1 | //===- GVN.cpp - Eliminate redundant values and loads ---------------------===// |
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 pass performs global value numbering to eliminate fully redundant |
10 | // instructions. It also performs simple dead load elimination. |
11 | // |
12 | // Note that this pass does the value numbering itself; it does not use the |
13 | // ValueNumbering analysis passes. |
14 | // |
15 | //===----------------------------------------------------------------------===// |
16 | |
17 | #include "llvm/Transforms/Scalar/GVN.h" |
18 | #include "llvm/ADT/DenseMap.h" |
19 | #include "llvm/ADT/DepthFirstIterator.h" |
20 | #include "llvm/ADT/Hashing.h" |
21 | #include "llvm/ADT/MapVector.h" |
22 | #include "llvm/ADT/PostOrderIterator.h" |
23 | #include "llvm/ADT/STLExtras.h" |
24 | #include "llvm/ADT/SetVector.h" |
25 | #include "llvm/ADT/SmallPtrSet.h" |
26 | #include "llvm/ADT/SmallVector.h" |
27 | #include "llvm/ADT/Statistic.h" |
28 | #include "llvm/Analysis/AliasAnalysis.h" |
29 | #include "llvm/Analysis/AssumeBundleQueries.h" |
30 | #include "llvm/Analysis/AssumptionCache.h" |
31 | #include "llvm/Analysis/CFG.h" |
32 | #include "llvm/Analysis/DomTreeUpdater.h" |
33 | #include "llvm/Analysis/GlobalsModRef.h" |
34 | #include "llvm/Analysis/InstructionPrecedenceTracking.h" |
35 | #include "llvm/Analysis/InstructionSimplify.h" |
36 | #include "llvm/Analysis/LoopInfo.h" |
37 | #include "llvm/Analysis/MemoryBuiltins.h" |
38 | #include "llvm/Analysis/MemoryDependenceAnalysis.h" |
39 | #include "llvm/Analysis/MemorySSA.h" |
40 | #include "llvm/Analysis/MemorySSAUpdater.h" |
41 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
42 | #include "llvm/Analysis/PHITransAddr.h" |
43 | #include "llvm/Analysis/TargetLibraryInfo.h" |
44 | #include "llvm/Analysis/ValueTracking.h" |
45 | #include "llvm/IR/Attributes.h" |
46 | #include "llvm/IR/BasicBlock.h" |
47 | #include "llvm/IR/Constant.h" |
48 | #include "llvm/IR/Constants.h" |
49 | #include "llvm/IR/DebugLoc.h" |
50 | #include "llvm/IR/Dominators.h" |
51 | #include "llvm/IR/Function.h" |
52 | #include "llvm/IR/InstrTypes.h" |
53 | #include "llvm/IR/Instruction.h" |
54 | #include "llvm/IR/Instructions.h" |
55 | #include "llvm/IR/IntrinsicInst.h" |
56 | #include "llvm/IR/LLVMContext.h" |
57 | #include "llvm/IR/Metadata.h" |
58 | #include "llvm/IR/Module.h" |
59 | #include "llvm/IR/PassManager.h" |
60 | #include "llvm/IR/PatternMatch.h" |
61 | #include "llvm/IR/Type.h" |
62 | #include "llvm/IR/Use.h" |
63 | #include "llvm/IR/Value.h" |
64 | #include "llvm/InitializePasses.h" |
65 | #include "llvm/Pass.h" |
66 | #include "llvm/Support/Casting.h" |
67 | #include "llvm/Support/CommandLine.h" |
68 | #include "llvm/Support/Compiler.h" |
69 | #include "llvm/Support/Debug.h" |
70 | #include "llvm/Support/raw_ostream.h" |
71 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
72 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
73 | #include "llvm/Transforms/Utils/Local.h" |
74 | #include "llvm/Transforms/Utils/SSAUpdater.h" |
75 | #include "llvm/Transforms/Utils/VNCoercion.h" |
76 | #include <algorithm> |
77 | #include <cassert> |
78 | #include <cstdint> |
79 | #include <optional> |
80 | #include <utility> |
81 | |
82 | using namespace llvm; |
83 | using namespace llvm::gvn; |
84 | using namespace llvm::VNCoercion; |
85 | using namespace PatternMatch; |
86 | |
87 | #define DEBUG_TYPE "gvn" |
88 | |
89 | STATISTIC(NumGVNInstr, "Number of instructions deleted" ); |
90 | STATISTIC(NumGVNLoad, "Number of loads deleted" ); |
91 | STATISTIC(NumGVNPRE, "Number of instructions PRE'd" ); |
92 | STATISTIC(NumGVNBlocks, "Number of blocks merged" ); |
93 | STATISTIC(NumGVNSimpl, "Number of instructions simplified" ); |
94 | STATISTIC(NumGVNEqProp, "Number of equalities propagated" ); |
95 | STATISTIC(NumPRELoad, "Number of loads PRE'd" ); |
96 | STATISTIC(NumPRELoopLoad, "Number of loop loads PRE'd" ); |
97 | STATISTIC(NumPRELoadMoved2CEPred, |
98 | "Number of loads moved to predecessor of a critical edge in PRE" ); |
99 | |
100 | STATISTIC(IsValueFullyAvailableInBlockNumSpeculationsMax, |
101 | "Number of blocks speculated as available in " |
102 | "IsValueFullyAvailableInBlock(), max" ); |
103 | STATISTIC(MaxBBSpeculationCutoffReachedTimes, |
104 | "Number of times we we reached gvn-max-block-speculations cut-off " |
105 | "preventing further exploration" ); |
106 | |
107 | static cl::opt<bool> GVNEnablePRE("enable-pre" , cl::init(Val: true), cl::Hidden); |
108 | static cl::opt<bool> GVNEnableLoadPRE("enable-load-pre" , cl::init(Val: true)); |
109 | static cl::opt<bool> GVNEnableLoadInLoopPRE("enable-load-in-loop-pre" , |
110 | cl::init(Val: true)); |
111 | static cl::opt<bool> |
112 | GVNEnableSplitBackedgeInLoadPRE("enable-split-backedge-in-load-pre" , |
113 | cl::init(Val: false)); |
114 | static cl::opt<bool> GVNEnableMemDep("enable-gvn-memdep" , cl::init(Val: true)); |
115 | |
116 | static cl::opt<uint32_t> MaxNumDeps( |
117 | "gvn-max-num-deps" , cl::Hidden, cl::init(Val: 100), |
118 | cl::desc("Max number of dependences to attempt Load PRE (default = 100)" )); |
119 | |
120 | // This is based on IsValueFullyAvailableInBlockNumSpeculationsMax stat. |
121 | static cl::opt<uint32_t> MaxBBSpeculations( |
122 | "gvn-max-block-speculations" , cl::Hidden, cl::init(Val: 600), |
123 | cl::desc("Max number of blocks we're willing to speculate on (and recurse " |
124 | "into) when deducing if a value is fully available or not in GVN " |
125 | "(default = 600)" )); |
126 | |
127 | static cl::opt<uint32_t> MaxNumVisitedInsts( |
128 | "gvn-max-num-visited-insts" , cl::Hidden, cl::init(Val: 100), |
129 | cl::desc("Max number of visited instructions when trying to find " |
130 | "dominating value of select dependency (default = 100)" )); |
131 | |
132 | static cl::opt<uint32_t> MaxNumInsnsPerBlock( |
133 | "gvn-max-num-insns" , cl::Hidden, cl::init(Val: 100), |
134 | cl::desc("Max number of instructions to scan in each basic block in GVN " |
135 | "(default = 100)" )); |
136 | |
137 | struct llvm::GVNPass::Expression { |
138 | uint32_t opcode; |
139 | bool commutative = false; |
140 | // The type is not necessarily the result type of the expression, it may be |
141 | // any additional type needed to disambiguate the expression. |
142 | Type *type = nullptr; |
143 | SmallVector<uint32_t, 4> varargs; |
144 | |
145 | Expression(uint32_t o = ~2U) : opcode(o) {} |
146 | |
147 | bool operator==(const Expression &other) const { |
148 | if (opcode != other.opcode) |
149 | return false; |
150 | if (opcode == ~0U || opcode == ~1U) |
151 | return true; |
152 | if (type != other.type) |
153 | return false; |
154 | if (varargs != other.varargs) |
155 | return false; |
156 | return true; |
157 | } |
158 | |
159 | friend hash_code hash_value(const Expression &Value) { |
160 | return hash_combine( |
161 | args: Value.opcode, args: Value.type, |
162 | args: hash_combine_range(first: Value.varargs.begin(), last: Value.varargs.end())); |
163 | } |
164 | }; |
165 | |
166 | namespace llvm { |
167 | |
168 | template <> struct DenseMapInfo<GVNPass::Expression> { |
169 | static inline GVNPass::Expression getEmptyKey() { return ~0U; } |
170 | static inline GVNPass::Expression getTombstoneKey() { return ~1U; } |
171 | |
172 | static unsigned getHashValue(const GVNPass::Expression &e) { |
173 | using llvm::hash_value; |
174 | |
175 | return static_cast<unsigned>(hash_value(Value: e)); |
176 | } |
177 | |
178 | static bool isEqual(const GVNPass::Expression &LHS, |
179 | const GVNPass::Expression &RHS) { |
180 | return LHS == RHS; |
181 | } |
182 | }; |
183 | |
184 | } // end namespace llvm |
185 | |
186 | /// Represents a particular available value that we know how to materialize. |
187 | /// Materialization of an AvailableValue never fails. An AvailableValue is |
188 | /// implicitly associated with a rematerialization point which is the |
189 | /// location of the instruction from which it was formed. |
190 | struct llvm::gvn::AvailableValue { |
191 | enum class ValType { |
192 | SimpleVal, // A simple offsetted value that is accessed. |
193 | LoadVal, // A value produced by a load. |
194 | MemIntrin, // A memory intrinsic which is loaded from. |
195 | UndefVal, // A UndefValue representing a value from dead block (which |
196 | // is not yet physically removed from the CFG). |
197 | SelectVal, // A pointer select which is loaded from and for which the load |
198 | // can be replace by a value select. |
199 | }; |
200 | |
201 | /// Val - The value that is live out of the block. |
202 | Value *Val; |
203 | /// Kind of the live-out value. |
204 | ValType Kind; |
205 | |
206 | /// Offset - The byte offset in Val that is interesting for the load query. |
207 | unsigned Offset = 0; |
208 | /// V1, V2 - The dominating non-clobbered values of SelectVal. |
209 | Value *V1 = nullptr, *V2 = nullptr; |
210 | |
211 | static AvailableValue get(Value *V, unsigned Offset = 0) { |
212 | AvailableValue Res; |
213 | Res.Val = V; |
214 | Res.Kind = ValType::SimpleVal; |
215 | Res.Offset = Offset; |
216 | return Res; |
217 | } |
218 | |
219 | static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) { |
220 | AvailableValue Res; |
221 | Res.Val = MI; |
222 | Res.Kind = ValType::MemIntrin; |
223 | Res.Offset = Offset; |
224 | return Res; |
225 | } |
226 | |
227 | static AvailableValue getLoad(LoadInst *Load, unsigned Offset = 0) { |
228 | AvailableValue Res; |
229 | Res.Val = Load; |
230 | Res.Kind = ValType::LoadVal; |
231 | Res.Offset = Offset; |
232 | return Res; |
233 | } |
234 | |
235 | static AvailableValue getUndef() { |
236 | AvailableValue Res; |
237 | Res.Val = nullptr; |
238 | Res.Kind = ValType::UndefVal; |
239 | Res.Offset = 0; |
240 | return Res; |
241 | } |
242 | |
243 | static AvailableValue getSelect(SelectInst *Sel, Value *V1, Value *V2) { |
244 | AvailableValue Res; |
245 | Res.Val = Sel; |
246 | Res.Kind = ValType::SelectVal; |
247 | Res.Offset = 0; |
248 | Res.V1 = V1; |
249 | Res.V2 = V2; |
250 | return Res; |
251 | } |
252 | |
253 | bool isSimpleValue() const { return Kind == ValType::SimpleVal; } |
254 | bool isCoercedLoadValue() const { return Kind == ValType::LoadVal; } |
255 | bool isMemIntrinValue() const { return Kind == ValType::MemIntrin; } |
256 | bool isUndefValue() const { return Kind == ValType::UndefVal; } |
257 | bool isSelectValue() const { return Kind == ValType::SelectVal; } |
258 | |
259 | Value *getSimpleValue() const { |
260 | assert(isSimpleValue() && "Wrong accessor" ); |
261 | return Val; |
262 | } |
263 | |
264 | LoadInst *getCoercedLoadValue() const { |
265 | assert(isCoercedLoadValue() && "Wrong accessor" ); |
266 | return cast<LoadInst>(Val); |
267 | } |
268 | |
269 | MemIntrinsic *getMemIntrinValue() const { |
270 | assert(isMemIntrinValue() && "Wrong accessor" ); |
271 | return cast<MemIntrinsic>(Val); |
272 | } |
273 | |
274 | SelectInst *getSelectValue() const { |
275 | assert(isSelectValue() && "Wrong accessor" ); |
276 | return cast<SelectInst>(Val); |
277 | } |
278 | |
279 | /// Emit code at the specified insertion point to adjust the value defined |
280 | /// here to the specified type. This handles various coercion cases. |
281 | Value *MaterializeAdjustedValue(LoadInst *Load, Instruction *InsertPt, |
282 | GVNPass &gvn) const; |
283 | }; |
284 | |
285 | /// Represents an AvailableValue which can be rematerialized at the end of |
286 | /// the associated BasicBlock. |
287 | struct llvm::gvn::AvailableValueInBlock { |
288 | /// BB - The basic block in question. |
289 | BasicBlock *BB = nullptr; |
290 | |
291 | /// AV - The actual available value |
292 | AvailableValue AV; |
293 | |
294 | static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) { |
295 | AvailableValueInBlock Res; |
296 | Res.BB = BB; |
297 | Res.AV = std::move(AV); |
298 | return Res; |
299 | } |
300 | |
301 | static AvailableValueInBlock get(BasicBlock *BB, Value *V, |
302 | unsigned Offset = 0) { |
303 | return get(BB, AV: AvailableValue::get(V, Offset)); |
304 | } |
305 | |
306 | static AvailableValueInBlock getUndef(BasicBlock *BB) { |
307 | return get(BB, AV: AvailableValue::getUndef()); |
308 | } |
309 | |
310 | static AvailableValueInBlock getSelect(BasicBlock *BB, SelectInst *Sel, |
311 | Value *V1, Value *V2) { |
312 | return get(BB, AV: AvailableValue::getSelect(Sel, V1, V2)); |
313 | } |
314 | |
315 | /// Emit code at the end of this block to adjust the value defined here to |
316 | /// the specified type. This handles various coercion cases. |
317 | Value *MaterializeAdjustedValue(LoadInst *Load, GVNPass &gvn) const { |
318 | return AV.MaterializeAdjustedValue(Load, InsertPt: BB->getTerminator(), gvn); |
319 | } |
320 | }; |
321 | |
322 | //===----------------------------------------------------------------------===// |
323 | // ValueTable Internal Functions |
324 | //===----------------------------------------------------------------------===// |
325 | |
326 | GVNPass::Expression GVNPass::ValueTable::createExpr(Instruction *I) { |
327 | Expression e; |
328 | e.type = I->getType(); |
329 | e.opcode = I->getOpcode(); |
330 | if (const GCRelocateInst *GCR = dyn_cast<GCRelocateInst>(Val: I)) { |
331 | // gc.relocate is 'special' call: its second and third operands are |
332 | // not real values, but indices into statepoint's argument list. |
333 | // Use the refered to values for purposes of identity. |
334 | e.varargs.push_back(Elt: lookupOrAdd(V: GCR->getOperand(i_nocapture: 0))); |
335 | e.varargs.push_back(Elt: lookupOrAdd(V: GCR->getBasePtr())); |
336 | e.varargs.push_back(Elt: lookupOrAdd(V: GCR->getDerivedPtr())); |
337 | } else { |
338 | for (Use &Op : I->operands()) |
339 | e.varargs.push_back(Elt: lookupOrAdd(V: Op)); |
340 | } |
341 | if (I->isCommutative()) { |
342 | // Ensure that commutative instructions that only differ by a permutation |
343 | // of their operands get the same value number by sorting the operand value |
344 | // numbers. Since commutative operands are the 1st two operands it is more |
345 | // efficient to sort by hand rather than using, say, std::sort. |
346 | assert(I->getNumOperands() >= 2 && "Unsupported commutative instruction!" ); |
347 | if (e.varargs[0] > e.varargs[1]) |
348 | std::swap(a&: e.varargs[0], b&: e.varargs[1]); |
349 | e.commutative = true; |
350 | } |
351 | |
352 | if (auto *C = dyn_cast<CmpInst>(Val: I)) { |
353 | // Sort the operand value numbers so x<y and y>x get the same value number. |
354 | CmpInst::Predicate Predicate = C->getPredicate(); |
355 | if (e.varargs[0] > e.varargs[1]) { |
356 | std::swap(a&: e.varargs[0], b&: e.varargs[1]); |
357 | Predicate = CmpInst::getSwappedPredicate(pred: Predicate); |
358 | } |
359 | e.opcode = (C->getOpcode() << 8) | Predicate; |
360 | e.commutative = true; |
361 | } else if (auto *E = dyn_cast<InsertValueInst>(Val: I)) { |
362 | e.varargs.append(in_start: E->idx_begin(), in_end: E->idx_end()); |
363 | } else if (auto *SVI = dyn_cast<ShuffleVectorInst>(Val: I)) { |
364 | ArrayRef<int> ShuffleMask = SVI->getShuffleMask(); |
365 | e.varargs.append(in_start: ShuffleMask.begin(), in_end: ShuffleMask.end()); |
366 | } |
367 | |
368 | return e; |
369 | } |
370 | |
371 | GVNPass::Expression GVNPass::ValueTable::createCmpExpr( |
372 | unsigned Opcode, CmpInst::Predicate Predicate, Value *LHS, Value *RHS) { |
373 | assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && |
374 | "Not a comparison!" ); |
375 | Expression e; |
376 | e.type = CmpInst::makeCmpResultType(opnd_type: LHS->getType()); |
377 | e.varargs.push_back(Elt: lookupOrAdd(V: LHS)); |
378 | e.varargs.push_back(Elt: lookupOrAdd(V: RHS)); |
379 | |
380 | // Sort the operand value numbers so x<y and y>x get the same value number. |
381 | if (e.varargs[0] > e.varargs[1]) { |
382 | std::swap(a&: e.varargs[0], b&: e.varargs[1]); |
383 | Predicate = CmpInst::getSwappedPredicate(pred: Predicate); |
384 | } |
385 | e.opcode = (Opcode << 8) | Predicate; |
386 | e.commutative = true; |
387 | return e; |
388 | } |
389 | |
390 | GVNPass::Expression |
391 | GVNPass::ValueTable::(ExtractValueInst *EI) { |
392 | assert(EI && "Not an ExtractValueInst?" ); |
393 | Expression e; |
394 | e.type = EI->getType(); |
395 | e.opcode = 0; |
396 | |
397 | WithOverflowInst *WO = dyn_cast<WithOverflowInst>(Val: EI->getAggregateOperand()); |
398 | if (WO != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) { |
399 | // EI is an extract from one of our with.overflow intrinsics. Synthesize |
400 | // a semantically equivalent expression instead of an extract value |
401 | // expression. |
402 | e.opcode = WO->getBinaryOp(); |
403 | e.varargs.push_back(Elt: lookupOrAdd(V: WO->getLHS())); |
404 | e.varargs.push_back(Elt: lookupOrAdd(V: WO->getRHS())); |
405 | return e; |
406 | } |
407 | |
408 | // Not a recognised intrinsic. Fall back to producing an extract value |
409 | // expression. |
410 | e.opcode = EI->getOpcode(); |
411 | for (Use &Op : EI->operands()) |
412 | e.varargs.push_back(Elt: lookupOrAdd(V: Op)); |
413 | |
414 | append_range(C&: e.varargs, R: EI->indices()); |
415 | |
416 | return e; |
417 | } |
418 | |
419 | GVNPass::Expression GVNPass::ValueTable::createGEPExpr(GetElementPtrInst *GEP) { |
420 | Expression E; |
421 | Type *PtrTy = GEP->getType()->getScalarType(); |
422 | const DataLayout &DL = GEP->getModule()->getDataLayout(); |
423 | unsigned BitWidth = DL.getIndexTypeSizeInBits(Ty: PtrTy); |
424 | MapVector<Value *, APInt> VariableOffsets; |
425 | APInt ConstantOffset(BitWidth, 0); |
426 | if (GEP->collectOffset(DL, BitWidth, VariableOffsets, ConstantOffset)) { |
427 | // Convert into offset representation, to recognize equivalent address |
428 | // calculations that use different type encoding. |
429 | LLVMContext &Context = GEP->getContext(); |
430 | E.opcode = GEP->getOpcode(); |
431 | E.type = nullptr; |
432 | E.varargs.push_back(Elt: lookupOrAdd(V: GEP->getPointerOperand())); |
