidestFixupType == OrigLU.WidestFixupType && |
2834 | LU.HasFormulaWithSameRegs(F: OrigF)) { |
2835 | // Scan through this use's formulae. |
2836 | for (const Formula &F : LU.Formulae) { |
2837 | // Check to see if this formula has the same registers and symbols |
2838 | // as OrigF. |
2839 | if (F.BaseRegs == OrigF.BaseRegs && |
2840 | F.ScaledReg == OrigF.ScaledReg && |
2841 | F.BaseGV == OrigF.BaseGV && |
2842 | F.Scale == OrigF.Scale && |
2843 | F.UnfoldedOffset == OrigF.UnfoldedOffset) { |
2844 | if (F.BaseOffset.isZero()) |
2845 | return &LU; |
2846 | // This is the formula where all the registers and symbols matched; |
2847 | // there aren't going to be any others. Since we declined it, we |
2848 | // can skip the rest of the formulae and proceed to the next LSRUse. |
2849 | break; |
2850 | } |
2851 | } |
2852 | } |
2853 | } |
2854 | |
2855 | // Nothing looked good. |
2856 | return nullptr; |
2857 | } |
2858 | |
2859 | void LSRInstance::CollectInterestingTypesAndFactors() { |
2860 | SmallSetVector<const SCEV *, 4> Strides; |
2861 | |
2862 | // Collect interesting types and strides. |
2863 | SmallVector<const SCEV *, 4> Worklist; |
2864 | for (const IVStrideUse &U : IU) { |
2865 | const SCEV *Expr = IU.getExpr(IU: U); |
2866 | if (!Expr) |
2867 | continue; |
2868 | |
2869 | // Collect interesting types. |
2870 | Types.insert(X: SE.getEffectiveSCEVType(Ty: Expr->getType())); |
2871 | |
2872 | // Add strides for mentioned loops. |
2873 | Worklist.push_back(Elt: Expr); |
2874 | do { |
2875 | const SCEV *S = Worklist.pop_back_val(); |
2876 | if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: S)) { |
2877 | if (AR->getLoop() == L) |
2878 | Strides.insert(X: AR->getStepRecurrence(SE)); |
2879 | Worklist.push_back(Elt: AR->getStart()); |
2880 | } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Val: S)) { |
2881 | append_range(C&: Worklist, R: Add->operands()); |
2882 | } |
2883 | } while (!Worklist.empty()); |
2884 | } |
2885 | |
2886 | // Compute interesting factors from the set of interesting strides. |
2887 | for (SmallSetVector<const SCEV *, 4>::const_iterator |
2888 | I = Strides.begin(), E = Strides.end(); I != E; ++I) |
2889 | for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter = |
2890 | std::next(x: I); NewStrideIter != E; ++NewStrideIter) { |
2891 | const SCEV *OldStride = *I; |
2892 | const SCEV *NewStride = *NewStrideIter; |
2893 | |
2894 | if (SE.getTypeSizeInBits(Ty: OldStride->getType()) != |
2895 | SE.getTypeSizeInBits(Ty: NewStride->getType())) { |
2896 | if (SE.getTypeSizeInBits(Ty: OldStride->getType()) > |
2897 | SE.getTypeSizeInBits(Ty: NewStride->getType())) |
2898 | NewStride = SE.getSignExtendExpr(Op: NewStride, Ty: OldStride->getType()); |
2899 | else |
2900 | OldStride = SE.getSignExtendExpr(Op: OldStride, Ty: NewStride->getType()); |
2901 | } |
2902 | if (const SCEVConstant *Factor = |
2903 | dyn_cast_or_null<SCEVConstant>(Val: getExactSDiv(LHS: NewStride, RHS: OldStride, |
2904 | SE, IgnoreSignificantBits: true))) { |
2905 | if (Factor->getAPInt().getSignificantBits() <= 64 && !Factor->isZero()) |
2906 | Factors.insert(X: Factor->getAPInt().getSExtValue()); |
2907 | } else if (const SCEVConstant *Factor = |
2908 | dyn_cast_or_null<SCEVConstant>(Val: getExactSDiv(LHS: OldStride, |
2909 | RHS: NewStride, |
2910 | SE, IgnoreSignificantBits: true))) { |
2911 | if (Factor->getAPInt().getSignificantBits() <= 64 && !Factor->isZero()) |
2912 | Factors.insert(X: Factor->getAPInt().getSExtValue()); |
2913 | } |
2914 | } |
2915 | |
2916 | // If all uses use the same type, don't bother looking for truncation-based |
2917 | // reuse. |
2918 | if (Types.size() == 1) |
2919 | Types.clear(); |
2920 | |
2921 | LLVM_DEBUG(print_factors_and_types(dbgs())); |
2922 | } |
2923 | |
2924 | /// Helper for CollectChains that finds an IV operand (computed by an AddRec in |
2925 | /// this loop) within [OI,OE) or returns OE. If IVUsers mapped Instructions to |
2926 | /// IVStrideUses, we could partially skip this. |
2927 | static User::op_iterator |
2928 | findIVOperand(User::op_iterator OI, User::op_iterator OE, |
2929 | Loop *L, ScalarEvolution &SE) { |
2930 | for(; OI != OE; ++OI) { |
2931 | if (Instruction *Oper = dyn_cast<Instruction>(Val&: *OI)) { |
2932 | if (!SE.isSCEVable(Ty: Oper->getType())) |
2933 | continue; |
2934 | |
2935 | if (const SCEVAddRecExpr *AR = |
2936 | dyn_cast<SCEVAddRecExpr>(Val: SE.getSCEV(V: Oper))) { |
2937 | if (AR->getLoop() == L) |
2938 | break; |
2939 | } |
2940 | } |
2941 | } |
2942 | return OI; |
2943 | } |
2944 | |
2945 | /// IVChain logic must consistently peek base TruncInst operands, so wrap it in |
2946 | /// a convenient helper. |
2947 | static Value *getWideOperand(Value *Oper) { |
2948 | if (TruncInst *Trunc = dyn_cast<TruncInst>(Val: Oper)) |
2949 | return Trunc->getOperand(i_nocapture: 0); |
2950 | return Oper; |
2951 | } |
2952 | |
2953 | /// Return an approximation of this SCEV expression's "base", or NULL for any |
2954 | /// constant. Returning the expression itself is conservative. Returning a |
2955 | /// deeper subexpression is more precise and valid as long as it isn't less |
2956 | /// complex than another subexpression. For expressions involving multiple |
2957 | /// unscaled values, we need to return the pointer-type SCEVUnknown. This avoids |
2958 | /// forming chains across objects, such as: PrevOper==a[i], IVOper==b[i], |
2959 | /// IVInc==b-a. |
2960 | /// |
2961 | /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost |
2962 | /// SCEVUnknown, we simply return the rightmost SCEV operand. |
2963 | static const SCEV *getExprBase(const SCEV *S) { |
2964 | switch (S->getSCEVType()) { |
2965 | default: // including scUnknown. |
2966 | return S; |
2967 | case scConstant: |
2968 | case scVScale: |
2969 | return nullptr; |
2970 | case scTruncate: |
2971 | return getExprBase(S: cast<SCEVTruncateExpr>(Val: S)->getOperand()); |
2972 | case scZeroExtend: |
2973 | return getExprBase(S: cast<SCEVZeroExtendExpr>(Val: S)->getOperand()); |
2974 | case scSignExtend: |
2975 | return getExprBase(S: cast<SCEVSignExtendExpr>(Val: S)->getOperand()); |
2976 | case scAddExpr: { |
2977 | // Skip over scaled operands (scMulExpr) to follow add operands as long as |
2978 | // there's nothing more complex. |
2979 | // FIXME: not sure if we want to recognize negation. |
2980 | const SCEVAddExpr *Add = cast<SCEVAddExpr>(Val: S); |
2981 | for (const SCEV *SubExpr : reverse(C: Add->operands())) { |
2982 | if (SubExpr->getSCEVType() == scAddExpr) |
2983 | return getExprBase(S: SubExpr); |
2984 | |
2985 | if (SubExpr->getSCEVType() != scMulExpr) |
2986 | return SubExpr; |
2987 | } |
2988 | return S; // all operands are scaled, be conservative. |
2989 | } |
2990 | case scAddRecExpr: |
2991 | return getExprBase(S: cast<SCEVAddRecExpr>(Val: S)->getStart()); |
2992 | } |
2993 | llvm_unreachable("Unknown SCEV kind!" ); |
2994 | } |
2995 | |
2996 | /// Return true if the chain increment is profitable to expand into a loop |
2997 | /// invariant value, which may require its own register. A profitable chain |
2998 | /// increment will be an offset relative to the same base. We allow such offsets |
2999 | /// to potentially be used as chain increment as long as it's not obviously |
3000 | /// expensive to expand using real instructions. |
3001 | bool IVChain::isProfitableIncrement(const SCEV *OperExpr, |
3002 | const SCEV *IncExpr, |
3003 | ScalarEvolution &SE) { |
3004 | // Aggressively form chains when -stress-ivchain. |
3005 | if (StressIVChain) |
3006 | return true; |
3007 | |
3008 | // Do not replace a constant offset from IV head with a nonconstant IV |
3009 | // increment. |
3010 | if (!isa<SCEVConstant>(Val: IncExpr)) { |
3011 | const SCEV *HeadExpr = SE.getSCEV(V: getWideOperand(Oper: Incs[0].IVOperand)); |
3012 | if (isa<SCEVConstant>(Val: SE.getMinusSCEV(LHS: OperExpr, RHS: HeadExpr))) |
3013 | return false; |
3014 | } |
3015 | |
3016 | SmallPtrSet<const SCEV*, 8> Processed; |
3017 | return !isHighCostExpansion(S: IncExpr, Processed, SE); |
3018 | } |
3019 | |
3020 | /// Return true if the number of registers needed for the chain is estimated to |
3021 | /// be less than the number required for the individual IV users. First prohibit |
3022 | /// any IV users that keep the IV live across increments (the Users set should |
3023 | /// be empty). Next count the number and type of increments in the chain. |
3024 | /// |
3025 | /// Chaining IVs can lead to considerable code bloat if ISEL doesn't |
3026 | /// effectively use postinc addressing modes. Only consider it profitable it the |
3027 | /// increments can be computed in fewer registers when chained. |
3028 | /// |
3029 | /// TODO: Consider IVInc free if it's already used in another chains. |
3030 | static bool isProfitableChain(IVChain &Chain, |
3031 | SmallPtrSetImpl<Instruction *> &Users, |
3032 | ScalarEvolution &SE, |
3033 | const TargetTransformInfo &TTI) { |
3034 | if (StressIVChain) |
3035 | return true; |
3036 | |
3037 | if (!Chain.hasIncs()) |
3038 | return false; |
3039 | |
3040 | if (!Users.empty()) { |
3041 | LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n" ; |
3042 | for (Instruction *Inst |
3043 | : Users) { dbgs() << " " << *Inst << "\n" ; }); |
3044 | return false; |
3045 | } |
3046 | assert(!Chain.Incs.empty() && "empty IV chains are not allowed" ); |
3047 | |
3048 | // The chain itself may require a register, so intialize cost to 1. |
3049 | int cost = 1; |
3050 | |
3051 | // A complete chain likely eliminates the need for keeping the original IV in |
3052 | // a register. LSR does not currently know how to form a complete chain unless |
3053 | // the header phi already exists. |
3054 | if (isa<PHINode>(Val: Chain.tailUserInst()) |
3055 | && SE.getSCEV(V: Chain.tailUserInst()) == Chain.Incs[0].IncExpr) { |
3056 | --cost; |
3057 | } |
3058 | const SCEV *LastIncExpr = nullptr; |
3059 | unsigned NumConstIncrements = 0; |
3060 | unsigned NumVarIncrements = 0; |
3061 | unsigned NumReusedIncrements = 0; |
3062 | |
3063 | if (TTI.isProfitableLSRChainElement(I: Chain.Incs[0].UserInst)) |
3064 | return true; |
3065 | |
3066 | for (const IVInc &Inc : Chain) { |
3067 | if (TTI.isProfitableLSRChainElement(I: Inc.UserInst)) |
3068 | return true; |
3069 | if (Inc.IncExpr->isZero()) |
3070 | continue; |
3071 | |
3072 | // Incrementing by zero or some constant is neutral. We assume constants can |
3073 | // be folded into an addressing mode or an add's immediate operand. |
3074 | if (isa<SCEVConstant>(Val: Inc.IncExpr)) { |
3075 | ++NumConstIncrements; |
3076 | continue; |
3077 | } |
3078 | |
3079 | if (Inc.IncExpr == LastIncExpr) |
3080 | ++NumReusedIncrements; |
3081 | else |
3082 | ++NumVarIncrements; |
3083 | |
3084 | LastIncExpr = Inc.IncExpr; |
3085 | } |
3086 | // An IV chain with a single increment is handled by LSR's postinc |
3087 | // uses. However, a chain with multiple increments requires keeping the IV's |
3088 | // value live longer than it needs to be if chained. |
3089 | if (NumConstIncrements > 1) |
3090 | --cost; |
3091 | |
3092 | // Materializing increment expressions in the preheader that didn't exist in |
3093 | // the original code may cost a register. For example, sign-extended array |
3094 | // indices can produce ridiculous increments like this: |
3095 | // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64))) |
3096 | cost += NumVarIncrements; |
3097 | |
3098 | // Reusing variable increments likely saves a register to hold the multiple of |
3099 | // the stride. |
3100 | cost -= NumReusedIncrements; |
3101 | |
3102 | LLVM_DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost |
3103 | << "\n" ); |
3104 | |
3105 | return cost < 0; |
3106 | } |
3107 | |
3108 | /// Add this IV user to an existing chain or make it the head of a new chain. |
3109 | void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper, |
3110 | SmallVectorImpl<ChainUsers> &ChainUsersVec) { |
3111 | // When IVs are used as types of varying widths, they are generally converted |
3112 | // to a wider type with some uses remaining narrow under a (free) trunc. |
3113 | Value *const NextIV = getWideOperand(Oper: IVOper); |
3114 | const SCEV *const OperExpr = SE.getSCEV(V: NextIV); |
3115 | const SCEV *const OperExprBase = getExprBase(S: OperExpr); |
3116 | |
3117 | // Visit all existing chains. Check if its IVOper can be computed as a |
3118 | // profitable loop invariant increment from the last link in the Chain. |
3119 | unsigned ChainIdx = 0, NChains = IVChainVec.size(); |
3120 | const SCEV *LastIncExpr = nullptr; |
3121 | for (; ChainIdx < NChains; ++ChainIdx) { |
3122 | IVChain &Chain = IVChainVec[ChainIdx]; |
3123 | |
3124 | // Prune the solution space aggressively by checking that both IV operands |
3125 | // are expressions that operate on the same unscaled SCEVUnknown. This |
3126 | // "base" will be canceled by the subsequent getMinusSCEV call. Checking |
3127 | // first avoids creating extra SCEV expressions. |
3128 | if (!StressIVChain && Chain.ExprBase != OperExprBase) |
3129 | continue; |
3130 | |
3131 | Value *PrevIV = getWideOperand(Oper: Chain.Incs.back().IVOperand); |
3132 | if (PrevIV->getType() != NextIV->getType()) |
3133 | continue; |
3134 | |
3135 | // A phi node terminates a chain. |
3136 | if (isa<PHINode>(Val: UserInst) && isa<PHINode>(Val: Chain.tailUserInst())) |
3137 | continue; |
3138 | |
3139 | // The increment must be loop-invariant so it can be kept in a register. |
3140 | const SCEV *PrevExpr = SE.getSCEV(V: PrevIV); |
3141 | const SCEV *IncExpr = SE.getMinusSCEV(LHS: OperExpr, RHS: PrevExpr); |
3142 | if (isa<SCEVCouldNotCompute>(Val: IncExpr) || !SE.isLoopInvariant(S: IncExpr, L)) |
3143 | continue; |
3144 | |
3145 | if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) { |
3146 | LastIncExpr = IncExpr; |
3147 | break; |
3148 | } |
3149 | } |
3150 | // If we haven't found a chain, create a new one, unless we hit the max. Don't |
3151 | // bother for phi nodes, because they must be last in the chain. |
3152 | if (ChainIdx == NChains) { |
3153 | if (isa<PHINode>(Val: UserInst)) |
3154 | return; |
3155 | if (NChains >= MaxChains && !StressIVChain) { |
3156 | LLVM_DEBUG(dbgs() << "IV Chain Limit\n" ); |
3157 | return; |
3158 | } |
3159 | LastIncExpr = OperExpr; |
3160 | // IVUsers may have skipped over sign/zero extensions. We don't currently |
3161 | // attempt to form chains involving extensions unless they can be hoisted |
3162 | // into this loop's AddRec. |
3163 | if (!isa<SCEVAddRecExpr>(Val: LastIncExpr)) |
3164 | return; |
3165 | ++NChains; |
3166 | IVChainVec.push_back(Elt: IVChain(IVInc(UserInst, IVOper, LastIncExpr), |
3167 | OperExprBase)); |
3168 | ChainUsersVec.resize(N: NChains); |
3169 | LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst |
3170 | << ") IV=" << *LastIncExpr << "\n" ); |
3171 | } else { |
3172 | LLVM_DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst |
3173 | << ") IV+" << *LastIncExpr << "\n" ); |
3174 | // Add this IV user to the end of the chain. |
3175 | IVChainVec[ChainIdx].add(X: IVInc(UserInst, IVOper, LastIncExpr)); |
3176 | } |
3177 | IVChain &Chain = IVChainVec[ChainIdx]; |
3178 | |
3179 | SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers; |
3180 | // This chain's NearUsers become FarUsers. |
3181 | if (!LastIncExpr->isZero()) { |
3182 | ChainUsersVec[ChainIdx].FarUsers.insert_range(R&: NearUsers); |
3183 | NearUsers.clear(); |
3184 | } |
3185 | |
3186 | // All other uses of IVOperand become near uses of the chain. |
3187 | // We currently ignore intermediate values within SCEV expressions, assuming |
3188 | // they will eventually be used be the current chain, or can be computed |
3189 | // from one of the chain increments. To be more precise we could |
3190 | // transitively follow its user and only add leaf IV users to the set. |
3191 | for (User *U : IVOper->users()) { |
3192 | Instruction *OtherUse = dyn_cast<Instruction>(Val: U); |
3193 | if (!OtherUse) |
3194 | continue; |
3195 | // Uses in the chain will no longer be uses if the chain is formed. |
3196 | // Include the head of the chain in this iteration (not Chain.begin()). |
3197 | IVChain::const_iterator IncIter = Chain.Incs.begin(); |
3198 | IVChain::const_iterator IncEnd = Chain.Incs.end(); |
3199 | for( ; IncIter != IncEnd; ++IncIter) { |
3200 | if (IncIter->UserInst == OtherUse) |
3201 | break; |
3202 | } |
3203 | if (IncIter != IncEnd) |
3204 | continue; |
3205 | |
3206 | if (SE.isSCEVable(Ty: OtherUse->getType()) |
3207 | && !isa<SCEVUnknown>(Val: SE.getSCEV(V: OtherUse)) |
3208 | && IU.isIVUserOrOperand(Inst: OtherUse)) { |
3209 | continue; |
3210 | } |
3211 | NearUsers.insert(Ptr: OtherUse); |
3212 | } |
3213 | |
3214 | // Since this user is part of the chain, it's no longer considered a use |
3215 | // of the chain. |
3216 | ChainUsersVec[ChainIdx].FarUsers.erase(Ptr: UserInst); |
3217 | } |
3218 | |
3219 | /// Populate the vector of Chains. |
3220 | /// |
3221 | /// This decreases ILP at the architecture level. Targets with ample registers, |
3222 | /// multiple memory ports, and no register renaming probably don't want |
3223 | /// this. However, such targets should probably disable LSR altogether. |
3224 | /// |
3225 | /// The job of LSR is to make a reasonable choice of induction variables across |
3226 | /// the loop. Subsequent passes can easily "unchain" computation exposing more |
3227 | /// ILP *within the loop* if the target wants it. |
3228 | /// |
3229 | /// Finding the best IV chain is potentially a scheduling problem. Since LSR |
3230 | /// will not reorder memory operations, it will recognize this as a chain, but |
3231 | /// will generate redundant IV increments. Ideally this would be corrected later |
3232 | /// by a smart scheduler: |
3233 | /// = A[i] |
3234 | /// = A[i+x] |
3235 | /// A[i] = |
3236 | /// A[i+x] = |
3237 | /// |
3238 | /// TODO: Walk the entire domtree within this loop, not just the path to the |
3239 | /// loop latch. This will discover chains on side paths, but requires |
3240 | /// maintaining multiple copies of the Chains state. |
3241 | void LSRInstance::CollectChains() { |
3242 | LLVM_DEBUG(dbgs() << "Collecting IV Chains.\n" ); |
3243 | SmallVector<ChainUsers, 8> ChainUsersVec; |
3244 | |
3245 | SmallVector<BasicBlock *,8> LatchPath; |
3246 | BasicBlock * = L->getHeader(); |
3247 | for (DomTreeNode *Rung = DT.getNode(BB: L->getLoopLatch()); |
3248 | Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) { |
3249 | LatchPath.push_back(Elt: Rung->getBlock()); |
3250 | } |
3251 | LatchPath.push_back(Elt: LoopHeader); |
3252 | |
3253 | // Walk the instruction stream from the loop header to the loop latch. |
3254 | for (BasicBlock *BB : reverse(C&: LatchPath)) { |
3255 | for (Instruction &I : *BB) { |
3256 | // Skip instructions that weren't seen by IVUsers analysis. |
3257 | if (isa<PHINode>(Val: I) || !IU.isIVUserOrOperand(Inst: &I)) |
3258 | continue; |
3259 | |
3260 | // Ignore users that are part of a SCEV expression. This way we only |
3261 | // consider leaf IV Users. This effectively rediscovers a portion of |
3262 | // IVUsers analysis but in program order this time. |
3263 | if (SE.isSCEVable(Ty: I.getType()) && !isa<SCEVUnknown>(Val: SE.getSCEV(V: &I))) |
3264 | continue; |
3265 | |
3266 | // Remove this instruction from any NearUsers set it may be in. |
3267 | for (unsigned ChainIdx = 0, NChains = IVChainVec.size(); |
3268 | ChainIdx < NChains; ++ChainIdx) { |
3269 | ChainUsersVec[ChainIdx].NearUsers.erase(Ptr: &I); |
3270 | } |
3271 | // Search for operands that can be chained. |
3272 | SmallPtrSet<Instruction*, 4> UniqueOperands; |
3273 | User::op_iterator IVOpEnd = I.op_end(); |
3274 | User::op_iterator IVOpIter = findIVOperand(OI: I.op_begin(), OE: IVOpEnd, L, SE); |
3275 | while (IVOpIter != IVOpEnd) { |
3276 | Instruction *IVOpInst = cast<Instruction>(Val&: *IVOpIter); |
3277 | if (UniqueOperands.insert(Ptr: IVOpInst).second) |
3278 | ChainInstruction(UserInst: &I, IVOper: IVOpInst, ChainUsersVec); |
3279 | IVOpIter = findIVOperand(OI: std::next(x: IVOpIter), OE: IVOpEnd, L, SE); |
3280 | } |
3281 | } // Continue walking down the instructions. |
3282 | } // Continue walking down the domtree. |
3283 | // Visit phi backedges to determine if the chain can generate the IV postinc. |
3284 | for (PHINode &PN : L->getHeader()->phis()) { |
3285 | if (!SE.isSCEVable(Ty: PN.getType())) |
3286 | continue; |
3287 | |
3288 | Instruction *IncV = |
3289 | dyn_cast<Instruction>(Val: PN.getIncomingValueForBlock(BB: L->getLoopLatch())); |
3290 | if (IncV) |
3291 | ChainInstruction(UserInst: &PN, IVOper: IncV, ChainUsersVec); |
3292 | } |
3293 | // Remove any unprofitable chains. |
3294 | unsigned ChainIdx = 0; |
3295 | for (unsigned UsersIdx = 0, NChains = IVChainVec.size(); |
3296 | UsersIdx < NChains; ++UsersIdx) { |
3297 | if (!isProfitableChain(Chain&: IVChainVec[UsersIdx], |
3298 | Users&: ChainUsersVec[UsersIdx].FarUsers, SE, TTI)) |
3299 | continue; |
3300 | // Preserve the chain at UsesIdx. |
3301 | if (ChainIdx != UsersIdx) |
3302 | IVChainVec[ChainIdx] = IVChainVec[UsersIdx]; |
3303 | FinalizeChain(Chain&: IVChainVec[ChainIdx]); |
3304 | ++ChainIdx; |
3305 | } |
3306 | IVChainVec.resize(N: ChainIdx); |
3307 | } |
3308 | |
3309 | void LSRInstance::FinalizeChain(IVChain &Chain) { |
3310 | assert(!Chain.Incs.empty() && "empty IV chains are not allowed" ); |
3311 | LLVM_DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n" ); |
3312 | |
3313 | for (const IVInc &Inc : Chain) { |
3314 | LLVM_DEBUG(dbgs() << " Inc: " << *Inc.UserInst << "\n" ); |
3315 | auto UseI = find(Range: Inc.UserInst->operands(), Val: Inc.IVOperand); |
3316 | assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand" ); |
3317 | IVIncSet.insert(Ptr: UseI); |
3318 | } |
3319 | } |
3320 | |
3321 | /// Return true if the IVInc can be folded into an addressing mode. |
3322 | static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst, |
3323 | Value *Operand, const TargetTransformInfo &TTI) { |
3324 | const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(Val: IncExpr); |
3325 | Immediate IncOffset = Immediate::getZero(); |
3326 | if (IncConst) { |
3327 | if (IncConst && IncConst->getAPInt().getSignificantBits() > 64) |
3328 | return false; |
3329 | IncOffset = Immediate::getFixed(MinVal: IncConst->getValue()->getSExtValue()); |
3330 | } else { |
3331 | // Look for mul(vscale, constant), to detect a scalable offset. |
3332 | const APInt *C; |
3333 | if (!match(S: IncExpr, P: m_scev_Mul(Op0: m_scev_APInt(C), Op1: m_SCEVVScale())) || |
3334 | C->getSignificantBits() > 64) |
3335 | return false; |
3336 | IncOffset = Immediate::getScalable(MinVal: C->getSExtValue()); |
3337 | } |
3338 | |
3339 | if (!isAddressUse(TTI, Inst: UserInst, OperandVal: Operand)) |
3340 | return false; |
3341 | |
3342 | MemAccessTy AccessTy = getAccessType(TTI, Inst: UserInst, OperandVal: Operand); |
3343 | if (!isAlwaysFoldable(TTI, Kind: LSRUse::Address, AccessTy, /*BaseGV=*/nullptr, |
3344 | BaseOffset: IncOffset, /*HasBaseReg=*/false)) |
3345 | return false; |
3346 | |
3347 | return true; |
3348 | } |
3349 | |
3350 | /// Generate an add or subtract for each IVInc in a chain to materialize the IV |
3351 | /// user's operand from the previous IV user's operand. |
3352 | void LSRInstance::GenerateIVChain(const IVChain &Chain, |
3353 | SmallVectorImpl<WeakTrackingVH> &DeadInsts) { |
3354 | // Find the new IVOperand for the head of the chain. It may have been replaced |
