| 1 | /* |
| 2 | * Copyright (C) 2015-2019 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 "B3LowerMacros.h" |
| 28 | |
| 29 | #if ENABLE(B3_JIT) |
| 30 | |
| 31 | #include "AllowMacroScratchRegisterUsage.h" |
| 32 | #include "B3AtomicValue.h" |
| 33 | #include "B3BasicBlockInlines.h" |
| 34 | #include "B3BlockInsertionSet.h" |
| 35 | #include "B3CCallValue.h" |
| 36 | #include "B3CaseCollectionInlines.h" |
| 37 | #include "B3ConstPtrValue.h" |
| 38 | #include "B3FenceValue.h" |
| 39 | #include "B3InsertionSetInlines.h" |
| 40 | #include "B3MemoryValueInlines.h" |
| 41 | #include "B3PatchpointValue.h" |
| 42 | #include "B3PhaseScope.h" |
| 43 | #include "B3ProcedureInlines.h" |
| 44 | #include "B3StackmapGenerationParams.h" |
| 45 | #include "B3SwitchValue.h" |
| 46 | #include "B3UpsilonValue.h" |
| 47 | #include "B3UseCounts.h" |
| 48 | #include "B3ValueInlines.h" |
| 49 | #include "CCallHelpers.h" |
| 50 | #include "LinkBuffer.h" |
| 51 | #include <cmath> |
| 52 | #include <wtf/BitVector.h> |
| 53 | |
| 54 | namespace JSC { namespace B3 { |
| 55 | |
| 56 | namespace { |
| 57 | |
| 58 | class LowerMacros { |
| 59 | public: |
| 60 | LowerMacros(Procedure& proc) |
| 61 | : m_proc(proc) |
| 62 | , m_blockInsertionSet(proc) |
| 63 | , m_insertionSet(proc) |
| 64 | , m_useCounts(proc) |
| 65 | { |
| 66 | } |
| 67 | |
| 68 | bool run() |
| 69 | { |
| 70 | RELEASE_ASSERT(!m_proc.hasQuirks()); |
| 71 | |
| 72 | for (BasicBlock* block : m_proc) { |
| 73 | m_block = block; |
| 74 | processCurrentBlock(); |
| 75 | } |
| 76 | m_changed |= m_blockInsertionSet.execute(); |
| 77 | if (m_changed) { |
| 78 | m_proc.resetReachability(); |
| 79 | m_proc.invalidateCFG(); |
| 80 | } |
| 81 | |
| 82 | // This indicates that we've |
| 83 | m_proc.setHasQuirks(true); |
| 84 | |
| 85 | return m_changed; |
| 86 | } |
| 87 | |
| 88 | private: |
| 89 | void processCurrentBlock() |
| 90 | { |
| 91 | for (m_index = 0; m_index < m_block->size(); ++m_index) { |
| 92 | m_value = m_block->at(m_index); |
| 93 | m_origin = m_value->origin(); |
| 94 | switch (m_value->opcode()) { |
| 95 | case Mod: { |
| 96 | if (m_value->isChill()) { |
| 97 | if (isARM64()) { |
| 98 | BasicBlock* before = m_blockInsertionSet.splitForward(m_block, m_index, &m_insertionSet); |
| 99 | BasicBlock* zeroDenCase = m_blockInsertionSet.insertBefore(m_block); |
| 100 | BasicBlock* normalModCase = m_blockInsertionSet.insertBefore(m_block); |
| 101 | |
| 102 | before->replaceLastWithNew<Value>(m_proc, Branch, m_origin, m_value->child(1)); |
| 103 | before->setSuccessors( |
| 104 | FrequentedBlock(normalModCase, FrequencyClass::Normal), |
| 105 | FrequentedBlock(zeroDenCase, FrequencyClass::Rare)); |
| 106 | |
| 107 | Value* divResult = normalModCase->appendNew<Value>(m_proc, chill(Div), m_origin, m_value->child(0), m_value->child(1)); |
| 108 | Value* multipliedBack = normalModCase->appendNew<Value>(m_proc, Mul, m_origin, divResult, m_value->child(1)); |
| 109 | Value* result = normalModCase->appendNew<Value>(m_proc, Sub, m_origin, m_value->child(0), multipliedBack); |
| 110 | UpsilonValue* normalResult = normalModCase->appendNew<UpsilonValue>(m_proc, m_origin, result); |
