File: | llvm/lib/Transforms/Scalar/GVN.cpp |
Warning: | line 2971, column 29 Although the value stored to 'P' is used in the enclosing expression, the value is never actually read from 'P' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
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/PointerIntPair.h" |
23 | #include "llvm/ADT/PostOrderIterator.h" |
24 | #include "llvm/ADT/STLExtras.h" |
25 | #include "llvm/ADT/SetVector.h" |
26 | #include "llvm/ADT/SmallPtrSet.h" |
27 | #include "llvm/ADT/SmallVector.h" |
28 | #include "llvm/ADT/Statistic.h" |
29 | #include "llvm/Analysis/AliasAnalysis.h" |
30 | #include "llvm/Analysis/AssumeBundleQueries.h" |
31 | #include "llvm/Analysis/AssumptionCache.h" |
32 | #include "llvm/Analysis/CFG.h" |
33 | #include "llvm/Analysis/DomTreeUpdater.h" |
34 | #include "llvm/Analysis/GlobalsModRef.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/Config/llvm-config.h" |
46 | #include "llvm/IR/Attributes.h" |
47 | #include "llvm/IR/BasicBlock.h" |
48 | #include "llvm/IR/Constant.h" |
49 | #include "llvm/IR/Constants.h" |
50 | #include "llvm/IR/DataLayout.h" |
51 | #include "llvm/IR/DebugLoc.h" |
52 | #include "llvm/IR/Dominators.h" |
53 | #include "llvm/IR/Function.h" |
54 | #include "llvm/IR/InstrTypes.h" |
55 | #include "llvm/IR/Instruction.h" |
56 | #include "llvm/IR/Instructions.h" |
57 | #include "llvm/IR/IntrinsicInst.h" |
58 | #include "llvm/IR/Intrinsics.h" |
59 | #include "llvm/IR/LLVMContext.h" |
60 | #include "llvm/IR/Metadata.h" |
61 | #include "llvm/IR/Module.h" |
62 | #include "llvm/IR/Operator.h" |
63 | #include "llvm/IR/PassManager.h" |
64 | #include "llvm/IR/PatternMatch.h" |
65 | #include "llvm/IR/Type.h" |
66 | #include "llvm/IR/Use.h" |
67 | #include "llvm/IR/Value.h" |
68 | #include "llvm/InitializePasses.h" |
69 | #include "llvm/Pass.h" |
70 | #include "llvm/Support/Casting.h" |
71 | #include "llvm/Support/CommandLine.h" |
72 | #include "llvm/Support/Compiler.h" |
73 | #include "llvm/Support/Debug.h" |
74 | #include "llvm/Support/raw_ostream.h" |
75 | #include "llvm/Transforms/Utils.h" |
76 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
77 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
78 | #include "llvm/Transforms/Utils/Local.h" |
79 | #include "llvm/Transforms/Utils/SSAUpdater.h" |
80 | #include "llvm/Transforms/Utils/VNCoercion.h" |
81 | #include <algorithm> |
82 | #include <cassert> |
83 | #include <cstdint> |
84 | #include <utility> |
85 | |
86 | using namespace llvm; |
87 | using namespace llvm::gvn; |
88 | using namespace llvm::VNCoercion; |
89 | using namespace PatternMatch; |
90 | |
91 | #define DEBUG_TYPE"gvn" "gvn" |
92 | |
93 | STATISTIC(NumGVNInstr, "Number of instructions deleted")static llvm::Statistic NumGVNInstr = {"gvn", "NumGVNInstr", "Number of instructions deleted" }; |
94 | STATISTIC(NumGVNLoad, "Number of loads deleted")static llvm::Statistic NumGVNLoad = {"gvn", "NumGVNLoad", "Number of loads deleted" }; |
95 | STATISTIC(NumGVNPRE, "Number of instructions PRE'd")static llvm::Statistic NumGVNPRE = {"gvn", "NumGVNPRE", "Number of instructions PRE'd" }; |
96 | STATISTIC(NumGVNBlocks, "Number of blocks merged")static llvm::Statistic NumGVNBlocks = {"gvn", "NumGVNBlocks", "Number of blocks merged"}; |
97 | STATISTIC(NumGVNSimpl, "Number of instructions simplified")static llvm::Statistic NumGVNSimpl = {"gvn", "NumGVNSimpl", "Number of instructions simplified" }; |
98 | STATISTIC(NumGVNEqProp, "Number of equalities propagated")static llvm::Statistic NumGVNEqProp = {"gvn", "NumGVNEqProp", "Number of equalities propagated"}; |
99 | STATISTIC(NumPRELoad, "Number of loads PRE'd")static llvm::Statistic NumPRELoad = {"gvn", "NumPRELoad", "Number of loads PRE'd" }; |
100 | STATISTIC(NumPRELoopLoad, "Number of loop loads PRE'd")static llvm::Statistic NumPRELoopLoad = {"gvn", "NumPRELoopLoad" , "Number of loop loads PRE'd"}; |
101 | |
102 | STATISTIC(IsValueFullyAvailableInBlockNumSpeculationsMax,static llvm::Statistic IsValueFullyAvailableInBlockNumSpeculationsMax = {"gvn", "IsValueFullyAvailableInBlockNumSpeculationsMax", "Number of blocks speculated as available in " "IsValueFullyAvailableInBlock(), max"} |
103 | "Number of blocks speculated as available in "static llvm::Statistic IsValueFullyAvailableInBlockNumSpeculationsMax = {"gvn", "IsValueFullyAvailableInBlockNumSpeculationsMax", "Number of blocks speculated as available in " "IsValueFullyAvailableInBlock(), max"} |
104 | "IsValueFullyAvailableInBlock(), max")static llvm::Statistic IsValueFullyAvailableInBlockNumSpeculationsMax = {"gvn", "IsValueFullyAvailableInBlockNumSpeculationsMax", "Number of blocks speculated as available in " "IsValueFullyAvailableInBlock(), max"}; |
105 | STATISTIC(MaxBBSpeculationCutoffReachedTimes,static llvm::Statistic MaxBBSpeculationCutoffReachedTimes = { "gvn", "MaxBBSpeculationCutoffReachedTimes", "Number of times we we reached gvn-max-block-speculations cut-off " "preventing further exploration"} |
106 | "Number of times we we reached gvn-max-block-speculations cut-off "static llvm::Statistic MaxBBSpeculationCutoffReachedTimes = { "gvn", "MaxBBSpeculationCutoffReachedTimes", "Number of times we we reached gvn-max-block-speculations cut-off " "preventing further exploration"} |
107 | "preventing further exploration")static llvm::Statistic MaxBBSpeculationCutoffReachedTimes = { "gvn", "MaxBBSpeculationCutoffReachedTimes", "Number of times we we reached gvn-max-block-speculations cut-off " "preventing further exploration"}; |
108 | |
109 | static cl::opt<bool> GVNEnablePRE("enable-pre", cl::init(true), cl::Hidden); |
110 | static cl::opt<bool> GVNEnableLoadPRE("enable-load-pre", cl::init(true)); |
111 | static cl::opt<bool> GVNEnableLoadInLoopPRE("enable-load-in-loop-pre", |
112 | cl::init(true)); |
113 | static cl::opt<bool> |
114 | GVNEnableSplitBackedgeInLoadPRE("enable-split-backedge-in-load-pre", |
115 | cl::init(true)); |
116 | static cl::opt<bool> GVNEnableMemDep("enable-gvn-memdep", cl::init(true)); |
117 | |
118 | static cl::opt<uint32_t> MaxNumDeps( |
119 | "gvn-max-num-deps", cl::Hidden, cl::init(100), cl::ZeroOrMore, |
120 | cl::desc("Max number of dependences to attempt Load PRE (default = 100)")); |
121 | |
122 | // This is based on IsValueFullyAvailableInBlockNumSpeculationsMax stat. |
123 | static cl::opt<uint32_t> MaxBBSpeculations( |
124 | "gvn-max-block-speculations", cl::Hidden, cl::init(600), cl::ZeroOrMore, |
125 | cl::desc("Max number of blocks we're willing to speculate on (and recurse " |
126 | "into) when deducing if a value is fully available or not in GVN " |
127 | "(default = 600)")); |
128 | |
129 | struct llvm::GVN::Expression { |
130 | uint32_t opcode; |
131 | bool commutative = false; |
132 | Type *type = nullptr; |
133 | SmallVector<uint32_t, 4> varargs; |
134 | |
135 | Expression(uint32_t o = ~2U) : opcode(o) {} |
136 | |
137 | bool operator==(const Expression &other) const { |
138 | if (opcode != other.opcode) |
139 | return false; |
140 | if (opcode == ~0U || opcode == ~1U) |
141 | return true; |
142 | if (type != other.type) |
143 | return false; |
144 | if (varargs != other.varargs) |
145 | return false; |
146 | return true; |
147 | } |
148 | |
149 | friend hash_code hash_value(const Expression &Value) { |
150 | return hash_combine( |
151 | Value.opcode, Value.type, |
152 | hash_combine_range(Value.varargs.begin(), Value.varargs.end())); |
153 | } |
154 | }; |
155 | |
156 | namespace llvm { |
157 | |
158 | template <> struct DenseMapInfo<GVN::Expression> { |
159 | static inline GVN::Expression getEmptyKey() { return ~0U; } |
160 | static inline GVN::Expression getTombstoneKey() { return ~1U; } |
161 | |
162 | static unsigned getHashValue(const GVN::Expression &e) { |
163 | using llvm::hash_value; |
164 | |
165 | return static_cast<unsigned>(hash_value(e)); |
166 | } |
167 | |
168 | static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) { |
169 | return LHS == RHS; |
170 | } |
171 | }; |
172 | |
173 | } // end namespace llvm |
174 | |
175 | /// Represents a particular available value that we know how to materialize. |
176 | /// Materialization of an AvailableValue never fails. An AvailableValue is |
177 | /// implicitly associated with a rematerialization point which is the |
178 | /// location of the instruction from which it was formed. |
179 | struct llvm::gvn::AvailableValue { |
180 | enum ValType { |
181 | SimpleVal, // A simple offsetted value that is accessed. |
182 | LoadVal, // A value produced by a load. |
183 | MemIntrin, // A memory intrinsic which is loaded from. |
184 | UndefVal // A UndefValue representing a value from dead block (which |
185 | // is not yet physically removed from the CFG). |
186 | }; |
187 | |
188 | /// V - The value that is live out of the block. |
189 | PointerIntPair<Value *, 2, ValType> Val; |
190 | |
191 | /// Offset - The byte offset in Val that is interesting for the load query. |
192 | unsigned Offset = 0; |
193 | |
194 | static AvailableValue get(Value *V, unsigned Offset = 0) { |
195 | AvailableValue Res; |
196 | Res.Val.setPointer(V); |
197 | Res.Val.setInt(SimpleVal); |
198 | Res.Offset = Offset; |
199 | return Res; |
200 | } |
201 | |
202 | static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) { |
203 | AvailableValue Res; |
204 | Res.Val.setPointer(MI); |
205 | Res.Val.setInt(MemIntrin); |
206 | Res.Offset = Offset; |
207 | return Res; |
208 | } |
209 | |
210 | static AvailableValue getLoad(LoadInst *Load, unsigned Offset = 0) { |
211 | AvailableValue Res; |
212 | Res.Val.setPointer(Load); |
213 | Res.Val.setInt(LoadVal); |
214 | Res.Offset = Offset; |
215 | return Res; |
216 | } |
217 | |
218 | static AvailableValue getUndef() { |
219 | AvailableValue Res; |
220 | Res.Val.setPointer(nullptr); |
221 | Res.Val.setInt(UndefVal); |
222 | Res.Offset = 0; |
223 | return Res; |
224 | } |
225 | |
226 | bool isSimpleValue() const { return Val.getInt() == SimpleVal; } |
227 | bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } |
228 | bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } |
229 | bool isUndefValue() const { return Val.getInt() == UndefVal; } |
230 | |
231 | Value *getSimpleValue() const { |
232 | assert(isSimpleValue() && "Wrong accessor")(static_cast <bool> (isSimpleValue() && "Wrong accessor" ) ? void (0) : __assert_fail ("isSimpleValue() && \"Wrong accessor\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 232, __extension__ __PRETTY_FUNCTION__)); |
233 | return Val.getPointer(); |
234 | } |
235 | |
236 | LoadInst *getCoercedLoadValue() const { |
237 | assert(isCoercedLoadValue() && "Wrong accessor")(static_cast <bool> (isCoercedLoadValue() && "Wrong accessor" ) ? void (0) : __assert_fail ("isCoercedLoadValue() && \"Wrong accessor\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 237, __extension__ __PRETTY_FUNCTION__)); |
238 | return cast<LoadInst>(Val.getPointer()); |
239 | } |
240 | |
241 | MemIntrinsic *getMemIntrinValue() const { |
242 | assert(isMemIntrinValue() && "Wrong accessor")(static_cast <bool> (isMemIntrinValue() && "Wrong accessor" ) ? void (0) : __assert_fail ("isMemIntrinValue() && \"Wrong accessor\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 242, __extension__ __PRETTY_FUNCTION__)); |
243 | return cast<MemIntrinsic>(Val.getPointer()); |
244 | } |
245 | |
246 | /// Emit code at the specified insertion point to adjust the value defined |
247 | /// here to the specified type. This handles various coercion cases. |
248 | Value *MaterializeAdjustedValue(LoadInst *Load, Instruction *InsertPt, |
249 | GVN &gvn) const; |
250 | }; |
251 | |
252 | /// Represents an AvailableValue which can be rematerialized at the end of |
253 | /// the associated BasicBlock. |
254 | struct llvm::gvn::AvailableValueInBlock { |
255 | /// BB - The basic block in question. |
256 | BasicBlock *BB = nullptr; |
257 | |
258 | /// AV - The actual available value |
259 | AvailableValue AV; |
260 | |
261 | static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) { |
262 | AvailableValueInBlock Res; |
263 | Res.BB = BB; |
264 | Res.AV = std::move(AV); |
265 | return Res; |
266 | } |
267 | |
268 | static AvailableValueInBlock get(BasicBlock *BB, Value *V, |
269 | unsigned Offset = 0) { |
270 | return get(BB, AvailableValue::get(V, Offset)); |
271 | } |
272 | |
273 | static AvailableValueInBlock getUndef(BasicBlock *BB) { |
274 | return get(BB, AvailableValue::getUndef()); |
275 | } |
276 | |
277 | /// Emit code at the end of this block to adjust the value defined here to |
278 | /// the specified type. This handles various coercion cases. |
279 | Value *MaterializeAdjustedValue(LoadInst *Load, GVN &gvn) const { |
280 | return AV.MaterializeAdjustedValue(Load, BB->getTerminator(), gvn); |
281 | } |
282 | }; |
283 | |
284 | //===----------------------------------------------------------------------===// |
285 | // ValueTable Internal Functions |
286 | //===----------------------------------------------------------------------===// |
287 | |
288 | GVN::Expression GVN::ValueTable::createExpr(Instruction *I) { |
289 | Expression e; |
290 | e.type = I->getType(); |
291 | e.opcode = I->getOpcode(); |
292 | if (const GCRelocateInst *GCR = dyn_cast<GCRelocateInst>(I)) { |
293 | // gc.relocate is 'special' call: its second and third operands are |
294 | // not real values, but indices into statepoint's argument list. |
295 | // Use the refered to values for purposes of identity. |
296 | e.varargs.push_back(lookupOrAdd(GCR->getOperand(0))); |
297 | e.varargs.push_back(lookupOrAdd(GCR->getBasePtr())); |
298 | e.varargs.push_back(lookupOrAdd(GCR->getDerivedPtr())); |
299 | } else { |
300 | for (Use &Op : I->operands()) |
301 | e.varargs.push_back(lookupOrAdd(Op)); |
302 | } |
303 | if (I->isCommutative()) { |
304 | // Ensure that commutative instructions that only differ by a permutation |
305 | // of their operands get the same value number by sorting the operand value |
306 | // numbers. Since commutative operands are the 1st two operands it is more |
307 | // efficient to sort by hand rather than using, say, std::sort. |
308 | assert(I->getNumOperands() >= 2 && "Unsupported commutative instruction!")(static_cast <bool> (I->getNumOperands() >= 2 && "Unsupported commutative instruction!") ? void (0) : __assert_fail ("I->getNumOperands() >= 2 && \"Unsupported commutative instruction!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 308, __extension__ __PRETTY_FUNCTION__)); |
309 | if (e.varargs[0] > e.varargs[1]) |
310 | std::swap(e.varargs[0], e.varargs[1]); |
311 | e.commutative = true; |
312 | } |
313 | |
314 | if (auto *C = dyn_cast<CmpInst>(I)) { |
315 | // Sort the operand value numbers so x<y and y>x get the same value number. |
316 | CmpInst::Predicate Predicate = C->getPredicate(); |
317 | if (e.varargs[0] > e.varargs[1]) { |
318 | std::swap(e.varargs[0], e.varargs[1]); |
319 | Predicate = CmpInst::getSwappedPredicate(Predicate); |
320 | } |
321 | e.opcode = (C->getOpcode() << 8) | Predicate; |
322 | e.commutative = true; |
323 | } else if (auto *E = dyn_cast<InsertValueInst>(I)) { |
324 | e.varargs.append(E->idx_begin(), E->idx_end()); |
325 | } else if (auto *SVI = dyn_cast<ShuffleVectorInst>(I)) { |
326 | ArrayRef<int> ShuffleMask = SVI->getShuffleMask(); |
327 | e.varargs.append(ShuffleMask.begin(), ShuffleMask.end()); |
328 | } |
329 | |
330 | return e; |
331 | } |
332 | |
333 | GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode, |
334 | CmpInst::Predicate Predicate, |
335 | Value *LHS, Value *RHS) { |
336 | assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&(static_cast <bool> ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && "Not a comparison!") ? void (0) : __assert_fail ("(Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && \"Not a comparison!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 337, __extension__ __PRETTY_FUNCTION__)) |
337 | "Not a comparison!")(static_cast <bool> ((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && "Not a comparison!") ? void (0) : __assert_fail ("(Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && \"Not a comparison!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 337, __extension__ __PRETTY_FUNCTION__)); |
338 | Expression e; |
339 | e.type = CmpInst::makeCmpResultType(LHS->getType()); |
340 | e.varargs.push_back(lookupOrAdd(LHS)); |
341 | e.varargs.push_back(lookupOrAdd(RHS)); |
342 | |
343 | // Sort the operand value numbers so x<y and y>x get the same value number. |
344 | if (e.varargs[0] > e.varargs[1]) { |
345 | std::swap(e.varargs[0], e.varargs[1]); |
346 | Predicate = CmpInst::getSwappedPredicate(Predicate); |
347 | } |
348 | e.opcode = (Opcode << 8) | Predicate; |
349 | e.commutative = true; |
350 | return e; |
351 | } |
352 | |
353 | GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) { |
354 | assert(EI && "Not an ExtractValueInst?")(static_cast <bool> (EI && "Not an ExtractValueInst?" ) ? void (0) : __assert_fail ("EI && \"Not an ExtractValueInst?\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 354, __extension__ __PRETTY_FUNCTION__)); |
355 | Expression e; |
356 | e.type = EI->getType(); |
357 | e.opcode = 0; |
358 | |
359 | WithOverflowInst *WO = dyn_cast<WithOverflowInst>(EI->getAggregateOperand()); |
360 | if (WO != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0) { |
361 | // EI is an extract from one of our with.overflow intrinsics. Synthesize |
362 | // a semantically equivalent expression instead of an extract value |
363 | // expression. |
364 | e.opcode = WO->getBinaryOp(); |
365 | e.varargs.push_back(lookupOrAdd(WO->getLHS())); |
366 | e.varargs.push_back(lookupOrAdd(WO->getRHS())); |
367 | return e; |
368 | } |
369 | |
370 | // Not a recognised intrinsic. Fall back to producing an extract value |
371 | // expression. |
372 | e.opcode = EI->getOpcode(); |
373 | for (Use &Op : EI->operands()) |
374 | e.varargs.push_back(lookupOrAdd(Op)); |
375 | |
376 | append_range(e.varargs, EI->indices()); |
377 | |
378 | return e; |
379 | } |
380 | |
381 | //===----------------------------------------------------------------------===// |
382 | // ValueTable External Functions |
383 | //===----------------------------------------------------------------------===// |
384 | |
385 | GVN::ValueTable::ValueTable() = default; |
386 | GVN::ValueTable::ValueTable(const ValueTable &) = default; |
387 | GVN::ValueTable::ValueTable(ValueTable &&) = default; |
388 | GVN::ValueTable::~ValueTable() = default; |
389 | GVN::ValueTable &GVN::ValueTable::operator=(const GVN::ValueTable &Arg) = default; |
390 | |
