| 1 | /* |
| 2 | * Copyright (C) 2016-2017 Apple Inc. All rights reserved. |
| 3 | * |
| 4 | * Redistribution and use in source and binary forms, with or without |
| 5 | * modification, are permitted provided that the following conditions |
| 6 | * are met: |
| 7 | * 1. Redistributions of source code must retain the above copyright |
| 8 | * notice, this list of conditions and the following disclaimer. |
| 9 | * 2. Redistributions in binary form must reproduce the above copyright |
| 10 | * notice, this list of conditions and the following disclaimer in the |
| 11 | * documentation and/or other materials provided with the distribution. |
| 12 | * |
| 13 | * THIS SOFTWARE IS PROVIDED BY APPLE INC. ``AS IS'' AND ANY |
| 14 | * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 15 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR |
| 16 | * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL APPLE INC. OR |
| 17 | * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, |
| 18 | * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
| 19 | * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
| 20 | * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY |
| 21 | * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
| 22 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
| 23 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 24 | */ |
| 25 | |
| 26 | #include "config.h" |
| 27 | #include "B3EliminateCommonSubexpressions.h" |
| 28 | |
| 29 | #if ENABLE(B3_JIT) |
| 30 | |
| 31 | #include "B3BasicBlockInlines.h" |
| 32 | #include "B3BlockWorklist.h" |
| 33 | #include "B3Dominators.h" |
| 34 | #include "B3HeapRange.h" |
| 35 | #include "B3InsertionSetInlines.h" |
| 36 | #include "B3MemoryValue.h" |
| 37 | #include "B3PhaseScope.h" |
| 38 | #include "B3ProcedureInlines.h" |
| 39 | #include "B3PureCSE.h" |
| 40 | #include "B3SlotBaseValue.h" |
| 41 | #include "B3StackSlot.h" |
| 42 | #include "B3ValueKey.h" |
| 43 | #include "B3ValueInlines.h" |
| 44 | #include "B3Variable.h" |
| 45 | #include "B3VariableValue.h" |
| 46 | #include "DFGGraph.h" |
| 47 | #include <wtf/CommaPrinter.h> |
| 48 | #include <wtf/HashMap.h> |
| 49 | #include <wtf/ListDump.h> |
| 50 | #include <wtf/RangeSet.h> |
| 51 | |
| 52 | namespace JSC { namespace B3 { |
| 53 | |
| 54 | namespace { |
| 55 | |
| 56 | namespace B3EliminateCommonSubexpressionsInternal { |
| 57 | static const bool verbose = false; |
| 58 | } |
| 59 | |
| 60 | // FIXME: We could treat Patchpoints with a non-empty set of reads as a "memory value" and somehow |
| 61 | // eliminate redundant ones. We would need some way of determining if two patchpoints are replacable. |
| 62 | // It doesn't seem right to use the reads set for this. We could use the generator, but that feels |
| 63 | // lame because the FTL will pretty much use a unique generator for each patchpoint even when two |
| 64 | // patchpoints have the same semantics as far as CSE would be concerned. We could invent something |
| 65 | // like a "value ID" for patchpoints. By default, each one gets a unique value ID, but FTL could force |
| 66 | // some patchpoints to share the same one as a signal that they will return the same value if executed |
| 67 | // in the same heap with the same inputs. |
| 68 | |
| 69 | typedef Vector<MemoryValue*, 1> MemoryMatches; |
| 70 | |
| 71 | class MemoryValueMap { |