| 111 | normalModCase->appendNew<Value>(m_proc, Jump, m_origin); |
| 112 | normalModCase->setSuccessors(FrequentedBlock(m_block)); |
| 113 | |
| 114 | UpsilonValue* zeroResult = zeroDenCase->appendNew<UpsilonValue>( |
| 115 | m_proc, m_origin, |
| 116 | zeroDenCase->appendIntConstant(m_proc, m_value, 0)); |
| 117 | zeroDenCase->appendNew<Value>(m_proc, Jump, m_origin); |
| 118 | zeroDenCase->setSuccessors(FrequentedBlock(m_block)); |
| 119 | |
| 120 | Value* phi = m_insertionSet.insert<Value>(m_index, Phi, m_value->type(), m_origin); |
| 121 | normalResult->setPhi(phi); |
| 122 | zeroResult->setPhi(phi); |
| 123 | m_value->replaceWithIdentity(phi); |
| 124 | before->updatePredecessorsAfter(); |
| 125 | m_changed = true; |
| 126 | } else |
| 127 | makeDivisionChill(Mod); |
| 128 | break; |
| 129 | } |
| 130 | |
| 131 | auto* fmodDouble = tagCFunctionPtr<double (*)(double, double)>(fmod, B3CCallPtrTag); |
| 132 | if (m_value->type() == Double) { |
| 133 | Value* functionAddress = m_insertionSet.insert<ConstPtrValue>(m_index, m_origin, fmodDouble); |
| 134 | Value* result = m_insertionSet.insert<CCallValue>(m_index, Double, m_origin, |
| 135 | Effects::none(), |
| 136 | functionAddress, |
| 137 | m_value->child(0), |
| 138 | m_value->child(1)); |
| 139 | m_value->replaceWithIdentity(result); |
| 140 | m_changed = true; |
| 141 | } else if (m_value->type() == Float) { |
| 142 | Value* numeratorAsDouble = m_insertionSet.insert<Value>(m_index, FloatToDouble, m_origin, m_value->child(0)); |
| 143 | Value* denominatorAsDouble = m_insertionSet.insert<Value>(m_index, FloatToDouble, m_origin, m_value->child(1)); |
| 144 | Value* functionAddress = m_insertionSet.insert<ConstPtrValue>(m_index, m_origin, fmodDouble); |
| 145 | Value* doubleMod = m_insertionSet.insert<CCallValue>(m_index, Double, m_origin, |
| 146 | Effects::none(), |
| 147 | functionAddress, |
| 148 | numeratorAsDouble, |
| 149 | denominatorAsDouble); |
| 150 | Value* result = m_insertionSet.insert<Value>(m_index, DoubleToFloat, m_origin, doubleMod); |
| 151 | m_value->replaceWithIdentity(result); |
| 152 | m_changed = true; |
| 153 | } else if (isARM64()) { |
| 154 | Value* divResult = m_insertionSet.insert<Value>(m_index, chill(Div), m_origin, m_value->child(0), m_value->child(1)); |
| 155 | Value* multipliedBack = m_insertionSet.insert<Value>(m_index, Mul, m_origin, divResult, m_value->child(1)); |
| 156 | Value* result = m_insertionSet.insert<Value>(m_index, Sub, m_origin, m_value->child(0), multipliedBack); |
| 157 | m_value->replaceWithIdentity(result); |
| 158 | m_changed = true; |
| 159 | } |
| 160 | break; |
| 161 | } |
| 162 | |
| 163 | case UMod: { |
| 164 | if (isARM64()) { |
| 165 | Value* divResult = m_insertionSet.insert<Value>(m_index, UDiv, m_origin, m_value->child(0), m_value->child(1)); |
| 166 | Value* multipliedBack = m_insertionSet.insert<Value>(m_index, Mul, m_origin, divResult, m_value->child(1)); |
| 167 | Value* result = m_insertionSet.insert<Value>(m_index, Sub, m_origin, m_value->child(0), multipliedBack); |
| 168 | m_value->replaceWithIdentity(result); |
| 169 | m_changed = true; |
| 170 | } |
| 171 | break; |
| 172 | } |
| 173 | |
| 174 | case Div: { |
| 175 | if (m_value->isChill()) |