433 | for (const auto &Pair : VariableOffsets) { |
434 | E.varargs.push_back(Elt: lookupOrAdd(V: Pair.first)); |
435 | E.varargs.push_back(Elt: lookupOrAdd(V: ConstantInt::get(Context, V: Pair.second))); |
436 | } |
437 | if (!ConstantOffset.isZero()) |
438 | E.varargs.push_back( |
439 | Elt: lookupOrAdd(V: ConstantInt::get(Context, V: ConstantOffset))); |
440 | } else { |
441 | // If converting to offset representation fails (for scalable vectors), |
442 | // fall back to type-based implementation: |
443 | E.opcode = GEP->getOpcode(); |
444 | E.type = GEP->getSourceElementType(); |
445 | for (Use &Op : GEP->operands()) |
446 | E.varargs.push_back(Elt: lookupOrAdd(V: Op)); |
447 | } |
448 | return E; |
449 | } |
450 | |
451 | //===----------------------------------------------------------------------===// |
452 | // ValueTable External Functions |
453 | //===----------------------------------------------------------------------===// |
454 | |
455 | GVNPass::ValueTable::ValueTable() = default; |
456 | GVNPass::ValueTable::ValueTable(const ValueTable &) = default; |
457 | GVNPass::ValueTable::ValueTable(ValueTable &&) = default; |
458 | GVNPass::ValueTable::~ValueTable() = default; |
459 | GVNPass::ValueTable & |
460 | GVNPass::ValueTable::operator=(const GVNPass::ValueTable &Arg) = default; |
461 | |
462 | /// add - Insert a value into the table with a specified value number. |
463 | void GVNPass::ValueTable::add(Value *V, uint32_t num) { |
464 | valueNumbering.insert(KV: std::make_pair(x&: V, y&: num)); |
465 | if (PHINode *PN = dyn_cast<PHINode>(Val: V)) |
466 | NumberingPhi[num] = PN; |
467 | } |
468 | |
469 | uint32_t GVNPass::ValueTable::lookupOrAddCall(CallInst *C) { |
470 | // FIXME: Currently the calls which may access the thread id may |
471 | // be considered as not accessing the memory. But this is |
472 | // problematic for coroutines, since coroutines may resume in a |
473 | // different thread. So we disable the optimization here for the |
474 | // correctness. However, it may block many other correct |
475 | // optimizations. Revert this one when we detect the memory |
476 | // accessing kind more precisely. |
477 | if (C->getFunction()->isPresplitCoroutine()) { |
478 | valueNumbering[C] = nextValueNumber; |
479 | return nextValueNumber++; |
480 | } |
481 | |
482 | // Do not combine convergent calls since they implicitly depend on the set of |
483 | // threads that is currently executing, and they might be in different basic |
484 | // blocks. |
485 | if (C->isConvergent()) { |
486 | valueNumbering[C] = nextValueNumber; |
487 | return nextValueNumber++; |
488 | } |
489 | |
490 | if (AA->doesNotAccessMemory(Call: C)) { |
491 | Expression exp = createExpr(I: C); |
492 | uint32_t e = assignExpNewValueNum(exp).first; |
493 | valueNumbering[C] = e; |
494 | return e; |
495 | } |
496 | |
497 | if (MD && AA->onlyReadsMemory(Call: C)) { |
498 | Expression exp = createExpr(I: C); |
499 | auto ValNum = assignExpNewValueNum(exp); |
500 | if (ValNum.second) { |
501 | valueNumbering[C] = ValNum.first; |
502 | return ValNum.first; |
503 | } |
504 | |
505 | MemDepResult local_dep = MD->getDependency(QueryInst: C); |
506 | |
507 | if (!local_dep.isDef() && !local_dep.isNonLocal()) { |
508 | valueNumbering[C] = nextValueNumber; |
509 | return nextValueNumber++; |
510 | } |
511 | |
512 | if (local_dep.isDef()) { |
513 | // For masked load/store intrinsics, the local_dep may actually be |
514 | // a normal load or store instruction. |
515 | CallInst *local_cdep = dyn_cast<CallInst>(Val: local_dep.getInst()); |
516 | |
517 | if (!local_cdep || local_cdep->arg_size() != C->arg_size()) { |
518 | valueNumbering[C] = nextValueNumber; |
519 | return nextValueNumber++; |
520 | } |
521 | |
522 | for (unsigned i = 0, e = C->arg_size(); i < e; ++i) { |
523 | uint32_t c_vn = lookupOrAdd(V: C->getArgOperand(i)); |
524 | uint32_t cd_vn = lookupOrAdd(V: local_cdep->getArgOperand(i)); |
525 | if (c_vn != cd_vn) { |
526 | valueNumbering[C] = nextValueNumber; |
527 | return nextValueNumber++; |
528 | } |
529 | } |
530 | |
531 | uint32_t v = lookupOrAdd(V: local_cdep); |
532 | valueNumbering[C] = v; |
533 | return v; |
534 | } |
535 | |
536 | // Non-local case. |
537 | const MemoryDependenceResults::NonLocalDepInfo &deps = |
538 | MD->getNonLocalCallDependency(QueryCall: C); |
539 | // FIXME: Move the checking logic to MemDep! |
540 | CallInst* cdep = nullptr; |
541 | |
542 | // Check to see if we have a single dominating call instruction that is |
543 | // identical to C. |
544 | for (const NonLocalDepEntry &I : deps) { |
545 | if (I.getResult().isNonLocal()) |
546 | continue; |
547 | |
548 | // We don't handle non-definitions. If we already have a call, reject |
549 | // instruction dependencies. |
550 | if (!I.getResult().isDef() || cdep != nullptr) { |
551 | cdep = nullptr; |
552 | break; |
553 | } |
554 | |
555 | CallInst *NonLocalDepCall = dyn_cast<CallInst>(Val: I.getResult().getInst()); |
556 | // FIXME: All duplicated with non-local case. |
557 | if (NonLocalDepCall && DT->properlyDominates(A: I.getBB(), B: C->getParent())) { |
558 | cdep = NonLocalDepCall; |
559 | continue; |
560 | } |
561 | |
562 | cdep = nullptr; |
563 | break; |
564 | } |
565 | |
566 | if (!cdep) { |
567 | valueNumbering[C] = nextValueNumber; |
568 | return nextValueNumber++; |
569 | } |
570 | |
571 | if (cdep->arg_size() != C->arg_size()) { |
572 | valueNumbering[C] = nextValueNumber; |
573 | return nextValueNumber++; |
574 | } |
575 | for (unsigned i = 0, e = C->arg_size(); i < e; ++i) { |
576 | uint32_t c_vn = lookupOrAdd(V: C->getArgOperand(i)); |
577 | uint32_t cd_vn = lookupOrAdd(V: cdep->getArgOperand(i)); |
578 | if (c_vn != cd_vn) { |
579 | valueNumbering[C] = nextValueNumber; |
580 | return nextValueNumber++; |
581 | } |
582 | } |
583 | |
584 | uint32_t v = lookupOrAdd(V: cdep); |
585 | valueNumbering[C] = v; |
586 | return v; |
587 | } |
588 | |
589 | valueNumbering[C] = nextValueNumber; |
590 | return nextValueNumber++; |
591 | } |
592 | |
593 | /// Returns true if a value number exists for the specified value. |
594 | bool GVNPass::ValueTable::exists(Value *V) const { |
595 | return valueNumbering.contains(Val: V); |
596 | } |
597 | |
598 | /// lookup_or_add - Returns the value number for the specified value, assigning |
599 | /// it a new number if it did not have one before. |
600 | uint32_t GVNPass::ValueTable::lookupOrAdd(Value *V) { |
601 | DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(Val: V); |
602 | if (VI != valueNumbering.end()) |
603 | return VI->second; |
604 | |
605 | auto *I = dyn_cast<Instruction>(Val: V); |
606 | if (!I) { |
607 | valueNumbering[V] = nextValueNumber; |
608 | return nextValueNumber++; |
609 | } |
610 | |
611 | Expression exp; |
612 | switch (I->getOpcode()) { |
613 | case Instruction::Call: |
614 | return lookupOrAddCall(C: cast<CallInst>(Val: I)); |
615 | case Instruction::FNeg: |
616 | case Instruction::Add: |
617 | case Instruction::FAdd: |
618 | case Instruction::Sub: |
619 | case Instruction::FSub: |
620 | case Instruction::Mul: |
621 | case Instruction::FMul: |
622 | case Instruction::UDiv: |
623 | case Instruction::SDiv: |
624 | case Instruction::FDiv: |
625 | case Instruction::URem: |
626 | case Instruction::SRem: |
627 | case Instruction::FRem: |
628 | case Instruction::Shl: |
629 | case Instruction::LShr: |
630 | case Instruction::AShr: |
631 | case Instruction::And: |
632 | case Instruction::Or: |
633 | case Instruction::Xor: |
634 | case Instruction::ICmp: |
635 | case Instruction::FCmp: |
636 | case Instruction::Trunc: |
637 | case Instruction::ZExt: |
638 | case Instruction::SExt: |
639 | case Instruction::FPToUI: |
640 | case Instruction::FPToSI: |
641 | case Instruction::UIToFP: |
642 | case Instruction::SIToFP: |
643 | case Instruction::FPTrunc: |
644 | case Instruction::FPExt: |
645 | case Instruction::PtrToInt: |
646 | case Instruction::IntToPtr: |
647 | case Instruction::AddrSpaceCast: |
648 | case Instruction::BitCast: |
649 | case Instruction::Select: |
650 | case Instruction::Freeze: |
651 | case Instruction::ExtractElement: |
652 | case Instruction::InsertElement: |
653 | case Instruction::ShuffleVector: |
654 | case Instruction::InsertValue: |
655 | exp = createExpr(I); |
656 | break; |
657 | case Instruction::GetElementPtr: |
658 | exp = createGEPExpr(GEP: cast<GetElementPtrInst>(Val: I)); |
659 | break; |
660 | case Instruction::ExtractValue: |
661 | exp = createExtractvalueExpr(EI: cast<ExtractValueInst>(Val: I)); |
662 | break; |
663 | case Instruction::PHI: |
664 | valueNumbering[V] = nextValueNumber; |
665 | NumberingPhi[nextValueNumber] = cast<PHINode>(Val: V); |
666 | return nextValueNumber++; |
667 | default: |
668 | valueNumbering[V] = nextValueNumber; |
669 | return nextValueNumber++; |
670 | } |
671 | |
672 | uint32_t e = assignExpNewValueNum(exp).first; |
673 | valueNumbering[V] = e; |
674 | return e; |
675 | } |
676 | |
677 | /// Returns the value number of the specified value. Fails if |
678 | /// the value has not yet been numbered. |
679 | uint32_t GVNPass::ValueTable::lookup(Value *V, bool Verify) const { |
680 | DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(Val: V); |
681 | if (Verify) { |
682 | assert(VI != valueNumbering.end() && "Value not numbered?" ); |
683 | return VI->second; |
684 | } |
685 | return (VI != valueNumbering.end()) ? VI->second : 0; |
686 | } |
687 | |
688 | /// Returns the value number of the given comparison, |
689 | /// assigning it a new number if it did not have one before. Useful when |
690 | /// we deduced the result of a comparison, but don't immediately have an |
691 | /// instruction realizing that comparison to hand. |
692 | uint32_t GVNPass::ValueTable::lookupOrAddCmp(unsigned Opcode, |
693 | CmpInst::Predicate Predicate, |
694 | Value *LHS, Value *RHS) { |
695 | Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS); |
696 | return assignExpNewValueNum(exp).first; |
697 | } |
698 | |
699 | /// Remove all entries from the ValueTable. |
700 | void GVNPass::ValueTable::clear() { |
701 | valueNumbering.clear(); |
702 | expressionNumbering.clear(); |
703 | NumberingPhi.clear(); |
704 | PhiTranslateTable.clear(); |
705 | nextValueNumber = 1; |
706 | Expressions.clear(); |
707 | ExprIdx.clear(); |
708 | nextExprNumber = 0; |
709 | } |
710 | |
711 | /// Remove a value from the value numbering. |
712 | void GVNPass::ValueTable::erase(Value *V) { |
713 | uint32_t Num = valueNumbering.lookup(Val: V); |
714 | valueNumbering.erase(Val: V); |
715 | // If V is PHINode, V <--> value number is an one-to-one mapping. |
716 | if (isa<PHINode>(Val: V)) |
717 | NumberingPhi.erase(Val: Num); |
718 | } |
719 | |
720 | /// verifyRemoved - Verify that the value is removed from all internal data |
721 | /// structures. |
722 | void GVNPass::ValueTable::verifyRemoved(const Value *V) const { |
723 | assert(!valueNumbering.contains(V) && |
724 | "Inst still occurs in value numbering map!" ); |
725 | } |
726 | |
727 | //===----------------------------------------------------------------------===// |
728 | // GVN Pass |
729 | //===----------------------------------------------------------------------===// |
730 | |
731 | bool GVNPass::isPREEnabled() const { |
732 | return Options.AllowPRE.value_or(u&: GVNEnablePRE); |
733 | } |
734 | |
735 | bool GVNPass::isLoadPREEnabled() const { |
736 | return Options.AllowLoadPRE.value_or(u&: GVNEnableLoadPRE); |
737 | } |
738 | |
739 | bool GVNPass::isLoadInLoopPREEnabled() const { |
740 | return Options.AllowLoadInLoopPRE.value_or(u&: GVNEnableLoadInLoopPRE); |
741 | } |
742 | |
743 | bool GVNPass::isLoadPRESplitBackedgeEnabled() const { |
744 | return Options.AllowLoadPRESplitBackedge.value_or( |
745 | u&: GVNEnableSplitBackedgeInLoadPRE); |
746 | } |
747 | |
748 | bool GVNPass::isMemDepEnabled() const { |
749 | return Options.AllowMemDep.value_or(u&: GVNEnableMemDep); |
750 | } |
751 | |
752 | PreservedAnalyses GVNPass::run(Function &F, FunctionAnalysisManager &AM) { |
753 | // FIXME: The order of evaluation of these 'getResult' calls is very |
754 | // significant! Re-ordering these variables will cause GVN when run alone to |
755 | // be less effective! We should fix memdep and basic-aa to not exhibit this |
756 | // behavior, but until then don't change the order here. |
757 | auto &AC = AM.getResult<AssumptionAnalysis>(IR&: F); |
758 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
759 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F); |
760 | auto &AA = AM.getResult<AAManager>(IR&: F); |
761 | auto *MemDep = |
762 | isMemDepEnabled() ? &AM.getResult<MemoryDependenceAnalysis>(IR&: F) : nullptr; |
763 | auto &LI = AM.getResult<LoopAnalysis>(IR&: F); |
764 | auto *MSSA = AM.getCachedResult<MemorySSAAnalysis>(IR&: F); |
765 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F); |
766 | bool Changed = runImpl(F, RunAC&: AC, RunDT&: DT, RunTLI: TLI, RunAA&: AA, RunMD: MemDep, LI, ORE: &ORE, |
767 | MSSA: MSSA ? &MSSA->getMSSA() : nullptr); |
768 | if (!Changed) |
769 | return PreservedAnalyses::all(); |
770 | PreservedAnalyses PA; |
771 | PA.preserve<DominatorTreeAnalysis>(); |
772 | PA.preserve<TargetLibraryAnalysis>(); |
773 | if (MSSA) |
774 | PA.preserve<MemorySSAAnalysis>(); |
775 | PA.preserve<LoopAnalysis>(); |
776 | return PA; |
777 | } |
778 | |
779 | void GVNPass::printPipeline( |
780 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
781 | static_cast<PassInfoMixin<GVNPass> *>(this)->printPipeline( |
782 | OS, MapClassName2PassName); |
783 | |
784 | OS << '<'; |
785 | if (Options.AllowPRE != std::nullopt) |
786 | OS << (*Options.AllowPRE ? "" : "no-" ) << "pre;" ; |
787 | if (Options.AllowLoadPRE != std::nullopt) |
788 | OS << (*Options.AllowLoadPRE ? "" : "no-" ) << "load-pre;" ; |
789 | if (Options.AllowLoadPRESplitBackedge != std::nullopt) |
790 | OS << (*Options.AllowLoadPRESplitBackedge ? "" : "no-" ) |
791 | << "split-backedge-load-pre;" ; |
792 | if (Options.AllowMemDep != std::nullopt) |
793 | OS << (*Options.AllowMemDep ? "" : "no-" ) << "memdep" ; |
794 | OS << '>'; |
795 | } |
796 | |
797 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
798 | LLVM_DUMP_METHOD void GVNPass::dump(DenseMap<uint32_t, Value *> &d) const { |
799 | errs() << "{\n" ; |
800 | for (auto &I : d) { |
801 | errs() << I.first << "\n" ; |
802 | I.second->dump(); |
803 | } |
804 | errs() << "}\n" ; |
805 | } |
806 | #endif |
807 | |
808 | enum class AvailabilityState : char { |
809 | /// We know the block *is not* fully available. This is a fixpoint. |
810 | Unavailable = 0, |
811 | /// We know the block *is* fully available. This is a fixpoint. |
812 | Available = 1, |
813 | /// We do not know whether the block is fully available or not, |
814 | /// but we are currently speculating that it will be. |
815 | /// If it would have turned out that the block was, in fact, not fully |
816 | /// available, this would have been cleaned up into an Unavailable. |
817 | SpeculativelyAvailable = 2, |
818 | }; |
819 | |
820 | /// Return true if we can prove that the value |
821 | /// we're analyzing is fully available in the specified block. As we go, keep |
822 | /// track of which blocks we know are fully alive in FullyAvailableBlocks. This |
823 | /// map is actually a tri-state map with the following values: |
824 | /// 0) we know the block *is not* fully available. |
825 | /// 1) we know the block *is* fully available. |
826 | /// 2) we do not know whether the block is fully available or not, but we are |
827 | /// currently speculating that it will be. |
828 | static bool IsValueFullyAvailableInBlock( |
829 | BasicBlock *BB, |
830 | DenseMap<BasicBlock *, AvailabilityState> &FullyAvailableBlocks) { |
831 | SmallVector<BasicBlock *, 32> Worklist; |
832 | std::optional<BasicBlock *> UnavailableBB; |
833 | |
834 | // The number of times we didn't find an entry for a block in a map and |
835 | // optimistically inserted an entry marking block as speculatively available. |
836 | unsigned NumNewNewSpeculativelyAvailableBBs = 0; |
837 | |
838 | #ifndef NDEBUG |
839 | SmallSet<BasicBlock *, 32> NewSpeculativelyAvailableBBs; |
840 | SmallVector<BasicBlock *, 32> AvailableBBs; |
841 | #endif |
842 | |
843 | Worklist.emplace_back(Args&: BB); |
844 | while (!Worklist.empty()) { |
845 | BasicBlock *CurrBB = Worklist.pop_back_val(); // LoadFO - depth-first! |
846 | // Optimistically assume that the block is Speculatively Available and check |
847 | // to see if we already know about this block in one lookup. |
848 | std::pair<DenseMap<BasicBlock *, AvailabilityState>::iterator, bool> IV = |
849 | FullyAvailableBlocks.try_emplace( |
850 | Key: CurrBB, Args: AvailabilityState::SpeculativelyAvailable); |
851 | AvailabilityState &State = IV.first->second; |
852 | |
853 | // Did the entry already exist for this block? |
854 | if (!IV.second) { |
855 | if (State == AvailabilityState::Unavailable) { |
856 | UnavailableBB = CurrBB; |
857 | break; // Backpropagate unavailability info. |
858 | } |
859 | |
860 | #ifndef NDEBUG |
861 | AvailableBBs.emplace_back(Args&: CurrBB); |
862 | #endif |
863 | continue; // Don't recurse further, but continue processing worklist. |
864 | } |
865 | |
866 | // No entry found for block. |
867 | ++NumNewNewSpeculativelyAvailableBBs; |
868 | bool OutOfBudget = NumNewNewSpeculativelyAvailableBBs > MaxBBSpeculations; |
869 | |
870 | // If we have exhausted our budget, mark this block as unavailable. |
871 | // Also, if this block has no predecessors, the value isn't live-in here. |
872 | if (OutOfBudget || pred_empty(BB: CurrBB)) { |
873 | MaxBBSpeculationCutoffReachedTimes += (int)OutOfBudget; |
874 | State = AvailabilityState::Unavailable; |
875 | UnavailableBB = CurrBB; |
876 | break; // Backpropagate unavailability info. |
877 | } |
878 | |
879 | // Tentatively consider this block as speculatively available. |
880 | #ifndef NDEBUG |