| 72 | public: |
| 73 | MemoryValueMap() { } |
| 74 | |
| 75 | void add(MemoryValue* memory) |
| 76 | { |
| 77 | Matches& matches = m_map.add(memory->lastChild(), Matches()).iterator->value; |
| 78 | if (matches.contains(memory)) |
| 79 | return; |
| 80 | matches.append(memory); |
| 81 | } |
| 82 | |
| 83 | template<typename Functor> |
| 84 | void removeIf(const Functor& functor) |
| 85 | { |
| 86 | m_map.removeIf( |
| 87 | [&] (HashMap<Value*, Matches>::KeyValuePairType& entry) -> bool { |
| 88 | entry.value.removeAllMatching( |
| 89 | [&] (Value* value) -> bool { |
| 90 | if (MemoryValue* memory = value->as<MemoryValue>()) |
| 91 | return functor(memory); |
| 92 | return true; |
| 93 | }); |
| 94 | return entry.value.isEmpty(); |
| 95 | }); |
| 96 | } |
| 97 | |
| 98 | Matches* find(Value* ptr) |
| 99 | { |
| 100 | auto iter = m_map.find(ptr); |
| 101 | if (iter == m_map.end()) |
| 102 | return nullptr; |
| 103 | return &iter->value; |
| 104 | } |
| 105 | |
| 106 | template<typename Functor> |
| 107 | MemoryValue* find(Value* ptr, const Functor& functor) |
| 108 | { |
| 109 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 110 | dataLog(" Looking for " , pointerDump(ptr), " in " , *this, "\n" ); |
| 111 | if (Matches* matches = find(ptr)) { |
| 112 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 113 | dataLog(" Matches: " , pointerListDump(*matches), "\n" ); |
| 114 | for (Value* candidateValue : *matches) { |
| 115 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 116 | dataLog(" Having candidate: " , pointerDump(candidateValue), "\n" ); |
| 117 | if (MemoryValue* candidateMemory = candidateValue->as<MemoryValue>()) { |
| 118 | if (functor(candidateMemory)) |
| 119 | return candidateMemory; |
| 120 | } |
| 121 | } |
| 122 | } |
| 123 | return nullptr; |
| 124 | } |
| 125 | |
| 126 | void dump(PrintStream& out) const |
| 127 | { |
| 128 | out.print("{" ); |
| 129 | CommaPrinter comma; |
| 130 | for (auto& entry : m_map) |
| 131 | out.print(comma, pointerDump(entry.key), "=>" , pointerListDump(entry.value)); |
| 132 | out.print("}" ); |
| 133 | } |
| 134 | |
| 135 | private: |
| 136 | // This uses Matches for two reasons: |
| 137 | // - It cannot be a MemoryValue* because the key is imprecise. Many MemoryValues could have the |
| 138 | // same key while being unaliased. |
| 139 | // - It can't be a MemoryMatches array because the MemoryValue*'s could be turned into Identity's. |
| 140 | HashMap<Value*, Matches> m_map; |
| 141 | }; |
| 142 | |
| 143 | struct ImpureBlockData { |
| 144 | void dump(PrintStream& out) const |
| 145 | { |
| 146 | out.print( |
| 147 | "{reads = " , reads, ", writes = " , writes, ", storesAtHead = " , storesAtHead, |
| 148 | ", memoryValuesAtTail = " , memoryValuesAtTail, "}" ); |
| 149 | } |
| 150 | |
| 151 | RangeSet<HeapRange> reads; // This only gets used for forward store elimination. |
| 152 | RangeSet<HeapRange> writes; // This gets used for both load and store elimination. |
| 153 | bool fence; |
| 154 | |
| 155 | MemoryValueMap storesAtHead; |
| 156 | MemoryValueMap memoryValuesAtTail; |
| 157 | }; |
| 158 | |
| 159 | class CSE { |
| 160 | public: |
| 161 | CSE(Procedure& proc) |
| 162 | : m_proc(proc) |
| 163 | , m_dominators(proc.dominators()) |
| 164 | , m_impureBlockData(proc.size()) |
| 165 | , m_insertionSet(proc) |
| 166 | { |