| 176 | makeDivisionChill(Div); |
| 177 | break; |
| 178 | } |
| 179 | |
| 180 | case CheckMul: { |
| 181 | if (isARM64() && m_value->child(0)->type() == Int32) { |
| 182 | CheckValue* checkMul = m_value->as<CheckValue>(); |
| 183 | |
| 184 | Value* left = m_insertionSet.insert<Value>(m_index, SExt32, m_origin, m_value->child(0)); |
| 185 | Value* right = m_insertionSet.insert<Value>(m_index, SExt32, m_origin, m_value->child(1)); |
| 186 | Value* mulResult = m_insertionSet.insert<Value>(m_index, Mul, m_origin, left, right); |
| 187 | Value* mulResult32 = m_insertionSet.insert<Value>(m_index, Trunc, m_origin, mulResult); |
| 188 | Value* upperResult = m_insertionSet.insert<Value>(m_index, Trunc, m_origin, |
| 189 | m_insertionSet.insert<Value>(m_index, SShr, m_origin, mulResult, m_insertionSet.insert<Const32Value>(m_index, m_origin, 32))); |
| 190 | Value* signBit = m_insertionSet.insert<Value>(m_index, SShr, m_origin, |
| 191 | mulResult32, |
| 192 | m_insertionSet.insert<Const32Value>(m_index, m_origin, 31)); |
| 193 | Value* hasOverflowed = m_insertionSet.insert<Value>(m_index, NotEqual, m_origin, upperResult, signBit); |
| 194 | |
| 195 | CheckValue* check = m_insertionSet.insert<CheckValue>(m_index, Check, m_origin, hasOverflowed); |
| 196 | check->setGenerator(checkMul->generator()); |
| 197 | check->clobberEarly(checkMul->earlyClobbered()); |
| 198 | check->clobberLate(checkMul->lateClobbered()); |
| 199 | auto children = checkMul->constrainedChildren(); |
| 200 | auto it = children.begin(); |
| 201 | for (std::advance(it, 2); it != children.end(); ++it) |
| 202 | check->append(*it); |
| 203 | |
| 204 | m_value->replaceWithIdentity(mulResult32); |
| 205 | m_changed = true; |
| 206 | } |
| 207 | break; |
| 208 | } |
| 209 | |
| 210 | case Switch: { |
| 211 | SwitchValue* switchValue = m_value->as<SwitchValue>(); |
| 212 | Vector<SwitchCase> cases; |
| 213 | for (SwitchCase switchCase : switchValue->cases(m_block)) |
| 214 | cases.append(switchCase); |
| 215 | std::sort( |
| 216 | cases.begin(), cases.end(), |
| 217 | [] (const SwitchCase& left, const SwitchCase& right) { |
| 218 | return left.caseValue() < right.caseValue(); |
| 219 | }); |
| 220 | FrequentedBlock fallThrough = m_block->fallThrough(); |
| 221 | m_block->values().removeLast(); |
| 222 | recursivelyBuildSwitch(cases, fallThrough, 0, false, cases.size(), m_block); |
| 223 | m_proc.deleteValue(switchValue); |
| 224 | m_block->updatePredecessorsAfter(); |
| 225 | m_changed = true; |
| 226 | break; |
| 227 | } |
| 228 | |
| 229 | case Depend: { |
| 230 | if (isX86()) { |
| 231 | // Create a load-load fence. This codegens to nothing on X86. We use it to tell the |
| 232 | // compiler not to block load motion. |
| 233 | FenceValue* fence = m_insertionSet.insert<FenceValue>(m_index, m_origin); |
| 234 | fence->read = HeapRange(); |
| 235 | fence->write = HeapRange::top(); |
| 236 | |
| 237 | // Kill the Depend, which should unlock a bunch of code simplification. |
| 238 | m_value->replaceWithBottom(m_insertionSet, m_index); |
| 239 | |
| 240 | m_changed = true; |
| 241 | } |
| 242 | break; |
| 243 | } |
| 244 | |
| 245 | case AtomicWeakCAS: |
| 246 | case AtomicStrongCAS: { |
