1 | /* |
2 | * Copyright (C) 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 "WasmAirIRGenerator.h" |
28 | |
29 | #if ENABLE(WEBASSEMBLY) |
30 | |
31 | #include "AirCode.h" |
32 | #include "AirGenerate.h" |
33 | #include "AirOpcodeUtils.h" |
34 | #include "AirValidate.h" |
35 | #include "AllowMacroScratchRegisterUsageIf.h" |
36 | #include "B3CCallValue.h" |
37 | #include "B3CheckSpecial.h" |
38 | #include "B3CheckValue.h" |
39 | #include "B3PatchpointSpecial.h" |
40 | #include "B3Procedure.h" |
41 | #include "B3ProcedureInlines.h" |
42 | #include "BinarySwitch.h" |
43 | #include "DisallowMacroScratchRegisterUsage.h" |
44 | #include "JSCInlines.h" |
45 | #include "JSWebAssemblyInstance.h" |
46 | #include "ScratchRegisterAllocator.h" |
47 | #include "VirtualRegister.h" |
48 | #include "WasmCallingConvention.h" |
49 | #include "WasmContextInlines.h" |
50 | #include "WasmExceptionType.h" |
51 | #include "WasmFunctionParser.h" |
52 | #include "WasmInstance.h" |
53 | #include "WasmMemory.h" |
54 | #include "WasmOMGPlan.h" |
55 | #include "WasmOpcodeOrigin.h" |
56 | #include "WasmSignatureInlines.h" |
57 | #include "WasmThunks.h" |
58 | #include <limits> |
59 | #include <wtf/Box.h> |
60 | #include <wtf/Optional.h> |
61 | #include <wtf/StdLibExtras.h> |
62 | |
63 | namespace JSC { namespace Wasm { |
64 | |
65 | using namespace B3::Air; |
66 | |
67 | struct ConstrainedTmp { |
68 | ConstrainedTmp(Tmp tmp) |
69 | : ConstrainedTmp(tmp, tmp.isReg() ? B3::ValueRep::reg(tmp.reg()) : B3::ValueRep::SomeRegister) |
70 | { } |
71 | |
72 | ConstrainedTmp(Tmp tmp, B3::ValueRep rep) |
73 | : tmp(tmp) |
74 | , rep(rep) |
75 | { |
76 | } |
77 | |
78 | Tmp tmp; |
79 | B3::ValueRep rep; |
80 | }; |
81 | |
82 | class TypedTmp { |
83 | public: |
84 | constexpr TypedTmp() |
85 | : m_tmp() |
86 | , m_type(Type::Void) |
87 | { } |
88 | |
89 | TypedTmp(Tmp tmp, Type type) |
90 | : m_tmp(tmp) |
91 | , m_type(type) |
92 | { } |
93 | |
94 | TypedTmp(const TypedTmp&) = default; |
95 | TypedTmp(TypedTmp&&) = default; |
96 | TypedTmp& operator=(TypedTmp&&) = default; |
97 | TypedTmp& operator=(const TypedTmp&) = default; |
98 | |
99 | bool operator==(const TypedTmp& other) const |
100 | { |
101 | return m_tmp == other.m_tmp && m_type == other.m_type; |
102 | } |
103 | bool operator!=(const TypedTmp& other) const |
104 | { |
105 | return !(*this == other); |
106 | } |
107 | |
108 | explicit operator bool() const { return !!tmp(); } |
109 | |
110 | operator Tmp() const { return tmp(); } |
111 | operator Arg() const { return Arg(tmp()); } |
112 | Tmp tmp() const { return m_tmp; } |
113 | Type type() const { return m_type; } |
114 | |
115 | private: |
116 | |
117 | Tmp m_tmp; |
118 | Type m_type; |
119 | }; |
120 | |
121 | class AirIRGenerator { |
122 | public: |
123 | struct ControlData { |
124 | ControlData(B3::Origin origin, Type returnType, TypedTmp resultTmp, BlockType type, BasicBlock* continuation, BasicBlock* special = nullptr) |
125 | : blockType(type) |
126 | , continuation(continuation) |
127 | , special(special) |
128 | , returnType(returnType) |
129 | { |
130 | UNUSED_PARAM(origin); // FIXME: Use origin. |
131 | if (resultTmp) { |
132 | ASSERT(returnType != Type::Void); |
133 | result.append(resultTmp); |
134 | } else |
135 | ASSERT(returnType == Type::Void); |
136 | } |
137 | |
138 | ControlData() |
139 | { |
140 | } |
141 | |
142 | void dump(PrintStream& out) const |
143 | { |
144 | switch (type()) { |
145 | case BlockType::If: |
146 | out.print("If: " ); |
147 | break; |
148 | case BlockType::Block: |
149 | out.print("Block: " ); |
150 | break; |
151 | case BlockType::Loop: |
152 | out.print("Loop: " ); |
153 | break; |
154 | case BlockType::TopLevel: |
155 | out.print("TopLevel: " ); |
156 | break; |
157 | } |
158 | out.print("Continuation: " , *continuation, ", Special: " ); |
159 | if (special) |
160 | out.print(*special); |
161 | else |
162 | out.print("None" ); |
163 | } |
164 | |
165 | BlockType type() const { return blockType; } |
166 | |
167 | Type signature() const { return returnType; } |
168 | |
169 | bool hasNonVoidSignature() const { return result.size(); } |
170 | |
171 | BasicBlock* targetBlockForBranch() |
172 | { |
173 | if (type() == BlockType::Loop) |
174 | return special; |
175 | return continuation; |
176 | } |
177 | |
178 | void convertIfToBlock() |
179 | { |
180 | ASSERT(type() == BlockType::If); |
181 | blockType = BlockType::Block; |
182 | special = nullptr; |
183 | } |
184 | |
185 | using ResultList = Vector<TypedTmp, 1>; |
186 | |
187 | ResultList resultForBranch() const |
188 | { |
189 | if (type() == BlockType::Loop) |
190 | return ResultList(); |
191 | return result; |
192 | } |
193 | |
194 | private: |
195 | friend class AirIRGenerator; |
196 | BlockType blockType; |
197 | BasicBlock* continuation; |
198 | BasicBlock* special; |
199 | ResultList result; |
200 | Type returnType; |
201 | }; |
202 | |
203 | using ExpressionType = TypedTmp; |
204 | using ControlType = ControlData; |
205 | using ExpressionList = Vector<ExpressionType, 1>; |
206 | using ResultList = ControlData::ResultList; |
207 | using ControlEntry = FunctionParser<AirIRGenerator>::ControlEntry; |
208 | |
209 | static ExpressionType emptyExpression() { return { }; }; |
210 | |
211 | using ErrorType = String; |
212 | using UnexpectedResult = Unexpected<ErrorType>; |
213 | using Result = Expected<std::unique_ptr<InternalFunction>, ErrorType>; |
214 | using PartialResult = Expected<void, ErrorType>; |
215 | |
216 | template <typename ...Args> |
217 | NEVER_INLINE UnexpectedResult WARN_UNUSED_RETURN fail(Args... args) const |
218 | { |
219 | using namespace FailureHelper; // See ADL comment in WasmParser.h. |
220 | return UnexpectedResult(makeString("WebAssembly.Module failed compiling: "_s , makeString(args)...)); |
221 | } |
222 | |
223 | #define WASM_COMPILE_FAIL_IF(condition, ...) do { \ |
224 | if (UNLIKELY(condition)) \ |
225 | return fail(__VA_ARGS__); \ |
226 | } while (0) |
227 | |
228 | AirIRGenerator(const ModuleInformation&, B3::Procedure&, InternalFunction*, Vector<UnlinkedWasmToWasmCall>&, MemoryMode, unsigned functionIndex, TierUpCount*, ThrowWasmException, const Signature&); |
229 | |
230 | PartialResult WARN_UNUSED_RETURN addArguments(const Signature&); |
231 | PartialResult WARN_UNUSED_RETURN addLocal(Type, uint32_t); |
232 | ExpressionType addConstant(Type, uint64_t); |
233 | ExpressionType addConstant(BasicBlock*, Type, uint64_t); |
234 | |
235 | // References |
236 | PartialResult WARN_UNUSED_RETURN addRefIsNull(ExpressionType& value, ExpressionType& result); |
237 | PartialResult WARN_UNUSED_RETURN addRefFunc(uint32_t index, ExpressionType& result); |
238 | |
239 | // Tables |
240 | PartialResult WARN_UNUSED_RETURN addTableGet(unsigned, ExpressionType& index, ExpressionType& result); |
241 | PartialResult WARN_UNUSED_RETURN addTableSet(unsigned, ExpressionType& index, ExpressionType& value); |
242 | PartialResult WARN_UNUSED_RETURN addTableSize(unsigned, ExpressionType& result); |
243 | PartialResult WARN_UNUSED_RETURN addTableGrow(unsigned, ExpressionType& fill, ExpressionType& delta, ExpressionType& result); |
244 | PartialResult WARN_UNUSED_RETURN addTableFill(unsigned, ExpressionType& offset, ExpressionType& fill, ExpressionType& count); |
245 | |
246 | // Locals |
247 | PartialResult WARN_UNUSED_RETURN getLocal(uint32_t index, ExpressionType& result); |
248 | PartialResult WARN_UNUSED_RETURN setLocal(uint32_t index, ExpressionType value); |
249 | |
250 | // Globals |
251 | PartialResult WARN_UNUSED_RETURN getGlobal(uint32_t index, ExpressionType& result); |
252 | PartialResult WARN_UNUSED_RETURN setGlobal(uint32_t index, ExpressionType value); |
253 | |
254 | // Memory |
255 | PartialResult WARN_UNUSED_RETURN load(LoadOpType, ExpressionType pointer, ExpressionType& result, uint32_t offset); |
256 | PartialResult WARN_UNUSED_RETURN store(StoreOpType, ExpressionType pointer, ExpressionType value, uint32_t offset); |
257 | PartialResult WARN_UNUSED_RETURN addGrowMemory(ExpressionType delta, ExpressionType& result); |
258 | PartialResult WARN_UNUSED_RETURN addCurrentMemory(ExpressionType& result); |
259 | |
260 | // Basic operators |
261 | template<OpType> |
262 | PartialResult WARN_UNUSED_RETURN addOp(ExpressionType arg, ExpressionType& result); |
263 | template<OpType> |
264 | PartialResult WARN_UNUSED_RETURN addOp(ExpressionType left, ExpressionType right, ExpressionType& result); |
265 | PartialResult WARN_UNUSED_RETURN addSelect(ExpressionType condition, ExpressionType nonZero, ExpressionType zero, ExpressionType& result); |
266 | |
267 | // Control flow |
268 | ControlData WARN_UNUSED_RETURN addTopLevel(Type signature); |
269 | ControlData WARN_UNUSED_RETURN addBlock(Type signature); |
270 | ControlData WARN_UNUSED_RETURN addLoop(Type signature); |
271 | PartialResult WARN_UNUSED_RETURN addIf(ExpressionType condition, Type signature, ControlData& result); |
272 | PartialResult WARN_UNUSED_RETURN addElse(ControlData&, const ExpressionList&); |
273 | PartialResult WARN_UNUSED_RETURN addElseToUnreachable(ControlData&); |
274 | |
275 | PartialResult WARN_UNUSED_RETURN addReturn(const ControlData&, const ExpressionList& returnValues); |
276 | PartialResult WARN_UNUSED_RETURN addBranch(ControlData&, ExpressionType condition, const ExpressionList& returnValues); |
277 | PartialResult WARN_UNUSED_RETURN addSwitch(ExpressionType condition, const Vector<ControlData*>& targets, ControlData& defaultTargets, const ExpressionList& expressionStack); |
278 | PartialResult WARN_UNUSED_RETURN endBlock(ControlEntry&, ExpressionList& expressionStack); |
279 | PartialResult WARN_UNUSED_RETURN addEndToUnreachable(ControlEntry&); |
280 | |
281 | // Calls |
282 | PartialResult WARN_UNUSED_RETURN addCall(uint32_t calleeIndex, const Signature&, Vector<ExpressionType>& args, ExpressionType& result); |
283 | PartialResult WARN_UNUSED_RETURN addCallIndirect(unsigned tableIndex, const Signature&, Vector<ExpressionType>& args, ExpressionType& result); |
284 | PartialResult WARN_UNUSED_RETURN addUnreachable(); |
285 | |
286 | PartialResult addShift(Type, B3::Air::Opcode, ExpressionType value, ExpressionType shift, ExpressionType& result); |
287 | PartialResult addIntegerSub(B3::Air::Opcode, ExpressionType lhs, ExpressionType rhs, ExpressionType& result); |
288 | PartialResult addFloatingPointAbs(B3::Air::Opcode, ExpressionType value, ExpressionType& result); |
289 | PartialResult addFloatingPointBinOp(Type, B3::Air::Opcode, ExpressionType lhs, ExpressionType rhs, ExpressionType& result); |
290 | |
291 | void dump(const Vector<ControlEntry>& controlStack, const ExpressionList* expressionStack); |
292 | void setParser(FunctionParser<AirIRGenerator>* parser) { m_parser = parser; }; |
293 | |
294 | static Vector<Tmp> toTmpVector(const Vector<TypedTmp>& vector) |
295 | { |
296 | Vector<Tmp> result; |
297 | for (const auto& item : vector) |
298 | result.append(item.tmp()); |
299 | return result; |
300 | } |
301 | |
302 | ALWAYS_INLINE void didKill(const ExpressionType& typedTmp) |
303 | { |
304 | Tmp tmp = typedTmp.tmp(); |
305 | if (!tmp) |
306 | return; |
307 | if (tmp.isGP()) |
308 | m_freeGPs.append(tmp); |
309 | else |
310 | m_freeFPs.append(tmp); |
311 | } |
312 | |
313 | private: |
314 | ALWAYS_INLINE void validateInst(Inst& inst) |
315 | { |
316 | if (!ASSERT_DISABLED) { |
317 | if (!inst.isValidForm()) { |
318 | dataLogLn(inst); |
319 | CRASH(); |
320 | } |
321 | } |
322 | } |
323 | |
324 | static Arg (const TypedTmp& tmp) { return tmp.tmp(); } |
325 | static Arg (const Tmp& tmp) { return Arg(tmp); } |
326 | static Arg (const Arg& arg) { return arg; } |
327 | |
328 | template<typename... Arguments> |
329 | void append(BasicBlock* block, Kind kind, Arguments&&... arguments) |
330 | { |
331 | // FIXME: Find a way to use origin here. |
332 | auto& inst = block->append(kind, nullptr, extractArg(arguments)...); |
333 | validateInst(inst); |
334 | } |
335 | |
336 | template<typename... Arguments> |
337 | void append(Kind kind, Arguments&&... arguments) |
338 | { |
339 | append(m_currentBlock, kind, std::forward<Arguments>(arguments)...); |
340 | } |
341 | |
342 | template<typename... Arguments> |
343 | void appendEffectful(B3::Air::Opcode op, Arguments&&... arguments) |
344 | { |
345 | Kind kind = op; |
346 | kind.effects = true; |
347 | append(m_currentBlock, kind, std::forward<Arguments>(arguments)...); |
348 | } |
349 | |
350 | Tmp newTmp(B3::Bank bank) |
351 | { |
352 | switch (bank) { |
353 | case B3::GP: |
354 | if (m_freeGPs.size()) |
355 | return m_freeGPs.takeLast(); |
356 | break; |
357 | case B3::FP: |
358 | if (m_freeFPs.size()) |
359 | return m_freeFPs.takeLast(); |
360 | break; |
361 | } |
362 | return m_code.newTmp(bank); |
363 | } |
364 | |
365 | TypedTmp g32() { return { newTmp(B3::GP), Type::I32 }; } |
366 | TypedTmp g64() { return { newTmp(B3::GP), Type::I64 }; } |
367 | TypedTmp gAnyref() { return { newTmp(B3::GP), Type::Anyref }; } |
368 | TypedTmp gFuncref() { return { newTmp(B3::GP), Type::Funcref }; } |
369 | TypedTmp f32() { return { newTmp(B3::FP), Type::F32 }; } |
370 | TypedTmp f64() { return { newTmp(B3::FP), Type::F64 }; } |
371 | |
372 | TypedTmp tmpForType(Type type) |
373 | { |
374 | switch (type) { |
375 | case Type::I32: |
376 | return g32(); |
377 | case Type::I64: |
378 | return g64(); |
379 | case Type::Funcref: |
380 | return gFuncref(); |
381 | case Type::Anyref: |
382 | return gAnyref(); |
383 | case Type::F32: |
384 | return f32(); |
385 | case Type::F64: |
386 | return f64(); |
387 | case Type::Void: |
388 | return { }; |
389 | default: |
390 | RELEASE_ASSERT_NOT_REACHED(); |
391 | } |
392 | } |
393 | |
394 | B3::PatchpointValue* addPatchpoint(B3::Type type) |
395 | { |
396 | return m_proc.add<B3::PatchpointValue>(type, B3::Origin()); |
397 | } |
398 | |
399 | template <typename ...Args> |
400 | void emitPatchpoint(B3::PatchpointValue* patch, Tmp result, Args... theArgs) |
401 | { |
402 | emitPatchpoint(m_currentBlock, patch, result, std::forward<Args>(theArgs)...); |
403 | } |
404 | |
405 | template <typename ...Args> |
406 | void emitPatchpoint(BasicBlock* basicBlock, B3::PatchpointValue* patch, Tmp result, Args... theArgs) |
407 | { |
408 | emitPatchpoint(basicBlock, patch, result, Vector<ConstrainedTmp, sizeof...(Args)>::from(theArgs...)); |
409 | } |
410 | |
411 | void emitPatchpoint(BasicBlock* basicBlock, B3::PatchpointValue* patch, Tmp result) |
412 | { |
413 | emitPatchpoint(basicBlock, patch, result, Vector<ConstrainedTmp>()); |
414 | } |
415 | |
416 | template <size_t inlineSize> |
417 | void emitPatchpoint(BasicBlock* basicBlock, B3::PatchpointValue* patch, Tmp result, Vector<ConstrainedTmp, inlineSize>&& args) |
418 | { |
419 | if (!m_patchpointSpecial) |
420 | m_patchpointSpecial = static_cast<B3::PatchpointSpecial*>(m_code.addSpecial(std::make_unique<B3::PatchpointSpecial>())); |
421 | |
422 | Inst inst(Patch, patch, Arg::special(m_patchpointSpecial)); |
423 | Inst resultMov; |
424 | if (result) { |
425 | ASSERT(patch->type() != B3::Void); |
426 | switch (patch->resultConstraint.kind()) { |
427 | case B3::ValueRep::Register: |
428 | inst.args.append(Tmp(patch->resultConstraint.reg())); |
429 | resultMov = Inst(result.isGP() ? Move : MoveDouble, nullptr, Tmp(patch->resultConstraint.reg()), result); |
430 | break; |
431 | case B3::ValueRep::SomeRegister: |
432 | inst.args.append(result); |
433 | break; |
434 | default: |
435 | RELEASE_ASSERT_NOT_REACHED(); |
436 | } |
437 | } else |
438 | ASSERT(patch->type() == B3::Void); |
439 | |
440 | for (ConstrainedTmp& tmp : args) { |
441 | // FIXME: This is less than ideal to create dummy values just to satisfy Air's |
442 | // validation. We should abstrcat Patch enough so ValueRep's don't need to be |
443 | // backed by Values. |
444 | // https://bugs.webkit.org/show_bug.cgi?id=194040 |
445 | B3::Value* dummyValue = m_proc.addConstant(B3::Origin(), tmp.tmp.isGP() ? B3::Int64 : B3::Double, 0); |
446 | patch->append(dummyValue, tmp.rep); |
447 | switch (tmp.rep.kind()) { |
448 | case B3::ValueRep::SomeRegister: |
449 | inst.args.append(tmp.tmp); |
450 | break; |
451 | case B3::ValueRep::Register: |
452 | patch->earlyClobbered().clear(tmp.rep.reg()); |
453 | append(basicBlock, tmp.tmp.isGP() ? Move : MoveDouble, tmp.tmp, tmp.rep.reg()); |
454 | inst.args.append(Tmp(tmp.rep.reg())); |
455 | break; |
456 | case B3::ValueRep::StackArgument: { |
457 | auto arg = Arg::callArg(tmp.rep.offsetFromSP()); |
458 | append(basicBlock, tmp.tmp.isGP() ? Move : MoveDouble, tmp.tmp, arg); |
459 | inst.args.append(arg); |
460 | break; |
461 | } |
462 | default: |
463 | RELEASE_ASSERT_NOT_REACHED(); |
464 | } |
465 | } |
466 | |
467 | if (patch->resultConstraint.isReg()) |
468 | patch->lateClobbered().clear(patch->resultConstraint.reg()); |
469 | for (unsigned i = patch->numGPScratchRegisters; i--;) |
470 | inst.args.append(g64().tmp()); |
471 | for (unsigned i = patch->numFPScratchRegisters; i--;) |
472 | inst.args.append(f64().tmp()); |
473 | |
474 | validateInst(inst); |
475 | basicBlock->append(WTFMove(inst)); |
476 | if (resultMov) { |
477 | validateInst(resultMov); |
478 | basicBlock->append(WTFMove(resultMov)); |
479 | } |
480 | } |
481 | |
482 | template <typename Branch, typename Generator> |
483 | void emitCheck(const Branch& makeBranch, const Generator& generator) |
484 | { |
485 | // We fail along the truthy edge of 'branch'. |
486 | Inst branch = makeBranch(); |
487 | |
488 | // FIXME: Make a hashmap of these. |
489 | B3::CheckSpecial::Key key(branch); |
490 | B3::CheckSpecial* special = static_cast<B3::CheckSpecial*>(m_code.addSpecial(std::make_unique<B3::CheckSpecial>(key))); |
491 | |
492 | // FIXME: Remove the need for dummy values |
493 | // https://bugs.webkit.org/show_bug.cgi?id=194040 |
494 | B3::Value* dummyPredicate = m_proc.addConstant(B3::Origin(), B3::Int32, 42); |
495 | B3::CheckValue* checkValue = m_proc.add<B3::CheckValue>(B3::Check, B3::Origin(), dummyPredicate); |
496 | checkValue->setGenerator(generator); |
497 | |
498 | Inst inst(Patch, checkValue, Arg::special(special)); |
499 | inst.args.appendVector(branch.args); |
500 | m_currentBlock->append(WTFMove(inst)); |
501 | } |
502 | |
503 | template <typename Func, typename ...Args> |
504 | void emitCCall(Func func, TypedTmp result, Args... args) |
505 | { |
506 | emitCCall(m_currentBlock, func, result, std::forward<Args>(args)...); |
507 | } |
508 | template <typename Func, typename ...Args> |
509 | void emitCCall(BasicBlock* block, Func func, TypedTmp result, Args... theArgs) |
510 | { |
511 | B3::Type resultType = B3::Void; |
512 | if (result) { |