| 167 | } |
| 168 | |
| 169 | bool run() |
| 170 | { |
| 171 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 172 | dataLog("B3 before CSE:\n" , m_proc); |
| 173 | |
| 174 | m_proc.resetValueOwners(); |
| 175 | |
| 176 | // Summarize the impure effects of each block, and the impure values available at the end of |
| 177 | // each block. This doesn't edit code yet. |
| 178 | for (BasicBlock* block : m_proc) { |
| 179 | ImpureBlockData& data = m_impureBlockData[block]; |
| 180 | for (Value* value : *block) { |
| 181 | Effects effects = value->effects(); |
| 182 | MemoryValue* memory = value->as<MemoryValue>(); |
| 183 | |
| 184 | if (memory && memory->isStore() |
| 185 | && !data.reads.overlaps(memory->range()) |
| 186 | && !data.writes.overlaps(memory->range()) |
| 187 | && (!data.fence || !memory->hasFence())) |
| 188 | data.storesAtHead.add(memory); |
| 189 | data.reads.add(effects.reads); |
| 190 | |
| 191 | if (HeapRange writes = effects.writes) |
| 192 | clobber(data, writes); |
| 193 | data.fence |= effects.fence; |
| 194 | |
| 195 | if (memory) |
| 196 | data.memoryValuesAtTail.add(memory); |
| 197 | } |
| 198 | |
| 199 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 200 | dataLog("Block " , *block, ": " , data, "\n" ); |
| 201 | } |
| 202 | |
| 203 | // Perform CSE. This edits code. |
| 204 | Vector<BasicBlock*> postOrder = m_proc.blocksInPostOrder(); |
| 205 | for (unsigned i = postOrder.size(); i--;) { |
| 206 | m_block = postOrder[i]; |
| 207 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 208 | dataLog("Looking at " , *m_block, ":\n" ); |
| 209 | |
| 210 | m_data = ImpureBlockData(); |
| 211 | for (m_index = 0; m_index < m_block->size(); ++m_index) { |
| 212 | m_value = m_block->at(m_index); |
| 213 | process(); |
| 214 | } |
| 215 | m_insertionSet.execute(m_block); |
| 216 | m_impureBlockData[m_block] = m_data; |
| 217 | } |
| 218 | |
| 219 | // The previous pass might have requested that we insert code in some basic block other than |
| 220 | // the one that it was looking at. This inserts them. |
| 221 | for (BasicBlock* block : m_proc) { |
| 222 | for (unsigned valueIndex = 0; valueIndex < block->size(); ++valueIndex) { |
| 223 | auto iter = m_sets.find(block->at(valueIndex)); |
| 224 | if (iter == m_sets.end()) |
| 225 | continue; |
| 226 | |
| 227 | for (Value* value : iter->value) |
| 228 | m_insertionSet.insertValue(valueIndex + 1, value); |
| 229 | } |
| 230 | m_insertionSet.execute(block); |
| 231 | } |
| 232 | |
| 233 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 234 | dataLog("B3 after CSE:\n" , m_proc); |
| 235 | |
| 236 | return m_changed; |
| 237 | } |
| 238 | |
| 239 | private: |
| 240 | void process() |
| 241 | { |
| 242 | m_value->performSubstitution(); |
| 243 | |
| 244 | if (m_pureCSE.process(m_value, m_dominators)) { |
| 245 | ASSERT(!m_value->effects().writes); |
| 246 | m_changed = true; |
| 247 | return; |
| 248 | } |
| 249 | |
| 250 | MemoryValue* memory = m_value->as<MemoryValue>(); |
| 251 | if (memory && processMemoryBeforeClobber(memory)) |
| 252 | return; |
| 253 | |
| 254 | if (HeapRange writes = m_value->effects().writes) |
| 255 | clobber(m_data, writes); |
| 256 | |
| 257 | if (memory) |
| 258 | processMemoryAfterClobber(memory); |
| 259 | } |
| 260 | |
| 261 | // Return true if we got rid of the operation. If you changed IR in this function, you have to |