3355 | // by LSR. |
3356 | const IVInc &Head = Chain.Incs[0]; |
3357 | User::op_iterator IVOpEnd = Head.UserInst->op_end(); |
3358 | // findIVOperand returns IVOpEnd if it can no longer find a valid IV user. |
3359 | User::op_iterator IVOpIter = findIVOperand(OI: Head.UserInst->op_begin(), |
3360 | OE: IVOpEnd, L, SE); |
3361 | Value *IVSrc = nullptr; |
3362 | while (IVOpIter != IVOpEnd) { |
3363 | IVSrc = getWideOperand(Oper: *IVOpIter); |
3364 | |
3365 | // If this operand computes the expression that the chain needs, we may use |
3366 | // it. (Check this after setting IVSrc which is used below.) |
3367 | // |
3368 | // Note that if Head.IncExpr is wider than IVSrc, then this phi is too |
3369 | // narrow for the chain, so we can no longer use it. We do allow using a |
3370 | // wider phi, assuming the LSR checked for free truncation. In that case we |
3371 | // should already have a truncate on this operand such that |
3372 | // getSCEV(IVSrc) == IncExpr. |
3373 | if (SE.getSCEV(V: *IVOpIter) == Head.IncExpr |
3374 | || SE.getSCEV(V: IVSrc) == Head.IncExpr) { |
3375 | break; |
3376 | } |
3377 | IVOpIter = findIVOperand(OI: std::next(x: IVOpIter), OE: IVOpEnd, L, SE); |
3378 | } |
3379 | if (IVOpIter == IVOpEnd) { |
3380 | // Gracefully give up on this chain. |
3381 | LLVM_DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n" ); |
3382 | return; |
3383 | } |
3384 | assert(IVSrc && "Failed to find IV chain source" ); |
3385 | |
3386 | LLVM_DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n" ); |
3387 | Type *IVTy = IVSrc->getType(); |
3388 | Type *IntTy = SE.getEffectiveSCEVType(Ty: IVTy); |
3389 | const SCEV *LeftOverExpr = nullptr; |
3390 | const SCEV *Accum = SE.getZero(Ty: IntTy); |
3391 | SmallVector<std::pair<const SCEV *, Value *>> Bases; |
3392 | Bases.emplace_back(Args&: Accum, Args&: IVSrc); |
3393 | |
3394 | for (const IVInc &Inc : Chain) { |
3395 | Instruction *InsertPt = Inc.UserInst; |
3396 | if (isa<PHINode>(Val: InsertPt)) |
3397 | InsertPt = L->getLoopLatch()->getTerminator(); |
3398 | |
3399 | // IVOper will replace the current IV User's operand. IVSrc is the IV |
3400 | // value currently held in a register. |
3401 | Value *IVOper = IVSrc; |
3402 | if (!Inc.IncExpr->isZero()) { |
3403 | // IncExpr was the result of subtraction of two narrow values, so must |
3404 | // be signed. |
3405 | const SCEV *IncExpr = SE.getNoopOrSignExtend(V: Inc.IncExpr, Ty: IntTy); |
3406 | Accum = SE.getAddExpr(LHS: Accum, RHS: IncExpr); |
3407 | LeftOverExpr = LeftOverExpr ? |
3408 | SE.getAddExpr(LHS: LeftOverExpr, RHS: IncExpr) : IncExpr; |
3409 | } |
3410 | |
3411 | // Look through each base to see if any can produce a nice addressing mode. |
3412 | bool FoundBase = false; |
3413 | for (auto [MapScev, MapIVOper] : reverse(C&: Bases)) { |
3414 | const SCEV *Remainder = SE.getMinusSCEV(LHS: Accum, RHS: MapScev); |
3415 | if (canFoldIVIncExpr(IncExpr: Remainder, UserInst: Inc.UserInst, Operand: Inc.IVOperand, TTI)) { |
3416 | if (!Remainder->isZero()) { |
3417 | Rewriter.clearPostInc(); |
3418 | Value *IncV = Rewriter.expandCodeFor(SH: Remainder, Ty: IntTy, I: InsertPt); |
3419 | const SCEV *IVOperExpr = |
3420 | SE.getAddExpr(LHS: SE.getUnknown(V: MapIVOper), RHS: SE.getUnknown(V: IncV)); |
3421 | IVOper = Rewriter.expandCodeFor(SH: IVOperExpr, Ty: IVTy, I: InsertPt); |
3422 | } else { |
3423 | IVOper = MapIVOper; |
3424 | } |
3425 | |
3426 | FoundBase = true; |
3427 | break; |
3428 | } |
3429 | } |
3430 | if (!FoundBase && LeftOverExpr && !LeftOverExpr->isZero()) { |
3431 | // Expand the IV increment. |
3432 | Rewriter.clearPostInc(); |
3433 | Value *IncV = Rewriter.expandCodeFor(SH: LeftOverExpr, Ty: IntTy, I: InsertPt); |
3434 | const SCEV *IVOperExpr = SE.getAddExpr(LHS: SE.getUnknown(V: IVSrc), |
3435 | RHS: SE.getUnknown(V: IncV)); |
3436 | IVOper = Rewriter.expandCodeFor(SH: IVOperExpr, Ty: IVTy, I: InsertPt); |
3437 | |
3438 | // If an IV increment can't be folded, use it as the next IV value. |
3439 | if (!canFoldIVIncExpr(IncExpr: LeftOverExpr, UserInst: Inc.UserInst, Operand: Inc.IVOperand, TTI)) { |
3440 | assert(IVTy == IVOper->getType() && "inconsistent IV increment type" ); |
3441 | Bases.emplace_back(Args&: Accum, Args&: IVOper); |
3442 | IVSrc = IVOper; |
3443 | LeftOverExpr = nullptr; |
3444 | } |
3445 | } |
3446 | Type *OperTy = Inc.IVOperand->getType(); |
3447 | if (IVTy != OperTy) { |
3448 | assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && |
3449 | "cannot extend a chained IV" ); |
3450 | IRBuilder<> Builder(InsertPt); |
3451 | IVOper = Builder.CreateTruncOrBitCast(V: IVOper, DestTy: OperTy, Name: "lsr.chain" ); |
3452 | } |
3453 | Inc.UserInst->replaceUsesOfWith(From: Inc.IVOperand, To: IVOper); |
3454 | if (auto *OperandIsInstr = dyn_cast<Instruction>(Val: Inc.IVOperand)) |
3455 | DeadInsts.emplace_back(Args&: OperandIsInstr); |
3456 | } |
3457 | // If LSR created a new, wider phi, we may also replace its postinc. We only |
3458 | // do this if we also found a wide value for the head of the chain. |
3459 | if (isa<PHINode>(Val: Chain.tailUserInst())) { |
3460 | for (PHINode &Phi : L->getHeader()->phis()) { |
3461 | if (Phi.getType() != IVSrc->getType()) |
3462 | continue; |
3463 | Instruction *PostIncV = dyn_cast<Instruction>( |
3464 | Val: Phi.getIncomingValueForBlock(BB: L->getLoopLatch())); |
3465 | if (!PostIncV || (SE.getSCEV(V: PostIncV) != SE.getSCEV(V: IVSrc))) |
3466 | continue; |
3467 | Value *IVOper = IVSrc; |
3468 | Type *PostIncTy = PostIncV->getType(); |
3469 | if (IVTy != PostIncTy) { |
3470 | assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types" ); |
3471 | IRBuilder<> Builder(L->getLoopLatch()->getTerminator()); |
3472 | Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc()); |
3473 | IVOper = Builder.CreatePointerCast(V: IVSrc, DestTy: PostIncTy, Name: "lsr.chain" ); |
3474 | } |
3475 | Phi.replaceUsesOfWith(From: PostIncV, To: IVOper); |
3476 | DeadInsts.emplace_back(Args&: PostIncV); |
3477 | } |
3478 | } |
3479 | } |
3480 | |
3481 | void LSRInstance::CollectFixupsAndInitialFormulae() { |
3482 | BranchInst *ExitBranch = nullptr; |
3483 | bool SaveCmp = TTI.canSaveCmp(L, BI: &ExitBranch, SE: &SE, LI: &LI, DT: &DT, AC: &AC, LibInfo: &TLI); |
3484 | |
3485 | // For calculating baseline cost |
3486 | SmallPtrSet<const SCEV *, 16> Regs; |
3487 | DenseSet<const SCEV *> VisitedRegs; |
3488 | DenseSet<size_t> VisitedLSRUse; |
3489 | |
3490 | for (const IVStrideUse &U : IU) { |
3491 | Instruction *UserInst = U.getUser(); |
3492 | // Skip IV users that are part of profitable IV Chains. |
3493 | User::op_iterator UseI = |
3494 | find(Range: UserInst->operands(), Val: U.getOperandValToReplace()); |
3495 | assert(UseI != UserInst->op_end() && "cannot find IV operand" ); |
3496 | if (IVIncSet.count(Ptr: UseI)) { |
3497 | LLVM_DEBUG(dbgs() << "Use is in profitable chain: " << **UseI << '\n'); |
3498 | continue; |
3499 | } |
3500 | |
3501 | LSRUse::KindType Kind = LSRUse::Basic; |
3502 | MemAccessTy AccessTy; |
3503 | if (isAddressUse(TTI, Inst: UserInst, OperandVal: U.getOperandValToReplace())) { |
3504 | Kind = LSRUse::Address; |
3505 | AccessTy = getAccessType(TTI, Inst: UserInst, OperandVal: U.getOperandValToReplace()); |
3506 | } |
3507 | |
3508 | const SCEV *S = IU.getExpr(IU: U); |
3509 | if (!S) |
3510 | continue; |
3511 | PostIncLoopSet TmpPostIncLoops = U.getPostIncLoops(); |
3512 | |
3513 | // Equality (== and !=) ICmps are special. We can rewrite (i == N) as |
3514 | // (N - i == 0), and this allows (N - i) to be the expression that we work |
3515 | // with rather than just N or i, so we can consider the register |
3516 | // requirements for both N and i at the same time. Limiting this code to |
3517 | // equality icmps is not a problem because all interesting loops use |
3518 | // equality icmps, thanks to IndVarSimplify. |
3519 | if (ICmpInst *CI = dyn_cast<ICmpInst>(Val: UserInst)) { |
3520 | // If CI can be saved in some target, like replaced inside hardware loop |
3521 | // in PowerPC, no need to generate initial formulae for it. |
3522 | if (SaveCmp && CI == dyn_cast<ICmpInst>(Val: ExitBranch->getCondition())) |
3523 | continue; |
3524 | if (CI->isEquality()) { |
3525 | // Swap the operands if needed to put the OperandValToReplace on the |
3526 | // left, for consistency. |
3527 | Value *NV = CI->getOperand(i_nocapture: 1); |
3528 | if (NV == U.getOperandValToReplace()) { |
3529 | CI->setOperand(i_nocapture: 1, Val_nocapture: CI->getOperand(i_nocapture: 0)); |
3530 | CI->setOperand(i_nocapture: 0, Val_nocapture: NV); |
3531 | NV = CI->getOperand(i_nocapture: 1); |
3532 | Changed = true; |
3533 | } |
3534 | |
3535 | // x == y --> x - y == 0 |
3536 | const SCEV *N = SE.getSCEV(V: NV); |
3537 | if (SE.isLoopInvariant(S: N, L) && Rewriter.isSafeToExpand(S: N) && |
3538 | (!NV->getType()->isPointerTy() || |
3539 | SE.getPointerBase(V: N) == SE.getPointerBase(V: S))) { |
3540 | // S is normalized, so normalize N before folding it into S |
3541 | // to keep the result normalized. |
3542 | N = normalizeForPostIncUse(S: N, Loops: TmpPostIncLoops, SE); |
3543 | if (!N) |
3544 | continue; |
3545 | Kind = LSRUse::ICmpZero; |
3546 | S = SE.getMinusSCEV(LHS: N, RHS: S); |
3547 | } else if (L->isLoopInvariant(V: NV) && |
3548 | (!isa<Instruction>(Val: NV) || |
3549 | DT.dominates(Def: cast<Instruction>(Val: NV), BB: L->getHeader())) && |
3550 | !NV->getType()->isPointerTy()) { |
3551 | // If we can't generally expand the expression (e.g. it contains |
3552 | // a divide), but it is already at a loop invariant point before the |
3553 | // loop, wrap it in an unknown (to prevent the expander from trying |
3554 | // to re-expand in a potentially unsafe way.) The restriction to |
3555 | // integer types is required because the unknown hides the base, and |
3556 | // SCEV can't compute the difference of two unknown pointers. |
3557 | N = SE.getUnknown(V: NV); |
3558 | N = normalizeForPostIncUse(S: N, Loops: TmpPostIncLoops, SE); |
3559 | if (!N) |
3560 | continue; |
3561 | Kind = LSRUse::ICmpZero; |
3562 | S = SE.getMinusSCEV(LHS: N, RHS: S); |
3563 | assert(!isa<SCEVCouldNotCompute>(S)); |
3564 | } |
3565 | |
3566 | // -1 and the negations of all interesting strides (except the negation |
3567 | // of -1) are now also interesting. |
3568 | for (size_t i = 0, e = Factors.size(); i != e; ++i) |
3569 | if (Factors[i] != -1) |
3570 | Factors.insert(X: -(uint64_t)Factors[i]); |
3571 | Factors.insert(X: -1); |
3572 | } |
3573 | } |
3574 | |
3575 | // Get or create an LSRUse. |
3576 | std::pair<size_t, Immediate> P = getUse(Expr&: S, Kind, AccessTy); |
3577 | size_t LUIdx = P.first; |
3578 | Immediate Offset = P.second; |
3579 | LSRUse &LU = Uses[LUIdx]; |
3580 | |
3581 | // Record the fixup. |
3582 | LSRFixup &LF = LU.getNewFixup(); |
3583 | LF.UserInst = UserInst; |
3584 | LF.OperandValToReplace = U.getOperandValToReplace(); |
3585 | LF.PostIncLoops = TmpPostIncLoops; |
3586 | LF.Offset = Offset; |
3587 | LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); |
3588 | |
3589 | // Create SCEV as Formula for calculating baseline cost |
3590 | if (!VisitedLSRUse.count(V: LUIdx) && !LF.isUseFullyOutsideLoop(L)) { |
3591 | Formula F; |
3592 | F.initialMatch(S, L, SE); |
3593 | BaselineCost.RateFormula(F, Regs, VisitedRegs, LU); |
3594 | VisitedLSRUse.insert(V: LUIdx); |
3595 | } |
3596 | |
3597 | if (!LU.WidestFixupType || |
3598 | SE.getTypeSizeInBits(Ty: LU.WidestFixupType) < |
3599 | SE.getTypeSizeInBits(Ty: LF.OperandValToReplace->getType())) |
3600 | LU.WidestFixupType = LF.OperandValToReplace->getType(); |
3601 | |
3602 | // If this is the first use of this LSRUse, give it a formula. |
3603 | if (LU.Formulae.empty()) { |
3604 | InsertInitialFormula(S, LU, LUIdx); |
3605 | CountRegisters(F: LU.Formulae.back(), LUIdx); |
3606 | } |
3607 | } |
3608 | |
3609 | LLVM_DEBUG(print_fixups(dbgs())); |
3610 | } |
3611 | |
3612 | /// Insert a formula for the given expression into the given use, separating out |
3613 | /// loop-variant portions from loop-invariant and loop-computable portions. |
3614 | void LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, |
3615 | size_t LUIdx) { |
3616 | // Mark uses whose expressions cannot be expanded. |
3617 | if (!Rewriter.isSafeToExpand(S)) |
3618 | LU.RigidFormula = true; |
3619 | |
3620 | Formula F; |
3621 | F.initialMatch(S, L, SE); |
3622 | bool Inserted = InsertFormula(LU, LUIdx, F); |
3623 | assert(Inserted && "Initial formula already exists!" ); (void)Inserted; |
3624 | } |
3625 | |
3626 | /// Insert a simple single-register formula for the given expression into the |
3627 | /// given use. |
3628 | void |
3629 | LSRInstance::InsertSupplementalFormula(const SCEV *S, |
3630 | LSRUse &LU, size_t LUIdx) { |
3631 | Formula F; |
3632 | F.BaseRegs.push_back(Elt: S); |
3633 | F.HasBaseReg = true; |
3634 | bool Inserted = InsertFormula(LU, LUIdx, F); |
3635 | assert(Inserted && "Supplemental formula already exists!" ); (void)Inserted; |
3636 | } |
3637 | |
3638 | /// Note which registers are used by the given formula, updating RegUses. |
3639 | void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) { |
3640 | if (F.ScaledReg) |
3641 | RegUses.countRegister(Reg: F.ScaledReg, LUIdx); |
3642 | for (const SCEV *BaseReg : F.BaseRegs) |
3643 | RegUses.countRegister(Reg: BaseReg, LUIdx); |
3644 | } |
3645 | |
3646 | /// If the given formula has not yet been inserted, add it to the list, and |
3647 | /// return true. Return false otherwise. |
3648 | bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) { |
3649 | // Do not insert formula that we will not be able to expand. |
3650 | assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && |
3651 | "Formula is illegal" ); |
3652 | |
3653 | if (!LU.InsertFormula(F, L: *L)) |
3654 | return false; |
3655 | |
3656 | CountRegisters(F, LUIdx); |
3657 | return true; |
3658 | } |
3659 | |
3660 | /// Check for other uses of loop-invariant values which we're tracking. These |
3661 | /// other uses will pin these values in registers, making them less profitable |
3662 | /// for elimination. |
3663 | /// TODO: This currently misses non-constant addrec step registers. |
3664 | /// TODO: Should this give more weight to users inside the loop? |
3665 | void |
3666 | LSRInstance::CollectLoopInvariantFixupsAndFormulae() { |
3667 | SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end()); |
3668 | SmallPtrSet<const SCEV *, 32> Visited; |
3669 | |
3670 | // Don't collect outside uses if we are favoring postinc - the instructions in |
3671 | // the loop are more important than the ones outside of it. |
3672 | if (AMK == TTI::AMK_PostIndexed) |
3673 | return; |
3674 | |
3675 | while (!Worklist.empty()) { |
3676 | const SCEV *S = Worklist.pop_back_val(); |
3677 | |
3678 | // Don't process the same SCEV twice |
3679 | if (!Visited.insert(Ptr: S).second) |
3680 | continue; |
3681 | |
3682 | if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(Val: S)) |
3683 | append_range(C&: Worklist, R: N->operands()); |
3684 | else if (const SCEVIntegralCastExpr *C = dyn_cast<SCEVIntegralCastExpr>(Val: S)) |
3685 | Worklist.push_back(Elt: C->getOperand()); |
3686 | else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(Val: S)) { |
3687 | Worklist.push_back(Elt: D->getLHS()); |
3688 | Worklist.push_back(Elt: D->getRHS()); |
3689 | } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(Val: S)) { |
3690 | const Value *V = US->getValue(); |
3691 | if (const Instruction *Inst = dyn_cast<Instruction>(Val: V)) { |
3692 | // Look for instructions defined outside the loop. |
3693 | if (L->contains(Inst)) continue; |
3694 | } else if (isa<Constant>(Val: V)) |
3695 | // Constants can be re-materialized. |
3696 | continue; |
3697 | for (const Use &U : V->uses()) { |
3698 | const Instruction *UserInst = dyn_cast<Instruction>(Val: U.getUser()); |
3699 | // Ignore non-instructions. |
3700 | if (!UserInst) |
3701 | continue; |
3702 | // Don't bother if the instruction is an EHPad. |
3703 | if (UserInst->isEHPad()) |
3704 | continue; |
3705 | // Ignore instructions in other functions (as can happen with |
3706 | // Constants). |
3707 | if (UserInst->getParent()->getParent() != L->getHeader()->getParent()) |
3708 | continue; |
3709 | // Ignore instructions not dominated by the loop. |
3710 | const BasicBlock *UseBB = !isa<PHINode>(Val: UserInst) ? |
3711 | UserInst->getParent() : |
3712 | cast<PHINode>(Val: UserInst)->getIncomingBlock( |
3713 | i: PHINode::getIncomingValueNumForOperand(i: U.getOperandNo())); |
3714 | if (!DT.dominates(A: L->getHeader(), B: UseBB)) |
3715 | continue; |
3716 | // Don't bother if the instruction is in a BB which ends in an EHPad. |
3717 | if (UseBB->getTerminator()->isEHPad()) |
3718 | continue; |
3719 | |
3720 | // Ignore cases in which the currently-examined value could come from |
3721 | // a basic block terminated with an EHPad. This checks all incoming |
3722 | // blocks of the phi node since it is possible that the same incoming |
3723 | // value comes from multiple basic blocks, only some of which may end |
3724 | // in an EHPad. If any of them do, a subsequent rewrite attempt by this |
3725 | // pass would try to insert instructions into an EHPad, hitting an |
3726 | // assertion. |
3727 | if (isa<PHINode>(Val: UserInst)) { |
3728 | const auto *PhiNode = cast<PHINode>(Val: UserInst); |
3729 | bool HasIncompatibleEHPTerminatedBlock = false; |
3730 | llvm::Value *ExpectedValue = U; |
3731 | for (unsigned int I = 0; I < PhiNode->getNumIncomingValues(); I++) { |
3732 | if (PhiNode->getIncomingValue(i: I) == ExpectedValue) { |
3733 | if (PhiNode->getIncomingBlock(i: I)->getTerminator()->isEHPad()) { |
3734 | HasIncompatibleEHPTerminatedBlock = true; |
3735 | break; |
3736 | } |
3737 | } |
3738 | } |
3739 | if (HasIncompatibleEHPTerminatedBlock) { |
3740 | continue; |
3741 | } |
3742 | } |
3743 | |
3744 | // Don't bother rewriting PHIs in catchswitch blocks. |
3745 | if (isa<CatchSwitchInst>(Val: UserInst->getParent()->getTerminator())) |
3746 | continue; |
3747 | // Ignore uses which are part of other SCEV expressions, to avoid |
3748 | // analyzing them multiple times. |
3749 | if (SE.isSCEVable(Ty: UserInst->getType())) { |
3750 | const SCEV *UserS = SE.getSCEV(V: const_cast<Instruction *>(UserInst)); |
3751 | // If the user is a no-op, look through to its uses. |
3752 | if (!isa<SCEVUnknown>(Val: UserS)) |
3753 | continue; |
3754 | if (UserS == US) { |
3755 | Worklist.push_back( |
3756 | Elt: SE.getUnknown(V: const_cast<Instruction *>(UserInst))); |
3757 | continue; |
3758 | } |
3759 | } |
3760 | // Ignore icmp instructions which are already being analyzed. |
3761 | if (const ICmpInst *ICI = dyn_cast<ICmpInst>(Val: UserInst)) { |
3762 | unsigned OtherIdx = !U.getOperandNo(); |
3763 | Value *OtherOp = const_cast<Value *>(ICI->getOperand(i_nocapture: OtherIdx)); |
3764 | if (SE.hasComputableLoopEvolution(S: SE.getSCEV(V: OtherOp), L)) |
3765 | continue; |
3766 | } |
3767 | |
3768 | std::pair<size_t, Immediate> P = |
3769 | getUse(Expr&: S, Kind: LSRUse::Basic, AccessTy: MemAccessTy()); |
3770 | size_t LUIdx = P.first; |
3771 | Immediate Offset = P.second; |
3772 | LSRUse &LU = Uses[LUIdx]; |
3773 | LSRFixup &LF = LU.getNewFixup(); |
3774 | LF.UserInst = const_cast<Instruction *>(UserInst); |
3775 | LF.OperandValToReplace = U; |
3776 | LF.Offset = Offset; |
3777 | LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); |
3778 | if (!LU.WidestFixupType || |
3779 | SE.getTypeSizeInBits(Ty: LU.WidestFixupType) < |
3780 | SE.getTypeSizeInBits(Ty: LF.OperandValToReplace->getType())) |
3781 | LU.WidestFixupType = LF.OperandValToReplace->getType(); |
3782 | InsertSupplementalFormula(S: US, LU, LUIdx); |
3783 | CountRegisters(F: LU.Formulae.back(), LUIdx: Uses.size() - 1); |
3784 | break; |
3785 | } |
3786 | } |
3787 | } |
3788 | } |
3789 | |
3790 | /// Split S into subexpressions which can be pulled out into separate |
3791 | /// registers. If C is non-null, multiply each subexpression by C. |
3792 | /// |
3793 | /// Return remainder expression after factoring the subexpressions captured by |
3794 | /// Ops. If Ops is complete, return NULL. |
3795 | static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C, |
3796 | SmallVectorImpl<const SCEV *> &Ops, |
3797 | const Loop *L, |
3798 | ScalarEvolution &SE, |
3799 | unsigned Depth = 0) { |
3800 | // Arbitrarily cap recursion to protect compile time. |
3801 | if (Depth >= 3) |
3802 | return S; |
3803 | |
3804 | if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Val: S)) { |
3805 | // Break out add operands. |
3806 | for (const SCEV *S : Add->operands()) { |
3807 | const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth: Depth+1); |
3808 | if (Remainder) |
3809 | Ops.push_back(Elt: C ? SE.getMulExpr(LHS: C, RHS: Remainder) : Remainder); |
3810 | } |
3811 | return nullptr; |
3812 | } |
3813 | const SCEV *Start, *Step; |
3814 | const SCEVConstant *Op0; |
3815 | const SCEV *Op1; |
3816 | if (match(S, P: m_scev_AffineAddRec(Op0: m_SCEV(V&: Start), Op1: m_SCEV(V&: Step)))) { |
3817 | // Split a non-zero base out of an addrec. |
3818 | if (Start->isZero()) |
3819 | return S; |
3820 | |
3821 | const SCEV *Remainder = CollectSubexprs(S: Start, C, Ops, L, SE, Depth: Depth + 1); |
3822 | // Split the non-zero AddRec unless it is part of a nested recurrence that |
3823 | // does not pertain to this loop. |
3824 | if (Remainder && (cast<SCEVAddRecExpr>(Val: S)->getLoop() == L || |
3825 | !isa<SCEVAddRecExpr>(Val: Remainder))) { |
3826 | Ops.push_back(Elt: C ? SE.getMulExpr(LHS: C, RHS: Remainder) : Remainder); |
3827 | Remainder = nullptr; |
3828 | } |
3829 | if (Remainder != Start) { |
3830 | if (!Remainder) |
3831 | Remainder = SE.getConstant(Ty: S->getType(), V: 0); |
3832 | return SE.getAddRecExpr(Start: Remainder, Step, |
3833 | L: cast<SCEVAddRecExpr>(Val: S)->getLoop(), |
3834 | // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) |
3835 | Flags: SCEV::FlagAnyWrap); |
3836 | } |
3837 | } else if (match(S, P: m_scev_Mul(Op0: m_SCEVConstant(V&: Op0), Op1: m_SCEV(V&: Op1)))) { |
3838 | // Break (C * (a + b + c)) into C*a + C*b + C*c. |
3839 | C = C ? cast<SCEVConstant>(Val: SE.getMulExpr(LHS: C, RHS: Op0)) : Op0; |
3840 | const SCEV *Remainder = CollectSubexprs(S: Op1, C, Ops, L, SE, Depth: Depth + 1); |
3841 | if (Remainder) |
3842 | Ops.push_back(Elt: SE.getMulExpr(LHS: C, RHS: Remainder)); |
3843 | return nullptr; |
3844 | } |
3845 | return S; |
3846 | } |
3847 | |
3848 | /// Return true if the SCEV represents a value that may end up as a |
3849 | /// post-increment operation. |
3850 | static bool mayUsePostIncMode(const TargetTransformInfo &TTI, |
3851 | LSRUse &LU, const SCEV *S, const Loop *L, |
3852 | ScalarEvolution &SE) { |
3853 | if (LU.Kind != LSRUse::Address || |
3854 | !LU.AccessTy.getType()->isIntOrIntVectorTy()) |
3855 | return false; |
3856 | const SCEV *Start; |
3857 | if (!match(S, P: m_scev_AffineAddRec(Op0: m_SCEV(V&: Start), Op1: m_SCEVConstant()))) |
3858 | return false; |
3859 | // Check if a post-indexed load/store can be used. |
3860 | if (TTI.isIndexedLoadLegal(Mode: TTI.MIM_PostInc, Ty: S->getType()) || |
3861 | TTI.isIndexedStoreLegal(Mode: TTI.MIM_PostInc, Ty: S->getType())) { |
3862 | if (!isa<SCEVConstant>(Val: Start) && SE.isLoopInvariant(S: Start, L)) |
3863 | return true; |
3864 | } |
3865 | return false; |
3866 | } |
3867 | |