881 | NewSpeculativelyAvailableBBs.insert(Ptr: CurrBB); |
882 | #endif |
883 | // And further recurse into block's predecessors, in depth-first order! |
884 | Worklist.append(in_start: pred_begin(BB: CurrBB), in_end: pred_end(BB: CurrBB)); |
885 | } |
886 | |
887 | #if LLVM_ENABLE_STATS |
888 | IsValueFullyAvailableInBlockNumSpeculationsMax.updateMax( |
889 | V: NumNewNewSpeculativelyAvailableBBs); |
890 | #endif |
891 | |
892 | // If the block isn't marked as fixpoint yet |
893 | // (the Unavailable and Available states are fixpoints) |
894 | auto MarkAsFixpointAndEnqueueSuccessors = |
895 | [&](BasicBlock *BB, AvailabilityState FixpointState) { |
896 | auto It = FullyAvailableBlocks.find(Val: BB); |
897 | if (It == FullyAvailableBlocks.end()) |
898 | return; // Never queried this block, leave as-is. |
899 | switch (AvailabilityState &State = It->second) { |
900 | case AvailabilityState::Unavailable: |
901 | case AvailabilityState::Available: |
902 | return; // Don't backpropagate further, continue processing worklist. |
903 | case AvailabilityState::SpeculativelyAvailable: // Fix it! |
904 | State = FixpointState; |
905 | #ifndef NDEBUG |
906 | assert(NewSpeculativelyAvailableBBs.erase(BB) && |
907 | "Found a speculatively available successor leftover?" ); |
908 | #endif |
909 | // Queue successors for further processing. |
910 | Worklist.append(in_start: succ_begin(BB), in_end: succ_end(BB)); |
911 | return; |
912 | } |
913 | }; |
914 | |
915 | if (UnavailableBB) { |
916 | // Okay, we have encountered an unavailable block. |
917 | // Mark speculatively available blocks reachable from UnavailableBB as |
918 | // unavailable as well. Paths are terminated when they reach blocks not in |
919 | // FullyAvailableBlocks or they are not marked as speculatively available. |
920 | Worklist.clear(); |
921 | Worklist.append(in_start: succ_begin(BB: *UnavailableBB), in_end: succ_end(BB: *UnavailableBB)); |
922 | while (!Worklist.empty()) |
923 | MarkAsFixpointAndEnqueueSuccessors(Worklist.pop_back_val(), |
924 | AvailabilityState::Unavailable); |
925 | } |
926 | |
927 | #ifndef NDEBUG |
928 | Worklist.clear(); |
929 | for (BasicBlock *AvailableBB : AvailableBBs) |
930 | Worklist.append(in_start: succ_begin(BB: AvailableBB), in_end: succ_end(BB: AvailableBB)); |
931 | while (!Worklist.empty()) |
932 | MarkAsFixpointAndEnqueueSuccessors(Worklist.pop_back_val(), |
933 | AvailabilityState::Available); |
934 | |
935 | assert(NewSpeculativelyAvailableBBs.empty() && |
936 | "Must have fixed all the new speculatively available blocks." ); |
937 | #endif |
938 | |
939 | return !UnavailableBB; |
940 | } |
941 | |
942 | /// If the specified OldValue exists in ValuesPerBlock, replace its value with |
943 | /// NewValue. |
944 | static void replaceValuesPerBlockEntry( |
945 | SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, Value *OldValue, |
946 | Value *NewValue) { |
947 | for (AvailableValueInBlock &V : ValuesPerBlock) { |
948 | if (V.AV.Val == OldValue) |
949 | V.AV.Val = NewValue; |
950 | if (V.AV.isSelectValue()) { |
951 | if (V.AV.V1 == OldValue) |
952 | V.AV.V1 = NewValue; |
953 | if (V.AV.V2 == OldValue) |
954 | V.AV.V2 = NewValue; |
955 | } |
956 | } |
957 | } |
958 | |
959 | /// Given a set of loads specified by ValuesPerBlock, |
960 | /// construct SSA form, allowing us to eliminate Load. This returns the value |
961 | /// that should be used at Load's definition site. |
962 | static Value * |
963 | ConstructSSAForLoadSet(LoadInst *Load, |
964 | SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, |
965 | GVNPass &gvn) { |
966 | // Check for the fully redundant, dominating load case. In this case, we can |
967 | // just use the dominating value directly. |
968 | if (ValuesPerBlock.size() == 1 && |
969 | gvn.getDominatorTree().properlyDominates(A: ValuesPerBlock[0].BB, |
970 | B: Load->getParent())) { |
971 | assert(!ValuesPerBlock[0].AV.isUndefValue() && |
972 | "Dead BB dominate this block" ); |
973 | return ValuesPerBlock[0].MaterializeAdjustedValue(Load, gvn); |
974 | } |
975 | |
976 | // Otherwise, we have to construct SSA form. |
977 | SmallVector<PHINode*, 8> NewPHIs; |
978 | SSAUpdater SSAUpdate(&NewPHIs); |
979 | SSAUpdate.Initialize(Ty: Load->getType(), Name: Load->getName()); |
980 | |
981 | for (const AvailableValueInBlock &AV : ValuesPerBlock) { |
982 | BasicBlock *BB = AV.BB; |
983 | |
984 | if (AV.AV.isUndefValue()) |
985 | continue; |
986 | |
987 | if (SSAUpdate.HasValueForBlock(BB)) |
988 | continue; |
989 | |
990 | // If the value is the load that we will be eliminating, and the block it's |
991 | // available in is the block that the load is in, then don't add it as |
992 | // SSAUpdater will resolve the value to the relevant phi which may let it |
993 | // avoid phi construction entirely if there's actually only one value. |
994 | if (BB == Load->getParent() && |
995 | ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == Load) || |
996 | (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == Load))) |
997 | continue; |
998 | |
999 | SSAUpdate.AddAvailableValue(BB, V: AV.MaterializeAdjustedValue(Load, gvn)); |
1000 | } |
1001 | |
1002 | // Perform PHI construction. |
1003 | return SSAUpdate.GetValueInMiddleOfBlock(BB: Load->getParent()); |
1004 | } |
1005 | |
1006 | Value *AvailableValue::MaterializeAdjustedValue(LoadInst *Load, |
1007 | Instruction *InsertPt, |
1008 | GVNPass &gvn) const { |
1009 | Value *Res; |
1010 | Type *LoadTy = Load->getType(); |
1011 | const DataLayout &DL = Load->getModule()->getDataLayout(); |
1012 | if (isSimpleValue()) { |
1013 | Res = getSimpleValue(); |
1014 | if (Res->getType() != LoadTy) { |
1015 | Res = getValueForLoad(SrcVal: Res, Offset, LoadTy, InsertPt, DL); |
1016 | |
1017 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset |
1018 | << " " << *getSimpleValue() << '\n' |
1019 | << *Res << '\n' |
1020 | << "\n\n\n" ); |
1021 | } |
1022 | } else if (isCoercedLoadValue()) { |
1023 | LoadInst *CoercedLoad = getCoercedLoadValue(); |
1024 | if (CoercedLoad->getType() == LoadTy && Offset == 0) { |
1025 | Res = CoercedLoad; |
1026 | combineMetadataForCSE(K: CoercedLoad, J: Load, DoesKMove: false); |
1027 | } else { |
1028 | Res = getValueForLoad(SrcVal: CoercedLoad, Offset, LoadTy, InsertPt, DL); |
1029 | // We are adding a new user for this load, for which the original |
1030 | // metadata may not hold. Additionally, the new load may have a different |
1031 | // size and type, so their metadata cannot be combined in any |
1032 | // straightforward way. |
1033 | // Drop all metadata that is not known to cause immediate UB on violation, |
1034 | // unless the load has !noundef, in which case all metadata violations |
1035 | // will be promoted to UB. |
1036 | // TODO: We can combine noalias/alias.scope metadata here, because it is |
1037 | // independent of the load type. |
1038 | if (!CoercedLoad->hasMetadata(KindID: LLVMContext::MD_noundef)) |
1039 | CoercedLoad->dropUnknownNonDebugMetadata( |
1040 | KnownIDs: {LLVMContext::MD_dereferenceable, |
1041 | LLVMContext::MD_dereferenceable_or_null, |
1042 | LLVMContext::MD_invariant_load, LLVMContext::MD_invariant_group}); |
1043 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset |
1044 | << " " << *getCoercedLoadValue() << '\n' |
1045 | << *Res << '\n' |
1046 | << "\n\n\n" ); |
1047 | } |
1048 | } else if (isMemIntrinValue()) { |
1049 | Res = getMemInstValueForLoad(SrcInst: getMemIntrinValue(), Offset, LoadTy, |
1050 | InsertPt, DL); |
1051 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset |
1052 | << " " << *getMemIntrinValue() << '\n' |
1053 | << *Res << '\n' |
1054 | << "\n\n\n" ); |
1055 | } else if (isSelectValue()) { |
1056 | // Introduce a new value select for a load from an eligible pointer select. |
1057 | SelectInst *Sel = getSelectValue(); |
1058 | assert(V1 && V2 && "both value operands of the select must be present" ); |
1059 | Res = SelectInst::Create(C: Sel->getCondition(), S1: V1, S2: V2, NameStr: "" , InsertBefore: Sel); |
1060 | } else { |
1061 | llvm_unreachable("Should not materialize value from dead block" ); |
1062 | } |
1063 | assert(Res && "failed to materialize?" ); |
1064 | return Res; |
1065 | } |
1066 | |
1067 | static bool isLifetimeStart(const Instruction *Inst) { |
1068 | if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Val: Inst)) |
1069 | return II->getIntrinsicID() == Intrinsic::lifetime_start; |
1070 | return false; |
1071 | } |
1072 | |
1073 | /// Assuming To can be reached from both From and Between, does Between lie on |
1074 | /// every path from From to To? |
1075 | static bool liesBetween(const Instruction *From, Instruction *Between, |
1076 | const Instruction *To, DominatorTree *DT) { |
1077 | if (From->getParent() == Between->getParent()) |
1078 | return DT->dominates(Def: From, User: Between); |
1079 | SmallSet<BasicBlock *, 1> Exclusion; |
1080 | Exclusion.insert(Ptr: Between->getParent()); |
1081 | return !isPotentiallyReachable(From, To, ExclusionSet: &Exclusion, DT); |
1082 | } |
1083 | |
1084 | /// Try to locate the three instruction involved in a missed |
1085 | /// load-elimination case that is due to an intervening store. |
1086 | static void (LoadInst *Load, MemDepResult DepInfo, |
1087 | DominatorTree *DT, |
1088 | OptimizationRemarkEmitter *ORE) { |
1089 | using namespace ore; |
1090 | |
1091 | Instruction *OtherAccess = nullptr; |
1092 | |
1093 | OptimizationRemarkMissed R(DEBUG_TYPE, "LoadClobbered" , Load); |
1094 | R << "load of type " << NV("Type" , Load->getType()) << " not eliminated" |
1095 | << setExtraArgs(); |
1096 | |
1097 | for (auto *U : Load->getPointerOperand()->users()) { |
1098 | if (U != Load && (isa<LoadInst>(Val: U) || isa<StoreInst>(Val: U))) { |
1099 | auto *I = cast<Instruction>(Val: U); |
1100 | if (I->getFunction() == Load->getFunction() && DT->dominates(Def: I, User: Load)) { |
1101 | // Use the most immediately dominating value |
1102 | if (OtherAccess) { |
1103 | if (DT->dominates(Def: OtherAccess, User: I)) |
1104 | OtherAccess = I; |
1105 | else |
1106 | assert(U == OtherAccess || DT->dominates(I, OtherAccess)); |
1107 | } else |
1108 | OtherAccess = I; |
1109 | } |
1110 | } |
1111 | } |
1112 | |
1113 | if (!OtherAccess) { |
1114 | // There is no dominating use, check if we can find a closest non-dominating |
1115 | // use that lies between any other potentially available use and Load. |
1116 | for (auto *U : Load->getPointerOperand()->users()) { |
1117 | if (U != Load && (isa<LoadInst>(Val: U) || isa<StoreInst>(Val: U))) { |
1118 | auto *I = cast<Instruction>(Val: U); |
1119 | if (I->getFunction() == Load->getFunction() && |
1120 | isPotentiallyReachable(From: I, To: Load, ExclusionSet: nullptr, DT)) { |
1121 | if (OtherAccess) { |
1122 | if (liesBetween(From: OtherAccess, Between: I, To: Load, DT)) { |
1123 | OtherAccess = I; |
1124 | } else if (!liesBetween(From: I, Between: OtherAccess, To: Load, DT)) { |
1125 | // These uses are both partially available at Load were it not for |
1126 | // the clobber, but neither lies strictly after the other. |
1127 | OtherAccess = nullptr; |
1128 | break; |
1129 | } // else: keep current OtherAccess since it lies between U and Load |
1130 | } else { |
1131 | OtherAccess = I; |
1132 | } |
1133 | } |
1134 | } |
1135 | } |
1136 | } |
1137 | |
1138 | if (OtherAccess) |
1139 | R << " in favor of " << NV("OtherAccess" , OtherAccess); |
1140 | |
1141 | R << " because it is clobbered by " << NV("ClobberedBy" , DepInfo.getInst()); |
1142 | |
1143 | ORE->emit(OptDiag&: R); |
1144 | } |
1145 | |
1146 | // Find non-clobbered value for Loc memory location in extended basic block |
1147 | // (chain of basic blocks with single predecessors) starting From instruction. |
1148 | static Value *findDominatingValue(const MemoryLocation &Loc, Type *LoadTy, |
1149 | Instruction *From, AAResults *AA) { |
1150 | uint32_t NumVisitedInsts = 0; |
1151 | BasicBlock *FromBB = From->getParent(); |
1152 | BatchAAResults BatchAA(*AA); |
1153 | for (BasicBlock *BB = FromBB; BB; BB = BB->getSinglePredecessor()) |
1154 | for (auto *Inst = BB == FromBB ? From : BB->getTerminator(); |
1155 | Inst != nullptr; Inst = Inst->getPrevNonDebugInstruction()) { |
1156 | // Stop the search if limit is reached. |
1157 | if (++NumVisitedInsts > MaxNumVisitedInsts) |
1158 | return nullptr; |
1159 | if (isModSet(MRI: BatchAA.getModRefInfo(I: Inst, OptLoc: Loc))) |
1160 | return nullptr; |
1161 | if (auto *LI = dyn_cast<LoadInst>(Val: Inst)) |
1162 | if (LI->getPointerOperand() == Loc.Ptr && LI->getType() == LoadTy) |
1163 | return LI; |
1164 | } |
1165 | return nullptr; |
1166 | } |
1167 | |
1168 | std::optional<AvailableValue> |
1169 | GVNPass::AnalyzeLoadAvailability(LoadInst *Load, MemDepResult DepInfo, |
1170 | Value *Address) { |
1171 | assert(Load->isUnordered() && "rules below are incorrect for ordered access" ); |
1172 | assert(DepInfo.isLocal() && "expected a local dependence" ); |
1173 | |
1174 | Instruction *DepInst = DepInfo.getInst(); |
1175 | |
1176 | const DataLayout &DL = Load->getModule()->getDataLayout(); |
1177 | if (DepInfo.isClobber()) { |
1178 | // If the dependence is to a store that writes to a superset of the bits |
1179 | // read by the load, we can extract the bits we need for the load from the |
1180 | // stored value. |
1181 | if (StoreInst *DepSI = dyn_cast<StoreInst>(Val: DepInst)) { |
1182 | // Can't forward from non-atomic to atomic without violating memory model. |
1183 | if (Address && Load->isAtomic() <= DepSI->isAtomic()) { |
1184 | int Offset = |
1185 | analyzeLoadFromClobberingStore(LoadTy: Load->getType(), LoadPtr: Address, DepSI, DL); |
1186 | if (Offset != -1) |
1187 | return AvailableValue::get(V: DepSI->getValueOperand(), Offset); |
1188 | } |
1189 | } |
1190 | |
1191 | // Check to see if we have something like this: |
1192 | // load i32* P |
1193 | // load i8* (P+1) |
1194 | // if we have this, replace the later with an extraction from the former. |
1195 | if (LoadInst *DepLoad = dyn_cast<LoadInst>(Val: DepInst)) { |
1196 | // If this is a clobber and L is the first instruction in its block, then |
1197 | // we have the first instruction in the entry block. |
1198 | // Can't forward from non-atomic to atomic without violating memory model. |
1199 | if (DepLoad != Load && Address && |
1200 | Load->isAtomic() <= DepLoad->isAtomic()) { |
1201 | Type *LoadType = Load->getType(); |
1202 | int Offset = -1; |
1203 | |
1204 | // If MD reported clobber, check it was nested. |
1205 | if (DepInfo.isClobber() && |
1206 | canCoerceMustAliasedValueToLoad(StoredVal: DepLoad, LoadTy: LoadType, DL)) { |
1207 | const auto ClobberOff = MD->getClobberOffset(DepInst: DepLoad); |
1208 | // GVN has no deal with a negative offset. |
1209 | Offset = (ClobberOff == std::nullopt || *ClobberOff < 0) |
1210 | ? -1 |
1211 | : *ClobberOff; |
1212 | } |
1213 | if (Offset == -1) |
1214 | Offset = |
1215 | analyzeLoadFromClobberingLoad(LoadTy: LoadType, LoadPtr: Address, DepLI: DepLoad, DL); |
1216 | if (Offset != -1) |
1217 | return AvailableValue::getLoad(Load: DepLoad, Offset); |
1218 | } |
1219 | } |
1220 | |
1221 | // If the clobbering value is a memset/memcpy/memmove, see if we can |
1222 | // forward a value on from it. |
1223 | if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Val: DepInst)) { |
1224 | if (Address && !Load->isAtomic()) { |
1225 | int Offset = analyzeLoadFromClobberingMemInst(LoadTy: Load->getType(), LoadPtr: Address, |
1226 | DepMI, DL); |
1227 | if (Offset != -1) |
1228 | return AvailableValue::getMI(MI: DepMI, Offset); |
1229 | } |
1230 | } |
1231 | |
1232 | // Nothing known about this clobber, have to be conservative |
1233 | LLVM_DEBUG( |
1234 | // fast print dep, using operator<< on instruction is too slow. |
1235 | dbgs() << "GVN: load " ; Load->printAsOperand(dbgs()); |
1236 | dbgs() << " is clobbered by " << *DepInst << '\n';); |
1237 | if (ORE->allowExtraAnalysis(DEBUG_TYPE)) |
1238 | reportMayClobberedLoad(Load, DepInfo, DT, ORE); |
1239 | |
1240 | return std::nullopt; |
1241 | } |
1242 | assert(DepInfo.isDef() && "follows from above" ); |
1243 | |
1244 | // Loading the alloca -> undef. |
1245 | // Loading immediately after lifetime begin -> undef. |
1246 | if (isa<AllocaInst>(Val: DepInst) || isLifetimeStart(Inst: DepInst)) |
1247 | return AvailableValue::get(V: UndefValue::get(T: Load->getType())); |
1248 | |
1249 | if (Constant *InitVal = |
1250 | getInitialValueOfAllocation(V: DepInst, TLI, Ty: Load->getType())) |
1251 | return AvailableValue::get(V: InitVal); |
1252 | |
1253 | if (StoreInst *S = dyn_cast<StoreInst>(Val: DepInst)) { |
1254 | // Reject loads and stores that are to the same address but are of |
1255 | // different types if we have to. If the stored value is convertable to |
1256 | // the loaded value, we can reuse it. |
1257 | if (!canCoerceMustAliasedValueToLoad(StoredVal: S->getValueOperand(), LoadTy: Load->getType(), |
1258 | DL)) |
1259 | return std::nullopt; |
1260 | |
1261 | // Can't forward from non-atomic to atomic without violating memory model. |
1262 | if (S->isAtomic() < Load->isAtomic()) |
1263 | return std::nullopt; |
1264 | |
1265 | return AvailableValue::get(V: S->getValueOperand()); |
1266 | } |
1267 | |
1268 | if (LoadInst *LD = dyn_cast<LoadInst>(Val: DepInst)) { |
1269 | // If the types mismatch and we can't handle it, reject reuse of the load. |
1270 | // If the stored value is larger or equal to the loaded value, we can reuse |
1271 | // it. |
1272 | if (!canCoerceMustAliasedValueToLoad(StoredVal: LD, LoadTy: Load->getType(), DL)) |
1273 | return std::nullopt; |
1274 | |
1275 | // Can't forward from non-atomic to atomic without violating memory model. |
1276 | if (LD->isAtomic() < Load->isAtomic()) |
1277 | return std::nullopt; |
1278 | |
1279 | return AvailableValue::getLoad(Load: LD); |
1280 | } |
1281 | |
1282 | // Check if load with Addr dependent from select can be converted to select |