| 247 | AtomicValue* atomic = m_value->as<AtomicValue>(); |
| 248 | Width width = atomic->accessWidth(); |
| 249 | |
| 250 | if (isCanonicalWidth(width)) |
| 251 | break; |
| 252 | |
| 253 | Value* expectedValue = atomic->child(0); |
| 254 | |
| 255 | if (!isX86()) { |
| 256 | // On ARM, the load part of the CAS does a load with zero extension. Therefore, we need |
| 257 | // to zero-extend the input. |
| 258 | Value* maskedExpectedValue = m_insertionSet.insert<Value>( |
| 259 | m_index, BitAnd, m_origin, expectedValue, |
| 260 | m_insertionSet.insertIntConstant(m_index, expectedValue, mask(width))); |
| 261 | |
| 262 | atomic->child(0) = maskedExpectedValue; |
| 263 | m_changed = true; |
| 264 | } |
| 265 | |
| 266 | if (atomic->opcode() == AtomicStrongCAS) { |
| 267 | Value* newValue = m_insertionSet.insert<Value>( |
| 268 | m_index, signExtendOpcode(width), m_origin, |
| 269 | m_insertionSet.insertClone(m_index, atomic)); |
| 270 | |
| 271 | atomic->replaceWithIdentity(newValue); |
| 272 | m_changed = true; |
| 273 | } |
| 274 | |
| 275 | break; |
| 276 | } |
| 277 | |
| 278 | case AtomicXchgAdd: |
| 279 | case AtomicXchgAnd: |
| 280 | case AtomicXchgOr: |
| 281 | case AtomicXchgSub: |
| 282 | case AtomicXchgXor: |
| 283 | case AtomicXchg: { |
| 284 | // On X86, these may actually return garbage in the high bits. On ARM64, these sorta |
| 285 | // zero-extend their high bits, except that the high bits might get polluted by high |
| 286 | // bits in the operand. So, either way, we need to throw a sign-extend on these |
| 287 | // things. |
| 288 | |
| 289 | if (isX86()) { |
| 290 | if (m_value->opcode() == AtomicXchgSub && m_useCounts.numUses(m_value)) { |
| 291 | // On x86, xchgadd is better than xchgsub if it has any users. |
| 292 | m_value->setOpcodeUnsafely(AtomicXchgAdd); |
| 293 | m_value->child(0) = m_insertionSet.insert<Value>( |
| 294 | m_index, Neg, m_origin, m_value->child(0)); |
| 295 | } |
| 296 | |
| 297 | bool exempt = false; |
| 298 | switch (m_value->opcode()) { |
| 299 | case AtomicXchgAnd: |
| 300 | case AtomicXchgOr: |
| 301 | case AtomicXchgSub: |
| 302 | case AtomicXchgXor: |
| 303 | exempt = true; |
| 304 | break; |
| 305 | default: |
| 306 | break; |
| 307 | } |
| 308 | if (exempt) |
| 309 | break; |
| 310 | } |
| 311 | |
| 312 | AtomicValue* atomic = m_value->as<AtomicValue>(); |
| 313 | Width width = atomic->accessWidth(); |
| 314 | |
| 315 | if (isCanonicalWidth(width)) |
| 316 | break; |
| 317 | |
| 318 | Value* newValue = m_insertionSet.insert<Value>( |
| 319 | m_index, signExtendOpcode(width), m_origin, |
| 320 | m_insertionSet.insertClone(m_index, atomic)); |
| 321 | |
| 322 | atomic->replaceWithIdentity(newValue); |
| 323 | m_changed = true; |
| 324 | break; |
| 325 | } |
| 326 | |
| 327 | case Load8Z: |
| 328 | case Load16Z: { |
| 329 | if (isX86()) |
| 330 | break; |
| 331 | |
| 332 | MemoryValue* memory = m_value->as<MemoryValue>(); |
| 333 | if (!memory->hasFence()) |
| 334 | break; |
| 335 | |
| 336 | // Sub-width load-acq on ARM64 always sign extends. |
| 337 | Value* newLoad = m_insertionSet.insertClone(m_index, memory); |
| 338 | newLoad->setOpcodeUnsafely(memory->opcode() == Load8Z ? Load8S : Load16S); |
| 339 | |
