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 "AirHelpers.h" |
34 | #include "AirOpcodeUtils.h" |
35 | #include "AirValidate.h" |
36 | #include "AllowMacroScratchRegisterUsageIf.h" |
37 | #include "B3CCallValue.h" |
38 | #include "B3CheckSpecial.h" |
39 | #include "B3CheckValue.h" |
40 | #include "B3PatchpointSpecial.h" |
41 | #include "B3Procedure.h" |
42 | #include "B3ProcedureInlines.h" |
43 | #include "BinarySwitch.h" |
44 | #include "DisallowMacroScratchRegisterUsage.h" |
45 | #include "JSCInlines.h" |
46 | #include "JSWebAssemblyInstance.h" |
47 | #include "ScratchRegisterAllocator.h" |
48 | #include "VirtualRegister.h" |
49 | #include "WasmCallingConvention.h" |
50 | #include "WasmContextInlines.h" |
51 | #include "WasmExceptionType.h" |
52 | #include "WasmFunctionParser.h" |
53 | #include "WasmInstance.h" |
54 | #include "WasmMemory.h" |
55 | #include "WasmOMGPlan.h" |
56 | #include "WasmOSREntryData.h" |
57 | #include "WasmOpcodeOrigin.h" |
58 | #include "WasmOperations.h" |
59 | #include "WasmSignatureInlines.h" |
60 | #include "WasmThunks.h" |
61 | #include <limits> |
62 | #include <wtf/Box.h> |
63 | #include <wtf/Optional.h> |
64 | #include <wtf/StdLibExtras.h> |
65 | |
66 | namespace JSC { namespace Wasm { |
67 | |
68 | using namespace B3::Air; |
69 | |
70 | struct ConstrainedTmp { |
71 | ConstrainedTmp() = default; |
72 | ConstrainedTmp(Tmp tmp) |
73 | : ConstrainedTmp(tmp, tmp.isReg() ? B3::ValueRep::reg(tmp.reg()) : B3::ValueRep::SomeRegister) |
74 | { } |
75 | |
76 | ConstrainedTmp(Tmp tmp, B3::ValueRep rep) |
77 | : tmp(tmp) |
78 | , rep(rep) |
79 | { |
80 | } |
81 | |
82 | explicit operator bool() const { return !!tmp; } |
83 | |
84 | Tmp tmp; |
85 | B3::ValueRep rep; |
86 | }; |
87 | |
88 | class TypedTmp { |
89 | public: |
90 | constexpr TypedTmp() |
91 | : m_tmp() |
92 | , m_type(Type::Void) |
93 | { } |
94 | |
95 | TypedTmp(Tmp tmp, Type type) |
96 | : m_tmp(tmp) |
97 | , m_type(type) |
98 | { } |
99 | |
100 | TypedTmp(const TypedTmp&) = default; |
101 | TypedTmp(TypedTmp&&) = default; |
102 | TypedTmp& operator=(TypedTmp&&) = default; |
103 | TypedTmp& operator=(const TypedTmp&) = default; |
104 | |
105 | bool operator==(const TypedTmp& other) const |
106 | { |
107 | return m_tmp == other.m_tmp && m_type == other.m_type; |
108 | } |
109 | bool operator!=(const TypedTmp& other) const |
110 | { |
111 | return !(*this == other); |
112 | } |
113 | |
114 | explicit operator bool() const { return !!tmp(); } |
115 | |
116 | operator Tmp() const { return tmp(); } |
117 | operator Arg() const { return Arg(tmp()); } |
118 | Tmp tmp() const { return m_tmp; } |
119 | Type type() const { return m_type; } |
120 | |
121 | void dump(PrintStream& out) const |
122 | { |
123 | out.print("(" , m_tmp, ", " , m_type, ")" ); |
124 | } |
125 | |
126 | private: |
127 | |
128 | Tmp m_tmp; |
129 | Type m_type; |
130 | }; |
131 | |
132 | using TypedTmps = Vector<TypedTmp, 1>; |
133 | |
134 | class AirIRGenerator { |
135 | public: |
136 | struct ControlData { |
137 | ControlData(B3::Origin origin, BlockSignature result, TypedTmps resultTmps, BlockType type, BasicBlock* continuation, BasicBlock* special = nullptr) |
138 | : controlBlockType(type) |
139 | , continuation(continuation) |
140 | , special(special) |
141 | , results(resultTmps) |
142 | , returnType(result) |
143 | { |
144 | UNUSED_PARAM(origin); |
145 | } |
146 | |
147 | ControlData() |
148 | { |
149 | } |
150 | |
151 | void dump(PrintStream& out) const |
152 | { |
153 | switch (blockType()) { |
154 | case BlockType::If: |
155 | out.print("If: " ); |
156 | break; |
157 | case BlockType::Block: |
158 | out.print("Block: " ); |
159 | break; |
160 | case BlockType::Loop: |
161 | out.print("Loop: " ); |
162 | break; |
163 | case BlockType::TopLevel: |
164 | out.print("TopLevel: " ); |
165 | break; |
166 | } |
167 | out.print("Continuation: " , *continuation, ", Special: " ); |
168 | if (special) |
169 | out.print(*special); |
170 | else |
171 | out.print("None" ); |
172 | |
173 | CommaPrinter comma(", " , " Result Tmps: [" ); |
174 | for (const auto& tmp : results) |
175 | out.print(comma, tmp); |
176 | if (comma.didPrint()) |
177 | out.print("]" ); |
178 | } |
179 | |
180 | BlockType blockType() const { return controlBlockType; } |
181 | BlockSignature signature() const { return returnType; } |
182 | |
183 | BasicBlock* targetBlockForBranch() |
184 | { |
185 | if (blockType() == BlockType::Loop) |
