1 | //===- GVNSink.cpp - sink expressions into successors ---------------------===// |
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 | /// \file GVNSink.cpp |
10 | /// This pass attempts to sink instructions into successors, reducing static |
11 | /// instruction count and enabling if-conversion. |
12 | /// |
13 | /// We use a variant of global value numbering to decide what can be sunk. |
14 | /// Consider: |
15 | /// |
16 | /// [ %a1 = add i32 %b, 1 ] [ %c1 = add i32 %d, 1 ] |
17 | /// [ %a2 = xor i32 %a1, 1 ] [ %c2 = xor i32 %c1, 1 ] |
18 | /// \ / |
19 | /// [ %e = phi i32 %a2, %c2 ] |
20 | /// [ add i32 %e, 4 ] |
21 | /// |
22 | /// |
23 | /// GVN would number %a1 and %c1 differently because they compute different |
24 | /// results - the VN of an instruction is a function of its opcode and the |
25 | /// transitive closure of its operands. This is the key property for hoisting |
26 | /// and CSE. |
27 | /// |
28 | /// What we want when sinking however is for a numbering that is a function of |
29 | /// the *uses* of an instruction, which allows us to answer the question "if I |
30 | /// replace %a1 with %c1, will it contribute in an equivalent way to all |
31 | /// successive instructions?". The PostValueTable class in GVN provides this |
32 | /// mapping. |
33 | // |
34 | //===----------------------------------------------------------------------===// |
35 | |
36 | #include "llvm/ADT/ArrayRef.h" |
37 | #include "llvm/ADT/DenseMap.h" |
38 | #include "llvm/ADT/DenseSet.h" |
39 | #include "llvm/ADT/Hashing.h" |
40 | #include "llvm/ADT/PostOrderIterator.h" |
41 | #include "llvm/ADT/STLExtras.h" |
42 | #include "llvm/ADT/SmallPtrSet.h" |
43 | #include "llvm/ADT/SmallVector.h" |
44 | #include "llvm/ADT/Statistic.h" |
45 | #include "llvm/Analysis/GlobalsModRef.h" |
46 | #include "llvm/IR/BasicBlock.h" |
47 | #include "llvm/IR/CFG.h" |
48 | #include "llvm/IR/Constants.h" |
49 | #include "llvm/IR/Function.h" |
50 | #include "llvm/IR/InstrTypes.h" |
51 | #include "llvm/IR/Instruction.h" |
52 | #include "llvm/IR/Instructions.h" |
53 | #include "llvm/IR/PassManager.h" |
54 | #include "llvm/IR/Type.h" |
55 | #include "llvm/IR/Use.h" |
56 | #include "llvm/IR/Value.h" |
57 | #include "llvm/Support/Allocator.h" |
58 | #include "llvm/Support/ArrayRecycler.h" |
59 | #include "llvm/Support/AtomicOrdering.h" |
60 | #include "llvm/Support/Casting.h" |
61 | #include "llvm/Support/Compiler.h" |
62 | #include "llvm/Support/Debug.h" |
63 | #include "llvm/Support/raw_ostream.h" |
64 | #include "llvm/Transforms/Scalar/GVN.h" |
65 | #include "llvm/Transforms/Scalar/GVNExpression.h" |
66 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
67 | #include "llvm/Transforms/Utils/Local.h" |
68 | #include "llvm/Transforms/Utils/LockstepReverseIterator.h" |
69 | #include <cassert> |
70 | #include <cstddef> |
71 | #include <cstdint> |
72 | #include <iterator> |
73 | #include <utility> |
74 | |
75 | using namespace llvm; |
76 | |
77 | #define DEBUG_TYPE "gvn-sink" |
78 | |
79 | STATISTIC(NumRemoved, "Number of instructions removed" ); |
80 | |
81 | namespace llvm { |
82 | namespace GVNExpression { |
83 | |
84 | LLVM_DUMP_METHOD void Expression::dump() const { |
85 | print(OS&: dbgs()); |
86 | dbgs() << "\n" ; |
87 | } |
88 | |
89 | } // end namespace GVNExpression |
90 | } // end namespace llvm |
91 | |
92 | namespace { |
93 | |
94 | static bool isMemoryInst(const Instruction *I) { |
95 | return isa<LoadInst>(Val: I) || isa<StoreInst>(Val: I) || |
96 | (isa<InvokeInst>(Val: I) && !cast<InvokeInst>(Val: I)->doesNotAccessMemory()) || |
97 | (isa<CallInst>(Val: I) && !cast<CallInst>(Val: I)->doesNotAccessMemory()); |
98 | } |
99 | |
100 | //===----------------------------------------------------------------------===// |
101 | |
102 | /// Candidate solution for sinking. There may be different ways to |
103 | /// sink instructions, differing in the number of instructions sunk, |
104 | /// the number of predecessors sunk from and the number of PHIs |
105 | /// required. |
106 | struct SinkingInstructionCandidate { |
107 | unsigned NumBlocks; |
108 | unsigned NumInstructions; |
109 | unsigned NumPHIs; |
110 | unsigned NumMemoryInsts; |