391 | /// add - Insert a value into the table with a specified value number. |
392 | void GVN::ValueTable::add(Value *V, uint32_t num) { |
393 | valueNumbering.insert(std::make_pair(V, num)); |
394 | if (PHINode *PN = dyn_cast<PHINode>(V)) |
395 | NumberingPhi[num] = PN; |
396 | } |
397 | |
398 | uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) { |
399 | if (AA->doesNotAccessMemory(C)) { |
400 | Expression exp = createExpr(C); |
401 | uint32_t e = assignExpNewValueNum(exp).first; |
402 | valueNumbering[C] = e; |
403 | return e; |
404 | } else if (MD && AA->onlyReadsMemory(C)) { |
405 | Expression exp = createExpr(C); |
406 | auto ValNum = assignExpNewValueNum(exp); |
407 | if (ValNum.second) { |
408 | valueNumbering[C] = ValNum.first; |
409 | return ValNum.first; |
410 | } |
411 | |
412 | MemDepResult local_dep = MD->getDependency(C); |
413 | |
414 | if (!local_dep.isDef() && !local_dep.isNonLocal()) { |
415 | valueNumbering[C] = nextValueNumber; |
416 | return nextValueNumber++; |
417 | } |
418 | |
419 | if (local_dep.isDef()) { |
420 | // For masked load/store intrinsics, the local_dep may actully be |
421 | // a normal load or store instruction. |
422 | CallInst *local_cdep = dyn_cast<CallInst>(local_dep.getInst()); |
423 | |
424 | if (!local_cdep || |
425 | local_cdep->getNumArgOperands() != C->getNumArgOperands()) { |
426 | valueNumbering[C] = nextValueNumber; |
427 | return nextValueNumber++; |
428 | } |
429 | |
430 | for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { |
431 | uint32_t c_vn = lookupOrAdd(C->getArgOperand(i)); |
432 | uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i)); |
433 | if (c_vn != cd_vn) { |
434 | valueNumbering[C] = nextValueNumber; |
435 | return nextValueNumber++; |
436 | } |
437 | } |
438 | |
439 | uint32_t v = lookupOrAdd(local_cdep); |
440 | valueNumbering[C] = v; |
441 | return v; |
442 | } |
443 | |
444 | // Non-local case. |
445 | const MemoryDependenceResults::NonLocalDepInfo &deps = |
446 | MD->getNonLocalCallDependency(C); |
447 | // FIXME: Move the checking logic to MemDep! |
448 | CallInst* cdep = nullptr; |
449 | |
450 | // Check to see if we have a single dominating call instruction that is |
451 | // identical to C. |
452 | for (unsigned i = 0, e = deps.size(); i != e; ++i) { |
453 | const NonLocalDepEntry *I = &deps[i]; |
454 | if (I->getResult().isNonLocal()) |
455 | continue; |
456 | |
457 | // We don't handle non-definitions. If we already have a call, reject |
458 | // instruction dependencies. |
459 | if (!I->getResult().isDef() || cdep != nullptr) { |
460 | cdep = nullptr; |
461 | break; |
462 | } |
463 | |
464 | CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); |
465 | // FIXME: All duplicated with non-local case. |
466 | if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ |
467 | cdep = NonLocalDepCall; |
468 | continue; |
469 | } |
470 | |
471 | cdep = nullptr; |
472 | break; |
473 | } |
474 | |
475 | if (!cdep) { |
476 | valueNumbering[C] = nextValueNumber; |
477 | return nextValueNumber++; |
478 | } |
479 | |
480 | if (cdep->getNumArgOperands() != C->getNumArgOperands()) { |
481 | valueNumbering[C] = nextValueNumber; |
482 | return nextValueNumber++; |
483 | } |
484 | for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { |
485 | uint32_t c_vn = lookupOrAdd(C->getArgOperand(i)); |
486 | uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i)); |
487 | if (c_vn != cd_vn) { |
488 | valueNumbering[C] = nextValueNumber; |
489 | return nextValueNumber++; |
490 | } |
491 | } |
492 | |
493 | uint32_t v = lookupOrAdd(cdep); |
494 | valueNumbering[C] = v; |
495 | return v; |
496 | } else { |
497 | valueNumbering[C] = nextValueNumber; |
498 | return nextValueNumber++; |
499 | } |
500 | } |
501 | |
502 | /// Returns true if a value number exists for the specified value. |
503 | bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; } |
504 | |
505 | /// lookup_or_add - Returns the value number for the specified value, assigning |
506 | /// it a new number if it did not have one before. |
507 | uint32_t GVN::ValueTable::lookupOrAdd(Value *V) { |
508 | DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); |
509 | if (VI != valueNumbering.end()) |
510 | return VI->second; |
511 | |
512 | if (!isa<Instruction>(V)) { |
513 | valueNumbering[V] = nextValueNumber; |
514 | return nextValueNumber++; |
515 | } |
516 | |
517 | Instruction* I = cast<Instruction>(V); |
518 | Expression exp; |
519 | switch (I->getOpcode()) { |
520 | case Instruction::Call: |
521 | return lookupOrAddCall(cast<CallInst>(I)); |
522 | case Instruction::FNeg: |
523 | case Instruction::Add: |
524 | case Instruction::FAdd: |
525 | case Instruction::Sub: |
526 | case Instruction::FSub: |
527 | case Instruction::Mul: |
528 | case Instruction::FMul: |
529 | case Instruction::UDiv: |
530 | case Instruction::SDiv: |
531 | case Instruction::FDiv: |
532 | case Instruction::URem: |
533 | case Instruction::SRem: |
534 | case Instruction::FRem: |
535 | case Instruction::Shl: |
536 | case Instruction::LShr: |
537 | case Instruction::AShr: |
538 | case Instruction::And: |
539 | case Instruction::Or: |
540 | case Instruction::Xor: |
541 | case Instruction::ICmp: |
542 | case Instruction::FCmp: |
543 | case Instruction::Trunc: |
544 | case Instruction::ZExt: |
545 | case Instruction::SExt: |
546 | case Instruction::FPToUI: |
547 | case Instruction::FPToSI: |
548 | case Instruction::UIToFP: |
549 | case Instruction::SIToFP: |
550 | case Instruction::FPTrunc: |
551 | case Instruction::FPExt: |
552 | case Instruction::PtrToInt: |
553 | case Instruction::IntToPtr: |
554 | case Instruction::AddrSpaceCast: |
555 | case Instruction::BitCast: |
556 | case Instruction::Select: |
557 | case Instruction::Freeze: |
558 | case Instruction::ExtractElement: |
559 | case Instruction::InsertElement: |
560 | case Instruction::ShuffleVector: |
561 | case Instruction::InsertValue: |
562 | case Instruction::GetElementPtr: |
563 | exp = createExpr(I); |
564 | break; |
565 | case Instruction::ExtractValue: |
566 | exp = createExtractvalueExpr(cast<ExtractValueInst>(I)); |
567 | break; |
568 | case Instruction::PHI: |
569 | valueNumbering[V] = nextValueNumber; |
570 | NumberingPhi[nextValueNumber] = cast<PHINode>(V); |
571 | return nextValueNumber++; |
572 | default: |
573 | valueNumbering[V] = nextValueNumber; |
574 | return nextValueNumber++; |
575 | } |
576 | |
577 | uint32_t e = assignExpNewValueNum(exp).first; |
578 | valueNumbering[V] = e; |
579 | return e; |
580 | } |
581 | |
582 | /// Returns the value number of the specified value. Fails if |
583 | /// the value has not yet been numbered. |
584 | uint32_t GVN::ValueTable::lookup(Value *V, bool Verify) const { |
585 | DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); |
586 | if (Verify) { |
587 | assert(VI != valueNumbering.end() && "Value not numbered?")(static_cast <bool> (VI != valueNumbering.end() && "Value not numbered?") ? void (0) : __assert_fail ("VI != valueNumbering.end() && \"Value not numbered?\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 587, __extension__ __PRETTY_FUNCTION__)); |
588 | return VI->second; |
589 | } |
590 | return (VI != valueNumbering.end()) ? VI->second : 0; |
591 | } |
592 | |
593 | /// Returns the value number of the given comparison, |
594 | /// assigning it a new number if it did not have one before. Useful when |
595 | /// we deduced the result of a comparison, but don't immediately have an |
596 | /// instruction realizing that comparison to hand. |
597 | uint32_t GVN::ValueTable::lookupOrAddCmp(unsigned Opcode, |
598 | CmpInst::Predicate Predicate, |
599 | Value *LHS, Value *RHS) { |
600 | Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS); |
601 | return assignExpNewValueNum(exp).first; |
602 | } |
603 | |
604 | /// Remove all entries from the ValueTable. |
605 | void GVN::ValueTable::clear() { |
606 | valueNumbering.clear(); |
607 | expressionNumbering.clear(); |
608 | NumberingPhi.clear(); |
609 | PhiTranslateTable.clear(); |
610 | nextValueNumber = 1; |
611 | Expressions.clear(); |
612 | ExprIdx.clear(); |
613 | nextExprNumber = 0; |
614 | } |
615 | |
616 | /// Remove a value from the value numbering. |
617 | void GVN::ValueTable::erase(Value *V) { |
618 | uint32_t Num = valueNumbering.lookup(V); |
619 | valueNumbering.erase(V); |
620 | // If V is PHINode, V <--> value number is an one-to-one mapping. |
621 | if (isa<PHINode>(V)) |
622 | NumberingPhi.erase(Num); |
623 | } |
624 | |
625 | /// verifyRemoved - Verify that the value is removed from all internal data |
626 | /// structures. |
627 | void GVN::ValueTable::verifyRemoved(const Value *V) const { |
628 | for (DenseMap<Value*, uint32_t>::const_iterator |
629 | I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { |
630 | assert(I->first != V && "Inst still occurs in value numbering map!")(static_cast <bool> (I->first != V && "Inst still occurs in value numbering map!" ) ? void (0) : __assert_fail ("I->first != V && \"Inst still occurs in value numbering map!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 630, __extension__ __PRETTY_FUNCTION__)); |
631 | } |
632 | } |
633 | |
634 | //===----------------------------------------------------------------------===// |
635 | // GVN Pass |
636 | //===----------------------------------------------------------------------===// |
637 | |
638 | bool GVN::isPREEnabled() const { |
639 | return Options.AllowPRE.getValueOr(GVNEnablePRE); |
640 | } |
641 | |
642 | bool GVN::isLoadPREEnabled() const { |
643 | return Options.AllowLoadPRE.getValueOr(GVNEnableLoadPRE); |
644 | } |
645 | |
646 | bool GVN::isLoadInLoopPREEnabled() const { |
647 | return Options.AllowLoadInLoopPRE.getValueOr(GVNEnableLoadInLoopPRE); |
648 | } |
649 | |
650 | bool GVN::isLoadPRESplitBackedgeEnabled() const { |
651 | return Options.AllowLoadPRESplitBackedge.getValueOr( |
652 | GVNEnableSplitBackedgeInLoadPRE); |
653 | } |
654 | |
655 | bool GVN::isMemDepEnabled() const { |
656 | return Options.AllowMemDep.getValueOr(GVNEnableMemDep); |
657 | } |
658 | |
659 | PreservedAnalyses GVN::run(Function &F, FunctionAnalysisManager &AM) { |
660 | // FIXME: The order of evaluation of these 'getResult' calls is very |
661 | // significant! Re-ordering these variables will cause GVN when run alone to |
662 | // be less effective! We should fix memdep and basic-aa to not exhibit this |
663 | // behavior, but until then don't change the order here. |
664 | auto &AC = AM.getResult<AssumptionAnalysis>(F); |
665 | auto &DT = AM.getResult<DominatorTreeAnalysis>(F); |
666 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); |
667 | auto &AA = AM.getResult<AAManager>(F); |
668 | auto *MemDep = |
669 | isMemDepEnabled() ? &AM.getResult<MemoryDependenceAnalysis>(F) : nullptr; |
670 | auto *LI = AM.getCachedResult<LoopAnalysis>(F); |
671 | auto *MSSA = AM.getCachedResult<MemorySSAAnalysis>(F); |
672 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); |
673 | bool Changed = runImpl(F, AC, DT, TLI, AA, MemDep, LI, &ORE, |
674 | MSSA ? &MSSA->getMSSA() : nullptr); |
675 | if (!Changed) |
676 | return PreservedAnalyses::all(); |
677 | PreservedAnalyses PA; |
678 | PA.preserve<DominatorTreeAnalysis>(); |
679 | PA.preserve<TargetLibraryAnalysis>(); |
680 | if (MSSA) |
681 | PA.preserve<MemorySSAAnalysis>(); |
682 | if (LI) |
683 | PA.preserve<LoopAnalysis>(); |
684 | return PA; |
685 | } |
686 | |
687 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
688 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void GVN::dump(DenseMap<uint32_t, Value*>& d) const { |
689 | errs() << "{\n"; |
690 | for (auto &I : d) { |
691 | errs() << I.first << "\n"; |
692 | I.second->dump(); |
693 | } |
694 | errs() << "}\n"; |
695 | } |
696 | #endif |
697 | |
698 | enum class AvailabilityState : char { |
699 | /// We know the block *is not* fully available. This is a fixpoint. |
700 | Unavailable = 0, |
701 | /// We know the block *is* fully available. This is a fixpoint. |
702 | Available = 1, |
703 | /// We do not know whether the block is fully available or not, |
704 | /// but we are currently speculating that it will be. |
705 | /// If it would have turned out that the block was, in fact, not fully |
706 | /// available, this would have been cleaned up into an Unavailable. |
707 | SpeculativelyAvailable = 2, |
708 | }; |
709 | |
710 | /// Return true if we can prove that the value |
711 | /// we're analyzing is fully available in the specified block. As we go, keep |
712 | /// track of which blocks we know are fully alive in FullyAvailableBlocks. This |
713 | /// map is actually a tri-state map with the following values: |
714 | /// 0) we know the block *is not* fully available. |
715 | /// 1) we know the block *is* fully available. |
716 | /// 2) we do not know whether the block is fully available or not, but we are |
717 | /// currently speculating that it will be. |
718 | static bool IsValueFullyAvailableInBlock( |
719 | BasicBlock *BB, |
720 | DenseMap<BasicBlock *, AvailabilityState> &FullyAvailableBlocks) { |
721 | SmallVector<BasicBlock *, 32> Worklist; |
722 | Optional<BasicBlock *> UnavailableBB; |
723 | |
724 | // The number of times we didn't find an entry for a block in a map and |
725 | // optimistically inserted an entry marking block as speculatively available. |
726 | unsigned NumNewNewSpeculativelyAvailableBBs = 0; |
727 | |
728 | #ifndef NDEBUG |
729 | SmallSet<BasicBlock *, 32> NewSpeculativelyAvailableBBs; |
730 | SmallVector<BasicBlock *, 32> AvailableBBs; |
731 | #endif |
732 | |
733 | Worklist.emplace_back(BB); |
734 | while (!Worklist.empty()) { |
735 | BasicBlock *CurrBB = Worklist.pop_back_val(); // LoadFO - depth-first! |
736 | // Optimistically assume that the block is Speculatively Available and check |
737 | // to see if we already know about this block in one lookup. |
738 | std::pair<DenseMap<BasicBlock *, AvailabilityState>::iterator, bool> IV = |
739 | FullyAvailableBlocks.try_emplace( |
740 | CurrBB, AvailabilityState::SpeculativelyAvailable); |
741 | AvailabilityState &State = IV.first->second; |
742 | |
743 | // Did the entry already exist for this block? |
744 | if (!IV.second) { |
745 | if (State == AvailabilityState::Unavailable) { |
746 | UnavailableBB = CurrBB; |
747 | break; // Backpropagate unavailability info. |
748 | } |
749 | |
750 | #ifndef NDEBUG |
751 | AvailableBBs.emplace_back(CurrBB); |
752 | #endif |
753 | continue; // Don't recurse further, but continue processing worklist. |
754 | } |
755 | |
756 | // No entry found for block. |
757 | ++NumNewNewSpeculativelyAvailableBBs; |
758 | bool OutOfBudget = NumNewNewSpeculativelyAvailableBBs > MaxBBSpeculations; |
759 | |
760 | // If we have exhausted our budget, mark this block as unavailable. |
761 | // Also, if this block has no predecessors, the value isn't live-in here. |
762 | if (OutOfBudget || pred_empty(CurrBB)) { |
763 | MaxBBSpeculationCutoffReachedTimes += (int)OutOfBudget; |
764 | State = AvailabilityState::Unavailable; |
765 | UnavailableBB = CurrBB; |
766 | break; // Backpropagate unavailability info. |
767 | } |
768 | |
769 | // Tentatively consider this block as speculatively available. |
770 | #ifndef NDEBUG |
771 | NewSpeculativelyAvailableBBs.insert(CurrBB); |
772 | #endif |
773 | // And further recurse into block's predecessors, in depth-first order! |
774 | Worklist.append(pred_begin(CurrBB), pred_end(CurrBB)); |
775 | } |
776 | |
777 | #if LLVM_ENABLE_STATS1 |
778 | IsValueFullyAvailableInBlockNumSpeculationsMax.updateMax( |
779 | NumNewNewSpeculativelyAvailableBBs); |
780 | #endif |
781 | |
782 | // If the block isn't marked as fixpoint yet |
783 | // (the Unavailable and Available states are fixpoints) |
784 | auto MarkAsFixpointAndEnqueueSuccessors = |
785 | [&](BasicBlock *BB, AvailabilityState FixpointState) { |
786 | auto It = FullyAvailableBlocks.find(BB); |
787 | if (It == FullyAvailableBlocks.end()) |
788 | return; // Never queried this block, leave as-is. |
789 | switch (AvailabilityState &State = It->second) { |
790 | case AvailabilityState::Unavailable: |
791 | case AvailabilityState::Available: |
792 | return; // Don't backpropagate further, continue processing worklist. |
793 | case AvailabilityState::SpeculativelyAvailable: // Fix it! |
794 | State = FixpointState; |
795 | #ifndef NDEBUG |
796 | assert(NewSpeculativelyAvailableBBs.erase(BB) &&(static_cast <bool> (NewSpeculativelyAvailableBBs.erase (BB) && "Found a speculatively available successor leftover?" ) ? void (0) : __assert_fail ("NewSpeculativelyAvailableBBs.erase(BB) && \"Found a speculatively available successor leftover?\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 797, __extension__ __PRETTY_FUNCTION__)) |
797 | "Found a speculatively available successor leftover?")(static_cast <bool> (NewSpeculativelyAvailableBBs.erase (BB) && "Found a speculatively available successor leftover?" ) ? void (0) : __assert_fail ("NewSpeculativelyAvailableBBs.erase(BB) && \"Found a speculatively available successor leftover?\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 797, __extension__ __PRETTY_FUNCTION__)); |
798 | #endif |
799 | // Queue successors for further processing. |
800 | Worklist.append(succ_begin(BB), succ_end(BB)); |
801 | return; |
802 | } |
803 | }; |
804 | |
805 | if (UnavailableBB) { |
806 | // Okay, we have encountered an unavailable block. |
807 | // Mark speculatively available blocks reachable from UnavailableBB as |
808 | // unavailable as well. Paths are terminated when they reach blocks not in |
809 | // FullyAvailableBlocks or they are not marked as speculatively available. |
810 | Worklist.clear(); |
811 | Worklist.append(succ_begin(*UnavailableBB), succ_end(*UnavailableBB)); |
812 | while (!Worklist.empty()) |
813 | MarkAsFixpointAndEnqueueSuccessors(Worklist.pop_back_val(), |
814 | AvailabilityState::Unavailable); |
815 | } |
816 | |
817 | #ifndef NDEBUG |
818 | Worklist.clear(); |
819 | for (BasicBlock *AvailableBB : AvailableBBs) |
820 | Worklist.append(succ_begin(AvailableBB), succ_end(AvailableBB)); |
821 | while (!Worklist.empty()) |
822 | MarkAsFixpointAndEnqueueSuccessors(Worklist.pop_back_val(), |
823 | AvailabilityState::Available); |
824 | |
825 | assert(NewSpeculativelyAvailableBBs.empty() &&(static_cast <bool> (NewSpeculativelyAvailableBBs.empty () && "Must have fixed all the new speculatively available blocks." ) ? void (0) : __assert_fail ("NewSpeculativelyAvailableBBs.empty() && \"Must have fixed all the new speculatively available blocks.\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 826, __extension__ __PRETTY_FUNCTION__)) |