513 | switch (result.type()) { |
514 | case Type::I32: |
515 | resultType = B3::Int32; |
516 | break; |
517 | case Type::I64: |
518 | case Type::Anyref: |
519 | case Type::Funcref: |
520 | resultType = B3::Int64; |
521 | break; |
522 | case Type::F32: |
523 | resultType = B3::Float; |
524 | break; |
525 | case Type::F64: |
526 | resultType = B3::Double; |
527 | break; |
528 | default: |
529 | RELEASE_ASSERT_NOT_REACHED(); |
530 | } |
531 | } |
532 | |
533 | auto makeDummyValue = [&] (Tmp tmp) { |
534 | // FIXME: This is less than ideal to create dummy values just to satisfy Air's |
535 | // validation. We should abstrcat CCall enough so we're not reliant on arguments |
536 | // to the B3::CCallValue. |
537 | // https://bugs.webkit.org/show_bug.cgi?id=194040 |
538 | if (tmp.isGP()) |
539 | return m_proc.addConstant(B3::Origin(), B3::Int64, 0); |
540 | return m_proc.addConstant(B3::Origin(), B3::Double, 0); |
541 | }; |
542 | |
543 | B3::Value* dummyFunc = m_proc.addConstant(B3::Origin(), B3::Int64, bitwise_cast<uintptr_t>(func)); |
544 | B3::Value* origin = m_proc.add<B3::CCallValue>(resultType, B3::Origin(), B3::Effects::none(), dummyFunc, makeDummyValue(theArgs)...); |
545 | |
546 | Inst inst(CCall, origin); |
547 | |
548 | Tmp callee = g64(); |
549 | append(block, Move, Arg::immPtr(tagCFunctionPtr<void*>(func, B3CCallPtrTag)), callee); |
550 | inst.args.append(callee); |
551 | |
552 | if (result) |
553 | inst.args.append(result.tmp()); |
554 | |
555 | for (Tmp tmp : Vector<Tmp, sizeof...(Args)>::from(theArgs.tmp()...)) |
556 | inst.args.append(tmp); |
557 | |
558 | block->append(WTFMove(inst)); |
559 | } |
560 | |
561 | static B3::Air::Opcode moveOpForValueType(Type type) |
562 | { |
563 | switch (type) { |
564 | case Type::I32: |
565 | return Move32; |
566 | case Type::I64: |
567 | case Type::Anyref: |
568 | case Type::Funcref: |
569 | return Move; |
570 | case Type::F32: |
571 | return MoveFloat; |
572 | case Type::F64: |
573 | return MoveDouble; |
574 | default: |
575 | RELEASE_ASSERT_NOT_REACHED(); |
576 | } |
577 | } |
578 | |
579 | void emitThrowException(CCallHelpers&, ExceptionType); |
580 | |
581 | void emitTierUpCheck(uint32_t decrementCount, B3::Origin); |
582 | |
583 | void emitWriteBarrierForJSWrapper(); |
584 | ExpressionType emitCheckAndPreparePointer(ExpressionType pointer, uint32_t offset, uint32_t sizeOfOp); |
585 | ExpressionType emitLoadOp(LoadOpType, ExpressionType pointer, uint32_t offset); |
586 | void emitStoreOp(StoreOpType, ExpressionType pointer, ExpressionType value, uint32_t offset); |
587 | |
588 | void unify(const ExpressionType& dst, const ExpressionType& source); |
589 | void unifyValuesWithBlock(const ExpressionList& resultStack, const ResultList& stack); |
590 | |
591 | template <typename IntType> |
592 | void emitChecksForModOrDiv(bool isSignedDiv, ExpressionType left, ExpressionType right); |
593 | |
594 | template <typename IntType> |
595 | void emitModOrDiv(bool isDiv, ExpressionType lhs, ExpressionType rhs, ExpressionType& result); |
596 | |
597 | enum class MinOrMax { Min, Max }; |
598 | |
599 | PartialResult addFloatingPointMinOrMax(Type, MinOrMax, ExpressionType lhs, ExpressionType rhs, ExpressionType& result); |
600 | |
601 | int32_t WARN_UNUSED_RETURN fixupPointerPlusOffset(ExpressionType&, uint32_t); |
602 | |
603 | void restoreWasmContextInstance(BasicBlock*, TypedTmp); |
604 | enum class RestoreCachedStackLimit { No, Yes }; |
605 | void restoreWebAssemblyGlobalState(RestoreCachedStackLimit, const MemoryInformation&, TypedTmp instance, BasicBlock*); |
606 | |
607 | B3::Origin origin(); |
608 | |
609 | FunctionParser<AirIRGenerator>* m_parser { nullptr }; |
610 | const ModuleInformation& m_info; |
611 | const MemoryMode m_mode { MemoryMode::BoundsChecking }; |
612 | const unsigned m_functionIndex { UINT_MAX }; |
613 | const TierUpCount* m_tierUp { nullptr }; |
614 | |
615 | B3::Procedure& m_proc; |
616 | Code& m_code; |
617 | BasicBlock* m_currentBlock { nullptr }; |
618 | BasicBlock* m_rootBlock { nullptr }; |
619 | Vector<TypedTmp> m_locals; |
620 | Vector<UnlinkedWasmToWasmCall>& m_unlinkedWasmToWasmCalls; // List each call site and the function index whose address it should be patched with. |
621 | GPRReg m_memoryBaseGPR { InvalidGPRReg }; |
622 | GPRReg m_memorySizeGPR { InvalidGPRReg }; |
623 | GPRReg m_wasmContextInstanceGPR { InvalidGPRReg }; |
624 | bool m_makesCalls { false }; |
625 | |
626 | Vector<Tmp, 8> m_freeGPs; |
627 | Vector<Tmp, 8> m_freeFPs; |
628 | |
629 | TypedTmp m_instanceValue; // Always use the accessor below to ensure the instance value is materialized when used. |
630 | bool m_usesInstanceValue { false }; |
631 | TypedTmp instanceValue() |
632 | { |
633 | m_usesInstanceValue = true; |
634 | return m_instanceValue; |
635 | } |
636 | |
637 | uint32_t m_maxNumJSCallArguments { 0 }; |
638 | unsigned m_numImportFunctions; |
639 | |
640 | B3::PatchpointSpecial* m_patchpointSpecial { nullptr }; |
641 | }; |
642 | |
643 | // Memory accesses in WebAssembly have unsigned 32-bit offsets, whereas they have signed 32-bit offsets in B3. |
644 | int32_t AirIRGenerator::fixupPointerPlusOffset(ExpressionType& ptr, uint32_t offset) |
645 | { |
646 | if (static_cast<uint64_t>(offset) > static_cast<uint64_t>(std::numeric_limits<int32_t>::max())) { |
647 | auto previousPtr = ptr; |
648 | ptr = g64(); |
649 | auto constant = g64(); |
650 | append(Move, Arg::bigImm(offset), constant); |
651 | append(Add64, constant, previousPtr, ptr); |
652 | return 0; |
653 | } |
654 | return offset; |
655 | } |
656 | |
657 | void AirIRGenerator::restoreWasmContextInstance(BasicBlock* block, TypedTmp instance) |
658 | { |
659 | if (Context::useFastTLS()) { |
660 | auto* patchpoint = addPatchpoint(B3::Void); |
661 | if (CCallHelpers::storeWasmContextInstanceNeedsMacroScratchRegister()) |
662 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
663 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
664 | AllowMacroScratchRegisterUsageIf allowScratch(jit, CCallHelpers::storeWasmContextInstanceNeedsMacroScratchRegister()); |
665 | jit.storeWasmContextInstance(params[0].gpr()); |
666 | }); |
667 | emitPatchpoint(block, patchpoint, Tmp(), instance); |
668 | return; |
669 | } |
670 | |
671 | // FIXME: Because WasmToWasm call clobbers wasmContextInstance register and does not restore it, we need to restore it in the caller side. |
672 | // This prevents us from using ArgumentReg to this (logically) immutable pinned register. |
673 | auto* patchpoint = addPatchpoint(B3::Void); |
674 | B3::Effects effects = B3::Effects::none(); |
675 | effects.writesPinned = true; |
676 | effects.reads = B3::HeapRange::top(); |
677 | patchpoint->effects = effects; |
678 | patchpoint->clobberLate(RegisterSet(m_wasmContextInstanceGPR)); |
679 | GPRReg wasmContextInstanceGPR = m_wasmContextInstanceGPR; |
680 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& param) { |
681 | jit.move(param[0].gpr(), wasmContextInstanceGPR); |
682 | }); |
683 | emitPatchpoint(block, patchpoint, Tmp(), instance); |
684 | } |
685 | |
686 | AirIRGenerator::AirIRGenerator(const ModuleInformation& info, B3::Procedure& procedure, InternalFunction* compilation, Vector<UnlinkedWasmToWasmCall>& unlinkedWasmToWasmCalls, MemoryMode mode, unsigned functionIndex, TierUpCount* tierUp, ThrowWasmException throwWasmException, const Signature& signature) |
687 | : m_info(info) |
688 | , m_mode(mode) |
689 | , m_functionIndex(functionIndex) |
690 | , m_tierUp(tierUp) |
691 | , m_proc(procedure) |
692 | , m_code(m_proc.code()) |
693 | , m_unlinkedWasmToWasmCalls(unlinkedWasmToWasmCalls) |
694 | , m_numImportFunctions(info.importFunctionCount()) |
695 | { |
696 | m_currentBlock = m_code.addBlock(); |
697 | m_rootBlock = m_currentBlock; |
698 | |
699 | // FIXME we don't really need to pin registers here if there's no memory. It makes wasm -> wasm thunks simpler for now. https://bugs.webkit.org/show_bug.cgi?id=166623 |
700 | const PinnedRegisterInfo& pinnedRegs = PinnedRegisterInfo::get(); |
701 | |
702 | m_memoryBaseGPR = pinnedRegs.baseMemoryPointer; |
703 | m_code.pinRegister(m_memoryBaseGPR); |
704 | |
705 | m_wasmContextInstanceGPR = pinnedRegs.wasmContextInstancePointer; |
706 | if (!Context::useFastTLS()) |
707 | m_code.pinRegister(m_wasmContextInstanceGPR); |
708 | |
709 | if (mode != MemoryMode::Signaling) { |
710 | m_memorySizeGPR = pinnedRegs.sizeRegister; |
711 | m_code.pinRegister(m_memorySizeGPR); |
712 | } |
713 | |
714 | if (throwWasmException) |
715 | Thunks::singleton().setThrowWasmException(throwWasmException); |
716 | |
717 | if (info.memory) { |
718 | switch (m_mode) { |
719 | case MemoryMode::BoundsChecking: |
720 | break; |
721 | case MemoryMode::Signaling: |
722 | // Most memory accesses in signaling mode don't do an explicit |
723 | // exception check because they can rely on fault handling to detect |
724 | // out-of-bounds accesses. FaultSignalHandler nonetheless needs the |
725 | // thunk to exist so that it can jump to that thunk. |
726 | if (UNLIKELY(!Thunks::singleton().stub(throwExceptionFromWasmThunkGenerator))) |
727 | CRASH(); |
728 | break; |
729 | } |
730 | } |
731 | |
732 | m_code.setNumEntrypoints(1); |
733 | |
734 | GPRReg contextInstance = Context::useFastTLS() ? wasmCallingConventionAir().prologueScratch(1) : m_wasmContextInstanceGPR; |
735 | |
736 | Ref<B3::Air::PrologueGenerator> prologueGenerator = createSharedTask<B3::Air::PrologueGeneratorFunction>([=] (CCallHelpers& jit, B3::Air::Code& code) { |
737 | AllowMacroScratchRegisterUsage allowScratch(jit); |
738 | code.emitDefaultPrologue(jit); |
739 | |
740 | { |
741 | GPRReg calleeGPR = wasmCallingConventionAir().prologueScratch(0); |
742 | auto moveLocation = jit.moveWithPatch(MacroAssembler::TrustedImmPtr(nullptr), calleeGPR); |
743 | jit.addLinkTask([compilation, moveLocation] (LinkBuffer& linkBuffer) { |
744 | compilation->calleeMoveLocation = linkBuffer.locationOf<WasmEntryPtrTag>(moveLocation); |
745 | }); |
746 | jit.emitPutToCallFrameHeader(calleeGPR, CallFrameSlot::callee); |
747 | jit.emitPutToCallFrameHeader(nullptr, CallFrameSlot::codeBlock); |
748 | } |
749 | |
750 | { |
751 | const Checked<int32_t> wasmFrameSize = m_code.frameSize(); |
752 | const unsigned minimumParentCheckSize = WTF::roundUpToMultipleOf(stackAlignmentBytes(), 1024); |
753 | const unsigned = WTF::roundUpToMultipleOf(stackAlignmentBytes(), std::max<uint32_t>( |
754 | // This allows us to elide stack checks for functions that are terminal nodes in the call |
755 | // tree, (e.g they don't make any calls) and have a small enough frame size. This works by |
756 | // having any such terminal node have its parent caller include some extra size in its |
757 | // own check for it. The goal here is twofold: |
758 | // 1. Emit less code. |
759 | // 2. Try to speed things up by skipping stack checks. |
760 | minimumParentCheckSize, |
761 | // This allows us to elide stack checks in the Wasm -> Embedder call IC stub. Since these will |
762 | // spill all arguments to the stack, we ensure that a stack check here covers the |
763 | // stack that such a stub would use. |
764 | (Checked<uint32_t>(m_maxNumJSCallArguments) * sizeof(Register) + jscCallingConvention().headerSizeInBytes()).unsafeGet() |
765 | )); |
766 | const int32_t checkSize = m_makesCalls ? (wasmFrameSize + extraFrameSize).unsafeGet() : wasmFrameSize.unsafeGet(); |
767 | bool needUnderflowCheck = static_cast<unsigned>(checkSize) > Options::reservedZoneSize(); |
768 | bool needsOverflowCheck = m_makesCalls || wasmFrameSize >= minimumParentCheckSize || needUnderflowCheck; |
769 | |
770 | // This allows leaf functions to not do stack checks if their frame size is within |
771 | // certain limits since their caller would have already done the check. |
772 | if (needsOverflowCheck) { |
773 | GPRReg scratch = wasmCallingConventionAir().prologueScratch(0); |
774 | |
775 | if (Context::useFastTLS()) |
776 | jit.loadWasmContextInstance(contextInstance); |
777 | |
778 | jit.addPtr(CCallHelpers::TrustedImm32(-checkSize), GPRInfo::callFrameRegister, scratch); |
779 | MacroAssembler::JumpList overflow; |
780 | if (UNLIKELY(needUnderflowCheck)) |
781 | overflow.append(jit.branchPtr(CCallHelpers::Above, scratch, GPRInfo::callFrameRegister)); |
782 | overflow.append(jit.branchPtr(CCallHelpers::Below, scratch, CCallHelpers::Address(contextInstance, Instance::offsetOfCachedStackLimit()))); |
783 | jit.addLinkTask([overflow] (LinkBuffer& linkBuffer) { |
784 | linkBuffer.link(overflow, CodeLocationLabel<JITThunkPtrTag>(Thunks::singleton().stub(throwStackOverflowFromWasmThunkGenerator).code())); |
785 | }); |
786 | } else if (m_usesInstanceValue && Context::useFastTLS()) { |
787 | // No overflow check is needed, but the instance values still needs to be correct. |
788 | jit.loadWasmContextInstance(contextInstance); |
789 | } |
790 | } |
791 | }); |
792 | |
793 | m_code.setPrologueForEntrypoint(0, WTFMove(prologueGenerator)); |
794 | |
795 | if (Context::useFastTLS()) { |
796 | m_instanceValue = g64(); |
797 | // FIXME: Would be nice to only do this if we use instance value. |
798 | append(Move, Tmp(contextInstance), m_instanceValue); |
799 | } else |
800 | m_instanceValue = { Tmp(contextInstance), Type::I64 }; |
801 | |
802 | ASSERT(!m_locals.size()); |
803 | m_locals.grow(signature.argumentCount()); |
804 | for (unsigned i = 0; i < signature.argumentCount(); ++i) { |
805 | Type type = signature.argument(i); |
806 | m_locals[i] = tmpForType(type); |
807 | } |
808 | |
809 | wasmCallingConventionAir().loadArguments(signature, [&] (const Arg& arg, unsigned i) { |
810 | switch (signature.argument(i)) { |
811 | case Type::I32: |
812 | append(Move32, arg, m_locals[i]); |
813 | break; |
814 | case Type::I64: |
815 | case Type::Anyref: |
816 | case Type::Funcref: |
817 | append(Move, arg, m_locals[i]); |
818 | break; |
819 | case Type::F32: |
820 | append(MoveFloat, arg, m_locals[i]); |
821 | break; |
822 | case Type::F64: |
823 | append(MoveDouble, arg, m_locals[i]); |
824 | break; |
825 | default: |
826 | RELEASE_ASSERT_NOT_REACHED(); |
827 | } |
828 | }); |
829 | |
830 | emitTierUpCheck(TierUpCount::functionEntryDecrement(), B3::Origin()); |
831 | } |
832 | |
833 | void AirIRGenerator::restoreWebAssemblyGlobalState(RestoreCachedStackLimit restoreCachedStackLimit, const MemoryInformation& memory, TypedTmp instance, BasicBlock* block) |
834 | { |
835 | restoreWasmContextInstance(block, instance); |
836 | |
837 | if (restoreCachedStackLimit == RestoreCachedStackLimit::Yes) { |
838 | // The Instance caches the stack limit, but also knows where its canonical location is. |
839 | static_assert(sizeof(decltype(static_cast<Instance*>(nullptr)->cachedStackLimit())) == sizeof(uint64_t), "" ); |
840 | |
841 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfPointerToActualStackLimit(), B3::Width64)); |
842 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfCachedStackLimit(), B3::Width64)); |
843 | auto temp = g64(); |
844 | append(block, Move, Arg::addr(instanceValue(), Instance::offsetOfPointerToActualStackLimit()), temp); |
845 | append(block, Move, Arg::addr(temp), temp); |
846 | append(block, Move, temp, Arg::addr(instanceValue(), Instance::offsetOfCachedStackLimit())); |
847 | } |
848 | |
849 | if (!!memory) { |
850 | const PinnedRegisterInfo* pinnedRegs = &PinnedRegisterInfo::get(); |
851 | RegisterSet clobbers; |
852 | clobbers.set(pinnedRegs->baseMemoryPointer); |
853 | clobbers.set(pinnedRegs->sizeRegister); |
854 | if (!isARM64()) |
855 | clobbers.set(RegisterSet::macroScratchRegisters()); |
856 | |
857 | auto* patchpoint = addPatchpoint(B3::Void); |
858 | B3::Effects effects = B3::Effects::none(); |
859 | effects.writesPinned = true; |
860 | effects.reads = B3::HeapRange::top(); |
861 | patchpoint->effects = effects; |
862 | patchpoint->clobber(clobbers); |
863 | patchpoint->numGPScratchRegisters = Gigacage::isEnabled(Gigacage::Primitive) ? 1 : 0; |
864 | |
865 | patchpoint->setGenerator([pinnedRegs] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
866 | AllowMacroScratchRegisterUsage allowScratch(jit); |
867 | GPRReg baseMemory = pinnedRegs->baseMemoryPointer; |
868 | GPRReg scratchOrSize = Gigacage::isEnabled(Gigacage::Primitive) ? params.gpScratch(0) : pinnedRegs->sizeRegister; |
869 | |
870 | jit.loadPtr(CCallHelpers::Address(params[0].gpr(), Instance::offsetOfCachedMemorySize()), pinnedRegs->sizeRegister); |
871 | jit.loadPtr(CCallHelpers::Address(params[0].gpr(), Instance::offsetOfCachedMemory()), baseMemory); |
872 | |
873 | jit.cageConditionally(Gigacage::Primitive, baseMemory, pinnedRegs->sizeRegister, scratchOrSize); |
874 | }); |
875 | |
876 | emitPatchpoint(block, patchpoint, Tmp(), instance); |
877 | } |
878 | } |
879 | |
880 | void AirIRGenerator::emitThrowException(CCallHelpers& jit, ExceptionType type) |
881 | { |
882 | jit.move(CCallHelpers::TrustedImm32(static_cast<uint32_t>(type)), GPRInfo::argumentGPR1); |
883 | auto jumpToExceptionStub = jit.jump(); |
884 | |
885 | jit.addLinkTask([jumpToExceptionStub] (LinkBuffer& linkBuffer) { |
886 | linkBuffer.link(jumpToExceptionStub, CodeLocationLabel<JITThunkPtrTag>(Thunks::singleton().stub(throwExceptionFromWasmThunkGenerator).code())); |
887 | }); |
888 | } |
889 | |
890 | auto AirIRGenerator::addLocal(Type type, uint32_t count) -> PartialResult |
891 | { |
892 | Checked<uint32_t, RecordOverflow> totalBytesChecked = count; |
893 | totalBytesChecked += m_locals.size(); |
894 | uint32_t totalBytes; |
895 | WASM_COMPILE_FAIL_IF((totalBytesChecked.safeGet(totalBytes) == CheckedState::DidOverflow) || !m_locals.tryReserveCapacity(totalBytes), "can't allocate memory for " , totalBytes, " locals" ); |
896 | |
897 | for (uint32_t i = 0; i < count; ++i) { |
898 | auto local = tmpForType(type); |
899 | m_locals.uncheckedAppend(local); |
900 | switch (type) { |
901 | case Type::Anyref: |
902 | case Type::Funcref: |
903 | append(Move, Arg::imm(JSValue::encode(jsNull())), local); |
904 | break; |
905 | case Type::I32: |
906 | case Type::I64: { |
907 | append(Xor64, local, local); |
908 | break; |
909 | } |
910 | case Type::F32: |
911 | case Type::F64: { |
912 | auto temp = g64(); |
913 | // IEEE 754 "0" is just int32/64 zero. |
914 | append(Xor64, temp, temp); |
915 | append(type == Type::F32 ? Move32ToFloat : Move64ToDouble, temp, local); |
916 | break; |
917 | } |
918 | default: |
919 | RELEASE_ASSERT_NOT_REACHED(); |
920 | } |
921 | } |
922 | return { }; |
923 | } |
924 | |
925 | auto AirIRGenerator::addConstant(Type type, uint64_t value) -> ExpressionType |
926 | { |
927 | return addConstant(m_currentBlock, type, value); |
928 | } |
929 | |