111 | int Cost = -1; |
112 | SmallVector<BasicBlock *, 4> Blocks; |
113 | |
114 | void calculateCost(unsigned NumOrigPHIs, unsigned NumOrigBlocks) { |
115 | unsigned = NumPHIs - NumOrigPHIs; |
116 | unsigned SplitEdgeCost = (NumOrigBlocks > NumBlocks) ? 2 : 0; |
117 | Cost = (NumInstructions * (NumBlocks - 1)) - |
118 | (NumExtraPHIs * |
119 | NumExtraPHIs) // PHIs are expensive, so make sure they're worth it. |
120 | - SplitEdgeCost; |
121 | } |
122 | |
123 | bool operator>(const SinkingInstructionCandidate &Other) const { |
124 | return Cost > Other.Cost; |
125 | } |
126 | }; |
127 | |
128 | #ifndef NDEBUG |
129 | raw_ostream &operator<<(raw_ostream &OS, const SinkingInstructionCandidate &C) { |
130 | OS << "<Candidate Cost=" << C.Cost << " #Blocks=" << C.NumBlocks |
131 | << " #Insts=" << C.NumInstructions << " #PHIs=" << C.NumPHIs << ">" ; |
132 | return OS; |
133 | } |
134 | #endif |
135 | |
136 | //===----------------------------------------------------------------------===// |
137 | |
138 | /// Describes a PHI node that may or may not exist. These track the PHIs |
139 | /// that must be created if we sunk a sequence of instructions. It provides |
140 | /// a hash function for efficient equality comparisons. |
141 | class ModelledPHI { |
142 | SmallVector<Value *, 4> Values; |
143 | SmallVector<BasicBlock *, 4> Blocks; |
144 | |
145 | public: |
146 | ModelledPHI() = default; |
147 | |
148 | ModelledPHI(const PHINode *PN, |
149 | const DenseMap<const BasicBlock *, unsigned> &BlockOrder) { |
150 | // BasicBlock comes first so we sort by basic block pointer order, |
151 | // then by value pointer order. No need to call `verifyModelledPHI` |
152 | // As the Values and Blocks are populated in a deterministic order. |
153 | using OpsType = std::pair<BasicBlock *, Value *>; |
154 | SmallVector<OpsType, 4> Ops; |
155 | for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) |
156 | Ops.push_back(Elt: {PN->getIncomingBlock(i: I), PN->getIncomingValue(i: I)}); |
157 | |
158 | auto ComesBefore = [BlockOrder](OpsType O1, OpsType O2) { |
159 | return BlockOrder.lookup(Val: O1.first) < BlockOrder.lookup(Val: O2.first); |
160 | }; |
161 | // Sort in a deterministic order. |
162 | llvm::sort(C&: Ops, Comp: ComesBefore); |
163 | |
164 | for (auto &P : Ops) { |
165 | Blocks.push_back(Elt: P.first); |
166 | Values.push_back(Elt: P.second); |
167 | } |
168 | } |
169 | |
170 | /// Create a dummy ModelledPHI that will compare unequal to any other ModelledPHI |
171 | /// without the same ID. |
172 | /// \note This is specifically for DenseMapInfo - do not use this! |
173 | static ModelledPHI createDummy(size_t ID) { |
174 | ModelledPHI M; |
175 | M.Values.push_back(Elt: reinterpret_cast<Value*>(ID)); |
176 | return M; |
177 | } |
178 | |
179 | void |
180 | verifyModelledPHI(const DenseMap<const BasicBlock *, unsigned> &BlockOrder) { |
181 | assert(Values.size() > 1 && Blocks.size() > 1 && |
182 | "Modelling PHI with less than 2 values" ); |
183 | auto ComesBefore = [BlockOrder](const BasicBlock *BB1, |
184 | const BasicBlock *BB2) { |
185 | return BlockOrder.lookup(Val: BB1) < BlockOrder.lookup(Val: BB2); |
186 | }; |
187 | assert(llvm::is_sorted(Blocks, ComesBefore)); |
188 | int C = 0; |
189 | for (const Value *V : Values) { |
190 | if (!isa<UndefValue>(Val: V)) { |
191 | assert(cast<Instruction>(V)->getParent() == Blocks[C]); |
192 | (void)C; |
193 | } |
194 | C++; |
195 | } |
196 | } |
197 | /// Create a PHI from an array of incoming values and incoming blocks. |
198 | ModelledPHI(SmallVectorImpl<Instruction *> &V, |
199 | SmallSetVector<BasicBlock *, 4> &B, |
200 | const DenseMap<const BasicBlock *, unsigned> &BlockOrder) { |
201 | // The order of Values and Blocks are already ordered by the caller. |
202 | llvm::append_range(C&: Values, R&: V); |
203 | llvm::append_range(C&: Blocks, R&: B); |
204 | verifyModelledPHI(BlockOrder); |
205 | } |
206 | |
207 | /// Create a PHI from [I[OpNum] for I in Insts]. |
208 | /// TODO: Figure out a way to verifyModelledPHI in this constructor. |
209 | ModelledPHI(ArrayRef<Instruction *> Insts, unsigned OpNum, |
210 | SmallSetVector<BasicBlock *, 4> &B) { |