| 262 | // set m_changed even if you also return true. |
| 263 | bool processMemoryBeforeClobber(MemoryValue* memory) |
| 264 | { |
| 265 | Value* value = memory->child(0); |
| 266 | Value* ptr = memory->lastChild(); |
| 267 | HeapRange range = memory->range(); |
| 268 | Value::OffsetType offset = memory->offset(); |
| 269 | |
| 270 | switch (memory->opcode()) { |
| 271 | case Store8: |
| 272 | return handleStoreBeforeClobber( |
| 273 | ptr, range, |
| 274 | [&] (MemoryValue* candidate) -> bool { |
| 275 | return candidate->offset() == offset |
| 276 | && ((candidate->opcode() == Store8 && candidate->child(0) == value) |
| 277 | || ((candidate->opcode() == Load8Z || candidate->opcode() == Load8S) |
| 278 | && candidate == value)); |
| 279 | }); |
| 280 | case Store16: |
| 281 | return handleStoreBeforeClobber( |
| 282 | ptr, range, |
| 283 | [&] (MemoryValue* candidate) -> bool { |
| 284 | return candidate->offset() == offset |
| 285 | && ((candidate->opcode() == Store16 && candidate->child(0) == value) |
| 286 | || ((candidate->opcode() == Load16Z || candidate->opcode() == Load16S) |
| 287 | && candidate == value)); |
| 288 | }); |
| 289 | case Store: |
| 290 | return handleStoreBeforeClobber( |
| 291 | ptr, range, |
| 292 | [&] (MemoryValue* candidate) -> bool { |
| 293 | return candidate->offset() == offset |
| 294 | && ((candidate->opcode() == Store && candidate->child(0) == value) |
| 295 | || (candidate->opcode() == Load && candidate == value)); |
| 296 | }); |
| 297 | default: |
| 298 | return false; |
| 299 | } |
| 300 | } |
| 301 | |
| 302 | void clobber(ImpureBlockData& data, HeapRange writes) |
| 303 | { |
| 304 | data.writes.add(writes); |
| 305 | |
| 306 | data.memoryValuesAtTail.removeIf( |
| 307 | [&] (MemoryValue* memory) { |
| 308 | return memory->range().overlaps(writes); |
| 309 | }); |
| 310 | } |
| 311 | |
| 312 | void processMemoryAfterClobber(MemoryValue* memory) |
| 313 | { |
| 314 | Value* ptr = memory->lastChild(); |
| 315 | HeapRange range = memory->range(); |
| 316 | Value::OffsetType offset = memory->offset(); |
| 317 | Type type = memory->type(); |
| 318 | |
| 319 | // FIXME: Empower this to insert more casts and shifts. For example, a Load8 could match a |
| 320 | // Store and mask the result. You could even have: |
| 321 | // |
| 322 | // Store(@value, @ptr, offset = 0) |
| 323 | // Load8Z(@ptr, offset = 2) |
| 324 | // |
| 325 | // Which could be turned into something like this: |
| 326 | // |
| 327 | // Store(@value, @ptr, offset = 0) |
| 328 | // ZShr(@value, 16) |
| 329 | |
| 330 | switch (memory->opcode()) { |
| 331 | case Load8Z: { |
| 332 | handleMemoryValue( |
| 333 | ptr, range, |
| 334 | [&] (MemoryValue* candidate) -> bool { |
| 335 | return candidate->offset() == offset |
| 336 | && (candidate->opcode() == Load8Z || candidate->opcode() == Store8); |
| 337 | }, |
| 338 | [&] (MemoryValue* match, Vector<Value*>& fixups) -> Value* { |
| 339 | if (match->opcode() == Store8) { |
| 340 | Value* mask = m_proc.add<Const32Value>(m_value->origin(), 0xff); |
| 341 | fixups.append(mask); |
| 342 | Value* zext = m_proc.add<Value>( |
| 343 | BitAnd, m_value->origin(), match->child(0), mask); |
| 344 | fixups.append(zext); |
| 345 | return zext; |
| 346 | } |
| 347 | return nullptr; |
| 348 | }); |
| 349 | break; |
| 350 | } |
| 351 | |