3868 | /// Helper function for LSRInstance::GenerateReassociations. |
3869 | void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx, |
3870 | const Formula &Base, |
3871 | unsigned Depth, size_t Idx, |
3872 | bool IsScaledReg) { |
3873 | const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx]; |
3874 | // Don't generate reassociations for the base register of a value that |
3875 | // may generate a post-increment operator. The reason is that the |
3876 | // reassociations cause extra base+register formula to be created, |
3877 | // and possibly chosen, but the post-increment is more efficient. |
3878 | if (AMK == TTI::AMK_PostIndexed && mayUsePostIncMode(TTI, LU, S: BaseReg, L, SE)) |
3879 | return; |
3880 | SmallVector<const SCEV *, 8> AddOps; |
3881 | const SCEV *Remainder = CollectSubexprs(S: BaseReg, C: nullptr, Ops&: AddOps, L, SE); |
3882 | if (Remainder) |
3883 | AddOps.push_back(Elt: Remainder); |
3884 | |
3885 | if (AddOps.size() == 1) |
3886 | return; |
3887 | |
3888 | for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(), |
3889 | JE = AddOps.end(); |
3890 | J != JE; ++J) { |
3891 | // Loop-variant "unknown" values are uninteresting; we won't be able to |
3892 | // do anything meaningful with them. |
3893 | if (isa<SCEVUnknown>(Val: *J) && !SE.isLoopInvariant(S: *J, L)) |
3894 | continue; |
3895 | |
3896 | // Don't pull a constant into a register if the constant could be folded |
3897 | // into an immediate field. |
3898 | if (isAlwaysFoldable(TTI, SE, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LU.Kind, |
3899 | AccessTy: LU.AccessTy, S: *J, HasBaseReg: Base.getNumRegs() > 1)) |
3900 | continue; |
3901 | |
3902 | // Collect all operands except *J. |
3903 | SmallVector<const SCEV *, 8> InnerAddOps(std::as_const(t&: AddOps).begin(), J); |
3904 | InnerAddOps.append(in_start: std::next(x: J), in_end: std::as_const(t&: AddOps).end()); |
3905 | |
3906 | // Don't leave just a constant behind in a register if the constant could |
3907 | // be folded into an immediate field. |
3908 | if (InnerAddOps.size() == 1 && |
3909 | isAlwaysFoldable(TTI, SE, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LU.Kind, |
3910 | AccessTy: LU.AccessTy, S: InnerAddOps[0], HasBaseReg: Base.getNumRegs() > 1)) |
3911 | continue; |
3912 | |
3913 | const SCEV *InnerSum = SE.getAddExpr(Ops&: InnerAddOps); |
3914 | if (InnerSum->isZero()) |
3915 | continue; |
3916 | Formula F = Base; |
3917 | |
3918 | if (F.UnfoldedOffset.isNonZero() && F.UnfoldedOffset.isScalable()) |
3919 | continue; |
3920 | |
3921 | // Add the remaining pieces of the add back into the new formula. |
3922 | const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(Val: InnerSum); |
3923 | if (InnerSumSC && SE.getTypeSizeInBits(Ty: InnerSumSC->getType()) <= 64 && |
3924 | TTI.isLegalAddImmediate(Imm: (uint64_t)F.UnfoldedOffset.getFixedValue() + |
3925 | InnerSumSC->getValue()->getZExtValue())) { |
3926 | F.UnfoldedOffset = |
3927 | Immediate::getFixed(MinVal: (uint64_t)F.UnfoldedOffset.getFixedValue() + |
3928 | InnerSumSC->getValue()->getZExtValue()); |
3929 | if (IsScaledReg) { |
3930 | F.ScaledReg = nullptr; |
3931 | F.Scale = 0; |
3932 | } else |
3933 | F.BaseRegs.erase(CI: F.BaseRegs.begin() + Idx); |
3934 | } else if (IsScaledReg) |
3935 | F.ScaledReg = InnerSum; |
3936 | else |
3937 | F.BaseRegs[Idx] = InnerSum; |
3938 | |
3939 | // Add J as its own register, or an unfolded immediate. |
3940 | const SCEVConstant *SC = dyn_cast<SCEVConstant>(Val: *J); |
3941 | if (SC && SE.getTypeSizeInBits(Ty: SC->getType()) <= 64 && |
3942 | TTI.isLegalAddImmediate(Imm: (uint64_t)F.UnfoldedOffset.getFixedValue() + |
3943 | SC->getValue()->getZExtValue())) |
3944 | F.UnfoldedOffset = |
3945 | Immediate::getFixed(MinVal: (uint64_t)F.UnfoldedOffset.getFixedValue() + |
3946 | SC->getValue()->getZExtValue()); |
3947 | else |
3948 | F.BaseRegs.push_back(Elt: *J); |
3949 | // We may have changed the number of register in base regs, adjust the |
3950 | // formula accordingly. |
3951 | F.canonicalize(L: *L); |
3952 | |
3953 | if (InsertFormula(LU, LUIdx, F)) |
3954 | // If that formula hadn't been seen before, recurse to find more like |
3955 | // it. |
3956 | // Add check on Log16(AddOps.size()) - same as Log2_32(AddOps.size()) >> 2) |
3957 | // Because just Depth is not enough to bound compile time. |
3958 | // This means that every time AddOps.size() is greater 16^x we will add |
3959 | // x to Depth. |
3960 | GenerateReassociations(LU, LUIdx, Base: LU.Formulae.back(), |
3961 | Depth: Depth + 1 + (Log2_32(Value: AddOps.size()) >> 2)); |
3962 | } |
3963 | } |
3964 | |
3965 | /// Split out subexpressions from adds and the bases of addrecs. |
3966 | void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx, |
3967 | Formula Base, unsigned Depth) { |
3968 | assert(Base.isCanonical(*L) && "Input must be in the canonical form" ); |
3969 | // Arbitrarily cap recursion to protect compile time. |
3970 | if (Depth >= 3) |
3971 | return; |
3972 | |
3973 | for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) |
3974 | GenerateReassociationsImpl(LU, LUIdx, Base, Depth, Idx: i); |
3975 | |
3976 | if (Base.Scale == 1) |
3977 | GenerateReassociationsImpl(LU, LUIdx, Base, Depth, |
3978 | /* Idx */ -1, /* IsScaledReg */ true); |
3979 | } |
3980 | |
3981 | /// Generate a formula consisting of all of the loop-dominating registers added |
3982 | /// into a single register. |
3983 | void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx, |
3984 | Formula Base) { |
3985 | // This method is only interesting on a plurality of registers. |
3986 | if (Base.BaseRegs.size() + (Base.Scale == 1) + |
3987 | (Base.UnfoldedOffset.isNonZero()) <= |
3988 | 1) |
3989 | return; |
3990 | |
3991 | // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before |
3992 | // processing the formula. |
3993 | Base.unscale(); |
3994 | SmallVector<const SCEV *, 4> Ops; |
3995 | Formula NewBase = Base; |
3996 | NewBase.BaseRegs.clear(); |
3997 | Type *CombinedIntegerType = nullptr; |
3998 | for (const SCEV *BaseReg : Base.BaseRegs) { |
3999 | if (SE.properlyDominates(S: BaseReg, BB: L->getHeader()) && |
4000 | !SE.hasComputableLoopEvolution(S: BaseReg, L)) { |
4001 | if (!CombinedIntegerType) |
4002 | CombinedIntegerType = SE.getEffectiveSCEVType(Ty: BaseReg->getType()); |
4003 | Ops.push_back(Elt: BaseReg); |
4004 | } |
4005 | else |
4006 | NewBase.BaseRegs.push_back(Elt: BaseReg); |
4007 | } |
4008 | |
4009 | // If no register is relevant, we're done. |
4010 | if (Ops.size() == 0) |
4011 | return; |
4012 | |
4013 | // Utility function for generating the required variants of the combined |
4014 | // registers. |
4015 | auto GenerateFormula = [&](const SCEV *Sum) { |
4016 | Formula F = NewBase; |
4017 | |
4018 | // TODO: If Sum is zero, it probably means ScalarEvolution missed an |
4019 | // opportunity to fold something. For now, just ignore such cases |
4020 | // rather than proceed with zero in a register. |
4021 | if (Sum->isZero()) |
4022 | return; |
4023 | |
4024 | F.BaseRegs.push_back(Elt: Sum); |
4025 | F.canonicalize(L: *L); |
4026 | (void)InsertFormula(LU, LUIdx, F); |
4027 | }; |
4028 | |
4029 | // If we collected at least two registers, generate a formula combining them. |
4030 | if (Ops.size() > 1) { |
4031 | SmallVector<const SCEV *, 4> OpsCopy(Ops); // Don't let SE modify Ops. |
4032 | GenerateFormula(SE.getAddExpr(Ops&: OpsCopy)); |
4033 | } |
4034 | |
4035 | // If we have an unfolded offset, generate a formula combining it with the |
4036 | // registers collected. |
4037 | if (NewBase.UnfoldedOffset.isNonZero() && NewBase.UnfoldedOffset.isFixed()) { |
4038 | assert(CombinedIntegerType && "Missing a type for the unfolded offset" ); |
4039 | Ops.push_back(Elt: SE.getConstant(Ty: CombinedIntegerType, |
4040 | V: NewBase.UnfoldedOffset.getFixedValue(), isSigned: true)); |
4041 | NewBase.UnfoldedOffset = Immediate::getFixed(MinVal: 0); |
4042 | GenerateFormula(SE.getAddExpr(Ops)); |
4043 | } |
4044 | } |
4045 | |
4046 | /// Helper function for LSRInstance::GenerateSymbolicOffsets. |
4047 | void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx, |
4048 | const Formula &Base, size_t Idx, |
4049 | bool IsScaledReg) { |
4050 | const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx]; |
4051 | GlobalValue *GV = ExtractSymbol(S&: G, SE); |
4052 | if (G->isZero() || !GV) |
4053 | return; |
4054 | Formula F = Base; |
4055 | F.BaseGV = GV; |
4056 | if (!isLegalUse(TTI, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LU.Kind, AccessTy: LU.AccessTy, F)) |
4057 | return; |
4058 | if (IsScaledReg) |
4059 | F.ScaledReg = G; |
4060 | else |
4061 | F.BaseRegs[Idx] = G; |
4062 | (void)InsertFormula(LU, LUIdx, F); |
4063 | } |
4064 | |
4065 | /// Generate reuse formulae using symbolic offsets. |
4066 | void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, |
4067 | Formula Base) { |
4068 | // We can't add a symbolic offset if the address already contains one. |
4069 | if (Base.BaseGV) return; |
4070 | |
4071 | for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) |
4072 | GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, Idx: i); |
4073 | if (Base.Scale == 1) |
4074 | GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1, |
4075 | /* IsScaledReg */ true); |
4076 | } |
4077 | |
4078 | /// Helper function for LSRInstance::GenerateConstantOffsets. |
4079 | void LSRInstance::GenerateConstantOffsetsImpl( |
4080 | LSRUse &LU, unsigned LUIdx, const Formula &Base, |
4081 | const SmallVectorImpl<Immediate> &Worklist, size_t Idx, bool IsScaledReg) { |
4082 | |
4083 | auto GenerateOffset = [&](const SCEV *G, Immediate Offset) { |
4084 | Formula F = Base; |
4085 | if (!Base.BaseOffset.isCompatibleImmediate(Imm: Offset)) |
4086 | return; |
4087 | F.BaseOffset = Base.BaseOffset.subUnsigned(RHS: Offset); |
4088 | |
4089 | if (isLegalUse(TTI, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LU.Kind, AccessTy: LU.AccessTy, F)) { |
4090 | // Add the offset to the base register. |
4091 | const SCEV *NewOffset = Offset.getSCEV(SE, Ty: G->getType()); |
4092 | const SCEV *NewG = SE.getAddExpr(LHS: NewOffset, RHS: G); |
4093 | // If it cancelled out, drop the base register, otherwise update it. |
4094 | if (NewG->isZero()) { |
4095 | if (IsScaledReg) { |
4096 | F.Scale = 0; |
4097 | F.ScaledReg = nullptr; |
4098 | } else |
4099 | F.deleteBaseReg(S&: F.BaseRegs[Idx]); |
4100 | F.canonicalize(L: *L); |
4101 | } else if (IsScaledReg) |
4102 | F.ScaledReg = NewG; |
4103 | else |
4104 | F.BaseRegs[Idx] = NewG; |
4105 | |
4106 | (void)InsertFormula(LU, LUIdx, F); |
4107 | } |
4108 | }; |
4109 | |
4110 | const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx]; |
4111 | |
4112 | // With constant offsets and constant steps, we can generate pre-inc |
4113 | // accesses by having the offset equal the step. So, for access #0 with a |
4114 | // step of 8, we generate a G - 8 base which would require the first access |
4115 | // to be ((G - 8) + 8),+,8. The pre-indexed access then updates the pointer |
4116 | // for itself and hopefully becomes the base for other accesses. This means |
4117 | // means that a single pre-indexed access can be generated to become the new |
4118 | // base pointer for each iteration of the loop, resulting in no extra add/sub |
4119 | // instructions for pointer updating. |
4120 | if (AMK == TTI::AMK_PreIndexed && LU.Kind == LSRUse::Address) { |
4121 | const APInt *StepInt; |
4122 | if (match(S: G, P: m_scev_AffineAddRec(Op0: m_SCEV(), Op1: m_scev_APInt(C&: StepInt)))) { |
4123 | int64_t Step = StepInt->isNegative() ? StepInt->getSExtValue() |
4124 | : StepInt->getZExtValue(); |
4125 | |
4126 | for (Immediate Offset : Worklist) { |
4127 | if (Offset.isFixed()) { |
4128 | Offset = Immediate::getFixed(MinVal: Offset.getFixedValue() - Step); |
4129 | GenerateOffset(G, Offset); |
4130 | } |
4131 | } |
4132 | } |
4133 | } |
4134 | for (Immediate Offset : Worklist) |
4135 | GenerateOffset(G, Offset); |
4136 | |
4137 | Immediate Imm = ExtractImmediate(S&: G, SE); |
4138 | if (G->isZero() || Imm.isZero() || |
4139 | !Base.BaseOffset.isCompatibleImmediate(Imm)) |
4140 | return; |
4141 | Formula F = Base; |
4142 | F.BaseOffset = F.BaseOffset.addUnsigned(RHS: Imm); |
4143 | if (!isLegalUse(TTI, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LU.Kind, AccessTy: LU.AccessTy, F)) |
4144 | return; |
4145 | if (IsScaledReg) { |
4146 | F.ScaledReg = G; |
4147 | } else { |
4148 | F.BaseRegs[Idx] = G; |
4149 | // We may generate non canonical Formula if G is a recurrent expr reg |
4150 | // related with current loop while F.ScaledReg is not. |
4151 | F.canonicalize(L: *L); |
4152 | } |
4153 | (void)InsertFormula(LU, LUIdx, F); |
4154 | } |
4155 | |
4156 | /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets. |
4157 | void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, |
4158 | Formula Base) { |
4159 | // TODO: For now, just add the min and max offset, because it usually isn't |
4160 | // worthwhile looking at everything inbetween. |
4161 | SmallVector<Immediate, 2> Worklist; |
4162 | Worklist.push_back(Elt: LU.MinOffset); |
4163 | if (LU.MaxOffset != LU.MinOffset) |
4164 | Worklist.push_back(Elt: LU.MaxOffset); |
4165 | |
4166 | for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) |
4167 | GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, Idx: i); |
4168 | if (Base.Scale == 1) |
4169 | GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1, |
4170 | /* IsScaledReg */ true); |
4171 | } |
4172 | |
4173 | /// For ICmpZero, check to see if we can scale up the comparison. For example, x |
4174 | /// == y -> x*c == y*c. |
4175 | void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, |
4176 | Formula Base) { |
4177 | if (LU.Kind != LSRUse::ICmpZero) return; |
4178 | |
4179 | // Determine the integer type for the base formula. |
4180 | Type *IntTy = Base.getType(); |
4181 | if (!IntTy) return; |
4182 | if (SE.getTypeSizeInBits(Ty: IntTy) > 64) return; |
4183 | |
4184 | // Don't do this if there is more than one offset. |
4185 | if (LU.MinOffset != LU.MaxOffset) return; |
4186 | |
4187 | // Check if transformation is valid. It is illegal to multiply pointer. |
4188 | if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy()) |
4189 | return; |
4190 | for (const SCEV *BaseReg : Base.BaseRegs) |
4191 | if (BaseReg->getType()->isPointerTy()) |
4192 | return; |
4193 | assert(!Base.BaseGV && "ICmpZero use is not legal!" ); |
4194 | |
4195 | // Check each interesting stride. |
4196 | for (int64_t Factor : Factors) { |
4197 | // Check that Factor can be represented by IntTy |
4198 | if (!ConstantInt::isValueValidForType(Ty: IntTy, V: Factor)) |
4199 | continue; |
4200 | // Check that the multiplication doesn't overflow. |
4201 | if (Base.BaseOffset.isMin() && Factor == -1) |
4202 | continue; |
4203 | // Not supporting scalable immediates. |
4204 | if (Base.BaseOffset.isNonZero() && Base.BaseOffset.isScalable()) |
4205 | continue; |
4206 | Immediate NewBaseOffset = Base.BaseOffset.mulUnsigned(RHS: Factor); |
4207 | assert(Factor != 0 && "Zero factor not expected!" ); |
4208 | if (NewBaseOffset.getFixedValue() / Factor != |
4209 | Base.BaseOffset.getFixedValue()) |
4210 | continue; |
4211 | // If the offset will be truncated at this use, check that it is in bounds. |
4212 | if (!IntTy->isPointerTy() && |
4213 | !ConstantInt::isValueValidForType(Ty: IntTy, V: NewBaseOffset.getFixedValue())) |
4214 | continue; |
4215 | |
4216 | // Check that multiplying with the use offset doesn't overflow. |
4217 | Immediate Offset = LU.MinOffset; |
4218 | if (Offset.isMin() && Factor == -1) |
4219 | continue; |
4220 | Offset = Offset.mulUnsigned(RHS: Factor); |
4221 | if (Offset.getFixedValue() / Factor != LU.MinOffset.getFixedValue()) |
4222 | continue; |
4223 | // If the offset will be truncated at this use, check that it is in bounds. |
4224 | if (!IntTy->isPointerTy() && |
4225 | !ConstantInt::isValueValidForType(Ty: IntTy, V: Offset.getFixedValue())) |
4226 | continue; |
4227 | |
4228 | Formula F = Base; |
4229 | F.BaseOffset = NewBaseOffset; |
4230 | |
4231 | // Check that this scale is legal. |
4232 | if (!isLegalUse(TTI, MinOffset: Offset, MaxOffset: Offset, Kind: LU.Kind, AccessTy: LU.AccessTy, F)) |
4233 | continue; |
4234 | |
4235 | // Compensate for the use having MinOffset built into it. |
4236 | F.BaseOffset = F.BaseOffset.addUnsigned(RHS: Offset).subUnsigned(RHS: LU.MinOffset); |
4237 | |
4238 | const SCEV *FactorS = SE.getConstant(Ty: IntTy, V: Factor); |
4239 | |
4240 | // Check that multiplying with each base register doesn't overflow. |
4241 | for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) { |
4242 | F.BaseRegs[i] = SE.getMulExpr(LHS: F.BaseRegs[i], RHS: FactorS); |
4243 | if (getExactSDiv(LHS: F.BaseRegs[i], RHS: FactorS, SE) != Base.BaseRegs[i]) |
4244 | goto next; |
4245 | } |
4246 | |
4247 | // Check that multiplying with the scaled register doesn't overflow. |
4248 | if (F.ScaledReg) { |
4249 | F.ScaledReg = SE.getMulExpr(LHS: F.ScaledReg, RHS: FactorS); |
4250 | if (getExactSDiv(LHS: F.ScaledReg, RHS: FactorS, SE) != Base.ScaledReg) |
4251 | continue; |
4252 | } |
4253 | |
4254 | // Check that multiplying with the unfolded offset doesn't overflow. |
4255 | if (F.UnfoldedOffset.isNonZero()) { |
4256 | if (F.UnfoldedOffset.isMin() && Factor == -1) |
4257 | continue; |
4258 | F.UnfoldedOffset = F.UnfoldedOffset.mulUnsigned(RHS: Factor); |
4259 | if (F.UnfoldedOffset.getFixedValue() / Factor != |
4260 | Base.UnfoldedOffset.getFixedValue()) |
4261 | continue; |
4262 | // If the offset will be truncated, check that it is in bounds. |
4263 | if (!IntTy->isPointerTy() && !ConstantInt::isValueValidForType( |
4264 | Ty: IntTy, V: F.UnfoldedOffset.getFixedValue())) |
4265 | continue; |
4266 | } |
4267 | |
4268 | // If we make it here and it's legal, add it. |
4269 | (void)InsertFormula(LU, LUIdx, F); |
4270 | next:; |
4271 | } |
4272 | } |
4273 | |
4274 | /// Generate stride factor reuse formulae by making use of scaled-offset address |
4275 | /// modes, for example. |
4276 | void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) { |
4277 | // Determine the integer type for the base formula. |
4278 | Type *IntTy = Base.getType(); |
4279 | if (!IntTy) return; |
4280 | |
4281 | // If this Formula already has a scaled register, we can't add another one. |
4282 | // Try to unscale the formula to generate a better scale. |
4283 | if (Base.Scale != 0 && !Base.unscale()) |
4284 | return; |
4285 | |
4286 | assert(Base.Scale == 0 && "unscale did not did its job!" ); |
4287 | |
4288 | // Check each interesting stride. |
4289 | for (int64_t Factor : Factors) { |
4290 | Base.Scale = Factor; |
4291 | Base.HasBaseReg = Base.BaseRegs.size() > 1; |
4292 | // Check whether this scale is going to be legal. |
4293 | if (!isLegalUse(TTI, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LU.Kind, AccessTy: LU.AccessTy, |
4294 | F: Base)) { |
4295 | // As a special-case, handle special out-of-loop Basic users specially. |
4296 | // TODO: Reconsider this special case. |
4297 | if (LU.Kind == LSRUse::Basic && |
4298 | isLegalUse(TTI, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LSRUse::Special, |
4299 | AccessTy: LU.AccessTy, F: Base) && |
4300 | LU.AllFixupsOutsideLoop) |
4301 | LU.Kind = LSRUse::Special; |
4302 | else |
4303 | continue; |
4304 | } |
4305 | // For an ICmpZero, negating a solitary base register won't lead to |
4306 | // new solutions. |
4307 | if (LU.Kind == LSRUse::ICmpZero && !Base.HasBaseReg && |
4308 | Base.BaseOffset.isZero() && !Base.BaseGV) |
4309 | continue; |
4310 | // For each addrec base reg, if its loop is current loop, apply the scale. |
4311 | for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { |
4312 | const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: Base.BaseRegs[i]); |
4313 | if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) { |
4314 | const SCEV *FactorS = SE.getConstant(Ty: IntTy, V: Factor); |
4315 | if (FactorS->isZero()) |
4316 | continue; |
4317 | // Divide out the factor, ignoring high bits, since we'll be |
4318 | // scaling the value back up in the end. |
4319 | if (const SCEV *Quotient = getExactSDiv(LHS: AR, RHS: FactorS, SE, IgnoreSignificantBits: true)) |
4320 | if (!Quotient->isZero()) { |
4321 | // TODO: This could be optimized to avoid all the copying. |
4322 | Formula F = Base; |
4323 | F.ScaledReg = Quotient; |
4324 | F.deleteBaseReg(S&: F.BaseRegs[i]); |
4325 | // The canonical representation of 1*reg is reg, which is already in |
4326 | // Base. In that case, do not try to insert the formula, it will be |
4327 | // rejected anyway. |
4328 | if (F.Scale == 1 && (F.BaseRegs.empty() || |
4329 | (AR->getLoop() != L && LU.AllFixupsOutsideLoop))) |
4330 | continue; |
4331 | // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate |
4332 | // non canonical Formula with ScaledReg's loop not being L. |
4333 | if (F.Scale == 1 && LU.AllFixupsOutsideLoop) |
4334 | F.canonicalize(L: *L); |
4335 | (void)InsertFormula(LU, LUIdx, F); |
4336 | } |
4337 | } |
4338 | } |
4339 | } |
4340 | } |
4341 | |
4342 | /// Extend/Truncate \p Expr to \p ToTy considering post-inc uses in \p Loops. |
4343 | /// For all PostIncLoopSets in \p Loops, first de-normalize \p Expr, then |
4344 | /// perform the extension/truncate and normalize again, as the normalized form |
4345 | /// can result in folds that are not valid in the post-inc use contexts. The |
4346 | /// expressions for all PostIncLoopSets must match, otherwise return nullptr. |
4347 | static const SCEV * |
4348 | getAnyExtendConsideringPostIncUses(ArrayRef<PostIncLoopSet> Loops, |
4349 | const SCEV *Expr, Type *ToTy, |
4350 | ScalarEvolution &SE) { |
4351 | const SCEV *Result = nullptr; |
4352 | for (auto &L : Loops) { |
4353 | auto *DenormExpr = denormalizeForPostIncUse(S: Expr, Loops: L, SE); |
4354 | const SCEV *NewDenormExpr = SE.getAnyExtendExpr(Op: DenormExpr, Ty: ToTy); |
4355 | const SCEV *New = normalizeForPostIncUse(S: NewDenormExpr, Loops: L, SE); |
4356 | if (!New || (Result && New != Result)) |
4357 | return nullptr; |
4358 | Result = New; |
4359 | } |
4360 | |
4361 | assert(Result && "failed to create expression" ); |
4362 | return Result; |
4363 | } |
4364 | |
4365 | /// Generate reuse formulae from different IV types. |
4366 | void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) { |
4367 | // Don't bother truncating symbolic values. |
4368 | if (Base.BaseGV) return; |
4369 | |
4370 | // Determine the integer type for the base formula. |
4371 | Type *DstTy = Base.getType(); |
4372 | if (!DstTy) return; |
4373 | if (DstTy->isPointerTy()) |
4374 | return; |
4375 | |
4376 | // It is invalid to extend a pointer type so exit early if ScaledReg or |
4377 | // any of the BaseRegs are pointers. |
4378 | if (Base.ScaledReg && Base.ScaledReg->getType()->isPointerTy()) |
4379 | return; |
4380 | if (any_of(Range&: Base.BaseRegs, |
4381 | P: [](const SCEV *S) { return S->getType()->isPointerTy(); })) |
4382 | return; |
4383 | |
4384 | SmallVector<PostIncLoopSet> Loops; |
4385 | for (auto &LF : LU.Fixups) |
4386 | Loops.push_back(Elt: LF.PostIncLoops); |
4387 | |
4388 | for (Type *SrcTy : Types) { |
4389 | if (SrcTy != DstTy && TTI.isTruncateFree(Ty1: SrcTy, Ty2: DstTy)) { |
4390 | Formula F = Base; |
4391 | |
4392 | // Sometimes SCEV is able to prove zero during ext transform. It may |
4393 | // happen if SCEV did not do all possible transforms while creating the |