826 | "Must have fixed all the new speculatively available blocks.")(static_cast <bool> (NewSpeculativelyAvailableBBs.empty () && "Must have fixed all the new speculatively available blocks." ) ? void (0) : __assert_fail ("NewSpeculativelyAvailableBBs.empty() && \"Must have fixed all the new speculatively available blocks.\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 826, __extension__ __PRETTY_FUNCTION__)); |
827 | #endif |
828 | |
829 | return !UnavailableBB; |
830 | } |
831 | |
832 | /// Given a set of loads specified by ValuesPerBlock, |
833 | /// construct SSA form, allowing us to eliminate Load. This returns the value |
834 | /// that should be used at Load's definition site. |
835 | static Value * |
836 | ConstructSSAForLoadSet(LoadInst *Load, |
837 | SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, |
838 | GVN &gvn) { |
839 | // Check for the fully redundant, dominating load case. In this case, we can |
840 | // just use the dominating value directly. |
841 | if (ValuesPerBlock.size() == 1 && |
842 | gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB, |
843 | Load->getParent())) { |
844 | assert(!ValuesPerBlock[0].AV.isUndefValue() &&(static_cast <bool> (!ValuesPerBlock[0].AV.isUndefValue () && "Dead BB dominate this block") ? void (0) : __assert_fail ("!ValuesPerBlock[0].AV.isUndefValue() && \"Dead BB dominate this block\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 845, __extension__ __PRETTY_FUNCTION__)) |
845 | "Dead BB dominate this block")(static_cast <bool> (!ValuesPerBlock[0].AV.isUndefValue () && "Dead BB dominate this block") ? void (0) : __assert_fail ("!ValuesPerBlock[0].AV.isUndefValue() && \"Dead BB dominate this block\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 845, __extension__ __PRETTY_FUNCTION__)); |
846 | return ValuesPerBlock[0].MaterializeAdjustedValue(Load, gvn); |
847 | } |
848 | |
849 | // Otherwise, we have to construct SSA form. |
850 | SmallVector<PHINode*, 8> NewPHIs; |
851 | SSAUpdater SSAUpdate(&NewPHIs); |
852 | SSAUpdate.Initialize(Load->getType(), Load->getName()); |
853 | |
854 | for (const AvailableValueInBlock &AV : ValuesPerBlock) { |
855 | BasicBlock *BB = AV.BB; |
856 | |
857 | if (AV.AV.isUndefValue()) |
858 | continue; |
859 | |
860 | if (SSAUpdate.HasValueForBlock(BB)) |
861 | continue; |
862 | |
863 | // If the value is the load that we will be eliminating, and the block it's |
864 | // available in is the block that the load is in, then don't add it as |
865 | // SSAUpdater will resolve the value to the relevant phi which may let it |
866 | // avoid phi construction entirely if there's actually only one value. |
867 | if (BB == Load->getParent() && |
868 | ((AV.AV.isSimpleValue() && AV.AV.getSimpleValue() == Load) || |
869 | (AV.AV.isCoercedLoadValue() && AV.AV.getCoercedLoadValue() == Load))) |
870 | continue; |
871 | |
872 | SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(Load, gvn)); |
873 | } |
874 | |
875 | // Perform PHI construction. |
876 | return SSAUpdate.GetValueInMiddleOfBlock(Load->getParent()); |
877 | } |
878 | |
879 | Value *AvailableValue::MaterializeAdjustedValue(LoadInst *Load, |
880 | Instruction *InsertPt, |
881 | GVN &gvn) const { |
882 | Value *Res; |
883 | Type *LoadTy = Load->getType(); |
884 | const DataLayout &DL = Load->getModule()->getDataLayout(); |
885 | if (isSimpleValue()) { |
886 | Res = getSimpleValue(); |
887 | if (Res->getType() != LoadTy) { |
888 | Res = getStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL); |
889 | |
890 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offsetdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " << *getSimpleValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
891 | << " " << *getSimpleValue() << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " << *getSimpleValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
892 | << *Res << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " << *getSimpleValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
893 | << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " << *getSimpleValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false); |
894 | } |
895 | } else if (isCoercedLoadValue()) { |
896 | LoadInst *CoercedLoad = getCoercedLoadValue(); |
897 | if (CoercedLoad->getType() == LoadTy && Offset == 0) { |
898 | Res = CoercedLoad; |
899 | } else { |
900 | Res = getLoadValueForLoad(CoercedLoad, Offset, LoadTy, InsertPt, DL); |
901 | // We would like to use gvn.markInstructionForDeletion here, but we can't |
902 | // because the load is already memoized into the leader map table that GVN |
903 | // tracks. It is potentially possible to remove the load from the table, |
904 | // but then there all of the operations based on it would need to be |
905 | // rehashed. Just leave the dead load around. |
906 | gvn.getMemDep().removeInstruction(CoercedLoad); |
907 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offsetdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " << *getCoercedLoadValue( ) << '\n' << *Res << '\n' << "\n\n\n" ; } } while (false) |
908 | << " " << *getCoercedLoadValue() << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " << *getCoercedLoadValue( ) << '\n' << *Res << '\n' << "\n\n\n" ; } } while (false) |
909 | << *Res << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " << *getCoercedLoadValue( ) << '\n' << *Res << '\n' << "\n\n\n" ; } } while (false) |
910 | << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " << *getCoercedLoadValue( ) << '\n' << *Res << '\n' << "\n\n\n" ; } } while (false); |
911 | } |
912 | } else if (isMemIntrinValue()) { |
913 | Res = getMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy, |
914 | InsertPt, DL); |
915 | LLVM_DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offsetdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset << " " << *getMemIntrinValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
916 | << " " << *getMemIntrinValue() << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset << " " << *getMemIntrinValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
917 | << *Res << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset << " " << *getMemIntrinValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false) |
918 | << "\n\n\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset << " " << *getMemIntrinValue() << '\n' << *Res << '\n' << "\n\n\n"; } } while (false); |
919 | } else { |
920 | llvm_unreachable("Should not materialize value from dead block")::llvm::llvm_unreachable_internal("Should not materialize value from dead block" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 920); |
921 | } |
922 | assert(Res && "failed to materialize?")(static_cast <bool> (Res && "failed to materialize?" ) ? void (0) : __assert_fail ("Res && \"failed to materialize?\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 922, __extension__ __PRETTY_FUNCTION__)); |
923 | return Res; |
924 | } |
925 | |
926 | static bool isLifetimeStart(const Instruction *Inst) { |
927 | if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) |
928 | return II->getIntrinsicID() == Intrinsic::lifetime_start; |
929 | return false; |
930 | } |
931 | |
932 | /// Assuming To can be reached from both From and Between, does Between lie on |
933 | /// every path from From to To? |
934 | static bool liesBetween(const Instruction *From, Instruction *Between, |
935 | const Instruction *To, DominatorTree *DT) { |
936 | if (From->getParent() == Between->getParent()) |
937 | return DT->dominates(From, Between); |
938 | SmallSet<BasicBlock *, 1> Exclusion; |
939 | Exclusion.insert(Between->getParent()); |
940 | return !isPotentiallyReachable(From, To, &Exclusion, DT); |
941 | } |
942 | |
943 | /// Try to locate the three instruction involved in a missed |
944 | /// load-elimination case that is due to an intervening store. |
945 | static void reportMayClobberedLoad(LoadInst *Load, MemDepResult DepInfo, |
946 | DominatorTree *DT, |
947 | OptimizationRemarkEmitter *ORE) { |
948 | using namespace ore; |
949 | |
950 | User *OtherAccess = nullptr; |
951 | |
952 | OptimizationRemarkMissed R(DEBUG_TYPE"gvn", "LoadClobbered", Load); |
953 | R << "load of type " << NV("Type", Load->getType()) << " not eliminated" |
954 | << setExtraArgs(); |
955 | |
956 | for (auto *U : Load->getPointerOperand()->users()) { |
957 | if (U != Load && (isa<LoadInst>(U) || isa<StoreInst>(U)) && |
958 | cast<Instruction>(U)->getFunction() == Load->getFunction() && |
959 | DT->dominates(cast<Instruction>(U), Load)) { |
960 | // Use the most immediately dominating value |
961 | if (OtherAccess) { |
962 | if (DT->dominates(cast<Instruction>(OtherAccess), cast<Instruction>(U))) |
963 | OtherAccess = U; |
964 | else |
965 | assert(DT->dominates(cast<Instruction>(U),(static_cast <bool> (DT->dominates(cast<Instruction >(U), cast<Instruction>(OtherAccess))) ? void (0) : __assert_fail ("DT->dominates(cast<Instruction>(U), cast<Instruction>(OtherAccess))" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 966, __extension__ __PRETTY_FUNCTION__)) |
966 | cast<Instruction>(OtherAccess)))(static_cast <bool> (DT->dominates(cast<Instruction >(U), cast<Instruction>(OtherAccess))) ? void (0) : __assert_fail ("DT->dominates(cast<Instruction>(U), cast<Instruction>(OtherAccess))" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 966, __extension__ __PRETTY_FUNCTION__)); |
967 | } else |
968 | OtherAccess = U; |
969 | } |
970 | } |
971 | |
972 | if (!OtherAccess) { |
973 | // There is no dominating use, check if we can find a closest non-dominating |
974 | // use that lies between any other potentially available use and Load. |
975 | for (auto *U : Load->getPointerOperand()->users()) { |
976 | if (U != Load && (isa<LoadInst>(U) || isa<StoreInst>(U)) && |
977 | cast<Instruction>(U)->getFunction() == Load->getFunction() && |
978 | isPotentiallyReachable(cast<Instruction>(U), Load, nullptr, DT)) { |
979 | if (OtherAccess) { |
980 | if (liesBetween(cast<Instruction>(OtherAccess), cast<Instruction>(U), |
981 | Load, DT)) { |
982 | OtherAccess = U; |
983 | } else if (!liesBetween(cast<Instruction>(U), |
984 | cast<Instruction>(OtherAccess), Load, DT)) { |
985 | // These uses are both partially available at Load were it not for |
986 | // the clobber, but neither lies strictly after the other. |
987 | OtherAccess = nullptr; |
988 | break; |
989 | } // else: keep current OtherAccess since it lies between U and Load |
990 | } else { |
991 | OtherAccess = U; |
992 | } |
993 | } |
994 | } |
995 | } |
996 | |
997 | if (OtherAccess) |
998 | R << " in favor of " << NV("OtherAccess", OtherAccess); |
999 | |
1000 | R << " because it is clobbered by " << NV("ClobberedBy", DepInfo.getInst()); |
1001 | |
1002 | ORE->emit(R); |
1003 | } |
1004 | |
1005 | bool GVN::AnalyzeLoadAvailability(LoadInst *Load, MemDepResult DepInfo, |
1006 | Value *Address, AvailableValue &Res) { |
1007 | assert((DepInfo.isDef() || DepInfo.isClobber()) &&(static_cast <bool> ((DepInfo.isDef() || DepInfo.isClobber ()) && "expected a local dependence") ? void (0) : __assert_fail ("(DepInfo.isDef() || DepInfo.isClobber()) && \"expected a local dependence\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1008, __extension__ __PRETTY_FUNCTION__)) |
1008 | "expected a local dependence")(static_cast <bool> ((DepInfo.isDef() || DepInfo.isClobber ()) && "expected a local dependence") ? void (0) : __assert_fail ("(DepInfo.isDef() || DepInfo.isClobber()) && \"expected a local dependence\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1008, __extension__ __PRETTY_FUNCTION__)); |
1009 | assert(Load->isUnordered() && "rules below are incorrect for ordered access")(static_cast <bool> (Load->isUnordered() && "rules below are incorrect for ordered access" ) ? void (0) : __assert_fail ("Load->isUnordered() && \"rules below are incorrect for ordered access\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1009, __extension__ __PRETTY_FUNCTION__)); |
1010 | |
1011 | const DataLayout &DL = Load->getModule()->getDataLayout(); |
1012 | |
1013 | Instruction *DepInst = DepInfo.getInst(); |
1014 | if (DepInfo.isClobber()) { |
1015 | // If the dependence is to a store that writes to a superset of the bits |
1016 | // read by the load, we can extract the bits we need for the load from the |
1017 | // stored value. |
1018 | if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { |
1019 | // Can't forward from non-atomic to atomic without violating memory model. |
1020 | if (Address && Load->isAtomic() <= DepSI->isAtomic()) { |
1021 | int Offset = |
1022 | analyzeLoadFromClobberingStore(Load->getType(), Address, DepSI, DL); |
1023 | if (Offset != -1) { |
1024 | Res = AvailableValue::get(DepSI->getValueOperand(), Offset); |
1025 | return true; |
1026 | } |
1027 | } |
1028 | } |
1029 | |
1030 | // Check to see if we have something like this: |
1031 | // load i32* P |
1032 | // load i8* (P+1) |
1033 | // if we have this, replace the later with an extraction from the former. |
1034 | if (LoadInst *DepLoad = dyn_cast<LoadInst>(DepInst)) { |
1035 | // If this is a clobber and L is the first instruction in its block, then |
1036 | // we have the first instruction in the entry block. |
1037 | // Can't forward from non-atomic to atomic without violating memory model. |
1038 | if (DepLoad != Load && Address && |
1039 | Load->isAtomic() <= DepLoad->isAtomic()) { |
1040 | Type *LoadType = Load->getType(); |
1041 | int Offset = -1; |
1042 | |
1043 | // If MD reported clobber, check it was nested. |
1044 | if (DepInfo.isClobber() && |
1045 | canCoerceMustAliasedValueToLoad(DepLoad, LoadType, DL)) { |
1046 | const auto ClobberOff = MD->getClobberOffset(DepLoad); |
1047 | // GVN has no deal with a negative offset. |
1048 | Offset = (ClobberOff == None || ClobberOff.getValue() < 0) |
1049 | ? -1 |
1050 | : ClobberOff.getValue(); |
1051 | } |
1052 | if (Offset == -1) |
1053 | Offset = |
1054 | analyzeLoadFromClobberingLoad(LoadType, Address, DepLoad, DL); |
1055 | if (Offset != -1) { |
1056 | Res = AvailableValue::getLoad(DepLoad, Offset); |
1057 | return true; |
1058 | } |
1059 | } |
1060 | } |
1061 | |
1062 | // If the clobbering value is a memset/memcpy/memmove, see if we can |
1063 | // forward a value on from it. |
1064 | if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInst)) { |
1065 | if (Address && !Load->isAtomic()) { |
1066 | int Offset = analyzeLoadFromClobberingMemInst(Load->getType(), Address, |
1067 | DepMI, DL); |
1068 | if (Offset != -1) { |
1069 | Res = AvailableValue::getMI(DepMI, Offset); |
1070 | return true; |
1071 | } |
1072 | } |
1073 | } |
1074 | // Nothing known about this clobber, have to be conservative |
1075 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " is clobbered by " << *DepInst << '\n';; } } while (false) |
1076 | // fast print dep, using operator<< on instruction is too slow.do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " is clobbered by " << *DepInst << '\n';; } } while (false) |
1077 | dbgs() << "GVN: load "; Load->printAsOperand(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " is clobbered by " << *DepInst << '\n';; } } while (false) |
1078 | dbgs() << " is clobbered by " << *DepInst << '\n';)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " is clobbered by " << *DepInst << '\n';; } } while (false); |
1079 | if (ORE->allowExtraAnalysis(DEBUG_TYPE"gvn")) |
1080 | reportMayClobberedLoad(Load, DepInfo, DT, ORE); |
1081 | |
1082 | return false; |
1083 | } |
1084 | assert(DepInfo.isDef() && "follows from above")(static_cast <bool> (DepInfo.isDef() && "follows from above" ) ? void (0) : __assert_fail ("DepInfo.isDef() && \"follows from above\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1084, __extension__ __PRETTY_FUNCTION__)); |
1085 | |
1086 | // Loading the allocation -> undef. |
1087 | if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) || |
1088 | isAlignedAllocLikeFn(DepInst, TLI) || |
1089 | // Loading immediately after lifetime begin -> undef. |
1090 | isLifetimeStart(DepInst)) { |
1091 | Res = AvailableValue::get(UndefValue::get(Load->getType())); |
1092 | return true; |
1093 | } |
1094 | |
1095 | // Loading from calloc (which zero initializes memory) -> zero |
1096 | if (isCallocLikeFn(DepInst, TLI)) { |
1097 | Res = AvailableValue::get(Constant::getNullValue(Load->getType())); |
1098 | return true; |
1099 | } |
1100 | |
1101 | if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { |
1102 | // Reject loads and stores that are to the same address but are of |
1103 | // different types if we have to. If the stored value is convertable to |
1104 | // the loaded value, we can reuse it. |
1105 | if (!canCoerceMustAliasedValueToLoad(S->getValueOperand(), Load->getType(), |
1106 | DL)) |
1107 | return false; |
1108 | |
1109 | // Can't forward from non-atomic to atomic without violating memory model. |
1110 | if (S->isAtomic() < Load->isAtomic()) |
1111 | return false; |
1112 | |
1113 | Res = AvailableValue::get(S->getValueOperand()); |
1114 | return true; |
1115 | } |
1116 | |
1117 | if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { |
1118 | // If the types mismatch and we can't handle it, reject reuse of the load. |
1119 | // If the stored value is larger or equal to the loaded value, we can reuse |
1120 | // it. |
1121 | if (!canCoerceMustAliasedValueToLoad(LD, Load->getType(), DL)) |
1122 | return false; |
1123 | |
1124 | // Can't forward from non-atomic to atomic without violating memory model. |
1125 | if (LD->isAtomic() < Load->isAtomic()) |
1126 | return false; |
1127 | |
1128 | Res = AvailableValue::getLoad(LD); |
1129 | return true; |
1130 | } |
1131 | |
1132 | // Unknown def - must be conservative |
1133 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown def " << *DepInst << '\n';; } } while (false) |
1134 | // fast print dep, using operator<< on instruction is too slow.do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown def " << *DepInst << '\n';; } } while (false) |
1135 | dbgs() << "GVN: load "; Load->printAsOperand(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown def " << *DepInst << '\n';; } } while (false) |
1136 | dbgs() << " has unknown def " << *DepInst << '\n';)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown def " << *DepInst << '\n';; } } while (false); |
1137 | return false; |
1138 | } |
1139 | |
1140 | void GVN::AnalyzeLoadAvailability(LoadInst *Load, LoadDepVect &Deps, |
1141 | AvailValInBlkVect &ValuesPerBlock, |
1142 | UnavailBlkVect &UnavailableBlocks) { |