| 352 | case Load8S: { |
| 353 | handleMemoryValue( |
| 354 | ptr, range, |
| 355 | [&] (MemoryValue* candidate) -> bool { |
| 356 | return candidate->offset() == offset |
| 357 | && (candidate->opcode() == Load8S || candidate->opcode() == Store8); |
| 358 | }, |
| 359 | [&] (MemoryValue* match, Vector<Value*>& fixups) -> Value* { |
| 360 | if (match->opcode() == Store8) { |
| 361 | Value* sext = m_proc.add<Value>( |
| 362 | SExt8, m_value->origin(), match->child(0)); |
| 363 | fixups.append(sext); |
| 364 | return sext; |
| 365 | } |
| 366 | return nullptr; |
| 367 | }); |
| 368 | break; |
| 369 | } |
| 370 | |
| 371 | case Load16Z: { |
| 372 | handleMemoryValue( |
| 373 | ptr, range, |
| 374 | [&] (MemoryValue* candidate) -> bool { |
| 375 | return candidate->offset() == offset |
| 376 | && (candidate->opcode() == Load16Z || candidate->opcode() == Store16); |
| 377 | }, |
| 378 | [&] (MemoryValue* match, Vector<Value*>& fixups) -> Value* { |
| 379 | if (match->opcode() == Store16) { |
| 380 | Value* mask = m_proc.add<Const32Value>(m_value->origin(), 0xffff); |
| 381 | fixups.append(mask); |
| 382 | Value* zext = m_proc.add<Value>( |
| 383 | BitAnd, m_value->origin(), match->child(0), mask); |
| 384 | fixups.append(zext); |
| 385 | return zext; |
| 386 | } |
| 387 | return nullptr; |
| 388 | }); |
| 389 | break; |
| 390 | } |
| 391 | |
| 392 | case Load16S: { |
| 393 | handleMemoryValue( |
| 394 | ptr, range, [&] (MemoryValue* candidate) -> bool { |
| 395 | return candidate->offset() == offset |
| 396 | && (candidate->opcode() == Load16S || candidate->opcode() == Store16); |
| 397 | }, |
| 398 | [&] (MemoryValue* match, Vector<Value*>& fixups) -> Value* { |
| 399 | if (match->opcode() == Store16) { |
| 400 | Value* sext = m_proc.add<Value>( |
| 401 | SExt16, m_value->origin(), match->child(0)); |
| 402 | fixups.append(sext); |
| 403 | return sext; |
| 404 | } |
| 405 | return nullptr; |
| 406 | }); |
| 407 | break; |
| 408 | } |
| 409 | |
| 410 | case Load: { |
| 411 | handleMemoryValue( |
| 412 | ptr, range, |
| 413 | [&] (MemoryValue* candidate) -> bool { |
| 414 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 415 | dataLog(" Consdering " , pointerDump(candidate), "\n" ); |
| 416 | if (candidate->offset() != offset) |
| 417 | return false; |
| 418 | |
| 419 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 420 | dataLog(" offset ok.\n" ); |
| 421 | |
| 422 | if (candidate->opcode() == Load && candidate->type() == type) |
| 423 | return true; |
| 424 | |
| 425 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 426 | dataLog(" not a load with ok type.\n" ); |
| 427 | |
| 428 | if (candidate->opcode() == Store && candidate->child(0)->type() == type) |
| 429 | return true; |
| 430 | |
| 431 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 432 | dataLog(" not a store with ok type.\n" ); |
| 433 | |
| 434 | return false; |
| 435 | }); |
| 436 | break; |
| 437 | } |
| 438 | |
| 439 | case Store8: { |
| 440 | handleStoreAfterClobber( |
| 441 | ptr, range, |
| 442 | [&] (MemoryValue* candidate) -> bool { |
| 443 | return candidate->opcode() == Store8 |
| 444 | && candidate->offset() == offset; |
| 445 | }); |
| 446 | break; |
| 447 | } |
| 448 | |
| 449 | case Store16: { |
| 450 | handleStoreAfterClobber( |
| 451 | ptr, range, |
| 452 | [&] (MemoryValue* candidate) -> bool { |