4394 | // initial node (maybe due to depth limitations), but it can do them while |
4395 | // taking ext. |
4396 | if (F.ScaledReg) { |
4397 | const SCEV *NewScaledReg = |
4398 | getAnyExtendConsideringPostIncUses(Loops, Expr: F.ScaledReg, ToTy: SrcTy, SE); |
4399 | if (!NewScaledReg || NewScaledReg->isZero()) |
4400 | continue; |
4401 | F.ScaledReg = NewScaledReg; |
4402 | } |
4403 | bool HasZeroBaseReg = false; |
4404 | for (const SCEV *&BaseReg : F.BaseRegs) { |
4405 | const SCEV *NewBaseReg = |
4406 | getAnyExtendConsideringPostIncUses(Loops, Expr: BaseReg, ToTy: SrcTy, SE); |
4407 | if (!NewBaseReg || NewBaseReg->isZero()) { |
4408 | HasZeroBaseReg = true; |
4409 | break; |
4410 | } |
4411 | BaseReg = NewBaseReg; |
4412 | } |
4413 | if (HasZeroBaseReg) |
4414 | continue; |
4415 | |
4416 | // TODO: This assumes we've done basic processing on all uses and |
4417 | // have an idea what the register usage is. |
4418 | if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses)) |
4419 | continue; |
4420 | |
4421 | F.canonicalize(L: *L); |
4422 | (void)InsertFormula(LU, LUIdx, F); |
4423 | } |
4424 | } |
4425 | } |
4426 | |
4427 | namespace { |
4428 | |
4429 | /// Helper class for GenerateCrossUseConstantOffsets. It's used to defer |
4430 | /// modifications so that the search phase doesn't have to worry about the data |
4431 | /// structures moving underneath it. |
4432 | struct WorkItem { |
4433 | size_t LUIdx; |
4434 | Immediate Imm; |
4435 | const SCEV *OrigReg; |
4436 | |
4437 | WorkItem(size_t LI, Immediate I, const SCEV *R) |
4438 | : LUIdx(LI), Imm(I), OrigReg(R) {} |
4439 | |
4440 | void print(raw_ostream &OS) const; |
4441 | void dump() const; |
4442 | }; |
4443 | |
4444 | } // end anonymous namespace |
4445 | |
4446 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
4447 | void WorkItem::print(raw_ostream &OS) const { |
4448 | OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx |
4449 | << " , add offset " << Imm; |
4450 | } |
4451 | |
4452 | LLVM_DUMP_METHOD void WorkItem::dump() const { |
4453 | print(errs()); errs() << '\n'; |
4454 | } |
4455 | #endif |
4456 | |
4457 | /// Look for registers which are a constant distance apart and try to form reuse |
4458 | /// opportunities between them. |
4459 | void LSRInstance::GenerateCrossUseConstantOffsets() { |
4460 | // Group the registers by their value without any added constant offset. |
4461 | using ImmMapTy = std::map<Immediate, const SCEV *, KeyOrderTargetImmediate>; |
4462 | |
4463 | DenseMap<const SCEV *, ImmMapTy> Map; |
4464 | DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap; |
4465 | SmallVector<const SCEV *, 8> Sequence; |
4466 | for (const SCEV *Use : RegUses) { |
4467 | const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify. |
4468 | Immediate Imm = ExtractImmediate(S&: Reg, SE); |
4469 | auto Pair = Map.try_emplace(Key: Reg); |
4470 | if (Pair.second) |
4471 | Sequence.push_back(Elt: Reg); |
4472 | Pair.first->second.insert(x: std::make_pair(x&: Imm, y&: Use)); |
4473 | UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Reg: Use); |
4474 | } |
4475 | |
4476 | // Now examine each set of registers with the same base value. Build up |
4477 | // a list of work to do and do the work in a separate step so that we're |
4478 | // not adding formulae and register counts while we're searching. |
4479 | SmallVector<WorkItem, 32> WorkItems; |
4480 | SmallSet<std::pair<size_t, Immediate>, 32, KeyOrderSizeTAndImmediate> |
4481 | UniqueItems; |
4482 | for (const SCEV *Reg : Sequence) { |
4483 | const ImmMapTy &Imms = Map.find(Val: Reg)->second; |
4484 | |
4485 | // It's not worthwhile looking for reuse if there's only one offset. |
4486 | if (Imms.size() == 1) |
4487 | continue; |
4488 | |
4489 | LLVM_DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':'; |
4490 | for (const auto &Entry |
4491 | : Imms) dbgs() |
4492 | << ' ' << Entry.first; |
4493 | dbgs() << '\n'); |
4494 | |
4495 | // Examine each offset. |
4496 | for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); |
4497 | J != JE; ++J) { |
4498 | const SCEV *OrigReg = J->second; |
4499 | |
4500 | Immediate JImm = J->first; |
4501 | const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg: OrigReg); |
4502 | |
4503 | if (!isa<SCEVConstant>(Val: OrigReg) && |
4504 | UsedByIndicesMap[Reg].count() == 1) { |
4505 | LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg |
4506 | << '\n'); |
4507 | continue; |
4508 | } |
4509 | |
4510 | // Conservatively examine offsets between this orig reg a few selected |
4511 | // other orig regs. |
4512 | Immediate First = Imms.begin()->first; |
4513 | Immediate Last = std::prev(x: Imms.end())->first; |
4514 | if (!First.isCompatibleImmediate(Imm: Last)) { |
4515 | LLVM_DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg |
4516 | << "\n" ); |
4517 | continue; |
4518 | } |
4519 | // Only scalable if both terms are scalable, or if one is scalable and |
4520 | // the other is 0. |
4521 | bool Scalable = First.isScalable() || Last.isScalable(); |
4522 | int64_t FI = First.getKnownMinValue(); |
4523 | int64_t LI = Last.getKnownMinValue(); |
4524 | // Compute (First + Last) / 2 without overflow using the fact that |
4525 | // First + Last = 2 * (First + Last) + (First ^ Last). |
4526 | int64_t Avg = (FI & LI) + ((FI ^ LI) >> 1); |
4527 | // If the result is negative and FI is odd and LI even (or vice versa), |
4528 | // we rounded towards -inf. Add 1 in that case, to round towards 0. |
4529 | Avg = Avg + ((FI ^ LI) & ((uint64_t)Avg >> 63)); |
4530 | ImmMapTy::const_iterator OtherImms[] = { |
4531 | Imms.begin(), std::prev(x: Imms.end()), |
4532 | Imms.lower_bound(x: Immediate::get(MinVal: Avg, Scalable))}; |
4533 | for (const auto &M : OtherImms) { |
4534 | if (M == J || M == JE) continue; |
4535 | if (!JImm.isCompatibleImmediate(Imm: M->first)) |
4536 | continue; |
4537 | |
4538 | // Compute the difference between the two. |
4539 | Immediate Imm = JImm.subUnsigned(RHS: M->first); |
4540 | for (unsigned LUIdx : UsedByIndices.set_bits()) |
4541 | // Make a memo of this use, offset, and register tuple. |
4542 | if (UniqueItems.insert(V: std::make_pair(x&: LUIdx, y&: Imm)).second) |
4543 | WorkItems.push_back(Elt: WorkItem(LUIdx, Imm, OrigReg)); |
4544 | } |
4545 | } |
4546 | } |
4547 | |
4548 | Map.clear(); |
4549 | Sequence.clear(); |
4550 | UsedByIndicesMap.clear(); |
4551 | UniqueItems.clear(); |
4552 | |
4553 | // Now iterate through the worklist and add new formulae. |
4554 | for (const WorkItem &WI : WorkItems) { |
4555 | size_t LUIdx = WI.LUIdx; |
4556 | LSRUse &LU = Uses[LUIdx]; |
4557 | Immediate Imm = WI.Imm; |
4558 | const SCEV *OrigReg = WI.OrigReg; |
4559 | |
4560 | Type *IntTy = SE.getEffectiveSCEVType(Ty: OrigReg->getType()); |
4561 | const SCEV *NegImmS = Imm.getNegativeSCEV(SE, Ty: IntTy); |
4562 | unsigned BitWidth = SE.getTypeSizeInBits(Ty: IntTy); |
4563 | |
4564 | // TODO: Use a more targeted data structure. |
4565 | for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) { |
4566 | Formula F = LU.Formulae[L]; |
4567 | // FIXME: The code for the scaled and unscaled registers looks |
4568 | // very similar but slightly different. Investigate if they |
4569 | // could be merged. That way, we would not have to unscale the |
4570 | // Formula. |
4571 | F.unscale(); |
4572 | // Use the immediate in the scaled register. |
4573 | if (F.ScaledReg == OrigReg) { |
4574 | if (!F.BaseOffset.isCompatibleImmediate(Imm)) |
4575 | continue; |
4576 | Immediate Offset = F.BaseOffset.addUnsigned(RHS: Imm.mulUnsigned(RHS: F.Scale)); |
4577 | // Don't create 50 + reg(-50). |
4578 | const SCEV *S = Offset.getNegativeSCEV(SE, Ty: IntTy); |
4579 | if (F.referencesReg(S)) |
4580 | continue; |
4581 | Formula NewF = F; |
4582 | NewF.BaseOffset = Offset; |
4583 | if (!isLegalUse(TTI, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, Kind: LU.Kind, AccessTy: LU.AccessTy, |
4584 | F: NewF)) |
4585 | continue; |
4586 | NewF.ScaledReg = SE.getAddExpr(LHS: NegImmS, RHS: NewF.ScaledReg); |
4587 | |
4588 | // If the new scale is a constant in a register, and adding the constant |
4589 | // value to the immediate would produce a value closer to zero than the |
4590 | // immediate itself, then the formula isn't worthwhile. |
4591 | if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Val: NewF.ScaledReg)) { |
4592 | // FIXME: Do we need to do something for scalable immediates here? |
4593 | // A scalable SCEV won't be constant, but we might still have |
4594 | // something in the offset? Bail out for now to be safe. |
4595 | if (NewF.BaseOffset.isNonZero() && NewF.BaseOffset.isScalable()) |
4596 | continue; |
4597 | if (C->getValue()->isNegative() != |
4598 | (NewF.BaseOffset.isLessThanZero()) && |
4599 | (C->getAPInt().abs() * APInt(BitWidth, F.Scale)) |
4600 | .ule(RHS: std::abs(i: NewF.BaseOffset.getFixedValue()))) |
4601 | continue; |
4602 | } |
4603 | |
4604 | // OK, looks good. |
4605 | NewF.canonicalize(L: *this->L); |
4606 | (void)InsertFormula(LU, LUIdx, F: NewF); |
4607 | } else { |
4608 | // Use the immediate in a base register. |
4609 | for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) { |
4610 | const SCEV *BaseReg = F.BaseRegs[N]; |
4611 | if (BaseReg != OrigReg) |
4612 | continue; |
4613 | Formula NewF = F; |
4614 | if (!NewF.BaseOffset.isCompatibleImmediate(Imm) || |
4615 | !NewF.UnfoldedOffset.isCompatibleImmediate(Imm) || |
4616 | !NewF.BaseOffset.isCompatibleImmediate(Imm: NewF.UnfoldedOffset)) |
4617 | continue; |
4618 | NewF.BaseOffset = NewF.BaseOffset.addUnsigned(RHS: Imm); |
4619 | if (!isLegalUse(TTI, MinOffset: LU.MinOffset, MaxOffset: LU.MaxOffset, |
4620 | Kind: LU.Kind, AccessTy: LU.AccessTy, F: NewF)) { |
4621 | if (AMK == TTI::AMK_PostIndexed && |
4622 | mayUsePostIncMode(TTI, LU, S: OrigReg, L: this->L, SE)) |
4623 | continue; |
4624 | Immediate NewUnfoldedOffset = NewF.UnfoldedOffset.addUnsigned(RHS: Imm); |
4625 | if (!isLegalAddImmediate(TTI, Offset: NewUnfoldedOffset)) |
4626 | continue; |
4627 | NewF = F; |
4628 | NewF.UnfoldedOffset = NewUnfoldedOffset; |
4629 | } |
4630 | NewF.BaseRegs[N] = SE.getAddExpr(LHS: NegImmS, RHS: BaseReg); |
4631 | |
4632 | // If the new formula has a constant in a register, and adding the |
4633 | // constant value to the immediate would produce a value closer to |
4634 | // zero than the immediate itself, then the formula isn't worthwhile. |
4635 | for (const SCEV *NewReg : NewF.BaseRegs) |
4636 | if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Val: NewReg)) { |
4637 | if (NewF.BaseOffset.isNonZero() && NewF.BaseOffset.isScalable()) |
4638 | goto skip_formula; |
4639 | if ((C->getAPInt() + NewF.BaseOffset.getFixedValue()) |
4640 | .abs() |
4641 | .slt(RHS: std::abs(i: NewF.BaseOffset.getFixedValue())) && |
4642 | (C->getAPInt() + NewF.BaseOffset.getFixedValue()) |
4643 | .countr_zero() >= |
4644 | (unsigned)llvm::countr_zero<uint64_t>( |
4645 | Val: NewF.BaseOffset.getFixedValue())) |
4646 | goto skip_formula; |
4647 | } |
4648 | |
4649 | // Ok, looks good. |
4650 | NewF.canonicalize(L: *this->L); |
4651 | (void)InsertFormula(LU, LUIdx, F: NewF); |
4652 | break; |
4653 | skip_formula:; |
4654 | } |
4655 | } |
4656 | } |
4657 | } |
4658 | } |
4659 | |
4660 | /// Generate formulae for each use. |
4661 | void |
4662 | LSRInstance::GenerateAllReuseFormulae() { |
4663 | // This is split into multiple loops so that hasRegsUsedByUsesOtherThan |
4664 | // queries are more precise. |
4665 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
4666 | LSRUse &LU = Uses[LUIdx]; |
4667 | for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) |
4668 | GenerateReassociations(LU, LUIdx, Base: LU.Formulae[i]); |
4669 | for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) |
4670 | GenerateCombinations(LU, LUIdx, Base: LU.Formulae[i]); |
4671 | } |
4672 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
4673 | LSRUse &LU = Uses[LUIdx]; |
4674 | for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) |
4675 | GenerateSymbolicOffsets(LU, LUIdx, Base: LU.Formulae[i]); |
4676 | for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) |
4677 | GenerateConstantOffsets(LU, LUIdx, Base: LU.Formulae[i]); |
4678 | for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) |
4679 | GenerateICmpZeroScales(LU, LUIdx, Base: LU.Formulae[i]); |
4680 | for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) |
4681 | GenerateScales(LU, LUIdx, Base: LU.Formulae[i]); |
4682 | } |
4683 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
4684 | LSRUse &LU = Uses[LUIdx]; |
4685 | for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) |
4686 | GenerateTruncates(LU, LUIdx, Base: LU.Formulae[i]); |
4687 | } |
4688 | |
4689 | GenerateCrossUseConstantOffsets(); |
4690 | |
4691 | LLVM_DEBUG(dbgs() << "\n" |
4692 | "After generating reuse formulae:\n" ; |
4693 | print_uses(dbgs())); |
4694 | } |
4695 | |
4696 | /// If there are multiple formulae with the same set of registers used |
4697 | /// by other uses, pick the best one and delete the others. |
4698 | void LSRInstance::FilterOutUndesirableDedicatedRegisters() { |
4699 | DenseSet<const SCEV *> VisitedRegs; |
4700 | SmallPtrSet<const SCEV *, 16> Regs; |
4701 | SmallPtrSet<const SCEV *, 16> LoserRegs; |
4702 | #ifndef NDEBUG |
4703 | bool ChangedFormulae = false; |
4704 | #endif |
4705 | |
4706 | // Collect the best formula for each unique set of shared registers. This |
4707 | // is reset for each use. |
4708 | using BestFormulaeTy = DenseMap<SmallVector<const SCEV *, 4>, size_t>; |
4709 | |
4710 | BestFormulaeTy BestFormulae; |
4711 | |
4712 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
4713 | LSRUse &LU = Uses[LUIdx]; |
4714 | LLVM_DEBUG(dbgs() << "Filtering for use " ; LU.print(dbgs()); |
4715 | dbgs() << '\n'); |
4716 | |
4717 | bool Any = false; |
4718 | for (size_t FIdx = 0, NumForms = LU.Formulae.size(); |
4719 | FIdx != NumForms; ++FIdx) { |
4720 | Formula &F = LU.Formulae[FIdx]; |
4721 | |
4722 | // Some formulas are instant losers. For example, they may depend on |
4723 | // nonexistent AddRecs from other loops. These need to be filtered |
4724 | // immediately, otherwise heuristics could choose them over others leading |
4725 | // to an unsatisfactory solution. Passing LoserRegs into RateFormula here |
4726 | // avoids the need to recompute this information across formulae using the |
4727 | // same bad AddRec. Passing LoserRegs is also essential unless we remove |
4728 | // the corresponding bad register from the Regs set. |
4729 | Cost CostF(L, SE, TTI, AMK); |
4730 | Regs.clear(); |
4731 | CostF.RateFormula(F, Regs, VisitedRegs, LU, LoserRegs: &LoserRegs); |
4732 | if (CostF.isLoser()) { |
4733 | // During initial formula generation, undesirable formulae are generated |
4734 | // by uses within other loops that have some non-trivial address mode or |
4735 | // use the postinc form of the IV. LSR needs to provide these formulae |
4736 | // as the basis of rediscovering the desired formula that uses an AddRec |
4737 | // corresponding to the existing phi. Once all formulae have been |
4738 | // generated, these initial losers may be pruned. |
4739 | LLVM_DEBUG(dbgs() << " Filtering loser " ; F.print(dbgs()); |
4740 | dbgs() << "\n" ); |
4741 | } |
4742 | else { |
4743 | SmallVector<const SCEV *, 4> Key; |
4744 | for (const SCEV *Reg : F.BaseRegs) { |
4745 | if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx)) |
4746 | Key.push_back(Elt: Reg); |
4747 | } |
4748 | if (F.ScaledReg && |
4749 | RegUses.isRegUsedByUsesOtherThan(Reg: F.ScaledReg, LUIdx)) |
4750 | Key.push_back(Elt: F.ScaledReg); |
4751 | // Unstable sort by host order ok, because this is only used for |
4752 | // uniquifying. |
4753 | llvm::sort(C&: Key); |
4754 | |
4755 | std::pair<BestFormulaeTy::const_iterator, bool> P = |
4756 | BestFormulae.insert(KV: std::make_pair(x&: Key, y&: FIdx)); |
4757 | if (P.second) |
4758 | continue; |
4759 | |
4760 | Formula &Best = LU.Formulae[P.first->second]; |
4761 | |
4762 | Cost CostBest(L, SE, TTI, AMK); |
4763 | Regs.clear(); |
4764 | CostBest.RateFormula(F: Best, Regs, VisitedRegs, LU); |
4765 | if (CostF.isLess(Other: CostBest)) |
4766 | std::swap(a&: F, b&: Best); |
4767 | LLVM_DEBUG(dbgs() << " Filtering out formula " ; F.print(dbgs()); |
4768 | dbgs() << "\n" |
4769 | " in favor of formula " ; |
4770 | Best.print(dbgs()); dbgs() << '\n'); |
4771 | } |
4772 | #ifndef NDEBUG |
4773 | ChangedFormulae = true; |
4774 | #endif |
4775 | LU.DeleteFormula(F); |
4776 | --FIdx; |
4777 | --NumForms; |
4778 | Any = true; |
4779 | } |
4780 | |
4781 | // Now that we've filtered out some formulae, recompute the Regs set. |
4782 | if (Any) |
4783 | LU.RecomputeRegs(LUIdx, RegUses); |
4784 | |
4785 | // Reset this to prepare for the next use. |
4786 | BestFormulae.clear(); |
4787 | } |
4788 | |
4789 | LLVM_DEBUG(if (ChangedFormulae) { |
4790 | dbgs() << "\n" |
4791 | "After filtering out undesirable candidates:\n" ; |
4792 | print_uses(dbgs()); |
4793 | }); |
4794 | } |
4795 | |
4796 | /// Estimate the worst-case number of solutions the solver might have to |
4797 | /// consider. It almost never considers this many solutions because it prune the |
4798 | /// search space, but the pruning isn't always sufficient. |
4799 | size_t LSRInstance::EstimateSearchSpaceComplexity() const { |
4800 | size_t Power = 1; |
4801 | for (const LSRUse &LU : Uses) { |
4802 | size_t FSize = LU.Formulae.size(); |
4803 | if (FSize >= ComplexityLimit) { |
4804 | Power = ComplexityLimit; |
4805 | break; |
4806 | } |
4807 | Power *= FSize; |
4808 | if (Power >= ComplexityLimit) |
4809 | break; |
4810 | } |
4811 | return Power; |
4812 | } |
4813 | |
4814 | /// When one formula uses a superset of the registers of another formula, it |
4815 | /// won't help reduce register pressure (though it may not necessarily hurt |
4816 | /// register pressure); remove it to simplify the system. |
4817 | void LSRInstance::NarrowSearchSpaceByDetectingSupersets() { |
4818 | if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { |
4819 | LLVM_DEBUG(dbgs() << "The search space is too complex.\n" ); |
4820 | |
4821 | LLVM_DEBUG(dbgs() << "Narrowing the search space by eliminating formulae " |
4822 | "which use a superset of registers used by other " |
4823 | "formulae.\n" ); |
4824 | |
4825 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
4826 | LSRUse &LU = Uses[LUIdx]; |
4827 | bool Any = false; |
4828 | for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { |
4829 | Formula &F = LU.Formulae[i]; |
4830 | if (F.BaseOffset.isNonZero() && F.BaseOffset.isScalable()) |
4831 | continue; |
4832 | // Look for a formula with a constant or GV in a register. If the use |
4833 | // also has a formula with that same value in an immediate field, |
4834 | // delete the one that uses a register. |
4835 | for (SmallVectorImpl<const SCEV *>::const_iterator |
4836 | I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) { |
4837 | if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Val: *I)) { |
4838 | Formula NewF = F; |
4839 | //FIXME: Formulas should store bitwidth to do wrapping properly. |
4840 | // See PR41034. |
4841 | NewF.BaseOffset = |
4842 | Immediate::getFixed(MinVal: NewF.BaseOffset.getFixedValue() + |
4843 | (uint64_t)C->getValue()->getSExtValue()); |
4844 | NewF.BaseRegs.erase(CI: NewF.BaseRegs.begin() + |
4845 | (I - F.BaseRegs.begin())); |
4846 | if (LU.HasFormulaWithSameRegs(F: NewF)) { |
4847 | LLVM_DEBUG(dbgs() << " Deleting " ; F.print(dbgs()); |
4848 | dbgs() << '\n'); |
4849 | LU.DeleteFormula(F); |
4850 | --i; |
4851 | --e; |
4852 | Any = true; |
4853 | break; |
4854 | } |
4855 | } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Val: *I)) { |
4856 | if (GlobalValue *GV = dyn_cast<GlobalValue>(Val: U->getValue())) |
4857 | if (!F.BaseGV) { |
4858 | Formula NewF = F; |
4859 | NewF.BaseGV = GV; |
4860 | NewF.BaseRegs.erase(CI: NewF.BaseRegs.begin() + |
4861 | (I - F.BaseRegs.begin())); |
4862 | if (LU.HasFormulaWithSameRegs(F: NewF)) { |
4863 | LLVM_DEBUG(dbgs() << " Deleting " ; F.print(dbgs()); |
4864 | dbgs() << '\n'); |
4865 | LU.DeleteFormula(F); |
4866 | --i; |
4867 | --e; |
4868 | Any = true; |
4869 | break; |
4870 | } |
4871 | } |
4872 | } |
4873 | } |
4874 | } |
4875 | if (Any) |
4876 | LU.RecomputeRegs(LUIdx, RegUses); |
4877 | } |
4878 | |
4879 | LLVM_DEBUG(dbgs() << "After pre-selection:\n" ; print_uses(dbgs())); |
4880 | } |
4881 | } |
4882 | |
4883 | /// When there are many registers for expressions like A, A+1, A+2, etc., |
4884 | /// allocate a single register for them. |
4885 | void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() { |
4886 | if (EstimateSearchSpaceComplexity() < ComplexityLimit) |
4887 | return; |
4888 | |
4889 | LLVM_DEBUG( |
4890 | dbgs() << "The search space is too complex.\n" |
4891 | "Narrowing the search space by assuming that uses separated " |
4892 | "by a constant offset will use the same registers.\n" ); |
4893 | |
4894 | // This is especially useful for unrolled loops. |
4895 | |
4896 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
4897 | LSRUse &LU = Uses[LUIdx]; |
4898 | for (const Formula &F : LU.Formulae) { |
4899 | if (F.BaseOffset.isZero() || (F.Scale != 0 && F.Scale != 1)) |
4900 | continue; |
4901 | |
4902 | LSRUse *LUThatHas = FindUseWithSimilarFormula(OrigF: F, OrigLU: LU); |
4903 | if (!LUThatHas) |
4904 | continue; |
4905 | |
4906 | if (!reconcileNewOffset(LU&: *LUThatHas, NewOffset: F.BaseOffset, /*HasBaseReg=*/ false, |
4907 | Kind: LU.Kind, AccessTy: LU.AccessTy)) |
4908 | continue; |
4909 | |
4910 | LLVM_DEBUG(dbgs() << " Deleting use " ; LU.print(dbgs()); dbgs() << '\n'); |
4911 | |
4912 | LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop; |
4913 | |
4914 | // Transfer the fixups of LU to LUThatHas. |
4915 | for (LSRFixup &Fixup : LU.Fixups) { |
4916 | Fixup.Offset += F.BaseOffset; |
4917 | LUThatHas->pushFixup(f&: Fixup); |
4918 | LLVM_DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n'); |
4919 | } |
4920 | |
4921 | // Delete formulae from the new use which are no longer legal. |
4922 | bool Any = false; |
4923 | for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) { |
4924 | Formula &F = LUThatHas->Formulae[i]; |
4925 | if (!isLegalUse(TTI, MinOffset: LUThatHas->MinOffset, MaxOffset: LUThatHas->MaxOffset, |
4926 | Kind: LUThatHas->Kind, AccessTy: LUThatHas->AccessTy, F)) { |
4927 | LLVM_DEBUG(dbgs() << " Deleting " ; F.print(dbgs()); dbgs() << '\n'); |
4928 | LUThatHas->DeleteFormula(F); |
4929 | --i; |
4930 | --e; |
4931 | Any = true; |
4932 | } |
4933 | } |
4934 | |
4935 | if (Any) |
4936 | LUThatHas->RecomputeRegs(LUIdx: LUThatHas - &Uses.front(), RegUses); |
4937 | |
4938 | // Delete the old use. |
4939 | DeleteUse(LU, LUIdx); |
4940 | --LUIdx; |
4941 | --NumUses; |
4942 | break; |
4943 | } |
4944 | } |
4945 | |
4946 | LLVM_DEBUG(dbgs() << "After pre-selection:\n" ; print_uses(dbgs())); |
4947 | } |
4948 | |
4949 | /// Call FilterOutUndesirableDedicatedRegisters again, if necessary, now that |
4950 | /// we've done more filtering, as it may be able to find more formulae to |
4951 | /// eliminate. |
4952 | void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){ |
4953 | if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { |
4954 | LLVM_DEBUG(dbgs() << "The search space is too complex.\n" ); |
4955 | |
4956 | LLVM_DEBUG(dbgs() << "Narrowing the search space by re-filtering out " |
4957 | "undesirable dedicated registers.\n" ); |
4958 | |