211 | llvm::append_range(C&: Blocks, R&: B); |
212 | for (auto *I : Insts) |
213 | Values.push_back(Elt: I->getOperand(i: OpNum)); |
214 | } |
215 | |
216 | /// Restrict the PHI's contents down to only \c NewBlocks. |
217 | /// \c NewBlocks must be a subset of \c this->Blocks. |
218 | void restrictToBlocks(const SmallSetVector<BasicBlock *, 4> &NewBlocks) { |
219 | auto BI = Blocks.begin(); |
220 | auto VI = Values.begin(); |
221 | while (BI != Blocks.end()) { |
222 | assert(VI != Values.end()); |
223 | if (!NewBlocks.contains(key: *BI)) { |
224 | BI = Blocks.erase(CI: BI); |
225 | VI = Values.erase(CI: VI); |
226 | } else { |
227 | ++BI; |
228 | ++VI; |
229 | } |
230 | } |
231 | assert(Blocks.size() == NewBlocks.size()); |
232 | } |
233 | |
234 | ArrayRef<Value *> getValues() const { return Values; } |
235 | |
236 | bool areAllIncomingValuesSame() const { |
237 | return llvm::all_equal(Range: Values); |
238 | } |
239 | |
240 | bool areAllIncomingValuesSameType() const { |
241 | return llvm::all_of( |
242 | Range: Values, P: [&](Value *V) { return V->getType() == Values[0]->getType(); }); |
243 | } |
244 | |
245 | bool areAnyIncomingValuesConstant() const { |
246 | return llvm::any_of(Range: Values, P: [&](Value *V) { return isa<Constant>(Val: V); }); |
247 | } |
248 | |
249 | // Hash functor |
250 | unsigned hash() const { |
251 | // Is deterministic because Values are saved in a specific order. |
252 | return (unsigned)hash_combine_range(R: Values); |
253 | } |
254 | |
255 | bool operator==(const ModelledPHI &Other) const { |
256 | return Values == Other.Values && Blocks == Other.Blocks; |
257 | } |
258 | }; |
259 | |
260 | template <typename ModelledPHI> struct DenseMapInfo { |
261 | static inline ModelledPHI &getEmptyKey() { |
262 | static ModelledPHI Dummy = ModelledPHI::createDummy(0); |
263 | return Dummy; |
264 | } |
265 | |
266 | static inline ModelledPHI &getTombstoneKey() { |
267 | static ModelledPHI Dummy = ModelledPHI::createDummy(1); |
268 | return Dummy; |
269 | } |
270 | |
271 | static unsigned getHashValue(const ModelledPHI &V) { return V.hash(); } |
272 | |
273 | static bool isEqual(const ModelledPHI &LHS, const ModelledPHI &RHS) { |
274 | return LHS == RHS; |
275 | } |
276 | }; |
277 | |
278 | using ModelledPHISet = DenseSet<ModelledPHI, DenseMapInfo<ModelledPHI>>; |
279 | |
280 | //===----------------------------------------------------------------------===// |
281 | // ValueTable |
282 | //===----------------------------------------------------------------------===// |
283 | // This is a value number table where the value number is a function of the |
284 | // *uses* of a value, rather than its operands. Thus, if VN(A) == VN(B) we know |
285 | // that the program would be equivalent if we replaced A with PHI(A, B). |
286 | //===----------------------------------------------------------------------===// |
287 | |
288 | /// A GVN expression describing how an instruction is used. The operands |
289 | /// field of BasicExpression is used to store uses, not operands. |
290 | /// |
291 | /// This class also contains fields for discriminators used when determining |
292 | /// equivalence of instructions with sideeffects. |
293 | class InstructionUseExpr : public GVNExpression::BasicExpression { |
294 | unsigned MemoryUseOrder = -1; |
295 | bool Volatile = false; |
296 | ArrayRef<int> ShuffleMask; |
297 | |
298 | public: |
299 | InstructionUseExpr(Instruction *I, ArrayRecycler<Value *> &R, |
300 | BumpPtrAllocator &A) |
301 | : GVNExpression::BasicExpression(I->getNumUses()) { |
302 | allocateOperands(Recycler&: R, Allocator&: A); |
303 | setOpcode(I->getOpcode()); |
304 | setType(I->getType()); |
305 | |
306 | if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Val: I)) |
307 | ShuffleMask = SVI->getShuffleMask().copy(A); |
308 | |
309 | for (auto &U : I->uses()) |
310 | op_push_back(Arg: U.getUser()); |
311 | llvm::sort(Start: op_begin(), End: op_end()); |
312 | } |
313 | |
314 | void setMemoryUseOrder(unsigned MUO) { MemoryUseOrder = MUO; } |
315 | void setVolatile(bool V) { Volatile = V; } |
316 | |
317 | hash_code getHashValue() const override { |