| 340 | Value* newValue = m_insertionSet.insert<Value>( |
| 341 | m_index, BitAnd, m_origin, newLoad, |
| 342 | m_insertionSet.insertIntConstant( |
| 343 | m_index, m_origin, Int32, mask(memory->accessWidth()))); |
| 344 | |
| 345 | m_value->replaceWithIdentity(newValue); |
| 346 | m_changed = true; |
| 347 | break; |
| 348 | } |
| 349 | |
| 350 | default: |
| 351 | break; |
| 352 | } |
| 353 | } |
| 354 | m_insertionSet.execute(m_block); |
| 355 | } |
| 356 | |
| 357 | void makeDivisionChill(Opcode nonChillOpcode) |
| 358 | { |
| 359 | ASSERT(nonChillOpcode == Div || nonChillOpcode == Mod); |
| 360 | |
| 361 | // ARM supports this instruction natively. |
| 362 | if (isARM64()) |
| 363 | return; |
| 364 | |
| 365 | // We implement "res = Div<Chill>/Mod<Chill>(num, den)" as follows: |
| 366 | // |
| 367 | // if (den + 1 <=_unsigned 1) { |
| 368 | // if (!den) { |
| 369 | // res = 0; |
| 370 | // goto done; |
| 371 | // } |
| 372 | // if (num == -2147483648) { |
| 373 | // res = isDiv ? num : 0; |
| 374 | // goto done; |
| 375 | // } |
| 376 | // } |
| 377 | // res = num (/ or %) dev; |
| 378 | // done: |
| 379 | m_changed = true; |
| 380 | |
| 381 | Value* num = m_value->child(0); |
| 382 | Value* den = m_value->child(1); |
| 383 | |
| 384 | Value* one = m_insertionSet.insertIntConstant(m_index, m_value, 1); |
| 385 | Value* isDenOK = m_insertionSet.insert<Value>( |
| 386 | m_index, Above, m_origin, |
| 387 | m_insertionSet.insert<Value>(m_index, Add, m_origin, den, one), |
| 388 | one); |
| 389 | |
| 390 | BasicBlock* before = m_blockInsertionSet.splitForward(m_block, m_index, &m_insertionSet); |
| 391 | |
| 392 | BasicBlock* normalDivCase = m_blockInsertionSet.insertBefore(m_block); |
| 393 | BasicBlock* shadyDenCase = m_blockInsertionSet.insertBefore(m_block); |
| 394 | BasicBlock* zeroDenCase = m_blockInsertionSet.insertBefore(m_block); |
| 395 | BasicBlock* neg1DenCase = m_blockInsertionSet.insertBefore(m_block); |
| 396 | BasicBlock* intMinCase = m_blockInsertionSet.insertBefore(m_block); |
| 397 | |
| 398 | before->replaceLastWithNew<Value>(m_proc, Branch, m_origin, isDenOK); |
| 399 | before->setSuccessors( |
| 400 | FrequentedBlock(normalDivCase, FrequencyClass::Normal), |
| 401 | FrequentedBlock(shadyDenCase, FrequencyClass::Rare)); |
| 402 | |
| 403 | UpsilonValue* normalResult = normalDivCase->appendNew<UpsilonValue>( |
| 404 | m_proc, m_origin, |
| 405 | normalDivCase->appendNew<Value>(m_proc, nonChillOpcode, m_origin, num, den)); |
| 406 | normalDivCase->appendNew<Value>(m_proc, Jump, m_origin); |
| 407 | normalDivCase->setSuccessors(FrequentedBlock(m_block)); |
| 408 | |
| 409 | shadyDenCase->appendNew<Value>(m_proc, Branch, m_origin, den); |
| 410 | shadyDenCase->setSuccessors( |
| 411 | FrequentedBlock(neg1DenCase, FrequencyClass::Normal), |
| 412 | FrequentedBlock(zeroDenCase, FrequencyClass::Rare)); |
| 413 | |
| 414 | UpsilonValue* zeroResult = zeroDenCase->appendNew<UpsilonValue>( |
| 415 | m_proc, m_origin, |
| 416 | zeroDenCase->appendIntConstant(m_proc, m_value, 0)); |
| 417 | zeroDenCase->appendNew<Value>(m_proc, Jump, m_origin); |
| 418 | zeroDenCase->setSuccessors(FrequentedBlock(m_block)); |
| 419 | |
| 420 | int64_t badNumeratorConst = 0; |