| 453 | return candidate->opcode() == Store16 |
| 454 | && candidate->offset() == offset; |
| 455 | }); |
| 456 | break; |
| 457 | } |
| 458 | |
| 459 | case Store: { |
| 460 | handleStoreAfterClobber( |
| 461 | ptr, range, |
| 462 | [&] (MemoryValue* candidate) -> bool { |
| 463 | return candidate->opcode() == Store |
| 464 | && candidate->offset() == offset; |
| 465 | }); |
| 466 | break; |
| 467 | } |
| 468 | |
| 469 | default: |
| 470 | break; |
| 471 | } |
| 472 | } |
| 473 | |
| 474 | template<typename Filter> |
| 475 | bool handleStoreBeforeClobber(Value* ptr, HeapRange range, const Filter& filter) |
| 476 | { |
| 477 | MemoryMatches matches = findMemoryValue(ptr, range, filter); |
| 478 | if (matches.isEmpty()) |
| 479 | return false; |
| 480 | |
| 481 | m_value->replaceWithNop(); |
| 482 | m_changed = true; |
| 483 | return true; |
| 484 | } |
| 485 | |
| 486 | template<typename Filter> |
| 487 | void handleStoreAfterClobber(Value* ptr, HeapRange range, const Filter& filter) |
| 488 | { |
| 489 | if (!m_value->traps() && findStoreAfterClobber(ptr, range, filter)) { |
| 490 | m_value->replaceWithNop(); |
| 491 | m_changed = true; |
| 492 | return; |
| 493 | } |
| 494 | |
| 495 | m_data.memoryValuesAtTail.add(m_value->as<MemoryValue>()); |
| 496 | } |
| 497 | |
| 498 | template<typename Filter> |
| 499 | bool findStoreAfterClobber(Value* ptr, HeapRange range, const Filter& filter) |
| 500 | { |
| 501 | if (m_value->as<MemoryValue>()->hasFence()) |
| 502 | return false; |
| 503 | |
| 504 | // We can eliminate a store if every forward path hits a store to the same location before |
| 505 | // hitting any operation that observes the store. This search seems like it should be |
| 506 | // expensive, but in the overwhelming majority of cases it will almost immediately hit an |
| 507 | // operation that interferes. |
| 508 | |
| 509 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 510 | dataLog(*m_value, ": looking forward for stores to " , *ptr, "...\n" ); |
| 511 | |
| 512 | // First search forward in this basic block. |
| 513 | // FIXME: It would be cool to get rid of this linear search. It's not super critical since |
| 514 | // we will probably bail out very quickly, but it *is* annoying. |
| 515 | for (unsigned index = m_index + 1; index < m_block->size(); ++index) { |
| 516 | Value* value = m_block->at(index); |
| 517 | |
| 518 | if (MemoryValue* memoryValue = value->as<MemoryValue>()) { |
| 519 | if (memoryValue->lastChild() == ptr && filter(memoryValue)) |
| 520 | return true; |
| 521 | } |
| 522 | |
| 523 | Effects effects = value->effects(); |
| 524 | if (effects.reads.overlaps(range) || effects.writes.overlaps(range)) |
| 525 | return false; |
| 526 | } |
| 527 | |
| 528 | if (!m_block->numSuccessors()) |
| 529 | return false; |
| 530 | |
| 531 | BlockWorklist worklist; |
| 532 | worklist.pushAll(m_block->successorBlocks()); |
| 533 | |
| 534 | while (BasicBlock* block = worklist.pop()) { |
| 535 | ImpureBlockData& data = m_impureBlockData[block]; |
| 536 | |
| 537 | MemoryValue* match = data.storesAtHead.find(ptr, filter); |
| 538 | if (match && match != m_value) |
| 539 | continue; |
| 540 | |
| 541 | if (data.writes.overlaps(range) || data.reads.overlaps(range)) |
| 542 | return false; |
| 543 | |
| 544 | if (!block->numSuccessors()) |
| 545 | return false; |
| 546 | |