1143 | // Filter out useless results (non-locals, etc). Keep track of the blocks |
1144 | // where we have a value available in repl, also keep track of whether we see |
1145 | // dependencies that produce an unknown value for the load (such as a call |
1146 | // that could potentially clobber the load). |
1147 | unsigned NumDeps = Deps.size(); |
1148 | for (unsigned i = 0, e = NumDeps; i != e; ++i) { |
1149 | BasicBlock *DepBB = Deps[i].getBB(); |
1150 | MemDepResult DepInfo = Deps[i].getResult(); |
1151 | |
1152 | if (DeadBlocks.count(DepBB)) { |
1153 | // Dead dependent mem-op disguise as a load evaluating the same value |
1154 | // as the load in question. |
1155 | ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB)); |
1156 | continue; |
1157 | } |
1158 | |
1159 | if (!DepInfo.isDef() && !DepInfo.isClobber()) { |
1160 | UnavailableBlocks.push_back(DepBB); |
1161 | continue; |
1162 | } |
1163 | |
1164 | // The address being loaded in this non-local block may not be the same as |
1165 | // the pointer operand of the load if PHI translation occurs. Make sure |
1166 | // to consider the right address. |
1167 | Value *Address = Deps[i].getAddress(); |
1168 | |
1169 | AvailableValue AV; |
1170 | if (AnalyzeLoadAvailability(Load, DepInfo, Address, AV)) { |
1171 | // subtlety: because we know this was a non-local dependency, we know |
1172 | // it's safe to materialize anywhere between the instruction within |
1173 | // DepInfo and the end of it's block. |
1174 | ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, |
1175 | std::move(AV))); |
1176 | } else { |
1177 | UnavailableBlocks.push_back(DepBB); |
1178 | } |
1179 | } |
1180 | |
1181 | assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&(static_cast <bool> (NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() && "post condition violation" ) ? void (0) : __assert_fail ("NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() && \"post condition violation\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1182, __extension__ __PRETTY_FUNCTION__)) |
1182 | "post condition violation")(static_cast <bool> (NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() && "post condition violation" ) ? void (0) : __assert_fail ("NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() && \"post condition violation\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1182, __extension__ __PRETTY_FUNCTION__)); |
1183 | } |
1184 | |
1185 | void GVN::eliminatePartiallyRedundantLoad( |
1186 | LoadInst *Load, AvailValInBlkVect &ValuesPerBlock, |
1187 | MapVector<BasicBlock *, Value *> &AvailableLoads) { |
1188 | for (const auto &AvailableLoad : AvailableLoads) { |
1189 | BasicBlock *UnavailableBlock = AvailableLoad.first; |
1190 | Value *LoadPtr = AvailableLoad.second; |
1191 | |
1192 | auto *NewLoad = |
1193 | new LoadInst(Load->getType(), LoadPtr, Load->getName() + ".pre", |
1194 | Load->isVolatile(), Load->getAlign(), Load->getOrdering(), |
1195 | Load->getSyncScopeID(), UnavailableBlock->getTerminator()); |
1196 | NewLoad->setDebugLoc(Load->getDebugLoc()); |
1197 | if (MSSAU) { |
1198 | auto *MSSA = MSSAU->getMemorySSA(); |
1199 | // Get the defining access of the original load or use the load if it is a |
1200 | // MemoryDef (e.g. because it is volatile). The inserted loads are |
1201 | // guaranteed to load from the same definition. |
1202 | auto *LoadAcc = MSSA->getMemoryAccess(Load); |
1203 | auto *DefiningAcc = |
1204 | isa<MemoryDef>(LoadAcc) ? LoadAcc : LoadAcc->getDefiningAccess(); |
1205 | auto *NewAccess = MSSAU->createMemoryAccessInBB( |
1206 | NewLoad, DefiningAcc, NewLoad->getParent(), |
1207 | MemorySSA::BeforeTerminator); |
1208 | if (auto *NewDef = dyn_cast<MemoryDef>(NewAccess)) |
1209 | MSSAU->insertDef(NewDef, /*RenameUses=*/true); |
1210 | else |
1211 | MSSAU->insertUse(cast<MemoryUse>(NewAccess), /*RenameUses=*/true); |
1212 | } |
1213 | |
1214 | // Transfer the old load's AA tags to the new load. |
1215 | AAMDNodes Tags; |
1216 | Load->getAAMetadata(Tags); |
1217 | if (Tags) |
1218 | NewLoad->setAAMetadata(Tags); |
1219 | |
1220 | if (auto *MD = Load->getMetadata(LLVMContext::MD_invariant_load)) |
1221 | NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD); |
1222 | if (auto *InvGroupMD = Load->getMetadata(LLVMContext::MD_invariant_group)) |
1223 | NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD); |
1224 | if (auto *RangeMD = Load->getMetadata(LLVMContext::MD_range)) |
1225 | NewLoad->setMetadata(LLVMContext::MD_range, RangeMD); |
1226 | if (auto *AccessMD = Load->getMetadata(LLVMContext::MD_access_group)) |
1227 | if (LI && |
1228 | LI->getLoopFor(Load->getParent()) == LI->getLoopFor(UnavailableBlock)) |
1229 | NewLoad->setMetadata(LLVMContext::MD_access_group, AccessMD); |
1230 | |
1231 | // We do not propagate the old load's debug location, because the new |
1232 | // load now lives in a different BB, and we want to avoid a jumpy line |
1233 | // table. |
1234 | // FIXME: How do we retain source locations without causing poor debugging |
1235 | // behavior? |
1236 | |
1237 | // Add the newly created load. |
1238 | ValuesPerBlock.push_back( |
1239 | AvailableValueInBlock::get(UnavailableBlock, NewLoad)); |
1240 | MD->invalidateCachedPointerInfo(LoadPtr); |
1241 | LLVM_DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN INSERTED " << *NewLoad << '\n'; } } while (false); |
1242 | } |
1243 | |
1244 | // Perform PHI construction. |
1245 | Value *V = ConstructSSAForLoadSet(Load, ValuesPerBlock, *this); |
1246 | Load->replaceAllUsesWith(V); |
1247 | if (isa<PHINode>(V)) |
1248 | V->takeName(Load); |
1249 | if (Instruction *I = dyn_cast<Instruction>(V)) |
1250 | I->setDebugLoc(Load->getDebugLoc()); |
1251 | if (V->getType()->isPtrOrPtrVectorTy()) |
1252 | MD->invalidateCachedPointerInfo(V); |
1253 | markInstructionForDeletion(Load); |
1254 | ORE->emit([&]() { |
1255 | return OptimizationRemark(DEBUG_TYPE"gvn", "LoadPRE", Load) |
1256 | << "load eliminated by PRE"; |
1257 | }); |
1258 | } |
1259 | |
1260 | bool GVN::PerformLoadPRE(LoadInst *Load, AvailValInBlkVect &ValuesPerBlock, |
1261 | UnavailBlkVect &UnavailableBlocks) { |
1262 | // Okay, we have *some* definitions of the value. This means that the value |
1263 | // is available in some of our (transitive) predecessors. Lets think about |
1264 | // doing PRE of this load. This will involve inserting a new load into the |
1265 | // predecessor when it's not available. We could do this in general, but |
1266 | // prefer to not increase code size. As such, we only do this when we know |
1267 | // that we only have to insert *one* load (which means we're basically moving |
1268 | // the load, not inserting a new one). |
1269 | |
1270 | SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(), |
1271 | UnavailableBlocks.end()); |
1272 | |
1273 | // Let's find the first basic block with more than one predecessor. Walk |
1274 | // backwards through predecessors if needed. |
1275 | BasicBlock *LoadBB = Load->getParent(); |
1276 | BasicBlock *TmpBB = LoadBB; |
1277 | |
1278 | // Check that there is no implicit control flow instructions above our load in |
1279 | // its block. If there is an instruction that doesn't always pass the |
1280 | // execution to the following instruction, then moving through it may become |
1281 | // invalid. For example: |
1282 | // |
1283 | // int arr[LEN]; |
1284 | // int index = ???; |
1285 | // ... |
1286 | // guard(0 <= index && index < LEN); |
1287 | // use(arr[index]); |
1288 | // |
1289 | // It is illegal to move the array access to any point above the guard, |
1290 | // because if the index is out of bounds we should deoptimize rather than |
1291 | // access the array. |
1292 | // Check that there is no guard in this block above our instruction. |
1293 | bool MustEnsureSafetyOfSpeculativeExecution = |
1294 | ICF->isDominatedByICFIFromSameBlock(Load); |
1295 | |
1296 | while (TmpBB->getSinglePredecessor()) { |
1297 | TmpBB = TmpBB->getSinglePredecessor(); |
1298 | if (TmpBB == LoadBB) // Infinite (unreachable) loop. |
1299 | return false; |
1300 | if (Blockers.count(TmpBB)) |
1301 | return false; |
1302 | |
1303 | // If any of these blocks has more than one successor (i.e. if the edge we |
1304 | // just traversed was critical), then there are other paths through this |
1305 | // block along which the load may not be anticipated. Hoisting the load |
1306 | // above this block would be adding the load to execution paths along |
1307 | // which it was not previously executed. |
1308 | if (TmpBB->getTerminator()->getNumSuccessors() != 1) |
1309 | return false; |
1310 | |
1311 | // Check that there is no implicit control flow in a block above. |
1312 | MustEnsureSafetyOfSpeculativeExecution = |
1313 | MustEnsureSafetyOfSpeculativeExecution || ICF->hasICF(TmpBB); |
1314 | } |
1315 | |
1316 | assert(TmpBB)(static_cast <bool> (TmpBB) ? void (0) : __assert_fail ( "TmpBB", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1316, __extension__ __PRETTY_FUNCTION__)); |
1317 | LoadBB = TmpBB; |
1318 | |
1319 | // Check to see how many predecessors have the loaded value fully |
1320 | // available. |
1321 | MapVector<BasicBlock *, Value *> PredLoads; |
1322 | DenseMap<BasicBlock *, AvailabilityState> FullyAvailableBlocks; |
1323 | for (const AvailableValueInBlock &AV : ValuesPerBlock) |
1324 | FullyAvailableBlocks[AV.BB] = AvailabilityState::Available; |
1325 | for (BasicBlock *UnavailableBB : UnavailableBlocks) |
1326 | FullyAvailableBlocks[UnavailableBB] = AvailabilityState::Unavailable; |
1327 | |
1328 | SmallVector<BasicBlock *, 4> CriticalEdgePred; |
1329 | for (BasicBlock *Pred : predecessors(LoadBB)) { |
1330 | // If any predecessor block is an EH pad that does not allow non-PHI |
1331 | // instructions before the terminator, we can't PRE the load. |
1332 | if (Pred->getTerminator()->isEHPad()) { |
1333 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1334 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1335 | << Pred->getName() << "': " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '" << Pred->getName() << "': " << *Load << '\n'; } } while (false); |
1336 | return false; |
1337 | } |
1338 | |
1339 | if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { |
1340 | continue; |
1341 | } |
1342 | |
1343 | if (Pred->getTerminator()->getNumSuccessors() != 1) { |
1344 | if (isa<IndirectBrInst>(Pred->getTerminator())) { |
1345 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1346 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1347 | << Pred->getName() << "': " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false); |
1348 | return false; |
1349 | } |
1350 | |
1351 | // FIXME: Can we support the fallthrough edge? |
1352 | if (isa<CallBrInst>(Pred->getTerminator())) { |
1353 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1354 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1355 | << Pred->getName() << "': " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF CALLBR CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false); |
1356 | return false; |
1357 | } |
1358 | |
1359 | if (LoadBB->isEHPad()) { |
1360 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1361 | dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1362 | << Pred->getName() << "': " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false); |
1363 | return false; |
1364 | } |
1365 | |
1366 | // Do not split backedge as it will break the canonical loop form. |
1367 | if (!isLoadPRESplitBackedgeEnabled()) |
1368 | if (DT->dominates(LoadBB, Pred)) { |
1369 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1370 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1371 | << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false) |
1372 | << Pred->getName() << "': " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULD NOT PRE LOAD BECAUSE OF A BACKEDGE CRITICAL EDGE '" << Pred->getName() << "': " << *Load << '\n'; } } while (false); |
1373 | return false; |
1374 | } |
1375 | |
1376 | CriticalEdgePred.push_back(Pred); |
1377 | } else { |
1378 | // Only add the predecessors that will not be split for now. |
1379 | PredLoads[Pred] = nullptr; |
1380 | } |
1381 | } |
1382 | |
1383 | // Decide whether PRE is profitable for this load. |
1384 | unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size(); |
1385 | assert(NumUnavailablePreds != 0 &&(static_cast <bool> (NumUnavailablePreds != 0 && "Fully available value should already be eliminated!") ? void (0) : __assert_fail ("NumUnavailablePreds != 0 && \"Fully available value should already be eliminated!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1386, __extension__ __PRETTY_FUNCTION__)) |
1386 | "Fully available value should already be eliminated!")(static_cast <bool> (NumUnavailablePreds != 0 && "Fully available value should already be eliminated!") ? void (0) : __assert_fail ("NumUnavailablePreds != 0 && \"Fully available value should already be eliminated!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1386, __extension__ __PRETTY_FUNCTION__)); |
1387 | |
1388 | // If this load is unavailable in multiple predecessors, reject it. |
1389 | // FIXME: If we could restructure the CFG, we could make a common pred with |
1390 | // all the preds that don't have an available Load and insert a new load into |
1391 | // that one block. |
1392 | if (NumUnavailablePreds != 1) |
1393 | return false; |
1394 | |
1395 | // Now we know where we will insert load. We must ensure that it is safe |
1396 | // to speculatively execute the load at that points. |
1397 | if (MustEnsureSafetyOfSpeculativeExecution) { |
1398 | if (CriticalEdgePred.size()) |
1399 | if (!isSafeToSpeculativelyExecute(Load, LoadBB->getFirstNonPHI(), DT)) |
1400 | return false; |
1401 | for (auto &PL : PredLoads) |
1402 | if (!isSafeToSpeculativelyExecute(Load, PL.first->getTerminator(), DT)) |
1403 | return false; |
1404 | } |
1405 | |
1406 | // Split critical edges, and update the unavailable predecessors accordingly. |
1407 | for (BasicBlock *OrigPred : CriticalEdgePred) { |
1408 | BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB); |
1409 | assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!")(static_cast <bool> (!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!") ? void (0) : __assert_fail ("!PredLoads.count(OrigPred) && \"Split edges shouldn't be in map!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 1409, __extension__ __PRETTY_FUNCTION__)); |
1410 | PredLoads[NewPred] = nullptr; |
1411 | LLVM_DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Split critical edge " << OrigPred ->getName() << "->" << LoadBB->getName() << '\n'; } } while (false) |
1412 | << LoadBB->getName() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Split critical edge " << OrigPred ->getName() << "->" << LoadBB->getName() << '\n'; } } while (false); |
1413 | } |
1414 | |
1415 | // Check if the load can safely be moved to all the unavailable predecessors. |
1416 | bool CanDoPRE = true; |
1417 | const DataLayout &DL = Load->getModule()->getDataLayout(); |
1418 | SmallVector<Instruction*, 8> NewInsts; |
1419 | for (auto &PredLoad : PredLoads) { |
1420 | BasicBlock *UnavailablePred = PredLoad.first; |
1421 | |
1422 | // Do PHI translation to get its value in the predecessor if necessary. The |
1423 | // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. |
1424 | // We do the translation for each edge we skipped by going from Load's block |
1425 | // to LoadBB, otherwise we might miss pieces needing translation. |
1426 | |
1427 | // If all preds have a single successor, then we know it is safe to insert |
1428 | // the load on the pred (?!?), so we can insert code to materialize the |
1429 | // pointer if it is not available. |
1430 | Value *LoadPtr = Load->getPointerOperand(); |
1431 | BasicBlock *Cur = Load->getParent(); |
1432 | while (Cur != LoadBB) { |
1433 | PHITransAddr Address(LoadPtr, DL, AC); |
1434 | LoadPtr = Address.PHITranslateWithInsertion( |
1435 | Cur, Cur->getSinglePredecessor(), *DT, NewInsts); |
1436 | if (!LoadPtr) { |
1437 | CanDoPRE = false; |
1438 | break; |
1439 | } |
1440 | Cur = Cur->getSinglePredecessor(); |
1441 | } |
1442 | |
1443 | if (LoadPtr) { |
1444 | PHITransAddr Address(LoadPtr, DL, AC); |
1445 | LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, *DT, |
1446 | NewInsts); |
1447 | } |
1448 | // If we couldn't find or insert a computation of this phi translated value, |
1449 | // we fail PRE. |
1450 | if (!LoadPtr) { |
1451 | LLVM_DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " << *Load->getPointerOperand() << "\n"; } } while (false) |
1452 | << *Load->getPointerOperand() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " << *Load->getPointerOperand() << "\n"; } } while (false); |
1453 | CanDoPRE = false; |
1454 | break; |
1455 | } |
1456 | |
1457 | PredLoad.second = LoadPtr; |
1458 | } |
1459 | |
1460 | if (!CanDoPRE) { |
1461 | while (!NewInsts.empty()) { |
1462 | // Erase instructions generated by the failed PHI translation before |
1463 | // trying to number them. PHI translation might insert instructions |
1464 | // in basic blocks other than the current one, and we delete them |
1465 | // directly, as markInstructionForDeletion only allows removing from the |
1466 | // current basic block. |
1467 | NewInsts.pop_back_val()->eraseFromParent(); |
1468 | } |
1469 | // HINT: Don't revert the edge-splitting as following transformation may |
1470 | // also need to split these critical edges. |
1471 | return !CriticalEdgePred.empty(); |
1472 | } |
1473 | |
1474 | // Okay, we can eliminate this load by inserting a reload in the predecessor |
1475 | // and using PHI construction to get the value in the other predecessors, do |
1476 | // it. |
1477 | LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN REMOVING PRE LOAD: " << *Load << '\n'; } } while (false); |
1478 | LLVM_DEBUG(if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts .back() << '\n'; } } while (false) |
1479 | << " INSTS: " << *NewInsts.back()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts .back() << '\n'; } } while (false) |
1480 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { if (!NewInsts.empty()) dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " << *NewInsts .back() << '\n'; } } while (false); |
1481 | |
1482 | // Assign value numbers to the new instructions. |
1483 | for (Instruction *I : NewInsts) { |
1484 | // Instructions that have been inserted in predecessor(s) to materialize |
1485 | // the load address do not retain their original debug locations. Doing |
1486 | // so could lead to confusing (but correct) source attributions. |