| 421 | switch (m_value->type().kind()) { |
| 422 | case Int32: |
| 423 | badNumeratorConst = std::numeric_limits<int32_t>::min(); |
| 424 | break; |
| 425 | case Int64: |
| 426 | badNumeratorConst = std::numeric_limits<int64_t>::min(); |
| 427 | break; |
| 428 | default: |
| 429 | ASSERT_NOT_REACHED(); |
| 430 | badNumeratorConst = 0; |
| 431 | } |
| 432 | |
| 433 | Value* badNumerator = |
| 434 | neg1DenCase->appendIntConstant(m_proc, m_value, badNumeratorConst); |
| 435 | |
| 436 | neg1DenCase->appendNew<Value>( |
| 437 | m_proc, Branch, m_origin, |
| 438 | neg1DenCase->appendNew<Value>( |
| 439 | m_proc, Equal, m_origin, num, badNumerator)); |
| 440 | neg1DenCase->setSuccessors( |
| 441 | FrequentedBlock(intMinCase, FrequencyClass::Rare), |
| 442 | FrequentedBlock(normalDivCase, FrequencyClass::Normal)); |
| 443 | |
| 444 | Value* intMinResult = nonChillOpcode == Div ? badNumerator : intMinCase->appendIntConstant(m_proc, m_value, 0); |
| 445 | UpsilonValue* intMinResultUpsilon = intMinCase->appendNew<UpsilonValue>( |
| 446 | m_proc, m_origin, intMinResult); |
| 447 | intMinCase->appendNew<Value>(m_proc, Jump, m_origin); |
| 448 | intMinCase->setSuccessors(FrequentedBlock(m_block)); |
| 449 | |
| 450 | Value* phi = m_insertionSet.insert<Value>( |
| 451 | m_index, Phi, m_value->type(), m_origin); |
| 452 | normalResult->setPhi(phi); |
| 453 | zeroResult->setPhi(phi); |
| 454 | intMinResultUpsilon->setPhi(phi); |
| 455 | |
| 456 | m_value->replaceWithIdentity(phi); |
| 457 | before->updatePredecessorsAfter(); |
| 458 | } |
| 459 | |
| 460 | void recursivelyBuildSwitch( |
| 461 | const Vector<SwitchCase>& cases, FrequentedBlock fallThrough, unsigned start, bool hardStart, |
| 462 | unsigned end, BasicBlock* before) |
| 463 | { |
| 464 | Value* child = m_value->child(0); |
| 465 | Type type = child->type(); |
| 466 | |
| 467 | // It's a good idea to use a table-based switch in some cases: the number of cases has to be |
| 468 | // large enough and they have to be dense enough. This could probably be improved a lot. For |
| 469 | // example, we could still use a jump table in cases where the inputs are sparse so long as we |
| 470 | // shift off the uninteresting bits. On the other hand, it's not clear that this would |
| 471 | // actually be any better than what we have done here and it's not clear that it would be |
| 472 | // better than a binary switch. |
| 473 | const unsigned minCasesForTable = 7; |
| 474 | const unsigned densityLimit = 4; |
| 475 | if (end - start >= minCasesForTable) { |
| 476 | int64_t firstValue = cases[start].caseValue(); |
| 477 | int64_t lastValue = cases[end - 1].caseValue(); |
| 478 | if ((lastValue - firstValue + 1) / (end - start) < densityLimit) { |
| 479 | BasicBlock* switchBlock = m_blockInsertionSet.insertAfter(m_block); |
| 480 | Value* index = before->appendNew<Value>( |
| 481 | m_proc, Sub, m_origin, child, |
| 482 | before->appendIntConstant(m_proc, m_origin, type, firstValue)); |
| 483 | before->appendNew<Value>( |
| 484 | m_proc, Branch, m_origin, |
| 485 | before->appendNew<Value>( |
| 486 | m_proc, Above, m_origin, index, |
| 487 | before->appendIntConstant(m_proc, m_origin, type, lastValue - firstValue))); |