930 | auto AirIRGenerator::addConstant(BasicBlock* block, Type type, uint64_t value) -> ExpressionType |
931 | { |
932 | auto result = tmpForType(type); |
933 | switch (type) { |
934 | case Type::I32: |
935 | case Type::I64: |
936 | case Type::Anyref: |
937 | case Type::Funcref: |
938 | append(block, Move, Arg::bigImm(value), result); |
939 | break; |
940 | case Type::F32: |
941 | case Type::F64: { |
942 | auto tmp = g64(); |
943 | append(block, Move, Arg::bigImm(value), tmp); |
944 | append(block, type == Type::F32 ? Move32ToFloat : Move64ToDouble, tmp, result); |
945 | break; |
946 | } |
947 | |
948 | default: |
949 | RELEASE_ASSERT_NOT_REACHED(); |
950 | } |
951 | |
952 | return result; |
953 | } |
954 | |
955 | auto AirIRGenerator::addArguments(const Signature& signature) -> PartialResult |
956 | { |
957 | RELEASE_ASSERT(m_locals.size() == signature.argumentCount()); // We handle arguments in the prologue |
958 | return { }; |
959 | } |
960 | |
961 | auto AirIRGenerator::addRefIsNull(ExpressionType& value, ExpressionType& result) -> PartialResult |
962 | { |
963 | ASSERT(value.tmp()); |
964 | result = tmpForType(Type::I32); |
965 | auto tmp = g64(); |
966 | |
967 | append(Move, Arg::bigImm(JSValue::encode(jsNull())), tmp); |
968 | append(Compare64, Arg::relCond(MacroAssembler::Equal), value, tmp, result); |
969 | |
970 | return { }; |
971 | } |
972 | |
973 | auto AirIRGenerator::addRefFunc(uint32_t index, ExpressionType& result) -> PartialResult |
974 | { |
975 | // FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>. |
976 | result = tmpForType(Type::Funcref); |
977 | emitCCall(&doWasmRefFunc, result, instanceValue(), addConstant(Type::I32, index)); |
978 | |
979 | return { }; |
980 | } |
981 | |
982 | auto AirIRGenerator::addTableGet(unsigned tableIndex, ExpressionType& index, ExpressionType& result) -> PartialResult |
983 | { |
984 | // FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>. |
985 | ASSERT(index.tmp()); |
986 | ASSERT(index.type() == Type::I32); |
987 | result = tmpForType(m_info.tables[tableIndex].wasmType()); |
988 | |
989 | emitCCall(&getWasmTableElement, result, instanceValue(), addConstant(Type::I32, tableIndex), index); |
990 | emitCheck([&] { |
991 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::Zero), result, result); |
992 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
993 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTableAccess); |
994 | }); |
995 | |
996 | return { }; |
997 | } |
998 | |
999 | auto AirIRGenerator::addTableSet(unsigned tableIndex, ExpressionType& index, ExpressionType& value) -> PartialResult |
1000 | { |
1001 | // FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>. |
1002 | ASSERT(index.tmp()); |
1003 | ASSERT(index.type() == Type::I32); |
1004 | ASSERT(value.tmp()); |
1005 | |
1006 | auto shouldThrow = g32(); |
1007 | emitCCall(&setWasmTableElement, shouldThrow, instanceValue(), addConstant(Type::I32, tableIndex), index, value); |
1008 | |
1009 | emitCheck([&] { |
1010 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::Zero), shouldThrow, shouldThrow); |
1011 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1012 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTableAccess); |
1013 | }); |
1014 | |
1015 | return { }; |
1016 | } |
1017 | |
1018 | auto AirIRGenerator::addTableSize(unsigned tableIndex, ExpressionType& result) -> PartialResult |
1019 | { |
1020 | // FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>. |
1021 | result = tmpForType(Type::I32); |
1022 | |
1023 | int32_t (*doSize)(Instance*, unsigned) = [] (Instance* instance, unsigned tableIndex) -> int32_t { |
1024 | return instance->table(tableIndex)->length(); |
1025 | }; |
1026 | |
1027 | emitCCall(doSize, result, instanceValue(), addConstant(Type::I32, tableIndex)); |
1028 | |
1029 | return { }; |
1030 | } |
1031 | |
1032 | auto AirIRGenerator::addTableGrow(unsigned tableIndex, ExpressionType& fill, ExpressionType& delta, ExpressionType& result) -> PartialResult |
1033 | { |
1034 | ASSERT(fill.tmp()); |
1035 | ASSERT(isSubtype(fill.type(), m_info.tables[tableIndex].wasmType())); |
1036 | ASSERT(delta.tmp()); |
1037 | ASSERT(delta.type() == Type::I32); |
1038 | result = tmpForType(Type::I32); |
1039 | |
1040 | emitCCall(&doWasmTableGrow, result, instanceValue(), addConstant(Type::I32, tableIndex), fill, delta); |
1041 | |
1042 | return { }; |
1043 | } |
1044 | |
1045 | auto AirIRGenerator::addTableFill(unsigned tableIndex, ExpressionType& offset, ExpressionType& fill, ExpressionType& count) -> PartialResult |
1046 | { |
1047 | ASSERT(fill.tmp()); |
1048 | ASSERT(isSubtype(fill.type(), m_info.tables[tableIndex].wasmType())); |
1049 | ASSERT(offset.tmp()); |
1050 | ASSERT(offset.type() == Type::I32); |
1051 | ASSERT(count.tmp()); |
1052 | ASSERT(count.type() == Type::I32); |
1053 | |
1054 | auto result = tmpForType(Type::I32); |
1055 | emitCCall(&doWasmTableFill, result, instanceValue(), addConstant(Type::I32, tableIndex), offset, fill, count); |
1056 | |
1057 | emitCheck([&] { |
1058 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::Zero), result, result); |
1059 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1060 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTableAccess); |
1061 | }); |
1062 | |
1063 | return { }; |
1064 | } |
1065 | |
1066 | auto AirIRGenerator::getLocal(uint32_t index, ExpressionType& result) -> PartialResult |
1067 | { |
1068 | ASSERT(m_locals[index].tmp()); |
1069 | result = tmpForType(m_locals[index].type()); |
1070 | append(moveOpForValueType(m_locals[index].type()), m_locals[index].tmp(), result); |
1071 | return { }; |
1072 | } |
1073 | |
1074 | auto AirIRGenerator::addUnreachable() -> PartialResult |
1075 | { |
1076 | B3::PatchpointValue* unreachable = addPatchpoint(B3::Void); |
1077 | unreachable->setGenerator([this] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1078 | this->emitThrowException(jit, ExceptionType::Unreachable); |
1079 | }); |
1080 | unreachable->effects.terminal = true; |
1081 | emitPatchpoint(unreachable, Tmp()); |
1082 | return { }; |
1083 | } |
1084 | |
1085 | auto AirIRGenerator::addGrowMemory(ExpressionType delta, ExpressionType& result) -> PartialResult |
1086 | { |
1087 | int32_t (*growMemory)(void*, Instance*, int32_t) = [] (void* callFrame, Instance* instance, int32_t delta) -> int32_t { |
1088 | instance->storeTopCallFrame(callFrame); |
1089 | |
1090 | if (delta < 0) |
1091 | return -1; |
1092 | |
1093 | auto grown = instance->memory()->grow(PageCount(delta)); |
1094 | if (!grown) { |
1095 | switch (grown.error()) { |
1096 | case Memory::GrowFailReason::InvalidDelta: |
1097 | case Memory::GrowFailReason::InvalidGrowSize: |
1098 | case Memory::GrowFailReason::WouldExceedMaximum: |
1099 | case Memory::GrowFailReason::OutOfMemory: |
1100 | return -1; |
1101 | } |
1102 | RELEASE_ASSERT_NOT_REACHED(); |
1103 | } |
1104 | |
1105 | return grown.value().pageCount(); |
1106 | }; |
1107 | |
1108 | result = g32(); |
1109 | emitCCall(growMemory, result, TypedTmp { Tmp(GPRInfo::callFrameRegister), Type::I64 }, instanceValue(), delta); |
1110 | restoreWebAssemblyGlobalState(RestoreCachedStackLimit::No, m_info.memory, instanceValue(), m_currentBlock); |
1111 | |
1112 | return { }; |
1113 | } |
1114 | |
1115 | auto AirIRGenerator::addCurrentMemory(ExpressionType& result) -> PartialResult |
1116 | { |
1117 | static_assert(sizeof(decltype(static_cast<Memory*>(nullptr)->size())) == sizeof(uint64_t), "codegen relies on this size" ); |
1118 | |
1119 | auto temp1 = g64(); |
1120 | auto temp2 = g64(); |
1121 | |
1122 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfCachedMemorySize(), B3::Width64)); |
1123 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfCachedMemorySize()), temp1); |
1124 | constexpr uint32_t shiftValue = 16; |
1125 | static_assert(PageCount::pageSize == 1ull << shiftValue, "This must hold for the code below to be correct." ); |
1126 | append(Move, Arg::imm(16), temp2); |
1127 | addShift(Type::I32, Urshift64, temp1, temp2, result); |
1128 | append(Move32, result, result); |
1129 | |
1130 | return { }; |
1131 | } |
1132 | |
1133 | auto AirIRGenerator::setLocal(uint32_t index, ExpressionType value) -> PartialResult |
1134 | { |
1135 | ASSERT(m_locals[index].tmp()); |
1136 | append(moveOpForValueType(m_locals[index].type()), value, m_locals[index].tmp()); |
1137 | return { }; |
1138 | } |
1139 | |
1140 | auto AirIRGenerator::getGlobal(uint32_t index, ExpressionType& result) -> PartialResult |
1141 | { |
1142 | Type type = m_info.globals[index].type; |
1143 | |
1144 | result = tmpForType(type); |
1145 | |
1146 | auto temp = g64(); |
1147 | |
1148 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfGlobals(), B3::Width64)); |
1149 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfGlobals()), temp); |
1150 | |
1151 | int32_t offset = safeCast<int32_t>(index * sizeof(Register)); |
1152 | if (Arg::isValidAddrForm(offset, B3::widthForType(toB3Type(type)))) |
1153 | append(moveOpForValueType(type), Arg::addr(temp, offset), result); |
1154 | else { |
1155 | auto temp2 = g64(); |
1156 | append(Move, Arg::bigImm(offset), temp2); |
1157 | append(Add64, temp2, temp, temp); |
1158 | append(moveOpForValueType(type), Arg::addr(temp), result); |
1159 | } |
1160 | return { }; |
1161 | } |
1162 | |
1163 | auto AirIRGenerator::setGlobal(uint32_t index, ExpressionType value) -> PartialResult |
1164 | { |
1165 | auto temp = g64(); |
1166 | |
1167 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfGlobals(), B3::Width64)); |
1168 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfGlobals()), temp); |
1169 | |
1170 | Type type = m_info.globals[index].type; |
1171 | |
1172 | int32_t offset = safeCast<int32_t>(index * sizeof(Register)); |
1173 | if (Arg::isValidAddrForm(offset, B3::widthForType(toB3Type(type)))) |
1174 | append(moveOpForValueType(type), value, Arg::addr(temp, offset)); |
1175 | else { |
1176 | auto temp2 = g64(); |
1177 | append(Move, Arg::bigImm(offset), temp2); |
1178 | append(Add64, temp2, temp, temp); |
1179 | append(moveOpForValueType(type), value, Arg::addr(temp)); |
1180 | } |
1181 | |
1182 | if (isSubtype(type, Anyref)) |
1183 | emitWriteBarrierForJSWrapper(); |
1184 | |
1185 | return { }; |
1186 | } |
1187 | |
1188 | inline void AirIRGenerator::emitWriteBarrierForJSWrapper() |
1189 | { |
1190 | auto cell = g64(); |
1191 | auto vm = g64(); |
1192 | auto cellState = g32(); |
1193 | auto threshold = g32(); |
1194 | |
1195 | BasicBlock* fenceCheckPath = m_code.addBlock(); |
1196 | BasicBlock* fencePath = m_code.addBlock(); |
1197 | BasicBlock* doSlowPath = m_code.addBlock(); |
1198 | BasicBlock* continuation = m_code.addBlock(); |
1199 | |
1200 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfOwner()), cell); |
1201 | append(Move, Arg::addr(cell, JSWebAssemblyInstance::offsetOfVM()), vm); |
1202 | append(Load8, Arg::addr(cell, JSCell::cellStateOffset()), cellState); |
1203 | append(Move32, Arg::addr(vm, VM::offsetOfHeapBarrierThreshold()), threshold); |
1204 | |
1205 | append(Branch32, Arg::relCond(MacroAssembler::Above), cellState, threshold); |
1206 | m_currentBlock->setSuccessors(continuation, fenceCheckPath); |
1207 | m_currentBlock = fenceCheckPath; |
1208 | |
1209 | append(Load8, Arg::addr(vm, VM::offsetOfHeapMutatorShouldBeFenced()), threshold); |
1210 | append(BranchTest32, Arg::resCond(MacroAssembler::Zero), threshold, threshold); |
1211 | m_currentBlock->setSuccessors(doSlowPath, fencePath); |
1212 | m_currentBlock = fencePath; |
1213 | |
1214 | auto* doFence = addPatchpoint(B3::Void); |
1215 | doFence->setGenerator([] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1216 | jit.memoryFence(); |
1217 | }); |
1218 | emitPatchpoint(doFence, Tmp()); |
1219 | |
1220 | append(Load8, Arg::addr(cell, JSCell::cellStateOffset()), cellState); |
1221 | append(Branch32, Arg::relCond(MacroAssembler::Above), cellState, Arg::imm(blackThreshold)); |
1222 | m_currentBlock->setSuccessors(continuation, doSlowPath); |
1223 | m_currentBlock = doSlowPath; |
1224 | |
1225 | void (*writeBarrier)(JSWebAssemblyInstance*, VM*) = [] (JSWebAssemblyInstance* cell, VM* vm) -> void { |
1226 | ASSERT(cell); |
1227 | ASSERT(vm); |
1228 | vm->heap.writeBarrierSlowPath(cell); |
1229 | }; |
1230 | emitCCall(writeBarrier, TypedTmp(), cell, vm); |
1231 | append(Jump); |
1232 | m_currentBlock->setSuccessors(continuation); |
1233 | m_currentBlock = continuation; |
1234 | } |
1235 | |
1236 | inline AirIRGenerator::ExpressionType AirIRGenerator::emitCheckAndPreparePointer(ExpressionType pointer, uint32_t offset, uint32_t sizeOfOperation) |
1237 | { |
1238 | ASSERT(m_memoryBaseGPR); |
1239 | |
1240 | auto result = g64(); |
1241 | append(Move32, pointer, result); |
1242 | |
1243 | switch (m_mode) { |
1244 | case MemoryMode::BoundsChecking: { |
1245 | // We're not using signal handling at all, we must therefore check that no memory access exceeds the current memory size. |
1246 | ASSERT(m_memorySizeGPR); |
1247 | ASSERT(sizeOfOperation + offset > offset); |
1248 | auto temp = g64(); |
1249 | append(Move, Arg::bigImm(static_cast<uint64_t>(sizeOfOperation) + offset - 1), temp); |
1250 | append(Add64, result, temp); |
1251 | |
1252 | emitCheck([&] { |
1253 | return Inst(Branch64, nullptr, Arg::relCond(MacroAssembler::AboveOrEqual), temp, Tmp(m_memorySizeGPR)); |
1254 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1255 | this->emitThrowException(jit, ExceptionType::OutOfBoundsMemoryAccess); |
1256 | }); |
1257 | break; |
1258 | } |
1259 | |
1260 | case MemoryMode::Signaling: { |
1261 | // We've virtually mapped 4GiB+redzone for this memory. Only the user-allocated pages are addressable, contiguously in range [0, current], |
1262 | // and everything above is mapped PROT_NONE. We don't need to perform any explicit bounds check in the 4GiB range because WebAssembly register |
1263 | // memory accesses are 32-bit. However WebAssembly register + offset accesses perform the addition in 64-bit which can push an access above |
1264 | // the 32-bit limit (the offset is unsigned 32-bit). The redzone will catch most small offsets, and we'll explicitly bounds check any |
1265 | // register + large offset access. We don't think this will be generated frequently. |
1266 | // |
1267 | // We could check that register + large offset doesn't exceed 4GiB+redzone since that's technically the limit we need to avoid overflowing the |
1268 | // PROT_NONE region, but it's better if we use a smaller immediate because it can codegens better. We know that anything equal to or greater |
1269 | // than the declared 'maximum' will trap, so we can compare against that number. If there was no declared 'maximum' then we still know that |
1270 | // any access equal to or greater than 4GiB will trap, no need to add the redzone. |
1271 | if (offset >= Memory::fastMappedRedzoneBytes()) { |
1272 | uint64_t maximum = m_info.memory.maximum() ? m_info.memory.maximum().bytes() : std::numeric_limits<uint32_t>::max(); |
1273 | auto temp = g64(); |
1274 | append(Move, Arg::bigImm(static_cast<uint64_t>(sizeOfOperation) + offset - 1), temp); |
1275 | append(Add64, result, temp); |
1276 | auto sizeMax = addConstant(Type::I64, maximum); |
1277 | |
1278 | emitCheck([&] { |
1279 | return Inst(Branch64, nullptr, Arg::relCond(MacroAssembler::AboveOrEqual), temp, sizeMax); |
1280 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1281 | this->emitThrowException(jit, ExceptionType::OutOfBoundsMemoryAccess); |
1282 | }); |
1283 | } |
1284 | break; |
1285 | } |
1286 | } |
1287 | |
1288 | append(Add64, Tmp(m_memoryBaseGPR), result); |
1289 | return result; |
1290 | } |
1291 | |
1292 | inline uint32_t sizeOfLoadOp(LoadOpType op) |
1293 | { |
1294 | switch (op) { |
1295 | case LoadOpType::I32Load8S: |
1296 | case LoadOpType::I32Load8U: |
1297 | case LoadOpType::I64Load8S: |
1298 | case LoadOpType::I64Load8U: |
1299 | return 1; |
1300 | case LoadOpType::I32Load16S: |
1301 | case LoadOpType::I64Load16S: |
1302 | case LoadOpType::I32Load16U: |
1303 | case LoadOpType::I64Load16U: |
1304 | return 2; |
1305 | case LoadOpType::I32Load: |
1306 | case LoadOpType::I64Load32S: |
1307 | case LoadOpType::I64Load32U: |
1308 | case LoadOpType::F32Load: |
1309 | return 4; |
1310 | case LoadOpType::I64Load: |
1311 | case LoadOpType::F64Load: |
1312 | return 8; |
1313 | } |
1314 | RELEASE_ASSERT_NOT_REACHED(); |
1315 | } |
1316 | |
1317 | inline TypedTmp AirIRGenerator::emitLoadOp(LoadOpType op, ExpressionType pointer, uint32_t uoffset) |
1318 | { |
1319 | uint32_t offset = fixupPointerPlusOffset(pointer, uoffset); |
1320 | |
1321 | TypedTmp immTmp; |
1322 | TypedTmp newPtr; |
1323 | TypedTmp result; |
1324 | |
1325 | Arg addrArg; |
1326 | if (Arg::isValidAddrForm(offset, B3::widthForBytes(sizeOfLoadOp(op)))) |
1327 | addrArg = Arg::addr(pointer, offset); |
1328 | else { |
1329 | immTmp = g64(); |
1330 | newPtr = g64(); |
1331 | append(Move, Arg::bigImm(offset), immTmp); |
1332 | append(Add64, immTmp, pointer, newPtr); |
1333 | addrArg = Arg::addr(newPtr); |
1334 | } |
1335 | |
1336 | switch (op) { |
1337 | case LoadOpType::I32Load8S: { |
1338 | result = g32(); |
1339 | appendEffectful(Load8SignedExtendTo32, addrArg, result); |
1340 | break; |
1341 | } |
1342 | |
1343 | case LoadOpType::I64Load8S: { |
1344 | result = g64(); |
1345 | appendEffectful(Load8SignedExtendTo32, addrArg, result); |
1346 | append(SignExtend32ToPtr, result, result); |
1347 | break; |
1348 | } |
1349 | |
1350 | case LoadOpType::I32Load8U: { |
1351 | result = g32(); |
1352 | appendEffectful(Load8, addrArg, result); |
1353 | break; |
1354 | } |
1355 | |
1356 | case LoadOpType::I64Load8U: { |
1357 | result = g64(); |
1358 | appendEffectful(Load8, addrArg, result); |
1359 | break; |
1360 | } |
1361 | |
1362 | case LoadOpType::I32Load16S: { |
1363 | result = g32(); |
1364 | appendEffectful(Load16SignedExtendTo32, addrArg, result); |
1365 | break; |
1366 | } |
1367 | |
1368 | case LoadOpType::I64Load16S: { |
1369 | result = g64(); |
1370 | appendEffectful(Load16SignedExtendTo32, addrArg, result); |
1371 | append(SignExtend32ToPtr, result, result); |
1372 | break; |
1373 | } |
1374 | |
1375 | case LoadOpType::I32Load16U: { |
1376 | result = g32(); |
1377 | appendEffectful(Load16, addrArg, result); |
1378 | break; |
1379 | } |
1380 | |
1381 | case LoadOpType::I64Load16U: { |
1382 | result = g64(); |
1383 | appendEffectful(Load16, addrArg, result); |
1384 | break; |
1385 | } |
1386 | |
1387 | case LoadOpType::I32Load: |
1388 | result = g32(); |
1389 | appendEffectful(Move32, addrArg, result); |
1390 | break; |
1391 | |
1392 | case LoadOpType::I64Load32U: { |
1393 | result = g64(); |