4959 | FilterOutUndesirableDedicatedRegisters(); |
4960 | |
4961 | LLVM_DEBUG(dbgs() << "After pre-selection:\n" ; print_uses(dbgs())); |
4962 | } |
4963 | } |
4964 | |
4965 | /// If a LSRUse has multiple formulae with the same ScaledReg and Scale. |
4966 | /// Pick the best one and delete the others. |
4967 | /// This narrowing heuristic is to keep as many formulae with different |
4968 | /// Scale and ScaledReg pair as possible while narrowing the search space. |
4969 | /// The benefit is that it is more likely to find out a better solution |
4970 | /// from a formulae set with more Scale and ScaledReg variations than |
4971 | /// a formulae set with the same Scale and ScaledReg. The picking winner |
4972 | /// reg heuristic will often keep the formulae with the same Scale and |
4973 | /// ScaledReg and filter others, and we want to avoid that if possible. |
4974 | void LSRInstance::NarrowSearchSpaceByFilterFormulaWithSameScaledReg() { |
4975 | if (EstimateSearchSpaceComplexity() < ComplexityLimit) |
4976 | return; |
4977 | |
4978 | LLVM_DEBUG( |
4979 | dbgs() << "The search space is too complex.\n" |
4980 | "Narrowing the search space by choosing the best Formula " |
4981 | "from the Formulae with the same Scale and ScaledReg.\n" ); |
4982 | |
4983 | // Map the "Scale * ScaledReg" pair to the best formula of current LSRUse. |
4984 | using BestFormulaeTy = DenseMap<std::pair<const SCEV *, int64_t>, size_t>; |
4985 | |
4986 | BestFormulaeTy BestFormulae; |
4987 | #ifndef NDEBUG |
4988 | bool ChangedFormulae = false; |
4989 | #endif |
4990 | DenseSet<const SCEV *> VisitedRegs; |
4991 | SmallPtrSet<const SCEV *, 16> Regs; |
4992 | |
4993 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
4994 | LSRUse &LU = Uses[LUIdx]; |
4995 | LLVM_DEBUG(dbgs() << "Filtering for use " ; LU.print(dbgs()); |
4996 | dbgs() << '\n'); |
4997 | |
4998 | // Return true if Formula FA is better than Formula FB. |
4999 | auto IsBetterThan = [&](Formula &FA, Formula &FB) { |
5000 | // First we will try to choose the Formula with fewer new registers. |
5001 | // For a register used by current Formula, the more the register is |
5002 | // shared among LSRUses, the less we increase the register number |
5003 | // counter of the formula. |
5004 | size_t FARegNum = 0; |
5005 | for (const SCEV *Reg : FA.BaseRegs) { |
5006 | const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg); |
5007 | FARegNum += (NumUses - UsedByIndices.count() + 1); |
5008 | } |
5009 | size_t FBRegNum = 0; |
5010 | for (const SCEV *Reg : FB.BaseRegs) { |
5011 | const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(Reg); |
5012 | FBRegNum += (NumUses - UsedByIndices.count() + 1); |
5013 | } |
5014 | if (FARegNum != FBRegNum) |
5015 | return FARegNum < FBRegNum; |
5016 | |
5017 | // If the new register numbers are the same, choose the Formula with |
5018 | // less Cost. |
5019 | Cost CostFA(L, SE, TTI, AMK); |
5020 | Cost CostFB(L, SE, TTI, AMK); |
5021 | Regs.clear(); |
5022 | CostFA.RateFormula(F: FA, Regs, VisitedRegs, LU); |
5023 | Regs.clear(); |
5024 | CostFB.RateFormula(F: FB, Regs, VisitedRegs, LU); |
5025 | return CostFA.isLess(Other: CostFB); |
5026 | }; |
5027 | |
5028 | bool Any = false; |
5029 | for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms; |
5030 | ++FIdx) { |
5031 | Formula &F = LU.Formulae[FIdx]; |
5032 | if (!F.ScaledReg) |
5033 | continue; |
5034 | auto P = BestFormulae.insert(KV: {{F.ScaledReg, F.Scale}, FIdx}); |
5035 | if (P.second) |
5036 | continue; |
5037 | |
5038 | Formula &Best = LU.Formulae[P.first->second]; |
5039 | if (IsBetterThan(F, Best)) |
5040 | std::swap(a&: F, b&: Best); |
5041 | LLVM_DEBUG(dbgs() << " Filtering out formula " ; F.print(dbgs()); |
5042 | dbgs() << "\n" |
5043 | " in favor of formula " ; |
5044 | Best.print(dbgs()); dbgs() << '\n'); |
5045 | #ifndef NDEBUG |
5046 | ChangedFormulae = true; |
5047 | #endif |
5048 | LU.DeleteFormula(F); |
5049 | --FIdx; |
5050 | --NumForms; |
5051 | Any = true; |
5052 | } |
5053 | if (Any) |
5054 | LU.RecomputeRegs(LUIdx, RegUses); |
5055 | |
5056 | // Reset this to prepare for the next use. |
5057 | BestFormulae.clear(); |
5058 | } |
5059 | |
5060 | LLVM_DEBUG(if (ChangedFormulae) { |
5061 | dbgs() << "\n" |
5062 | "After filtering out undesirable candidates:\n" ; |
5063 | print_uses(dbgs()); |
5064 | }); |
5065 | } |
5066 | |
5067 | /// If we are over the complexity limit, filter out any post-inc prefering |
5068 | /// variables to only post-inc values. |
5069 | void LSRInstance::NarrowSearchSpaceByFilterPostInc() { |
5070 | if (AMK != TTI::AMK_PostIndexed) |
5071 | return; |
5072 | if (EstimateSearchSpaceComplexity() < ComplexityLimit) |
5073 | return; |
5074 | |
5075 | LLVM_DEBUG(dbgs() << "The search space is too complex.\n" |
5076 | "Narrowing the search space by choosing the lowest " |
5077 | "register Formula for PostInc Uses.\n" ); |
5078 | |
5079 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
5080 | LSRUse &LU = Uses[LUIdx]; |
5081 | |
5082 | if (LU.Kind != LSRUse::Address) |
5083 | continue; |
5084 | if (!TTI.isIndexedLoadLegal(Mode: TTI.MIM_PostInc, Ty: LU.AccessTy.getType()) && |
5085 | !TTI.isIndexedStoreLegal(Mode: TTI.MIM_PostInc, Ty: LU.AccessTy.getType())) |
5086 | continue; |
5087 | |
5088 | size_t MinRegs = std::numeric_limits<size_t>::max(); |
5089 | for (const Formula &F : LU.Formulae) |
5090 | MinRegs = std::min(a: F.getNumRegs(), b: MinRegs); |
5091 | |
5092 | bool Any = false; |
5093 | for (size_t FIdx = 0, NumForms = LU.Formulae.size(); FIdx != NumForms; |
5094 | ++FIdx) { |
5095 | Formula &F = LU.Formulae[FIdx]; |
5096 | if (F.getNumRegs() > MinRegs) { |
5097 | LLVM_DEBUG(dbgs() << " Filtering out formula " ; F.print(dbgs()); |
5098 | dbgs() << "\n" ); |
5099 | LU.DeleteFormula(F); |
5100 | --FIdx; |
5101 | --NumForms; |
5102 | Any = true; |
5103 | } |
5104 | } |
5105 | if (Any) |
5106 | LU.RecomputeRegs(LUIdx, RegUses); |
5107 | |
5108 | if (EstimateSearchSpaceComplexity() < ComplexityLimit) |
5109 | break; |
5110 | } |
5111 | |
5112 | LLVM_DEBUG(dbgs() << "After pre-selection:\n" ; print_uses(dbgs())); |
5113 | } |
5114 | |
5115 | /// The function delete formulas with high registers number expectation. |
5116 | /// Assuming we don't know the value of each formula (already delete |
5117 | /// all inefficient), generate probability of not selecting for each |
5118 | /// register. |
5119 | /// For example, |
5120 | /// Use1: |
5121 | /// reg(a) + reg({0,+,1}) |
5122 | /// reg(a) + reg({-1,+,1}) + 1 |
5123 | /// reg({a,+,1}) |
5124 | /// Use2: |
5125 | /// reg(b) + reg({0,+,1}) |
5126 | /// reg(b) + reg({-1,+,1}) + 1 |
5127 | /// reg({b,+,1}) |
5128 | /// Use3: |
5129 | /// reg(c) + reg(b) + reg({0,+,1}) |
5130 | /// reg(c) + reg({b,+,1}) |
5131 | /// |
5132 | /// Probability of not selecting |
5133 | /// Use1 Use2 Use3 |
5134 | /// reg(a) (1/3) * 1 * 1 |
5135 | /// reg(b) 1 * (1/3) * (1/2) |
5136 | /// reg({0,+,1}) (2/3) * (2/3) * (1/2) |
5137 | /// reg({-1,+,1}) (2/3) * (2/3) * 1 |
5138 | /// reg({a,+,1}) (2/3) * 1 * 1 |
5139 | /// reg({b,+,1}) 1 * (2/3) * (2/3) |
5140 | /// reg(c) 1 * 1 * 0 |
5141 | /// |
5142 | /// Now count registers number mathematical expectation for each formula: |
5143 | /// Note that for each use we exclude probability if not selecting for the use. |
5144 | /// For example for Use1 probability for reg(a) would be just 1 * 1 (excluding |
5145 | /// probabilty 1/3 of not selecting for Use1). |
5146 | /// Use1: |
5147 | /// reg(a) + reg({0,+,1}) 1 + 1/3 -- to be deleted |
5148 | /// reg(a) + reg({-1,+,1}) + 1 1 + 4/9 -- to be deleted |
5149 | /// reg({a,+,1}) 1 |
5150 | /// Use2: |
5151 | /// reg(b) + reg({0,+,1}) 1/2 + 1/3 -- to be deleted |
5152 | /// reg(b) + reg({-1,+,1}) + 1 1/2 + 2/3 -- to be deleted |
5153 | /// reg({b,+,1}) 2/3 |
5154 | /// Use3: |
5155 | /// reg(c) + reg(b) + reg({0,+,1}) 1 + 1/3 + 4/9 -- to be deleted |
5156 | /// reg(c) + reg({b,+,1}) 1 + 2/3 |
5157 | void LSRInstance::NarrowSearchSpaceByDeletingCostlyFormulas() { |
5158 | if (EstimateSearchSpaceComplexity() < ComplexityLimit) |
5159 | return; |
5160 | // Ok, we have too many of formulae on our hands to conveniently handle. |
5161 | // Use a rough heuristic to thin out the list. |
5162 | |
5163 | // Set of Regs wich will be 100% used in final solution. |
5164 | // Used in each formula of a solution (in example above this is reg(c)). |
5165 | // We can skip them in calculations. |
5166 | SmallPtrSet<const SCEV *, 4> UniqRegs; |
5167 | LLVM_DEBUG(dbgs() << "The search space is too complex.\n" ); |
5168 | |
5169 | // Map each register to probability of not selecting |
5170 | DenseMap <const SCEV *, float> RegNumMap; |
5171 | for (const SCEV *Reg : RegUses) { |
5172 | if (UniqRegs.count(Ptr: Reg)) |
5173 | continue; |
5174 | float PNotSel = 1; |
5175 | for (const LSRUse &LU : Uses) { |
5176 | if (!LU.Regs.count(Ptr: Reg)) |
5177 | continue; |
5178 | float P = LU.getNotSelectedProbability(Reg); |
5179 | if (P != 0.0) |
5180 | PNotSel *= P; |
5181 | else |
5182 | UniqRegs.insert(Ptr: Reg); |
5183 | } |
5184 | RegNumMap.insert(KV: std::make_pair(x&: Reg, y&: PNotSel)); |
5185 | } |
5186 | |
5187 | LLVM_DEBUG( |
5188 | dbgs() << "Narrowing the search space by deleting costly formulas\n" ); |
5189 | |
5190 | // Delete formulas where registers number expectation is high. |
5191 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
5192 | LSRUse &LU = Uses[LUIdx]; |
5193 | // If nothing to delete - continue. |
5194 | if (LU.Formulae.size() < 2) |
5195 | continue; |
5196 | // This is temporary solution to test performance. Float should be |
5197 | // replaced with round independent type (based on integers) to avoid |
5198 | // different results for different target builds. |
5199 | float FMinRegNum = LU.Formulae[0].getNumRegs(); |
5200 | float FMinARegNum = LU.Formulae[0].getNumRegs(); |
5201 | size_t MinIdx = 0; |
5202 | for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { |
5203 | Formula &F = LU.Formulae[i]; |
5204 | float FRegNum = 0; |
5205 | float FARegNum = 0; |
5206 | for (const SCEV *BaseReg : F.BaseRegs) { |
5207 | if (UniqRegs.count(Ptr: BaseReg)) |
5208 | continue; |
5209 | FRegNum += RegNumMap[BaseReg] / LU.getNotSelectedProbability(Reg: BaseReg); |
5210 | if (isa<SCEVAddRecExpr>(Val: BaseReg)) |
5211 | FARegNum += |
5212 | RegNumMap[BaseReg] / LU.getNotSelectedProbability(Reg: BaseReg); |
5213 | } |
5214 | if (const SCEV *ScaledReg = F.ScaledReg) { |
5215 | if (!UniqRegs.count(Ptr: ScaledReg)) { |
5216 | FRegNum += |
5217 | RegNumMap[ScaledReg] / LU.getNotSelectedProbability(Reg: ScaledReg); |
5218 | if (isa<SCEVAddRecExpr>(Val: ScaledReg)) |
5219 | FARegNum += |
5220 | RegNumMap[ScaledReg] / LU.getNotSelectedProbability(Reg: ScaledReg); |
5221 | } |
5222 | } |
5223 | if (FMinRegNum > FRegNum || |
5224 | (FMinRegNum == FRegNum && FMinARegNum > FARegNum)) { |
5225 | FMinRegNum = FRegNum; |
5226 | FMinARegNum = FARegNum; |
5227 | MinIdx = i; |
5228 | } |
5229 | } |
5230 | LLVM_DEBUG(dbgs() << " The formula " ; LU.Formulae[MinIdx].print(dbgs()); |
5231 | dbgs() << " with min reg num " << FMinRegNum << '\n'); |
5232 | if (MinIdx != 0) |
5233 | std::swap(a&: LU.Formulae[MinIdx], b&: LU.Formulae[0]); |
5234 | while (LU.Formulae.size() != 1) { |
5235 | LLVM_DEBUG(dbgs() << " Deleting " ; LU.Formulae.back().print(dbgs()); |
5236 | dbgs() << '\n'); |
5237 | LU.Formulae.pop_back(); |
5238 | } |
5239 | LU.RecomputeRegs(LUIdx, RegUses); |
5240 | assert(LU.Formulae.size() == 1 && "Should be exactly 1 min regs formula" ); |
5241 | Formula &F = LU.Formulae[0]; |
5242 | LLVM_DEBUG(dbgs() << " Leaving only " ; F.print(dbgs()); dbgs() << '\n'); |
5243 | // When we choose the formula, the regs become unique. |
5244 | UniqRegs.insert_range(R&: F.BaseRegs); |
5245 | if (F.ScaledReg) |
5246 | UniqRegs.insert(Ptr: F.ScaledReg); |
5247 | } |
5248 | LLVM_DEBUG(dbgs() << "After pre-selection:\n" ; print_uses(dbgs())); |
5249 | } |
5250 | |
5251 | // Check if Best and Reg are SCEVs separated by a constant amount C, and if so |
5252 | // would the addressing offset +C would be legal where the negative offset -C is |
5253 | // not. |
5254 | static bool IsSimplerBaseSCEVForTarget(const TargetTransformInfo &TTI, |
5255 | ScalarEvolution &SE, const SCEV *Best, |
5256 | const SCEV *Reg, |
5257 | MemAccessTy AccessType) { |
5258 | if (Best->getType() != Reg->getType() || |
5259 | (isa<SCEVAddRecExpr>(Val: Best) && isa<SCEVAddRecExpr>(Val: Reg) && |
5260 | cast<SCEVAddRecExpr>(Val: Best)->getLoop() != |
5261 | cast<SCEVAddRecExpr>(Val: Reg)->getLoop())) |
5262 | return false; |
5263 | std::optional<APInt> Diff = SE.computeConstantDifference(LHS: Best, RHS: Reg); |
5264 | if (!Diff) |
5265 | return false; |
5266 | |
5267 | return TTI.isLegalAddressingMode( |
5268 | Ty: AccessType.MemTy, /*BaseGV=*/nullptr, |
5269 | /*BaseOffset=*/Diff->getSExtValue(), |
5270 | /*HasBaseReg=*/true, /*Scale=*/0, AddrSpace: AccessType.AddrSpace) && |
5271 | !TTI.isLegalAddressingMode( |
5272 | Ty: AccessType.MemTy, /*BaseGV=*/nullptr, |
5273 | /*BaseOffset=*/-Diff->getSExtValue(), |
5274 | /*HasBaseReg=*/true, /*Scale=*/0, AddrSpace: AccessType.AddrSpace); |
5275 | } |
5276 | |
5277 | /// Pick a register which seems likely to be profitable, and then in any use |
5278 | /// which has any reference to that register, delete all formulae which do not |
5279 | /// reference that register. |
5280 | void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() { |
5281 | // With all other options exhausted, loop until the system is simple |
5282 | // enough to handle. |
5283 | SmallPtrSet<const SCEV *, 4> Taken; |
5284 | while (EstimateSearchSpaceComplexity() >= ComplexityLimit) { |
5285 | // Ok, we have too many of formulae on our hands to conveniently handle. |
5286 | // Use a rough heuristic to thin out the list. |
5287 | LLVM_DEBUG(dbgs() << "The search space is too complex.\n" ); |
5288 | |
5289 | // Pick the register which is used by the most LSRUses, which is likely |
5290 | // to be a good reuse register candidate. |
5291 | const SCEV *Best = nullptr; |
5292 | unsigned BestNum = 0; |
5293 | for (const SCEV *Reg : RegUses) { |
5294 | if (Taken.count(Ptr: Reg)) |
5295 | continue; |
5296 | if (!Best) { |
5297 | Best = Reg; |
5298 | BestNum = RegUses.getUsedByIndices(Reg).count(); |
5299 | } else { |
5300 | unsigned Count = RegUses.getUsedByIndices(Reg).count(); |
5301 | if (Count > BestNum) { |
5302 | Best = Reg; |
5303 | BestNum = Count; |
5304 | } |
5305 | |
5306 | // If the scores are the same, but the Reg is simpler for the target |
5307 | // (for example {x,+,1} as opposed to {x+C,+,1}, where the target can |
5308 | // handle +C but not -C), opt for the simpler formula. |
5309 | if (Count == BestNum) { |
5310 | int LUIdx = RegUses.getUsedByIndices(Reg).find_first(); |
5311 | if (LUIdx >= 0 && Uses[LUIdx].Kind == LSRUse::Address && |
5312 | IsSimplerBaseSCEVForTarget(TTI, SE, Best, Reg, |
5313 | AccessType: Uses[LUIdx].AccessTy)) { |
5314 | Best = Reg; |
5315 | BestNum = Count; |
5316 | } |
5317 | } |
5318 | } |
5319 | } |
5320 | assert(Best && "Failed to find best LSRUse candidate" ); |
5321 | |
5322 | LLVM_DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best |
5323 | << " will yield profitable reuse.\n" ); |
5324 | Taken.insert(Ptr: Best); |
5325 | |
5326 | // In any use with formulae which references this register, delete formulae |
5327 | // which don't reference it. |
5328 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { |
5329 | LSRUse &LU = Uses[LUIdx]; |
5330 | if (!LU.Regs.count(Ptr: Best)) continue; |
5331 | |
5332 | bool Any = false; |
5333 | for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { |
5334 | Formula &F = LU.Formulae[i]; |
5335 | if (!F.referencesReg(S: Best)) { |
5336 | LLVM_DEBUG(dbgs() << " Deleting " ; F.print(dbgs()); dbgs() << '\n'); |
5337 | LU.DeleteFormula(F); |
5338 | --e; |
5339 | --i; |
5340 | Any = true; |
5341 | assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?" ); |
5342 | continue; |
5343 | } |
5344 | } |
5345 | |
5346 | if (Any) |
5347 | LU.RecomputeRegs(LUIdx, RegUses); |
5348 | } |
5349 | |
5350 | LLVM_DEBUG(dbgs() << "After pre-selection:\n" ; print_uses(dbgs())); |
5351 | } |
5352 | } |
5353 | |
5354 | /// If there are an extraordinary number of formulae to choose from, use some |
5355 | /// rough heuristics to prune down the number of formulae. This keeps the main |
5356 | /// solver from taking an extraordinary amount of time in some worst-case |
5357 | /// scenarios. |
5358 | void LSRInstance::NarrowSearchSpaceUsingHeuristics() { |
5359 | NarrowSearchSpaceByDetectingSupersets(); |
5360 | NarrowSearchSpaceByCollapsingUnrolledCode(); |
5361 | NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); |
5362 | if (FilterSameScaledReg) |
5363 | NarrowSearchSpaceByFilterFormulaWithSameScaledReg(); |
5364 | NarrowSearchSpaceByFilterPostInc(); |
5365 | if (LSRExpNarrow) |
5366 | NarrowSearchSpaceByDeletingCostlyFormulas(); |
5367 | else |
5368 | NarrowSearchSpaceByPickingWinnerRegs(); |
5369 | } |
5370 | |
5371 | /// This is the recursive solver. |
5372 | void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution, |
5373 | Cost &SolutionCost, |
5374 | SmallVectorImpl<const Formula *> &Workspace, |
5375 | const Cost &CurCost, |
5376 | const SmallPtrSet<const SCEV *, 16> &CurRegs, |
5377 | DenseSet<const SCEV *> &VisitedRegs) const { |
5378 | // Some ideas: |
5379 | // - prune more: |
5380 | // - use more aggressive filtering |
5381 | // - sort the formula so that the most profitable solutions are found first |
5382 | // - sort the uses too |
5383 | // - search faster: |
5384 | // - don't compute a cost, and then compare. compare while computing a cost |
5385 | // and bail early. |
5386 | // - track register sets with SmallBitVector |
5387 | |
5388 | const LSRUse &LU = Uses[Workspace.size()]; |
5389 | |
5390 | // If this use references any register that's already a part of the |
5391 | // in-progress solution, consider it a requirement that a formula must |
5392 | // reference that register in order to be considered. This prunes out |
5393 | // unprofitable searching. |
5394 | SmallSetVector<const SCEV *, 4> ReqRegs; |
5395 | for (const SCEV *S : CurRegs) |
5396 | if (LU.Regs.count(Ptr: S)) |
5397 | ReqRegs.insert(X: S); |
5398 | |
5399 | SmallPtrSet<const SCEV *, 16> NewRegs; |
5400 | Cost NewCost(L, SE, TTI, AMK); |
5401 | for (const Formula &F : LU.Formulae) { |
5402 | // Ignore formulae which may not be ideal in terms of register reuse of |
5403 | // ReqRegs. The formula should use all required registers before |
5404 | // introducing new ones. |
5405 | // This can sometimes (notably when trying to favour postinc) lead to |
5406 | // sub-optimial decisions. There it is best left to the cost modelling to |
5407 | // get correct. |
5408 | if (AMK != TTI::AMK_PostIndexed || LU.Kind != LSRUse::Address) { |
5409 | int NumReqRegsToFind = std::min(a: F.getNumRegs(), b: ReqRegs.size()); |
5410 | for (const SCEV *Reg : ReqRegs) { |
5411 | if ((F.ScaledReg && F.ScaledReg == Reg) || |
5412 | is_contained(Range: F.BaseRegs, Element: Reg)) { |
5413 | --NumReqRegsToFind; |
5414 | if (NumReqRegsToFind == 0) |
5415 | break; |
5416 | } |
5417 | } |
5418 | if (NumReqRegsToFind != 0) { |
5419 | // If none of the formulae satisfied the required registers, then we could |
5420 | // clear ReqRegs and try again. Currently, we simply give up in this case. |
5421 | continue; |
5422 | } |
5423 | } |
5424 | |
5425 | // Evaluate the cost of the current formula. If it's already worse than |
5426 | // the current best, prune the search at that point. |
5427 | NewCost = CurCost; |
5428 | NewRegs = CurRegs; |
5429 | NewCost.RateFormula(F, Regs&: NewRegs, VisitedRegs, LU); |
5430 | if (NewCost.isLess(Other: SolutionCost)) { |
5431 | Workspace.push_back(Elt: &F); |
5432 | if (Workspace.size() != Uses.size()) { |
5433 | SolveRecurse(Solution, SolutionCost, Workspace, CurCost: NewCost, |
5434 | CurRegs: NewRegs, VisitedRegs); |
5435 | if (F.getNumRegs() == 1 && Workspace.size() == 1) |
5436 | VisitedRegs.insert(V: F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]); |
5437 | } else { |
5438 | LLVM_DEBUG(dbgs() << "New best at " ; NewCost.print(dbgs()); |
5439 | dbgs() << ".\nRegs:\n" ; |
5440 | for (const SCEV *S : NewRegs) dbgs() |
5441 | << "- " << *S << "\n" ; |
5442 | dbgs() << '\n'); |
5443 | |
5444 | SolutionCost = NewCost; |
5445 | Solution = Workspace; |
5446 | } |
5447 | Workspace.pop_back(); |
5448 | } |
5449 | } |
5450 | } |
5451 | |
5452 | /// Choose one formula from each use. Return the results in the given Solution |
5453 | /// vector. |
5454 | void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const { |
5455 | SmallVector<const Formula *, 8> Workspace; |
5456 | Cost SolutionCost(L, SE, TTI, AMK); |
5457 | SolutionCost.Lose(); |
5458 | Cost CurCost(L, SE, TTI, AMK); |
5459 | SmallPtrSet<const SCEV *, 16> CurRegs; |
5460 | DenseSet<const SCEV *> VisitedRegs; |
5461 | Workspace.reserve(N: Uses.size()); |
5462 | |
5463 | // SolveRecurse does all the work. |
5464 | SolveRecurse(Solution, SolutionCost, Workspace, CurCost, |
5465 | CurRegs, VisitedRegs); |
5466 | if (Solution.empty()) { |
5467 | LLVM_DEBUG(dbgs() << "\nNo Satisfactory Solution\n" ); |
5468 | return; |
5469 | } |
5470 | |
5471 | // Ok, we've now made all our decisions. |
5472 | LLVM_DEBUG(dbgs() << "\n" |
5473 | "The chosen solution requires " ; |
5474 | SolutionCost.print(dbgs()); dbgs() << ":\n" ; |
5475 | for (size_t i = 0, e = Uses.size(); i != e; ++i) { |
5476 | dbgs() << " " ; |
5477 | Uses[i].print(dbgs()); |
5478 | dbgs() << "\n" |
5479 | " " ; |
5480 | Solution[i]->print(dbgs()); |
5481 | dbgs() << '\n'; |
5482 | }); |
5483 | |
5484 | assert(Solution.size() == Uses.size() && "Malformed solution!" ); |
5485 | |
5486 | const bool EnableDropUnprofitableSolution = [&] { |
5487 | switch (AllowDropSolutionIfLessProfitable) { |
5488 | case cl::BOU_TRUE: |
5489 | return true; |
5490 | case cl::BOU_FALSE: |
5491 | return false; |
5492 | case cl::BOU_UNSET: |
5493 | return TTI.shouldDropLSRSolutionIfLessProfitable(); |
5494 | } |
5495 | llvm_unreachable("Unhandled cl::boolOrDefault enum" ); |
5496 | }(); |
5497 | |
5498 | if (BaselineCost.isLess(Other: SolutionCost)) { |
5499 | if (!EnableDropUnprofitableSolution) |
5500 | LLVM_DEBUG( |
5501 | dbgs() << "Baseline is more profitable than chosen solution, " |
5502 | "add option 'lsr-drop-solution' to drop LSR solution.\n" ); |
5503 | else { |
5504 | LLVM_DEBUG(dbgs() << "Baseline is more profitable than chosen " |
5505 | "solution, dropping LSR solution.\n" ;); |
5506 | Solution.clear(); |
5507 | } |
5508 | } |
5509 | } |
5510 | |
5511 | /// Helper for AdjustInsertPositionForExpand. Climb up the dominator tree far as |
5512 | /// we can go while still being dominated by the input positions. This helps |
5513 | /// canonicalize the insert position, which encourages sharing. |
5514 | BasicBlock::iterator |
5515 | LSRInstance::HoistInsertPosition(BasicBlock::iterator IP, |
5516 | const SmallVectorImpl<Instruction *> &Inputs) |
5517 | const { |
5518 | Instruction *Tentative = &*IP; |
5519 | while (true) { |
5520 | bool AllDominate = true; |
5521 | Instruction *BetterPos = nullptr; |
5522 | // Don't bother attempting to insert before a catchswitch, their basic block |
5523 | // cannot have other non-PHI instructions. |
5524 | if (isa<CatchSwitchInst>(Val: Tentative)) |
5525 | return IP; |
5526 | |