| 488 | before->setSuccessors(fallThrough, FrequentedBlock(switchBlock)); |
| 489 | |
| 490 | size_t tableSize = lastValue - firstValue + 1; |
| 491 | |
| 492 | if (index->type() != pointerType() && index->type() == Int32) |
| 493 | index = switchBlock->appendNew<Value>(m_proc, ZExt32, m_origin, index); |
| 494 | |
| 495 | PatchpointValue* patchpoint = |
| 496 | switchBlock->appendNew<PatchpointValue>(m_proc, Void, m_origin); |
| 497 | |
| 498 | // Even though this loads from the jump table, the jump table is immutable. For the |
| 499 | // purpose of alias analysis, reading something immutable is like reading nothing. |
| 500 | patchpoint->effects = Effects(); |
| 501 | patchpoint->effects.terminal = true; |
| 502 | |
| 503 | patchpoint->appendSomeRegister(index); |
| 504 | patchpoint->numGPScratchRegisters = 2; |
| 505 | // Technically, we don't have to clobber macro registers on X86_64. This is probably |
| 506 | // OK though. |
| 507 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
| 508 | |
| 509 | BitVector handledIndices; |
| 510 | for (unsigned i = start; i < end; ++i) { |
| 511 | FrequentedBlock block = cases[i].target(); |
| 512 | int64_t value = cases[i].caseValue(); |
| 513 | switchBlock->appendSuccessor(block); |
| 514 | size_t index = value - firstValue; |
| 515 | ASSERT(!handledIndices.get(index)); |
| 516 | handledIndices.set(index); |
| 517 | } |
| 518 | |
| 519 | bool hasUnhandledIndex = false; |
| 520 | for (unsigned i = 0; i < tableSize; ++i) { |
| 521 | if (!handledIndices.get(i)) { |
| 522 | hasUnhandledIndex = true; |
| 523 | break; |
| 524 | } |
| 525 | } |
| 526 | |
| 527 | if (hasUnhandledIndex) |
| 528 | switchBlock->appendSuccessor(fallThrough); |
| 529 | |
| 530 | patchpoint->setGenerator( |
| 531 | [=] (CCallHelpers& jit, const StackmapGenerationParams& params) { |
| 532 | AllowMacroScratchRegisterUsage allowScratch(jit); |
| 533 | |
| 534 | using JumpTableCodePtr = MacroAssemblerCodePtr<JSSwitchPtrTag>; |
| 535 | JumpTableCodePtr* jumpTable = static_cast<JumpTableCodePtr*>( |
| 536 | params.proc().addDataSection(sizeof(JumpTableCodePtr) * tableSize)); |
| 537 | |
| 538 | GPRReg index = params[0].gpr(); |
| 539 | GPRReg scratch = params.gpScratch(0); |
| 540 | |
| 541 | jit.move(CCallHelpers::TrustedImmPtr(jumpTable), scratch); |
| 542 | jit.load64(CCallHelpers::BaseIndex(scratch, index, CCallHelpers::timesPtr()), scratch); |
| 543 | jit.farJump(scratch, JSSwitchPtrTag); |
| 544 | |
| 545 | // These labels are guaranteed to be populated before either late paths or |
| 546 | // link tasks run. |
| 547 | Vector<Box<CCallHelpers::Label>> labels = params.successorLabels(); |
| 548 | |
| 549 | jit.addLinkTask( |
| 550 | [=] (LinkBuffer& linkBuffer) { |
| 551 | if (hasUnhandledIndex) { |
| 552 | JumpTableCodePtr fallThrough = linkBuffer.locationOf<JSSwitchPtrTag>(*labels.last()); |
| 553 | for (unsigned i = tableSize; i--;) |
| 554 | jumpTable[i] = fallThrough; |
| 555 | } |
| 556 | |
| 557 | unsigned labelIndex = 0; |
| 558 | for (unsigned tableIndex : handledIndices) |
| 559 | jumpTable[tableIndex] = linkBuffer.locationOf<JSSwitchPtrTag>(*labels[labelIndex++]); |
| 560 | }); |
| 561 | }); |