1394 | appendEffectful(Move32, addrArg, result); |
1395 | break; |
1396 | } |
1397 | |
1398 | case LoadOpType::I64Load32S: { |
1399 | result = g64(); |
1400 | appendEffectful(Move32, addrArg, result); |
1401 | append(SignExtend32ToPtr, result, result); |
1402 | break; |
1403 | } |
1404 | |
1405 | case LoadOpType::I64Load: { |
1406 | result = g64(); |
1407 | appendEffectful(Move, addrArg, result); |
1408 | break; |
1409 | } |
1410 | |
1411 | case LoadOpType::F32Load: { |
1412 | result = f32(); |
1413 | appendEffectful(MoveFloat, addrArg, result); |
1414 | break; |
1415 | } |
1416 | |
1417 | case LoadOpType::F64Load: { |
1418 | result = f64(); |
1419 | appendEffectful(MoveDouble, addrArg, result); |
1420 | break; |
1421 | } |
1422 | } |
1423 | |
1424 | return result; |
1425 | } |
1426 | |
1427 | auto AirIRGenerator::load(LoadOpType op, ExpressionType pointer, ExpressionType& result, uint32_t offset) -> PartialResult |
1428 | { |
1429 | ASSERT(pointer.tmp().isGP()); |
1430 | |
1431 | if (UNLIKELY(sumOverflows<uint32_t>(offset, sizeOfLoadOp(op)))) { |
1432 | // FIXME: Even though this is provably out of bounds, it's not a validation error, so we have to handle it |
1433 | // as a runtime exception. However, this may change: https://bugs.webkit.org/show_bug.cgi?id=166435 |
1434 | auto* patch = addPatchpoint(B3::Void); |
1435 | patch->setGenerator([this] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1436 | this->emitThrowException(jit, ExceptionType::OutOfBoundsMemoryAccess); |
1437 | }); |
1438 | emitPatchpoint(patch, Tmp()); |
1439 | |
1440 | // We won't reach here, so we just pick a random reg. |
1441 | switch (op) { |
1442 | case LoadOpType::I32Load8S: |
1443 | case LoadOpType::I32Load16S: |
1444 | case LoadOpType::I32Load: |
1445 | case LoadOpType::I32Load16U: |
1446 | case LoadOpType::I32Load8U: |
1447 | result = g32(); |
1448 | break; |
1449 | case LoadOpType::I64Load8S: |
1450 | case LoadOpType::I64Load8U: |
1451 | case LoadOpType::I64Load16S: |
1452 | case LoadOpType::I64Load32U: |
1453 | case LoadOpType::I64Load32S: |
1454 | case LoadOpType::I64Load: |
1455 | case LoadOpType::I64Load16U: |
1456 | result = g64(); |
1457 | break; |
1458 | case LoadOpType::F32Load: |
1459 | result = f32(); |
1460 | break; |
1461 | case LoadOpType::F64Load: |
1462 | result = f64(); |
1463 | break; |
1464 | } |
1465 | } else |
1466 | result = emitLoadOp(op, emitCheckAndPreparePointer(pointer, offset, sizeOfLoadOp(op)), offset); |
1467 | |
1468 | return { }; |
1469 | } |
1470 | |
1471 | inline uint32_t sizeOfStoreOp(StoreOpType op) |
1472 | { |
1473 | switch (op) { |
1474 | case StoreOpType::I32Store8: |
1475 | case StoreOpType::I64Store8: |
1476 | return 1; |
1477 | case StoreOpType::I32Store16: |
1478 | case StoreOpType::I64Store16: |
1479 | return 2; |
1480 | case StoreOpType::I32Store: |
1481 | case StoreOpType::I64Store32: |
1482 | case StoreOpType::F32Store: |
1483 | return 4; |
1484 | case StoreOpType::I64Store: |
1485 | case StoreOpType::F64Store: |
1486 | return 8; |
1487 | } |
1488 | RELEASE_ASSERT_NOT_REACHED(); |
1489 | } |
1490 | |
1491 | |
1492 | inline void AirIRGenerator::emitStoreOp(StoreOpType op, ExpressionType pointer, ExpressionType value, uint32_t uoffset) |
1493 | { |
1494 | uint32_t offset = fixupPointerPlusOffset(pointer, uoffset); |
1495 | |
1496 | TypedTmp immTmp; |
1497 | TypedTmp newPtr; |
1498 | |
1499 | Arg addrArg; |
1500 | if (Arg::isValidAddrForm(offset, B3::widthForBytes(sizeOfStoreOp(op)))) |
1501 | addrArg = Arg::addr(pointer, offset); |
1502 | else { |
1503 | immTmp = g64(); |
1504 | newPtr = g64(); |
1505 | append(Move, Arg::bigImm(offset), immTmp); |
1506 | append(Add64, immTmp, pointer, newPtr); |
1507 | addrArg = Arg::addr(newPtr); |
1508 | } |
1509 | |
1510 | switch (op) { |
1511 | case StoreOpType::I64Store8: |
1512 | case StoreOpType::I32Store8: |
1513 | append(Store8, value, addrArg); |
1514 | return; |
1515 | |
1516 | case StoreOpType::I64Store16: |
1517 | case StoreOpType::I32Store16: |
1518 | append(Store16, value, addrArg); |
1519 | return; |
1520 | |
1521 | case StoreOpType::I64Store32: |
1522 | case StoreOpType::I32Store: |
1523 | append(Move32, value, addrArg); |
1524 | return; |
1525 | |
1526 | case StoreOpType::I64Store: |
1527 | append(Move, value, addrArg); |
1528 | return; |
1529 | |
1530 | case StoreOpType::F32Store: |
1531 | append(MoveFloat, value, addrArg); |
1532 | return; |
1533 | |
1534 | case StoreOpType::F64Store: |
1535 | append(MoveDouble, value, addrArg); |
1536 | return; |
1537 | } |
1538 | |
1539 | RELEASE_ASSERT_NOT_REACHED(); |
1540 | } |
1541 | |
1542 | auto AirIRGenerator::store(StoreOpType op, ExpressionType pointer, ExpressionType value, uint32_t offset) -> PartialResult |
1543 | { |
1544 | ASSERT(pointer.tmp().isGP()); |
1545 | |
1546 | if (UNLIKELY(sumOverflows<uint32_t>(offset, sizeOfStoreOp(op)))) { |
1547 | // FIXME: Even though this is provably out of bounds, it's not a validation error, so we have to handle it |
1548 | // as a runtime exception. However, this may change: https://bugs.webkit.org/show_bug.cgi?id=166435 |
1549 | auto* throwException = addPatchpoint(B3::Void); |
1550 | throwException->setGenerator([this] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1551 | this->emitThrowException(jit, ExceptionType::OutOfBoundsMemoryAccess); |
1552 | }); |
1553 | emitPatchpoint(throwException, Tmp()); |
1554 | } else |
1555 | emitStoreOp(op, emitCheckAndPreparePointer(pointer, offset, sizeOfStoreOp(op)), value, offset); |
1556 | |
1557 | return { }; |
1558 | } |
1559 | |
1560 | auto AirIRGenerator::addSelect(ExpressionType condition, ExpressionType nonZero, ExpressionType zero, ExpressionType& result) -> PartialResult |
1561 | { |
1562 | ASSERT(nonZero.type() == zero.type()); |
1563 | result = tmpForType(nonZero.type()); |
1564 | append(moveOpForValueType(nonZero.type()), nonZero, result); |
1565 | |
1566 | BasicBlock* isZero = m_code.addBlock(); |
1567 | BasicBlock* continuation = m_code.addBlock(); |
1568 | |
1569 | append(BranchTest32, Arg::resCond(MacroAssembler::Zero), condition, condition); |
1570 | m_currentBlock->setSuccessors(isZero, continuation); |
1571 | |
1572 | append(isZero, moveOpForValueType(zero.type()), zero, result); |
1573 | append(isZero, Jump); |
1574 | isZero->setSuccessors(continuation); |
1575 | |
1576 | m_currentBlock = continuation; |
1577 | |
1578 | return { }; |
1579 | } |
1580 | |
1581 | void AirIRGenerator::emitTierUpCheck(uint32_t decrementCount, B3::Origin origin) |
1582 | { |
1583 | UNUSED_PARAM(origin); |
1584 | |
1585 | if (!m_tierUp) |
1586 | return; |
1587 | |
1588 | auto countdownPtr = g64(); |
1589 | auto oldCountdown = g64(); |
1590 | auto newCountdown = g64(); |
1591 | |
1592 | append(Move, Arg::bigImm(reinterpret_cast<uint64_t>(m_tierUp)), countdownPtr); |
1593 | append(Move32, Arg::addr(countdownPtr), oldCountdown); |
1594 | |
1595 | RELEASE_ASSERT(Arg::isValidImmForm(decrementCount)); |
1596 | append(Move32, oldCountdown, newCountdown); |
1597 | append(Sub32, Arg::imm(decrementCount), newCountdown); |
1598 | append(Move32, newCountdown, Arg::addr(countdownPtr)); |
1599 | |
1600 | auto* patch = addPatchpoint(B3::Void); |
1601 | B3::Effects effects = B3::Effects::none(); |
1602 | effects.reads = B3::HeapRange::top(); |
1603 | effects.writes = B3::HeapRange::top(); |
1604 | patch->effects = effects; |
1605 | |
1606 | patch->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
1607 | MacroAssembler::Jump tierUp = jit.branch32(MacroAssembler::Above, params[0].gpr(), params[1].gpr()); |
1608 | MacroAssembler::Label tierUpResume = jit.label(); |
1609 | |
1610 | params.addLatePath([=] (CCallHelpers& jit) { |
1611 | tierUp.link(&jit); |
1612 | |
1613 | const unsigned = 0; |
1614 | RegisterSet registersToSpill = { }; |
1615 | registersToSpill.add(GPRInfo::argumentGPR1); |
1616 | unsigned numberOfStackBytesUsedForRegisterPreservation = ScratchRegisterAllocator::preserveRegistersToStackForCall(jit, registersToSpill, extraPaddingBytes); |
1617 | |
1618 | jit.move(MacroAssembler::TrustedImm32(m_functionIndex), GPRInfo::argumentGPR1); |
1619 | MacroAssembler::Call call = jit.nearCall(); |
1620 | |
1621 | ScratchRegisterAllocator::restoreRegistersFromStackForCall(jit, registersToSpill, RegisterSet(), numberOfStackBytesUsedForRegisterPreservation, extraPaddingBytes); |
1622 | jit.jump(tierUpResume); |
1623 | |
1624 | jit.addLinkTask([=] (LinkBuffer& linkBuffer) { |
1625 | MacroAssembler::repatchNearCall(linkBuffer.locationOfNearCall<NoPtrTag>(call), CodeLocationLabel<JITThunkPtrTag>(Thunks::singleton().stub(triggerOMGTierUpThunkGenerator).code())); |
1626 | |
1627 | }); |
1628 | }); |
1629 | }); |
1630 | |
1631 | emitPatchpoint(patch, Tmp(), newCountdown, oldCountdown); |
1632 | } |
1633 | |
1634 | AirIRGenerator::ControlData AirIRGenerator::addLoop(Type signature) |
1635 | { |
1636 | BasicBlock* body = m_code.addBlock(); |
1637 | BasicBlock* continuation = m_code.addBlock(); |
1638 | |
1639 | append(Jump); |
1640 | m_currentBlock->setSuccessors(body); |
1641 | |
1642 | m_currentBlock = body; |
1643 | emitTierUpCheck(TierUpCount::loopDecrement(), origin()); |
1644 | |
1645 | return ControlData(origin(), signature, tmpForType(signature), BlockType::Loop, continuation, body); |
1646 | } |
1647 | |
1648 | AirIRGenerator::ControlData AirIRGenerator::addTopLevel(Type signature) |
1649 | { |
1650 | return ControlData(B3::Origin(), signature, tmpForType(signature), BlockType::TopLevel, m_code.addBlock()); |
1651 | } |
1652 | |
1653 | AirIRGenerator::ControlData AirIRGenerator::addBlock(Type signature) |
1654 | { |
1655 | return ControlData(origin(), signature, tmpForType(signature), BlockType::Block, m_code.addBlock()); |
1656 | } |
1657 | |
1658 | auto AirIRGenerator::addIf(ExpressionType condition, Type signature, ControlType& result) -> PartialResult |
1659 | { |
1660 | BasicBlock* taken = m_code.addBlock(); |
1661 | BasicBlock* notTaken = m_code.addBlock(); |
1662 | BasicBlock* continuation = m_code.addBlock(); |
1663 | |
1664 | // Wasm bools are i32. |
1665 | append(BranchTest32, Arg::resCond(MacroAssembler::NonZero), condition, condition); |
1666 | m_currentBlock->setSuccessors(taken, notTaken); |
1667 | |
1668 | m_currentBlock = taken; |
1669 | result = ControlData(origin(), signature, tmpForType(signature), BlockType::If, continuation, notTaken); |
1670 | return { }; |
1671 | } |
1672 | |
1673 | auto AirIRGenerator::addElse(ControlData& data, const ExpressionList& currentStack) -> PartialResult |
1674 | { |
1675 | unifyValuesWithBlock(currentStack, data.result); |
1676 | append(Jump); |
1677 | m_currentBlock->setSuccessors(data.continuation); |
1678 | return addElseToUnreachable(data); |
1679 | } |
1680 | |
1681 | auto AirIRGenerator::addElseToUnreachable(ControlData& data) -> PartialResult |
1682 | { |
1683 | ASSERT(data.type() == BlockType::If); |
1684 | m_currentBlock = data.special; |
1685 | data.convertIfToBlock(); |
1686 | return { }; |
1687 | } |
1688 | |
1689 | auto AirIRGenerator::addReturn(const ControlData& data, const ExpressionList& returnValues) -> PartialResult |
1690 | { |
1691 | ASSERT(returnValues.size() <= 1); |
1692 | if (returnValues.size()) { |
1693 | Tmp returnValueGPR = Tmp(GPRInfo::returnValueGPR); |
1694 | Tmp returnValueFPR = Tmp(FPRInfo::returnValueFPR); |
1695 | switch (data.signature()) { |
1696 | case Type::I32: |
1697 | append(Move32, returnValues[0], returnValueGPR); |
1698 | append(Ret32, returnValueGPR); |
1699 | break; |
1700 | case Type::I64: |
1701 | case Type::Anyref: |
1702 | case Type::Funcref: |
1703 | append(Move, returnValues[0], returnValueGPR); |
1704 | append(Ret64, returnValueGPR); |
1705 | break; |
1706 | case Type::F32: |
1707 | append(MoveFloat, returnValues[0], returnValueFPR); |
1708 | append(RetFloat, returnValueFPR); |
1709 | break; |
1710 | case Type::F64: |
1711 | append(MoveDouble, returnValues[0], returnValueFPR); |
1712 | append(RetFloat, returnValueFPR); |
1713 | break; |
1714 | default: |
1715 | RELEASE_ASSERT_NOT_REACHED(); |
1716 | } |
1717 | } else |
1718 | append(RetVoid); |
1719 | return { }; |
1720 | } |
1721 | |
1722 | // NOTE: All branches in Wasm are on 32-bit ints |
1723 | |
1724 | auto AirIRGenerator::addBranch(ControlData& data, ExpressionType condition, const ExpressionList& returnValues) -> PartialResult |
1725 | { |
1726 | unifyValuesWithBlock(returnValues, data.resultForBranch()); |
1727 | |
1728 | BasicBlock* target = data.targetBlockForBranch(); |
1729 | if (condition) { |
1730 | BasicBlock* continuation = m_code.addBlock(); |
1731 | append(BranchTest32, Arg::resCond(MacroAssembler::NonZero), condition, condition); |
1732 | m_currentBlock->setSuccessors(target, continuation); |
1733 | m_currentBlock = continuation; |
1734 | } else { |
1735 | append(Jump); |
1736 | m_currentBlock->setSuccessors(target); |
1737 | } |
1738 | |
1739 | return { }; |
1740 | } |
1741 | |
1742 | auto AirIRGenerator::addSwitch(ExpressionType condition, const Vector<ControlData*>& targets, ControlData& defaultTarget, const ExpressionList& expressionStack) -> PartialResult |
1743 | { |
1744 | auto& successors = m_currentBlock->successors(); |
1745 | ASSERT(successors.isEmpty()); |
1746 | for (const auto& target : targets) { |
1747 | unifyValuesWithBlock(expressionStack, target->resultForBranch()); |
1748 | successors.append(target->targetBlockForBranch()); |
1749 | } |
1750 | unifyValuesWithBlock(expressionStack, defaultTarget.resultForBranch()); |
1751 | successors.append(defaultTarget.targetBlockForBranch()); |
1752 | |
1753 | ASSERT(condition.type() == Type::I32); |
1754 | |
1755 | // FIXME: We should consider dynamically switching between a jump table |
1756 | // and a binary switch depending on the number of successors. |
1757 | // https://bugs.webkit.org/show_bug.cgi?id=194477 |
1758 | |
1759 | size_t numTargets = targets.size(); |
1760 | |
1761 | auto* patchpoint = addPatchpoint(B3::Void); |
1762 | patchpoint->effects = B3::Effects::none(); |
1763 | patchpoint->effects.terminal = true; |
1764 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
1765 | |
1766 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
1767 | AllowMacroScratchRegisterUsage allowScratch(jit); |
1768 | |
1769 | Vector<int64_t> cases; |
1770 | cases.reserveInitialCapacity(numTargets); |
1771 | for (size_t i = 0; i < numTargets; ++i) |
1772 | cases.uncheckedAppend(i); |
1773 | |
1774 | GPRReg valueReg = params[0].gpr(); |
1775 | BinarySwitch binarySwitch(valueReg, cases, BinarySwitch::Int32); |
1776 | |
1777 | Vector<CCallHelpers::Jump> caseJumps; |
1778 | caseJumps.resize(numTargets); |
1779 | |
1780 | while (binarySwitch.advance(jit)) { |
1781 | unsigned value = binarySwitch.caseValue(); |
1782 | unsigned index = binarySwitch.caseIndex(); |
1783 | ASSERT_UNUSED(value, value == index); |
1784 | ASSERT(index < numTargets); |
1785 | caseJumps[index] = jit.jump(); |
1786 | } |
1787 | |
1788 | CCallHelpers::JumpList fallThrough = binarySwitch.fallThrough(); |
1789 | |
1790 | Vector<Box<CCallHelpers::Label>> successorLabels = params.successorLabels(); |
1791 | ASSERT(successorLabels.size() == caseJumps.size() + 1); |
1792 | |
1793 | params.addLatePath([=, caseJumps = WTFMove(caseJumps), successorLabels = WTFMove(successorLabels)] (CCallHelpers& jit) { |
1794 | for (size_t i = 0; i < numTargets; ++i) |
1795 | caseJumps[i].linkTo(*successorLabels[i], &jit); |
1796 | fallThrough.linkTo(*successorLabels[numTargets], &jit); |
1797 | }); |
1798 | }); |
1799 | |
1800 | emitPatchpoint(patchpoint, TypedTmp(), condition); |
1801 | |
1802 | return { }; |
1803 | } |
1804 | |
1805 | auto AirIRGenerator::endBlock(ControlEntry& entry, ExpressionList& expressionStack) -> PartialResult |
1806 | { |
1807 | ControlData& data = entry.controlData; |
1808 | |
1809 | unifyValuesWithBlock(expressionStack, data.result); |
1810 | append(Jump); |
1811 | m_currentBlock->setSuccessors(data.continuation); |
1812 | |
1813 | return addEndToUnreachable(entry); |
1814 | } |
1815 | |
1816 | |
1817 | auto AirIRGenerator::addEndToUnreachable(ControlEntry& entry) -> PartialResult |
1818 | { |
1819 | ControlData& data = entry.controlData; |
1820 | m_currentBlock = data.continuation; |
1821 | |
1822 | if (data.type() == BlockType::If) { |
1823 | append(data.special, Jump); |
1824 | data.special->setSuccessors(m_currentBlock); |
1825 | } |
1826 | |
1827 | for (const auto& result : data.result) |
1828 | entry.enclosedExpressionStack.append(result); |
1829 | |
1830 | // TopLevel does not have any code after this so we need to make sure we emit a return here. |
1831 | if (data.type() == BlockType::TopLevel) |
1832 | return addReturn(data, entry.enclosedExpressionStack); |
1833 | |
1834 | return { }; |
1835 | } |
1836 | |
1837 | auto AirIRGenerator::addCall(uint32_t functionIndex, const Signature& signature, Vector<ExpressionType>& args, ExpressionType& result) -> PartialResult |
1838 | { |
1839 | ASSERT(signature.argumentCount() == args.size()); |
1840 | |
1841 | m_makesCalls = true; |
1842 | |
1843 | Type returnType = signature.returnType(); |
1844 | if (returnType != Type::Void) |
1845 | result = tmpForType(returnType); |
1846 | |
1847 | Vector<UnlinkedWasmToWasmCall>* unlinkedWasmToWasmCalls = &m_unlinkedWasmToWasmCalls; |
1848 | |
1849 | if (m_info.isImportedFunctionFromFunctionIndexSpace(functionIndex)) { |
1850 | m_maxNumJSCallArguments = std::max(m_maxNumJSCallArguments, static_cast<uint32_t>(args.size())); |
1851 | |
1852 | auto currentInstance = g64(); |
1853 | append(Move, instanceValue(), currentInstance); |
1854 | |
1855 | auto targetInstance = g64(); |
1856 | |
1857 | // FIXME: We should have better isel here. |
1858 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
1859 | append(Move, Arg::bigImm(Instance::offsetOfTargetInstance(functionIndex)), targetInstance); |
1860 | append(Add64, instanceValue(), targetInstance); |
1861 | append(Move, Arg::addr(targetInstance), targetInstance); |
1862 | |
1863 | BasicBlock* isWasmBlock = m_code.addBlock(); |
1864 | BasicBlock* isEmbedderBlock = m_code.addBlock(); |
1865 | BasicBlock* continuation = m_code.addBlock(); |
1866 | |
1867 | append(BranchTest64, Arg::resCond(MacroAssembler::NonZero), targetInstance, targetInstance); |
1868 | m_currentBlock->setSuccessors(isWasmBlock, isEmbedderBlock); |
1869 | |
1870 | { |
1871 | auto* patchpoint = addPatchpoint(toB3Type(returnType)); |
1872 | patchpoint->effects.writesPinned = true; |
1873 | patchpoint->effects.readsPinned = true; |
1874 | // We need to clobber all potential pinned registers since we might be leaving the instance. |
1875 | // We pessimistically assume we could be calling to something that is bounds checking. |