1283 | // between load values. There must be no instructions between the found |
1284 | // loads and DepInst that may clobber the loads. |
1285 | if (auto *Sel = dyn_cast<SelectInst>(Val: DepInst)) { |
1286 | assert(Sel->getType() == Load->getPointerOperandType()); |
1287 | auto Loc = MemoryLocation::get(LI: Load); |
1288 | Value *V1 = |
1289 | findDominatingValue(Loc: Loc.getWithNewPtr(NewPtr: Sel->getTrueValue()), |
1290 | LoadTy: Load->getType(), From: DepInst, AA: getAliasAnalysis()); |
1291 | if (!V1) |
1292 | return std::nullopt; |
1293 | Value *V2 = |
1294 | findDominatingValue(Loc: Loc.getWithNewPtr(NewPtr: Sel->getFalseValue()), |
1295 | LoadTy: Load->getType(), From: DepInst, AA: getAliasAnalysis()); |
1296 | if (!V2) |
1297 | return std::nullopt; |
1298 | return AvailableValue::getSelect(Sel, V1, V2); |
1299 | } |
1300 | |
1301 | // Unknown def - must be conservative |
1302 | LLVM_DEBUG( |
1303 | // fast print dep, using operator<< on instruction is too slow. |
1304 | dbgs() << "GVN: load " ; Load->printAsOperand(dbgs()); |
1305 | dbgs() << " has unknown def " << *DepInst << '\n';); |
1306 | return std::nullopt; |
1307 | } |
1308 | |
1309 | void GVNPass::AnalyzeLoadAvailability(LoadInst *Load, LoadDepVect &Deps, |
1310 | AvailValInBlkVect &ValuesPerBlock, |
1311 | UnavailBlkVect &UnavailableBlocks) { |
1312 | // Filter out useless results (non-locals, etc). Keep track of the blocks |
1313 | // where we have a value available in repl, also keep track of whether we see |
1314 | // dependencies that produce an unknown value for the load (such as a call |
1315 | // that could potentially clobber the load). |
1316 | for (const auto &Dep : Deps) { |
1317 | BasicBlock *DepBB = Dep.getBB(); |
1318 | MemDepResult DepInfo = Dep.getResult(); |
1319 | |
1320 | if (DeadBlocks.count(key: DepBB)) { |
1321 | // Dead dependent mem-op disguise as a load evaluating the same value |
1322 | // as the load in question. |
1323 | ValuesPerBlock.push_back(Elt: AvailableValueInBlock::getUndef(BB: DepBB)); |
1324 | continue; |
1325 | } |
1326 | |
1327 | if (!DepInfo.isLocal()) { |
1328 | UnavailableBlocks.push_back(Elt: DepBB); |
1329 | continue; |
1330 | } |
1331 | |
1332 | // The address being loaded in this non-local block may not be the same as |
1333 | // the pointer operand of the load if PHI translation occurs. Make sure |
1334 | // to consider the right address. |
1335 | if (auto AV = AnalyzeLoadAvailability(Load, DepInfo, Address: Dep.getAddress())) { |
1336 | // subtlety: because we know this was a non-local dependency, we know |
1337 | // it's safe to materialize anywhere between the instruction within |
1338 | // DepInfo and the end of it's block. |
1339 | ValuesPerBlock.push_back( |
1340 | Elt: AvailableValueInBlock::get(BB: DepBB, AV: std::move(*AV))); |
1341 | } else { |
1342 | UnavailableBlocks.push_back(Elt: DepBB); |
1343 | } |
1344 | } |
1345 | |
1346 | assert(Deps.size() == ValuesPerBlock.size() + UnavailableBlocks.size() && |
1347 | "post condition violation" ); |
1348 | } |
1349 | |
1350 | /// Given the following code, v1 is partially available on some edges, but not |
1351 | /// available on the edge from PredBB. This function tries to find if there is |
1352 | /// another identical load in the other successor of PredBB. |
1353 | /// |
1354 | /// v0 = load %addr |
1355 | /// br %LoadBB |
1356 | /// |
1357 | /// LoadBB: |
1358 | /// v1 = load %addr |
1359 | /// ... |
1360 | /// |
1361 | /// PredBB: |
1362 | /// ... |
1363 | /// br %cond, label %LoadBB, label %SuccBB |
1364 | /// |
1365 | /// SuccBB: |
1366 | /// v2 = load %addr |
1367 | /// ... |
1368 | /// |
1369 | LoadInst *GVNPass::findLoadToHoistIntoPred(BasicBlock *Pred, BasicBlock *LoadBB, |
1370 | LoadInst *Load) { |
1371 | // For simplicity we handle a Pred has 2 successors only. |
1372 | auto *Term = Pred->getTerminator(); |
1373 | if (Term->getNumSuccessors() != 2 || Term->isSpecialTerminator()) |
1374 | return nullptr; |
1375 | auto *SuccBB = Term->getSuccessor(Idx: 0); |
1376 | if (SuccBB == LoadBB) |
1377 | SuccBB = Term->getSuccessor(Idx: 1); |
1378 | if (!SuccBB->getSinglePredecessor()) |
1379 | return nullptr; |
1380 | |
1381 | unsigned int NumInsts = MaxNumInsnsPerBlock; |
1382 | for (Instruction &Inst : *SuccBB) { |
1383 | if (Inst.isDebugOrPseudoInst()) |
1384 | continue; |
1385 | if (--NumInsts == 0) |
1386 | return nullptr; |
1387 | |
1388 | if (!Inst.isIdenticalTo(I: Load)) |
1389 | continue; |
1390 | |
1391 | MemDepResult Dep = MD->getDependency(QueryInst: &Inst); |
1392 | // If an identical load doesn't depends on any local instructions, it can |
1393 | // be safely moved to PredBB. |
1394 | // Also check for the implicit control flow instructions. See the comments |
1395 | // in PerformLoadPRE for details. |
1396 | if (Dep.isNonLocal() && !ICF->isDominatedByICFIFromSameBlock(Insn: &Inst)) |
1397 | return cast<LoadInst>(Val: &Inst); |
1398 | |
1399 | // Otherwise there is something in the same BB clobbers the memory, we can't |
1400 | // move this and later load to PredBB. |
1401 | return nullptr; |
1402 | } |
1403 | |
1404 | return nullptr; |
1405 | } |
1406 | |
1407 | void GVNPass::eliminatePartiallyRedundantLoad( |
1408 | LoadInst *Load, AvailValInBlkVect &ValuesPerBlock, |
1409 | MapVector<BasicBlock *, Value *> &AvailableLoads, |
1410 | MapVector<BasicBlock *, LoadInst *> *CriticalEdgePredAndLoad) { |
1411 | for (const auto &AvailableLoad : AvailableLoads) { |
1412 | BasicBlock *UnavailableBlock = AvailableLoad.first; |
1413 | Value *LoadPtr = AvailableLoad.second; |
1414 | |
1415 | auto *NewLoad = |
1416 | new LoadInst(Load->getType(), LoadPtr, Load->getName() + ".pre" , |
1417 | Load->isVolatile(), Load->getAlign(), Load->getOrdering(), |
1418 | Load->getSyncScopeID(), UnavailableBlock->getTerminator()); |
1419 | NewLoad->setDebugLoc(Load->getDebugLoc()); |
1420 | if (MSSAU) { |
1421 | auto *NewAccess = MSSAU->createMemoryAccessInBB( |
1422 | I: NewLoad, Definition: nullptr, BB: NewLoad->getParent(), Point: MemorySSA::BeforeTerminator); |
1423 | if (auto *NewDef = dyn_cast<MemoryDef>(Val: NewAccess)) |
1424 | MSSAU->insertDef(Def: NewDef, /*RenameUses=*/true); |
1425 | else |
1426 | MSSAU->insertUse(Use: cast<MemoryUse>(Val: NewAccess), /*RenameUses=*/true); |
1427 | } |
1428 | |
1429 | // Transfer the old load's AA tags to the new load. |
1430 | AAMDNodes Tags = Load->getAAMetadata(); |
1431 | if (Tags) |
1432 | NewLoad->setAAMetadata(Tags); |
1433 | |
1434 | if (auto *MD = Load->getMetadata(KindID: LLVMContext::MD_invariant_load)) |
1435 | NewLoad->setMetadata(KindID: LLVMContext::MD_invariant_load, Node: MD); |
1436 | if (auto *InvGroupMD = Load->getMetadata(KindID: LLVMContext::MD_invariant_group)) |
1437 | NewLoad->setMetadata(KindID: LLVMContext::MD_invariant_group, Node: InvGroupMD); |
1438 | if (auto *RangeMD = Load->getMetadata(KindID: LLVMContext::MD_range)) |
1439 | NewLoad->setMetadata(KindID: LLVMContext::MD_range, Node: RangeMD); |
1440 | if (auto *AccessMD = Load->getMetadata(KindID: LLVMContext::MD_access_group)) |
1441 | if (LI->getLoopFor(BB: Load->getParent()) == LI->getLoopFor(BB: UnavailableBlock)) |
1442 | NewLoad->setMetadata(KindID: LLVMContext::MD_access_group, Node: AccessMD); |
1443 | |
1444 | // We do not propagate the old load's debug location, because the new |
1445 | // load now lives in a different BB, and we want to avoid a jumpy line |
1446 | // table. |
1447 | // FIXME: How do we retain source locations without causing poor debugging |
1448 | // behavior? |
1449 | |
1450 | // Add the newly created load. |
1451 | ValuesPerBlock.push_back( |
1452 | Elt: AvailableValueInBlock::get(BB: UnavailableBlock, V: NewLoad)); |
1453 | MD->invalidateCachedPointerInfo(Ptr: LoadPtr); |
1454 | LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); |
1455 | |
1456 | // For PredBB in CriticalEdgePredAndLoad we need to replace the uses of old |
1457 | // load instruction with the new created load instruction. |
1458 | if (CriticalEdgePredAndLoad) { |
1459 | auto I = CriticalEdgePredAndLoad->find(Key: UnavailableBlock); |
1460 | if (I != CriticalEdgePredAndLoad->end()) { |
1461 | ++NumPRELoadMoved2CEPred; |
1462 | ICF->insertInstructionTo(Inst: NewLoad, BB: UnavailableBlock); |
1463 | LoadInst *OldLoad = I->second; |
1464 | combineMetadataForCSE(K: NewLoad, J: OldLoad, DoesKMove: false); |
1465 | OldLoad->replaceAllUsesWith(V: NewLoad); |
1466 | replaceValuesPerBlockEntry(ValuesPerBlock, OldValue: OldLoad, NewValue: NewLoad); |
1467 | if (uint32_t ValNo = VN.lookup(V: OldLoad, Verify: false)) |
1468 | removeFromLeaderTable(N: ValNo, I: OldLoad, BB: OldLoad->getParent()); |
1469 | VN.erase(V: OldLoad); |
1470 | removeInstruction(I: OldLoad); |
1471 | } |
1472 | } |
1473 | } |
1474 | |
1475 | // Perform PHI construction. |
1476 | Value *V = ConstructSSAForLoadSet(Load, ValuesPerBlock, gvn&: *this); |
1477 | // ConstructSSAForLoadSet is responsible for combining metadata. |
1478 | ICF->removeUsersOf(Inst: Load); |
1479 | Load->replaceAllUsesWith(V); |
1480 | if (isa<PHINode>(Val: V)) |
1481 | V->takeName(V: Load); |
1482 | if (Instruction *I = dyn_cast<Instruction>(Val: V)) |
1483 | I->setDebugLoc(Load->getDebugLoc()); |
1484 | if (V->getType()->isPtrOrPtrVectorTy()) |
1485 | MD->invalidateCachedPointerInfo(Ptr: V); |
1486 | markInstructionForDeletion(I: Load); |
1487 | ORE->emit(RemarkBuilder: [&]() { |
1488 | return OptimizationRemark(DEBUG_TYPE, "LoadPRE" , Load) |
1489 | << "load eliminated by PRE" ; |
1490 | }); |
1491 | } |
1492 | |
1493 | bool GVNPass::PerformLoadPRE(LoadInst *Load, AvailValInBlkVect &ValuesPerBlock, |
1494 | UnavailBlkVect &UnavailableBlocks) { |
1495 | // Okay, we have *some* definitions of the value. This means that the value |
1496 | // is available in some of our (transitive) predecessors. Lets think about |
1497 | // doing PRE of this load. This will involve inserting a new load into the |
1498 | // predecessor when it's not available. We could do this in general, but |
1499 | // prefer to not increase code size. As such, we only do this when we know |
1500 | // that we only have to insert *one* load (which means we're basically moving |
1501 | // the load, not inserting a new one). |
1502 | |
1503 | SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(), |
1504 | UnavailableBlocks.end()); |
1505 | |
1506 | // Let's find the first basic block with more than one predecessor. Walk |
1507 | // backwards through predecessors if needed. |
1508 | BasicBlock *LoadBB = Load->getParent(); |
1509 | BasicBlock *TmpBB = LoadBB; |
1510 | |
1511 | // Check that there is no implicit control flow instructions above our load in |
1512 | // its block. If there is an instruction that doesn't always pass the |
1513 | // execution to the following instruction, then moving through it may become |
1514 | // invalid. For example: |
1515 | // |
1516 | // int arr[LEN]; |
1517 | // int index = ???; |
1518 | // ... |
1519 | // guard(0 <= index && index < LEN); |
1520 | // use(arr[index]); |
1521 | // |
1522 | // It is illegal to move the array access to any point above the guard, |
1523 | // because if the index is out of bounds we should deoptimize rather than |
1524 | // access the array. |
1525 | // Check that there is no guard in this block above our instruction. |
1526 | bool MustEnsureSafetyOfSpeculativeExecution = |
1527 | ICF->isDominatedByICFIFromSameBlock(Insn: Load); |
1528 | |
1529 | while (TmpBB->getSinglePredecessor()) { |
1530 | TmpBB = TmpBB->getSinglePredecessor(); |
1531 | if (TmpBB == LoadBB) // Infinite (unreachable) loop. |
1532 | return false; |
1533 | if (Blockers.count(Ptr: TmpBB)) |
1534 | return false; |
1535 | |
1536 | // If any of these blocks has more than one successor (i.e. if the edge we |
1537 | // just traversed was critical), then there are other paths through this |
1538 | // block along which the load may not be anticipated. Hoisting the load |
1539 | // above this block would be adding the load to execution paths along |
1540 | // which it was not previously executed. |
1541 | if (TmpBB->getTerminator()->getNumSuccessors() != 1) |
1542 | return false; |
1543 | |
1544 | // Check that there is no implicit control flow in a block above. |
1545 | MustEnsureSafetyOfSpeculativeExecution = |
1546 | MustEnsureSafetyOfSpeculativeExecution || ICF->hasICF(BB: TmpBB); |
1547 | } |
1548 | |
1549 | assert(TmpBB); |
1550 | LoadBB = TmpBB; |
1551 | |
1552 | // Check to see how many predecessors have the loaded value fully |
1553 | // available. |
1554 | MapVector<BasicBlock *, Value *> PredLoads; |
1555 | DenseMap<BasicBlock *, AvailabilityState> FullyAvailableBlocks; |
1556 | for (const AvailableValueInBlock &AV : ValuesPerBlock) |
1557 | FullyAvailableBlocks[AV.BB] = AvailabilityState::Available; |
1558 | for (BasicBlock *UnavailableBB : UnavailableBlocks) |
1559 | FullyAvailableBlocks[UnavailableBB] = AvailabilityState::Unavailable; |
1560 | |
1561 | // The edge from Pred to LoadBB is a critical edge will be splitted. |
1562 | SmallVector<BasicBlock *, 4> CriticalEdgePredSplit; |
1563 | // The edge from Pred to LoadBB is a critical edge, another successor of Pred |
1564 | // contains a load can be moved to Pred. This data structure maps the Pred to |
1565 | // the movable load. |
1566 | MapVector<BasicBlock *, LoadInst *> CriticalEdgePredAndLoad; |
1567 | for (BasicBlock *Pred : predecessors(BB: LoadBB)) { |
1568 | // If any predecessor block is an EH pad that does not allow non-PHI |
1569 | // instructions before the terminator, we can't PRE the load. |
1570 | if (Pred->getTerminator()->isEHPad()) { |
1571 | LLVM_DEBUG( |
1572 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '" |
1573 | << Pred->getName() << "': " << *Load << '\n'); |
1574 | return false; |
1575 | } |
1576 | |
1577 | if (IsValueFullyAvailableInBlock(BB: Pred, FullyAvailableBlocks)) { |
1578 | continue; |
1579 | } |
1580 | |
1581 | if (Pred->getTerminator()->getNumSuccessors() != 1) { |
1582 | if (isa<IndirectBrInst>(Val: Pred->getTerminator())) { |
1583 | LLVM_DEBUG( |
1584 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" |
1585 | << Pred->getName() << "': " << *Load << '\n'); |
1586 | return false; |
1587 | } |
1588 | |
1589 | if (LoadBB->isEHPad()) { |
1590 | LLVM_DEBUG( |
1591 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '" |
1592 | << Pred->getName() << "': " << *Load << '\n'); |
1593 | return false; |
1594 | } |
1595 | |
1596 | // Do not split backedge as it will break the canonical loop form. |
1597 | if (!isLoadPRESplitBackedgeEnabled()) |
1598 | if (DT->dominates(A: LoadBB, B: Pred)) { |
1599 | LLVM_DEBUG( |
1600 | dbgs() |
1601 | << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" |
1602 | << Pred->getName() << "': " << *Load << '\n'); |
1603 | return false; |
1604 | } |
1605 | |
1606 | if (LoadInst *LI = findLoadToHoistIntoPred(Pred, LoadBB, Load)) |
1607 | CriticalEdgePredAndLoad[Pred] = LI; |
1608 | else |
1609 | CriticalEdgePredSplit.push_back(Elt: Pred); |
1610 | } else { |
1611 | // Only add the predecessors that will not be split for now. |
1612 | PredLoads[Pred] = nullptr; |
1613 | } |
1614 | } |
1615 | |
1616 | // Decide whether PRE is profitable for this load. |
1617 | unsigned NumInsertPreds = PredLoads.size() + CriticalEdgePredSplit.size(); |
1618 | unsigned NumUnavailablePreds = NumInsertPreds + |
1619 | CriticalEdgePredAndLoad.size(); |
1620 | assert(NumUnavailablePreds != 0 && |
1621 | "Fully available value should already be eliminated!" ); |
1622 | (void)NumUnavailablePreds; |
1623 | |
1624 | // If we need to insert new load in multiple predecessors, reject it. |
1625 | // FIXME: If we could restructure the CFG, we could make a common pred with |
1626 | // all the preds that don't have an available Load and insert a new load into |
1627 | // that one block. |
1628 | if (NumInsertPreds > 1) |
1629 | return false; |
1630 | |
1631 | // Now we know where we will insert load. We must ensure that it is safe |
1632 | // to speculatively execute the load at that points. |
1633 | if (MustEnsureSafetyOfSpeculativeExecution) { |
1634 | if (CriticalEdgePredSplit.size()) |
1635 | if (!isSafeToSpeculativelyExecute(I: Load, CtxI: LoadBB->getFirstNonPHI(), AC, DT)) |
1636 | return false; |
1637 | for (auto &PL : PredLoads) |
1638 | if (!isSafeToSpeculativelyExecute(I: Load, CtxI: PL.first->getTerminator(), AC, |
1639 | DT)) |
1640 | return false; |
1641 | for (auto &CEP : CriticalEdgePredAndLoad) |
1642 | if (!isSafeToSpeculativelyExecute(I: Load, CtxI: CEP.first->getTerminator(), AC, |
1643 | DT)) |
1644 | return false; |
1645 | } |
1646 | |
1647 | // Split critical edges, and update the unavailable predecessors accordingly. |
1648 | for (BasicBlock *OrigPred : CriticalEdgePredSplit) { |
1649 | BasicBlock *NewPred = splitCriticalEdges(Pred: OrigPred, Succ: LoadBB); |
1650 | assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!" ); |
1651 | PredLoads[NewPred] = nullptr; |
1652 | LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->" |
1653 | << LoadBB->getName() << '\n'); |
1654 | } |
1655 | |
1656 | for (auto &CEP : CriticalEdgePredAndLoad) |
1657 | PredLoads[CEP.first] = nullptr; |
1658 | |
1659 | // Check if the load can safely be moved to all the unavailable predecessors. |
1660 | bool CanDoPRE = true; |
1661 | const DataLayout &DL = Load->getModule()->getDataLayout(); |
1662 | SmallVector<Instruction*, 8> NewInsts; |
1663 | for (auto &PredLoad : PredLoads) { |
1664 | BasicBlock *UnavailablePred = PredLoad.first; |
1665 | |
1666 | // Do PHI translation to get its value in the predecessor if necessary. The |
1667 | // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. |
1668 | // We do the translation for each edge we skipped by going from Load's block |
1669 | // to LoadBB, otherwise we might miss pieces needing translation. |
1670 | |
1671 | // If all preds have a single successor, then we know it is safe to insert |