318 | return hash_combine(args: GVNExpression::BasicExpression::getHashValue(), |
319 | args: MemoryUseOrder, args: Volatile, args: ShuffleMask); |
320 | } |
321 | |
322 | template <typename Function> hash_code getHashValue(Function MapFn) { |
323 | hash_code H = hash_combine(args: getOpcode(), args: getType(), args: MemoryUseOrder, args: Volatile, |
324 | args: ShuffleMask); |
325 | for (auto *V : operands()) |
326 | H = hash_combine(H, MapFn(V)); |
327 | return H; |
328 | } |
329 | }; |
330 | |
331 | using BasicBlocksSet = SmallPtrSet<const BasicBlock *, 32>; |
332 | |
333 | class ValueTable { |
334 | DenseMap<Value *, uint32_t> ValueNumbering; |
335 | DenseMap<GVNExpression::Expression *, uint32_t> ExpressionNumbering; |
336 | DenseMap<size_t, uint32_t> HashNumbering; |
337 | BumpPtrAllocator Allocator; |
338 | ArrayRecycler<Value *> Recycler; |
339 | uint32_t nextValueNumber = 1; |
340 | BasicBlocksSet ReachableBBs; |
341 | |
342 | /// Create an expression for I based on its opcode and its uses. If I |
343 | /// touches or reads memory, the expression is also based upon its memory |
344 | /// order - see \c getMemoryUseOrder(). |
345 | InstructionUseExpr *createExpr(Instruction *I) { |
346 | InstructionUseExpr *E = |
347 | new (Allocator) InstructionUseExpr(I, Recycler, Allocator); |
348 | if (isMemoryInst(I)) |
349 | E->setMemoryUseOrder(getMemoryUseOrder(Inst: I)); |
350 | |
351 | if (CmpInst *C = dyn_cast<CmpInst>(Val: I)) { |
352 | CmpInst::Predicate Predicate = C->getPredicate(); |
353 | E->setOpcode((C->getOpcode() << 8) | Predicate); |
354 | } |
355 | return E; |
356 | } |
357 | |
358 | /// Helper to compute the value number for a memory instruction |
359 | /// (LoadInst/StoreInst), including checking the memory ordering and |
360 | /// volatility. |
361 | template <class Inst> InstructionUseExpr *createMemoryExpr(Inst *I) { |
362 | if (isStrongerThanUnordered(I->getOrdering()) || I->isAtomic()) |
363 | return nullptr; |
364 | InstructionUseExpr *E = createExpr(I); |
365 | E->setVolatile(I->isVolatile()); |
366 | return E; |
367 | } |
368 | |
369 | public: |
370 | ValueTable() = default; |
371 | |
372 | /// Set basic blocks reachable from entry block. |
373 | void setReachableBBs(const BasicBlocksSet &ReachableBBs) { |
374 | this->ReachableBBs = ReachableBBs; |
375 | } |
376 | |
377 | /// Returns the value number for the specified value, assigning |
378 | /// it a new number if it did not have one before. |
379 | uint32_t lookupOrAdd(Value *V) { |
380 | auto VI = ValueNumbering.find(Val: V); |
381 | if (VI != ValueNumbering.end()) |
382 | return VI->second; |
383 | |
384 | if (!isa<Instruction>(Val: V)) { |
385 | ValueNumbering[V] = nextValueNumber; |
386 | return nextValueNumber++; |
387 | } |
388 | |
389 | Instruction *I = cast<Instruction>(Val: V); |
390 | if (!ReachableBBs.contains(Ptr: I->getParent())) |
391 | return ~0U; |
392 | |
393 | InstructionUseExpr *exp = nullptr; |
394 | switch (I->getOpcode()) { |
395 | case Instruction::Load: |
396 | exp = createMemoryExpr(I: cast<LoadInst>(Val: I)); |
397 | break; |
398 | case Instruction::Store: |
399 | exp = createMemoryExpr(I: cast<StoreInst>(Val: I)); |
400 | break; |
401 | case Instruction::Call: |
402 | case Instruction::Invoke: |
403 | case Instruction::FNeg: |
404 | case Instruction::Add: |
405 | case Instruction::FAdd: |
406 | case Instruction::Sub: |
407 | case Instruction::FSub: |
408 | case Instruction::Mul: |
409 | case Instruction::FMul: |
410 | case Instruction::UDiv: |
411 | case Instruction::SDiv: |
412 | case Instruction::FDiv: |
413 | case Instruction::URem: |
414 | case Instruction::SRem: |
415 | case Instruction::FRem: |
416 | case Instruction::Shl: |
417 | case Instruction::LShr: |
418 | case Instruction::AShr: |
419 | case Instruction::And: |
420 | case Instruction::Or: |
421 | case Instruction::Xor: |
422 | case Instruction::ICmp: |
423 | case Instruction::FCmp: |
424 | case Instruction::Trunc: |
425 | case Instruction::ZExt: |
426 | case Instruction::SExt: |
427 | case Instruction::FPToUI: |
428 | case Instruction::FPToSI: |
429 | case Instruction::UIToFP: |
430 | case Instruction::SIToFP: |
431 | case Instruction::FPTrunc: |