1876 | // FIXME: We shouldn't have to do this: https://bugs.webkit.org/show_bug.cgi?id=172181 |
1877 | patchpoint->clobberLate(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking)); |
1878 | |
1879 | Vector<ConstrainedTmp> patchArgs; |
1880 | wasmCallingConventionAir().setupCall(m_code, returnType, patchpoint, toTmpVector(args), [&] (Tmp tmp, B3::ValueRep rep) { |
1881 | patchArgs.append({ tmp, rep }); |
1882 | }); |
1883 | |
1884 | patchpoint->setGenerator([unlinkedWasmToWasmCalls, functionIndex] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1885 | AllowMacroScratchRegisterUsage allowScratch(jit); |
1886 | CCallHelpers::Call call = jit.threadSafePatchableNearCall(); |
1887 | jit.addLinkTask([unlinkedWasmToWasmCalls, call, functionIndex] (LinkBuffer& linkBuffer) { |
1888 | unlinkedWasmToWasmCalls->append({ linkBuffer.locationOfNearCall<WasmEntryPtrTag>(call), functionIndex }); |
1889 | }); |
1890 | }); |
1891 | |
1892 | emitPatchpoint(isWasmBlock, patchpoint, result, WTFMove(patchArgs)); |
1893 | append(isWasmBlock, Jump); |
1894 | isWasmBlock->setSuccessors(continuation); |
1895 | } |
1896 | |
1897 | { |
1898 | auto jumpDestination = g64(); |
1899 | append(isEmbedderBlock, Move, Arg::bigImm(Instance::offsetOfWasmToEmbedderStub(functionIndex)), jumpDestination); |
1900 | append(isEmbedderBlock, Add64, instanceValue(), jumpDestination); |
1901 | append(isEmbedderBlock, Move, Arg::addr(jumpDestination), jumpDestination); |
1902 | |
1903 | auto* patchpoint = addPatchpoint(toB3Type(returnType)); |
1904 | patchpoint->effects.writesPinned = true; |
1905 | patchpoint->effects.readsPinned = true; |
1906 | // We need to clobber all potential pinned registers since we might be leaving the instance. |
1907 | // We pessimistically assume we could be calling to something that is bounds checking. |
1908 | // FIXME: We shouldn't have to do this: https://bugs.webkit.org/show_bug.cgi?id=172181 |
1909 | patchpoint->clobberLate(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking)); |
1910 | |
1911 | Vector<ConstrainedTmp> patchArgs; |
1912 | patchArgs.append(jumpDestination); |
1913 | |
1914 | wasmCallingConventionAir().setupCall(m_code, returnType, patchpoint, toTmpVector(args), [&] (Tmp tmp, B3::ValueRep rep) { |
1915 | patchArgs.append({ tmp, rep }); |
1916 | }); |
1917 | |
1918 | patchpoint->setGenerator([returnType] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
1919 | AllowMacroScratchRegisterUsage allowScratch(jit); |
1920 | jit.call(params[returnType == Void ? 0 : 1].gpr(), WasmEntryPtrTag); |
1921 | }); |
1922 | |
1923 | emitPatchpoint(isEmbedderBlock, patchpoint, result, WTFMove(patchArgs)); |
1924 | append(isEmbedderBlock, Jump); |
1925 | isEmbedderBlock->setSuccessors(continuation); |
1926 | } |
1927 | |
1928 | m_currentBlock = continuation; |
1929 | // The call could have been to another WebAssembly instance, and / or could have modified our Memory. |
1930 | restoreWebAssemblyGlobalState(RestoreCachedStackLimit::Yes, m_info.memory, currentInstance, continuation); |
1931 | } else { |
1932 | auto* patchpoint = addPatchpoint(toB3Type(returnType)); |
1933 | patchpoint->effects.writesPinned = true; |
1934 | patchpoint->effects.readsPinned = true; |
1935 | |
1936 | Vector<ConstrainedTmp> patchArgs; |
1937 | wasmCallingConventionAir().setupCall(m_code, returnType, patchpoint, toTmpVector(args), [&] (Tmp tmp, B3::ValueRep rep) { |
1938 | patchArgs.append({ tmp, rep }); |
1939 | }); |
1940 | |
1941 | patchpoint->setGenerator([unlinkedWasmToWasmCalls, functionIndex] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1942 | AllowMacroScratchRegisterUsage allowScratch(jit); |
1943 | CCallHelpers::Call call = jit.threadSafePatchableNearCall(); |
1944 | jit.addLinkTask([unlinkedWasmToWasmCalls, call, functionIndex] (LinkBuffer& linkBuffer) { |
1945 | unlinkedWasmToWasmCalls->append({ linkBuffer.locationOfNearCall<WasmEntryPtrTag>(call), functionIndex }); |
1946 | }); |
1947 | }); |
1948 | |
1949 | emitPatchpoint(m_currentBlock, patchpoint, result, WTFMove(patchArgs)); |
1950 | } |
1951 | |
1952 | return { }; |
1953 | } |
1954 | |
1955 | auto AirIRGenerator::addCallIndirect(unsigned tableIndex, const Signature& signature, Vector<ExpressionType>& args, ExpressionType& result) -> PartialResult |
1956 | { |
1957 | ExpressionType calleeIndex = args.takeLast(); |
1958 | ASSERT(signature.argumentCount() == args.size()); |
1959 | ASSERT(m_info.tableCount() > tableIndex); |
1960 | ASSERT(m_info.tables[tableIndex].type() == TableElementType::Funcref); |
1961 | |
1962 | m_makesCalls = true; |
1963 | // Note: call indirect can call either WebAssemblyFunction or WebAssemblyWrapperFunction. Because |
1964 | // WebAssemblyWrapperFunction is like calling into the embedder, we conservatively assume all call indirects |
1965 | // can be to the embedder for our stack check calculation. |
1966 | m_maxNumJSCallArguments = std::max(m_maxNumJSCallArguments, static_cast<uint32_t>(args.size())); |
1967 | |
1968 | auto currentInstance = g64(); |
1969 | append(Move, instanceValue(), currentInstance); |
1970 | |
1971 | ExpressionType callableFunctionBuffer = g64(); |
1972 | ExpressionType instancesBuffer = g64(); |
1973 | ExpressionType callableFunctionBufferLength = g64(); |
1974 | { |
1975 | RELEASE_ASSERT(Arg::isValidAddrForm(FuncRefTable::offsetOfFunctions(), B3::Width64)); |
1976 | RELEASE_ASSERT(Arg::isValidAddrForm(FuncRefTable::offsetOfInstances(), B3::Width64)); |
1977 | RELEASE_ASSERT(Arg::isValidAddrForm(FuncRefTable::offsetOfLength(), B3::Width64)); |
1978 | |
1979 | if (UNLIKELY(!Arg::isValidAddrForm(Instance::offsetOfTablePtr(m_numImportFunctions, tableIndex), B3::Width64))) { |
1980 | append(Move, Arg::bigImm(Instance::offsetOfTablePtr(m_numImportFunctions, tableIndex)), callableFunctionBufferLength); |
1981 | append(Add64, instanceValue(), callableFunctionBufferLength); |
1982 | append(Move, Arg::addr(callableFunctionBufferLength), callableFunctionBufferLength); |
1983 | } else |
1984 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfTablePtr(m_numImportFunctions, tableIndex)), callableFunctionBufferLength); |
1985 | append(Move, Arg::addr(callableFunctionBufferLength, FuncRefTable::offsetOfFunctions()), callableFunctionBuffer); |
1986 | append(Move, Arg::addr(callableFunctionBufferLength, FuncRefTable::offsetOfInstances()), instancesBuffer); |
1987 | append(Move32, Arg::addr(callableFunctionBufferLength, Table::offsetOfLength()), callableFunctionBufferLength); |
1988 | } |
1989 | |
1990 | append(Move32, calleeIndex, calleeIndex); |
1991 | |
1992 | // Check the index we are looking for is valid. |
1993 | emitCheck([&] { |
1994 | return Inst(Branch32, nullptr, Arg::relCond(MacroAssembler::AboveOrEqual), calleeIndex, callableFunctionBufferLength); |
1995 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1996 | this->emitThrowException(jit, ExceptionType::OutOfBoundsCallIndirect); |
1997 | }); |
1998 | |
1999 | ExpressionType calleeCode = g64(); |
2000 | { |
2001 | ExpressionType calleeSignatureIndex = g64(); |
2002 | // Compute the offset in the table index space we are looking for. |
2003 | append(Move, Arg::imm(sizeof(WasmToWasmImportableFunction)), calleeSignatureIndex); |
2004 | append(Mul64, calleeIndex, calleeSignatureIndex); |
2005 | append(Add64, callableFunctionBuffer, calleeSignatureIndex); |
2006 | |
2007 | append(Move, Arg::addr(calleeSignatureIndex, WasmToWasmImportableFunction::offsetOfEntrypointLoadLocation()), calleeCode); // Pointer to callee code. |
2008 | |
2009 | // Check that the WasmToWasmImportableFunction is initialized. We trap if it isn't. An "invalid" SignatureIndex indicates it's not initialized. |
2010 | // FIXME: when we have trap handlers, we can just let the call fail because Signature::invalidIndex is 0. https://bugs.webkit.org/show_bug.cgi?id=177210 |
2011 | static_assert(sizeof(WasmToWasmImportableFunction::signatureIndex) == sizeof(uint64_t), "Load codegen assumes i64" ); |
2012 | |
2013 | // FIXME: This seems dumb to do two checks just for a nicer error message. |
2014 | // We should move just to use a single branch and then figure out what |
2015 | // error to use in the exception handler. |
2016 | |
2017 | append(Move, Arg::addr(calleeSignatureIndex, WasmToWasmImportableFunction::offsetOfSignatureIndex()), calleeSignatureIndex); |
2018 | |
2019 | emitCheck([&] { |
2020 | static_assert(Signature::invalidIndex == 0, "" ); |
2021 | return Inst(BranchTest64, nullptr, Arg::resCond(MacroAssembler::Zero), calleeSignatureIndex, calleeSignatureIndex); |
2022 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2023 | this->emitThrowException(jit, ExceptionType::NullTableEntry); |
2024 | }); |
2025 | |
2026 | ExpressionType expectedSignatureIndex = g64(); |
2027 | append(Move, Arg::bigImm(SignatureInformation::get(signature)), expectedSignatureIndex); |
2028 | emitCheck([&] { |
2029 | return Inst(Branch64, nullptr, Arg::relCond(MacroAssembler::NotEqual), calleeSignatureIndex, expectedSignatureIndex); |
2030 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2031 | this->emitThrowException(jit, ExceptionType::BadSignature); |
2032 | }); |
2033 | } |
2034 | |
2035 | // Do a context switch if needed. |
2036 | { |
2037 | auto newContextInstance = g64(); |
2038 | append(Move, Arg::index(instancesBuffer, calleeIndex, 8, 0), newContextInstance); |
2039 | |
2040 | BasicBlock* doContextSwitch = m_code.addBlock(); |
2041 | BasicBlock* continuation = m_code.addBlock(); |
2042 | |
2043 | append(Branch64, Arg::relCond(MacroAssembler::Equal), newContextInstance, instanceValue()); |
2044 | m_currentBlock->setSuccessors(continuation, doContextSwitch); |
2045 | |
2046 | auto* patchpoint = addPatchpoint(B3::Void); |
2047 | patchpoint->effects.writesPinned = true; |
2048 | // We pessimistically assume we're calling something with BoundsChecking memory. |
2049 | // FIXME: We shouldn't have to do this: https://bugs.webkit.org/show_bug.cgi?id=172181 |
2050 | patchpoint->clobber(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking)); |
2051 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2052 | patchpoint->numGPScratchRegisters = Gigacage::isEnabled(Gigacage::Primitive) ? 1 : 0; |
2053 | |
2054 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2055 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2056 | GPRReg newContextInstance = params[0].gpr(); |
2057 | GPRReg oldContextInstance = params[1].gpr(); |
2058 | const PinnedRegisterInfo& pinnedRegs = PinnedRegisterInfo::get(); |
2059 | GPRReg baseMemory = pinnedRegs.baseMemoryPointer; |
2060 | ASSERT(newContextInstance != baseMemory); |
2061 | jit.loadPtr(CCallHelpers::Address(oldContextInstance, Instance::offsetOfCachedStackLimit()), baseMemory); |
2062 | jit.storePtr(baseMemory, CCallHelpers::Address(newContextInstance, Instance::offsetOfCachedStackLimit())); |
2063 | jit.storeWasmContextInstance(newContextInstance); |
2064 | // FIXME: We should support more than one memory size register |
2065 | // see: https://bugs.webkit.org/show_bug.cgi?id=162952 |
2066 | ASSERT(pinnedRegs.sizeRegister != newContextInstance); |
2067 | GPRReg scratchOrSize = Gigacage::isEnabled(Gigacage::Primitive) ? params.gpScratch(0) : pinnedRegs.sizeRegister; |
2068 | |
2069 | jit.loadPtr(CCallHelpers::Address(newContextInstance, Instance::offsetOfCachedMemorySize()), pinnedRegs.sizeRegister); // Memory size. |
2070 | jit.loadPtr(CCallHelpers::Address(newContextInstance, Instance::offsetOfCachedMemory()), baseMemory); // Memory::void*. |
2071 | |
2072 | jit.cageConditionally(Gigacage::Primitive, baseMemory, pinnedRegs.sizeRegister, scratchOrSize); |
2073 | }); |
2074 | |
2075 | emitPatchpoint(doContextSwitch, patchpoint, Tmp(), newContextInstance, instanceValue()); |
2076 | append(doContextSwitch, Jump); |
2077 | doContextSwitch->setSuccessors(continuation); |
2078 | |
2079 | m_currentBlock = continuation; |
2080 | } |
2081 | |
2082 | append(Move, Arg::addr(calleeCode), calleeCode); |
2083 | |
2084 | Type returnType = signature.returnType(); |
2085 | if (returnType != Type::Void) |
2086 | result = tmpForType(returnType); |
2087 | |
2088 | auto* patch = addPatchpoint(toB3Type(returnType)); |
2089 | patch->effects.writesPinned = true; |
2090 | patch->effects.readsPinned = true; |
2091 | // We need to clobber all potential pinned registers since we might be leaving the instance. |
2092 | // We pessimistically assume we're always calling something that is bounds checking so |
2093 | // because the wasm->wasm thunk unconditionally overrides the size registers. |
2094 | // FIXME: We should not have to do this, but the wasm->wasm stub assumes it can |
2095 | // use all the pinned registers as scratch: https://bugs.webkit.org/show_bug.cgi?id=172181 |
2096 | patch->clobberLate(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking)); |
2097 | |
2098 | Vector<ConstrainedTmp> emitArgs; |
2099 | emitArgs.append(calleeCode); |
2100 | wasmCallingConventionAir().setupCall(m_code, returnType, patch, toTmpVector(args), [&] (Tmp tmp, B3::ValueRep rep) { |
2101 | emitArgs.append({ tmp, rep }); |
2102 | }); |
2103 | patch->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2104 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2105 | jit.call(params[returnType == Void ? 0 : 1].gpr(), WasmEntryPtrTag); |
2106 | }); |
2107 | |
2108 | emitPatchpoint(m_currentBlock, patch, result, WTFMove(emitArgs)); |
2109 | |
2110 | // The call could have been to another WebAssembly instance, and / or could have modified our Memory. |
2111 | restoreWebAssemblyGlobalState(RestoreCachedStackLimit::Yes, m_info.memory, currentInstance, m_currentBlock); |
2112 | |
2113 | return { }; |
2114 | } |
2115 | |
2116 | void AirIRGenerator::unify(const ExpressionType& dst, const ExpressionType& source) |
2117 | { |
2118 | ASSERT(isSubtype(source.type(), dst.type())); |
2119 | append(moveOpForValueType(dst.type()), source, dst); |
2120 | } |
2121 | |
2122 | void AirIRGenerator::unifyValuesWithBlock(const ExpressionList& resultStack, const ResultList& result) |
2123 | { |
2124 | ASSERT(result.size() <= resultStack.size()); |
2125 | |
2126 | for (size_t i = 0; i < result.size(); ++i) |
2127 | unify(result[result.size() - 1 - i], resultStack[resultStack.size() - 1 - i]); |
2128 | } |
2129 | |
2130 | void AirIRGenerator::dump(const Vector<ControlEntry>&, const ExpressionList*) |
2131 | { |
2132 | } |
2133 | |
2134 | auto AirIRGenerator::origin() -> B3::Origin |
2135 | { |
2136 | // FIXME: We should implement a way to give Inst's an origin. |
2137 | return B3::Origin(); |
2138 | } |
2139 | |
2140 | Expected<std::unique_ptr<InternalFunction>, String> parseAndCompileAir(CompilationContext& compilationContext, const uint8_t* functionStart, size_t functionLength, const Signature& signature, Vector<UnlinkedWasmToWasmCall>& unlinkedWasmToWasmCalls, const ModuleInformation& info, MemoryMode mode, uint32_t functionIndex, TierUpCount* tierUp, ThrowWasmException throwWasmException) |
2141 | { |
2142 | auto result = std::make_unique<InternalFunction>(); |
2143 | |
2144 | compilationContext.embedderEntrypointJIT = std::make_unique<CCallHelpers>(); |
2145 | compilationContext.wasmEntrypointJIT = std::make_unique<CCallHelpers>(); |
2146 | |
2147 | B3::Procedure procedure; |
2148 | Code& code = procedure.code(); |
2149 | |
2150 | procedure.setOriginPrinter([] (PrintStream& out, B3::Origin origin) { |
2151 | if (origin.data()) |
2152 | out.print("Wasm: " , bitwise_cast<OpcodeOrigin>(origin)); |
2153 | }); |
2154 | |
2155 | // This means we cannot use either StackmapGenerationParams::usedRegisters() or |
2156 | // StackmapGenerationParams::unavailableRegisters(). In exchange for this concession, we |
2157 | // don't strictly need to run Air::reportUsedRegisters(), which saves a bit of CPU time at |
2158 | // optLevel=1. |
2159 | procedure.setNeedsUsedRegisters(false); |
2160 | |
2161 | procedure.setOptLevel(Options::webAssemblyBBQOptimizationLevel()); |
2162 | |
2163 | AirIRGenerator irGenerator(info, procedure, result.get(), unlinkedWasmToWasmCalls, mode, functionIndex, tierUp, throwWasmException, signature); |
2164 | FunctionParser<AirIRGenerator> parser(irGenerator, functionStart, functionLength, signature, info); |
2165 | WASM_FAIL_IF_HELPER_FAILS(parser.parse()); |
2166 | |
2167 | |
2168 | for (BasicBlock* block : code) { |
2169 | for (size_t i = 0; i < block->numSuccessors(); ++i) |
2170 | block->successorBlock(i)->addPredecessor(block); |
2171 | } |
2172 | |
2173 | { |
2174 | B3::Air::prepareForGeneration(code); |
2175 | B3::Air::generate(code, *compilationContext.wasmEntrypointJIT); |
2176 | compilationContext.wasmEntrypointByproducts = procedure.releaseByproducts(); |
2177 | result->entrypoint.calleeSaveRegisters = code.calleeSaveRegisterAtOffsetList(); |
2178 | } |
2179 | |
2180 | return result; |
2181 | } |
2182 | |
2183 | template <typename IntType> |
2184 | void AirIRGenerator::emitChecksForModOrDiv(bool isSignedDiv, ExpressionType left, ExpressionType right) |
2185 | { |
2186 | static_assert(sizeof(IntType) == 4 || sizeof(IntType) == 8, "" ); |
2187 | |
2188 | emitCheck([&] { |
2189 | return Inst(sizeof(IntType) == 4 ? BranchTest32 : BranchTest64, nullptr, Arg::resCond(MacroAssembler::Zero), right, right); |
2190 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2191 | this->emitThrowException(jit, ExceptionType::DivisionByZero); |
2192 | }); |
2193 | |
2194 | if (isSignedDiv) { |
2195 | ASSERT(std::is_signed<IntType>::value); |
2196 | IntType min = std::numeric_limits<IntType>::min(); |
2197 | |
2198 | // FIXME: Better isel for compare with imms here. |
2199 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
2200 | auto minTmp = sizeof(IntType) == 4 ? g32() : g64(); |
2201 | auto negOne = sizeof(IntType) == 4 ? g32() : g64(); |
2202 | |
2203 | B3::Air::Opcode op = sizeof(IntType) == 4 ? Compare32 : Compare64; |
2204 | append(Move, Arg::bigImm(static_cast<uint64_t>(min)), minTmp); |
2205 | append(op, Arg::relCond(MacroAssembler::Equal), left, minTmp, minTmp); |
2206 | |
2207 | append(Move, Arg::imm(-1), negOne); |
2208 | append(op, Arg::relCond(MacroAssembler::Equal), right, negOne, negOne); |
2209 | |
2210 | emitCheck([&] { |
2211 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), minTmp, negOne); |
2212 | }, |
2213 | [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2214 | this->emitThrowException(jit, ExceptionType::IntegerOverflow); |
2215 | }); |
2216 | } |
2217 | } |
2218 | |
2219 | template <typename IntType> |
2220 | void AirIRGenerator::emitModOrDiv(bool isDiv, ExpressionType lhs, ExpressionType rhs, ExpressionType& result) |
2221 | { |