1487 | I->updateLocationAfterHoist(); |
1488 | |
1489 | // FIXME: We really _ought_ to insert these value numbers into their |
1490 | // parent's availability map. However, in doing so, we risk getting into |
1491 | // ordering issues. If a block hasn't been processed yet, we would be |
1492 | // marking a value as AVAIL-IN, which isn't what we intend. |
1493 | VN.lookupOrAdd(I); |
1494 | } |
1495 | |
1496 | eliminatePartiallyRedundantLoad(Load, ValuesPerBlock, PredLoads); |
1497 | ++NumPRELoad; |
1498 | return true; |
1499 | } |
1500 | |
1501 | bool GVN::performLoopLoadPRE(LoadInst *Load, AvailValInBlkVect &ValuesPerBlock, |
1502 | UnavailBlkVect &UnavailableBlocks) { |
1503 | if (!LI) |
1504 | return false; |
1505 | |
1506 | const Loop *L = LI->getLoopFor(Load->getParent()); |
1507 | // TODO: Generalize to other loop blocks that dominate the latch. |
1508 | if (!L || L->getHeader() != Load->getParent()) |
1509 | return false; |
1510 | |
1511 | BasicBlock *Preheader = L->getLoopPreheader(); |
1512 | BasicBlock *Latch = L->getLoopLatch(); |
1513 | if (!Preheader || !Latch) |
1514 | return false; |
1515 | |
1516 | Value *LoadPtr = Load->getPointerOperand(); |
1517 | // Must be available in preheader. |
1518 | if (!L->isLoopInvariant(LoadPtr)) |
1519 | return false; |
1520 | |
1521 | // We plan to hoist the load to preheader without introducing a new fault. |
1522 | // In order to do it, we need to prove that we cannot side-exit the loop |
1523 | // once loop header is first entered before execution of the load. |
1524 | if (ICF->isDominatedByICFIFromSameBlock(Load)) |
1525 | return false; |
1526 | |
1527 | BasicBlock *LoopBlock = nullptr; |
1528 | for (auto *Blocker : UnavailableBlocks) { |
1529 | // Blockers from outside the loop are handled in preheader. |
1530 | if (!L->contains(Blocker)) |
1531 | continue; |
1532 | |
1533 | // Only allow one loop block. Loop header is not less frequently executed |
1534 | // than each loop block, and likely it is much more frequently executed. But |
1535 | // in case of multiple loop blocks, we need extra information (such as block |
1536 | // frequency info) to understand whether it is profitable to PRE into |
1537 | // multiple loop blocks. |
1538 | if (LoopBlock) |
1539 | return false; |
1540 | |
1541 | // Do not sink into inner loops. This may be non-profitable. |
1542 | if (L != LI->getLoopFor(Blocker)) |
1543 | return false; |
1544 | |
1545 | // Blocks that dominate the latch execute on every single iteration, maybe |
1546 | // except the last one. So PREing into these blocks doesn't make much sense |
1547 | // in most cases. But the blocks that do not necessarily execute on each |
1548 | // iteration are sometimes much colder than the header, and this is when |
1549 | // PRE is potentially profitable. |
1550 | if (DT->dominates(Blocker, Latch)) |
1551 | return false; |
1552 | |
1553 | // Make sure that the terminator itself doesn't clobber. |
1554 | if (Blocker->getTerminator()->mayWriteToMemory()) |
1555 | return false; |
1556 | |
1557 | LoopBlock = Blocker; |
1558 | } |
1559 | |
1560 | if (!LoopBlock) |
1561 | return false; |
1562 | |
1563 | // Make sure the memory at this pointer cannot be freed, therefore we can |
1564 | // safely reload from it after clobber. |
1565 | if (LoadPtr->canBeFreed()) |
1566 | return false; |
1567 | |
1568 | // TODO: Support critical edge splitting if blocker has more than 1 successor. |
1569 | MapVector<BasicBlock *, Value *> AvailableLoads; |
1570 | AvailableLoads[LoopBlock] = LoadPtr; |
1571 | AvailableLoads[Preheader] = LoadPtr; |
1572 | |
1573 | LLVM_DEBUG(dbgs() << "GVN REMOVING PRE LOOP LOAD: " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN REMOVING PRE LOOP LOAD: " << *Load << '\n'; } } while (false); |
1574 | eliminatePartiallyRedundantLoad(Load, ValuesPerBlock, AvailableLoads); |
1575 | ++NumPRELoopLoad; |
1576 | return true; |
1577 | } |
1578 | |
1579 | static void reportLoadElim(LoadInst *Load, Value *AvailableValue, |
1580 | OptimizationRemarkEmitter *ORE) { |
1581 | using namespace ore; |
1582 | |
1583 | ORE->emit([&]() { |
1584 | return OptimizationRemark(DEBUG_TYPE"gvn", "LoadElim", Load) |
1585 | << "load of type " << NV("Type", Load->getType()) << " eliminated" |
1586 | << setExtraArgs() << " in favor of " |
1587 | << NV("InfavorOfValue", AvailableValue); |
1588 | }); |
1589 | } |
1590 | |
1591 | /// Attempt to eliminate a load whose dependencies are |
1592 | /// non-local by performing PHI construction. |
1593 | bool GVN::processNonLocalLoad(LoadInst *Load) { |
1594 | // non-local speculations are not allowed under asan. |
1595 | if (Load->getParent()->getParent()->hasFnAttribute( |
1596 | Attribute::SanitizeAddress) || |
1597 | Load->getParent()->getParent()->hasFnAttribute( |
1598 | Attribute::SanitizeHWAddress)) |
1599 | return false; |
1600 | |
1601 | // Step 1: Find the non-local dependencies of the load. |
1602 | LoadDepVect Deps; |
1603 | MD->getNonLocalPointerDependency(Load, Deps); |
1604 | |
1605 | // If we had to process more than one hundred blocks to find the |
1606 | // dependencies, this load isn't worth worrying about. Optimizing |
1607 | // it will be too expensive. |
1608 | unsigned NumDeps = Deps.size(); |
1609 | if (NumDeps > MaxNumDeps) |
1610 | return false; |
1611 | |
1612 | // If we had a phi translation failure, we'll have a single entry which is a |
1613 | // clobber in the current block. Reject this early. |
1614 | if (NumDeps == 1 && |
1615 | !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) { |
1616 | LLVM_DEBUG(dbgs() << "GVN: non-local load "; Load->printAsOperand(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: non-local load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown dependencies\n";; } } while (false) |
1617 | dbgs() << " has unknown dependencies\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: non-local load "; Load->printAsOperand (dbgs()); dbgs() << " has unknown dependencies\n";; } } while (false); |
1618 | return false; |
1619 | } |
1620 | |
1621 | bool Changed = false; |
1622 | // If this load follows a GEP, see if we can PRE the indices before analyzing. |
1623 | if (GetElementPtrInst *GEP = |
1624 | dyn_cast<GetElementPtrInst>(Load->getOperand(0))) { |
1625 | for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(), |
1626 | OE = GEP->idx_end(); |
1627 | OI != OE; ++OI) |
1628 | if (Instruction *I = dyn_cast<Instruction>(OI->get())) |
1629 | Changed |= performScalarPRE(I); |
1630 | } |
1631 | |
1632 | // Step 2: Analyze the availability of the load |
1633 | AvailValInBlkVect ValuesPerBlock; |
1634 | UnavailBlkVect UnavailableBlocks; |
1635 | AnalyzeLoadAvailability(Load, Deps, ValuesPerBlock, UnavailableBlocks); |
1636 | |
1637 | // If we have no predecessors that produce a known value for this load, exit |
1638 | // early. |
1639 | if (ValuesPerBlock.empty()) |
1640 | return Changed; |
1641 | |
1642 | // Step 3: Eliminate fully redundancy. |
1643 | // |
1644 | // If all of the instructions we depend on produce a known value for this |
1645 | // load, then it is fully redundant and we can use PHI insertion to compute |
1646 | // its value. Insert PHIs and remove the fully redundant value now. |
1647 | if (UnavailableBlocks.empty()) { |
1648 | LLVM_DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *Load << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *Load << '\n'; } } while (false); |
1649 | |
1650 | // Perform PHI construction. |
1651 | Value *V = ConstructSSAForLoadSet(Load, ValuesPerBlock, *this); |
1652 | Load->replaceAllUsesWith(V); |
1653 | |
1654 | if (isa<PHINode>(V)) |
1655 | V->takeName(Load); |
1656 | if (Instruction *I = dyn_cast<Instruction>(V)) |
1657 | // If instruction I has debug info, then we should not update it. |
1658 | // Also, if I has a null DebugLoc, then it is still potentially incorrect |
1659 | // to propagate Load's DebugLoc because Load may not post-dominate I. |
1660 | if (Load->getDebugLoc() && Load->getParent() == I->getParent()) |
1661 | I->setDebugLoc(Load->getDebugLoc()); |
1662 | if (V->getType()->isPtrOrPtrVectorTy()) |
1663 | MD->invalidateCachedPointerInfo(V); |
1664 | markInstructionForDeletion(Load); |
1665 | ++NumGVNLoad; |
1666 | reportLoadElim(Load, V, ORE); |
1667 | return true; |
1668 | } |
1669 | |
1670 | // Step 4: Eliminate partial redundancy. |
1671 | if (!isPREEnabled() || !isLoadPREEnabled()) |
1672 | return Changed; |
1673 | if (!isLoadInLoopPREEnabled() && LI && LI->getLoopFor(Load->getParent())) |
1674 | return Changed; |
1675 | |
1676 | if (performLoopLoadPRE(Load, ValuesPerBlock, UnavailableBlocks) || |
1677 | PerformLoadPRE(Load, ValuesPerBlock, UnavailableBlocks)) |
1678 | return true; |
1679 | |
1680 | return Changed; |
1681 | } |
1682 | |
1683 | static bool impliesEquivalanceIfTrue(CmpInst* Cmp) { |
1684 | if (Cmp->getPredicate() == CmpInst::Predicate::ICMP_EQ) |
1685 | return true; |
1686 | |
1687 | // Floating point comparisons can be equal, but not equivalent. Cases: |
1688 | // NaNs for unordered operators |
1689 | // +0.0 vs 0.0 for all operators |
1690 | if (Cmp->getPredicate() == CmpInst::Predicate::FCMP_OEQ || |
1691 | (Cmp->getPredicate() == CmpInst::Predicate::FCMP_UEQ && |
1692 | Cmp->getFastMathFlags().noNaNs())) { |
1693 | Value *LHS = Cmp->getOperand(0); |
1694 | Value *RHS = Cmp->getOperand(1); |
1695 | // If we can prove either side non-zero, then equality must imply |
1696 | // equivalence. |
1697 | // FIXME: We should do this optimization if 'no signed zeros' is |
1698 | // applicable via an instruction-level fast-math-flag or some other |
1699 | // indicator that relaxed FP semantics are being used. |
1700 | if (isa<ConstantFP>(LHS) && !cast<ConstantFP>(LHS)->isZero()) |
1701 | return true; |
1702 | if (isa<ConstantFP>(RHS) && !cast<ConstantFP>(RHS)->isZero()) |
1703 | return true;; |
1704 | // TODO: Handle vector floating point constants |
1705 | } |
1706 | return false; |
1707 | } |
1708 | |
1709 | static bool impliesEquivalanceIfFalse(CmpInst* Cmp) { |
1710 | if (Cmp->getPredicate() == CmpInst::Predicate::ICMP_NE) |
1711 | return true; |
1712 | |
1713 | // Floating point comparisons can be equal, but not equivelent. Cases: |
1714 | // NaNs for unordered operators |
1715 | // +0.0 vs 0.0 for all operators |
1716 | if ((Cmp->getPredicate() == CmpInst::Predicate::FCMP_ONE && |
1717 | Cmp->getFastMathFlags().noNaNs()) || |
1718 | Cmp->getPredicate() == CmpInst::Predicate::FCMP_UNE) { |
1719 | Value *LHS = Cmp->getOperand(0); |
1720 | Value *RHS = Cmp->getOperand(1); |
1721 | // If we can prove either side non-zero, then equality must imply |
1722 | // equivalence. |
1723 | // FIXME: We should do this optimization if 'no signed zeros' is |
1724 | // applicable via an instruction-level fast-math-flag or some other |
1725 | // indicator that relaxed FP semantics are being used. |
1726 | if (isa<ConstantFP>(LHS) && !cast<ConstantFP>(LHS)->isZero()) |
1727 | return true; |
1728 | if (isa<ConstantFP>(RHS) && !cast<ConstantFP>(RHS)->isZero()) |
1729 | return true;; |
1730 | // TODO: Handle vector floating point constants |
1731 | } |
1732 | return false; |
1733 | } |
1734 | |
1735 | |
1736 | static bool hasUsersIn(Value *V, BasicBlock *BB) { |
1737 | for (User *U : V->users()) |
1738 | if (isa<Instruction>(U) && |
1739 | cast<Instruction>(U)->getParent() == BB) |
1740 | return true; |
1741 | return false; |
1742 | } |
1743 | |
1744 | bool GVN::processAssumeIntrinsic(AssumeInst *IntrinsicI) { |
1745 | Value *V = IntrinsicI->getArgOperand(0); |
1746 | |
1747 | if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) { |
1748 | if (Cond->isZero()) { |
1749 | Type *Int8Ty = Type::getInt8Ty(V->getContext()); |
1750 | // Insert a new store to null instruction before the load to indicate that |
1751 | // this code is not reachable. FIXME: We could insert unreachable |
1752 | // instruction directly because we can modify the CFG. |
1753 | auto *NewS = new StoreInst(UndefValue::get(Int8Ty), |
1754 | Constant::getNullValue(Int8Ty->getPointerTo()), |
1755 | IntrinsicI); |
1756 | if (MSSAU) { |
1757 | const MemoryUseOrDef *FirstNonDom = nullptr; |
1758 | const auto *AL = |
1759 | MSSAU->getMemorySSA()->getBlockAccesses(IntrinsicI->getParent()); |
1760 | |
1761 | // If there are accesses in the current basic block, find the first one |
1762 | // that does not come before NewS. The new memory access is inserted |
1763 | // after the found access or before the terminator if no such access is |
1764 | // found. |
1765 | if (AL) { |
1766 | for (auto &Acc : *AL) { |
1767 | if (auto *Current = dyn_cast<MemoryUseOrDef>(&Acc)) |
1768 | if (!Current->getMemoryInst()->comesBefore(NewS)) { |
1769 | FirstNonDom = Current; |
1770 | break; |
1771 | } |
1772 | } |
1773 | } |
1774 | |
1775 | // This added store is to null, so it will never executed and we can |
1776 | // just use the LiveOnEntry def as defining access. |
1777 | auto *NewDef = |
1778 | FirstNonDom ? MSSAU->createMemoryAccessBefore( |
1779 | NewS, MSSAU->getMemorySSA()->getLiveOnEntryDef(), |
1780 | const_cast<MemoryUseOrDef *>(FirstNonDom)) |
1781 | : MSSAU->createMemoryAccessInBB( |
1782 | NewS, MSSAU->getMemorySSA()->getLiveOnEntryDef(), |
1783 | NewS->getParent(), MemorySSA::BeforeTerminator); |
1784 | |
1785 | MSSAU->insertDef(cast<MemoryDef>(NewDef), /*RenameUses=*/false); |
1786 | } |
1787 | } |
1788 | if (isAssumeWithEmptyBundle(*IntrinsicI)) |
1789 | markInstructionForDeletion(IntrinsicI); |
1790 | return false; |
1791 | } else if (isa<Constant>(V)) { |
1792 | // If it's not false, and constant, it must evaluate to true. This means our |
1793 | // assume is assume(true), and thus, pointless, and we don't want to do |
1794 | // anything more here. |
1795 | return false; |
1796 | } |
1797 | |
1798 | Constant *True = ConstantInt::getTrue(V->getContext()); |
1799 | bool Changed = false; |
1800 | |
1801 | for (BasicBlock *Successor : successors(IntrinsicI->getParent())) { |
1802 | BasicBlockEdge Edge(IntrinsicI->getParent(), Successor); |
1803 | |
1804 | // This property is only true in dominated successors, propagateEquality |
1805 | // will check dominance for us. |
1806 | Changed |= propagateEquality(V, True, Edge, false); |
1807 | } |
1808 | |
1809 | // We can replace assume value with true, which covers cases like this: |
1810 | // call void @llvm.assume(i1 %cmp) |
1811 | // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true |
1812 | ReplaceOperandsWithMap[V] = True; |
1813 | |
1814 | // Similarly, after assume(!NotV) we know that NotV == false. |
1815 | Value *NotV; |
1816 | if (match(V, m_Not(m_Value(NotV)))) |
1817 | ReplaceOperandsWithMap[NotV] = ConstantInt::getFalse(V->getContext()); |
1818 | |
1819 | // If we find an equality fact, canonicalize all dominated uses in this block |
1820 | // to one of the two values. We heuristically choice the "oldest" of the |
1821 | // two where age is determined by value number. (Note that propagateEquality |
1822 | // above handles the cross block case.) |
1823 | // |
1824 | // Key case to cover are: |
1825 | // 1) |
1826 | // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen |
1827 | // call void @llvm.assume(i1 %cmp) |
1828 | // ret float %0 ; will change it to ret float 3.000000e+00 |
1829 | // 2) |
1830 | // %load = load float, float* %addr |
1831 | // %cmp = fcmp oeq float %load, %0 |
1832 | // call void @llvm.assume(i1 %cmp) |
1833 | // ret float %load ; will change it to ret float %0 |
1834 | if (auto *CmpI = dyn_cast<CmpInst>(V)) { |
1835 | if (impliesEquivalanceIfTrue(CmpI)) { |
1836 | Value *CmpLHS = CmpI->getOperand(0); |
1837 | Value *CmpRHS = CmpI->getOperand(1); |
1838 | // Heuristically pick the better replacement -- the choice of heuristic |
1839 | // isn't terribly important here, but the fact we canonicalize on some |
1840 | // replacement is for exposing other simplifications. |
1841 | // TODO: pull this out as a helper function and reuse w/existing |
1842 | // (slightly different) logic. |
1843 | if (isa<Constant>(CmpLHS) && !isa<Constant>(CmpRHS)) |
1844 | std::swap(CmpLHS, CmpRHS); |
1845 | if (!isa<Instruction>(CmpLHS) && isa<Instruction>(CmpRHS)) |
1846 | std::swap(CmpLHS, CmpRHS); |
1847 | if ((isa<Argument>(CmpLHS) && isa<Argument>(CmpRHS)) || |
1848 | (isa<Instruction>(CmpLHS) && isa<Instruction>(CmpRHS))) { |
1849 | // Move the 'oldest' value to the right-hand side, using the value |
1850 | // number as a proxy for age. |
1851 | uint32_t LVN = VN.lookupOrAdd(CmpLHS); |
1852 | uint32_t RVN = VN.lookupOrAdd(CmpRHS); |
1853 | if (LVN < RVN) |
1854 | std::swap(CmpLHS, CmpRHS); |
1855 | } |
1856 | |
1857 | // Handle degenerate case where we either haven't pruned a dead path or a |
1858 | // removed a trivial assume yet. |
1859 | if (isa<Constant>(CmpLHS) && isa<Constant>(CmpRHS)) |
1860 | return Changed; |
1861 | |
1862 | LLVM_DEBUG(dbgs() << "Replacing dominated uses of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Replacing dominated uses of " << *CmpLHS << " with " << *CmpRHS << " in block " << IntrinsicI->getParent()->getName() << "\n" ; } } while (false) |
1863 | << *CmpLHS << " with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Replacing dominated uses of " << *CmpLHS << " with " << *CmpRHS << " in block " << IntrinsicI->getParent()->getName() << "\n" ; } } while (false) |
1864 | << *CmpRHS << " in block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Replacing dominated uses of " << *CmpLHS << " with " << *CmpRHS << " in block " << IntrinsicI->getParent()->getName() << "\n" ; } } while (false) |
1865 | << IntrinsicI->getParent()->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "Replacing dominated uses of " << *CmpLHS << " with " << *CmpRHS << " in block " << IntrinsicI->getParent()->getName() << "\n" ; } } while (false); |
1866 | |
1867 | |
1868 | // Setup the replacement map - this handles uses within the same block |
1869 | if (hasUsersIn(CmpLHS, IntrinsicI->getParent())) |
1870 | ReplaceOperandsWithMap[CmpLHS] = CmpRHS; |
1871 | |
1872 | // NOTE: The non-block local cases are handled by the call to |
1873 | // propagateEquality above; this block is just about handling the block |
1874 | // local cases. TODO: There's a bunch of logic in propagateEqualiy which |
1875 | // isn't duplicated for the block local case, can we share it somehow? |
1876 | } |
1877 | } |
1878 | return Changed; |
1879 | } |
1880 | |
1881 | static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) { |
1882 | patchReplacementInstruction(I, Repl); |
1883 | I->replaceAllUsesWith(Repl); |
1884 | } |
1885 | |