432 | case Instruction::FPExt: |
433 | case Instruction::PtrToInt: |
434 | case Instruction::IntToPtr: |
435 | case Instruction::BitCast: |
436 | case Instruction::AddrSpaceCast: |
437 | case Instruction::Select: |
438 | case Instruction::ExtractElement: |
439 | case Instruction::InsertElement: |
440 | case Instruction::ShuffleVector: |
441 | case Instruction::InsertValue: |
442 | case Instruction::GetElementPtr: |
443 | exp = createExpr(I); |
444 | break; |
445 | default: |
446 | break; |
447 | } |
448 | |
449 | if (!exp) { |
450 | ValueNumbering[V] = nextValueNumber; |
451 | return nextValueNumber++; |
452 | } |
453 | |
454 | uint32_t e = ExpressionNumbering[exp]; |
455 | if (!e) { |
456 | hash_code H = exp->getHashValue(MapFn: [=](Value *V) { return lookupOrAdd(V); }); |
457 | auto [I, Inserted] = HashNumbering.try_emplace(Key: H, Args&: nextValueNumber); |
458 | e = I->second; |
459 | if (Inserted) |
460 | ExpressionNumbering[exp] = nextValueNumber++; |
461 | } |
462 | ValueNumbering[V] = e; |
463 | return e; |
464 | } |
465 | |
466 | /// Returns the value number of the specified value. Fails if the value has |
467 | /// not yet been numbered. |
468 | uint32_t lookup(Value *V) const { |
469 | auto VI = ValueNumbering.find(Val: V); |
470 | assert(VI != ValueNumbering.end() && "Value not numbered?" ); |
471 | return VI->second; |
472 | } |
473 | |
474 | /// Removes all value numberings and resets the value table. |
475 | void clear() { |
476 | ValueNumbering.clear(); |
477 | ExpressionNumbering.clear(); |
478 | HashNumbering.clear(); |
479 | Recycler.clear(Allocator); |
480 | nextValueNumber = 1; |
481 | } |
482 | |
483 | /// \c Inst uses or touches memory. Return an ID describing the memory state |
484 | /// at \c Inst such that if getMemoryUseOrder(I1) == getMemoryUseOrder(I2), |
485 | /// the exact same memory operations happen after I1 and I2. |
486 | /// |
487 | /// This is a very hard problem in general, so we use domain-specific |
488 | /// knowledge that we only ever check for equivalence between blocks sharing a |
489 | /// single immediate successor that is common, and when determining if I1 == |
490 | /// I2 we will have already determined that next(I1) == next(I2). This |
491 | /// inductive property allows us to simply return the value number of the next |
492 | /// instruction that defines memory. |
493 | uint32_t getMemoryUseOrder(Instruction *Inst) { |
494 | auto *BB = Inst->getParent(); |
495 | for (auto I = std::next(x: Inst->getIterator()), E = BB->end(); |
496 | I != E && !I->isTerminator(); ++I) { |
497 | if (!isMemoryInst(I: &*I)) |
498 | continue; |
499 | if (isa<LoadInst>(Val: &*I)) |
500 | continue; |
501 | CallInst *CI = dyn_cast<CallInst>(Val: &*I); |
502 | if (CI && CI->onlyReadsMemory()) |
503 | continue; |
504 | InvokeInst *II = dyn_cast<InvokeInst>(Val: &*I); |
505 | if (II && II->onlyReadsMemory()) |
506 | continue; |
507 | return lookupOrAdd(V: &*I); |
508 | } |
509 | return 0; |
510 | } |
511 | }; |
512 | |
513 | //===----------------------------------------------------------------------===// |
514 | |
515 | class GVNSink { |
516 | public: |
517 | GVNSink() {} |
518 | |
519 | bool run(Function &F) { |
520 | LLVM_DEBUG(dbgs() << "GVNSink: running on function @" << F.getName() |
521 | << "\n" ); |
522 | |
523 | unsigned NumSunk = 0; |
524 | ReversePostOrderTraversal<Function*> RPOT(&F); |
525 | VN.setReachableBBs(BasicBlocksSet(llvm::from_range, RPOT)); |
526 | // Populate reverse post-order to order basic blocks in deterministic |
527 | // order. Any arbitrary ordering will work in this case as long as they are |
528 | // deterministic. The node ordering of newly created basic blocks |
529 | // are irrelevant because RPOT(for computing sinkable candidates) is also |
530 | // obtained ahead of time and only their order are relevant for this pass. |
531 | unsigned NodeOrdering = 0; |
532 | RPOTOrder[*RPOT.begin()] = ++NodeOrdering; |
533 | for (auto *BB : RPOT) |
534 | if (!pred_empty(BB)) |
535 | RPOTOrder[BB] = ++NodeOrdering; |
536 | for (auto *N : RPOT) |
537 | NumSunk += sinkBB(BBEnd: N); |
538 | |
539 | return NumSunk > 0; |
540 | } |
541 | |
542 | private: |