1672 | // the load on the pred (?!?), so we can insert code to materialize the |
1673 | // pointer if it is not available. |
1674 | Value *LoadPtr = Load->getPointerOperand(); |
1675 | BasicBlock *Cur = Load->getParent(); |
1676 | while (Cur != LoadBB) { |
1677 | PHITransAddr Address(LoadPtr, DL, AC); |
1678 | LoadPtr = Address.translateWithInsertion(CurBB: Cur, PredBB: Cur->getSinglePredecessor(), |
1679 | DT: *DT, NewInsts); |
1680 | if (!LoadPtr) { |
1681 | CanDoPRE = false; |
1682 | break; |
1683 | } |
1684 | Cur = Cur->getSinglePredecessor(); |
1685 | } |
1686 | |
1687 | if (LoadPtr) { |
1688 | PHITransAddr Address(LoadPtr, DL, AC); |
1689 | LoadPtr = Address.translateWithInsertion(CurBB: LoadBB, PredBB: UnavailablePred, DT: *DT, |
1690 | NewInsts); |
1691 | } |
1692 | // If we couldn't find or insert a computation of this phi translated value, |
1693 | // we fail PRE. |
1694 | if (!LoadPtr) { |
1695 | LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " |
1696 | << *Load->getPointerOperand() << "\n" ); |
1697 | CanDoPRE = false; |
1698 | break; |
1699 | } |
1700 | |
1701 | PredLoad.second = LoadPtr; |
1702 | } |
1703 | |
1704 | if (!CanDoPRE) { |
1705 | while (!NewInsts.empty()) { |
1706 | // Erase instructions generated by the failed PHI translation before |
1707 | // trying to number them. PHI translation might insert instructions |
1708 | // in basic blocks other than the current one, and we delete them |
1709 | // directly, as markInstructionForDeletion only allows removing from the |
1710 | // current basic block. |
1711 | NewInsts.pop_back_val()->eraseFromParent(); |
1712 | } |
1713 | // HINT: Don't revert the edge-splitting as following transformation may |
1714 | // also need to split these critical edges. |
1715 | return !CriticalEdgePredSplit.empty(); |
1716 | } |
1717 | |
1718 | // Okay, we can eliminate this load by inserting a reload in the predecessor |
1719 | // and using PHI construction to get the value in the other predecessors, do |
1720 | // it. |
1721 | LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *Load << '\n'); |
1722 | LLVM_DEBUG(if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size() |
1723 | << " INSTS: " << *NewInsts.back() |
1724 | << '\n'); |
1725 | |
1726 | // Assign value numbers to the new instructions. |
1727 | for (Instruction *I : NewInsts) { |
1728 | // Instructions that have been inserted in predecessor(s) to materialize |
1729 | // the load address do not retain their original debug locations. Doing |
1730 | // so could lead to confusing (but correct) source attributions. |
1731 | I->updateLocationAfterHoist(); |
1732 | |
1733 | // FIXME: We really _ought_ to insert these value numbers into their |
1734 | // parent's availability map. However, in doing so, we risk getting into |
1735 | // ordering issues. If a block hasn't been processed yet, we would be |
1736 | // marking a value as AVAIL-IN, which isn't what we intend. |
1737 | VN.lookupOrAdd(V: I); |
1738 | } |
1739 | |
1740 | eliminatePartiallyRedundantLoad(Load, ValuesPerBlock, AvailableLoads&: PredLoads, |
1741 | CriticalEdgePredAndLoad: &CriticalEdgePredAndLoad); |
1742 | ++NumPRELoad; |
1743 | return true; |
1744 | } |
1745 | |
1746 | bool GVNPass::performLoopLoadPRE(LoadInst *Load, |
1747 | AvailValInBlkVect &ValuesPerBlock, |
1748 | UnavailBlkVect &UnavailableBlocks) { |
1749 | const Loop *L = LI->getLoopFor(BB: Load->getParent()); |
1750 | // TODO: Generalize to other loop blocks that dominate the latch. |
1751 | if (!L || L->getHeader() != Load->getParent()) |
1752 | return false; |
1753 | |
1754 | BasicBlock * = L->getLoopPreheader(); |
1755 | BasicBlock *Latch = L->getLoopLatch(); |
1756 | if (!Preheader || !Latch) |
1757 | return false; |
1758 | |
1759 | Value *LoadPtr = Load->getPointerOperand(); |
1760 | // Must be available in preheader. |
1761 | if (!L->isLoopInvariant(V: LoadPtr)) |
1762 | return false; |
1763 | |
1764 | // We plan to hoist the load to preheader without introducing a new fault. |
1765 | // In order to do it, we need to prove that we cannot side-exit the loop |
1766 | // once loop header is first entered before execution of the load. |
1767 | if (ICF->isDominatedByICFIFromSameBlock(Insn: Load)) |
1768 | return false; |
1769 | |
1770 | BasicBlock *LoopBlock = nullptr; |
1771 | for (auto *Blocker : UnavailableBlocks) { |
1772 | // Blockers from outside the loop are handled in preheader. |
1773 | if (!L->contains(BB: Blocker)) |
1774 | continue; |
1775 | |
1776 | // Only allow one loop block. Loop header is not less frequently executed |
1777 | // than each loop block, and likely it is much more frequently executed. But |
1778 | // in case of multiple loop blocks, we need extra information (such as block |
1779 | // frequency info) to understand whether it is profitable to PRE into |
1780 | // multiple loop blocks. |
1781 | if (LoopBlock) |
1782 | return false; |
1783 | |
1784 | // Do not sink into inner loops. This may be non-profitable. |
1785 | if (L != LI->getLoopFor(BB: Blocker)) |
1786 | return false; |
1787 | |
1788 | // Blocks that dominate the latch execute on every single iteration, maybe |
1789 | // except the last one. So PREing into these blocks doesn't make much sense |
1790 | // in most cases. But the blocks that do not necessarily execute on each |
1791 | // iteration are sometimes much colder than the header, and this is when |
1792 | // PRE is potentially profitable. |
1793 | if (DT->dominates(A: Blocker, B: Latch)) |
1794 | return false; |
1795 | |
1796 | // Make sure that the terminator itself doesn't clobber. |
1797 | if (Blocker->getTerminator()->mayWriteToMemory()) |
1798 | return false; |
1799 | |
1800 | LoopBlock = Blocker; |
1801 | } |
1802 | |
1803 | if (!LoopBlock) |
1804 | return false; |
1805 | |
1806 | // Make sure the memory at this pointer cannot be freed, therefore we can |
1807 | // safely reload from it after clobber. |
1808 | if (LoadPtr->canBeFreed()) |
1809 | return false; |
1810 | |
1811 | // TODO: Support critical edge splitting if blocker has more than 1 successor. |
1812 | MapVector<BasicBlock *, Value *> AvailableLoads; |
1813 | AvailableLoads[LoopBlock] = LoadPtr; |
1814 | AvailableLoads[Preheader] = LoadPtr; |
1815 | |
1816 | LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOOP LOAD: " << *Load << '\n'); |
1817 | eliminatePartiallyRedundantLoad(Load, ValuesPerBlock, AvailableLoads, |
1818 | /*CriticalEdgePredAndLoad*/ nullptr); |
1819 | ++NumPRELoopLoad; |
1820 | return true; |
1821 | } |
1822 | |
1823 | static void (LoadInst *Load, Value *AvailableValue, |
1824 | OptimizationRemarkEmitter *ORE) { |
1825 | using namespace ore; |
1826 | |
1827 | ORE->emit(RemarkBuilder: [&]() { |
1828 | return OptimizationRemark(DEBUG_TYPE, "LoadElim" , Load) |
1829 | << "load of type " << NV("Type" , Load->getType()) << " eliminated" |
1830 | << setExtraArgs() << " in favor of " |
1831 | << NV("InfavorOfValue" , AvailableValue); |
1832 | }); |
1833 | } |
1834 | |
1835 | /// Attempt to eliminate a load whose dependencies are |
1836 | /// non-local by performing PHI construction. |
1837 | bool GVNPass::processNonLocalLoad(LoadInst *Load) { |
1838 | // non-local speculations are not allowed under asan. |
1839 | if (Load->getParent()->getParent()->hasFnAttribute( |
1840 | Attribute::SanitizeAddress) || |
1841 | Load->getParent()->getParent()->hasFnAttribute( |
1842 | Attribute::SanitizeHWAddress)) |
1843 | return false; |
1844 | |
1845 | // Step 1: Find the non-local dependencies of the load. |
1846 | LoadDepVect Deps; |
1847 | MD->getNonLocalPointerDependency(QueryInst: Load, Result&: Deps); |
1848 | |
1849 | // If we had to process more than one hundred blocks to find the |
1850 | // dependencies, this load isn't worth worrying about. Optimizing |
1851 | // it will be too expensive. |
1852 | unsigned NumDeps = Deps.size(); |
1853 | if (NumDeps > MaxNumDeps) |
1854 | return false; |
1855 | |
1856 | // If we had a phi translation failure, we'll have a single entry which is a |
1857 | // clobber in the current block. Reject this early. |
1858 | if (NumDeps == 1 && |
1859 | !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { |
1860 | LLVM_DEBUG(dbgs() << "GVN: non-local load " ; Load->printAsOperand(dbgs()); |
1861 | dbgs() << " has unknown dependencies\n" ;); |
1862 | return false; |
1863 | } |
1864 | |
1865 | bool Changed = false; |
1866 | // If this load follows a GEP, see if we can PRE the indices before analyzing. |
1867 | if (GetElementPtrInst *GEP = |
1868 | dyn_cast<GetElementPtrInst>(Val: Load->getOperand(i_nocapture: 0))) { |
1869 | for (Use &U : GEP->indices()) |
1870 | if (Instruction *I = dyn_cast<Instruction>(Val: U.get())) |
1871 | Changed |= performScalarPRE(I); |
1872 | } |
1873 | |
1874 | // Step 2: Analyze the availability of the load |
1875 | AvailValInBlkVect ValuesPerBlock; |
1876 | UnavailBlkVect UnavailableBlocks; |
1877 | AnalyzeLoadAvailability(Load, Deps, ValuesPerBlock, UnavailableBlocks); |
1878 | |
1879 | // If we have no predecessors that produce a known value for this load, exit |
1880 | // early. |
1881 | if (ValuesPerBlock.empty()) |
1882 | return Changed; |
1883 | |
1884 | // Step 3: Eliminate fully redundancy. |
1885 | // |
1886 | // If all of the instructions we depend on produce a known value for this |
1887 | // load, then it is fully redundant and we can use PHI insertion to compute |
1888 | // its value. Insert PHIs and remove the fully redundant value now. |
1889 | if (UnavailableBlocks.empty()) { |
1890 | LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *Load << '\n'); |
1891 | |
1892 | // Perform PHI construction. |
1893 | Value *V = ConstructSSAForLoadSet(Load, ValuesPerBlock, gvn&: *this); |
1894 | // ConstructSSAForLoadSet is responsible for combining metadata. |
1895 | ICF->removeUsersOf(Inst: Load); |
1896 | Load->replaceAllUsesWith(V); |
1897 | |
1898 | if (isa<PHINode>(Val: V)) |
1899 | V->takeName(V: Load); |
1900 | if (Instruction *I = dyn_cast<Instruction>(Val: V)) |
1901 | // If instruction I has debug info, then we should not update it. |
1902 | // Also, if I has a null DebugLoc, then it is still potentially incorrect |
1903 | // to propagate Load's DebugLoc because Load may not post-dominate I. |
1904 | if (Load->getDebugLoc() && Load->getParent() == I->getParent()) |
1905 | I->setDebugLoc(Load->getDebugLoc()); |
1906 | if (V->getType()->isPtrOrPtrVectorTy()) |
1907 | MD->invalidateCachedPointerInfo(Ptr: V); |
1908 | markInstructionForDeletion(I: Load); |
1909 | ++NumGVNLoad; |
1910 | reportLoadElim(Load, AvailableValue: V, ORE); |
1911 | return true; |
1912 | } |
1913 | |
1914 | // Step 4: Eliminate partial redundancy. |
1915 | if (!isPREEnabled() || !isLoadPREEnabled()) |
1916 | return Changed; |
1917 | if (!isLoadInLoopPREEnabled() && LI->getLoopFor(BB: Load->getParent())) |
1918 | return Changed; |
1919 | |
1920 | if (performLoopLoadPRE(Load, ValuesPerBlock, UnavailableBlocks) || |
1921 | PerformLoadPRE(Load, ValuesPerBlock, UnavailableBlocks)) |
1922 | return true; |
1923 | |
1924 | return Changed; |
1925 | } |
1926 | |
1927 | static bool impliesEquivalanceIfTrue(CmpInst* Cmp) { |
1928 | if (Cmp->getPredicate() == CmpInst::Predicate::ICMP_EQ) |
1929 | return true; |
1930 | |
1931 | // Floating point comparisons can be equal, but not equivalent. Cases: |
1932 | // NaNs for unordered operators |
1933 | // +0.0 vs 0.0 for all operators |
1934 | if (Cmp->getPredicate() == CmpInst::Predicate::FCMP_OEQ || |
1935 | (Cmp->getPredicate() == CmpInst::Predicate::FCMP_UEQ && |
1936 | Cmp->getFastMathFlags().noNaNs())) { |
1937 | Value *LHS = Cmp->getOperand(i_nocapture: 0); |
1938 | Value *RHS = Cmp->getOperand(i_nocapture: 1); |
1939 | // If we can prove either side non-zero, then equality must imply |
1940 | // equivalence. |
1941 | // FIXME: We should do this optimization if 'no signed zeros' is |
1942 | // applicable via an instruction-level fast-math-flag or some other |
1943 | // indicator that relaxed FP semantics are being used. |
1944 | if (isa<ConstantFP>(Val: LHS) && !cast<ConstantFP>(Val: LHS)->isZero()) |
1945 | return true; |
1946 | if (isa<ConstantFP>(Val: RHS) && !cast<ConstantFP>(Val: RHS)->isZero()) |
1947 | return true; |
1948 | // TODO: Handle vector floating point constants |
1949 | } |
1950 | return false; |
1951 | } |
1952 | |
1953 | static bool impliesEquivalanceIfFalse(CmpInst* Cmp) { |
1954 | if (Cmp->getPredicate() == CmpInst::Predicate::ICMP_NE) |
1955 | return true; |
1956 | |
1957 | // Floating point comparisons can be equal, but not equivelent. Cases: |
1958 | // NaNs for unordered operators |
1959 | // +0.0 vs 0.0 for all operators |
1960 | if ((Cmp->getPredicate() == CmpInst::Predicate::FCMP_ONE && |
1961 | Cmp->getFastMathFlags().noNaNs()) || |
1962 | Cmp->getPredicate() == CmpInst::Predicate::FCMP_UNE) { |
1963 | Value *LHS = Cmp->getOperand(i_nocapture: 0); |
1964 | Value *RHS = Cmp->getOperand(i_nocapture: 1); |
1965 | // If we can prove either side non-zero, then equality must imply |
1966 | // equivalence. |
1967 | // FIXME: We should do this optimization if 'no signed zeros' is |
1968 | // applicable via an instruction-level fast-math-flag or some other |
1969 | // indicator that relaxed FP semantics are being used. |
1970 | if (isa<ConstantFP>(Val: LHS) && !cast<ConstantFP>(Val: LHS)->isZero()) |
1971 | return true; |
1972 | if (isa<ConstantFP>(Val: RHS) && !cast<ConstantFP>(Val: RHS)->isZero()) |
1973 | return true; |
1974 | // TODO: Handle vector floating point constants |
1975 | } |
1976 | return false; |
1977 | } |
1978 | |
1979 | |
1980 | static bool hasUsersIn(Value *V, BasicBlock *BB) { |
1981 | return llvm::any_of(Range: V->users(), P: [BB](User *U) { |
1982 | auto *I = dyn_cast<Instruction>(Val: U); |
1983 | return I && I->getParent() == BB; |
1984 | }); |
1985 | } |
1986 | |
1987 | bool GVNPass::processAssumeIntrinsic(AssumeInst *IntrinsicI) { |
1988 | Value *V = IntrinsicI->getArgOperand(i: 0); |
1989 | |
1990 | if (ConstantInt *Cond = dyn_cast<ConstantInt>(Val: V)) { |
1991 | if (Cond->isZero()) { |
1992 | Type *Int8Ty = Type::getInt8Ty(C&: V->getContext()); |
1993 | Type *PtrTy = PointerType::get(C&: V->getContext(), AddressSpace: 0); |
1994 | // Insert a new store to null instruction before the load to indicate that |
1995 | // this code is not reachable. FIXME: We could insert unreachable |
1996 | // instruction directly because we can modify the CFG. |
1997 | auto *NewS = new StoreInst(PoisonValue::get(T: Int8Ty), |
1998 | Constant::getNullValue(Ty: PtrTy), IntrinsicI); |
1999 | if (MSSAU) { |
2000 | const MemoryUseOrDef *FirstNonDom = nullptr; |
2001 | const auto *AL = |
2002 | MSSAU->getMemorySSA()->getBlockAccesses(BB: IntrinsicI->getParent()); |
2003 | |
2004 | // If there are accesses in the current basic block, find the first one |
2005 | // that does not come before NewS. The new memory access is inserted |
2006 | // after the found access or before the terminator if no such access is |
2007 | // found. |
2008 | if (AL) { |
2009 | for (const auto &Acc : *AL) { |
2010 | if (auto *Current = dyn_cast<MemoryUseOrDef>(Val: &Acc)) |
2011 | if (!Current->getMemoryInst()->comesBefore(Other: NewS)) { |
2012 | FirstNonDom = Current; |
2013 | break; |
2014 | } |
2015 | } |
2016 | } |
2017 | |
2018 | auto *NewDef = |
2019 | FirstNonDom ? MSSAU->createMemoryAccessBefore( |
2020 | I: NewS, Definition: nullptr, |
2021 | InsertPt: const_cast<MemoryUseOrDef *>(FirstNonDom)) |
2022 | : MSSAU->createMemoryAccessInBB( |
2023 | I: NewS, Definition: nullptr, |
2024 | BB: NewS->getParent(), Point: MemorySSA::BeforeTerminator); |
2025 | |
2026 | MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewDef), /*RenameUses=*/false); |
2027 | } |
2028 | } |
2029 | if (isAssumeWithEmptyBundle(Assume: *IntrinsicI)) { |
2030 | markInstructionForDeletion(I: IntrinsicI); |
2031 | return true; |
2032 | } |
2033 | return false; |
2034 | } |
2035 | |
2036 | if (isa<Constant>(Val: V)) { |
2037 | // If it's not false, and constant, it must evaluate to true. This means our |
2038 | // assume is assume(true), and thus, pointless, and we don't want to do |
2039 | // anything more here. |
2040 | return false; |
2041 | } |
2042 | |
2043 | Constant *True = ConstantInt::getTrue(Context&: V->getContext()); |
2044 | bool Changed = false; |
2045 | |
2046 | for (BasicBlock *Successor : successors(BB: IntrinsicI->getParent())) { |
2047 | BasicBlockEdge Edge(IntrinsicI->getParent(), Successor); |
2048 | |
2049 | // This property is only true in dominated successors, propagateEquality |
2050 | // will check dominance for us. |
2051 | Changed |= propagateEquality(LHS: V, RHS: True, Root: Edge, DominatesByEdge: false); |
2052 | } |
2053 | |
2054 | // We can replace assume value with true, which covers cases like this: |
2055 | // call void @llvm.assume(i1 %cmp) |
2056 | // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true |
2057 | ReplaceOperandsWithMap[V] = True; |
2058 | |
2059 | // Similarly, after assume(!NotV) we know that NotV == false. |
2060 | Value *NotV; |
2061 | if (match(V, P: m_Not(V: m_Value(V&: NotV)))) |
2062 | ReplaceOperandsWithMap[NotV] = ConstantInt::getFalse(Context&: V->getContext()); |
2063 | |
2064 | // If we find an equality fact, canonicalize all dominated uses in this block |
2065 | // to one of the two values. We heuristically choice the "oldest" of the |
2066 | // two where age is determined by value number. (Note that propagateEquality |
2067 | // above handles the cross block case.) |
2068 | // |
2069 | // Key case to cover are: |
2070 | // 1) |
2071 | // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen |
2072 | // call void @llvm.assume(i1 %cmp) |
2073 | // ret float %0 ; will change it to ret float 3.000000e+00 |
2074 | // 2) |
2075 | // %load = load float, float* %addr |
2076 | // %cmp = fcmp oeq float %load, %0 |
2077 | // call void @llvm.assume(i1 %cmp) |
2078 | // ret float %load ; will change it to ret float %0 |