5527 | for (Instruction *Inst : Inputs) { |
5528 | if (Inst == Tentative || !DT.dominates(Def: Inst, User: Tentative)) { |
5529 | AllDominate = false; |
5530 | break; |
5531 | } |
5532 | // Attempt to find an insert position in the middle of the block, |
5533 | // instead of at the end, so that it can be used for other expansions. |
5534 | if (Tentative->getParent() == Inst->getParent() && |
5535 | (!BetterPos || !DT.dominates(Def: Inst, User: BetterPos))) |
5536 | BetterPos = &*std::next(x: BasicBlock::iterator(Inst)); |
5537 | } |
5538 | if (!AllDominate) |
5539 | break; |
5540 | if (BetterPos) |
5541 | IP = BetterPos->getIterator(); |
5542 | else |
5543 | IP = Tentative->getIterator(); |
5544 | |
5545 | const Loop *IPLoop = LI.getLoopFor(BB: IP->getParent()); |
5546 | unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0; |
5547 | |
5548 | BasicBlock *IDom; |
5549 | for (DomTreeNode *Rung = DT.getNode(BB: IP->getParent()); ; ) { |
5550 | if (!Rung) return IP; |
5551 | Rung = Rung->getIDom(); |
5552 | if (!Rung) return IP; |
5553 | IDom = Rung->getBlock(); |
5554 | |
5555 | // Don't climb into a loop though. |
5556 | const Loop *IDomLoop = LI.getLoopFor(BB: IDom); |
5557 | unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0; |
5558 | if (IDomDepth <= IPLoopDepth && |
5559 | (IDomDepth != IPLoopDepth || IDomLoop == IPLoop)) |
5560 | break; |
5561 | } |
5562 | |
5563 | Tentative = IDom->getTerminator(); |
5564 | } |
5565 | |
5566 | return IP; |
5567 | } |
5568 | |
5569 | /// Determine an input position which will be dominated by the operands and |
5570 | /// which will dominate the result. |
5571 | BasicBlock::iterator LSRInstance::AdjustInsertPositionForExpand( |
5572 | BasicBlock::iterator LowestIP, const LSRFixup &LF, const LSRUse &LU) const { |
5573 | // Collect some instructions which must be dominated by the |
5574 | // expanding replacement. These must be dominated by any operands that |
5575 | // will be required in the expansion. |
5576 | SmallVector<Instruction *, 4> Inputs; |
5577 | if (Instruction *I = dyn_cast<Instruction>(Val: LF.OperandValToReplace)) |
5578 | Inputs.push_back(Elt: I); |
5579 | if (LU.Kind == LSRUse::ICmpZero) |
5580 | if (Instruction *I = |
5581 | dyn_cast<Instruction>(Val: cast<ICmpInst>(Val: LF.UserInst)->getOperand(i_nocapture: 1))) |
5582 | Inputs.push_back(Elt: I); |
5583 | if (LF.PostIncLoops.count(Ptr: L)) { |
5584 | if (LF.isUseFullyOutsideLoop(L)) |
5585 | Inputs.push_back(Elt: L->getLoopLatch()->getTerminator()); |
5586 | else |
5587 | Inputs.push_back(Elt: IVIncInsertPos); |
5588 | } |
5589 | // The expansion must also be dominated by the increment positions of any |
5590 | // loops it for which it is using post-inc mode. |
5591 | for (const Loop *PIL : LF.PostIncLoops) { |
5592 | if (PIL == L) continue; |
5593 | |
5594 | // Be dominated by the loop exit. |
5595 | SmallVector<BasicBlock *, 4> ExitingBlocks; |
5596 | PIL->getExitingBlocks(ExitingBlocks); |
5597 | if (!ExitingBlocks.empty()) { |
5598 | BasicBlock *BB = ExitingBlocks[0]; |
5599 | for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i) |
5600 | BB = DT.findNearestCommonDominator(A: BB, B: ExitingBlocks[i]); |
5601 | Inputs.push_back(Elt: BB->getTerminator()); |
5602 | } |
5603 | } |
5604 | |
5605 | assert(!isa<PHINode>(LowestIP) && !LowestIP->isEHPad() && |
5606 | "Insertion point must be a normal instruction" ); |
5607 | |
5608 | // Then, climb up the immediate dominator tree as far as we can go while |
5609 | // still being dominated by the input positions. |
5610 | BasicBlock::iterator IP = HoistInsertPosition(IP: LowestIP, Inputs); |
5611 | |
5612 | // Don't insert instructions before PHI nodes. |
5613 | while (isa<PHINode>(Val: IP)) ++IP; |
5614 | |
5615 | // Ignore landingpad instructions. |
5616 | while (IP->isEHPad()) ++IP; |
5617 | |
5618 | // Set IP below instructions recently inserted by SCEVExpander. This keeps the |
5619 | // IP consistent across expansions and allows the previously inserted |
5620 | // instructions to be reused by subsequent expansion. |
5621 | while (Rewriter.isInsertedInstruction(I: &*IP) && IP != LowestIP) |
5622 | ++IP; |
5623 | |
5624 | return IP; |
5625 | } |
5626 | |
5627 | /// Emit instructions for the leading candidate expression for this LSRUse (this |
5628 | /// is called "expanding"). |
5629 | Value *LSRInstance::Expand(const LSRUse &LU, const LSRFixup &LF, |
5630 | const Formula &F, BasicBlock::iterator IP, |
5631 | SmallVectorImpl<WeakTrackingVH> &DeadInsts) const { |
5632 | if (LU.RigidFormula) |
5633 | return LF.OperandValToReplace; |
5634 | |
5635 | // Determine an input position which will be dominated by the operands and |
5636 | // which will dominate the result. |
5637 | IP = AdjustInsertPositionForExpand(LowestIP: IP, LF, LU); |
5638 | Rewriter.setInsertPoint(&*IP); |
5639 | |
5640 | // Inform the Rewriter if we have a post-increment use, so that it can |
5641 | // perform an advantageous expansion. |
5642 | Rewriter.setPostInc(LF.PostIncLoops); |
5643 | |
5644 | // This is the type that the user actually needs. |
5645 | Type *OpTy = LF.OperandValToReplace->getType(); |
5646 | // This will be the type that we'll initially expand to. |
5647 | Type *Ty = F.getType(); |
5648 | if (!Ty) |
5649 | // No type known; just expand directly to the ultimate type. |
5650 | Ty = OpTy; |
5651 | else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(Ty: OpTy)) |
5652 | // Expand directly to the ultimate type if it's the right size. |
5653 | Ty = OpTy; |
5654 | // This is the type to do integer arithmetic in. |
5655 | Type *IntTy = SE.getEffectiveSCEVType(Ty); |
5656 | |
5657 | // Build up a list of operands to add together to form the full base. |
5658 | SmallVector<const SCEV *, 8> Ops; |
5659 | |
5660 | // Expand the BaseRegs portion. |
5661 | for (const SCEV *Reg : F.BaseRegs) { |
5662 | assert(!Reg->isZero() && "Zero allocated in a base register!" ); |
5663 | |
5664 | // If we're expanding for a post-inc user, make the post-inc adjustment. |
5665 | Reg = denormalizeForPostIncUse(S: Reg, Loops: LF.PostIncLoops, SE); |
5666 | Ops.push_back(Elt: SE.getUnknown(V: Rewriter.expandCodeFor(SH: Reg, Ty: nullptr))); |
5667 | } |
5668 | |
5669 | // Expand the ScaledReg portion. |
5670 | Value *ICmpScaledV = nullptr; |
5671 | if (F.Scale != 0) { |
5672 | const SCEV *ScaledS = F.ScaledReg; |
5673 | |
5674 | // If we're expanding for a post-inc user, make the post-inc adjustment. |
5675 | PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); |
5676 | ScaledS = denormalizeForPostIncUse(S: ScaledS, Loops, SE); |
5677 | |
5678 | if (LU.Kind == LSRUse::ICmpZero) { |
5679 | // Expand ScaleReg as if it was part of the base regs. |
5680 | if (F.Scale == 1) |
5681 | Ops.push_back( |
5682 | Elt: SE.getUnknown(V: Rewriter.expandCodeFor(SH: ScaledS, Ty: nullptr))); |
5683 | else { |
5684 | // An interesting way of "folding" with an icmp is to use a negated |
5685 | // scale, which we'll implement by inserting it into the other operand |
5686 | // of the icmp. |
5687 | assert(F.Scale == -1 && |
5688 | "The only scale supported by ICmpZero uses is -1!" ); |
5689 | ICmpScaledV = Rewriter.expandCodeFor(SH: ScaledS, Ty: nullptr); |
5690 | } |
5691 | } else { |
5692 | // Otherwise just expand the scaled register and an explicit scale, |
5693 | // which is expected to be matched as part of the address. |
5694 | |
5695 | // Flush the operand list to suppress SCEVExpander hoisting address modes. |
5696 | // Unless the addressing mode will not be folded. |
5697 | if (!Ops.empty() && LU.Kind == LSRUse::Address && |
5698 | isAMCompletelyFolded(TTI, LU, F)) { |
5699 | Value *FullV = Rewriter.expandCodeFor(SH: SE.getAddExpr(Ops), Ty: nullptr); |
5700 | Ops.clear(); |
5701 | Ops.push_back(Elt: SE.getUnknown(V: FullV)); |
5702 | } |
5703 | ScaledS = SE.getUnknown(V: Rewriter.expandCodeFor(SH: ScaledS, Ty: nullptr)); |
5704 | if (F.Scale != 1) |
5705 | ScaledS = |
5706 | SE.getMulExpr(LHS: ScaledS, RHS: SE.getConstant(Ty: ScaledS->getType(), V: F.Scale)); |
5707 | Ops.push_back(Elt: ScaledS); |
5708 | } |
5709 | } |
5710 | |
5711 | // Expand the GV portion. |
5712 | if (F.BaseGV) { |
5713 | // Flush the operand list to suppress SCEVExpander hoisting. |
5714 | if (!Ops.empty()) { |
5715 | Value *FullV = Rewriter.expandCodeFor(SH: SE.getAddExpr(Ops), Ty: IntTy); |
5716 | Ops.clear(); |
5717 | Ops.push_back(Elt: SE.getUnknown(V: FullV)); |
5718 | } |
5719 | Ops.push_back(Elt: SE.getUnknown(V: F.BaseGV)); |
5720 | } |
5721 | |
5722 | // Flush the operand list to suppress SCEVExpander hoisting of both folded and |
5723 | // unfolded offsets. LSR assumes they both live next to their uses. |
5724 | if (!Ops.empty()) { |
5725 | Value *FullV = Rewriter.expandCodeFor(SH: SE.getAddExpr(Ops), Ty); |
5726 | Ops.clear(); |
5727 | Ops.push_back(Elt: SE.getUnknown(V: FullV)); |
5728 | } |
5729 | |
5730 | // FIXME: Are we sure we won't get a mismatch here? Is there a way to bail |
5731 | // out at this point, or should we generate a SCEV adding together mixed |
5732 | // offsets? |
5733 | assert(F.BaseOffset.isCompatibleImmediate(LF.Offset) && |
5734 | "Expanding mismatched offsets\n" ); |
5735 | // Expand the immediate portion. |
5736 | Immediate Offset = F.BaseOffset.addUnsigned(RHS: LF.Offset); |
5737 | if (Offset.isNonZero()) { |
5738 | if (LU.Kind == LSRUse::ICmpZero) { |
5739 | // The other interesting way of "folding" with an ICmpZero is to use a |
5740 | // negated immediate. |
5741 | if (!ICmpScaledV) |
5742 | ICmpScaledV = |
5743 | ConstantInt::get(Ty: IntTy, V: -(uint64_t)Offset.getFixedValue()); |
5744 | else { |
5745 | Ops.push_back(Elt: SE.getUnknown(V: ICmpScaledV)); |
5746 | ICmpScaledV = ConstantInt::get(Ty: IntTy, V: Offset.getFixedValue()); |
5747 | } |
5748 | } else { |
5749 | // Just add the immediate values. These again are expected to be matched |
5750 | // as part of the address. |
5751 | Ops.push_back(Elt: Offset.getUnknownSCEV(SE, Ty: IntTy)); |
5752 | } |
5753 | } |
5754 | |
5755 | // Expand the unfolded offset portion. |
5756 | Immediate UnfoldedOffset = F.UnfoldedOffset; |
5757 | if (UnfoldedOffset.isNonZero()) { |
5758 | // Just add the immediate values. |
5759 | Ops.push_back(Elt: UnfoldedOffset.getUnknownSCEV(SE, Ty: IntTy)); |
5760 | } |
5761 | |
5762 | // Emit instructions summing all the operands. |
5763 | const SCEV *FullS = Ops.empty() ? |
5764 | SE.getConstant(Ty: IntTy, V: 0) : |
5765 | SE.getAddExpr(Ops); |
5766 | Value *FullV = Rewriter.expandCodeFor(SH: FullS, Ty); |
5767 | |
5768 | // We're done expanding now, so reset the rewriter. |
5769 | Rewriter.clearPostInc(); |
5770 | |
5771 | // An ICmpZero Formula represents an ICmp which we're handling as a |
5772 | // comparison against zero. Now that we've expanded an expression for that |
5773 | // form, update the ICmp's other operand. |
5774 | if (LU.Kind == LSRUse::ICmpZero) { |
5775 | ICmpInst *CI = cast<ICmpInst>(Val: LF.UserInst); |
5776 | if (auto *OperandIsInstr = dyn_cast<Instruction>(Val: CI->getOperand(i_nocapture: 1))) |
5777 | DeadInsts.emplace_back(Args&: OperandIsInstr); |
5778 | assert(!F.BaseGV && "ICmp does not support folding a global value and " |
5779 | "a scale at the same time!" ); |
5780 | if (F.Scale == -1) { |
5781 | if (ICmpScaledV->getType() != OpTy) { |
5782 | Instruction *Cast = CastInst::Create( |
5783 | CastInst::getCastOpcode(Val: ICmpScaledV, SrcIsSigned: false, Ty: OpTy, DstIsSigned: false), |
5784 | S: ICmpScaledV, Ty: OpTy, Name: "tmp" , InsertBefore: CI->getIterator()); |
5785 | ICmpScaledV = Cast; |
5786 | } |
5787 | CI->setOperand(i_nocapture: 1, Val_nocapture: ICmpScaledV); |
5788 | } else { |
5789 | // A scale of 1 means that the scale has been expanded as part of the |
5790 | // base regs. |
5791 | assert((F.Scale == 0 || F.Scale == 1) && |
5792 | "ICmp does not support folding a global value and " |
5793 | "a scale at the same time!" ); |
5794 | Constant *C = ConstantInt::getSigned(Ty: SE.getEffectiveSCEVType(Ty: OpTy), |
5795 | V: -(uint64_t)Offset.getFixedValue()); |
5796 | if (C->getType() != OpTy) { |
5797 | C = ConstantFoldCastOperand( |
5798 | Opcode: CastInst::getCastOpcode(Val: C, SrcIsSigned: false, Ty: OpTy, DstIsSigned: false), C, DestTy: OpTy, |
5799 | DL: CI->getDataLayout()); |
5800 | assert(C && "Cast of ConstantInt should have folded" ); |
5801 | } |
5802 | |
5803 | CI->setOperand(i_nocapture: 1, Val_nocapture: C); |
5804 | } |
5805 | } |
5806 | |
5807 | return FullV; |
5808 | } |
5809 | |
5810 | /// Helper for Rewrite. PHI nodes are special because the use of their operands |
5811 | /// effectively happens in their predecessor blocks, so the expression may need |
5812 | /// to be expanded in multiple places. |
5813 | void LSRInstance::RewriteForPHI(PHINode *PN, const LSRUse &LU, |
5814 | const LSRFixup &LF, const Formula &F, |
5815 | SmallVectorImpl<WeakTrackingVH> &DeadInsts) { |
5816 | DenseMap<BasicBlock *, Value *> Inserted; |
5817 | |
5818 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) |
5819 | if (PN->getIncomingValue(i) == LF.OperandValToReplace) { |
5820 | bool needUpdateFixups = false; |
5821 | BasicBlock *BB = PN->getIncomingBlock(i); |
5822 | |
5823 | // If this is a critical edge, split the edge so that we do not insert |
5824 | // the code on all predecessor/successor paths. We do this unless this |
5825 | // is the canonical backedge for this loop, which complicates post-inc |
5826 | // users. |
5827 | if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 && |
5828 | !isa<IndirectBrInst>(Val: BB->getTerminator()) && |
5829 | !isa<CatchSwitchInst>(Val: BB->getTerminator())) { |
5830 | BasicBlock *Parent = PN->getParent(); |
5831 | Loop *PNLoop = LI.getLoopFor(BB: Parent); |
5832 | if (!PNLoop || Parent != PNLoop->getHeader()) { |
5833 | // Split the critical edge. |
5834 | BasicBlock *NewBB = nullptr; |
5835 | if (!Parent->isLandingPad()) { |
5836 | NewBB = |
5837 | SplitCriticalEdge(Src: BB, Dst: Parent, |
5838 | Options: CriticalEdgeSplittingOptions(&DT, &LI, MSSAU) |
5839 | .setMergeIdenticalEdges() |
5840 | .setKeepOneInputPHIs()); |
5841 | } else { |
5842 | SmallVector<BasicBlock*, 2> NewBBs; |
5843 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); |
5844 | SplitLandingPadPredecessors(OrigBB: Parent, Preds: BB, Suffix: "" , Suffix2: "" , NewBBs, DTU: &DTU, LI: &LI); |
5845 | NewBB = NewBBs[0]; |
5846 | } |
5847 | // If NewBB==NULL, then SplitCriticalEdge refused to split because all |
5848 | // phi predecessors are identical. The simple thing to do is skip |
5849 | // splitting in this case rather than complicate the API. |
5850 | if (NewBB) { |
5851 | // If PN is outside of the loop and BB is in the loop, we want to |
5852 | // move the block to be immediately before the PHI block, not |
5853 | // immediately after BB. |
5854 | if (L->contains(BB) && !L->contains(Inst: PN)) |
5855 | NewBB->moveBefore(MovePos: PN->getParent()); |
5856 | |
5857 | // Splitting the edge can reduce the number of PHI entries we have. |
5858 | e = PN->getNumIncomingValues(); |
5859 | BB = NewBB; |
5860 | i = PN->getBasicBlockIndex(BB); |
5861 | |
5862 | needUpdateFixups = true; |
5863 | } |
5864 | } |
5865 | } |
5866 | |
5867 | std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair = |
5868 | Inserted.try_emplace(Key: BB); |
5869 | if (!Pair.second) |
5870 | PN->setIncomingValue(i, V: Pair.first->second); |
5871 | else { |
5872 | Value *FullV = |
5873 | Expand(LU, LF, F, IP: BB->getTerminator()->getIterator(), DeadInsts); |
5874 | |
5875 | // If this is reuse-by-noop-cast, insert the noop cast. |
5876 | Type *OpTy = LF.OperandValToReplace->getType(); |
5877 | if (FullV->getType() != OpTy) |
5878 | FullV = CastInst::Create( |
5879 | CastInst::getCastOpcode(Val: FullV, SrcIsSigned: false, Ty: OpTy, DstIsSigned: false), S: FullV, |
5880 | Ty: LF.OperandValToReplace->getType(), Name: "tmp" , |
5881 | InsertBefore: BB->getTerminator()->getIterator()); |
5882 | |
5883 | // If the incoming block for this value is not in the loop, it means the |
5884 | // current PHI is not in a loop exit, so we must create a LCSSA PHI for |
5885 | // the inserted value. |
5886 | if (auto *I = dyn_cast<Instruction>(Val: FullV)) |
5887 | if (L->contains(Inst: I) && !L->contains(BB)) |
5888 | InsertedNonLCSSAInsts.insert(X: I); |
5889 | |
5890 | PN->setIncomingValue(i, V: FullV); |
5891 | Pair.first->second = FullV; |
5892 | } |
5893 | |
5894 | // If LSR splits critical edge and phi node has other pending |
5895 | // fixup operands, we need to update those pending fixups. Otherwise |
5896 | // formulae will not be implemented completely and some instructions |
5897 | // will not be eliminated. |
5898 | if (needUpdateFixups) { |
5899 | for (LSRUse &LU : Uses) |
5900 | for (LSRFixup &Fixup : LU.Fixups) |
5901 | // If fixup is supposed to rewrite some operand in the phi |
5902 | // that was just updated, it may be already moved to |
5903 | // another phi node. Such fixup requires update. |
5904 | if (Fixup.UserInst == PN) { |
5905 | // Check if the operand we try to replace still exists in the |
5906 | // original phi. |
5907 | bool foundInOriginalPHI = false; |
5908 | for (const auto &val : PN->incoming_values()) |
5909 | if (val == Fixup.OperandValToReplace) { |
5910 | foundInOriginalPHI = true; |
5911 | break; |
5912 | } |
5913 | |
5914 | // If fixup operand found in original PHI - nothing to do. |
5915 | if (foundInOriginalPHI) |
5916 | continue; |
5917 | |
5918 | // Otherwise it might be moved to another PHI and requires update. |
5919 | // If fixup operand not found in any of the incoming blocks that |
5920 | // means we have already rewritten it - nothing to do. |
5921 | for (const auto &Block : PN->blocks()) |
5922 | for (BasicBlock::iterator I = Block->begin(); isa<PHINode>(Val: I); |
5923 | ++I) { |
5924 | PHINode *NewPN = cast<PHINode>(Val&: I); |
5925 | for (const auto &val : NewPN->incoming_values()) |
5926 | if (val == Fixup.OperandValToReplace) |
5927 | Fixup.UserInst = NewPN; |
5928 | } |
5929 | } |
5930 | } |
5931 | } |
5932 | } |
5933 | |
5934 | /// Emit instructions for the leading candidate expression for this LSRUse (this |
5935 | /// is called "expanding"), and update the UserInst to reference the newly |
5936 | /// expanded value. |
5937 | void LSRInstance::Rewrite(const LSRUse &LU, const LSRFixup &LF, |
5938 | const Formula &F, |
5939 | SmallVectorImpl<WeakTrackingVH> &DeadInsts) { |
5940 | // First, find an insertion point that dominates UserInst. For PHI nodes, |
5941 | // find the nearest block which dominates all the relevant uses. |
5942 | if (PHINode *PN = dyn_cast<PHINode>(Val: LF.UserInst)) { |
5943 | RewriteForPHI(PN, LU, LF, F, DeadInsts); |
5944 | } else { |
5945 | Value *FullV = Expand(LU, LF, F, IP: LF.UserInst->getIterator(), DeadInsts); |
5946 | |
5947 | // If this is reuse-by-noop-cast, insert the noop cast. |
5948 | Type *OpTy = LF.OperandValToReplace->getType(); |
5949 | if (FullV->getType() != OpTy) { |
5950 | Instruction *Cast = |
5951 | CastInst::Create(CastInst::getCastOpcode(Val: FullV, SrcIsSigned: false, Ty: OpTy, DstIsSigned: false), |
5952 | S: FullV, Ty: OpTy, Name: "tmp" , InsertBefore: LF.UserInst->getIterator()); |
5953 | FullV = Cast; |
5954 | } |
5955 | |
5956 | // Update the user. ICmpZero is handled specially here (for now) because |
5957 | // Expand may have updated one of the operands of the icmp already, and |
5958 | // its new value may happen to be equal to LF.OperandValToReplace, in |
5959 | // which case doing replaceUsesOfWith leads to replacing both operands |
5960 | // with the same value. TODO: Reorganize this. |
5961 | if (LU.Kind == LSRUse::ICmpZero) |
5962 | LF.UserInst->setOperand(i: 0, Val: FullV); |
5963 | else |
5964 | LF.UserInst->replaceUsesOfWith(From: LF.OperandValToReplace, To: FullV); |
5965 | } |
5966 | |
5967 | if (auto *OperandIsInstr = dyn_cast<Instruction>(Val: LF.OperandValToReplace)) |
5968 | DeadInsts.emplace_back(Args&: OperandIsInstr); |
5969 | } |
5970 | |
5971 | // Trying to hoist the IVInc to loop header if all IVInc users are in |
5972 | // the loop header. It will help backend to generate post index load/store |
5973 | // when the latch block is different from loop header block. |
5974 | static bool canHoistIVInc(const TargetTransformInfo &TTI, const LSRFixup &Fixup, |
5975 | const LSRUse &LU, Instruction *IVIncInsertPos, |
5976 | Loop *L) { |
5977 | if (LU.Kind != LSRUse::Address) |
5978 | return false; |
5979 | |
5980 | // For now this code do the conservative optimization, only work for |
5981 | // the header block. Later we can hoist the IVInc to the block post |
5982 | // dominate all users. |
5983 | BasicBlock * = L->getHeader(); |
5984 | if (IVIncInsertPos->getParent() == LHeader) |
5985 | return false; |
5986 | |
5987 | if (!Fixup.OperandValToReplace || |
5988 | any_of(Range: Fixup.OperandValToReplace->users(), P: [&LHeader](User *U) { |
5989 | Instruction *UI = cast<Instruction>(Val: U); |
5990 | return UI->getParent() != LHeader; |
5991 | })) |
5992 | return false; |
5993 | |
5994 | Instruction *I = Fixup.UserInst; |
5995 | Type *Ty = I->getType(); |
5996 | return (isa<LoadInst>(Val: I) && TTI.isIndexedLoadLegal(Mode: TTI.MIM_PostInc, Ty)) || |
5997 | (isa<StoreInst>(Val: I) && TTI.isIndexedStoreLegal(Mode: TTI.MIM_PostInc, Ty)); |
5998 | } |
5999 | |
6000 | /// Rewrite all the fixup locations with new values, following the chosen |
6001 | /// solution. |
6002 | void LSRInstance::ImplementSolution( |
6003 | const SmallVectorImpl<const Formula *> &Solution) { |
6004 | // Keep track of instructions we may have made dead, so that |
6005 | // we can remove them after we are done working. |
6006 | SmallVector<WeakTrackingVH, 16> DeadInsts; |
6007 | |
6008 | // Mark phi nodes that terminate chains so the expander tries to reuse them. |
6009 | for (const IVChain &Chain : IVChainVec) { |
6010 | if (PHINode *PN = dyn_cast<PHINode>(Val: Chain.tailUserInst())) |
6011 | Rewriter.setChainedPhi(PN); |
6012 | } |
6013 | |
6014 | // Expand the new value definitions and update the users. |
6015 | for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) |
6016 | for (const LSRFixup &Fixup : Uses[LUIdx].Fixups) { |
6017 | Instruction *InsertPos = |
6018 | canHoistIVInc(TTI, Fixup, LU: Uses[LUIdx], IVIncInsertPos, L) |
6019 | ? L->getHeader()->getTerminator() |
6020 | : IVIncInsertPos; |
6021 | Rewriter.setIVIncInsertPos(L, Pos: InsertPos); |
6022 | Rewrite(LU: Uses[LUIdx], LF: Fixup, F: *Solution[LUIdx], DeadInsts); |
6023 | Changed = true; |
6024 | } |
6025 | |
6026 | auto InsertedInsts = InsertedNonLCSSAInsts.takeVector(); |
6027 | formLCSSAForInstructions(Worklist&: InsertedInsts, DT, LI, SE: &SE); |
6028 | |
6029 | for (const IVChain &Chain : IVChainVec) { |
6030 | GenerateIVChain(Chain, DeadInsts); |
6031 | Changed = true; |
6032 | } |
6033 | |
6034 | for (const WeakVH &IV : Rewriter.getInsertedIVs()) |
6035 | if (IV && dyn_cast<Instruction>(Val: &*IV)->getParent()) |
6036 | ScalarEvolutionIVs.push_back(Elt: IV); |
6037 | |
6038 | // Clean up after ourselves. This must be done before deleting any |
6039 | // instructions. |
6040 | Rewriter.clear(); |
6041 | |
6042 | Changed |= RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts, |
6043 | TLI: &TLI, MSSAU); |
6044 | |
6045 | // In our cost analysis above, we assume that each addrec consumes exactly |
6046 | // one register, and arrange to have increments inserted just before the |