543 | ValueTable VN; |
544 | DenseMap<const BasicBlock *, unsigned> RPOTOrder; |
545 | |
546 | bool shouldAvoidSinkingInstruction(Instruction *I) { |
547 | // These instructions may change or break semantics if moved. |
548 | if (isa<PHINode>(Val: I) || I->isEHPad() || isa<AllocaInst>(Val: I) || |
549 | I->getType()->isTokenTy()) |
550 | return true; |
551 | return false; |
552 | } |
553 | |
554 | /// The main heuristic function. Analyze the set of instructions pointed to by |
555 | /// LRI and return a candidate solution if these instructions can be sunk, or |
556 | /// std::nullopt otherwise. |
557 | std::optional<SinkingInstructionCandidate> |
558 | analyzeInstructionForSinking(LockstepReverseIterator<false> &LRI, |
559 | unsigned &InstNum, unsigned &MemoryInstNum, |
560 | ModelledPHISet &NeededPHIs, |
561 | SmallPtrSetImpl<Value *> &PHIContents); |
562 | |
563 | /// Create a ModelledPHI for each PHI in BB, adding to PHIs. |
564 | void analyzeInitialPHIs(BasicBlock *BB, ModelledPHISet &PHIs, |
565 | SmallPtrSetImpl<Value *> &PHIContents) { |
566 | for (PHINode &PN : BB->phis()) { |
567 | auto MPHI = ModelledPHI(&PN, RPOTOrder); |
568 | PHIs.insert(V: MPHI); |
569 | PHIContents.insert_range(R: MPHI.getValues()); |
570 | } |
571 | } |
572 | |
573 | /// The main instruction sinking driver. Set up state and try and sink |
574 | /// instructions into BBEnd from its predecessors. |
575 | unsigned sinkBB(BasicBlock *BBEnd); |
576 | |
577 | /// Perform the actual mechanics of sinking an instruction from Blocks into |
578 | /// BBEnd, which is their only successor. |
579 | void sinkLastInstruction(ArrayRef<BasicBlock *> Blocks, BasicBlock *BBEnd); |
580 | |
581 | /// Remove PHIs that all have the same incoming value. |
582 | void foldPointlessPHINodes(BasicBlock *BB) { |
583 | auto I = BB->begin(); |
584 | while (PHINode *PN = dyn_cast<PHINode>(Val: I++)) { |
585 | if (!llvm::all_of(Range: PN->incoming_values(), P: [&](const Value *V) { |
586 | return V == PN->getIncomingValue(i: 0); |
587 | })) |
588 | continue; |
589 | if (PN->getIncomingValue(i: 0) != PN) |
590 | PN->replaceAllUsesWith(V: PN->getIncomingValue(i: 0)); |
591 | else |
592 | PN->replaceAllUsesWith(V: PoisonValue::get(T: PN->getType())); |
593 | PN->eraseFromParent(); |
594 | } |
595 | } |
596 | }; |
597 | |
598 | std::optional<SinkingInstructionCandidate> |
599 | GVNSink::analyzeInstructionForSinking(LockstepReverseIterator<false> &LRI, |
600 | unsigned &InstNum, |
601 | unsigned &MemoryInstNum, |
602 | ModelledPHISet &NeededPHIs, |
603 | SmallPtrSetImpl<Value *> &PHIContents) { |
604 | auto Insts = *LRI; |
605 | LLVM_DEBUG(dbgs() << " -- Analyzing instruction set: [\n" ; for (auto *I |
606 | : Insts) { |
607 | I->dump(); |
608 | } dbgs() << " ]\n" ;); |
609 | |
610 | DenseMap<uint32_t, unsigned> VNums; |
611 | for (auto *I : Insts) { |
612 | uint32_t N = VN.lookupOrAdd(V: I); |
613 | LLVM_DEBUG(dbgs() << " VN=" << Twine::utohexstr(N) << " for" << *I << "\n" ); |
614 | if (N == ~0U) |
615 | return std::nullopt; |
616 | VNums[N]++; |
617 | } |
618 | unsigned VNumToSink = llvm::max_element(Range&: VNums, C: llvm::less_second())->first; |
619 | |
620 | if (VNums[VNumToSink] == 1) |
621 | // Can't sink anything! |
622 | return std::nullopt; |
623 | |
624 | // Now restrict the number of incoming blocks down to only those with |
625 | // VNumToSink. |
626 | auto &ActivePreds = LRI.getActiveBlocks(); |
627 | unsigned InitialActivePredSize = ActivePreds.size(); |
628 | SmallVector<Instruction *, 4> NewInsts; |
629 | for (auto *I : Insts) { |
630 | if (VN.lookup(V: I) != VNumToSink) |
631 | ActivePreds.remove(X: I->getParent()); |
632 | else |
633 | NewInsts.push_back(Elt: I); |
634 | } |
635 | for (auto *I : NewInsts) |
636 | if (shouldAvoidSinkingInstruction(I)) |
637 | return std::nullopt; |
638 | |
639 | // If we've restricted the incoming blocks, restrict all needed PHIs also |
640 | // to that set. |
641 | bool RecomputePHIContents = false; |
642 | if (ActivePreds.size() != InitialActivePredSize) { |
643 | ModelledPHISet NewNeededPHIs; |
644 | for (auto P : NeededPHIs) { |
645 | P.restrictToBlocks(NewBlocks: ActivePreds); |