| 547 | worklist.pushAll(block->successorBlocks()); |
| 548 | } |
| 549 | |
| 550 | return true; |
| 551 | } |
| 552 | |
| 553 | template<typename Filter> |
| 554 | void handleMemoryValue(Value* ptr, HeapRange range, const Filter& filter) |
| 555 | { |
| 556 | handleMemoryValue( |
| 557 | ptr, range, filter, |
| 558 | [] (MemoryValue*, Vector<Value*>&) -> Value* { |
| 559 | return nullptr; |
| 560 | }); |
| 561 | } |
| 562 | |
| 563 | template<typename Filter, typename Replace> |
| 564 | void handleMemoryValue( |
| 565 | Value* ptr, HeapRange range, const Filter& filter, const Replace& replace) |
| 566 | { |
| 567 | MemoryMatches matches = findMemoryValue(ptr, range, filter); |
| 568 | if (replaceMemoryValue(matches, replace)) |
| 569 | return; |
| 570 | m_data.memoryValuesAtTail.add(m_value->as<MemoryValue>()); |
| 571 | } |
| 572 | |
| 573 | template<typename Replace> |
| 574 | bool replaceMemoryValue(const MemoryMatches& matches, const Replace& replace) |
| 575 | { |
| 576 | if (matches.isEmpty()) |
| 577 | return false; |
| 578 | |
| 579 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 580 | dataLog("Eliminating " , *m_value, " due to " , pointerListDump(matches), "\n" ); |
| 581 | |
| 582 | m_changed = true; |
| 583 | |
| 584 | if (matches.size() == 1) { |
| 585 | MemoryValue* dominatingMatch = matches[0]; |
| 586 | RELEASE_ASSERT(m_dominators.dominates(dominatingMatch->owner, m_block)); |
| 587 | |
| 588 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 589 | dataLog(" Eliminating using " , *dominatingMatch, "\n" ); |
| 590 | Vector<Value*> ; |
| 591 | if (Value* value = replace(dominatingMatch, extraValues)) { |
| 592 | for (Value* : extraValues) |
| 593 | m_insertionSet.insertValue(m_index, extraValue); |
| 594 | m_value->replaceWithIdentity(value); |
| 595 | } else { |
| 596 | if (dominatingMatch->isStore()) |
| 597 | m_value->replaceWithIdentity(dominatingMatch->child(0)); |
| 598 | else |
| 599 | m_value->replaceWithIdentity(dominatingMatch); |
| 600 | } |
| 601 | return true; |
| 602 | } |
| 603 | |
| 604 | // FIXME: It would be way better if this phase just did SSA calculation directly. |
| 605 | // Right now we're relying on the fact that CSE's position in the phase order is |
| 606 | // almost right before SSA fixup. |
| 607 | |
| 608 | Variable* variable = m_proc.addVariable(m_value->type()); |
| 609 | |
| 610 | VariableValue* get = m_insertionSet.insert<VariableValue>( |
| 611 | m_index, Get, m_value->origin(), variable); |
| 612 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 613 | dataLog(" Inserting get of value: " , *get, "\n" ); |
| 614 | m_value->replaceWithIdentity(get); |
| 615 | |
| 616 | for (MemoryValue* match : matches) { |
| 617 | Vector<Value*>& sets = m_sets.add(match, Vector<Value*>()).iterator->value; |
| 618 | |
| 619 | Value* value = replace(match, sets); |
| 620 | if (!value) { |
| 621 | if (match->isStore()) |
| 622 | value = match->child(0); |
| 623 | else |
| 624 | value = match; |
| 625 | } |
| 626 | |
| 627 | Value* set = m_proc.add<VariableValue>(Set, m_value->origin(), variable, value); |
| 628 | sets.append(set); |
| 629 | } |
| 630 | |
| 631 | return true; |
| 632 | } |
| 633 | |
| 634 | template<typename Filter> |
| 635 | MemoryMatches findMemoryValue(Value* ptr, HeapRange range, const Filter& filter) |
| 636 | { |