2222 | static_assert(sizeof(IntType) == 4 || sizeof(IntType) == 8, "" ); |
2223 | |
2224 | result = sizeof(IntType) == 4 ? g32() : g64(); |
2225 | |
2226 | bool isSigned = std::is_signed<IntType>::value; |
2227 | |
2228 | if (isARM64()) { |
2229 | B3::Air::Opcode div; |
2230 | switch (sizeof(IntType)) { |
2231 | case 4: |
2232 | div = isSigned ? Div32 : UDiv32; |
2233 | break; |
2234 | case 8: |
2235 | div = isSigned ? Div64 : UDiv64; |
2236 | break; |
2237 | } |
2238 | |
2239 | append(div, lhs, rhs, result); |
2240 | |
2241 | if (!isDiv) { |
2242 | append(sizeof(IntType) == 4 ? Mul32 : Mul64, result, rhs, result); |
2243 | append(sizeof(IntType) == 4 ? Sub32 : Sub64, lhs, result, result); |
2244 | } |
2245 | |
2246 | return; |
2247 | } |
2248 | |
2249 | #if CPU(X86) || CPU(X86_64) |
2250 | Tmp eax(X86Registers::eax); |
2251 | Tmp edx(X86Registers::edx); |
2252 | |
2253 | if (isSigned) { |
2254 | B3::Air::Opcode convertToDoubleWord; |
2255 | B3::Air::Opcode div; |
2256 | switch (sizeof(IntType)) { |
2257 | case 4: |
2258 | convertToDoubleWord = X86ConvertToDoubleWord32; |
2259 | div = X86Div32; |
2260 | break; |
2261 | case 8: |
2262 | convertToDoubleWord = X86ConvertToQuadWord64; |
2263 | div = X86Div64; |
2264 | break; |
2265 | default: |
2266 | RELEASE_ASSERT_NOT_REACHED(); |
2267 | } |
2268 | |
2269 | // We implement "res = Div<Chill>/Mod<Chill>(num, den)" as follows: |
2270 | // |
2271 | // if (den + 1 <=_unsigned 1) { |
2272 | // if (!den) { |
2273 | // res = 0; |
2274 | // goto done; |
2275 | // } |
2276 | // if (num == -2147483648) { |
2277 | // res = isDiv ? num : 0; |
2278 | // goto done; |
2279 | // } |
2280 | // } |
2281 | // res = num (/ or %) dev; |
2282 | // done: |
2283 | |
2284 | BasicBlock* denIsGood = m_code.addBlock(); |
2285 | BasicBlock* denMayBeBad = m_code.addBlock(); |
2286 | BasicBlock* denNotZero = m_code.addBlock(); |
2287 | BasicBlock* continuation = m_code.addBlock(); |
2288 | |
2289 | auto temp = sizeof(IntType) == 4 ? g32() : g64(); |
2290 | auto one = addConstant(sizeof(IntType) == 4 ? Type::I32 : Type::I64, 1); |
2291 | |
2292 | append(sizeof(IntType) == 4 ? Add32 : Add64, rhs, one, temp); |
2293 | append(sizeof(IntType) == 4 ? Branch32 : Branch64, Arg::relCond(MacroAssembler::Above), temp, one); |
2294 | m_currentBlock->setSuccessors(denIsGood, denMayBeBad); |
2295 | |
2296 | append(denMayBeBad, Xor64, result, result); |
2297 | append(denMayBeBad, sizeof(IntType) == 4 ? BranchTest32 : BranchTest64, Arg::resCond(MacroAssembler::Zero), rhs, rhs); |
2298 | denMayBeBad->setSuccessors(continuation, denNotZero); |
2299 | |
2300 | auto min = addConstant(denNotZero, sizeof(IntType) == 4 ? Type::I32 : Type::I64, std::numeric_limits<IntType>::min()); |
2301 | if (isDiv) |
2302 | append(denNotZero, sizeof(IntType) == 4 ? Move32 : Move, min, result); |
2303 | else { |
2304 | // Result is zero, as set above... |
2305 | } |
2306 | append(denNotZero, sizeof(IntType) == 4 ? Branch32 : Branch64, Arg::relCond(MacroAssembler::Equal), lhs, min); |
2307 | denNotZero->setSuccessors(continuation, denIsGood); |
2308 | |
2309 | auto divResult = isDiv ? eax : edx; |
2310 | append(denIsGood, Move, lhs, eax); |
2311 | append(denIsGood, convertToDoubleWord, eax, edx); |
2312 | append(denIsGood, div, eax, edx, rhs); |
2313 | append(denIsGood, sizeof(IntType) == 4 ? Move32 : Move, divResult, result); |
2314 | append(denIsGood, Jump); |
2315 | denIsGood->setSuccessors(continuation); |
2316 | |
2317 | m_currentBlock = continuation; |
2318 | return; |
2319 | } |
2320 | |
2321 | B3::Air::Opcode div = sizeof(IntType) == 4 ? X86UDiv32 : X86UDiv64; |
2322 | |
2323 | Tmp divResult = isDiv ? eax : edx; |
2324 | |
2325 | append(Move, lhs, eax); |
2326 | append(Xor64, edx, edx); |
2327 | append(div, eax, edx, rhs); |
2328 | append(sizeof(IntType) == 4 ? Move32 : Move, divResult, result); |
2329 | #else |
2330 | RELEASE_ASSERT_NOT_REACHED(); |
2331 | #endif |
2332 | } |
2333 | |
2334 | template<> |
2335 | auto AirIRGenerator::addOp<OpType::I32DivS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2336 | { |
2337 | emitChecksForModOrDiv<int32_t>(true, left, right); |
2338 | emitModOrDiv<int32_t>(true, left, right, result); |
2339 | return { }; |
2340 | } |
2341 | |
2342 | template<> |
2343 | auto AirIRGenerator::addOp<OpType::I32RemS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2344 | { |
2345 | emitChecksForModOrDiv<int32_t>(false, left, right); |
2346 | emitModOrDiv<int32_t>(false, left, right, result); |
2347 | return { }; |
2348 | } |
2349 | |
2350 | template<> |
2351 | auto AirIRGenerator::addOp<OpType::I32DivU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2352 | { |
2353 | emitChecksForModOrDiv<uint32_t>(false, left, right); |
2354 | emitModOrDiv<uint32_t>(true, left, right, result); |
2355 | return { }; |
2356 | } |
2357 | |
2358 | template<> |
2359 | auto AirIRGenerator::addOp<OpType::I32RemU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2360 | { |
2361 | emitChecksForModOrDiv<uint32_t>(false, left, right); |
2362 | emitModOrDiv<uint32_t>(false, left, right, result); |
2363 | return { }; |
2364 | } |
2365 | |
2366 | template<> |
2367 | auto AirIRGenerator::addOp<OpType::I64DivS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2368 | { |
2369 | emitChecksForModOrDiv<int64_t>(true, left, right); |
2370 | emitModOrDiv<int64_t>(true, left, right, result); |
2371 | return { }; |
2372 | } |
2373 | |
2374 | template<> |
2375 | auto AirIRGenerator::addOp<OpType::I64RemS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2376 | { |
2377 | emitChecksForModOrDiv<int64_t>(false, left, right); |
2378 | emitModOrDiv<int64_t>(false, left, right, result); |
2379 | return { }; |
2380 | } |
2381 | |
2382 | template<> |
2383 | auto AirIRGenerator::addOp<OpType::I64DivU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2384 | { |
2385 | emitChecksForModOrDiv<uint64_t>(false, left, right); |
2386 | emitModOrDiv<uint64_t>(true, left, right, result); |
2387 | return { }; |
2388 | } |
2389 | |
2390 | template<> |
2391 | auto AirIRGenerator::addOp<OpType::I64RemU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2392 | { |
2393 | emitChecksForModOrDiv<uint64_t>(false, left, right); |
2394 | emitModOrDiv<uint64_t>(false, left, right, result); |
2395 | return { }; |
2396 | } |
2397 | |
2398 | template<> |
2399 | auto AirIRGenerator::addOp<OpType::I32Ctz>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2400 | { |
2401 | auto* patchpoint = addPatchpoint(B3::Int32); |
2402 | patchpoint->effects = B3::Effects::none(); |
2403 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2404 | jit.countTrailingZeros32(params[1].gpr(), params[0].gpr()); |
2405 | }); |
2406 | result = g32(); |
2407 | emitPatchpoint(patchpoint, result, arg); |
2408 | return { }; |
2409 | } |
2410 | |
2411 | template<> |
2412 | auto AirIRGenerator::addOp<OpType::I64Ctz>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2413 | { |
2414 | auto* patchpoint = addPatchpoint(B3::Int64); |
2415 | patchpoint->effects = B3::Effects::none(); |
2416 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2417 | jit.countTrailingZeros64(params[1].gpr(), params[0].gpr()); |
2418 | }); |
2419 | result = g64(); |
2420 | emitPatchpoint(patchpoint, result, arg); |
2421 | return { }; |
2422 | } |
2423 | |
2424 | template<> |
2425 | auto AirIRGenerator::addOp<OpType::I32Popcnt>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2426 | { |
2427 | result = g32(); |
2428 | |
2429 | #if CPU(X86_64) |
2430 | if (MacroAssembler::supportsCountPopulation()) { |
2431 | auto* patchpoint = addPatchpoint(B3::Int32); |
2432 | patchpoint->effects = B3::Effects::none(); |
2433 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2434 | jit.countPopulation32(params[1].gpr(), params[0].gpr()); |
2435 | }); |
2436 | emitPatchpoint(patchpoint, result, arg); |
2437 | return { }; |
2438 | } |
2439 | #endif |
2440 | |
2441 | uint32_t (*popcount)(int32_t) = [] (int32_t value) -> uint32_t { return __builtin_popcount(value); }; |
2442 | emitCCall(popcount, result, arg); |
2443 | return { }; |
2444 | } |
2445 | |
2446 | template<> |
2447 | auto AirIRGenerator::addOp<OpType::I64Popcnt>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2448 | { |
2449 | result = g64(); |
2450 | |
2451 | #if CPU(X86_64) |
2452 | if (MacroAssembler::supportsCountPopulation()) { |
2453 | auto* patchpoint = addPatchpoint(B3::Int64); |
2454 | patchpoint->effects = B3::Effects::none(); |
2455 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2456 | jit.countPopulation64(params[1].gpr(), params[0].gpr()); |
2457 | }); |
2458 | emitPatchpoint(patchpoint, result, arg); |
2459 | return { }; |
2460 | } |
2461 | #endif |
2462 | |
2463 | uint64_t (*popcount)(int64_t) = [] (int64_t value) -> uint64_t { return __builtin_popcountll(value); }; |
2464 | emitCCall(popcount, result, arg); |
2465 | return { }; |
2466 | } |
2467 | |
2468 | template<> |
2469 | auto AirIRGenerator::addOp<F64ConvertUI64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2470 | { |
2471 | auto* patchpoint = addPatchpoint(B3::Double); |
2472 | patchpoint->effects = B3::Effects::none(); |
2473 | if (isX86()) |
2474 | patchpoint->numGPScratchRegisters = 1; |
2475 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2476 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2477 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2478 | #if CPU(X86_64) |
2479 | jit.convertUInt64ToDouble(params[1].gpr(), params[0].fpr(), params.gpScratch(0)); |
2480 | #else |
2481 | jit.convertUInt64ToDouble(params[1].gpr(), params[0].fpr()); |
2482 | #endif |
2483 | }); |
2484 | result = f64(); |
2485 | emitPatchpoint(patchpoint, result, arg); |
2486 | return { }; |
2487 | } |
2488 | |
2489 | template<> |
2490 | auto AirIRGenerator::addOp<OpType::F32ConvertUI64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2491 | { |
2492 | auto* patchpoint = addPatchpoint(B3::Float); |
2493 | patchpoint->effects = B3::Effects::none(); |
2494 | if (isX86()) |
2495 | patchpoint->numGPScratchRegisters = 1; |
2496 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2497 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2498 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2499 | #if CPU(X86_64) |
2500 | jit.convertUInt64ToFloat(params[1].gpr(), params[0].fpr(), params.gpScratch(0)); |
2501 | #else |
2502 | jit.convertUInt64ToFloat(params[1].gpr(), params[0].fpr()); |
2503 | #endif |
2504 | }); |
2505 | result = f32(); |
2506 | emitPatchpoint(patchpoint, result, arg); |
2507 | return { }; |
2508 | } |
2509 | |
2510 | template<> |
2511 | auto AirIRGenerator::addOp<OpType::F64Nearest>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2512 | { |
2513 | auto* patchpoint = addPatchpoint(B3::Double); |
2514 | patchpoint->effects = B3::Effects::none(); |
2515 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2516 | jit.roundTowardNearestIntDouble(params[1].fpr(), params[0].fpr()); |
2517 | }); |
2518 | result = f64(); |
2519 | emitPatchpoint(patchpoint, result, arg); |
2520 | return { }; |
2521 | } |
2522 | |
2523 | template<> |
2524 | auto AirIRGenerator::addOp<OpType::F32Nearest>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2525 | { |
2526 | auto* patchpoint = addPatchpoint(B3::Float); |
2527 | patchpoint->effects = B3::Effects::none(); |
2528 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2529 | jit.roundTowardNearestIntFloat(params[1].fpr(), params[0].fpr()); |
2530 | }); |
2531 | result = f32(); |
2532 | emitPatchpoint(patchpoint, result, arg); |
2533 | return { }; |
2534 | } |
2535 | |
2536 | template<> |
2537 | auto AirIRGenerator::addOp<OpType::F64Trunc>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2538 | { |
2539 | auto* patchpoint = addPatchpoint(B3::Double); |
2540 | patchpoint->effects = B3::Effects::none(); |
2541 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2542 | jit.roundTowardZeroDouble(params[1].fpr(), params[0].fpr()); |
2543 | }); |
2544 | result = f64(); |
2545 | emitPatchpoint(patchpoint, result, arg); |
2546 | return { }; |
2547 | } |
2548 | |
2549 | template<> |
2550 | auto AirIRGenerator::addOp<OpType::F32Trunc>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2551 | { |
2552 | auto* patchpoint = addPatchpoint(B3::Float); |
2553 | patchpoint->effects = B3::Effects::none(); |
2554 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2555 | jit.roundTowardZeroFloat(params[1].fpr(), params[0].fpr()); |
2556 | }); |
2557 | result = f32(); |
2558 | emitPatchpoint(patchpoint, result, arg); |
2559 | return { }; |
2560 | } |
2561 | |
2562 | template<> |
2563 | auto AirIRGenerator::addOp<OpType::I32TruncSF64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2564 | { |
2565 | auto max = addConstant(Type::F64, bitwise_cast<uint64_t>(-static_cast<double>(std::numeric_limits<int32_t>::min()))); |
2566 | auto min = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int32_t>::min()))); |
2567 | |
2568 | auto temp1 = g32(); |
2569 | auto temp2 = g32(); |
2570 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrUnordered), arg, min, temp1); |
2571 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2572 | append(Or32, temp1, temp2); |
2573 | |
2574 | emitCheck([&] { |
2575 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2576 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2577 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2578 | }); |
2579 | |
2580 | auto* patchpoint = addPatchpoint(B3::Int32); |
2581 | patchpoint->effects = B3::Effects::none(); |
2582 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2583 | jit.truncateDoubleToInt32(params[1].fpr(), params[0].gpr()); |
2584 | }); |
2585 | result = g32(); |
2586 | emitPatchpoint(patchpoint, result, arg); |
2587 | |
2588 | return { }; |
2589 | } |
2590 | |
2591 | template<> |
2592 | auto AirIRGenerator::addOp<OpType::I32TruncSF32>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2593 | { |
2594 | auto max = addConstant(Type::F32, bitwise_cast<uint32_t>(-static_cast<float>(std::numeric_limits<int32_t>::min()))); |
2595 | auto min = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int32_t>::min()))); |
2596 | |
2597 | auto temp1 = g32(); |
2598 | auto temp2 = g32(); |
2599 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrUnordered), arg, min, temp1); |
2600 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2601 | append(Or32, temp1, temp2); |
2602 | |
2603 | emitCheck([&] { |
2604 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2605 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2606 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2607 | }); |
2608 | |
2609 | auto* patchpoint = addPatchpoint(B3::Int32); |
2610 | patchpoint->effects = B3::Effects::none(); |
2611 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2612 | jit.truncateFloatToInt32(params[1].fpr(), params[0].gpr()); |
2613 | }); |
2614 | result = g32(); |
2615 | emitPatchpoint(patchpoint, result, arg); |
2616 | return { }; |
2617 | } |
2618 | |
2619 | |
2620 | template<> |
2621 | auto AirIRGenerator::addOp<OpType::I32TruncUF64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2622 | { |
2623 | auto max = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int32_t>::min()) * -2.0)); |
2624 | auto min = addConstant(Type::F64, bitwise_cast<uint64_t>(-1.0)); |
2625 | |
2626 | auto temp1 = g32(); |
2627 | auto temp2 = g32(); |
2628 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqualOrUnordered), arg, min, temp1); |
2629 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2630 | append(Or32, temp1, temp2); |
2631 | |
2632 | emitCheck([&] { |
2633 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2634 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2635 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2636 | }); |
2637 | |
2638 | auto* patchpoint = addPatchpoint(B3::Int32); |
2639 | patchpoint->effects = B3::Effects::none(); |
2640 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2641 | jit.truncateDoubleToUint32(params[1].fpr(), params[0].gpr()); |
2642 | }); |
2643 | result = g32(); |
2644 | emitPatchpoint(patchpoint, result, arg); |
2645 | return { }; |
2646 | } |
2647 | |
2648 | template<> |
2649 | auto AirIRGenerator::addOp<OpType::I32TruncUF32>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2650 | { |
2651 | auto max = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int32_t>::min()) * static_cast<float>(-2.0))); |
2652 | auto min = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(-1.0))); |
2653 | |
2654 | auto temp1 = g32(); |
2655 | auto temp2 = g32(); |
2656 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqualOrUnordered), arg, min, temp1); |
2657 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2658 | append(Or32, temp1, temp2); |
2659 | |
2660 | emitCheck([&] { |
2661 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2662 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2663 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2664 | }); |
2665 | |
2666 | auto* patchpoint = addPatchpoint(B3::Int32); |
2667 | patchpoint->effects = B3::Effects::none(); |
2668 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2669 | jit.truncateFloatToUint32(params[1].fpr(), params[0].gpr()); |
2670 | }); |
2671 | result = g32(); |
2672 | emitPatchpoint(patchpoint, result, arg); |
2673 | return { }; |
2674 | } |
2675 | |
2676 | template<> |
2677 | auto AirIRGenerator::addOp<OpType::I64TruncSF64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2678 | { |
2679 | auto max = addConstant(Type::F64, bitwise_cast<uint64_t>(-static_cast<double>(std::numeric_limits<int64_t>::min()))); |
2680 | auto min = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int64_t>::min()))); |
2681 | |
2682 | auto temp1 = g32(); |
2683 | auto temp2 = g32(); |
2684 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrUnordered), arg, min, temp1); |
2685 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2686 | append(Or32, temp1, temp2); |
2687 | |
2688 | emitCheck([&] { |
2689 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2690 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2691 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2692 | }); |
2693 | |
2694 | auto* patchpoint = addPatchpoint(B3::Int64); |
2695 | patchpoint->effects = B3::Effects::none(); |
2696 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2697 | jit.truncateDoubleToInt64(params[1].fpr(), params[0].gpr()); |
2698 | }); |
2699 | |
2700 | result = g64(); |
2701 | emitPatchpoint(patchpoint, result, arg); |
2702 | return { }; |
2703 | } |
2704 | |
2705 | template<> |
2706 | auto AirIRGenerator::addOp<OpType::I64TruncUF64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2707 | { |