646 | NewNeededPHIs.insert(V: P); |
647 | } |
648 | NeededPHIs = NewNeededPHIs; |
649 | LRI.restrictToBlocks(Blocks&: ActivePreds); |
650 | RecomputePHIContents = true; |
651 | } |
652 | |
653 | // The sunk instruction's results. |
654 | ModelledPHI NewPHI(NewInsts, ActivePreds, RPOTOrder); |
655 | |
656 | // Does sinking this instruction render previous PHIs redundant? |
657 | if (NeededPHIs.erase(V: NewPHI)) |
658 | RecomputePHIContents = true; |
659 | |
660 | if (RecomputePHIContents) { |
661 | // The needed PHIs have changed, so recompute the set of all needed |
662 | // values. |
663 | PHIContents.clear(); |
664 | for (auto &PHI : NeededPHIs) |
665 | PHIContents.insert_range(R: PHI.getValues()); |
666 | } |
667 | |
668 | // Is this instruction required by a later PHI that doesn't match this PHI? |
669 | // if so, we can't sink this instruction. |
670 | for (auto *V : NewPHI.getValues()) |
671 | if (PHIContents.count(Ptr: V)) |
672 | // V exists in this PHI, but the whole PHI is different to NewPHI |
673 | // (else it would have been removed earlier). We cannot continue |
674 | // because this isn't representable. |
675 | return std::nullopt; |
676 | |
677 | // Which operands need PHIs? |
678 | // FIXME: If any of these fail, we should partition up the candidates to |
679 | // try and continue making progress. |
680 | Instruction *I0 = NewInsts[0]; |
681 | |
682 | auto isNotSameOperation = [&I0](Instruction *I) { |
683 | return !I0->isSameOperationAs(I); |
684 | }; |
685 | |
686 | if (any_of(Range&: NewInsts, P: isNotSameOperation)) |
687 | return std::nullopt; |
688 | |
689 | for (unsigned OpNum = 0, E = I0->getNumOperands(); OpNum != E; ++OpNum) { |
690 | ModelledPHI PHI(NewInsts, OpNum, ActivePreds); |
691 | if (PHI.areAllIncomingValuesSame()) |
692 | continue; |
693 | if (!canReplaceOperandWithVariable(I: I0, OpIdx: OpNum)) |
694 | // We can 't create a PHI from this instruction! |
695 | return std::nullopt; |
696 | if (NeededPHIs.count(V: PHI)) |
697 | continue; |
698 | if (!PHI.areAllIncomingValuesSameType()) |
699 | return std::nullopt; |
700 | // Don't create indirect calls! The called value is the final operand. |
701 | if ((isa<CallInst>(Val: I0) || isa<InvokeInst>(Val: I0)) && OpNum == E - 1 && |
702 | PHI.areAnyIncomingValuesConstant()) |
703 | return std::nullopt; |
704 | |
705 | NeededPHIs.reserve(Size: NeededPHIs.size()); |
706 | NeededPHIs.insert(V: PHI); |
707 | PHIContents.insert_range(R: PHI.getValues()); |
708 | } |
709 | |
710 | if (isMemoryInst(I: NewInsts[0])) |
711 | ++MemoryInstNum; |
712 | |
713 | SinkingInstructionCandidate Cand; |
714 | Cand.NumInstructions = ++InstNum; |
715 | Cand.NumMemoryInsts = MemoryInstNum; |
716 | Cand.NumBlocks = ActivePreds.size(); |
717 | Cand.NumPHIs = NeededPHIs.size(); |
718 | append_range(C&: Cand.Blocks, R&: ActivePreds); |
719 | |
720 | return Cand; |
721 | } |
722 | |
723 | unsigned GVNSink::sinkBB(BasicBlock *BBEnd) { |
724 | LLVM_DEBUG(dbgs() << "GVNSink: running on basic block " ; |
725 | BBEnd->printAsOperand(dbgs()); dbgs() << "\n" ); |
726 | SmallVector<BasicBlock *, 4> Preds; |
727 | for (auto *B : predecessors(BB: BBEnd)) { |
728 | // Bailout on basic blocks without predecessor(PR42346). |
729 | if (!RPOTOrder.count(Val: B)) |
730 | return 0; |
731 | auto *T = B->getTerminator(); |
732 | if (isa<BranchInst>(Val: T) || isa<SwitchInst>(Val: T)) |
733 | Preds.push_back(Elt: B); |
734 | else |
735 | return 0; |
736 | } |
737 | if (Preds.size() < 2) |
738 | return 0; |
739 | auto ComesBefore = [this](const BasicBlock *BB1, const BasicBlock *BB2) { |
740 | return RPOTOrder.lookup(Val: BB1) < RPOTOrder.lookup(Val: BB2); |
741 | }; |
742 | // Sort in a deterministic order. |
743 | llvm::sort(C&: Preds, Comp: ComesBefore); |
744 | |
745 | unsigned NumOrigPreds = Preds.size(); |
746 | // We can only sink instructions through unconditional branches. |
747 | llvm::erase_if(C&: Preds, P: [](BasicBlock *BB) { |
748 | return BB->getTerminator()->getNumSuccessors() != 1; |
749 | }); |
750 | |
751 | LockstepReverseIterator<false> LRI(Preds); |
752 | SmallVector<SinkingInstructionCandidate, 4> Candidates; |
753 | unsigned InstNum = 0, MemoryInstNum = 0; |