1886 | /// Attempt to eliminate a load, first by eliminating it |
1887 | /// locally, and then attempting non-local elimination if that fails. |
1888 | bool GVN::processLoad(LoadInst *L) { |
1889 | if (!MD) |
1890 | return false; |
1891 | |
1892 | // This code hasn't been audited for ordered or volatile memory access |
1893 | if (!L->isUnordered()) |
1894 | return false; |
1895 | |
1896 | if (L->use_empty()) { |
1897 | markInstructionForDeletion(L); |
1898 | return true; |
1899 | } |
1900 | |
1901 | // ... to a pointer that has been loaded from before... |
1902 | MemDepResult Dep = MD->getDependency(L); |
1903 | |
1904 | // If it is defined in another block, try harder. |
1905 | if (Dep.isNonLocal()) |
1906 | return processNonLocalLoad(L); |
1907 | |
1908 | // Only handle the local case below |
1909 | if (!Dep.isDef() && !Dep.isClobber()) { |
1910 | // This might be a NonFuncLocal or an Unknown |
1911 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; L->printAsOperand (dbgs()); dbgs() << " has unknown dependence\n";; } } while (false) |
1912 | // fast print dep, using operator<< on instruction is too slow.do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; L->printAsOperand (dbgs()); dbgs() << " has unknown dependence\n";; } } while (false) |
1913 | dbgs() << "GVN: load "; L->printAsOperand(dbgs());do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; L->printAsOperand (dbgs()); dbgs() << " has unknown dependence\n";; } } while (false) |
1914 | dbgs() << " has unknown dependence\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN: load "; L->printAsOperand (dbgs()); dbgs() << " has unknown dependence\n";; } } while (false); |
1915 | return false; |
1916 | } |
1917 | |
1918 | AvailableValue AV; |
1919 | if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) { |
1920 | Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this); |
1921 | |
1922 | // Replace the load! |
1923 | patchAndReplaceAllUsesWith(L, AvailableValue); |
1924 | markInstructionForDeletion(L); |
1925 | if (MSSAU) |
1926 | MSSAU->removeMemoryAccess(L); |
1927 | ++NumGVNLoad; |
1928 | reportLoadElim(L, AvailableValue, ORE); |
1929 | // Tell MDA to reexamine the reused pointer since we might have more |
1930 | // information after forwarding it. |
1931 | if (MD && AvailableValue->getType()->isPtrOrPtrVectorTy()) |
1932 | MD->invalidateCachedPointerInfo(AvailableValue); |
1933 | return true; |
1934 | } |
1935 | |
1936 | return false; |
1937 | } |
1938 | |
1939 | /// Return a pair the first field showing the value number of \p Exp and the |
1940 | /// second field showing whether it is a value number newly created. |
1941 | std::pair<uint32_t, bool> |
1942 | GVN::ValueTable::assignExpNewValueNum(Expression &Exp) { |
1943 | uint32_t &e = expressionNumbering[Exp]; |
1944 | bool CreateNewValNum = !e; |
1945 | if (CreateNewValNum) { |
1946 | Expressions.push_back(Exp); |
1947 | if (ExprIdx.size() < nextValueNumber + 1) |
1948 | ExprIdx.resize(nextValueNumber * 2); |
1949 | e = nextValueNumber; |
1950 | ExprIdx[nextValueNumber++] = nextExprNumber++; |
1951 | } |
1952 | return {e, CreateNewValNum}; |
1953 | } |
1954 | |
1955 | /// Return whether all the values related with the same \p num are |
1956 | /// defined in \p BB. |
1957 | bool GVN::ValueTable::areAllValsInBB(uint32_t Num, const BasicBlock *BB, |
1958 | GVN &Gvn) { |
1959 | LeaderTableEntry *Vals = &Gvn.LeaderTable[Num]; |
1960 | while (Vals && Vals->BB == BB) |
1961 | Vals = Vals->Next; |
1962 | return !Vals; |
1963 | } |
1964 | |
1965 | /// Wrap phiTranslateImpl to provide caching functionality. |
1966 | uint32_t GVN::ValueTable::phiTranslate(const BasicBlock *Pred, |
1967 | const BasicBlock *PhiBlock, uint32_t Num, |
1968 | GVN &Gvn) { |
1969 | auto FindRes = PhiTranslateTable.find({Num, Pred}); |
1970 | if (FindRes != PhiTranslateTable.end()) |
1971 | return FindRes->second; |
1972 | uint32_t NewNum = phiTranslateImpl(Pred, PhiBlock, Num, Gvn); |
1973 | PhiTranslateTable.insert({{Num, Pred}, NewNum}); |
1974 | return NewNum; |
1975 | } |
1976 | |
1977 | // Return true if the value number \p Num and NewNum have equal value. |
1978 | // Return false if the result is unknown. |
1979 | bool GVN::ValueTable::areCallValsEqual(uint32_t Num, uint32_t NewNum, |
1980 | const BasicBlock *Pred, |
1981 | const BasicBlock *PhiBlock, GVN &Gvn) { |
1982 | CallInst *Call = nullptr; |
1983 | LeaderTableEntry *Vals = &Gvn.LeaderTable[Num]; |
1984 | while (Vals) { |
1985 | Call = dyn_cast<CallInst>(Vals->Val); |
1986 | if (Call && Call->getParent() == PhiBlock) |
1987 | break; |
1988 | Vals = Vals->Next; |
1989 | } |
1990 | |
1991 | if (AA->doesNotAccessMemory(Call)) |
1992 | return true; |
1993 | |
1994 | if (!MD || !AA->onlyReadsMemory(Call)) |
1995 | return false; |
1996 | |
1997 | MemDepResult local_dep = MD->getDependency(Call); |
1998 | if (!local_dep.isNonLocal()) |
1999 | return false; |
2000 | |
2001 | const MemoryDependenceResults::NonLocalDepInfo &deps = |
2002 | MD->getNonLocalCallDependency(Call); |
2003 | |
2004 | // Check to see if the Call has no function local clobber. |
2005 | for (const NonLocalDepEntry &D : deps) { |
2006 | if (D.getResult().isNonFuncLocal()) |
2007 | return true; |
2008 | } |
2009 | return false; |
2010 | } |
2011 | |
2012 | /// Translate value number \p Num using phis, so that it has the values of |
2013 | /// the phis in BB. |
2014 | uint32_t GVN::ValueTable::phiTranslateImpl(const BasicBlock *Pred, |
2015 | const BasicBlock *PhiBlock, |
2016 | uint32_t Num, GVN &Gvn) { |
2017 | if (PHINode *PN = NumberingPhi[Num]) { |
2018 | for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) { |
2019 | if (PN->getParent() == PhiBlock && PN->getIncomingBlock(i) == Pred) |
2020 | if (uint32_t TransVal = lookup(PN->getIncomingValue(i), false)) |
2021 | return TransVal; |
2022 | } |
2023 | return Num; |
2024 | } |
2025 | |
2026 | // If there is any value related with Num is defined in a BB other than |
2027 | // PhiBlock, it cannot depend on a phi in PhiBlock without going through |
2028 | // a backedge. We can do an early exit in that case to save compile time. |
2029 | if (!areAllValsInBB(Num, PhiBlock, Gvn)) |
2030 | return Num; |
2031 | |
2032 | if (Num >= ExprIdx.size() || ExprIdx[Num] == 0) |
2033 | return Num; |
2034 | Expression Exp = Expressions[ExprIdx[Num]]; |
2035 | |
2036 | for (unsigned i = 0; i < Exp.varargs.size(); i++) { |
2037 | // For InsertValue and ExtractValue, some varargs are index numbers |
2038 | // instead of value numbers. Those index numbers should not be |
2039 | // translated. |
2040 | if ((i > 1 && Exp.opcode == Instruction::InsertValue) || |
2041 | (i > 0 && Exp.opcode == Instruction::ExtractValue) || |
2042 | (i > 1 && Exp.opcode == Instruction::ShuffleVector)) |
2043 | continue; |
2044 | Exp.varargs[i] = phiTranslate(Pred, PhiBlock, Exp.varargs[i], Gvn); |
2045 | } |
2046 | |
2047 | if (Exp.commutative) { |
2048 | assert(Exp.varargs.size() >= 2 && "Unsupported commutative instruction!")(static_cast <bool> (Exp.varargs.size() >= 2 && "Unsupported commutative instruction!") ? void (0) : __assert_fail ("Exp.varargs.size() >= 2 && \"Unsupported commutative instruction!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2048, __extension__ __PRETTY_FUNCTION__)); |
2049 | if (Exp.varargs[0] > Exp.varargs[1]) { |
2050 | std::swap(Exp.varargs[0], Exp.varargs[1]); |
2051 | uint32_t Opcode = Exp.opcode >> 8; |
2052 | if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) |
2053 | Exp.opcode = (Opcode << 8) | |
2054 | CmpInst::getSwappedPredicate( |
2055 | static_cast<CmpInst::Predicate>(Exp.opcode & 255)); |
2056 | } |
2057 | } |
2058 | |
2059 | if (uint32_t NewNum = expressionNumbering[Exp]) { |
2060 | if (Exp.opcode == Instruction::Call && NewNum != Num) |
2061 | return areCallValsEqual(Num, NewNum, Pred, PhiBlock, Gvn) ? NewNum : Num; |
2062 | return NewNum; |
2063 | } |
2064 | return Num; |
2065 | } |
2066 | |
2067 | /// Erase stale entry from phiTranslate cache so phiTranslate can be computed |
2068 | /// again. |
2069 | void GVN::ValueTable::eraseTranslateCacheEntry(uint32_t Num, |
2070 | const BasicBlock &CurrBlock) { |
2071 | for (const BasicBlock *Pred : predecessors(&CurrBlock)) |
2072 | PhiTranslateTable.erase({Num, Pred}); |
2073 | } |
2074 | |
2075 | // In order to find a leader for a given value number at a |
2076 | // specific basic block, we first obtain the list of all Values for that number, |
2077 | // and then scan the list to find one whose block dominates the block in |
2078 | // question. This is fast because dominator tree queries consist of only |
2079 | // a few comparisons of DFS numbers. |
2080 | Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) { |
2081 | LeaderTableEntry Vals = LeaderTable[num]; |
2082 | if (!Vals.Val) return nullptr; |
2083 | |
2084 | Value *Val = nullptr; |
2085 | if (DT->dominates(Vals.BB, BB)) { |
2086 | Val = Vals.Val; |
2087 | if (isa<Constant>(Val)) return Val; |
2088 | } |
2089 | |
2090 | LeaderTableEntry* Next = Vals.Next; |
2091 | while (Next) { |
2092 | if (DT->dominates(Next->BB, BB)) { |
2093 | if (isa<Constant>(Next->Val)) return Next->Val; |
2094 | if (!Val) Val = Next->Val; |
2095 | } |
2096 | |
2097 | Next = Next->Next; |
2098 | } |
2099 | |
2100 | return Val; |
2101 | } |
2102 | |
2103 | /// There is an edge from 'Src' to 'Dst'. Return |
2104 | /// true if every path from the entry block to 'Dst' passes via this edge. In |
2105 | /// particular 'Dst' must not be reachable via another edge from 'Src'. |
2106 | static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E, |
2107 | DominatorTree *DT) { |
2108 | // While in theory it is interesting to consider the case in which Dst has |
2109 | // more than one predecessor, because Dst might be part of a loop which is |
2110 | // only reachable from Src, in practice it is pointless since at the time |
2111 | // GVN runs all such loops have preheaders, which means that Dst will have |
2112 | // been changed to have only one predecessor, namely Src. |
2113 | const BasicBlock *Pred = E.getEnd()->getSinglePredecessor(); |
2114 | assert((!Pred || Pred == E.getStart()) &&(static_cast <bool> ((!Pred || Pred == E.getStart()) && "No edge between these basic blocks!") ? void (0) : __assert_fail ("(!Pred || Pred == E.getStart()) && \"No edge between these basic blocks!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2115, __extension__ __PRETTY_FUNCTION__)) |
2115 | "No edge between these basic blocks!")(static_cast <bool> ((!Pred || Pred == E.getStart()) && "No edge between these basic blocks!") ? void (0) : __assert_fail ("(!Pred || Pred == E.getStart()) && \"No edge between these basic blocks!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2115, __extension__ __PRETTY_FUNCTION__)); |
2116 | return Pred != nullptr; |
2117 | } |
2118 | |
2119 | void GVN::assignBlockRPONumber(Function &F) { |
2120 | BlockRPONumber.clear(); |
2121 | uint32_t NextBlockNumber = 1; |
2122 | ReversePostOrderTraversal<Function *> RPOT(&F); |
2123 | for (BasicBlock *BB : RPOT) |
2124 | BlockRPONumber[BB] = NextBlockNumber++; |
2125 | InvalidBlockRPONumbers = false; |
2126 | } |
2127 | |
2128 | bool GVN::replaceOperandsForInBlockEquality(Instruction *Instr) const { |
2129 | bool Changed = false; |
2130 | for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) { |
2131 | Value *Operand = Instr->getOperand(OpNum); |
2132 | auto it = ReplaceOperandsWithMap.find(Operand); |
2133 | if (it != ReplaceOperandsWithMap.end()) { |
2134 | LLVM_DEBUG(dbgs() << "GVN replacing: " << *Operand << " with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN replacing: " << *Operand << " with " << *it->second << " in instruction " << *Instr << '\n'; } } while (false) |
2135 | << *it->second << " in instruction " << *Instr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN replacing: " << *Operand << " with " << *it->second << " in instruction " << *Instr << '\n'; } } while (false); |
2136 | Instr->setOperand(OpNum, it->second); |
2137 | Changed = true; |
2138 | } |
2139 | } |
2140 | return Changed; |
2141 | } |
2142 | |
2143 | /// The given values are known to be equal in every block |
2144 | /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with |
2145 | /// 'RHS' everywhere in the scope. Returns whether a change was made. |
2146 | /// If DominatesByEdge is false, then it means that we will propagate the RHS |
2147 | /// value starting from the end of Root.Start. |
2148 | bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root, |
2149 | bool DominatesByEdge) { |
2150 | SmallVector<std::pair<Value*, Value*>, 4> Worklist; |
2151 | Worklist.push_back(std::make_pair(LHS, RHS)); |
2152 | bool Changed = false; |
2153 | // For speed, compute a conservative fast approximation to |
2154 | // DT->dominates(Root, Root.getEnd()); |
2155 | const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT); |
2156 | |
2157 | while (!Worklist.empty()) { |
2158 | std::pair<Value*, Value*> Item = Worklist.pop_back_val(); |
2159 | LHS = Item.first; RHS = Item.second; |
2160 | |
2161 | if (LHS == RHS) |
2162 | continue; |
2163 | assert(LHS->getType() == RHS->getType() && "Equality but unequal types!")(static_cast <bool> (LHS->getType() == RHS->getType () && "Equality but unequal types!") ? void (0) : __assert_fail ("LHS->getType() == RHS->getType() && \"Equality but unequal types!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2163, __extension__ __PRETTY_FUNCTION__)); |
2164 | |
2165 | // Don't try to propagate equalities between constants. |
2166 | if (isa<Constant>(LHS) && isa<Constant>(RHS)) |
2167 | continue; |
2168 | |
2169 | // Prefer a constant on the right-hand side, or an Argument if no constants. |
2170 | if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS))) |
2171 | std::swap(LHS, RHS); |
2172 | assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!")(static_cast <bool> ((isa<Argument>(LHS) || isa< Instruction>(LHS)) && "Unexpected value!") ? void ( 0) : __assert_fail ("(isa<Argument>(LHS) || isa<Instruction>(LHS)) && \"Unexpected value!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2172, __extension__ __PRETTY_FUNCTION__)); |
2173 | |
2174 | // If there is no obvious reason to prefer the left-hand side over the |
2175 | // right-hand side, ensure the longest lived term is on the right-hand side, |
2176 | // so the shortest lived term will be replaced by the longest lived. |
2177 | // This tends to expose more simplifications. |
2178 | uint32_t LVN = VN.lookupOrAdd(LHS); |
2179 | if ((isa<Argument>(LHS) && isa<Argument>(RHS)) || |
2180 | (isa<Instruction>(LHS) && isa<Instruction>(RHS))) { |
2181 | // Move the 'oldest' value to the right-hand side, using the value number |
2182 | // as a proxy for age. |
2183 | uint32_t RVN = VN.lookupOrAdd(RHS); |
2184 | if (LVN < RVN) { |
2185 | std::swap(LHS, RHS); |
2186 | LVN = RVN; |
2187 | } |
2188 | } |
2189 | |
2190 | // If value numbering later sees that an instruction in the scope is equal |
2191 | // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve |
2192 | // the invariant that instructions only occur in the leader table for their |
2193 | // own value number (this is used by removeFromLeaderTable), do not do this |
2194 | // if RHS is an instruction (if an instruction in the scope is morphed into |
2195 | // LHS then it will be turned into RHS by the next GVN iteration anyway, so |
2196 | // using the leader table is about compiling faster, not optimizing better). |
2197 | // The leader table only tracks basic blocks, not edges. Only add to if we |
2198 | // have the simple case where the edge dominates the end. |
2199 | if (RootDominatesEnd && !isa<Instruction>(RHS)) |
2200 | addToLeaderTable(LVN, RHS, Root.getEnd()); |
2201 | |
2202 | // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As |
2203 | // LHS always has at least one use that is not dominated by Root, this will |
2204 | // never do anything if LHS has only one use. |
2205 | if (!LHS->hasOneUse()) { |
2206 | unsigned NumReplacements = |
2207 | DominatesByEdge |
2208 | ? replaceDominatedUsesWith(LHS, RHS, *DT, Root) |
2209 | : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart()); |
2210 | |
2211 | Changed |= NumReplacements > 0; |
2212 | NumGVNEqProp += NumReplacements; |
2213 | // Cached information for anything that uses LHS will be invalid. |
2214 | if (MD) |
2215 | MD->invalidateCachedPointerInfo(LHS); |
2216 | } |
2217 | |
2218 | // Now try to deduce additional equalities from this one. For example, if |
2219 | // the known equality was "(A != B)" == "false" then it follows that A and B |
2220 | // are equal in the scope. Only boolean equalities with an explicit true or |
2221 | // false RHS are currently supported. |
2222 | if (!RHS->getType()->isIntegerTy(1)) |
2223 | // Not a boolean equality - bail out. |
2224 | continue; |
2225 | ConstantInt *CI = dyn_cast<ConstantInt>(RHS); |
2226 | if (!CI) |
2227 | // RHS neither 'true' nor 'false' - bail out. |
2228 | continue; |
2229 | // Whether RHS equals 'true'. Otherwise it equals 'false'. |
2230 | bool isKnownTrue = CI->isMinusOne(); |
2231 | bool isKnownFalse = !isKnownTrue; |
2232 | |
2233 | // If "A && B" is known true then both A and B are known true. If "A || B" |
2234 | // is known false then both A and B are known false. |
2235 | Value *A, *B; |
2236 | if ((isKnownTrue && match(LHS, m_LogicalAnd(m_Value(A), m_Value(B)))) || |
2237 | (isKnownFalse && match(LHS, m_LogicalOr(m_Value(A), m_Value(B))))) { |
2238 | Worklist.push_back(std::make_pair(A, RHS)); |
2239 | Worklist.push_back(std::make_pair(B, RHS)); |
2240 | continue; |
2241 | } |
2242 | |
2243 | // If we are propagating an equality like "(A == B)" == "true" then also |
2244 | // propagate the equality A == B. When propagating a comparison such as |
2245 | // "(A >= B)" == "true", replace all instances of "A < B" with "false". |
2246 | if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) { |
2247 | Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1); |
2248 | |
2249 | // If "A == B" is known true, or "A != B" is known false, then replace |
2250 | // A with B everywhere in the scope. For floating point operations, we |
2251 | // have to be careful since equality does not always imply equivalance. |
2252 | if ((isKnownTrue && impliesEquivalanceIfTrue(Cmp)) || |
2253 | (isKnownFalse && impliesEquivalanceIfFalse(Cmp))) |
2254 | Worklist.push_back(std::make_pair(Op0, Op1)); |
2255 | |
2256 | // If "A >= B" is known true, replace "A < B" with false everywhere. |
2257 | CmpInst::Predicate NotPred = Cmp->getInversePredicate(); |