2079 | if (auto *CmpI = dyn_cast<CmpInst>(Val: V)) { |
2080 | if (impliesEquivalanceIfTrue(Cmp: CmpI)) { |
2081 | Value *CmpLHS = CmpI->getOperand(i_nocapture: 0); |
2082 | Value *CmpRHS = CmpI->getOperand(i_nocapture: 1); |
2083 | // Heuristically pick the better replacement -- the choice of heuristic |
2084 | // isn't terribly important here, but the fact we canonicalize on some |
2085 | // replacement is for exposing other simplifications. |
2086 | // TODO: pull this out as a helper function and reuse w/existing |
2087 | // (slightly different) logic. |
2088 | if (isa<Constant>(Val: CmpLHS) && !isa<Constant>(Val: CmpRHS)) |
2089 | std::swap(a&: CmpLHS, b&: CmpRHS); |
2090 | if (!isa<Instruction>(Val: CmpLHS) && isa<Instruction>(Val: CmpRHS)) |
2091 | std::swap(a&: CmpLHS, b&: CmpRHS); |
2092 | if ((isa<Argument>(Val: CmpLHS) && isa<Argument>(Val: CmpRHS)) || |
2093 | (isa<Instruction>(Val: CmpLHS) && isa<Instruction>(Val: CmpRHS))) { |
2094 | // Move the 'oldest' value to the right-hand side, using the value |
2095 | // number as a proxy for age. |
2096 | uint32_t LVN = VN.lookupOrAdd(V: CmpLHS); |
2097 | uint32_t RVN = VN.lookupOrAdd(V: CmpRHS); |
2098 | if (LVN < RVN) |
2099 | std::swap(a&: CmpLHS, b&: CmpRHS); |
2100 | } |
2101 | |
2102 | // Handle degenerate case where we either haven't pruned a dead path or a |
2103 | // removed a trivial assume yet. |
2104 | if (isa<Constant>(Val: CmpLHS) && isa<Constant>(Val: CmpRHS)) |
2105 | return Changed; |
2106 | |
2107 | LLVM_DEBUG(dbgs() << "Replacing dominated uses of " |
2108 | << *CmpLHS << " with " |
2109 | << *CmpRHS << " in block " |
2110 | << IntrinsicI->getParent()->getName() << "\n" ); |
2111 | |
2112 | |
2113 | // Setup the replacement map - this handles uses within the same block |
2114 | if (hasUsersIn(V: CmpLHS, BB: IntrinsicI->getParent())) |
2115 | ReplaceOperandsWithMap[CmpLHS] = CmpRHS; |
2116 | |
2117 | // NOTE: The non-block local cases are handled by the call to |
2118 | // propagateEquality above; this block is just about handling the block |
2119 | // local cases. TODO: There's a bunch of logic in propagateEqualiy which |
2120 | // isn't duplicated for the block local case, can we share it somehow? |
2121 | } |
2122 | } |
2123 | return Changed; |
2124 | } |
2125 | |
2126 | static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { |
2127 | patchReplacementInstruction(I, Repl); |
2128 | I->replaceAllUsesWith(V: Repl); |
2129 | } |
2130 | |
2131 | /// Attempt to eliminate a load, first by eliminating it |
2132 | /// locally, and then attempting non-local elimination if that fails. |
2133 | bool GVNPass::processLoad(LoadInst *L) { |
2134 | if (!MD) |
2135 | return false; |
2136 | |
2137 | // This code hasn't been audited for ordered or volatile memory access |
2138 | if (!L->isUnordered()) |
2139 | return false; |
2140 | |
2141 | if (L->use_empty()) { |
2142 | markInstructionForDeletion(I: L); |
2143 | return true; |
2144 | } |
2145 | |
2146 | // ... to a pointer that has been loaded from before... |
2147 | MemDepResult Dep = MD->getDependency(QueryInst: L); |
2148 | |
2149 | // If it is defined in another block, try harder. |
2150 | if (Dep.isNonLocal()) |
2151 | return processNonLocalLoad(Load: L); |
2152 | |
2153 | // Only handle the local case below |
2154 | if (!Dep.isLocal()) { |
2155 | // This might be a NonFuncLocal or an Unknown |
2156 | LLVM_DEBUG( |
2157 | // fast print dep, using operator<< on instruction is too slow. |
2158 | dbgs() << "GVN: load " ; L->printAsOperand(dbgs()); |
2159 | dbgs() << " has unknown dependence\n" ;); |
2160 | return false; |
2161 | } |
2162 | |
2163 | auto AV = AnalyzeLoadAvailability(Load: L, DepInfo: Dep, Address: L->getPointerOperand()); |
2164 | if (!AV) |
2165 | return false; |
2166 | |
2167 | Value *AvailableValue = AV->MaterializeAdjustedValue(Load: L, InsertPt: L, gvn&: *this); |
2168 | |
2169 | // MaterializeAdjustedValue is responsible for combining metadata. |
2170 | ICF->removeUsersOf(Inst: L); |
2171 | L->replaceAllUsesWith(V: AvailableValue); |
2172 | markInstructionForDeletion(I: L); |
2173 | if (MSSAU) |
2174 | MSSAU->removeMemoryAccess(I: L); |
2175 | ++NumGVNLoad; |
2176 | reportLoadElim(Load: L, AvailableValue, ORE); |
2177 | // Tell MDA to reexamine the reused pointer since we might have more |
2178 | // information after forwarding it. |
2179 | if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy()) |
2180 | MD->invalidateCachedPointerInfo(Ptr: AvailableValue); |
2181 | return true; |
2182 | } |
2183 | |
2184 | /// Return a pair the first field showing the value number of \p Exp and the |
2185 | /// second field showing whether it is a value number newly created. |
2186 | std::pair<uint32_t, bool> |
2187 | GVNPass::ValueTable::assignExpNewValueNum(Expression &Exp) { |
2188 | uint32_t &e = expressionNumbering[Exp]; |
2189 | bool CreateNewValNum = !e; |
2190 | if (CreateNewValNum) { |
2191 | Expressions.push_back(x: Exp); |
2192 | if (ExprIdx.size() < nextValueNumber + 1) |
2193 | ExprIdx.resize(new_size: nextValueNumber * 2); |
2194 | e = nextValueNumber; |
2195 | ExprIdx[nextValueNumber++] = nextExprNumber++; |
2196 | } |
2197 | return {e, CreateNewValNum}; |
2198 | } |
2199 | |
2200 | /// Return whether all the values related with the same \p num are |
2201 | /// defined in \p BB. |
2202 | bool GVNPass::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB, |
2203 | GVNPass &Gvn) { |
2204 | LeaderTableEntry *Vals = &Gvn.LeaderTable[Num]; |
2205 | while (Vals && Vals->BB == BB) |
2206 | Vals = Vals->Next; |
2207 | return !Vals; |
2208 | } |
2209 | |
2210 | /// Wrap phiTranslateImpl to provide caching functionality. |
2211 | uint32_t GVNPass::ValueTable::phiTranslate(const BasicBlock *Pred, |
2212 | const BasicBlock *PhiBlock, |
2213 | uint32_t Num, GVNPass &Gvn) { |
2214 | auto FindRes = PhiTranslateTable.find(Val: {Num, Pred}); |
2215 | if (FindRes != PhiTranslateTable.end()) |
2216 | return FindRes->second; |
2217 | uint32_t NewNum = phiTranslateImpl(BB: Pred, PhiBlock, Num, Gvn); |
2218 | PhiTranslateTable.insert(KV: {{Num, Pred}, NewNum}); |
2219 | return NewNum; |
2220 | } |
2221 | |
2222 | // Return true if the value number \p Num and NewNum have equal value. |
2223 | // Return false if the result is unknown. |
2224 | bool GVNPass::ValueTable::areCallValsEqual(uint32_t Num, uint32_t NewNum, |
2225 | const BasicBlock *Pred, |
2226 | const BasicBlock *PhiBlock, |
2227 | GVNPass &Gvn) { |
2228 | CallInst *Call = nullptr; |
2229 | LeaderTableEntry *Vals = &Gvn.LeaderTable[Num]; |
2230 | while (Vals) { |
2231 | Call = dyn_cast<CallInst>(Val: Vals->Val); |
2232 | if (Call && Call->getParent() == PhiBlock) |
2233 | break; |
2234 | Vals = Vals->Next; |
2235 | } |
2236 | |
2237 | if (AA->doesNotAccessMemory(Call)) |
2238 | return true; |
2239 | |
2240 | if (!MD || !AA->onlyReadsMemory(Call)) |
2241 | return false; |
2242 | |
2243 | MemDepResult local_dep = MD->getDependency(QueryInst: Call); |
2244 | if (!local_dep.isNonLocal()) |
2245 | return false; |
2246 | |
2247 | const MemoryDependenceResults::NonLocalDepInfo &deps = |
2248 | MD->getNonLocalCallDependency(QueryCall: Call); |
2249 | |
2250 | // Check to see if the Call has no function local clobber. |
2251 | for (const NonLocalDepEntry &D : deps) { |
2252 | if (D.getResult().isNonFuncLocal()) |
2253 | return true; |
2254 | } |
2255 | return false; |
2256 | } |
2257 | |
2258 | /// Translate value number \p Num using phis, so that it has the values of |
2259 | /// the phis in BB. |
2260 | uint32_t GVNPass::ValueTable::phiTranslateImpl(const BasicBlock *Pred, |
2261 | const BasicBlock *PhiBlock, |
2262 | uint32_t Num, GVNPass &Gvn) { |
2263 | if (PHINode *PN = NumberingPhi[Num]) { |
2264 | for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) { |
2265 | if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred) |
2266 | if (uint32_t TransVal = lookup(V: PN->getIncomingValue(i), Verify: false)) |
2267 | return TransVal; |
2268 | } |
2269 | return Num; |
2270 | } |
2271 | |
2272 | // If there is any value related with Num is defined in a BB other than |
2273 | // PhiBlock, it cannot depend on a phi in PhiBlock without going through |
2274 | // a backedge. We can do an early exit in that case to save compile time. |
2275 | if (!areAllValsInBB(Num, BB: PhiBlock, Gvn)) |
2276 | return Num; |
2277 | |
2278 | if (Num >= ExprIdx.size() || ExprIdx[Num] == 0) |
2279 | return Num; |
2280 | Expression Exp = Expressions[ExprIdx[Num]]; |
2281 | |
2282 | for (unsigned i = 0; i < Exp.varargs.size(); i++) { |
2283 | // For InsertValue and ExtractValue, some varargs are index numbers |
2284 | // instead of value numbers. Those index numbers should not be |
2285 | // translated. |
2286 | if ((i > 1 && Exp.opcode == Instruction::InsertValue) || |
2287 | (i > 0 && Exp.opcode == Instruction::ExtractValue) || |
2288 | (i > 1 && Exp.opcode == Instruction::ShuffleVector)) |
2289 | continue; |
2290 | Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Num: Exp.varargs[i], Gvn); |
2291 | } |
2292 | |
2293 | if (Exp.commutative) { |
2294 | assert(Exp.varargs.size() >= 2 && "Unsupported commutative instruction!" ); |
2295 | if (Exp.varargs[0] > Exp.varargs[1]) { |
2296 | std::swap(a&: Exp.varargs[0], b&: Exp.varargs[1]); |
2297 | uint32_t Opcode = Exp.opcode >> 8; |
2298 | if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) |
2299 | Exp.opcode = (Opcode << 8) | |
2300 | CmpInst::getSwappedPredicate( |
2301 | pred: static_cast<CmpInst::Predicate>(Exp.opcode & 255)); |
2302 | } |
2303 | } |
2304 | |
2305 | if (uint32_t NewNum = expressionNumbering[Exp]) { |
2306 | if (Exp.opcode == Instruction::Call && NewNum != Num) |
2307 | return areCallValsEqual(Num, NewNum, Pred, PhiBlock, Gvn) ? NewNum : Num; |
2308 | return NewNum; |
2309 | } |
2310 | return Num; |
2311 | } |
2312 | |
2313 | /// Erase stale entry from phiTranslate cache so phiTranslate can be computed |
2314 | /// again. |
2315 | void GVNPass::ValueTable::eraseTranslateCacheEntry( |
2316 | uint32_t Num, const BasicBlock &CurrBlock) { |
2317 | for (const BasicBlock *Pred : predecessors(BB: &CurrBlock)) |
2318 | PhiTranslateTable.erase(Val: {Num, Pred}); |
2319 | } |
2320 | |
2321 | // In order to find a leader for a given value number at a |
2322 | // specific basic block, we first obtain the list of all Values for that number, |
2323 | // and then scan the list to find one whose block dominates the block in |
2324 | // question. This is fast because dominator tree queries consist of only |
2325 | // a few comparisons of DFS numbers. |
2326 | Value *GVNPass::findLeader(const BasicBlock *BB, uint32_t num) { |
2327 | LeaderTableEntry Vals = LeaderTable[num]; |
2328 | if (!Vals.Val) return nullptr; |
2329 | |
2330 | Value *Val = nullptr; |
2331 | if (DT->dominates(A: Vals.BB, B: BB)) { |
2332 | Val = Vals.Val; |
2333 | if (isa<Constant>(Val)) return Val; |
2334 | } |
2335 | |
2336 | LeaderTableEntry* Next = Vals.Next; |
2337 | while (Next) { |
2338 | if (DT->dominates(A: Next->BB, B: BB)) { |
2339 | if (isa<Constant>(Val: Next->Val)) return Next->Val; |
2340 | if (!Val) Val = Next->Val; |
2341 | } |
2342 | |
2343 | Next = Next->Next; |
2344 | } |
2345 | |
2346 | return Val; |
2347 | } |
2348 | |
2349 | /// There is an edge from 'Src' to 'Dst'. Return |
2350 | /// true if every path from the entry block to 'Dst' passes via this edge. In |
2351 | /// particular 'Dst' must not be reachable via another edge from 'Src'. |
2352 | static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, |
2353 | DominatorTree *DT) { |
2354 | // While in theory it is interesting to consider the case in which Dst has |
2355 | // more than one predecessor, because Dst might be part of a loop which is |
2356 | // only reachable from Src, in practice it is pointless since at the time |
2357 | // GVN runs all such loops have preheaders, which means that Dst will have |
2358 | // been changed to have only one predecessor, namely Src. |
2359 | const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); |
2360 | assert((!Pred || Pred == E.getStart()) && |
2361 | "No edge between these basic blocks!" ); |
2362 | return Pred != nullptr; |
2363 | } |
2364 | |
2365 | void GVNPass::assignBlockRPONumber(Function &F) { |
2366 | BlockRPONumber.clear(); |
2367 | uint32_t NextBlockNumber = 1; |
2368 | ReversePostOrderTraversal<Function *> RPOT(&F); |
2369 | for (BasicBlock *BB : RPOT) |
2370 | BlockRPONumber[BB] = NextBlockNumber++; |
2371 | InvalidBlockRPONumbers = false; |
2372 | } |
2373 | |
2374 | bool GVNPass::replaceOperandsForInBlockEquality(Instruction *Instr) const { |
2375 | bool Changed = false; |
2376 | for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) { |
2377 | Value *Operand = Instr->getOperand(i: OpNum); |
2378 | auto it = ReplaceOperandsWithMap.find(Key: Operand); |
2379 | if (it != ReplaceOperandsWithMap.end()) { |
2380 | LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with " |
2381 | << *it->second << " in instruction " << *Instr << '\n'); |
2382 | Instr->setOperand(i: OpNum, Val: it->second); |
2383 | Changed = true; |
2384 | } |
2385 | } |
2386 | return Changed; |
2387 | } |
2388 | |
2389 | /// The given values are known to be equal in every block |
2390 | /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with |
2391 | /// 'RHS' everywhere in the scope. Returns whether a change was made. |
2392 | /// If DominatesByEdge is false, then it means that we will propagate the RHS |
2393 | /// value starting from the end of Root.Start. |
2394 | bool GVNPass::propagateEquality(Value *LHS, Value *RHS, |
2395 | const BasicBlockEdge &Root, |
2396 | bool DominatesByEdge) { |
2397 | SmallVector<std::pair<Value*, Value*>, 4> Worklist; |
2398 | Worklist.push_back(Elt: std::make_pair(x&: LHS, y&: RHS)); |
2399 | bool Changed = false; |
2400 | // For speed, compute a conservative fast approximation to |
2401 | // DT->dominates(Root, Root.getEnd()); |
2402 | const bool RootDominatesEnd = isOnlyReachableViaThisEdge(E: Root, DT); |
2403 | |
2404 | while (!Worklist.empty()) { |
2405 | std::pair<Value*, Value*> Item = Worklist.pop_back_val(); |
2406 | LHS = Item.first; RHS = Item.second; |
2407 | |
2408 | if (LHS == RHS) |
2409 | continue; |
2410 | assert(LHS->getType() == RHS->getType() && "Equality but unequal types!" ); |
2411 | |
2412 | // Don't try to propagate equalities between constants. |
2413 | if (isa<Constant>(Val: LHS) && isa<Constant>(Val: RHS)) |
2414 | continue; |
2415 | |
2416 | // Prefer a constant on the right-hand side, or an Argument if no constants. |
2417 | if (isa<Constant>(Val: LHS) || (isa<Argument>(Val: LHS) && !isa<Constant>(Val: RHS))) |
2418 | std::swap(a&: LHS, b&: RHS); |
2419 | assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!" ); |
2420 | |
2421 | // If there is no obvious reason to prefer the left-hand side over the |
2422 | // right-hand side, ensure the longest lived term is on the right-hand side, |
2423 | // so the shortest lived term will be replaced by the longest lived. |
2424 | // This tends to expose more simplifications. |
2425 | uint32_t LVN = VN.lookupOrAdd(V: LHS); |
2426 | if ((isa<Argument>(Val: LHS) && isa<Argument>(Val: RHS)) || |
2427 | (isa<Instruction>(Val: LHS) && isa<Instruction>(Val: RHS))) { |
2428 | // Move the 'oldest' value to the right-hand side, using the value number |
2429 | // as a proxy for age. |
2430 | uint32_t RVN = VN.lookupOrAdd(V: RHS); |
2431 | if (LVN < RVN) { |
2432 | std::swap(a&: LHS, b&: RHS); |
2433 | LVN = RVN; |
2434 | } |
2435 | } |
2436 | |
2437 | // If value numbering later sees that an instruction in the scope is equal |
2438 | // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve |
2439 | // the invariant that instructions only occur in the leader table for their |
2440 | // own value number (this is used by removeFromLeaderTable), do not do this |
2441 | // if RHS is an instruction (if an instruction in the scope is morphed into |
2442 | // LHS then it will be turned into RHS by the next GVN iteration anyway, so |
2443 | // using the leader table is about compiling faster, not optimizing better). |
2444 | // The leader table only tracks basic blocks, not edges. Only add to if we |
2445 | // have the simple case where the edge dominates the end. |
2446 | if (RootDominatesEnd && !isa<Instruction>(Val: RHS)) |
2447 | addToLeaderTable(N: LVN, V: RHS, BB: Root.getEnd()); |
2448 | |
2449 | // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As |
2450 | // LHS always has at least one use that is not dominated by Root, this will |
2451 | // never do anything if LHS has only one use. |
2452 | if (!LHS->hasOneUse()) { |
2453 | unsigned NumReplacements = |
2454 | DominatesByEdge |
2455 | ? replaceDominatedUsesWith(From: LHS, To: RHS, DT&: *DT, Edge: Root) |
2456 | : replaceDominatedUsesWith(From: LHS, To: RHS, DT&: *DT, BB: Root.getStart()); |
2457 | |
2458 | Changed |= NumReplacements > 0; |
2459 | NumGVNEqProp += NumReplacements; |
2460 | // Cached information for anything that uses LHS will be invalid. |
2461 | if (MD) |
2462 | MD->invalidateCachedPointerInfo(Ptr: LHS); |
2463 | } |
2464 | |
2465 | // Now try to deduce additional equalities from this one. For example, if |
2466 | // the known equality was "(A != B)" == "false" then it follows that A and B |
2467 | // are equal in the scope. Only boolean equalities with an explicit true or |
2468 | // false RHS are currently supported. |
2469 | if (!RHS->getType()->isIntegerTy(Bitwidth: 1)) |
2470 | // Not a boolean equality - bail out. |
2471 | continue; |
2472 | ConstantInt *CI = dyn_cast<ConstantInt>(Val: RHS); |
2473 | if (!CI) |
2474 | // RHS neither 'true' nor 'false' - bail out. |
2475 | continue; |
2476 | // Whether RHS equals 'true'. Otherwise it equals 'false'. |
2477 | bool isKnownTrue = CI->isMinusOne(); |
2478 | bool isKnownFalse = !isKnownTrue; |
2479 | |
2480 | // If "A && B" is known true then both A and B are known true. If "A || B" |