2708 | auto max = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int64_t>::min()) * -2.0)); |
2709 | auto min = addConstant(Type::F64, bitwise_cast<uint64_t>(-1.0)); |
2710 | |
2711 | auto temp1 = g32(); |
2712 | auto temp2 = g32(); |
2713 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqualOrUnordered), arg, min, temp1); |
2714 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2715 | append(Or32, temp1, temp2); |
2716 | |
2717 | emitCheck([&] { |
2718 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2719 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2720 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2721 | }); |
2722 | |
2723 | TypedTmp signBitConstant; |
2724 | if (isX86()) |
2725 | signBitConstant = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<uint64_t>::max() - std::numeric_limits<int64_t>::max()))); |
2726 | |
2727 | Vector<ConstrainedTmp> args; |
2728 | auto* patchpoint = addPatchpoint(B3::Int64); |
2729 | patchpoint->effects = B3::Effects::none(); |
2730 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2731 | args.append(arg); |
2732 | if (isX86()) { |
2733 | args.append(signBitConstant); |
2734 | patchpoint->numFPScratchRegisters = 1; |
2735 | } |
2736 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2737 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2738 | FPRReg scratch = InvalidFPRReg; |
2739 | FPRReg constant = InvalidFPRReg; |
2740 | if (isX86()) { |
2741 | scratch = params.fpScratch(0); |
2742 | constant = params[2].fpr(); |
2743 | } |
2744 | jit.truncateDoubleToUint64(params[1].fpr(), params[0].gpr(), scratch, constant); |
2745 | }); |
2746 | |
2747 | result = g64(); |
2748 | emitPatchpoint(m_currentBlock, patchpoint, result, WTFMove(args)); |
2749 | return { }; |
2750 | } |
2751 | |
2752 | template<> |
2753 | auto AirIRGenerator::addOp<OpType::I64TruncSF32>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2754 | { |
2755 | auto max = addConstant(Type::F32, bitwise_cast<uint32_t>(-static_cast<float>(std::numeric_limits<int64_t>::min()))); |
2756 | auto min = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int64_t>::min()))); |
2757 | |
2758 | auto temp1 = g32(); |
2759 | auto temp2 = g32(); |
2760 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrUnordered), arg, min, temp1); |
2761 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2762 | append(Or32, temp1, temp2); |
2763 | |
2764 | emitCheck([&] { |
2765 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2766 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2767 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2768 | }); |
2769 | |
2770 | auto* patchpoint = addPatchpoint(B3::Int64); |
2771 | patchpoint->effects = B3::Effects::none(); |
2772 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2773 | jit.truncateFloatToInt64(params[1].fpr(), params[0].gpr()); |
2774 | }); |
2775 | result = g64(); |
2776 | emitPatchpoint(patchpoint, result, arg); |
2777 | return { }; |
2778 | } |
2779 | |
2780 | template<> |
2781 | auto AirIRGenerator::addOp<OpType::I64TruncUF32>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2782 | { |
2783 | auto max = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int64_t>::min()) * static_cast<float>(-2.0))); |
2784 | auto min = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(-1.0))); |
2785 | |
2786 | auto temp1 = g32(); |
2787 | auto temp2 = g32(); |
2788 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqualOrUnordered), arg, min, temp1); |
2789 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2790 | append(Or32, temp1, temp2); |
2791 | |
2792 | emitCheck([&] { |
2793 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2794 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2795 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2796 | }); |
2797 | |
2798 | TypedTmp signBitConstant; |
2799 | if (isX86()) |
2800 | signBitConstant = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<uint64_t>::max() - std::numeric_limits<int64_t>::max()))); |
2801 | |
2802 | auto* patchpoint = addPatchpoint(B3::Int64); |
2803 | patchpoint->effects = B3::Effects::none(); |
2804 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2805 | Vector<ConstrainedTmp> args; |
2806 | args.append(arg); |
2807 | if (isX86()) { |
2808 | args.append(signBitConstant); |
2809 | patchpoint->numFPScratchRegisters = 1; |
2810 | } |
2811 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2812 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2813 | FPRReg scratch = InvalidFPRReg; |
2814 | FPRReg constant = InvalidFPRReg; |
2815 | if (isX86()) { |
2816 | scratch = params.fpScratch(0); |
2817 | constant = params[2].fpr(); |
2818 | } |
2819 | jit.truncateFloatToUint64(params[1].fpr(), params[0].gpr(), scratch, constant); |
2820 | }); |
2821 | |
2822 | result = g64(); |
2823 | emitPatchpoint(m_currentBlock, patchpoint, result, WTFMove(args)); |
2824 | |
2825 | return { }; |
2826 | } |
2827 | |
2828 | auto AirIRGenerator::addShift(Type type, B3::Air::Opcode op, ExpressionType value, ExpressionType shift, ExpressionType& result) -> PartialResult |
2829 | { |
2830 | ASSERT(type == Type::I64 || type == Type::I32); |
2831 | result = tmpForType(type); |
2832 | |
2833 | if (isValidForm(op, Arg::Tmp, Arg::Tmp, Arg::Tmp)) { |
2834 | append(op, value, shift, result); |
2835 | return { }; |
2836 | } |
2837 | |
2838 | #if CPU(X86_64) |
2839 | Tmp ecx = Tmp(X86Registers::ecx); |
2840 | append(Move, value, result); |
2841 | append(Move, shift, ecx); |
2842 | append(op, ecx, result); |
2843 | #else |
2844 | RELEASE_ASSERT_NOT_REACHED(); |
2845 | #endif |
2846 | return { }; |
2847 | } |
2848 | |
2849 | auto AirIRGenerator::addIntegerSub(B3::Air::Opcode op, ExpressionType lhs, ExpressionType rhs, ExpressionType& result) -> PartialResult |
2850 | { |
2851 | ASSERT(op == Sub32 || op == Sub64); |
2852 | |
2853 | result = op == Sub32 ? g32() : g64(); |
2854 | |
2855 | if (isValidForm(op, Arg::Tmp, Arg::Tmp, Arg::Tmp)) { |
2856 | append(op, lhs, rhs, result); |
2857 | return { }; |
2858 | } |
2859 | |
2860 | RELEASE_ASSERT(isX86()); |
2861 | // Sub a, b |
2862 | // means |
2863 | // b = b Sub a |
2864 | append(Move, lhs, result); |
2865 | append(op, rhs, result); |
2866 | return { }; |
2867 | } |
2868 | |
2869 | auto AirIRGenerator::addFloatingPointAbs(B3::Air::Opcode op, ExpressionType value, ExpressionType& result) -> PartialResult |
2870 | { |
2871 | RELEASE_ASSERT(op == AbsFloat || op == AbsDouble); |
2872 | |
2873 | result = op == AbsFloat ? f32() : f64(); |
2874 | |
2875 | if (isValidForm(op, Arg::Tmp, Arg::Tmp)) { |
2876 | append(op, value, result); |
2877 | return { }; |
2878 | } |
2879 | |
2880 | RELEASE_ASSERT(isX86()); |
2881 | |
2882 | if (op == AbsFloat) { |
2883 | auto constant = g32(); |
2884 | append(Move, Arg::imm(static_cast<uint32_t>(~(1ull << 31))), constant); |
2885 | append(Move32ToFloat, constant, result); |
2886 | append(AndFloat, value, result); |
2887 | } else { |
2888 | auto constant = g64(); |
2889 | append(Move, Arg::bigImm(~(1ull << 63)), constant); |
2890 | append(Move64ToDouble, constant, result); |
2891 | append(AndDouble, value, result); |
2892 | } |
2893 | return { }; |
2894 | } |
2895 | |
2896 | auto AirIRGenerator::addFloatingPointBinOp(Type type, B3::Air::Opcode op, ExpressionType lhs, ExpressionType rhs, ExpressionType& result) -> PartialResult |
2897 | { |
2898 | ASSERT(type == Type::F32 || type == Type::F64); |
2899 | result = tmpForType(type); |
2900 | |
2901 | if (isValidForm(op, Arg::Tmp, Arg::Tmp, Arg::Tmp)) { |
2902 | append(op, lhs, rhs, result); |
2903 | return { }; |
2904 | } |
2905 | |
2906 | RELEASE_ASSERT(isX86()); |
2907 | |
2908 | // Op a, b |
2909 | // means |
2910 | // b = b Op a |
2911 | append(moveOpForValueType(type), lhs, result); |
2912 | append(op, rhs, result); |
2913 | return { }; |
2914 | } |
2915 | |
2916 | template<> auto AirIRGenerator::addOp<OpType::F32Ceil>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
2917 | { |
2918 | result = f32(); |
2919 | append(CeilFloat, arg0, result); |
2920 | return { }; |
2921 | } |
2922 | |
2923 | template<> auto AirIRGenerator::addOp<OpType::I32Mul>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
2924 | { |
2925 | result = g32(); |
2926 | append(Mul32, arg0, arg1, result); |
2927 | return { }; |
2928 | } |
2929 | |
2930 | template<> auto AirIRGenerator::addOp<OpType::I32Sub>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
2931 | { |
2932 | return addIntegerSub(Sub32, arg0, arg1, result); |
2933 | } |
2934 | |
2935 | template<> auto AirIRGenerator::addOp<OpType::F64Le>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
2936 | { |
2937 | result = g32(); |
2938 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqual), arg0, arg1, result); |
2939 | return { }; |
2940 | } |
2941 | |
2942 | template<> auto AirIRGenerator::addOp<OpType::F32DemoteF64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
2943 | { |
2944 | result = f32(); |
2945 | append(ConvertDoubleToFloat, arg0, result); |
2946 | return { }; |
2947 | } |
2948 | |
2949 | template<> auto AirIRGenerator::addOp<OpType::F32Min>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
2950 | { |
2951 | return addFloatingPointMinOrMax(F32, MinOrMax::Min, arg0, arg1, result); |
2952 | } |
2953 | |
2954 | template<> auto AirIRGenerator::addOp<OpType::F64Ne>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
2955 | { |
2956 | result = g32(); |
2957 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleNotEqualOrUnordered), arg0, arg1, result); |
2958 | return { }; |
2959 | } |
2960 | |
2961 | template<> auto AirIRGenerator::addOp<OpType::F64Lt>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
2962 | { |
2963 | result = g32(); |
2964 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThan), arg0, arg1, result); |
2965 | return { }; |
2966 | } |
2967 | |
2968 | auto AirIRGenerator::addFloatingPointMinOrMax(Type floatType, MinOrMax minOrMax, ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
2969 | { |
2970 | ASSERT(floatType == F32 || floatType == F64); |
2971 | result = tmpForType(floatType); |
2972 | |
2973 | BasicBlock* isEqual = m_code.addBlock(); |
2974 | BasicBlock* notEqual = m_code.addBlock(); |
2975 | BasicBlock* isLessThan = m_code.addBlock(); |
2976 | BasicBlock* notLessThan = m_code.addBlock(); |
2977 | BasicBlock* isGreaterThan = m_code.addBlock(); |
2978 | BasicBlock* isNaN = m_code.addBlock(); |
2979 | BasicBlock* continuation = m_code.addBlock(); |
2980 | |
2981 | auto branchOp = floatType == F32 ? BranchFloat : BranchDouble; |
2982 | append(m_currentBlock, branchOp, Arg::doubleCond(MacroAssembler::DoubleEqual), arg0, arg1); |
2983 | m_currentBlock->setSuccessors(isEqual, notEqual); |
2984 | |
2985 | append(notEqual, branchOp, Arg::doubleCond(MacroAssembler::DoubleLessThan), arg0, arg1); |
2986 | notEqual->setSuccessors(isLessThan, notLessThan); |
2987 | |
2988 | append(notLessThan, branchOp, Arg::doubleCond(MacroAssembler::DoubleGreaterThan), arg0, arg1); |
2989 | notLessThan->setSuccessors(isGreaterThan, isNaN); |
2990 | |
2991 | auto andOp = floatType == F32 ? AndFloat : AndDouble; |
2992 | auto orOp = floatType == F32 ? OrFloat : OrDouble; |
2993 | append(isEqual, minOrMax == MinOrMax::Max ? andOp : orOp, arg0, arg1, result); |
2994 | append(isEqual, Jump); |
2995 | isEqual->setSuccessors(continuation); |
2996 | |
2997 | auto isLessThanResult = minOrMax == MinOrMax::Max ? arg1 : arg0; |
2998 | append(isLessThan, moveOpForValueType(floatType), isLessThanResult, result); |
2999 | append(isLessThan, Jump); |
3000 | isLessThan->setSuccessors(continuation); |
3001 | |
3002 | auto isGreaterThanResult = minOrMax == MinOrMax::Max ? arg0 : arg1; |
3003 | append(isGreaterThan, moveOpForValueType(floatType), isGreaterThanResult, result); |
3004 | append(isGreaterThan, Jump); |
3005 | isGreaterThan->setSuccessors(continuation); |
3006 | |
3007 | auto addOp = floatType == F32 ? AddFloat : AddDouble; |
3008 | append(isNaN, addOp, arg0, arg1, result); |
3009 | append(isNaN, Jump); |
3010 | isNaN->setSuccessors(continuation); |
3011 | |
3012 | m_currentBlock = continuation; |
3013 | |
3014 | return { }; |
3015 | } |
3016 | |
3017 | template<> auto AirIRGenerator::addOp<OpType::F32Max>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3018 | { |
3019 | return addFloatingPointMinOrMax(F32, MinOrMax::Max, arg0, arg1, result); |
3020 | } |
3021 | |
3022 | template<> auto AirIRGenerator::addOp<OpType::F64Mul>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3023 | { |
3024 | return addFloatingPointBinOp(Type::F64, MulDouble, arg0, arg1, result); |
3025 | } |
3026 | |
3027 | template<> auto AirIRGenerator::addOp<OpType::F32Div>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3028 | { |
3029 | return addFloatingPointBinOp(Type::F32, DivFloat, arg0, arg1, result); |
3030 | } |
3031 | |
3032 | template<> auto AirIRGenerator::addOp<OpType::I32Clz>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3033 | { |
3034 | result = g32(); |
3035 | append(CountLeadingZeros32, arg0, result); |
3036 | return { }; |
3037 | } |
3038 | |
3039 | template<> auto AirIRGenerator::addOp<OpType::F32Copysign>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3040 | { |
3041 | // FIXME: We can have better codegen here for the imms and two operand forms on x86 |
3042 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
3043 | result = f32(); |
3044 | auto temp1 = g32(); |
3045 | auto sign = g32(); |
3046 | auto value = g32(); |
3047 | |
3048 | // FIXME: Try to use Imm where possible: |
3049 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
3050 | append(MoveFloatTo32, arg1, temp1); |
3051 | append(Move, Arg::bigImm(0x80000000), sign); |
3052 | append(And32, temp1, sign, sign); |
3053 | |
3054 | append(MoveDoubleTo64, arg0, temp1); |
3055 | append(Move, Arg::bigImm(0x7fffffff), value); |
3056 | append(And32, temp1, value, value); |
3057 | |
3058 | append(Or32, sign, value, value); |
3059 | append(Move32ToFloat, value, result); |
3060 | |
3061 | return { }; |
3062 | } |
3063 | |
3064 | template<> auto AirIRGenerator::addOp<OpType::F64ConvertUI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3065 | { |
3066 | result = f64(); |
3067 | auto temp = g64(); |
3068 | append(Move32, arg0, temp); |
3069 | append(ConvertInt64ToDouble, temp, result); |
3070 | return { }; |
3071 | } |
3072 | |
3073 | template<> auto AirIRGenerator::addOp<OpType::F32ReinterpretI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3074 | { |
3075 | result = f32(); |
3076 | append(Move32ToFloat, arg0, result); |
3077 | return { }; |
3078 | } |
3079 | |
3080 | template<> auto AirIRGenerator::addOp<OpType::I64And>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3081 | { |
3082 | result = g64(); |
3083 | append(And64, arg0, arg1, result); |
3084 | return { }; |
3085 | } |
3086 | |
3087 | template<> auto AirIRGenerator::addOp<OpType::F32Ne>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3088 | { |
3089 | result = g32(); |
3090 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleNotEqualOrUnordered), arg0, arg1, result); |
3091 | return { }; |
3092 | } |
3093 | |
3094 | template<> auto AirIRGenerator::addOp<OpType::F64Gt>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3095 | { |
3096 | result = g32(); |
3097 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThan), arg0, arg1, result); |
3098 | return { }; |
3099 | } |
3100 | |
3101 | template<> auto AirIRGenerator::addOp<OpType::F32Sqrt>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3102 | { |
3103 | result = f32(); |
3104 | append(SqrtFloat, arg0, result); |
3105 | return { }; |
3106 | } |
3107 | |
3108 | template<> auto AirIRGenerator::addOp<OpType::F64Ge>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3109 | { |
3110 | result = g32(); |
3111 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqual), arg0, arg1, result); |
3112 | return { }; |
3113 | } |
3114 | |
3115 | template<> auto AirIRGenerator::addOp<OpType::I64GtS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3116 | { |
3117 | result = g32(); |
3118 | append(Compare64, Arg::relCond(MacroAssembler::GreaterThan), arg0, arg1, result); |
3119 | return { }; |
3120 | } |
3121 | |
3122 | template<> auto AirIRGenerator::addOp<OpType::I64GtU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3123 | { |
3124 | result = g32(); |
3125 | append(Compare64, Arg::relCond(MacroAssembler::Above), arg0, arg1, result); |
3126 | return { }; |
3127 | } |
3128 | |
3129 | template<> auto AirIRGenerator::addOp<OpType::I64Eqz>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3130 | { |
3131 | result = g32(); |
3132 | append(Test64, Arg::resCond(MacroAssembler::Zero), arg0, arg0, result); |
3133 | return { }; |
3134 | } |
3135 | |
3136 | template<> auto AirIRGenerator::addOp<OpType::F64Div>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3137 | { |
3138 | return addFloatingPointBinOp(Type::F64, DivDouble, arg0, arg1, result); |
3139 | } |
3140 | |
3141 | template<> auto AirIRGenerator::addOp<OpType::F32Add>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3142 | { |
3143 | result = f32(); |
3144 | append(AddFloat, arg0, arg1, result); |
3145 | return { }; |
3146 | } |
3147 | |
3148 | template<> auto AirIRGenerator::addOp<OpType::I64Or>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3149 | { |
3150 | result = g64(); |
3151 | append(Or64, arg0, arg1, result); |
3152 | return { }; |
3153 | } |
3154 | |
3155 | template<> auto AirIRGenerator::addOp<OpType::I32LeU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3156 | { |
3157 | result = g32(); |
3158 | append(Compare32, Arg::relCond(MacroAssembler::BelowOrEqual), arg0, arg1, result); |
3159 | return { }; |
3160 | } |
3161 | |
3162 | template<> auto AirIRGenerator::addOp<OpType::I32LeS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3163 | { |
3164 | result = g32(); |
3165 | append(Compare32, Arg::relCond(MacroAssembler::LessThanOrEqual), arg0, arg1, result); |
3166 | return { }; |
3167 | } |
3168 | |
3169 | template<> auto AirIRGenerator::addOp<OpType::I64Ne>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3170 | { |
3171 | result = g32(); |
3172 | append(Compare64, Arg::relCond(MacroAssembler::NotEqual), arg0, arg1, result); |
3173 | return { }; |
3174 | } |
3175 | |
3176 | template<> auto AirIRGenerator::addOp<OpType::I64Clz>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3177 | { |
3178 | result = g64(); |
3179 | append(CountLeadingZeros64, arg0, result); |
3180 | return { }; |
3181 | } |
3182 | |
3183 | template<> auto AirIRGenerator::addOp<OpType::F32Neg>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3184 | { |