2258 | Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse); |
2259 | // Since we don't have the instruction "A < B" immediately to hand, work |
2260 | // out the value number that it would have and use that to find an |
2261 | // appropriate instruction (if any). |
2262 | uint32_t NextNum = VN.getNextUnusedValueNumber(); |
2263 | uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1); |
2264 | // If the number we were assigned was brand new then there is no point in |
2265 | // looking for an instruction realizing it: there cannot be one! |
2266 | if (Num < NextNum) { |
2267 | Value *NotCmp = findLeader(Root.getEnd(), Num); |
2268 | if (NotCmp && isa<Instruction>(NotCmp)) { |
2269 | unsigned NumReplacements = |
2270 | DominatesByEdge |
2271 | ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root) |
2272 | : replaceDominatedUsesWith(NotCmp, NotVal, *DT, |
2273 | Root.getStart()); |
2274 | Changed |= NumReplacements > 0; |
2275 | NumGVNEqProp += NumReplacements; |
2276 | // Cached information for anything that uses NotCmp will be invalid. |
2277 | if (MD) |
2278 | MD->invalidateCachedPointerInfo(NotCmp); |
2279 | } |
2280 | } |
2281 | // Ensure that any instruction in scope that gets the "A < B" value number |
2282 | // is replaced with false. |
2283 | // The leader table only tracks basic blocks, not edges. Only add to if we |
2284 | // have the simple case where the edge dominates the end. |
2285 | if (RootDominatesEnd) |
2286 | addToLeaderTable(Num, NotVal, Root.getEnd()); |
2287 | |
2288 | continue; |
2289 | } |
2290 | } |
2291 | |
2292 | return Changed; |
2293 | } |
2294 | |
2295 | /// When calculating availability, handle an instruction |
2296 | /// by inserting it into the appropriate sets |
2297 | bool GVN::processInstruction(Instruction *I) { |
2298 | // Ignore dbg info intrinsics. |
2299 | if (isa<DbgInfoIntrinsic>(I)) |
2300 | return false; |
2301 | |
2302 | // If the instruction can be easily simplified then do so now in preference |
2303 | // to value numbering it. Value numbering often exposes redundancies, for |
2304 | // example if it determines that %y is equal to %x then the instruction |
2305 | // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. |
2306 | const DataLayout &DL = I->getModule()->getDataLayout(); |
2307 | if (Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC})) { |
2308 | bool Changed = false; |
2309 | if (!I->use_empty()) { |
2310 | // Simplification can cause a special instruction to become not special. |
2311 | // For example, devirtualization to a willreturn function. |
2312 | ICF->removeUsersOf(I); |
2313 | I->replaceAllUsesWith(V); |
2314 | Changed = true; |
2315 | } |
2316 | if (isInstructionTriviallyDead(I, TLI)) { |
2317 | markInstructionForDeletion(I); |
2318 | Changed = true; |
2319 | } |
2320 | if (Changed) { |
2321 | if (MD && V->getType()->isPtrOrPtrVectorTy()) |
2322 | MD->invalidateCachedPointerInfo(V); |
2323 | ++NumGVNSimpl; |
2324 | return true; |
2325 | } |
2326 | } |
2327 | |
2328 | if (auto *Assume = dyn_cast<AssumeInst>(I)) |
2329 | return processAssumeIntrinsic(Assume); |
2330 | |
2331 | if (LoadInst *Load = dyn_cast<LoadInst>(I)) { |
2332 | if (processLoad(Load)) |
2333 | return true; |
2334 | |
2335 | unsigned Num = VN.lookupOrAdd(Load); |
2336 | addToLeaderTable(Num, Load, Load->getParent()); |
2337 | return false; |
2338 | } |
2339 | |
2340 | // For conditional branches, we can perform simple conditional propagation on |
2341 | // the condition value itself. |
2342 | if (BranchInst *BI = dyn_cast<BranchInst>(I)) { |
2343 | if (!BI->isConditional()) |
2344 | return false; |
2345 | |
2346 | if (isa<Constant>(BI->getCondition())) |
2347 | return processFoldableCondBr(BI); |
2348 | |
2349 | Value *BranchCond = BI->getCondition(); |
2350 | BasicBlock *TrueSucc = BI->getSuccessor(0); |
2351 | BasicBlock *FalseSucc = BI->getSuccessor(1); |
2352 | // Avoid multiple edges early. |
2353 | if (TrueSucc == FalseSucc) |
2354 | return false; |
2355 | |
2356 | BasicBlock *Parent = BI->getParent(); |
2357 | bool Changed = false; |
2358 | |
2359 | Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext()); |
2360 | BasicBlockEdge TrueE(Parent, TrueSucc); |
2361 | Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true); |
2362 | |
2363 | Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext()); |
2364 | BasicBlockEdge FalseE(Parent, FalseSucc); |
2365 | Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true); |
2366 | |
2367 | return Changed; |
2368 | } |
2369 | |
2370 | // For switches, propagate the case values into the case destinations. |
2371 | if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) { |
2372 | Value *SwitchCond = SI->getCondition(); |
2373 | BasicBlock *Parent = SI->getParent(); |
2374 | bool Changed = false; |
2375 | |
2376 | // Remember how many outgoing edges there are to every successor. |
2377 | SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges; |
2378 | for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i) |
2379 | ++SwitchEdges[SI->getSuccessor(i)]; |
2380 | |
2381 | for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); |
2382 | i != e; ++i) { |
2383 | BasicBlock *Dst = i->getCaseSuccessor(); |
2384 | // If there is only a single edge, propagate the case value into it. |
2385 | if (SwitchEdges.lookup(Dst) == 1) { |
2386 | BasicBlockEdge E(Parent, Dst); |
2387 | Changed |= propagateEquality(SwitchCond, i->getCaseValue(), E, true); |
2388 | } |
2389 | } |
2390 | return Changed; |
2391 | } |
2392 | |
2393 | // Instructions with void type don't return a value, so there's |
2394 | // no point in trying to find redundancies in them. |
2395 | if (I->getType()->isVoidTy()) |
2396 | return false; |
2397 | |
2398 | uint32_t NextNum = VN.getNextUnusedValueNumber(); |
2399 | unsigned Num = VN.lookupOrAdd(I); |
2400 | |
2401 | // Allocations are always uniquely numbered, so we can save time and memory |
2402 | // by fast failing them. |
2403 | if (isa<AllocaInst>(I) || I->isTerminator() || isa<PHINode>(I)) { |
2404 | addToLeaderTable(Num, I, I->getParent()); |
2405 | return false; |
2406 | } |
2407 | |
2408 | // If the number we were assigned was a brand new VN, then we don't |
2409 | // need to do a lookup to see if the number already exists |
2410 | // somewhere in the domtree: it can't! |
2411 | if (Num >= NextNum) { |
2412 | addToLeaderTable(Num, I, I->getParent()); |
2413 | return false; |
2414 | } |
2415 | |
2416 | // Perform fast-path value-number based elimination of values inherited from |
2417 | // dominators. |
2418 | Value *Repl = findLeader(I->getParent(), Num); |
2419 | if (!Repl) { |
2420 | // Failure, just remember this instance for future use. |
2421 | addToLeaderTable(Num, I, I->getParent()); |
2422 | return false; |
2423 | } else if (Repl == I) { |
2424 | // If I was the result of a shortcut PRE, it might already be in the table |
2425 | // and the best replacement for itself. Nothing to do. |
2426 | return false; |
2427 | } |
2428 | |
2429 | // Remove it! |
2430 | patchAndReplaceAllUsesWith(I, Repl); |
2431 | if (MD && Repl->getType()->isPtrOrPtrVectorTy()) |
2432 | MD->invalidateCachedPointerInfo(Repl); |
2433 | markInstructionForDeletion(I); |
2434 | return true; |
2435 | } |
2436 | |
2437 | /// runOnFunction - This is the main transformation entry point for a function. |
2438 | bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT, |
2439 | const TargetLibraryInfo &RunTLI, AAResults &RunAA, |
2440 | MemoryDependenceResults *RunMD, LoopInfo *LI, |
2441 | OptimizationRemarkEmitter *RunORE, MemorySSA *MSSA) { |
2442 | AC = &RunAC; |
2443 | DT = &RunDT; |
2444 | VN.setDomTree(DT); |
2445 | TLI = &RunTLI; |
2446 | VN.setAliasAnalysis(&RunAA); |
2447 | MD = RunMD; |
2448 | ImplicitControlFlowTracking ImplicitCFT; |
2449 | ICF = &ImplicitCFT; |
2450 | this->LI = LI; |
2451 | VN.setMemDep(MD); |
2452 | ORE = RunORE; |
2453 | InvalidBlockRPONumbers = true; |
2454 | MemorySSAUpdater Updater(MSSA); |
2455 | MSSAU = MSSA ? &Updater : nullptr; |
2456 | |
2457 | bool Changed = false; |
2458 | bool ShouldContinue = true; |
2459 | |
2460 | DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); |
2461 | // Merge unconditional branches, allowing PRE to catch more |
2462 | // optimization opportunities. |
2463 | for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { |
2464 | BasicBlock *BB = &*FI++; |
2465 | |
2466 | bool removedBlock = MergeBlockIntoPredecessor(BB, &DTU, LI, MSSAU, MD); |
2467 | if (removedBlock) |
2468 | ++NumGVNBlocks; |
2469 | |
2470 | Changed |= removedBlock; |
2471 | } |
2472 | |
2473 | unsigned Iteration = 0; |
2474 | while (ShouldContinue) { |
2475 | LLVM_DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN iteration: " << Iteration << "\n"; } } while (false); |
2476 | ShouldContinue = iterateOnFunction(F); |
2477 | Changed |= ShouldContinue; |
2478 | ++Iteration; |
2479 | } |
2480 | |
2481 | if (isPREEnabled()) { |
2482 | // Fabricate val-num for dead-code in order to suppress assertion in |
2483 | // performPRE(). |
2484 | assignValNumForDeadCode(); |
2485 | bool PREChanged = true; |
2486 | while (PREChanged) { |
2487 | PREChanged = performPRE(F); |
2488 | Changed |= PREChanged; |
2489 | } |
2490 | } |
2491 | |
2492 | // FIXME: Should perform GVN again after PRE does something. PRE can move |
2493 | // computations into blocks where they become fully redundant. Note that |
2494 | // we can't do this until PRE's critical edge splitting updates memdep. |
2495 | // Actually, when this happens, we should just fully integrate PRE into GVN. |
2496 | |
2497 | cleanupGlobalSets(); |
2498 | // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each |
2499 | // iteration. |
2500 | DeadBlocks.clear(); |
2501 | |
2502 | if (MSSA && VerifyMemorySSA) |
2503 | MSSA->verifyMemorySSA(); |
2504 | |
2505 | return Changed; |
2506 | } |
2507 | |
2508 | bool GVN::processBlock(BasicBlock *BB) { |
2509 | // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function |
2510 | // (and incrementing BI before processing an instruction). |
2511 | assert(InstrsToErase.empty() &&(static_cast <bool> (InstrsToErase.empty() && "We expect InstrsToErase to be empty across iterations" ) ? void (0) : __assert_fail ("InstrsToErase.empty() && \"We expect InstrsToErase to be empty across iterations\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2512, __extension__ __PRETTY_FUNCTION__)) |
2512 | "We expect InstrsToErase to be empty across iterations")(static_cast <bool> (InstrsToErase.empty() && "We expect InstrsToErase to be empty across iterations" ) ? void (0) : __assert_fail ("InstrsToErase.empty() && \"We expect InstrsToErase to be empty across iterations\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2512, __extension__ __PRETTY_FUNCTION__)); |
2513 | if (DeadBlocks.count(BB)) |
2514 | return false; |
2515 | |
2516 | // Clearing map before every BB because it can be used only for single BB. |
2517 | ReplaceOperandsWithMap.clear(); |
2518 | bool ChangedFunction = false; |
2519 | |
2520 | // Since we may not have visited the input blocks of the phis, we can't |
2521 | // use our normal hash approach for phis. Instead, simply look for |
2522 | // obvious duplicates. The first pass of GVN will tend to create |
2523 | // identical phis, and the second or later passes can eliminate them. |
2524 | ChangedFunction |= EliminateDuplicatePHINodes(BB); |
2525 | |
2526 | for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); |
2527 | BI != BE;) { |
2528 | if (!ReplaceOperandsWithMap.empty()) |
2529 | ChangedFunction |= replaceOperandsForInBlockEquality(&*BI); |
2530 | ChangedFunction |= processInstruction(&*BI); |
2531 | |
2532 | if (InstrsToErase.empty()) { |
2533 | ++BI; |
2534 | continue; |
2535 | } |
2536 | |
2537 | // If we need some instructions deleted, do it now. |
2538 | NumGVNInstr += InstrsToErase.size(); |
2539 | |
2540 | // Avoid iterator invalidation. |
2541 | bool AtStart = BI == BB->begin(); |
2542 | if (!AtStart) |
2543 | --BI; |
2544 | |
2545 | for (auto *I : InstrsToErase) { |
2546 | assert(I->getParent() == BB && "Removing instruction from wrong block?")(static_cast <bool> (I->getParent() == BB && "Removing instruction from wrong block?") ? void (0) : __assert_fail ("I->getParent() == BB && \"Removing instruction from wrong block?\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2546, __extension__ __PRETTY_FUNCTION__)); |
2547 | LLVM_DEBUG(dbgs() << "GVN removed: " << *I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN removed: " << *I << '\n'; } } while (false); |
2548 | salvageKnowledge(I, AC); |
2549 | salvageDebugInfo(*I); |
2550 | if (MD) MD->removeInstruction(I); |
2551 | if (MSSAU) |
2552 | MSSAU->removeMemoryAccess(I); |
2553 | LLVM_DEBUG(verifyRemoved(I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { verifyRemoved(I); } } while (false); |
2554 | ICF->removeInstruction(I); |
2555 | I->eraseFromParent(); |
2556 | } |
2557 | InstrsToErase.clear(); |
2558 | |
2559 | if (AtStart) |
2560 | BI = BB->begin(); |
2561 | else |
2562 | ++BI; |
2563 | } |
2564 | |
2565 | return ChangedFunction; |
2566 | } |
2567 | |
2568 | // Instantiate an expression in a predecessor that lacked it. |
2569 | bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred, |
2570 | BasicBlock *Curr, unsigned int ValNo) { |
2571 | // Because we are going top-down through the block, all value numbers |
2572 | // will be available in the predecessor by the time we need them. Any |
2573 | // that weren't originally present will have been instantiated earlier |
2574 | // in this loop. |
2575 | bool success = true; |
2576 | for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) { |
2577 | Value *Op = Instr->getOperand(i); |
2578 | if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) |
2579 | continue; |
2580 | // This could be a newly inserted instruction, in which case, we won't |
2581 | // find a value number, and should give up before we hurt ourselves. |
2582 | // FIXME: Rewrite the infrastructure to let it easier to value number |
2583 | // and process newly inserted instructions. |
2584 | if (!VN.exists(Op)) { |
2585 | success = false; |
2586 | break; |
2587 | } |
2588 | uint32_t TValNo = |
2589 | VN.phiTranslate(Pred, Curr, VN.lookup(Op), *this); |
2590 | if (Value *V = findLeader(Pred, TValNo)) { |
2591 | Instr->setOperand(i, V); |
2592 | } else { |
2593 | success = false; |
2594 | break; |
2595 | } |
2596 | } |
2597 | |
2598 | // Fail out if we encounter an operand that is not available in |
2599 | // the PRE predecessor. This is typically because of loads which |
2600 | // are not value numbered precisely. |
2601 | if (!success) |
2602 | return false; |
2603 | |
2604 | Instr->insertBefore(Pred->getTerminator()); |
2605 | Instr->setName(Instr->getName() + ".pre"); |
2606 | Instr->setDebugLoc(Instr->getDebugLoc()); |
2607 | |
2608 | ICF->insertInstructionTo(Instr, Pred); |
2609 | |
2610 | unsigned Num = VN.lookupOrAdd(Instr); |
2611 | VN.add(Instr, Num); |
2612 | |
2613 | // Update the availability map to include the new instruction. |
2614 | addToLeaderTable(Num, Instr, Pred); |
2615 | return true; |
2616 | } |
2617 | |
2618 | bool GVN::performScalarPRE(Instruction *CurInst) { |
2619 | if (isa<AllocaInst>(CurInst) || CurInst->isTerminator() || |
2620 | isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() || |
2621 | CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || |
2622 | isa<DbgInfoIntrinsic>(CurInst)) |
2623 | return false; |
2624 | |
2625 | // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from |
2626 | // sinking the compare again, and it would force the code generator to |
2627 | // move the i1 from processor flags or predicate registers into a general |
2628 | // purpose register. |
2629 | if (isa<CmpInst>(CurInst)) |
2630 | return false; |
2631 | |
2632 | // Don't do PRE on GEPs. The inserted PHI would prevent CodeGenPrepare from |
2633 | // sinking the addressing mode computation back to its uses. Extending the |
2634 | // GEP's live range increases the register pressure, and therefore it can |
2635 | // introduce unnecessary spills. |
2636 | // |
2637 | // This doesn't prevent Load PRE. PHI translation will make the GEP available |
2638 | // to the load by moving it to the predecessor block if necessary. |
2639 | if (isa<GetElementPtrInst>(CurInst)) |
2640 | return false; |
2641 | |
2642 | if (auto *CallB = dyn_cast<CallBase>(CurInst)) { |
2643 | // We don't currently value number ANY inline asm calls. |
2644 | if (CallB->isInlineAsm()) |
2645 | return false; |
2646 | // Don't do PRE on convergent calls. |
2647 | if (CallB->isConvergent()) |
2648 | return false; |
2649 | } |
2650 | |
2651 | uint32_t ValNo = VN.lookup(CurInst); |
2652 | |
2653 | // Look for the predecessors for PRE opportunities. We're |
2654 | // only trying to solve the basic diamond case, where |
2655 | // a value is computed in the successor and one predecessor, |
2656 | // but not the other. We also explicitly disallow cases |
2657 | // where the successor is its own predecessor, because they're |
2658 | // more complicated to get right. |
2659 | unsigned NumWith = 0; |
2660 | unsigned NumWithout = 0; |
2661 | BasicBlock *PREPred = nullptr; |
2662 | BasicBlock *CurrentBlock = CurInst->getParent(); |
2663 | |
2664 | // Update the RPO numbers for this function. |
2665 | if (InvalidBlockRPONumbers) |
2666 | assignBlockRPONumber(*CurrentBlock->getParent()); |
2667 | |
2668 | SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap; |
2669 | for (BasicBlock *P : predecessors(CurrentBlock)) { |
2670 | // We're not interested in PRE where blocks with predecessors that are |
2671 | // not reachable. |
2672 | if (!DT->isReachableFromEntry(P)) { |
2673 | NumWithout = 2; |
2674 | break; |
2675 | } |
2676 | // It is not safe to do PRE when P->CurrentBlock is a loop backedge, and |
2677 | // when CurInst has operand defined in CurrentBlock (so it may be defined |
2678 | // by phi in the loop header). |
2679 | assert(BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) &&(static_cast <bool> (BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && "Invalid BlockRPONumber map." ) ? void (0) : __assert_fail ("BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && \"Invalid BlockRPONumber map.\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2680, __extension__ __PRETTY_FUNCTION__)) |
2680 | "Invalid BlockRPONumber map.")(static_cast <bool> (BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && "Invalid BlockRPONumber map." ) ? void (0) : __assert_fail ("BlockRPONumber.count(P) && BlockRPONumber.count(CurrentBlock) && \"Invalid BlockRPONumber map.\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2680, __extension__ __PRETTY_FUNCTION__)); |