| 562 | return; |
| 563 | } |
| 564 | } |
| 565 | |
| 566 | // See comments in jit/BinarySwitch.cpp for a justification of this algorithm. The only |
| 567 | // thing we do differently is that we don't use randomness. |
| 568 | |
| 569 | const unsigned leafThreshold = 3; |
| 570 | |
| 571 | unsigned size = end - start; |
| 572 | |
| 573 | if (size <= leafThreshold) { |
| 574 | bool allConsecutive = false; |
| 575 | |
| 576 | if ((hardStart || (start && cases[start - 1].caseValue() == cases[start].caseValue() - 1)) |
| 577 | && end < cases.size() |
| 578 | && cases[end - 1].caseValue() == cases[end].caseValue() - 1) { |
| 579 | allConsecutive = true; |
| 580 | for (unsigned i = 0; i < size - 1; ++i) { |
| 581 | if (cases[start + i].caseValue() + 1 != cases[start + i + 1].caseValue()) { |
| 582 | allConsecutive = false; |
| 583 | break; |
| 584 | } |
| 585 | } |
| 586 | } |
| 587 | |
| 588 | unsigned limit = allConsecutive ? size - 1 : size; |
| 589 | |
| 590 | for (unsigned i = 0; i < limit; ++i) { |
| 591 | BasicBlock* nextCheck = m_blockInsertionSet.insertAfter(m_block); |
| 592 | before->appendNew<Value>( |
| 593 | m_proc, Branch, m_origin, |
| 594 | before->appendNew<Value>( |
| 595 | m_proc, Equal, m_origin, child, |
| 596 | before->appendIntConstant( |
| 597 | m_proc, m_origin, type, |
| 598 | cases[start + i].caseValue()))); |
| 599 | before->setSuccessors(cases[start + i].target(), FrequentedBlock(nextCheck)); |
| 600 | |
| 601 | before = nextCheck; |
| 602 | } |
| 603 | |
| 604 | before->appendNew<Value>(m_proc, Jump, m_origin); |
| 605 | if (allConsecutive) |
| 606 | before->setSuccessors(cases[end - 1].target()); |
| 607 | else |
| 608 | before->setSuccessors(fallThrough); |
| 609 | return; |
| 610 | } |
| 611 | |
| 612 | unsigned medianIndex = (start + end) / 2; |
| 613 | |
| 614 | BasicBlock* left = m_blockInsertionSet.insertAfter(m_block); |
| 615 | BasicBlock* right = m_blockInsertionSet.insertAfter(m_block); |
| 616 | |
| 617 | before->appendNew<Value>( |
| 618 | m_proc, Branch, m_origin, |
| 619 | before->appendNew<Value>( |
| 620 | m_proc, LessThan, m_origin, child, |
| 621 | before->appendIntConstant( |
| 622 | m_proc, m_origin, type, |
| 623 | cases[medianIndex].caseValue()))); |
| 624 | before->setSuccessors(FrequentedBlock(left), FrequentedBlock(right)); |
| 625 | |
| 626 | recursivelyBuildSwitch(cases, fallThrough, start, hardStart, medianIndex, left); |
| 627 | recursivelyBuildSwitch(cases, fallThrough, medianIndex, true, end, right); |
| 628 | } |
| 629 | |
| 630 | Procedure& m_proc; |
| 631 | BlockInsertionSet m_blockInsertionSet; |
| 632 | InsertionSet m_insertionSet; |
| 633 | UseCounts m_useCounts; |
| 634 | BasicBlock* m_block; |
| 635 | unsigned m_index; |
| 636 | Value* m_value; |
| 637 | Origin m_origin; |
| 638 | bool m_changed { false }; |
| 639 | }; |
| 640 | |
| 641 | } // anonymous namespace |
| 642 | |
| 643 | bool lowerMacros(Procedure& proc) |
| 644 | { |
| 645 | PhaseScope phaseScope(proc, "B3::lowerMacros" ); |
| 646 | LowerMacros lowerMacros(proc); |
| 647 | return lowerMacros.run(); |
| 648 | } |
| 649 | |
| 650 | } } // namespace JSC::B3 |
| 651 | |
| 652 | #endif // ENABLE(B3_JIT) |
| 653 | |
| 654 | |