6047 | // latch to maximimize the chance this is true. However, if we reused |
6048 | // existing IVs, we now need to move the increments to match our |
6049 | // expectations. Otherwise, our cost modeling results in us having a |
6050 | // chosen a non-optimal result for the actual schedule. (And yes, this |
6051 | // scheduling decision does impact later codegen.) |
6052 | for (PHINode &PN : L->getHeader()->phis()) { |
6053 | BinaryOperator *BO = nullptr; |
6054 | Value *Start = nullptr, *Step = nullptr; |
6055 | if (!matchSimpleRecurrence(P: &PN, BO, Start, Step)) |
6056 | continue; |
6057 | |
6058 | switch (BO->getOpcode()) { |
6059 | case Instruction::Sub: |
6060 | if (BO->getOperand(i_nocapture: 0) != &PN) |
6061 | // sub is non-commutative - match handling elsewhere in LSR |
6062 | continue; |
6063 | break; |
6064 | case Instruction::Add: |
6065 | break; |
6066 | default: |
6067 | continue; |
6068 | }; |
6069 | |
6070 | if (!isa<Constant>(Val: Step)) |
6071 | // If not a constant step, might increase register pressure |
6072 | // (We assume constants have been canonicalized to RHS) |
6073 | continue; |
6074 | |
6075 | if (BO->getParent() == IVIncInsertPos->getParent()) |
6076 | // Only bother moving across blocks. Isel can handle block local case. |
6077 | continue; |
6078 | |
6079 | // Can we legally schedule inc at the desired point? |
6080 | if (!llvm::all_of(Range: BO->uses(), |
6081 | P: [&](Use &U) {return DT.dominates(Def: IVIncInsertPos, U);})) |
6082 | continue; |
6083 | BO->moveBefore(InsertPos: IVIncInsertPos->getIterator()); |
6084 | Changed = true; |
6085 | } |
6086 | |
6087 | |
6088 | } |
6089 | |
6090 | LSRInstance::LSRInstance(Loop *L, IVUsers &IU, ScalarEvolution &SE, |
6091 | DominatorTree &DT, LoopInfo &LI, |
6092 | const TargetTransformInfo &TTI, AssumptionCache &AC, |
6093 | TargetLibraryInfo &TLI, MemorySSAUpdater *MSSAU) |
6094 | : IU(IU), SE(SE), DT(DT), LI(LI), AC(AC), TLI(TLI), TTI(TTI), L(L), |
6095 | MSSAU(MSSAU), AMK(PreferredAddresingMode.getNumOccurrences() > 0 |
6096 | ? PreferredAddresingMode |
6097 | : TTI.getPreferredAddressingMode(L, SE: &SE)), |
6098 | Rewriter(SE, L->getHeader()->getDataLayout(), "lsr" , false), |
6099 | BaselineCost(L, SE, TTI, AMK) { |
6100 | // If LoopSimplify form is not available, stay out of trouble. |
6101 | if (!L->isLoopSimplifyForm()) |
6102 | return; |
6103 | |
6104 | // If there's no interesting work to be done, bail early. |
6105 | if (IU.empty()) return; |
6106 | |
6107 | // If there's too much analysis to be done, bail early. We won't be able to |
6108 | // model the problem anyway. |
6109 | unsigned NumUsers = 0; |
6110 | for (const IVStrideUse &U : IU) { |
6111 | if (++NumUsers > MaxIVUsers) { |
6112 | (void)U; |
6113 | LLVM_DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U |
6114 | << "\n" ); |
6115 | return; |
6116 | } |
6117 | // Bail out if we have a PHI on an EHPad that gets a value from a |
6118 | // CatchSwitchInst. Because the CatchSwitchInst cannot be split, there is |
6119 | // no good place to stick any instructions. |
6120 | if (auto *PN = dyn_cast<PHINode>(Val: U.getUser())) { |
6121 | auto FirstNonPHI = PN->getParent()->getFirstNonPHIIt(); |
6122 | if (isa<FuncletPadInst>(Val: FirstNonPHI) || |
6123 | isa<CatchSwitchInst>(Val: FirstNonPHI)) |
6124 | for (BasicBlock *PredBB : PN->blocks()) |
6125 | if (isa<CatchSwitchInst>(Val: PredBB->getFirstNonPHIIt())) |
6126 | return; |
6127 | } |
6128 | } |
6129 | |
6130 | LLVM_DEBUG(dbgs() << "\nLSR on loop " ; |
6131 | L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false); |
6132 | dbgs() << ":\n" ); |
6133 | |
6134 | // Configure SCEVExpander already now, so the correct mode is used for |
6135 | // isSafeToExpand() checks. |
6136 | #if LLVM_ENABLE_ABI_BREAKING_CHECKS |
6137 | Rewriter.setDebugType(DEBUG_TYPE); |
6138 | #endif |
6139 | Rewriter.disableCanonicalMode(); |
6140 | Rewriter.enableLSRMode(); |
6141 | |
6142 | // First, perform some low-level loop optimizations. |
6143 | OptimizeShadowIV(); |
6144 | OptimizeLoopTermCond(); |
6145 | |
6146 | // If loop preparation eliminates all interesting IV users, bail. |
6147 | if (IU.empty()) return; |
6148 | |
6149 | // Skip nested loops until we can model them better with formulae. |
6150 | if (!L->isInnermost()) { |
6151 | LLVM_DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n" ); |
6152 | return; |
6153 | } |
6154 | |
6155 | // Start collecting data and preparing for the solver. |
6156 | // If number of registers is not the major cost, we cannot benefit from the |
6157 | // current profitable chain optimization which is based on number of |
6158 | // registers. |
6159 | // FIXME: add profitable chain optimization for other kinds major cost, for |
6160 | // example number of instructions. |
6161 | if (TTI.isNumRegsMajorCostOfLSR() || StressIVChain) |
6162 | CollectChains(); |
6163 | CollectInterestingTypesAndFactors(); |
6164 | CollectFixupsAndInitialFormulae(); |
6165 | CollectLoopInvariantFixupsAndFormulae(); |
6166 | |
6167 | if (Uses.empty()) |
6168 | return; |
6169 | |
6170 | LLVM_DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n" ; |
6171 | print_uses(dbgs())); |
6172 | LLVM_DEBUG(dbgs() << "The baseline solution requires " ; |
6173 | BaselineCost.print(dbgs()); dbgs() << "\n" ); |
6174 | |
6175 | // Now use the reuse data to generate a bunch of interesting ways |
6176 | // to formulate the values needed for the uses. |
6177 | GenerateAllReuseFormulae(); |
6178 | |
6179 | FilterOutUndesirableDedicatedRegisters(); |
6180 | NarrowSearchSpaceUsingHeuristics(); |
6181 | |
6182 | SmallVector<const Formula *, 8> Solution; |
6183 | Solve(Solution); |
6184 | |
6185 | // Release memory that is no longer needed. |
6186 | Factors.clear(); |
6187 | Types.clear(); |
6188 | RegUses.clear(); |
6189 | |
6190 | if (Solution.empty()) |
6191 | return; |
6192 | |
6193 | #ifndef NDEBUG |
6194 | // Formulae should be legal. |
6195 | for (const LSRUse &LU : Uses) { |
6196 | for (const Formula &F : LU.Formulae) |
6197 | assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, |
6198 | F) && "Illegal formula generated!" ); |
6199 | }; |
6200 | #endif |
6201 | |
6202 | // Now that we've decided what we want, make it so. |
6203 | ImplementSolution(Solution); |
6204 | } |
6205 | |
6206 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
6207 | void LSRInstance::print_factors_and_types(raw_ostream &OS) const { |
6208 | if (Factors.empty() && Types.empty()) return; |
6209 | |
6210 | OS << "LSR has identified the following interesting factors and types: " ; |
6211 | bool First = true; |
6212 | |
6213 | for (int64_t Factor : Factors) { |
6214 | if (!First) OS << ", " ; |
6215 | First = false; |
6216 | OS << '*' << Factor; |
6217 | } |
6218 | |
6219 | for (Type *Ty : Types) { |
6220 | if (!First) OS << ", " ; |
6221 | First = false; |
6222 | OS << '(' << *Ty << ')'; |
6223 | } |
6224 | OS << '\n'; |
6225 | } |
6226 | |
6227 | void LSRInstance::print_fixups(raw_ostream &OS) const { |
6228 | OS << "LSR is examining the following fixup sites:\n" ; |
6229 | for (const LSRUse &LU : Uses) |
6230 | for (const LSRFixup &LF : LU.Fixups) { |
6231 | dbgs() << " " ; |
6232 | LF.print(OS); |
6233 | OS << '\n'; |
6234 | } |
6235 | } |
6236 | |
6237 | void LSRInstance::print_uses(raw_ostream &OS) const { |
6238 | OS << "LSR is examining the following uses:\n" ; |
6239 | for (const LSRUse &LU : Uses) { |
6240 | dbgs() << " " ; |
6241 | LU.print(OS); |
6242 | OS << '\n'; |
6243 | for (const Formula &F : LU.Formulae) { |
6244 | OS << " " ; |
6245 | F.print(OS); |
6246 | OS << '\n'; |
6247 | } |
6248 | } |
6249 | } |
6250 | |
6251 | void LSRInstance::print(raw_ostream &OS) const { |
6252 | print_factors_and_types(OS); |
6253 | print_fixups(OS); |
6254 | print_uses(OS); |
6255 | } |
6256 | |
6257 | LLVM_DUMP_METHOD void LSRInstance::dump() const { |
6258 | print(errs()); errs() << '\n'; |
6259 | } |
6260 | #endif |
6261 | |
6262 | namespace { |
6263 | |
6264 | class LoopStrengthReduce : public LoopPass { |
6265 | public: |
6266 | static char ID; // Pass ID, replacement for typeid |
6267 | |
6268 | LoopStrengthReduce(); |
6269 | |
6270 | private: |
6271 | bool runOnLoop(Loop *L, LPPassManager &LPM) override; |
6272 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
6273 | }; |
6274 | |
6275 | } // end anonymous namespace |
6276 | |
6277 | LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) { |
6278 | initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry()); |
6279 | } |
6280 | |
6281 | void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const { |
6282 | // We split critical edges, so we change the CFG. However, we do update |
6283 | // many analyses if they are around. |
6284 | AU.addPreservedID(ID&: LoopSimplifyID); |
6285 | |
6286 | AU.addRequired<LoopInfoWrapperPass>(); |
6287 | AU.addPreserved<LoopInfoWrapperPass>(); |
6288 | AU.addRequiredID(ID&: LoopSimplifyID); |
6289 | AU.addRequired<DominatorTreeWrapperPass>(); |
6290 | AU.addPreserved<DominatorTreeWrapperPass>(); |
6291 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
6292 | AU.addPreserved<ScalarEvolutionWrapperPass>(); |
6293 | AU.addRequired<AssumptionCacheTracker>(); |
6294 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
6295 | // Requiring LoopSimplify a second time here prevents IVUsers from running |
6296 | // twice, since LoopSimplify was invalidated by running ScalarEvolution. |
6297 | AU.addRequiredID(ID&: LoopSimplifyID); |
6298 | AU.addRequired<IVUsersWrapperPass>(); |
6299 | AU.addPreserved<IVUsersWrapperPass>(); |
6300 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
6301 | AU.addPreserved<MemorySSAWrapperPass>(); |
6302 | } |
6303 | |
6304 | namespace { |
6305 | |
6306 | /// Enables more convenient iteration over a DWARF expression vector. |
6307 | static iterator_range<llvm::DIExpression::expr_op_iterator> |
6308 | ToDwarfOpIter(SmallVectorImpl<uint64_t> &Expr) { |
6309 | llvm::DIExpression::expr_op_iterator Begin = |
6310 | llvm::DIExpression::expr_op_iterator(Expr.begin()); |
6311 | llvm::DIExpression::expr_op_iterator End = |
6312 | llvm::DIExpression::expr_op_iterator(Expr.end()); |
6313 | return {Begin, End}; |
6314 | } |
6315 | |
6316 | struct SCEVDbgValueBuilder { |
6317 | SCEVDbgValueBuilder() = default; |
6318 | SCEVDbgValueBuilder(const SCEVDbgValueBuilder &Base) { clone(Base); } |
6319 | |
6320 | void clone(const SCEVDbgValueBuilder &Base) { |
6321 | LocationOps = Base.LocationOps; |
6322 | Expr = Base.Expr; |
6323 | } |
6324 | |
6325 | void clear() { |
6326 | LocationOps.clear(); |
6327 | Expr.clear(); |
6328 | } |
6329 | |
6330 | /// The DIExpression as we translate the SCEV. |
6331 | SmallVector<uint64_t, 6> Expr; |
6332 | /// The location ops of the DIExpression. |
6333 | SmallVector<Value *, 2> LocationOps; |
6334 | |
6335 | void pushOperator(uint64_t Op) { Expr.push_back(Elt: Op); } |
6336 | void pushUInt(uint64_t Operand) { Expr.push_back(Elt: Operand); } |
6337 | |
6338 | /// Add a DW_OP_LLVM_arg to the expression, followed by the index of the value |
6339 | /// in the set of values referenced by the expression. |
6340 | void pushLocation(llvm::Value *V) { |
6341 | Expr.push_back(Elt: llvm::dwarf::DW_OP_LLVM_arg); |
6342 | auto *It = llvm::find(Range&: LocationOps, Val: V); |
6343 | unsigned ArgIndex = 0; |
6344 | if (It != LocationOps.end()) { |
6345 | ArgIndex = std::distance(first: LocationOps.begin(), last: It); |
6346 | } else { |
6347 | ArgIndex = LocationOps.size(); |
6348 | LocationOps.push_back(Elt: V); |
6349 | } |
6350 | Expr.push_back(Elt: ArgIndex); |
6351 | } |
6352 | |
6353 | void pushValue(const SCEVUnknown *U) { |
6354 | llvm::Value *V = cast<SCEVUnknown>(Val: U)->getValue(); |
6355 | pushLocation(V); |
6356 | } |
6357 | |
6358 | bool pushConst(const SCEVConstant *C) { |
6359 | if (C->getAPInt().getSignificantBits() > 64) |
6360 | return false; |
6361 | Expr.push_back(Elt: llvm::dwarf::DW_OP_consts); |
6362 | Expr.push_back(Elt: C->getAPInt().getSExtValue()); |
6363 | return true; |
6364 | } |
6365 | |
6366 | // Iterating the expression as DWARF ops is convenient when updating |
6367 | // DWARF_OP_LLVM_args. |
6368 | iterator_range<llvm::DIExpression::expr_op_iterator> expr_ops() { |
6369 | return ToDwarfOpIter(Expr); |
6370 | } |
6371 | |
6372 | /// Several SCEV types are sequences of the same arithmetic operator applied |
6373 | /// to constants and values that may be extended or truncated. |
6374 | bool pushArithmeticExpr(const llvm::SCEVCommutativeExpr *CommExpr, |
6375 | uint64_t DwarfOp) { |
6376 | assert((isa<llvm::SCEVAddExpr>(CommExpr) || isa<SCEVMulExpr>(CommExpr)) && |
6377 | "Expected arithmetic SCEV type" ); |
6378 | bool Success = true; |
6379 | unsigned EmitOperator = 0; |
6380 | for (const auto &Op : CommExpr->operands()) { |
6381 | Success &= pushSCEV(S: Op); |
6382 | |
6383 | if (EmitOperator >= 1) |
6384 | pushOperator(Op: DwarfOp); |
6385 | ++EmitOperator; |
6386 | } |
6387 | return Success; |
6388 | } |
6389 | |
6390 | // TODO: Identify and omit noop casts. |
6391 | bool pushCast(const llvm::SCEVCastExpr *C, bool IsSigned) { |
6392 | const llvm::SCEV *Inner = C->getOperand(i: 0); |
6393 | const llvm::Type *Type = C->getType(); |
6394 | uint64_t ToWidth = Type->getIntegerBitWidth(); |
6395 | bool Success = pushSCEV(S: Inner); |
6396 | uint64_t CastOps[] = {dwarf::DW_OP_LLVM_convert, ToWidth, |
6397 | IsSigned ? llvm::dwarf::DW_ATE_signed |
6398 | : llvm::dwarf::DW_ATE_unsigned}; |
6399 | for (const auto &Op : CastOps) |
6400 | pushOperator(Op); |
6401 | return Success; |
6402 | } |
6403 | |
6404 | // TODO: MinMax - although these haven't been encountered in the test suite. |
6405 | bool pushSCEV(const llvm::SCEV *S) { |
6406 | bool Success = true; |
6407 | if (const SCEVConstant *StartInt = dyn_cast<SCEVConstant>(Val: S)) { |
6408 | Success &= pushConst(C: StartInt); |
6409 | |
6410 | } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Val: S)) { |
6411 | if (!U->getValue()) |
6412 | return false; |
6413 | pushLocation(V: U->getValue()); |
6414 | |
6415 | } else if (const SCEVMulExpr *MulRec = dyn_cast<SCEVMulExpr>(Val: S)) { |
6416 | Success &= pushArithmeticExpr(CommExpr: MulRec, DwarfOp: llvm::dwarf::DW_OP_mul); |
6417 | |
6418 | } else if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(Val: S)) { |
6419 | Success &= pushSCEV(S: UDiv->getLHS()); |
6420 | Success &= pushSCEV(S: UDiv->getRHS()); |
6421 | pushOperator(Op: llvm::dwarf::DW_OP_div); |
6422 | |
6423 | } else if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(Val: S)) { |
6424 | // Assert if a new and unknown SCEVCastEXpr type is encountered. |
6425 | assert((isa<SCEVZeroExtendExpr>(Cast) || isa<SCEVTruncateExpr>(Cast) || |
6426 | isa<SCEVPtrToIntExpr>(Cast) || isa<SCEVSignExtendExpr>(Cast)) && |
6427 | "Unexpected cast type in SCEV." ); |
6428 | Success &= pushCast(C: Cast, IsSigned: (isa<SCEVSignExtendExpr>(Val: Cast))); |
6429 | |
6430 | } else if (const SCEVAddExpr *AddExpr = dyn_cast<SCEVAddExpr>(Val: S)) { |
6431 | Success &= pushArithmeticExpr(CommExpr: AddExpr, DwarfOp: llvm::dwarf::DW_OP_plus); |
6432 | |
6433 | } else if (isa<SCEVAddRecExpr>(Val: S)) { |
6434 | // Nested SCEVAddRecExpr are generated by nested loops and are currently |
6435 | // unsupported. |
6436 | return false; |
6437 | |
6438 | } else { |
6439 | return false; |
6440 | } |
6441 | return Success; |
6442 | } |
6443 | |
6444 | /// Return true if the combination of arithmetic operator and underlying |
6445 | /// SCEV constant value is an identity function. |
6446 | bool isIdentityFunction(uint64_t Op, const SCEV *S) { |
6447 | if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Val: S)) { |
6448 | if (C->getAPInt().getSignificantBits() > 64) |
6449 | return false; |
6450 | int64_t I = C->getAPInt().getSExtValue(); |
6451 | switch (Op) { |
6452 | case llvm::dwarf::DW_OP_plus: |
6453 | case llvm::dwarf::DW_OP_minus: |
6454 | return I == 0; |
6455 | case llvm::dwarf::DW_OP_mul: |
6456 | case llvm::dwarf::DW_OP_div: |
6457 | return I == 1; |
6458 | } |
6459 | } |
6460 | return false; |
6461 | } |
6462 | |
6463 | /// Convert a SCEV of a value to a DIExpression that is pushed onto the |
6464 | /// builder's expression stack. The stack should already contain an |
6465 | /// expression for the iteration count, so that it can be multiplied by |
6466 | /// the stride and added to the start. |
6467 | /// Components of the expression are omitted if they are an identity function. |
6468 | /// Chain (non-affine) SCEVs are not supported. |
6469 | bool SCEVToValueExpr(const llvm::SCEVAddRecExpr &SAR, ScalarEvolution &SE) { |
6470 | assert(SAR.isAffine() && "Expected affine SCEV" ); |
6471 | const SCEV *Start = SAR.getStart(); |
6472 | const SCEV *Stride = SAR.getStepRecurrence(SE); |
6473 | |
6474 | // Skip pushing arithmetic noops. |
6475 | if (!isIdentityFunction(Op: llvm::dwarf::DW_OP_mul, S: Stride)) { |
6476 | if (!pushSCEV(S: Stride)) |
6477 | return false; |
6478 | pushOperator(Op: llvm::dwarf::DW_OP_mul); |
6479 | } |
6480 | if (!isIdentityFunction(Op: llvm::dwarf::DW_OP_plus, S: Start)) { |
6481 | if (!pushSCEV(S: Start)) |
6482 | return false; |
6483 | pushOperator(Op: llvm::dwarf::DW_OP_plus); |
6484 | } |
6485 | return true; |
6486 | } |
6487 | |
6488 | /// Create an expression that is an offset from a value (usually the IV). |
6489 | void createOffsetExpr(int64_t Offset, Value *OffsetValue) { |
6490 | pushLocation(V: OffsetValue); |
6491 | DIExpression::appendOffset(Ops&: Expr, Offset); |
6492 | LLVM_DEBUG( |
6493 | dbgs() << "scev-salvage: Generated IV offset expression. Offset: " |
6494 | << std::to_string(Offset) << "\n" ); |
6495 | } |
6496 | |
6497 | /// Combine a translation of the SCEV and the IV to create an expression that |
6498 | /// recovers a location's value. |
6499 | /// returns true if an expression was created. |
6500 | bool createIterCountExpr(const SCEV *S, |
6501 | const SCEVDbgValueBuilder &IterationCount, |
6502 | ScalarEvolution &SE) { |
6503 | // SCEVs for SSA values are most frquently of the form |
6504 | // {start,+,stride}, but sometimes they are ({start,+,stride} + %a + ..). |
6505 | // This is because %a is a PHI node that is not the IV. However, these |
6506 | // SCEVs have not been observed to result in debuginfo-lossy optimisations, |
6507 | // so its not expected this point will be reached. |
6508 | if (!isa<SCEVAddRecExpr>(Val: S)) |
6509 | return false; |
6510 | |
6511 | LLVM_DEBUG(dbgs() << "scev-salvage: Location to salvage SCEV: " << *S |
6512 | << '\n'); |
6513 | |
6514 | const auto *Rec = cast<SCEVAddRecExpr>(Val: S); |
6515 | if (!Rec->isAffine()) |
6516 | return false; |
6517 | |
6518 | if (S->getExpressionSize() > MaxSCEVSalvageExpressionSize) |
6519 | return false; |
6520 | |
6521 | // Initialise a new builder with the iteration count expression. In |
6522 | // combination with the value's SCEV this enables recovery. |
6523 | clone(Base: IterationCount); |
6524 | if (!SCEVToValueExpr(SAR: *Rec, SE)) |
6525 | return false; |
6526 | |
6527 | return true; |
6528 | } |
6529 | |
6530 | /// Convert a SCEV of a value to a DIExpression that is pushed onto the |
6531 | /// builder's expression stack. The stack should already contain an |
6532 | /// expression for the iteration count, so that it can be multiplied by |
6533 | /// the stride and added to the start. |
6534 | /// Components of the expression are omitted if they are an identity function. |
6535 | bool SCEVToIterCountExpr(const llvm::SCEVAddRecExpr &SAR, |
6536 | ScalarEvolution &SE) { |
6537 | assert(SAR.isAffine() && "Expected affine SCEV" ); |
6538 | const SCEV *Start = SAR.getStart(); |
6539 | const SCEV *Stride = SAR.getStepRecurrence(SE); |
6540 | |
6541 | // Skip pushing arithmetic noops. |
6542 | if (!isIdentityFunction(Op: llvm::dwarf::DW_OP_minus, S: Start)) { |
6543 | if (!pushSCEV(S: Start)) |
6544 | return false; |
6545 | pushOperator(Op: llvm::dwarf::DW_OP_minus); |
6546 | } |
6547 | if (!isIdentityFunction(Op: llvm::dwarf::DW_OP_div, S: Stride)) { |
6548 | if (!pushSCEV(S: Stride)) |
6549 | return false; |
6550 | pushOperator(Op: llvm::dwarf::DW_OP_div); |
6551 | } |
6552 | return true; |
6553 | } |
6554 | |
6555 | // Append the current expression and locations to a location list and an |
6556 | // expression list. Modify the DW_OP_LLVM_arg indexes to account for |
6557 | // the locations already present in the destination list. |
6558 | void appendToVectors(SmallVectorImpl<uint64_t> &DestExpr, |
6559 | SmallVectorImpl<Value *> &DestLocations) { |
6560 | assert(!DestLocations.empty() && |
6561 | "Expected the locations vector to contain the IV" ); |
6562 | // The DWARF_OP_LLVM_arg arguments of the expression being appended must be |
6563 | // modified to account for the locations already in the destination vector. |
6564 | // All builders contain the IV as the first location op. |
6565 | assert(!LocationOps.empty() && |
6566 | "Expected the location ops to contain the IV." ); |
6567 | // DestIndexMap[n] contains the index in DestLocations for the nth |
6568 | // location in this SCEVDbgValueBuilder. |
6569 | SmallVector<uint64_t, 2> DestIndexMap; |
6570 | for (const auto &Op : LocationOps) { |
6571 | auto It = find(Range&: DestLocations, Val: Op); |
6572 | if (It != DestLocations.end()) { |
6573 | // Location already exists in DestLocations, reuse existing ArgIndex. |
6574 | DestIndexMap.push_back(Elt: std::distance(first: DestLocations.begin(), last: It)); |
6575 | continue; |
6576 | } |
6577 | // Location is not in DestLocations, add it. |
6578 | DestIndexMap.push_back(Elt: DestLocations.size()); |
6579 | DestLocations.push_back(Elt: Op); |
6580 | } |
6581 | |
6582 | for (const auto &Op : expr_ops()) { |
6583 | if (Op.getOp() != dwarf::DW_OP_LLVM_arg) { |
6584 | Op.appendToVector(V&: DestExpr); |
6585 | continue; |
6586 | } |
6587 | |
6588 | DestExpr.push_back(Elt: dwarf::DW_OP_LLVM_arg); |
6589 | // `DW_OP_LLVM_arg n` represents the nth LocationOp in this SCEV, |
6590 | // DestIndexMap[n] contains its new index in DestLocations. |
6591 | uint64_t NewIndex = DestIndexMap[Op.getArg(I: 0)]; |
6592 | DestExpr.push_back(Elt: NewIndex); |
6593 | } |
6594 | } |
6595 | }; |
6596 | |
6597 | /// Holds all the required data to salvage a dbg.value using the pre-LSR SCEVs |
6598 | /// and DIExpression. |
6599 | struct DVIRecoveryRec { |
6600 | DVIRecoveryRec(DbgValueInst *DbgValue) |
6601 | : DbgRef(DbgValue), Expr(DbgValue->getExpression()), |
6602 | HadLocationArgList(false) {} |
6603 | DVIRecoveryRec(DbgVariableRecord *DVR) |
6604 | : DbgRef(DVR), Expr(DVR->getExpression()), HadLocationArgList(false) {} |
6605 | |
6606 | PointerUnion<DbgValueInst *, DbgVariableRecord *> DbgRef; |
6607 | DIExpression *Expr; |
6608 | bool HadLocationArgList; |
6609 | SmallVector<WeakVH, 2> LocationOps; |
6610 | SmallVector<const llvm::SCEV *, 2> SCEVs; |
6611 | SmallVector<std::unique_ptr<SCEVDbgValueBuilder>, 2> RecoveryExprs; |
6612 | |
6613 | void clear() { |
6614 | for (auto &RE : RecoveryExprs) |
6615 | RE.reset(); |
6616 | RecoveryExprs.clear(); |
6617 | } |
6618 | |
6619 | ~DVIRecoveryRec() { clear(); } |
6620 | }; |
6621 | } // namespace |
6622 | |
6623 | /// Returns the total number of DW_OP_llvm_arg operands in the expression. |
6624 | /// This helps in determining if a DIArglist is necessary or can be omitted from |
6625 | /// the dbg.value. |
6626 | static unsigned numLLVMArgOps(SmallVectorImpl<uint64_t> &Expr) { |
6627 | auto expr_ops = ToDwarfOpIter(Expr); |