2681 | if (BlockRPONumber[P] >= BlockRPONumber[CurrentBlock] && |
2682 | llvm::any_of(CurInst->operands(), [&](const Use &U) { |
2683 | if (auto *Inst = dyn_cast<Instruction>(U.get())) |
2684 | return Inst->getParent() == CurrentBlock; |
2685 | return false; |
2686 | })) { |
2687 | NumWithout = 2; |
2688 | break; |
2689 | } |
2690 | |
2691 | uint32_t TValNo = VN.phiTranslate(P, CurrentBlock, ValNo, *this); |
2692 | Value *predV = findLeader(P, TValNo); |
2693 | if (!predV) { |
2694 | predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P)); |
2695 | PREPred = P; |
2696 | ++NumWithout; |
2697 | } else if (predV == CurInst) { |
2698 | /* CurInst dominates this predecessor. */ |
2699 | NumWithout = 2; |
2700 | break; |
2701 | } else { |
2702 | predMap.push_back(std::make_pair(predV, P)); |
2703 | ++NumWith; |
2704 | } |
2705 | } |
2706 | |
2707 | // Don't do PRE when it might increase code size, i.e. when |
2708 | // we would need to insert instructions in more than one pred. |
2709 | if (NumWithout > 1 || NumWith == 0) |
2710 | return false; |
2711 | |
2712 | // We may have a case where all predecessors have the instruction, |
2713 | // and we just need to insert a phi node. Otherwise, perform |
2714 | // insertion. |
2715 | Instruction *PREInstr = nullptr; |
2716 | |
2717 | if (NumWithout != 0) { |
2718 | if (!isSafeToSpeculativelyExecute(CurInst)) { |
2719 | // It is only valid to insert a new instruction if the current instruction |
2720 | // is always executed. An instruction with implicit control flow could |
2721 | // prevent us from doing it. If we cannot speculate the execution, then |
2722 | // PRE should be prohibited. |
2723 | if (ICF->isDominatedByICFIFromSameBlock(CurInst)) |
2724 | return false; |
2725 | } |
2726 | |
2727 | // Don't do PRE across indirect branch. |
2728 | if (isa<IndirectBrInst>(PREPred->getTerminator())) |
2729 | return false; |
2730 | |
2731 | // Don't do PRE across callbr. |
2732 | // FIXME: Can we do this across the fallthrough edge? |
2733 | if (isa<CallBrInst>(PREPred->getTerminator())) |
2734 | return false; |
2735 | |
2736 | // We can't do PRE safely on a critical edge, so instead we schedule |
2737 | // the edge to be split and perform the PRE the next time we iterate |
2738 | // on the function. |
2739 | unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); |
2740 | if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { |
2741 | toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); |
2742 | return false; |
2743 | } |
2744 | // We need to insert somewhere, so let's give it a shot |
2745 | PREInstr = CurInst->clone(); |
2746 | if (!performScalarPREInsertion(PREInstr, PREPred, CurrentBlock, ValNo)) { |
2747 | // If we failed insertion, make sure we remove the instruction. |
2748 | LLVM_DEBUG(verifyRemoved(PREInstr))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { verifyRemoved(PREInstr); } } while (false); |
2749 | PREInstr->deleteValue(); |
2750 | return false; |
2751 | } |
2752 | } |
2753 | |
2754 | // Either we should have filled in the PRE instruction, or we should |
2755 | // not have needed insertions. |
2756 | assert(PREInstr != nullptr || NumWithout == 0)(static_cast <bool> (PREInstr != nullptr || NumWithout == 0) ? void (0) : __assert_fail ("PREInstr != nullptr || NumWithout == 0" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2756, __extension__ __PRETTY_FUNCTION__)); |
2757 | |
2758 | ++NumGVNPRE; |
2759 | |
2760 | // Create a PHI to make the value available in this block. |
2761 | PHINode *Phi = |
2762 | PHINode::Create(CurInst->getType(), predMap.size(), |
2763 | CurInst->getName() + ".pre-phi", &CurrentBlock->front()); |
2764 | for (unsigned i = 0, e = predMap.size(); i != e; ++i) { |
2765 | if (Value *V = predMap[i].first) { |
2766 | // If we use an existing value in this phi, we have to patch the original |
2767 | // value because the phi will be used to replace a later value. |
2768 | patchReplacementInstruction(CurInst, V); |
2769 | Phi->addIncoming(V, predMap[i].second); |
2770 | } else |
2771 | Phi->addIncoming(PREInstr, PREPred); |
2772 | } |
2773 | |
2774 | VN.add(Phi, ValNo); |
2775 | // After creating a new PHI for ValNo, the phi translate result for ValNo will |
2776 | // be changed, so erase the related stale entries in phi translate cache. |
2777 | VN.eraseTranslateCacheEntry(ValNo, *CurrentBlock); |
2778 | addToLeaderTable(ValNo, Phi, CurrentBlock); |
2779 | Phi->setDebugLoc(CurInst->getDebugLoc()); |
2780 | CurInst->replaceAllUsesWith(Phi); |
2781 | if (MD && Phi->getType()->isPtrOrPtrVectorTy()) |
2782 | MD->invalidateCachedPointerInfo(Phi); |
2783 | VN.erase(CurInst); |
2784 | removeFromLeaderTable(ValNo, CurInst, CurrentBlock); |
2785 | |
2786 | LLVM_DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { dbgs() << "GVN PRE removed: " << *CurInst << '\n'; } } while (false); |
2787 | if (MD) |
2788 | MD->removeInstruction(CurInst); |
2789 | if (MSSAU) |
2790 | MSSAU->removeMemoryAccess(CurInst); |
2791 | LLVM_DEBUG(verifyRemoved(CurInst))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("gvn")) { verifyRemoved(CurInst); } } while (false); |
2792 | // FIXME: Intended to be markInstructionForDeletion(CurInst), but it causes |
2793 | // some assertion failures. |
2794 | ICF->removeInstruction(CurInst); |
2795 | CurInst->eraseFromParent(); |
2796 | ++NumGVNInstr; |
2797 | |
2798 | return true; |
2799 | } |
2800 | |
2801 | /// Perform a purely local form of PRE that looks for diamond |
2802 | /// control flow patterns and attempts to perform simple PRE at the join point. |
2803 | bool GVN::performPRE(Function &F) { |
2804 | bool Changed = false; |
2805 | for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) { |
2806 | // Nothing to PRE in the entry block. |
2807 | if (CurrentBlock == &F.getEntryBlock()) |
2808 | continue; |
2809 | |
2810 | // Don't perform PRE on an EH pad. |
2811 | if (CurrentBlock->isEHPad()) |
2812 | continue; |
2813 | |
2814 | for (BasicBlock::iterator BI = CurrentBlock->begin(), |
2815 | BE = CurrentBlock->end(); |
2816 | BI != BE;) { |
2817 | Instruction *CurInst = &*BI++; |
2818 | Changed |= performScalarPRE(CurInst); |
2819 | } |
2820 | } |
2821 | |
2822 | if (splitCriticalEdges()) |
2823 | Changed = true; |
2824 | |
2825 | return Changed; |
2826 | } |
2827 | |
2828 | /// Split the critical edge connecting the given two blocks, and return |
2829 | /// the block inserted to the critical edge. |
2830 | BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) { |
2831 | // GVN does not require loop-simplify, do not try to preserve it if it is not |
2832 | // possible. |
2833 | BasicBlock *BB = SplitCriticalEdge( |
2834 | Pred, Succ, |
2835 | CriticalEdgeSplittingOptions(DT, LI, MSSAU).unsetPreserveLoopSimplify()); |
2836 | if (BB) { |
2837 | if (MD) |
2838 | MD->invalidateCachedPredecessors(); |
2839 | InvalidBlockRPONumbers = true; |
2840 | } |
2841 | return BB; |
2842 | } |
2843 | |
2844 | /// Split critical edges found during the previous |
2845 | /// iteration that may enable further optimization. |
2846 | bool GVN::splitCriticalEdges() { |
2847 | if (toSplit.empty()) |
2848 | return false; |
2849 | |
2850 | bool Changed = false; |
2851 | do { |
2852 | std::pair<Instruction *, unsigned> Edge = toSplit.pop_back_val(); |
2853 | Changed |= SplitCriticalEdge(Edge.first, Edge.second, |
2854 | CriticalEdgeSplittingOptions(DT, LI, MSSAU)) != |
2855 | nullptr; |
2856 | } while (!toSplit.empty()); |
2857 | if (Changed) { |
2858 | if (MD) |
2859 | MD->invalidateCachedPredecessors(); |
2860 | InvalidBlockRPONumbers = true; |
2861 | } |
2862 | return Changed; |
2863 | } |
2864 | |
2865 | /// Executes one iteration of GVN |
2866 | bool GVN::iterateOnFunction(Function &F) { |
2867 | cleanupGlobalSets(); |
2868 | |
2869 | // Top-down walk of the dominator tree |
2870 | bool Changed = false; |
2871 | // Needed for value numbering with phi construction to work. |
2872 | // RPOT walks the graph in its constructor and will not be invalidated during |
2873 | // processBlock. |
2874 | ReversePostOrderTraversal<Function *> RPOT(&F); |
2875 | |
2876 | for (BasicBlock *BB : RPOT) |
2877 | Changed |= processBlock(BB); |
2878 | |
2879 | return Changed; |
2880 | } |
2881 | |
2882 | void GVN::cleanupGlobalSets() { |
2883 | VN.clear(); |
2884 | LeaderTable.clear(); |
2885 | BlockRPONumber.clear(); |
2886 | TableAllocator.Reset(); |
2887 | ICF->clear(); |
2888 | InvalidBlockRPONumbers = true; |
2889 | } |
2890 | |
2891 | /// Verify that the specified instruction does not occur in our |
2892 | /// internal data structures. |
2893 | void GVN::verifyRemoved(const Instruction *Inst) const { |
2894 | VN.verifyRemoved(Inst); |
2895 | |
2896 | // Walk through the value number scope to make sure the instruction isn't |
2897 | // ferreted away in it. |
2898 | for (const auto &I : LeaderTable) { |
2899 | const LeaderTableEntry *Node = &I.second; |
2900 | assert(Node->Val != Inst && "Inst still in value numbering scope!")(static_cast <bool> (Node->Val != Inst && "Inst still in value numbering scope!" ) ? void (0) : __assert_fail ("Node->Val != Inst && \"Inst still in value numbering scope!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2900, __extension__ __PRETTY_FUNCTION__)); |
2901 | |
2902 | while (Node->Next) { |
2903 | Node = Node->Next; |
2904 | assert(Node->Val != Inst && "Inst still in value numbering scope!")(static_cast <bool> (Node->Val != Inst && "Inst still in value numbering scope!" ) ? void (0) : __assert_fail ("Node->Val != Inst && \"Inst still in value numbering scope!\"" , "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/Scalar/GVN.cpp" , 2904, __extension__ __PRETTY_FUNCTION__)); |
2905 | } |
2906 | } |
2907 | } |
2908 | |
2909 | /// BB is declared dead, which implied other blocks become dead as well. This |
2910 | /// function is to add all these blocks to "DeadBlocks". For the dead blocks' |
2911 | /// live successors, update their phi nodes by replacing the operands |
2912 | /// corresponding to dead blocks with UndefVal. |
2913 | void GVN::addDeadBlock(BasicBlock *BB) { |
2914 | SmallVector<BasicBlock *, 4> NewDead; |
2915 | SmallSetVector<BasicBlock *, 4> DF; |
2916 | |
2917 | NewDead.push_back(BB); |
2918 | while (!NewDead.empty()) { |
2919 | BasicBlock *D = NewDead.pop_back_val(); |
2920 | if (DeadBlocks.count(D)) |
2921 | continue; |
2922 | |
2923 | // All blocks dominated by D are dead. |
2924 | SmallVector<BasicBlock *, 8> Dom; |
2925 | DT->getDescendants(D, Dom); |
2926 | DeadBlocks.insert(Dom.begin(), Dom.end()); |
2927 | |
2928 | // Figure out the dominance-frontier(D). |
2929 | for (BasicBlock *B : Dom) { |
2930 | for (BasicBlock *S : successors(B)) { |
2931 | if (DeadBlocks.count(S)) |
2932 | continue; |
2933 | |
2934 | bool AllPredDead = true; |
2935 | for (BasicBlock *P : predecessors(S)) |
2936 | if (!DeadBlocks.count(P)) { |
2937 | AllPredDead = false; |
2938 | break; |
2939 | } |
2940 | |
2941 | if (!AllPredDead) { |
2942 | // S could be proved dead later on. That is why we don't update phi |
2943 | // operands at this moment. |
2944 | DF.insert(S); |
2945 | } else { |
2946 | // While S is not dominated by D, it is dead by now. This could take |
2947 | // place if S already have a dead predecessor before D is declared |
2948 | // dead. |
2949 | NewDead.push_back(S); |
2950 | } |
2951 | } |
2952 | } |
2953 | } |
2954 | |
2955 | // For the dead blocks' live successors, update their phi nodes by replacing |
2956 | // the operands corresponding to dead blocks with UndefVal. |
2957 | for (BasicBlock *B : DF) { |
2958 | if (DeadBlocks.count(B)) |
2959 | continue; |
2960 | |
2961 | // First, split the critical edges. This might also create additional blocks |
2962 | // to preserve LoopSimplify form and adjust edges accordingly. |
2963 | SmallVector<BasicBlock *, 4> Preds(predecessors(B)); |
2964 | for (BasicBlock *P : Preds) { |
2965 | if (!DeadBlocks.count(P)) |
2966 | continue; |
2967 | |
2968 | if (llvm::is_contained(successors(P), B) && |
2969 | isCriticalEdge(P->getTerminator(), B)) { |
2970 | if (BasicBlock *S = splitCriticalEdges(P, B)) |
2971 | DeadBlocks.insert(P = S); |
Although the value stored to 'P' is used in the enclosing expression, the value is never actually read from 'P' | |
2972 | } |
2973 | } |
2974 | |
2975 | // Now undef the incoming values from the dead predecessors. |
2976 | for (BasicBlock *P : predecessors(B)) { |
2977 | if (!DeadBlocks.count(P)) |
2978 | continue; |
2979 | for (PHINode &Phi : B->phis()) { |
2980 | Phi.setIncomingValueForBlock(P, UndefValue::get(Phi.getType())); |
2981 | if (MD) |
2982 | MD->invalidateCachedPointerInfo(&Phi); |
2983 | } |
2984 | } |
2985 | } |
2986 | } |
2987 | |
2988 | // If the given branch is recognized as a foldable branch (i.e. conditional |
2989 | // branch with constant condition), it will perform following analyses and |
2990 | // transformation. |
2991 | // 1) If the dead out-coming edge is a critical-edge, split it. Let |
2992 | // R be the target of the dead out-coming edge. |
2993 | // 1) Identify the set of dead blocks implied by the branch's dead outcoming |
2994 | // edge. The result of this step will be {X| X is dominated by R} |
2995 | // 2) Identify those blocks which haves at least one dead predecessor. The |
2996 | // result of this step will be dominance-frontier(R). |
2997 | // 3) Update the PHIs in DF(R) by replacing the operands corresponding to |
2998 | // dead blocks with "UndefVal" in an hope these PHIs will optimized away. |
2999 | // |
3000 | // Return true iff *NEW* dead code are found. |
3001 | bool GVN::processFoldableCondBr(BranchInst *BI) { |
3002 | if (!BI || BI->isUnconditional()) |
3003 | return false; |
3004 | |
3005 | // If a branch has two identical successors, we cannot declare either dead. |
3006 | if (BI->getSuccessor(0) == BI->getSuccessor(1)) |
3007 | return false; |
3008 | |
3009 | ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition()); |
3010 | if (!Cond) |
3011 | return false; |
3012 | |
3013 | BasicBlock *DeadRoot = |
3014 | Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0); |
3015 | if (DeadBlocks.count(DeadRoot)) |
3016 | return false; |
3017 | |
3018 | if (!DeadRoot->getSinglePredecessor()) |
3019 | DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot); |
3020 | |
3021 | addDeadBlock(DeadRoot); |
3022 | return true; |
3023 | } |
3024 | |
3025 | // performPRE() will trigger assert if it comes across an instruction without |
3026 | // associated val-num. As it normally has far more live instructions than dead |
3027 | // instructions, it makes more sense just to "fabricate" a val-number for the |
3028 | // dead code than checking if instruction involved is dead or not. |
3029 | void GVN::assignValNumForDeadCode() { |
3030 | for (BasicBlock *BB : DeadBlocks) { |
3031 | for (Instruction &Inst : *BB) { |
3032 | unsigned ValNum = VN.lookupOrAdd(&Inst); |
3033 | addToLeaderTable(ValNum, &Inst, BB); |
3034 | } |
3035 | } |
3036 | } |
3037 | |
3038 | class llvm::gvn::GVNLegacyPass : public FunctionPass { |
3039 | public: |
3040 | static char ID; // Pass identification, replacement for typeid |
3041 | |
3042 | explicit GVNLegacyPass(bool NoMemDepAnalysis = !GVNEnableMemDep) |
3043 | : FunctionPass(ID), Impl(GVNOptions().setMemDep(!NoMemDepAnalysis)) { |
3044 | initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry()); |
3045 | } |
3046 | |
3047 | bool runOnFunction(Function &F) override { |
3048 | if (skipFunction(F)) |
3049 | return false; |
3050 | |
3051 | auto *LIWP = getAnalysisIfAvailable<LoopInfoWrapperPass>(); |
3052 | |
3053 | auto *MSSAWP = getAnalysisIfAvailable<MemorySSAWrapperPass>(); |
3054 | return Impl.runImpl( |
3055 | F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F), |
3056 | getAnalysis<DominatorTreeWrapperPass>().getDomTree(), |
3057 | getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F), |
3058 | getAnalysis<AAResultsWrapperPass>().getAAResults(), |
3059 | Impl.isMemDepEnabled() |
3060 | ? &getAnalysis<MemoryDependenceWrapperPass>().getMemDep() |
3061 | : nullptr, |
3062 | LIWP ? &LIWP->getLoopInfo() : nullptr, |
3063 | &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(), |
3064 | MSSAWP ? &MSSAWP->getMSSA() : nullptr); |
3065 | } |
3066 | |
3067 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
3068 | AU.addRequired<AssumptionCacheTracker>(); |
3069 | AU.addRequired<DominatorTreeWrapperPass>(); |
3070 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
3071 | AU.addRequired<LoopInfoWrapperPass>(); |
3072 | if (Impl.isMemDepEnabled()) |
3073 | AU.addRequired<MemoryDependenceWrapperPass>(); |
3074 | AU.addRequired<AAResultsWrapperPass>(); |
3075 | AU.addPreserved<DominatorTreeWrapperPass>(); |
3076 | AU.addPreserved<GlobalsAAWrapperPass>(); |
3077 | AU.addPreserved<TargetLibraryInfoWrapperPass>(); |
3078 | AU.addPreserved<LoopInfoWrapperPass>(); |
3079 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); |
3080 | AU.addPreserved<MemorySSAWrapperPass>(); |
3081 | } |
3082 | |
3083 | private: |
3084 | GVN Impl; |
3085 | }; |
3086 | |
3087 | char GVNLegacyPass::ID = 0; |
3088 | |
3089 | INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)static void *initializeGVNLegacyPassPassOnce(PassRegistry & Registry) { |
3090 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); |
3091 | INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)initializeMemoryDependenceWrapperPassPass(Registry); |
3092 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); |
3093 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); |
3094 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); |
3095 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry); |
3096 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); |
3097 | INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)PassInfo *PI = new PassInfo( "Global Value Numbering", "gvn", &GVNLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <GVNLegacyPass>), false, false); Registry.registerPass( *PI, true); return PI; } static llvm::once_flag InitializeGVNLegacyPassPassFlag ; void llvm::initializeGVNLegacyPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeGVNLegacyPassPassFlag, initializeGVNLegacyPassPassOnce , std::ref(Registry)); } |
3098 | |
3099 | // The public interface to this file... |
3100 | FunctionPass *llvm::createGVNPass(bool NoMemDepAnalysis) { |
3101 | return new GVNLegacyPass(NoMemDepAnalysis); |
3102 | } |