2481 | // is known false then both A and B are known false. |
2482 | Value *A, *B; |
2483 | if ((isKnownTrue && match(V: LHS, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B)))) || |
2484 | (isKnownFalse && match(V: LHS, P: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))))) { |
2485 | Worklist.push_back(Elt: std::make_pair(x&: A, y&: RHS)); |
2486 | Worklist.push_back(Elt: std::make_pair(x&: B, y&: RHS)); |
2487 | continue; |
2488 | } |
2489 | |
2490 | // If we are propagating an equality like "(A == B)" == "true" then also |
2491 | // propagate the equality A == B. When propagating a comparison such as |
2492 | // "(A >= B)" == "true", replace all instances of "A < B" with "false". |
2493 | if (CmpInst *Cmp = dyn_cast<CmpInst>(Val: LHS)) { |
2494 | Value *Op0 = Cmp->getOperand(i_nocapture: 0), *Op1 = Cmp->getOperand(i_nocapture: 1); |
2495 | |
2496 | // If "A == B" is known true, or "A != B" is known false, then replace |
2497 | // A with B everywhere in the scope. For floating point operations, we |
2498 | // have to be careful since equality does not always imply equivalance. |
2499 | if ((isKnownTrue && impliesEquivalanceIfTrue(Cmp)) || |
2500 | (isKnownFalse && impliesEquivalanceIfFalse(Cmp))) |
2501 | Worklist.push_back(Elt: std::make_pair(x&: Op0, y&: Op1)); |
2502 | |
2503 | // If "A >= B" is known true, replace "A < B" with false everywhere. |
2504 | CmpInst::Predicate NotPred = Cmp->getInversePredicate(); |
2505 | Constant *NotVal = ConstantInt::get(Ty: Cmp->getType(), V: isKnownFalse); |
2506 | // Since we don't have the instruction "A < B" immediately to hand, work |
2507 | // out the value number that it would have and use that to find an |
2508 | // appropriate instruction (if any). |
2509 | uint32_t NextNum = VN.getNextUnusedValueNumber(); |
2510 | uint32_t Num = VN.lookupOrAddCmp(Opcode: Cmp->getOpcode(), Predicate: NotPred, LHS: Op0, RHS: Op1); |
2511 | // If the number we were assigned was brand new then there is no point in |
2512 | // looking for an instruction realizing it: there cannot be one! |
2513 | if (Num < NextNum) { |
2514 | Value *NotCmp = findLeader(BB: Root.getEnd(), num: Num); |
2515 | if (NotCmp && isa<Instruction>(Val: NotCmp)) { |
2516 | unsigned NumReplacements = |
2517 | DominatesByEdge |
2518 | ? replaceDominatedUsesWith(From: NotCmp, To: NotVal, DT&: *DT, Edge: Root) |
2519 | : replaceDominatedUsesWith(From: NotCmp, To: NotVal, DT&: *DT, |
2520 | BB: Root.getStart()); |
2521 | Changed |= NumReplacements > 0; |
2522 | NumGVNEqProp += NumReplacements; |
2523 | // Cached information for anything that uses NotCmp will be invalid. |
2524 | if (MD) |
2525 | MD->invalidateCachedPointerInfo(Ptr: NotCmp); |
2526 | } |
2527 | } |
2528 | // Ensure that any instruction in scope that gets the "A < B" value number |
2529 | // is replaced with false. |
2530 | // The leader table only tracks basic blocks, not edges. Only add to if we |
2531 | // have the simple case where the edge dominates the end. |
2532 | if (RootDominatesEnd) |
2533 | addToLeaderTable(N: Num, V: NotVal, BB: Root.getEnd()); |
2534 | |
2535 | continue; |
2536 | } |
2537 | } |
2538 | |
2539 | return Changed; |
2540 | } |
2541 | |
2542 | /// When calculating availability, handle an instruction |
2543 | /// by inserting it into the appropriate sets |
2544 | bool GVNPass::processInstruction(Instruction *I) { |
2545 | // Ignore dbg info intrinsics. |
2546 | if (isa<DbgInfoIntrinsic>(Val: I)) |
2547 | return false; |
2548 | |
2549 | // If the instruction can be easily simplified then do so now in preference |
2550 | // to value numbering it. Value numbering often exposes redundancies, for |
2551 | // example if it determines that %y is equal to %x then the instruction |
2552 | // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. |
2553 | const DataLayout &DL = I->getModule()->getDataLayout(); |
2554 | if (Value *V = simplifyInstruction(I, Q: {DL, TLI, DT, AC})) { |
2555 | bool Changed = false; |
2556 | if (!I->use_empty()) { |
2557 | // Simplification can cause a special instruction to become not special. |
2558 | // For example, devirtualization to a willreturn function. |
2559 | ICF->removeUsersOf(Inst: I); |
2560 | I->replaceAllUsesWith(V); |
2561 | Changed = true; |
2562 | } |
2563 | if (isInstructionTriviallyDead(I, TLI)) { |
2564 | markInstructionForDeletion(I); |
2565 | Changed = true; |
2566 | } |
2567 | if (Changed) { |
2568 | if (MD && V->getType()->isPtrOrPtrVectorTy()) |
2569 | MD->invalidateCachedPointerInfo(Ptr: V); |
2570 | ++NumGVNSimpl; |
2571 | return true; |
2572 | } |
2573 | } |
2574 | |
2575 | if (auto *Assume = dyn_cast<AssumeInst>(Val: I)) |
2576 | return processAssumeIntrinsic(IntrinsicI: Assume); |
2577 | |
2578 | if (LoadInst *Load = dyn_cast<LoadInst>(Val: I)) { |
2579 | if (processLoad(L: Load)) |
2580 | return true; |
2581 | |
2582 | unsigned Num = VN.lookupOrAdd(V: Load); |
2583 | addToLeaderTable(N: Num, V: Load, BB: Load->getParent()); |
2584 | return false; |
2585 | } |
2586 | |
2587 | // For conditional branches, we can perform simple conditional propagation on |
2588 | // the condition value itself. |
2589 | if (BranchInst *BI = dyn_cast<BranchInst>(Val: I)) { |
2590 | if (!BI->isConditional()) |
2591 | return false; |
2592 | |
2593 | if (isa<Constant>(Val: BI->getCondition())) |
2594 | return processFoldableCondBr(BI); |
2595 | |
2596 | Value *BranchCond = BI->getCondition(); |
2597 | BasicBlock *TrueSucc = BI->getSuccessor(i: 0); |
2598 | BasicBlock *FalseSucc = BI->getSuccessor(i: 1); |
2599 | // Avoid multiple edges early. |
2600 | if (TrueSucc == FalseSucc) |
2601 | return false; |
2602 | |
2603 | BasicBlock *Parent = BI->getParent(); |
2604 | bool Changed = false; |
2605 | |
2606 | Value *TrueVal = ConstantInt::getTrue(Context&: TrueSucc->getContext()); |
2607 | BasicBlockEdge TrueE(Parent, TrueSucc); |
2608 | Changed |= propagateEquality(LHS: BranchCond, RHS: TrueVal, Root: TrueE, DominatesByEdge: true); |
2609 | |
2610 | Value *FalseVal = ConstantInt::getFalse(Context&: FalseSucc->getContext()); |
2611 | BasicBlockEdge FalseE(Parent, FalseSucc); |
2612 | Changed |= propagateEquality(LHS: BranchCond, RHS: FalseVal, Root: FalseE, DominatesByEdge: true); |
2613 | |
2614 | return Changed; |
2615 | } |
2616 | |
2617 | // For switches, propagate the case values into the case destinations. |
2618 | if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: I)) { |
2619 | Value *SwitchCond = SI->getCondition(); |
2620 | BasicBlock *Parent = SI->getParent(); |
2621 | bool Changed = false; |
2622 | |
2623 | // Remember how many outgoing edges there are to every successor. |
2624 | SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; |
2625 | for (BasicBlock *Succ : successors(BB: Parent)) |
2626 | ++SwitchEdges[Succ]; |
2627 | |
2628 | for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); |
2629 | i != e; ++i) { |
2630 | BasicBlock *Dst = i->getCaseSuccessor(); |
2631 | // If there is only a single edge, propagate the case value into it. |
2632 | if (SwitchEdges.lookup(Val: Dst) == 1) { |
2633 | BasicBlockEdge E(Parent, Dst); |
2634 | Changed |= propagateEquality(LHS: SwitchCond, RHS: i->getCaseValue(), Root: E, DominatesByEdge: true); |
2635 | } |
2636 | } |
2637 | return Changed; |
2638 | } |
2639 | |
2640 | // Instructions with void type don't return a value, so there's |
2641 | // no point in trying to find redundancies in them. |
2642 | if (I->getType()->isVoidTy()) |
2643 | return false; |
2644 | |
2645 | uint32_t NextNum = VN.getNextUnusedValueNumber(); |
2646 | unsigned Num = VN.lookupOrAdd(V: I); |
2647 | |
2648 | // Allocations are always uniquely numbered, so we can save time and memory |
2649 | // by fast failing them. |
2650 | if (isa<AllocaInst>(Val: I) || I->isTerminator() || isa<PHINode>(Val: I)) { |
2651 | addToLeaderTable(N: Num, V: I, BB: I->getParent()); |
2652 | return false; |
2653 | } |
2654 | |
2655 | // If the number we were assigned was a brand new VN, then we don't |
2656 | // need to do a lookup to see if the number already exists |
2657 | // somewhere in the domtree: it can't! |
2658 | if (Num >= NextNum) { |
2659 | addToLeaderTable(N: Num, V: I, BB: I->getParent()); |
2660 | return false; |
2661 | } |
2662 | |
2663 | // Perform fast-path value-number based elimination of values inherited from |
2664 | // dominators. |
2665 | Value *Repl = findLeader(BB: I->getParent(), num: Num); |
2666 | if (!Repl) { |
2667 | // Failure, just remember this instance for future use. |
2668 | addToLeaderTable(N: Num, V: I, BB: I->getParent()); |
2669 | return false; |
2670 | } |
2671 | |
2672 | if (Repl == I) { |
2673 | // If I was the result of a shortcut PRE, it might already be in the table |
2674 | // and the best replacement for itself. Nothing to do. |
2675 | return false; |
2676 | } |
2677 | |
2678 | // Remove it! |
2679 | patchAndReplaceAllUsesWith(I, Repl); |
2680 | if (MD && Repl->getType()->isPtrOrPtrVectorTy()) |
2681 | MD->invalidateCachedPointerInfo(Ptr: Repl); |
2682 | markInstructionForDeletion(I); |
2683 | return true; |
2684 | } |
2685 | |
2686 | /// runOnFunction - This is the main transformation entry point for a function. |
2687 | bool GVNPass::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT, |
2688 | const TargetLibraryInfo &RunTLI, AAResults &RunAA, |
2689 | MemoryDependenceResults *RunMD, LoopInfo &LI, |
2690 | OptimizationRemarkEmitter *RunORE, MemorySSA *MSSA) { |
2691 | AC = &RunAC; |
2692 | DT = &RunDT; |
2693 | VN.setDomTree(DT); |
2694 | TLI = &RunTLI; |
2695 | VN.setAliasAnalysis(&RunAA); |
2696 | MD = RunMD; |
2697 | ImplicitControlFlowTracking ImplicitCFT; |
2698 | ICF = &ImplicitCFT; |
2699 | this->LI = &LI; |
2700 | VN.setMemDep(MD); |
2701 | ORE = RunORE; |
2702 | InvalidBlockRPONumbers = true; |
2703 | MemorySSAUpdater Updater(MSSA); |
2704 | MSSAU = MSSA ? &Updater : nullptr; |
2705 | |
2706 | bool Changed = false; |
2707 | bool ShouldContinue = true; |
2708 | |
2709 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); |
2710 | // Merge unconditional branches, allowing PRE to catch more |
2711 | // optimization opportunities. |
2712 | for (BasicBlock &BB : llvm::make_early_inc_range(Range&: F)) { |
2713 | bool removedBlock = MergeBlockIntoPredecessor(BB: &BB, DTU: &DTU, LI: &LI, MSSAU, MemDep: MD); |
2714 | if (removedBlock) |
2715 | ++NumGVNBlocks; |
2716 | |
2717 | Changed |= removedBlock; |
2718 | } |
2719 | |
2720 | unsigned Iteration = 0; |
2721 | while (ShouldContinue) { |
2722 | LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n" ); |
2723 | (void) Iteration; |
2724 | ShouldContinue = iterateOnFunction(F); |
2725 | Changed |= ShouldContinue; |
2726 | ++Iteration; |
2727 | } |
2728 | |
2729 | if (isPREEnabled()) { |
2730 | // Fabricate val-num for dead-code in order to suppress assertion in |
2731 | // performPRE(). |
2732 | assignValNumForDeadCode(); |
2733 | bool PREChanged = true; |
2734 | while (PREChanged) { |
2735 | PREChanged = performPRE(F); |
2736 | Changed |= PREChanged; |
2737 | } |
2738 | } |
2739 | |
2740 | // FIXME: Should perform GVN again after PRE does something. PRE can move |
2741 | // computations into blocks where they become fully redundant. Note that |
2742 | // we can't do this until PRE's critical edge splitting updates memdep. |
2743 | // Actually, when this happens, we should just fully integrate PRE into GVN. |
2744 | |
2745 | cleanupGlobalSets(); |
2746 | // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each |
2747 | // iteration. |
2748 | DeadBlocks.clear(); |
2749 | |
2750 | if (MSSA && VerifyMemorySSA) |
2751 | MSSA->verifyMemorySSA(); |
2752 | |
2753 | return Changed; |
2754 | } |
2755 | |
2756 | bool GVNPass::processBlock(BasicBlock *BB) { |
2757 | // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function |
2758 | // (and incrementing BI before processing an instruction). |
2759 | assert(InstrsToErase.empty() && |
2760 | "We expect InstrsToErase to be empty across iterations" ); |
2761 | if (DeadBlocks.count(key: BB)) |
2762 | return false; |
2763 | |
2764 | // Clearing map before every BB because it can be used only for single BB. |
2765 | ReplaceOperandsWithMap.clear(); |
2766 | bool ChangedFunction = false; |
2767 | |
2768 | // Since we may not have visited the input blocks of the phis, we can't |
2769 | // use our normal hash approach for phis. Instead, simply look for |
2770 | // obvious duplicates. The first pass of GVN will tend to create |
2771 | // identical phis, and the second or later passes can eliminate them. |
2772 | SmallPtrSet<PHINode *, 8> PHINodesToRemove; |
2773 | ChangedFunction |= EliminateDuplicatePHINodes(BB, ToRemove&: PHINodesToRemove); |
2774 | for (PHINode *PN : PHINodesToRemove) { |
2775 | VN.erase(V: PN); |
2776 | removeInstruction(I: PN); |
2777 | } |
2778 | |
2779 | for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); |
2780 | BI != BE;) { |
2781 | if (!ReplaceOperandsWithMap.empty()) |
2782 | ChangedFunction |= replaceOperandsForInBlockEquality(Instr: &*BI); |
2783 | ChangedFunction |= processInstruction(I: &*BI); |
2784 | |
2785 | if (InstrsToErase.empty()) { |
2786 | ++BI; |
2787 | continue; |
2788 | } |
2789 | |
2790 | // If we need some instructions deleted, do it now. |
2791 | NumGVNInstr += InstrsToErase.size(); |
2792 | |
2793 | // Avoid iterator invalidation. |
2794 | bool AtStart = BI == BB->begin(); |
2795 | if (!AtStart) |
2796 | --BI; |
2797 | |
2798 | for (auto *I : InstrsToErase) { |
2799 | assert(I->getParent() == BB && "Removing instruction from wrong block?" ); |
2800 | LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n'); |
2801 | salvageKnowledge(I, AC); |
2802 | salvageDebugInfo(I&: *I); |
2803 | removeInstruction(I); |
2804 | } |
2805 | InstrsToErase.clear(); |
2806 | |
2807 | if (AtStart) |
2808 | BI = BB->begin(); |
2809 | else |
2810 | ++BI; |
2811 | } |
2812 | |
2813 | return ChangedFunction; |
2814 | } |
2815 | |
2816 | // Instantiate an expression in a predecessor that lacked it. |
2817 | bool GVNPass::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred, |
2818 | BasicBlock *Curr, unsigned int ValNo) { |
2819 | // Because we are going top-down through the block, all value numbers |
2820 | // will be available in the predecessor by the time we need them. Any |
2821 | // that weren't originally present will have been instantiated earlier |
2822 | // in this loop. |
2823 | bool success = true; |
2824 | for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) { |
2825 | Value *Op = Instr->getOperand(i); |
2826 | if (isa<Argument>(Val: Op) || isa<Constant>(Val: Op) || isa<GlobalValue>(Val: Op)) |
2827 | continue; |
2828 | // This could be a newly inserted instruction, in which case, we won't |
2829 | // find a value number, and should give up before we hurt ourselves. |
2830 | // FIXME: Rewrite the infrastructure to let it easier to value number |
2831 | // and process newly inserted instructions. |
2832 | if (!VN.exists(V: Op)) { |
2833 | success = false; |
2834 | break; |
2835 | } |
2836 | uint32_t TValNo = |
2837 | VN.phiTranslate(Pred, PhiBlock: Curr, Num: VN.lookup(V: Op), Gvn&: *this); |
2838 | if (Value *V = findLeader(BB: Pred, num: TValNo)) { |
2839 | Instr->setOperand(i, Val: V); |
2840 | } else { |
2841 | success = false; |
2842 | break; |
2843 | } |
2844 | } |
2845 | |
2846 | // Fail out if we encounter an operand that is not available in |
2847 | // the PRE predecessor. This is typically because of loads which |
2848 | // are not value numbered precisely. |
2849 | if (!success) |
2850 | return false; |
2851 | |
2852 | Instr->insertBefore(InsertPos: Pred->getTerminator()); |
2853 | Instr->setName(Instr->getName() + ".pre" ); |
2854 | Instr->setDebugLoc(Instr->getDebugLoc()); |
2855 | |
2856 | ICF->insertInstructionTo(Inst: Instr, BB: Pred); |
2857 | |
2858 | unsigned Num = VN.lookupOrAdd(V: Instr); |
2859 | VN.add(V: Instr, num: Num); |
2860 | |
2861 | // Update the availability map to include the new instruction. |
2862 | addToLeaderTable(N: Num, V: Instr, BB: Pred); |
2863 | return true; |
2864 | } |
2865 | |
2866 | bool GVNPass::performScalarPRE(Instruction *CurInst) { |
2867 | if (isa<AllocaInst>(Val: CurInst) || CurInst->isTerminator() || |
2868 | isa<PHINode>(Val: CurInst) || CurInst->getType()->isVoidTy() || |
2869 | CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || |
2870 | isa<DbgInfoIntrinsic>(Val: CurInst)) |
2871 | return false; |
2872 | |
2873 | // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from |
2874 | // sinking the compare again, and it would force the code generator to |
2875 | // move the i1 from processor flags or predicate registers into a general |
2876 | // purpose register. |
2877 | if (isa<CmpInst>(Val: CurInst)) |
2878 | return false; |
2879 | |
2880 | // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from |
2881 | // sinking the addressing mode computation back to its uses. Extending the |
2882 | // GEP's live range increases the register pressure, and therefore it can |
2883 | // introduce unnecessary spills. |
2884 | // |
2885 | // This doesn't prevent Load PRE. PHI translation will make the GEP available |
2886 | // to the load by moving it to the predecessor block if necessary. |
2887 | if (isa<GetElementPtrInst>(Val: CurInst)) |
2888 | return false; |
2889 | |
2890 | if (auto *CallB = dyn_cast<CallBase>(Val: CurInst)) { |
2891 | // We don't currently value number ANY inline asm calls. |
2892 | if (CallB->isInlineAsm()) |
2893 | return false; |
2894 | } |
2895 | |
2896 | uint32_t ValNo = VN.lookup(V: CurInst); |
2897 | |
2898 | // Look for the predecessors for PRE opportunities. We're |
2899 | // only trying to solve the basic diamond case, where |
2900 | // a value is computed in the successor and one predecessor, |
2901 | // but not the other. We also explicitly disallow cases |
2902 | // where the successor is its own predecessor, because they're |