6628 | unsigned Count = 0; |
6629 | for (auto Op : expr_ops) |
6630 | if (Op.getOp() == dwarf::DW_OP_LLVM_arg) |
6631 | Count++; |
6632 | return Count; |
6633 | } |
6634 | |
6635 | /// Overwrites DVI with the location and Ops as the DIExpression. This will |
6636 | /// create an invalid expression if Ops has any dwarf::DW_OP_llvm_arg operands, |
6637 | /// because a DIArglist is not created for the first argument of the dbg.value. |
6638 | template <typename T> |
6639 | static void updateDVIWithLocation(T &DbgVal, Value *Location, |
6640 | SmallVectorImpl<uint64_t> &Ops) { |
6641 | assert(numLLVMArgOps(Ops) == 0 && "Expected expression that does not " |
6642 | "contain any DW_OP_llvm_arg operands." ); |
6643 | DbgVal.setRawLocation(ValueAsMetadata::get(V: Location)); |
6644 | DbgVal.setExpression(DIExpression::get(Context&: DbgVal.getContext(), Elements: Ops)); |
6645 | DbgVal.setExpression(DIExpression::get(Context&: DbgVal.getContext(), Elements: Ops)); |
6646 | } |
6647 | |
6648 | /// Overwrite DVI with locations placed into a DIArglist. |
6649 | template <typename T> |
6650 | static void updateDVIWithLocations(T &DbgVal, |
6651 | SmallVectorImpl<Value *> &Locations, |
6652 | SmallVectorImpl<uint64_t> &Ops) { |
6653 | assert(numLLVMArgOps(Ops) != 0 && |
6654 | "Expected expression that references DIArglist locations using " |
6655 | "DW_OP_llvm_arg operands." ); |
6656 | SmallVector<ValueAsMetadata *, 3> MetadataLocs; |
6657 | for (Value *V : Locations) |
6658 | MetadataLocs.push_back(Elt: ValueAsMetadata::get(V)); |
6659 | auto ValArrayRef = llvm::ArrayRef<llvm::ValueAsMetadata *>(MetadataLocs); |
6660 | DbgVal.setRawLocation(llvm::DIArgList::get(Context&: DbgVal.getContext(), Args: ValArrayRef)); |
6661 | DbgVal.setExpression(DIExpression::get(Context&: DbgVal.getContext(), Elements: Ops)); |
6662 | } |
6663 | |
6664 | /// Write the new expression and new location ops for the dbg.value. If possible |
6665 | /// reduce the szie of the dbg.value intrinsic by omitting DIArglist. This |
6666 | /// can be omitted if: |
6667 | /// 1. There is only a single location, refenced by a single DW_OP_llvm_arg. |
6668 | /// 2. The DW_OP_LLVM_arg is the first operand in the expression. |
6669 | static void UpdateDbgValueInst(DVIRecoveryRec &DVIRec, |
6670 | SmallVectorImpl<Value *> &NewLocationOps, |
6671 | SmallVectorImpl<uint64_t> &NewExpr) { |
6672 | auto UpdateDbgValueInstImpl = [&](auto *DbgVal) { |
6673 | unsigned NumLLVMArgs = numLLVMArgOps(Expr&: NewExpr); |
6674 | if (NumLLVMArgs == 0) { |
6675 | // Location assumed to be on the stack. |
6676 | updateDVIWithLocation(*DbgVal, NewLocationOps[0], NewExpr); |
6677 | } else if (NumLLVMArgs == 1 && NewExpr[0] == dwarf::DW_OP_LLVM_arg) { |
6678 | // There is only a single DW_OP_llvm_arg at the start of the expression, |
6679 | // so it can be omitted along with DIArglist. |
6680 | assert(NewExpr[1] == 0 && |
6681 | "Lone LLVM_arg in a DIExpression should refer to location-op 0." ); |
6682 | llvm::SmallVector<uint64_t, 6> ShortenedOps(llvm::drop_begin(RangeOrContainer&: NewExpr, N: 2)); |
6683 | updateDVIWithLocation(*DbgVal, NewLocationOps[0], ShortenedOps); |
6684 | } else { |
6685 | // Multiple DW_OP_llvm_arg, so DIArgList is strictly necessary. |
6686 | updateDVIWithLocations(*DbgVal, NewLocationOps, NewExpr); |
6687 | } |
6688 | |
6689 | // If the DIExpression was previously empty then add the stack terminator. |
6690 | // Non-empty expressions have only had elements inserted into them and so |
6691 | // the terminator should already be present e.g. stack_value or fragment. |
6692 | DIExpression *SalvageExpr = DbgVal->getExpression(); |
6693 | if (!DVIRec.Expr->isComplex() && SalvageExpr->isComplex()) { |
6694 | SalvageExpr = |
6695 | DIExpression::append(Expr: SalvageExpr, Ops: {dwarf::DW_OP_stack_value}); |
6696 | DbgVal->setExpression(SalvageExpr); |
6697 | } |
6698 | }; |
6699 | if (isa<DbgValueInst *>(Val: DVIRec.DbgRef)) |
6700 | UpdateDbgValueInstImpl(cast<DbgValueInst *>(Val&: DVIRec.DbgRef)); |
6701 | else |
6702 | UpdateDbgValueInstImpl(cast<DbgVariableRecord *>(Val&: DVIRec.DbgRef)); |
6703 | } |
6704 | |
6705 | /// Cached location ops may be erased during LSR, in which case a poison is |
6706 | /// required when restoring from the cache. The type of that location is no |
6707 | /// longer available, so just use int8. The poison will be replaced by one or |
6708 | /// more locations later when a SCEVDbgValueBuilder selects alternative |
6709 | /// locations to use for the salvage. |
6710 | static Value *getValueOrPoison(WeakVH &VH, LLVMContext &C) { |
6711 | return (VH) ? VH : PoisonValue::get(T: llvm::Type::getInt8Ty(C)); |
6712 | } |
6713 | |
6714 | /// Restore the DVI's pre-LSR arguments. Substitute undef for any erased values. |
6715 | static void restorePreTransformState(DVIRecoveryRec &DVIRec) { |
6716 | auto RestorePreTransformStateImpl = [&](auto *DbgVal) { |
6717 | LLVM_DEBUG(dbgs() << "scev-salvage: restore dbg.value to pre-LSR state\n" |
6718 | << "scev-salvage: post-LSR: " << *DbgVal << '\n'); |
6719 | assert(DVIRec.Expr && "Expected an expression" ); |
6720 | DbgVal->setExpression(DVIRec.Expr); |
6721 | |
6722 | // Even a single location-op may be inside a DIArgList and referenced with |
6723 | // DW_OP_LLVM_arg, which is valid only with a DIArgList. |
6724 | if (!DVIRec.HadLocationArgList) { |
6725 | assert(DVIRec.LocationOps.size() == 1 && |
6726 | "Unexpected number of location ops." ); |
6727 | // LSR's unsuccessful salvage attempt may have added DIArgList, which in |
6728 | // this case was not present before, so force the location back to a |
6729 | // single uncontained Value. |
6730 | Value *CachedValue = |
6731 | getValueOrPoison(DVIRec.LocationOps[0], DbgVal->getContext()); |
6732 | DbgVal->setRawLocation(ValueAsMetadata::get(V: CachedValue)); |
6733 | } else { |
6734 | SmallVector<ValueAsMetadata *, 3> MetadataLocs; |
6735 | for (WeakVH VH : DVIRec.LocationOps) { |
6736 | Value *CachedValue = getValueOrPoison(VH, DbgVal->getContext()); |
6737 | MetadataLocs.push_back(Elt: ValueAsMetadata::get(V: CachedValue)); |
6738 | } |
6739 | auto ValArrayRef = llvm::ArrayRef<llvm::ValueAsMetadata *>(MetadataLocs); |
6740 | DbgVal->setRawLocation( |
6741 | llvm::DIArgList::get(Context&: DbgVal->getContext(), Args: ValArrayRef)); |
6742 | } |
6743 | LLVM_DEBUG(dbgs() << "scev-salvage: pre-LSR: " << *DbgVal << '\n'); |
6744 | }; |
6745 | if (isa<DbgValueInst *>(Val: DVIRec.DbgRef)) |
6746 | RestorePreTransformStateImpl(cast<DbgValueInst *>(Val&: DVIRec.DbgRef)); |
6747 | else |
6748 | RestorePreTransformStateImpl(cast<DbgVariableRecord *>(Val&: DVIRec.DbgRef)); |
6749 | } |
6750 | |
6751 | static bool SalvageDVI(llvm::Loop *L, ScalarEvolution &SE, |
6752 | llvm::PHINode *LSRInductionVar, DVIRecoveryRec &DVIRec, |
6753 | const SCEV *SCEVInductionVar, |
6754 | SCEVDbgValueBuilder IterCountExpr) { |
6755 | |
6756 | if (isa<DbgValueInst *>(Val: DVIRec.DbgRef) |
6757 | ? !cast<DbgValueInst *>(Val&: DVIRec.DbgRef)->isKillLocation() |
6758 | : !cast<DbgVariableRecord *>(Val&: DVIRec.DbgRef)->isKillLocation()) |
6759 | return false; |
6760 | |
6761 | // LSR may have caused several changes to the dbg.value in the failed salvage |
6762 | // attempt. So restore the DIExpression, the location ops and also the |
6763 | // location ops format, which is always DIArglist for multiple ops, but only |
6764 | // sometimes for a single op. |
6765 | restorePreTransformState(DVIRec); |
6766 | |
6767 | // LocationOpIndexMap[i] will store the post-LSR location index of |
6768 | // the non-optimised out location at pre-LSR index i. |
6769 | SmallVector<int64_t, 2> LocationOpIndexMap; |
6770 | LocationOpIndexMap.assign(NumElts: DVIRec.LocationOps.size(), Elt: -1); |
6771 | SmallVector<Value *, 2> NewLocationOps; |
6772 | NewLocationOps.push_back(Elt: LSRInductionVar); |
6773 | |
6774 | for (unsigned i = 0; i < DVIRec.LocationOps.size(); i++) { |
6775 | WeakVH VH = DVIRec.LocationOps[i]; |
6776 | // Place the locations not optimised out in the list first, avoiding |
6777 | // inserts later. The map is used to update the DIExpression's |
6778 | // DW_OP_LLVM_arg arguments as the expression is updated. |
6779 | if (VH && !isa<UndefValue>(Val: VH)) { |
6780 | NewLocationOps.push_back(Elt: VH); |
6781 | LocationOpIndexMap[i] = NewLocationOps.size() - 1; |
6782 | LLVM_DEBUG(dbgs() << "scev-salvage: Location index " << i |
6783 | << " now at index " << LocationOpIndexMap[i] << "\n" ); |
6784 | continue; |
6785 | } |
6786 | |
6787 | // It's possible that a value referred to in the SCEV may have been |
6788 | // optimised out by LSR. |
6789 | if (SE.containsErasedValue(S: DVIRec.SCEVs[i]) || |
6790 | SE.containsUndefs(S: DVIRec.SCEVs[i])) { |
6791 | LLVM_DEBUG(dbgs() << "scev-salvage: SCEV for location at index: " << i |
6792 | << " refers to a location that is now undef or erased. " |
6793 | "Salvage abandoned.\n" ); |
6794 | return false; |
6795 | } |
6796 | |
6797 | LLVM_DEBUG(dbgs() << "scev-salvage: salvaging location at index " << i |
6798 | << " with SCEV: " << *DVIRec.SCEVs[i] << "\n" ); |
6799 | |
6800 | DVIRec.RecoveryExprs[i] = std::make_unique<SCEVDbgValueBuilder>(); |
6801 | SCEVDbgValueBuilder *SalvageExpr = DVIRec.RecoveryExprs[i].get(); |
6802 | |
6803 | // Create an offset-based salvage expression if possible, as it requires |
6804 | // less DWARF ops than an iteration count-based expression. |
6805 | if (std::optional<APInt> Offset = |
6806 | SE.computeConstantDifference(LHS: DVIRec.SCEVs[i], RHS: SCEVInductionVar)) { |
6807 | if (Offset->getSignificantBits() <= 64) |
6808 | SalvageExpr->createOffsetExpr(Offset: Offset->getSExtValue(), OffsetValue: LSRInductionVar); |
6809 | else |
6810 | return false; |
6811 | } else if (!SalvageExpr->createIterCountExpr(S: DVIRec.SCEVs[i], IterationCount: IterCountExpr, |
6812 | SE)) |
6813 | return false; |
6814 | } |
6815 | |
6816 | // Merge the DbgValueBuilder generated expressions and the original |
6817 | // DIExpression, place the result into an new vector. |
6818 | SmallVector<uint64_t, 3> NewExpr; |
6819 | if (DVIRec.Expr->getNumElements() == 0) { |
6820 | assert(DVIRec.RecoveryExprs.size() == 1 && |
6821 | "Expected only a single recovery expression for an empty " |
6822 | "DIExpression." ); |
6823 | assert(DVIRec.RecoveryExprs[0] && |
6824 | "Expected a SCEVDbgSalvageBuilder for location 0" ); |
6825 | SCEVDbgValueBuilder *B = DVIRec.RecoveryExprs[0].get(); |
6826 | B->appendToVectors(DestExpr&: NewExpr, DestLocations&: NewLocationOps); |
6827 | } |
6828 | for (const auto &Op : DVIRec.Expr->expr_ops()) { |
6829 | // Most Ops needn't be updated. |
6830 | if (Op.getOp() != dwarf::DW_OP_LLVM_arg) { |
6831 | Op.appendToVector(V&: NewExpr); |
6832 | continue; |
6833 | } |
6834 | |
6835 | uint64_t LocationArgIndex = Op.getArg(I: 0); |
6836 | SCEVDbgValueBuilder *DbgBuilder = |
6837 | DVIRec.RecoveryExprs[LocationArgIndex].get(); |
6838 | // The location doesn't have s SCEVDbgValueBuilder, so LSR did not |
6839 | // optimise it away. So just translate the argument to the updated |
6840 | // location index. |
6841 | if (!DbgBuilder) { |
6842 | NewExpr.push_back(Elt: dwarf::DW_OP_LLVM_arg); |
6843 | assert(LocationOpIndexMap[Op.getArg(0)] != -1 && |
6844 | "Expected a positive index for the location-op position." ); |
6845 | NewExpr.push_back(Elt: LocationOpIndexMap[Op.getArg(I: 0)]); |
6846 | continue; |
6847 | } |
6848 | // The location has a recovery expression. |
6849 | DbgBuilder->appendToVectors(DestExpr&: NewExpr, DestLocations&: NewLocationOps); |
6850 | } |
6851 | |
6852 | UpdateDbgValueInst(DVIRec, NewLocationOps, NewExpr); |
6853 | if (isa<DbgValueInst *>(Val: DVIRec.DbgRef)) |
6854 | LLVM_DEBUG(dbgs() << "scev-salvage: Updated DVI: " |
6855 | << *cast<DbgValueInst *>(DVIRec.DbgRef) << "\n" ); |
6856 | else |
6857 | LLVM_DEBUG(dbgs() << "scev-salvage: Updated DVI: " |
6858 | << *cast<DbgVariableRecord *>(DVIRec.DbgRef) << "\n" ); |
6859 | return true; |
6860 | } |
6861 | |
6862 | /// Obtain an expression for the iteration count, then attempt to salvage the |
6863 | /// dbg.value intrinsics. |
6864 | static void DbgRewriteSalvageableDVIs( |
6865 | llvm::Loop *L, ScalarEvolution &SE, llvm::PHINode *LSRInductionVar, |
6866 | SmallVector<std::unique_ptr<DVIRecoveryRec>, 2> &DVIToUpdate) { |
6867 | if (DVIToUpdate.empty()) |
6868 | return; |
6869 | |
6870 | const llvm::SCEV *SCEVInductionVar = SE.getSCEV(V: LSRInductionVar); |
6871 | assert(SCEVInductionVar && |
6872 | "Anticipated a SCEV for the post-LSR induction variable" ); |
6873 | |
6874 | if (const SCEVAddRecExpr *IVAddRec = |
6875 | dyn_cast<SCEVAddRecExpr>(Val: SCEVInductionVar)) { |
6876 | if (!IVAddRec->isAffine()) |
6877 | return; |
6878 | |
6879 | // Prevent translation using excessive resources. |
6880 | if (IVAddRec->getExpressionSize() > MaxSCEVSalvageExpressionSize) |
6881 | return; |
6882 | |
6883 | // The iteration count is required to recover location values. |
6884 | SCEVDbgValueBuilder IterCountExpr; |
6885 | IterCountExpr.pushLocation(V: LSRInductionVar); |
6886 | if (!IterCountExpr.SCEVToIterCountExpr(SAR: *IVAddRec, SE)) |
6887 | return; |
6888 | |
6889 | LLVM_DEBUG(dbgs() << "scev-salvage: IV SCEV: " << *SCEVInductionVar |
6890 | << '\n'); |
6891 | |
6892 | for (auto &DVIRec : DVIToUpdate) { |
6893 | SalvageDVI(L, SE, LSRInductionVar, DVIRec&: *DVIRec, SCEVInductionVar, |
6894 | IterCountExpr); |
6895 | } |
6896 | } |
6897 | } |
6898 | |
6899 | /// Identify and cache salvageable DVI locations and expressions along with the |
6900 | /// corresponding SCEV(s). Also ensure that the DVI is not deleted between |
6901 | /// cacheing and salvaging. |
6902 | static void DbgGatherSalvagableDVI( |
6903 | Loop *L, ScalarEvolution &SE, |
6904 | SmallVector<std::unique_ptr<DVIRecoveryRec>, 2> &SalvageableDVISCEVs, |
6905 | SmallSet<AssertingVH<DbgValueInst>, 2> &DVIHandles) { |
6906 | for (const auto &B : L->getBlocks()) { |
6907 | for (auto &I : *B) { |
6908 | auto ProcessDbgValue = [&](auto *DbgVal) -> bool { |
6909 | // Ensure that if any location op is undef that the dbg.vlue is not |
6910 | // cached. |
6911 | if (DbgVal->isKillLocation()) |
6912 | return false; |
6913 | |
6914 | // Check that the location op SCEVs are suitable for translation to |
6915 | // DIExpression. |
6916 | const auto &HasTranslatableLocationOps = |
6917 | [&](const auto *DbgValToTranslate) -> bool { |
6918 | for (const auto LocOp : DbgValToTranslate->location_ops()) { |
6919 | if (!LocOp) |
6920 | return false; |
6921 | |
6922 | if (!SE.isSCEVable(Ty: LocOp->getType())) |
6923 | return false; |
6924 | |
6925 | const SCEV *S = SE.getSCEV(V: LocOp); |
6926 | if (SE.containsUndefs(S)) |
6927 | return false; |
6928 | } |
6929 | return true; |
6930 | }; |
6931 | |
6932 | if (!HasTranslatableLocationOps(DbgVal)) |
6933 | return false; |
6934 | |
6935 | std::unique_ptr<DVIRecoveryRec> NewRec = |
6936 | std::make_unique<DVIRecoveryRec>(DbgVal); |
6937 | // Each location Op may need a SCEVDbgValueBuilder in order to recover |
6938 | // it. Pre-allocating a vector will enable quick lookups of the builder |
6939 | // later during the salvage. |
6940 | NewRec->RecoveryExprs.resize(DbgVal->getNumVariableLocationOps()); |
6941 | for (const auto LocOp : DbgVal->location_ops()) { |
6942 | NewRec->SCEVs.push_back(Elt: SE.getSCEV(V: LocOp)); |
6943 | NewRec->LocationOps.push_back(LocOp); |
6944 | NewRec->HadLocationArgList = DbgVal->hasArgList(); |
6945 | } |
6946 | SalvageableDVISCEVs.push_back(Elt: std::move(NewRec)); |
6947 | return true; |
6948 | }; |
6949 | for (DbgVariableRecord &DVR : filterDbgVars(R: I.getDbgRecordRange())) { |
6950 | if (DVR.isDbgValue() || DVR.isDbgAssign()) |
6951 | ProcessDbgValue(&DVR); |
6952 | } |
6953 | auto DVI = dyn_cast<DbgValueInst>(Val: &I); |
6954 | if (!DVI) |
6955 | continue; |
6956 | if (ProcessDbgValue(DVI)) |
6957 | DVIHandles.insert(V: DVI); |
6958 | } |
6959 | } |
6960 | } |
6961 | |
6962 | /// Ideally pick the PHI IV inserted by ScalarEvolutionExpander. As a fallback |
6963 | /// any PHi from the loop header is usable, but may have less chance of |
6964 | /// surviving subsequent transforms. |
6965 | static llvm::PHINode *GetInductionVariable(const Loop &L, ScalarEvolution &SE, |
6966 | const LSRInstance &LSR) { |
6967 | |
6968 | auto IsSuitableIV = [&](PHINode *P) { |
6969 | if (!SE.isSCEVable(Ty: P->getType())) |
6970 | return false; |
6971 | if (const SCEVAddRecExpr *Rec = dyn_cast<SCEVAddRecExpr>(Val: SE.getSCEV(V: P))) |
6972 | return Rec->isAffine() && !SE.containsUndefs(S: SE.getSCEV(V: P)); |
6973 | return false; |
6974 | }; |
6975 | |
6976 | // For now, just pick the first IV that was generated and inserted by |
6977 | // ScalarEvolution. Ideally pick an IV that is unlikely to be optimised away |
6978 | // by subsequent transforms. |
6979 | for (const WeakVH &IV : LSR.getScalarEvolutionIVs()) { |
6980 | if (!IV) |
6981 | continue; |
6982 | |
6983 | // There should only be PHI node IVs. |
6984 | PHINode *P = cast<PHINode>(Val: &*IV); |
6985 | |
6986 | if (IsSuitableIV(P)) |
6987 | return P; |
6988 | } |
6989 | |
6990 | for (PHINode &P : L.getHeader()->phis()) { |
6991 | if (IsSuitableIV(&P)) |
6992 | return &P; |
6993 | } |
6994 | return nullptr; |
6995 | } |
6996 | |
6997 | static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE, |
6998 | DominatorTree &DT, LoopInfo &LI, |
6999 | const TargetTransformInfo &TTI, |
7000 | AssumptionCache &AC, TargetLibraryInfo &TLI, |
7001 | MemorySSA *MSSA) { |
7002 | |
7003 | // Debug preservation - before we start removing anything identify which DVI |
7004 | // meet the salvageable criteria and store their DIExpression and SCEVs. |
7005 | SmallVector<std::unique_ptr<DVIRecoveryRec>, 2> SalvageableDVIRecords; |
7006 | SmallSet<AssertingVH<DbgValueInst>, 2> DVIHandles; |
7007 | DbgGatherSalvagableDVI(L, SE, SalvageableDVISCEVs&: SalvageableDVIRecords, DVIHandles); |
7008 | |
7009 | bool Changed = false; |
7010 | std::unique_ptr<MemorySSAUpdater> MSSAU; |
7011 | if (MSSA) |
7012 | MSSAU = std::make_unique<MemorySSAUpdater>(args&: MSSA); |
7013 | |
7014 | // Run the main LSR transformation. |
7015 | const LSRInstance &Reducer = |
7016 | LSRInstance(L, IU, SE, DT, LI, TTI, AC, TLI, MSSAU.get()); |
7017 | Changed |= Reducer.getChanged(); |
7018 | |
7019 | // Remove any extra phis created by processing inner loops. |
7020 | Changed |= DeleteDeadPHIs(BB: L->getHeader(), TLI: &TLI, MSSAU: MSSAU.get()); |
7021 | if (EnablePhiElim && L->isLoopSimplifyForm()) { |
7022 | SmallVector<WeakTrackingVH, 16> DeadInsts; |
7023 | const DataLayout &DL = L->getHeader()->getDataLayout(); |
7024 | SCEVExpander Rewriter(SE, DL, "lsr" , false); |
7025 | #if LLVM_ENABLE_ABI_BREAKING_CHECKS |
7026 | Rewriter.setDebugType(DEBUG_TYPE); |
7027 | #endif |
7028 | unsigned numFolded = Rewriter.replaceCongruentIVs(L, DT: &DT, DeadInsts, TTI: &TTI); |
7029 | Rewriter.clear(); |
7030 | if (numFolded) { |
7031 | Changed = true; |
7032 | RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts, TLI: &TLI, |
7033 | MSSAU: MSSAU.get()); |
7034 | DeleteDeadPHIs(BB: L->getHeader(), TLI: &TLI, MSSAU: MSSAU.get()); |
7035 | } |
7036 | } |
7037 | // LSR may at times remove all uses of an induction variable from a loop. |
7038 | // The only remaining use is the PHI in the exit block. |
7039 | // When this is the case, if the exit value of the IV can be calculated using |
7040 | // SCEV, we can replace the exit block PHI with the final value of the IV and |
7041 | // skip the updates in each loop iteration. |
7042 | if (L->isRecursivelyLCSSAForm(DT, LI) && L->getExitBlock()) { |
7043 | SmallVector<WeakTrackingVH, 16> DeadInsts; |
7044 | const DataLayout &DL = L->getHeader()->getDataLayout(); |
7045 | SCEVExpander Rewriter(SE, DL, "lsr" , true); |
7046 | int Rewrites = rewriteLoopExitValues(L, LI: &LI, TLI: &TLI, SE: &SE, TTI: &TTI, Rewriter, DT: &DT, |
7047 | ReplaceExitValue: UnusedIndVarInLoop, DeadInsts); |
7048 | Rewriter.clear(); |
7049 | if (Rewrites) { |
7050 | Changed = true; |
7051 | RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts, TLI: &TLI, |
7052 | MSSAU: MSSAU.get()); |
7053 | DeleteDeadPHIs(BB: L->getHeader(), TLI: &TLI, MSSAU: MSSAU.get()); |
7054 | } |
7055 | } |
7056 | |
7057 | if (SalvageableDVIRecords.empty()) |
7058 | return Changed; |
7059 | |
7060 | // Obtain relevant IVs and attempt to rewrite the salvageable DVIs with |
7061 | // expressions composed using the derived iteration count. |
7062 | // TODO: Allow for multiple IV references for nested AddRecSCEVs |
7063 | for (const auto &L : LI) { |
7064 | if (llvm::PHINode *IV = GetInductionVariable(L: *L, SE, LSR: Reducer)) |
7065 | DbgRewriteSalvageableDVIs(L, SE, LSRInductionVar: IV, DVIToUpdate&: SalvageableDVIRecords); |
7066 | else { |
7067 | LLVM_DEBUG(dbgs() << "scev-salvage: SCEV salvaging not possible. An IV " |
7068 | "could not be identified.\n" ); |
7069 | } |
7070 | } |
7071 | |
7072 | for (auto &Rec : SalvageableDVIRecords) |
7073 | Rec->clear(); |
7074 | SalvageableDVIRecords.clear(); |
7075 | DVIHandles.clear(); |
7076 | return Changed; |
7077 | } |
7078 | |
7079 | bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) { |
7080 | if (skipLoop(L)) |
7081 | return false; |
7082 | |
7083 | auto &IU = getAnalysis<IVUsersWrapperPass>().getIU(); |
7084 | auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE(); |
7085 | auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
7086 | auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); |
7087 | const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI( |
7088 | F: *L->getHeader()->getParent()); |
7089 | auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache( |
7090 | F&: *L->getHeader()->getParent()); |
7091 | auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI( |
7092 | F: *L->getHeader()->getParent()); |
7093 | auto *MSSAAnalysis = getAnalysisIfAvailable<MemorySSAWrapperPass>(); |
7094 | MemorySSA *MSSA = nullptr; |
7095 | if (MSSAAnalysis) |
7096 | MSSA = &MSSAAnalysis->getMSSA(); |
7097 | return ReduceLoopStrength(L, IU, SE, DT, LI, TTI, AC, TLI, MSSA); |
7098 | } |
7099 | |
7100 | PreservedAnalyses LoopStrengthReducePass::run(Loop &L, LoopAnalysisManager &AM, |
7101 | LoopStandardAnalysisResults &AR, |
7102 | LPMUpdater &) { |
7103 | if (!ReduceLoopStrength(L: &L, IU&: AM.getResult<IVUsersAnalysis>(IR&: L, ExtraArgs&: AR), SE&: AR.SE, |
7104 | DT&: AR.DT, LI&: AR.LI, TTI: AR.TTI, AC&: AR.AC, TLI&: AR.TLI, MSSA: AR.MSSA)) |
7105 | return PreservedAnalyses::all(); |
7106 | |
7107 | auto PA = getLoopPassPreservedAnalyses(); |
7108 | if (AR.MSSA) |
7109 | PA.preserve<MemorySSAAnalysis>(); |
7110 | return PA; |
7111 | } |
7112 | |
7113 | char LoopStrengthReduce::ID = 0; |
7114 | |
7115 | INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce" , |
7116 | "Loop Strength Reduction" , false, false) |
7117 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
7118 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
7119 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) |
7120 | INITIALIZE_PASS_DEPENDENCY(IVUsersWrapperPass) |
7121 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) |
7122 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify) |
7123 | INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce" , |
7124 | "Loop Strength Reduction" , false, false) |
7125 | |
7126 | Pass *llvm::createLoopStrengthReducePass() { return new LoopStrengthReduce(); } |
7127 | |