| 637 | if (B3EliminateCommonSubexpressionsInternal::verbose) { |
| 638 | dataLog(*m_value, ": looking backward for " , *ptr, "...\n" ); |
| 639 | dataLog(" Full value: " , deepDump(m_value), "\n" ); |
| 640 | } |
| 641 | |
| 642 | if (m_value->as<MemoryValue>()->hasFence()) { |
| 643 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 644 | dataLog(" Giving up because fences.\n" ); |
| 645 | return { }; |
| 646 | } |
| 647 | |
| 648 | if (MemoryValue* match = m_data.memoryValuesAtTail.find(ptr, filter)) { |
| 649 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 650 | dataLog(" Found " , *match, " locally.\n" ); |
| 651 | return { match }; |
| 652 | } |
| 653 | |
| 654 | if (m_data.writes.overlaps(range)) { |
| 655 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 656 | dataLog(" Giving up because of writes.\n" ); |
| 657 | return { }; |
| 658 | } |
| 659 | |
| 660 | BlockWorklist worklist; |
| 661 | worklist.pushAll(m_block->predecessors()); |
| 662 | |
| 663 | MemoryMatches matches; |
| 664 | |
| 665 | while (BasicBlock* block = worklist.pop()) { |
| 666 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 667 | dataLog(" Looking at " , *block, "\n" ); |
| 668 | |
| 669 | ImpureBlockData& data = m_impureBlockData[block]; |
| 670 | |
| 671 | MemoryValue* match = data.memoryValuesAtTail.find(ptr, filter); |
| 672 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 673 | dataLog(" Consdering match: " , pointerDump(match), "\n" ); |
| 674 | if (match && match != m_value) { |
| 675 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 676 | dataLog(" Found match: " , *match, "\n" ); |
| 677 | matches.append(match); |
| 678 | continue; |
| 679 | } |
| 680 | |
| 681 | if (data.writes.overlaps(range)) { |
| 682 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 683 | dataLog(" Giving up because of writes.\n" ); |
| 684 | return { }; |
| 685 | } |
| 686 | |
| 687 | if (!block->numPredecessors()) { |
| 688 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 689 | dataLog(" Giving up because it's live at root.\n" ); |
| 690 | // This essentially proves that this is live at the prologue. That means that we |
| 691 | // cannot reliably optimize this case. |
| 692 | return { }; |
| 693 | } |
| 694 | |
| 695 | worklist.pushAll(block->predecessors()); |
| 696 | } |
| 697 | |
| 698 | if (B3EliminateCommonSubexpressionsInternal::verbose) |
| 699 | dataLog(" Got matches: " , pointerListDump(matches), "\n" ); |
| 700 | return matches; |
| 701 | } |
| 702 | |
| 703 | Procedure& m_proc; |
| 704 | |
| 705 | Dominators& m_dominators; |
| 706 | PureCSE m_pureCSE; |
| 707 | |
| 708 | IndexMap<BasicBlock*, ImpureBlockData> m_impureBlockData; |
| 709 | |
| 710 | ImpureBlockData m_data; |
| 711 | |
| 712 | BasicBlock* m_block; |
| 713 | unsigned m_index; |
| 714 | Value* m_value; |
| 715 | |
| 716 | HashMap<Value*, Vector<Value*>> m_sets; |
| 717 | |
| 718 | InsertionSet m_insertionSet; |
| 719 | |
| 720 | bool m_changed { false }; |
| 721 | }; |
| 722 | |
| 723 | } // anonymous namespace |
| 724 | |
| 725 | bool eliminateCommonSubexpressions(Procedure& proc) |
| 726 | { |
| 727 | PhaseScope phaseScope(proc, "eliminateCommonSubexpressions" ); |
| 728 | |
| 729 | CSE cse(proc); |
| 730 | return cse.run(); |
| 731 | } |
| 732 | |
| 733 | } } // namespace JSC::B3 |
| 734 | |
| 735 | #endif // ENABLE(B3_JIT) |
| 736 | |
| 737 | |