2903 | // more complicated to get right. |
2904 | unsigned NumWith = 0; |
2905 | unsigned NumWithout = 0; |
2906 | BasicBlock *PREPred = nullptr; |
2907 | BasicBlock *CurrentBlock = CurInst->getParent(); |
2908 | |
2909 | // Update the RPO numbers for this function. |
2910 | if (InvalidBlockRPONumbers) |
2911 | assignBlockRPONumber(F&: *CurrentBlock->getParent()); |
2912 | |
2913 | SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap; |
2914 | for (BasicBlock *P : predecessors(BB: CurrentBlock)) { |
2915 | // We're not interested in PRE where blocks with predecessors that are |
2916 | // not reachable. |
2917 | if (!DT->isReachableFromEntry(A: P)) { |
2918 | NumWithout = 2; |
2919 | break; |
2920 | } |
2921 | // It is not safe to do PRE when P->CurrentBlock is a loop backedge. |
2922 | assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && |
2923 | "Invalid BlockRPONumber map." ); |
2924 | if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock]) { |
2925 | NumWithout = 2; |
2926 | break; |
2927 | } |
2928 | |
2929 | uint32_t TValNo = VN.phiTranslate(Pred: P, PhiBlock: CurrentBlock, Num: ValNo, Gvn&: *this); |
2930 | Value *predV = findLeader(BB: P, num: TValNo); |
2931 | if (!predV) { |
2932 | predMap.push_back(Elt: std::make_pair(x: static_cast<Value *>(nullptr), y&: P)); |
2933 | PREPred = P; |
2934 | ++NumWithout; |
2935 | } else if (predV == CurInst) { |
2936 | /* CurInst dominates this predecessor. */ |
2937 | NumWithout = 2; |
2938 | break; |
2939 | } else { |
2940 | predMap.push_back(Elt: std::make_pair(x&: predV, y&: P)); |
2941 | ++NumWith; |
2942 | } |
2943 | } |
2944 | |
2945 | // Don't do PRE when it might increase code size, i.e. when |
2946 | // we would need to insert instructions in more than one pred. |
2947 | if (NumWithout > 1 || NumWith == 0) |
2948 | return false; |
2949 | |
2950 | // We may have a case where all predecessors have the instruction, |
2951 | // and we just need to insert a phi node. Otherwise, perform |
2952 | // insertion. |
2953 | Instruction *PREInstr = nullptr; |
2954 | |
2955 | if (NumWithout != 0) { |
2956 | if (!isSafeToSpeculativelyExecute(I: CurInst)) { |
2957 | // It is only valid to insert a new instruction if the current instruction |
2958 | // is always executed. An instruction with implicit control flow could |
2959 | // prevent us from doing it. If we cannot speculate the execution, then |
2960 | // PRE should be prohibited. |
2961 | if (ICF->isDominatedByICFIFromSameBlock(Insn: CurInst)) |
2962 | return false; |
2963 | } |
2964 | |
2965 | // Don't do PRE across indirect branch. |
2966 | if (isa<IndirectBrInst>(Val: PREPred->getTerminator())) |
2967 | return false; |
2968 | |
2969 | // We can't do PRE safely on a critical edge, so instead we schedule |
2970 | // the edge to be split and perform the PRE the next time we iterate |
2971 | // on the function. |
2972 | unsigned SuccNum = GetSuccessorNumber(BB: PREPred, Succ: CurrentBlock); |
2973 | if (isCriticalEdge(TI: PREPred->getTerminator(), SuccNum)) { |
2974 | toSplit.push_back(Elt: std::make_pair(x: PREPred->getTerminator(), y&: SuccNum)); |
2975 | return false; |
2976 | } |
2977 | // We need to insert somewhere, so let's give it a shot |
2978 | PREInstr = CurInst->clone(); |
2979 | if (!performScalarPREInsertion(Instr: PREInstr, Pred: PREPred, Curr: CurrentBlock, ValNo)) { |
2980 | // If we failed insertion, make sure we remove the instruction. |
2981 | #ifndef NDEBUG |
2982 | verifyRemoved(I: PREInstr); |
2983 | #endif |
2984 | PREInstr->deleteValue(); |
2985 | return false; |
2986 | } |
2987 | } |
2988 | |
2989 | // Either we should have filled in the PRE instruction, or we should |
2990 | // not have needed insertions. |
2991 | assert(PREInstr != nullptr || NumWithout == 0); |
2992 | |
2993 | ++NumGVNPRE; |
2994 | |
2995 | // Create a PHI to make the value available in this block. |
2996 | PHINode *Phi = PHINode::Create(Ty: CurInst->getType(), NumReservedValues: predMap.size(), |
2997 | NameStr: CurInst->getName() + ".pre-phi" ); |
2998 | Phi->insertBefore(InsertPos: CurrentBlock->begin()); |
2999 | for (unsigned i = 0, e = predMap.size(); i != e; ++i) { |
3000 | if (Value *V = predMap[i].first) { |
3001 | // If we use an existing value in this phi, we have to patch the original |
3002 | // value because the phi will be used to replace a later value. |
3003 | patchReplacementInstruction(I: CurInst, Repl: V); |
3004 | Phi->addIncoming(V, BB: predMap[i].second); |
3005 | } else |
3006 | Phi->addIncoming(V: PREInstr, BB: PREPred); |
3007 | } |
3008 | |
3009 | VN.add(V: Phi, num: ValNo); |
3010 | // After creating a new PHI for ValNo, the phi translate result for ValNo will |
3011 | // be changed, so erase the related stale entries in phi translate cache. |
3012 | VN.eraseTranslateCacheEntry(Num: ValNo, CurrBlock: *CurrentBlock); |
3013 | addToLeaderTable(N: ValNo, V: Phi, BB: CurrentBlock); |
3014 | Phi->setDebugLoc(CurInst->getDebugLoc()); |
3015 | CurInst->replaceAllUsesWith(V: Phi); |
3016 | if (MD && Phi->getType()->isPtrOrPtrVectorTy()) |
3017 | MD->invalidateCachedPointerInfo(Ptr: Phi); |
3018 | VN.erase(V: CurInst); |
3019 | removeFromLeaderTable(N: ValNo, I: CurInst, BB: CurrentBlock); |
3020 | |
3021 | LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); |
3022 | removeInstruction(I: CurInst); |
3023 | ++NumGVNInstr; |
3024 | |
3025 | return true; |
3026 | } |
3027 | |
3028 | /// Perform a purely local form of PRE that looks for diamond |
3029 | /// control flow patterns and attempts to perform simple PRE at the join point. |
3030 | bool GVNPass::performPRE(Function &F) { |
3031 | bool Changed = false; |
3032 | for (BasicBlock *CurrentBlock : depth_first(G: &F.getEntryBlock())) { |
3033 | // Nothing to PRE in the entry block. |
3034 | if (CurrentBlock == &F.getEntryBlock()) |
3035 | continue; |
3036 | |
3037 | // Don't perform PRE on an EH pad. |
3038 | if (CurrentBlock->isEHPad()) |
3039 | continue; |
3040 | |
3041 | for (BasicBlock::iterator BI = CurrentBlock->begin(), |
3042 | BE = CurrentBlock->end(); |
3043 | BI != BE;) { |
3044 | Instruction *CurInst = &*BI++; |
3045 | Changed |= performScalarPRE(CurInst); |
3046 | } |
3047 | } |
3048 | |
3049 | if (splitCriticalEdges()) |
3050 | Changed = true; |
3051 | |
3052 | return Changed; |
3053 | } |
3054 | |
3055 | /// Split the critical edge connecting the given two blocks, and return |
3056 | /// the block inserted to the critical edge. |
3057 | BasicBlock *GVNPass::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) { |
3058 | // GVN does not require loop-simplify, do not try to preserve it if it is not |
3059 | // possible. |
3060 | BasicBlock *BB = SplitCriticalEdge( |
3061 | Src: Pred, Dst: Succ, |
3062 | Options: CriticalEdgeSplittingOptions(DT, LI, MSSAU).unsetPreserveLoopSimplify()); |
3063 | if (BB) { |
3064 | if (MD) |
3065 | MD->invalidateCachedPredecessors(); |
3066 | InvalidBlockRPONumbers = true; |
3067 | } |
3068 | return BB; |
3069 | } |
3070 | |
3071 | /// Split critical edges found during the previous |
3072 | /// iteration that may enable further optimization. |
3073 | bool GVNPass::splitCriticalEdges() { |
3074 | if (toSplit.empty()) |
3075 | return false; |
3076 | |
3077 | bool Changed = false; |
3078 | do { |
3079 | std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val(); |
3080 | Changed |= SplitCriticalEdge(TI: Edge.first, SuccNum: Edge.second, |
3081 | Options: CriticalEdgeSplittingOptions(DT, LI, MSSAU)) != |
3082 | nullptr; |
3083 | } while (!toSplit.empty()); |
3084 | if (Changed) { |
3085 | if (MD) |
3086 | MD->invalidateCachedPredecessors(); |
3087 | InvalidBlockRPONumbers = true; |
3088 | } |
3089 | return Changed; |
3090 | } |
3091 | |
3092 | /// Executes one iteration of GVN |
3093 | bool GVNPass::iterateOnFunction(Function &F) { |
3094 | cleanupGlobalSets(); |
3095 | |
3096 | // Top-down walk of the dominator tree |
3097 | bool Changed = false; |
3098 | // Needed for value numbering with phi construction to work. |
3099 | // RPOT walks the graph in its constructor and will not be invalidated during |
3100 | // processBlock. |
3101 | ReversePostOrderTraversal<Function *> RPOT(&F); |
3102 | |
3103 | for (BasicBlock *BB : RPOT) |
3104 | Changed |= processBlock(BB); |
3105 | |
3106 | return Changed; |
3107 | } |
3108 | |
3109 | void GVNPass::cleanupGlobalSets() { |
3110 | VN.clear(); |
3111 | LeaderTable.clear(); |
3112 | BlockRPONumber.clear(); |
3113 | TableAllocator.Reset(); |
3114 | ICF->clear(); |
3115 | InvalidBlockRPONumbers = true; |
3116 | } |
3117 | |
3118 | void GVNPass::removeInstruction(Instruction *I) { |
3119 | if (MD) MD->removeInstruction(InstToRemove: I); |
3120 | if (MSSAU) |
3121 | MSSAU->removeMemoryAccess(I); |
3122 | #ifndef NDEBUG |
3123 | verifyRemoved(I); |
3124 | #endif |
3125 | ICF->removeInstruction(Inst: I); |
3126 | I->eraseFromParent(); |
3127 | } |
3128 | |
3129 | /// Verify that the specified instruction does not occur in our |
3130 | /// internal data structures. |
3131 | void GVNPass::verifyRemoved(const Instruction *Inst) const { |
3132 | VN.verifyRemoved(V: Inst); |
3133 | |
3134 | // Walk through the value number scope to make sure the instruction isn't |
3135 | // ferreted away in it. |
3136 | for (const auto &I : LeaderTable) { |
3137 | const LeaderTableEntry *Node = &I.second; |
3138 | assert(Node->Val != Inst && "Inst still in value numbering scope!" ); |
3139 | |
3140 | while (Node->Next) { |
3141 | Node = Node->Next; |
3142 | assert(Node->Val != Inst && "Inst still in value numbering scope!" ); |
3143 | } |
3144 | } |
3145 | } |
3146 | |
3147 | /// BB is declared dead, which implied other blocks become dead as well. This |
3148 | /// function is to add all these blocks to "DeadBlocks". For the dead blocks' |
3149 | /// live successors, update their phi nodes by replacing the operands |
3150 | /// corresponding to dead blocks with UndefVal. |
3151 | void GVNPass::addDeadBlock(BasicBlock *BB) { |
3152 | SmallVector<BasicBlock *, 4> NewDead; |
3153 | SmallSetVector<BasicBlock *, 4> DF; |
3154 | |
3155 | NewDead.push_back(Elt: BB); |
3156 | while (!NewDead.empty()) { |
3157 | BasicBlock *D = NewDead.pop_back_val(); |
3158 | if (DeadBlocks.count(key: D)) |
3159 | continue; |
3160 | |
3161 | // All blocks dominated by D are dead. |
3162 | SmallVector<BasicBlock *, 8> Dom; |
3163 | DT->getDescendants(R: D, Result&: Dom); |
3164 | DeadBlocks.insert(Start: Dom.begin(), End: Dom.end()); |
3165 | |
3166 | // Figure out the dominance-frontier(D). |
3167 | for (BasicBlock *B : Dom) { |
3168 | for (BasicBlock *S : successors(BB: B)) { |
3169 | if (DeadBlocks.count(key: S)) |
3170 | continue; |
3171 | |
3172 | bool AllPredDead = true; |
3173 | for (BasicBlock *P : predecessors(BB: S)) |
3174 | if (!DeadBlocks.count(key: P)) { |
3175 | AllPredDead = false; |
3176 | break; |
3177 | } |
3178 | |
3179 | if (!AllPredDead) { |
3180 | // S could be proved dead later on. That is why we don't update phi |
3181 | // operands at this moment. |
3182 | DF.insert(X: S); |
3183 | } else { |
3184 | // While S is not dominated by D, it is dead by now. This could take |
3185 | // place if S already have a dead predecessor before D is declared |
3186 | // dead. |
3187 | NewDead.push_back(Elt: S); |
3188 | } |
3189 | } |
3190 | } |
3191 | } |
3192 | |
3193 | // For the dead blocks' live successors, update their phi nodes by replacing |
3194 | // the operands corresponding to dead blocks with UndefVal. |
3195 | for (BasicBlock *B : DF) { |
3196 | if (DeadBlocks.count(key: B)) |
3197 | continue; |
3198 | |
3199 | // First, split the critical edges. This might also create additional blocks |
3200 | // to preserve LoopSimplify form and adjust edges accordingly. |
3201 | SmallVector<BasicBlock *, 4> Preds(predecessors(BB: B)); |
3202 | for (BasicBlock *P : Preds) { |
3203 | if (!DeadBlocks.count(key: P)) |
3204 | continue; |
3205 | |
3206 | if (llvm::is_contained(Range: successors(BB: P), Element: B) && |
3207 | isCriticalEdge(TI: P->getTerminator(), Succ: B)) { |
3208 | if (BasicBlock *S = splitCriticalEdges(Pred: P, Succ: B)) |
3209 | DeadBlocks.insert(X: P = S); |
3210 | } |
3211 | } |
3212 | |
3213 | // Now poison the incoming values from the dead predecessors. |
3214 | for (BasicBlock *P : predecessors(BB: B)) { |
3215 | if (!DeadBlocks.count(key: P)) |
3216 | continue; |
3217 | for (PHINode &Phi : B->phis()) { |
3218 | Phi.setIncomingValueForBlock(BB: P, V: PoisonValue::get(T: Phi.getType())); |
3219 | if (MD) |
3220 | MD->invalidateCachedPointerInfo(Ptr: &Phi); |
3221 | } |
3222 | } |
3223 | } |
3224 | } |
3225 | |
3226 | // If the given branch is recognized as a foldable branch (i.e. conditional |
3227 | // branch with constant condition), it will perform following analyses and |
3228 | // transformation. |
3229 | // 1) If the dead out-coming edge is a critical-edge, split it. Let |
3230 | // R be the target of the dead out-coming edge. |
3231 | // 1) Identify the set of dead blocks implied by the branch's dead outcoming |
3232 | // edge. The result of this step will be {X| X is dominated by R} |
3233 | // 2) Identify those blocks which haves at least one dead predecessor. The |
3234 | // result of this step will be dominance-frontier(R). |
3235 | // 3) Update the PHIs in DF(R) by replacing the operands corresponding to |
3236 | // dead blocks with "UndefVal" in an hope these PHIs will optimized away. |
3237 | // |
3238 | // Return true iff *NEW* dead code are found. |
3239 | bool GVNPass::processFoldableCondBr(BranchInst *BI) { |
3240 | if (!BI || BI->isUnconditional()) |
3241 | return false; |
3242 | |
3243 | // If a branch has two identical successors, we cannot declare either dead. |
3244 | if (BI->getSuccessor(i: 0) == BI->getSuccessor(i: 1)) |
3245 | return false; |
3246 | |
3247 | ConstantInt *Cond = dyn_cast<ConstantInt>(Val: BI->getCondition()); |
3248 | if (!Cond) |
3249 | return false; |
3250 | |
3251 | BasicBlock *DeadRoot = |
3252 | Cond->getZExtValue() ? BI->getSuccessor(i: 1) : BI->getSuccessor(i: 0); |
3253 | if (DeadBlocks.count(key: DeadRoot)) |
3254 | return false; |
3255 | |
3256 | if (!DeadRoot->getSinglePredecessor()) |
3257 | DeadRoot = splitCriticalEdges(Pred: BI->getParent(), Succ: DeadRoot); |
3258 | |
3259 | addDeadBlock(BB: DeadRoot); |
3260 | return true; |
3261 | } |
3262 | |
3263 | // performPRE() will trigger assert if it comes across an instruction without |
3264 | // associated val-num. As it normally has far more live instructions than dead |
3265 | // instructions, it makes more sense just to "fabricate" a val-number for the |
3266 | // dead code than checking if instruction involved is dead or not. |
3267 | void GVNPass::assignValNumForDeadCode() { |
3268 | for (BasicBlock *BB : DeadBlocks) { |
3269 | for (Instruction &Inst : *BB) { |
3270 | unsigned ValNum = VN.lookupOrAdd(V: &Inst); |
3271 | addToLeaderTable(N: ValNum, V: &Inst, BB); |
3272 | } |
3273 | } |
3274 | } |
3275 | |
3276 | class llvm::gvn::GVNLegacyPass : public FunctionPass { |
3277 | public: |
3278 | static char ID; // Pass identification, replacement for typeid |
3279 | |
3280 | explicit GVNLegacyPass(bool NoMemDepAnalysis = !GVNEnableMemDep) |
3281 | : FunctionPass(ID), Impl(GVNOptions().setMemDep(!NoMemDepAnalysis)) { |
3282 | initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry()); |
3283 | } |
3284 | |
3285 | bool runOnFunction(Function &F) override { |
3286 | if (skipFunction(F)) |
3287 | return false; |
3288 | |
3289 | auto *MSSAWP = getAnalysisIfAvailable<MemorySSAWrapperPass>(); |
3290 | return Impl.runImpl( |
3291 | F, RunAC&: getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F), |
3292 | RunDT&: getAnalysis<DominatorTreeWrapperPass>().getDomTree(), |
3293 | RunTLI: getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F), |
3294 | RunAA&: getAnalysis<AAResultsWrapperPass>().getAAResults(), |
3295 | RunMD: Impl.isMemDepEnabled() |
3296 | ? &getAnalysis<MemoryDependenceWrapperPass>().getMemDep() |
3297 | : nullptr, |
3298 | LI&: getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), |
3299 | RunORE: &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(), |
3300 | MSSA: MSSAWP ? &MSSAWP->getMSSA() : nullptr); |
3301 | } |
3302 | |
3303 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
3304 | AU.addRequired<AssumptionCacheTracker>(); |
3305 | AU.addRequired<DominatorTreeWrapperPass>(); |
3306 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
3307 | AU.addRequired<LoopInfoWrapperPass>(); |
3308 | if (Impl.isMemDepEnabled()) |
3309 | AU.addRequired<MemoryDependenceWrapperPass>(); |
3310 | AU.addRequired<AAResultsWrapperPass>(); |
3311 | AU.addPreserved<DominatorTreeWrapperPass>(); |
3312 | AU.addPreserved<GlobalsAAWrapperPass>(); |
3313 | AU.addPreserved<TargetLibraryInfoWrapperPass>(); |
3314 | AU.addPreserved<LoopInfoWrapperPass>(); |
3315 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); |
3316 | AU.addPreserved<MemorySSAWrapperPass>(); |
3317 | } |
3318 | |
3319 | private: |
3320 | GVNPass Impl; |
3321 | }; |
3322 | |
3323 | char GVNLegacyPass::ID = 0; |
3324 | |
3325 | INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn" , "Global Value Numbering" , false, false) |
3326 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) |
3327 | INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass) |
3328 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
3329 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
3330 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
3331 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass) |
3332 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass) |
3333 | INITIALIZE_PASS_END(GVNLegacyPass, "gvn" , "Global Value Numbering" , false, false) |
3334 | |
3335 | // The public interface to this file... |
3336 | FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) { |
3337 | return new GVNLegacyPass(NoMemDepAnalysis); |
3338 | } |
3339 | |