3185 | result = f32(); |
3186 | if (isValidForm(NegateFloat, Arg::Tmp, Arg::Tmp)) |
3187 | append(NegateFloat, arg0, result); |
3188 | else { |
3189 | auto constant = addConstant(Type::I32, bitwise_cast<uint32_t>(static_cast<float>(-0.0))); |
3190 | auto temp = g32(); |
3191 | append(MoveFloatTo32, arg0, temp); |
3192 | append(Xor32, constant, temp); |
3193 | append(Move32ToFloat, temp, result); |
3194 | } |
3195 | return { }; |
3196 | } |
3197 | |
3198 | template<> auto AirIRGenerator::addOp<OpType::I32And>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3199 | { |
3200 | result = g32(); |
3201 | append(And32, arg0, arg1, result); |
3202 | return { }; |
3203 | } |
3204 | |
3205 | template<> auto AirIRGenerator::addOp<OpType::I32LtU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3206 | { |
3207 | result = g32(); |
3208 | append(Compare32, Arg::relCond(MacroAssembler::Below), arg0, arg1, result); |
3209 | return { }; |
3210 | } |
3211 | |
3212 | template<> auto AirIRGenerator::addOp<OpType::I64Rotr>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3213 | { |
3214 | return addShift(Type::I64, RotateRight64, arg0, arg1, result); |
3215 | } |
3216 | |
3217 | template<> auto AirIRGenerator::addOp<OpType::F64Abs>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3218 | { |
3219 | return addFloatingPointAbs(AbsDouble, arg0, result); |
3220 | } |
3221 | |
3222 | template<> auto AirIRGenerator::addOp<OpType::I32LtS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3223 | { |
3224 | result = g32(); |
3225 | append(Compare32, Arg::relCond(MacroAssembler::LessThan), arg0, arg1, result); |
3226 | return { }; |
3227 | } |
3228 | |
3229 | template<> auto AirIRGenerator::addOp<OpType::I32Eq>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3230 | { |
3231 | result = g32(); |
3232 | append(Compare32, Arg::relCond(MacroAssembler::Equal), arg0, arg1, result); |
3233 | return { }; |
3234 | } |
3235 | |
3236 | template<> auto AirIRGenerator::addOp<OpType::F64Copysign>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3237 | { |
3238 | // FIXME: We can have better codegen here for the imms and two operand forms on x86 |
3239 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
3240 | result = f64(); |
3241 | auto temp1 = g64(); |
3242 | auto sign = g64(); |
3243 | auto value = g64(); |
3244 | |
3245 | append(MoveDoubleTo64, arg1, temp1); |
3246 | append(Move, Arg::bigImm(0x8000000000000000), sign); |
3247 | append(And64, temp1, sign, sign); |
3248 | |
3249 | append(MoveDoubleTo64, arg0, temp1); |
3250 | append(Move, Arg::bigImm(0x7fffffffffffffff), value); |
3251 | append(And64, temp1, value, value); |
3252 | |
3253 | append(Or64, sign, value, value); |
3254 | append(Move64ToDouble, value, result); |
3255 | |
3256 | return { }; |
3257 | } |
3258 | |
3259 | template<> auto AirIRGenerator::addOp<OpType::F32ConvertSI64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3260 | { |
3261 | result = f32(); |
3262 | append(ConvertInt64ToFloat, arg0, result); |
3263 | return { }; |
3264 | } |
3265 | |
3266 | template<> auto AirIRGenerator::addOp<OpType::I64Rotl>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3267 | { |
3268 | if (isARM64()) { |
3269 | // ARM64 doesn't have a rotate left. |
3270 | auto newShift = g64(); |
3271 | append(Move, arg1, newShift); |
3272 | append(Neg64, newShift); |
3273 | return addShift(Type::I64, RotateRight64, arg0, newShift, result); |
3274 | } else |
3275 | return addShift(Type::I64, RotateLeft64, arg0, arg1, result); |
3276 | } |
3277 | |
3278 | template<> auto AirIRGenerator::addOp<OpType::F32Lt>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3279 | { |
3280 | result = g32(); |
3281 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThan), arg0, arg1, result); |
3282 | return { }; |
3283 | } |
3284 | |
3285 | template<> auto AirIRGenerator::addOp<OpType::F64ConvertSI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3286 | { |
3287 | result = f64(); |
3288 | append(ConvertInt32ToDouble, arg0, result); |
3289 | return { }; |
3290 | } |
3291 | |
3292 | template<> auto AirIRGenerator::addOp<OpType::F64Eq>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3293 | { |
3294 | result = g32(); |
3295 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleEqual), arg0, arg1, result); |
3296 | return { }; |
3297 | } |
3298 | |
3299 | template<> auto AirIRGenerator::addOp<OpType::F32Le>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3300 | { |
3301 | result = g32(); |
3302 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqual), arg0, arg1, result); |
3303 | return { }; |
3304 | } |
3305 | |
3306 | template<> auto AirIRGenerator::addOp<OpType::F32Ge>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3307 | { |
3308 | result = g32(); |
3309 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqual), arg0, arg1, result); |
3310 | return { }; |
3311 | } |
3312 | |
3313 | template<> auto AirIRGenerator::addOp<OpType::I32ShrU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3314 | { |
3315 | return addShift(Type::I32, Urshift32, arg0, arg1, result); |
3316 | } |
3317 | |
3318 | template<> auto AirIRGenerator::addOp<OpType::F32ConvertUI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3319 | { |
3320 | result = f32(); |
3321 | auto temp = g64(); |
3322 | append(Move32, arg0, temp); |
3323 | append(ConvertInt64ToFloat, temp, result); |
3324 | return { }; |
3325 | } |
3326 | |
3327 | template<> auto AirIRGenerator::addOp<OpType::I32ShrS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3328 | { |
3329 | return addShift(Type::I32, Rshift32, arg0, arg1, result); |
3330 | } |
3331 | |
3332 | template<> auto AirIRGenerator::addOp<OpType::I32GeU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3333 | { |
3334 | result = g32(); |
3335 | append(Compare32, Arg::relCond(MacroAssembler::AboveOrEqual), arg0, arg1, result); |
3336 | return { }; |
3337 | } |
3338 | |
3339 | template<> auto AirIRGenerator::addOp<OpType::F64Ceil>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3340 | { |
3341 | result = f64(); |
3342 | append(CeilDouble, arg0, result); |
3343 | return { }; |
3344 | } |
3345 | |
3346 | template<> auto AirIRGenerator::addOp<OpType::I32GeS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3347 | { |
3348 | result = g32(); |
3349 | append(Compare32, Arg::relCond(MacroAssembler::GreaterThanOrEqual), arg0, arg1, result); |
3350 | return { }; |
3351 | } |
3352 | |
3353 | template<> auto AirIRGenerator::addOp<OpType::I32Shl>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3354 | { |
3355 | return addShift(Type::I32, Lshift32, arg0, arg1, result); |
3356 | } |
3357 | |
3358 | template<> auto AirIRGenerator::addOp<OpType::F64Floor>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3359 | { |
3360 | result = f64(); |
3361 | append(FloorDouble, arg0, result); |
3362 | return { }; |
3363 | } |
3364 | |
3365 | template<> auto AirIRGenerator::addOp<OpType::I32Xor>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3366 | { |
3367 | result = g32(); |
3368 | append(Xor32, arg0, arg1, result); |
3369 | return { }; |
3370 | } |
3371 | |
3372 | template<> auto AirIRGenerator::addOp<OpType::F32Abs>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3373 | { |
3374 | return addFloatingPointAbs(AbsFloat, arg0, result); |
3375 | } |
3376 | |
3377 | template<> auto AirIRGenerator::addOp<OpType::F64Min>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3378 | { |
3379 | return addFloatingPointMinOrMax(F64, MinOrMax::Min, arg0, arg1, result); |
3380 | } |
3381 | |
3382 | template<> auto AirIRGenerator::addOp<OpType::F32Mul>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3383 | { |
3384 | result = f32(); |
3385 | append(MulFloat, arg0, arg1, result); |
3386 | return { }; |
3387 | } |
3388 | |
3389 | template<> auto AirIRGenerator::addOp<OpType::I64Sub>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3390 | { |
3391 | return addIntegerSub(Sub64, arg0, arg1, result); |
3392 | } |
3393 | |
3394 | template<> auto AirIRGenerator::addOp<OpType::I32ReinterpretF32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3395 | { |
3396 | result = g32(); |
3397 | append(MoveFloatTo32, arg0, result); |
3398 | return { }; |
3399 | } |
3400 | |
3401 | template<> auto AirIRGenerator::addOp<OpType::I32Add>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3402 | { |
3403 | result = g32(); |
3404 | append(Add32, arg0, arg1, result); |
3405 | return { }; |
3406 | } |
3407 | |
3408 | template<> auto AirIRGenerator::addOp<OpType::F64Sub>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3409 | { |
3410 | return addFloatingPointBinOp(Type::F64, SubDouble, arg0, arg1, result); |
3411 | } |
3412 | |
3413 | template<> auto AirIRGenerator::addOp<OpType::I32Or>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3414 | { |
3415 | result = g32(); |
3416 | append(Or32, arg0, arg1, result); |
3417 | return { }; |
3418 | } |
3419 | |
3420 | template<> auto AirIRGenerator::addOp<OpType::I64LtU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3421 | { |
3422 | result = g32(); |
3423 | append(Compare64, Arg::relCond(MacroAssembler::Below), arg0, arg1, result); |
3424 | return { }; |
3425 | } |
3426 | |
3427 | template<> auto AirIRGenerator::addOp<OpType::I64LtS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3428 | { |
3429 | result = g32(); |
3430 | append(Compare64, Arg::relCond(MacroAssembler::LessThan), arg0, arg1, result); |
3431 | return { }; |
3432 | } |
3433 | |
3434 | template<> auto AirIRGenerator::addOp<OpType::F64ConvertSI64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3435 | { |
3436 | result = f64(); |
3437 | append(ConvertInt64ToDouble, arg0, result); |
3438 | return { }; |
3439 | } |
3440 | |
3441 | template<> auto AirIRGenerator::addOp<OpType::I64Xor>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3442 | { |
3443 | result = g64(); |
3444 | append(Xor64, arg0, arg1, result); |
3445 | return { }; |
3446 | } |
3447 | |
3448 | template<> auto AirIRGenerator::addOp<OpType::I64GeU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3449 | { |
3450 | result = g32(); |
3451 | append(Compare64, Arg::relCond(MacroAssembler::AboveOrEqual), arg0, arg1, result); |
3452 | return { }; |
3453 | } |
3454 | |
3455 | template<> auto AirIRGenerator::addOp<OpType::I64Mul>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3456 | { |
3457 | result = g64(); |
3458 | append(Mul64, arg0, arg1, result); |
3459 | return { }; |
3460 | } |
3461 | |
3462 | template<> auto AirIRGenerator::addOp<OpType::F32Sub>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3463 | { |
3464 | result = f32(); |
3465 | if (isValidForm(SubFloat, Arg::Tmp, Arg::Tmp, Arg::Tmp)) |
3466 | append(SubFloat, arg0, arg1, result); |
3467 | else { |
3468 | RELEASE_ASSERT(isX86()); |
3469 | append(MoveFloat, arg0, result); |
3470 | append(SubFloat, arg1, result); |
3471 | } |
3472 | return { }; |
3473 | } |
3474 | |
3475 | template<> auto AirIRGenerator::addOp<OpType::F64PromoteF32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3476 | { |
3477 | result = f64(); |
3478 | append(ConvertFloatToDouble, arg0, result); |
3479 | return { }; |
3480 | } |
3481 | |
3482 | template<> auto AirIRGenerator::addOp<OpType::F64Add>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3483 | { |
3484 | result = f64(); |
3485 | append(AddDouble, arg0, arg1, result); |
3486 | return { }; |
3487 | } |
3488 | |
3489 | template<> auto AirIRGenerator::addOp<OpType::I64GeS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3490 | { |
3491 | result = g32(); |
3492 | append(Compare64, Arg::relCond(MacroAssembler::GreaterThanOrEqual), arg0, arg1, result); |
3493 | return { }; |
3494 | } |
3495 | |
3496 | template<> auto AirIRGenerator::addOp<OpType::I64ExtendUI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3497 | { |
3498 | result = g64(); |
3499 | append(Move32, arg0, result); |
3500 | return { }; |
3501 | } |
3502 | |
3503 | template<> auto AirIRGenerator::addOp<OpType::I32Ne>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3504 | { |
3505 | result = g32(); |
3506 | RELEASE_ASSERT(arg0 && arg1); |
3507 | append(Compare32, Arg::relCond(MacroAssembler::NotEqual), arg0, arg1, result); |
3508 | return { }; |
3509 | } |
3510 | |
3511 | template<> auto AirIRGenerator::addOp<OpType::F64ReinterpretI64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3512 | { |
3513 | result = f64(); |
3514 | append(Move64ToDouble, arg0, result); |
3515 | return { }; |
3516 | } |
3517 | |
3518 | template<> auto AirIRGenerator::addOp<OpType::F32Eq>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3519 | { |
3520 | result = g32(); |
3521 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleEqual), arg0, arg1, result); |
3522 | return { }; |
3523 | } |
3524 | |
3525 | template<> auto AirIRGenerator::addOp<OpType::I64Eq>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3526 | { |
3527 | result = g32(); |
3528 | append(Compare64, Arg::relCond(MacroAssembler::Equal), arg0, arg1, result); |
3529 | return { }; |
3530 | } |
3531 | |
3532 | template<> auto AirIRGenerator::addOp<OpType::F32Floor>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3533 | { |
3534 | result = f32(); |
3535 | append(FloorFloat, arg0, result); |
3536 | return { }; |
3537 | } |
3538 | |
3539 | template<> auto AirIRGenerator::addOp<OpType::F32ConvertSI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3540 | { |
3541 | result = f32(); |
3542 | append(ConvertInt32ToFloat, arg0, result); |
3543 | return { }; |
3544 | } |
3545 | |
3546 | template<> auto AirIRGenerator::addOp<OpType::I32Eqz>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3547 | { |
3548 | result = g32(); |
3549 | append(Test32, Arg::resCond(MacroAssembler::Zero), arg0, arg0, result); |
3550 | return { }; |
3551 | } |
3552 | |
3553 | template<> auto AirIRGenerator::addOp<OpType::I64ReinterpretF64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3554 | { |
3555 | result = g64(); |
3556 | append(MoveDoubleTo64, arg0, result); |
3557 | return { }; |
3558 | } |
3559 | |
3560 | template<> auto AirIRGenerator::addOp<OpType::I64ShrS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3561 | { |
3562 | return addShift(Type::I64, Rshift64, arg0, arg1, result); |
3563 | } |
3564 | |
3565 | template<> auto AirIRGenerator::addOp<OpType::I64ShrU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3566 | { |
3567 | return addShift(Type::I64, Urshift64, arg0, arg1, result); |
3568 | } |
3569 | |
3570 | template<> auto AirIRGenerator::addOp<OpType::F64Sqrt>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3571 | { |
3572 | result = f64(); |
3573 | append(SqrtDouble, arg0, result); |
3574 | return { }; |
3575 | } |
3576 | |
3577 | template<> auto AirIRGenerator::addOp<OpType::I64Shl>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3578 | { |
3579 | return addShift(Type::I64, Lshift64, arg0, arg1, result); |
3580 | } |
3581 | |
3582 | template<> auto AirIRGenerator::addOp<OpType::F32Gt>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3583 | { |
3584 | result = g32(); |
3585 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThan), arg0, arg1, result); |
3586 | return { }; |
3587 | } |
3588 | |
3589 | template<> auto AirIRGenerator::addOp<OpType::I32WrapI64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3590 | { |
3591 | result = g32(); |
3592 | append(Move32, arg0, result); |
3593 | return { }; |
3594 | } |
3595 | |
3596 | template<> auto AirIRGenerator::addOp<OpType::I32Rotl>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3597 | { |
3598 | if (isARM64()) { |
3599 | // ARM64 doesn't have a rotate left. |
3600 | auto newShift = g64(); |
3601 | append(Move, arg1, newShift); |
3602 | append(Neg64, newShift); |
3603 | return addShift(Type::I32, RotateRight32, arg0, newShift, result); |
3604 | } else |
3605 | return addShift(Type::I32, RotateLeft32, arg0, arg1, result); |
3606 | } |
3607 | |
3608 | template<> auto AirIRGenerator::addOp<OpType::I32Rotr>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3609 | { |
3610 | return addShift(Type::I32, RotateRight32, arg0, arg1, result); |
3611 | } |
3612 | |
3613 | template<> auto AirIRGenerator::addOp<OpType::I32GtU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3614 | { |
3615 | result = g32(); |
3616 | append(Compare32, Arg::relCond(MacroAssembler::Above), arg0, arg1, result); |
3617 | return { }; |
3618 | } |
3619 | |
3620 | template<> auto AirIRGenerator::addOp<OpType::I64ExtendSI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3621 | { |
3622 | result = g64(); |
3623 | append(SignExtend32ToPtr, arg0, result); |
3624 | return { }; |
3625 | } |
3626 | |
3627 | template<> auto AirIRGenerator::addOp<OpType::I32GtS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3628 | { |
3629 | result = g32(); |
3630 | append(Compare32, Arg::relCond(MacroAssembler::GreaterThan), arg0, arg1, result); |
3631 | return { }; |
3632 | } |
3633 | |
3634 | template<> auto AirIRGenerator::addOp<OpType::F64Neg>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3635 | { |
3636 | result = f64(); |
3637 | if (isValidForm(NegateDouble, Arg::Tmp, Arg::Tmp)) |
3638 | append(NegateDouble, arg0, result); |
3639 | else { |
3640 | auto constant = addConstant(Type::I64, bitwise_cast<uint64_t>(static_cast<double>(-0.0))); |
3641 | auto temp = g64(); |
3642 | append(MoveDoubleTo64, arg0, temp); |
3643 | append(Xor64, constant, temp); |
3644 | append(Move64ToDouble, temp, result); |
3645 | } |
3646 | return { }; |
3647 | } |
3648 | |
3649 | template<> auto AirIRGenerator::addOp<OpType::F64Max>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3650 | { |
3651 | return addFloatingPointMinOrMax(F64, MinOrMax::Max, arg0, arg1, result); |
3652 | } |
3653 | |
3654 | template<> auto AirIRGenerator::addOp<OpType::I64LeU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3655 | { |
3656 | result = g32(); |
3657 | append(Compare64, Arg::relCond(MacroAssembler::BelowOrEqual), arg0, arg1, result); |
3658 | return { }; |
3659 | } |
3660 | |
3661 | template<> auto AirIRGenerator::addOp<OpType::I64LeS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3662 | { |
3663 | result = g32(); |
3664 | append(Compare64, Arg::relCond(MacroAssembler::LessThanOrEqual), arg0, arg1, result); |
3665 | return { }; |
3666 | } |
3667 | |
3668 | template<> auto AirIRGenerator::addOp<OpType::I64Add>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3669 | { |
3670 | result = g64(); |
3671 | append(Add64, arg0, arg1, result); |
3672 | return { }; |
3673 | } |
3674 | |
3675 | } } // namespace JSC::Wasm |
3676 | |
3677 | #endif // ENABLE(WEBASSEMBLY) |
3678 | |