754 | ModelledPHISet NeededPHIs; |
755 | SmallPtrSet<Value *, 4> PHIContents; |
756 | analyzeInitialPHIs(BB: BBEnd, PHIs&: NeededPHIs, PHIContents); |
757 | unsigned NumOrigPHIs = NeededPHIs.size(); |
758 | |
759 | while (LRI.isValid()) { |
760 | auto Cand = analyzeInstructionForSinking(LRI, InstNum, MemoryInstNum, |
761 | NeededPHIs, PHIContents); |
762 | if (!Cand) |
763 | break; |
764 | Cand->calculateCost(NumOrigPHIs, NumOrigBlocks: Preds.size()); |
765 | Candidates.emplace_back(Args&: *Cand); |
766 | --LRI; |
767 | } |
768 | |
769 | llvm::stable_sort(Range&: Candidates, C: std::greater<SinkingInstructionCandidate>()); |
770 | LLVM_DEBUG(dbgs() << " -- Sinking candidates:\n" ; for (auto &C |
771 | : Candidates) dbgs() |
772 | << " " << C << "\n" ;); |
773 | |
774 | // Pick the top candidate, as long it is positive! |
775 | if (Candidates.empty() || Candidates.front().Cost <= 0) |
776 | return 0; |
777 | auto C = Candidates.front(); |
778 | |
779 | LLVM_DEBUG(dbgs() << " -- Sinking: " << C << "\n" ); |
780 | BasicBlock *InsertBB = BBEnd; |
781 | if (C.Blocks.size() < NumOrigPreds) { |
782 | LLVM_DEBUG(dbgs() << " -- Splitting edge to " ; |
783 | BBEnd->printAsOperand(dbgs()); dbgs() << "\n" ); |
784 | InsertBB = SplitBlockPredecessors(BB: BBEnd, Preds: C.Blocks, Suffix: ".gvnsink.split" ); |
785 | if (!InsertBB) { |
786 | LLVM_DEBUG(dbgs() << " -- FAILED to split edge!\n" ); |
787 | // Edge couldn't be split. |
788 | return 0; |
789 | } |
790 | } |
791 | |
792 | for (unsigned I = 0; I < C.NumInstructions; ++I) |
793 | sinkLastInstruction(Blocks: C.Blocks, BBEnd: InsertBB); |
794 | |
795 | return C.NumInstructions; |
796 | } |
797 | |
798 | void GVNSink::sinkLastInstruction(ArrayRef<BasicBlock *> Blocks, |
799 | BasicBlock *BBEnd) { |
800 | SmallVector<Instruction *, 4> Insts; |
801 | for (BasicBlock *BB : Blocks) |
802 | Insts.push_back(Elt: BB->getTerminator()->getPrevNonDebugInstruction()); |
803 | Instruction *I0 = Insts.front(); |
804 | |
805 | SmallVector<Value *, 4> NewOperands; |
806 | for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) { |
807 | bool NeedPHI = llvm::any_of(Range&: Insts, P: [&I0, O](const Instruction *I) { |
808 | return I->getOperand(i: O) != I0->getOperand(i: O); |
809 | }); |
810 | if (!NeedPHI) { |
811 | NewOperands.push_back(Elt: I0->getOperand(i: O)); |
812 | continue; |
813 | } |
814 | |
815 | // Create a new PHI in the successor block and populate it. |
816 | auto *Op = I0->getOperand(i: O); |
817 | assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!" ); |
818 | auto *PN = |
819 | PHINode::Create(Ty: Op->getType(), NumReservedValues: Insts.size(), NameStr: Op->getName() + ".sink" ); |
820 | PN->insertBefore(InsertPos: BBEnd->begin()); |
821 | for (auto *I : Insts) |
822 | PN->addIncoming(V: I->getOperand(i: O), BB: I->getParent()); |
823 | NewOperands.push_back(Elt: PN); |
824 | } |
825 | |
826 | // Arbitrarily use I0 as the new "common" instruction; remap its operands |
827 | // and move it to the start of the successor block. |
828 | for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) |
829 | I0->getOperandUse(i: O).set(NewOperands[O]); |
830 | I0->moveBefore(InsertPos: BBEnd->getFirstInsertionPt()); |
831 | |
832 | // Update metadata and IR flags. |
833 | for (auto *I : Insts) |
834 | if (I != I0) { |
835 | combineMetadataForCSE(K: I0, J: I, DoesKMove: true); |
836 | I0->andIRFlags(V: I); |
837 | } |
838 | |
839 | for (auto *I : Insts) |
840 | if (I != I0) { |
841 | I->replaceAllUsesWith(V: I0); |
842 | I0->applyMergedLocation(LocA: I0->getDebugLoc(), LocB: I->getDebugLoc()); |
843 | } |
844 | foldPointlessPHINodes(BB: BBEnd); |
845 | |
846 | // Finally nuke all instructions apart from the common instruction. |
847 | for (auto *I : Insts) |
848 | if (I != I0) |
849 | I->eraseFromParent(); |
850 | |
851 | NumRemoved += Insts.size() - 1; |
852 | } |
853 | |
854 | } // end anonymous namespace |
855 | |
856 | PreservedAnalyses GVNSinkPass::run(Function &F, FunctionAnalysisManager &AM) { |
857 | GVNSink G; |
858 | if (!G.run(F)) |
859 | return PreservedAnalyses::all(); |
860 | |
861 | return PreservedAnalyses::none(); |
862 | } |
863 | |