186 | return special; |
187 | return continuation; |
188 | } |
189 | |
190 | void convertIfToBlock() |
191 | { |
192 | ASSERT(blockType() == BlockType::If); |
193 | controlBlockType = BlockType::Block; |
194 | special = nullptr; |
195 | } |
196 | |
197 | private: |
198 | friend class AirIRGenerator; |
199 | BlockType controlBlockType; |
200 | BasicBlock* continuation; |
201 | BasicBlock* special; |
202 | TypedTmps results; |
203 | BlockSignature returnType; |
204 | }; |
205 | |
206 | using ExpressionType = TypedTmp; |
207 | using ControlType = ControlData; |
208 | using ExpressionList = Vector<ExpressionType, 1>; |
209 | using Stack = ExpressionList; |
210 | using ResultList = TypedTmps; |
211 | using ControlEntry = FunctionParser<AirIRGenerator>::ControlEntry; |
212 | |
213 | static ExpressionType emptyExpression() { return { }; }; |
214 | Stack createStack() { return Stack(); } |
215 | bool isControlTypeIf(const ControlType& control) { return control.blockType() == BlockType::If; } |
216 | |
217 | using ErrorType = String; |
218 | using UnexpectedResult = Unexpected<ErrorType>; |
219 | using Result = Expected<std::unique_ptr<InternalFunction>, ErrorType>; |
220 | using PartialResult = Expected<void, ErrorType>; |
221 | |
222 | template <typename ...Args> |
223 | NEVER_INLINE UnexpectedResult WARN_UNUSED_RETURN fail(Args... args) const |
224 | { |
225 | using namespace FailureHelper; // See ADL comment in WasmParser.h. |
226 | return UnexpectedResult(makeString("WebAssembly.Module failed compiling: "_s , makeString(args)...)); |
227 | } |
228 | |
229 | #define WASM_COMPILE_FAIL_IF(condition, ...) do { \ |
230 | if (UNLIKELY(condition)) \ |
231 | return fail(__VA_ARGS__); \ |
232 | } while (0) |
233 | |
234 | AirIRGenerator(const ModuleInformation&, B3::Procedure&, InternalFunction*, Vector<UnlinkedWasmToWasmCall>&, MemoryMode, unsigned functionIndex, TierUpCount*, ThrowWasmException, const Signature&); |
235 | |
236 | PartialResult WARN_UNUSED_RETURN addArguments(const Signature&); |
237 | PartialResult WARN_UNUSED_RETURN addLocal(Type, uint32_t); |
238 | ExpressionType addConstant(Type, uint64_t); |
239 | ExpressionType addConstant(BasicBlock*, Type, uint64_t); |
240 | ExpressionType addBottom(BasicBlock*, Type); |
241 | |
242 | // References |
243 | PartialResult WARN_UNUSED_RETURN addRefIsNull(ExpressionType& value, ExpressionType& result); |
244 | PartialResult WARN_UNUSED_RETURN addRefFunc(uint32_t index, ExpressionType& result); |
245 | |
246 | // Tables |
247 | PartialResult WARN_UNUSED_RETURN addTableGet(unsigned, ExpressionType& index, ExpressionType& result); |
248 | PartialResult WARN_UNUSED_RETURN addTableSet(unsigned, ExpressionType& index, ExpressionType& value); |
249 | PartialResult WARN_UNUSED_RETURN addTableSize(unsigned, ExpressionType& result); |
250 | PartialResult WARN_UNUSED_RETURN addTableGrow(unsigned, ExpressionType& fill, ExpressionType& delta, ExpressionType& result); |
251 | PartialResult WARN_UNUSED_RETURN addTableFill(unsigned, ExpressionType& offset, ExpressionType& fill, ExpressionType& count); |
252 | |
253 | // Locals |
254 | PartialResult WARN_UNUSED_RETURN getLocal(uint32_t index, ExpressionType& result); |
255 | PartialResult WARN_UNUSED_RETURN setLocal(uint32_t index, ExpressionType value); |
256 | |
257 | // Globals |
258 | PartialResult WARN_UNUSED_RETURN getGlobal(uint32_t index, ExpressionType& result); |
259 | PartialResult WARN_UNUSED_RETURN setGlobal(uint32_t index, ExpressionType value); |
260 | |
261 | // Memory |
262 | PartialResult WARN_UNUSED_RETURN load(LoadOpType, ExpressionType pointer, ExpressionType& result, uint32_t offset); |
263 | PartialResult WARN_UNUSED_RETURN store(StoreOpType, ExpressionType pointer, ExpressionType value, uint32_t offset); |
264 | PartialResult WARN_UNUSED_RETURN addGrowMemory(ExpressionType delta, ExpressionType& result); |
265 | PartialResult WARN_UNUSED_RETURN addCurrentMemory(ExpressionType& result); |
266 | |
267 | // Basic operators |
268 | template<OpType> |
269 | PartialResult WARN_UNUSED_RETURN addOp(ExpressionType arg, ExpressionType& result); |
270 | template<OpType> |
271 | PartialResult WARN_UNUSED_RETURN addOp(ExpressionType left, ExpressionType right, ExpressionType& result); |
272 | PartialResult WARN_UNUSED_RETURN addSelect(ExpressionType condition, ExpressionType nonZero, ExpressionType zero, ExpressionType& result); |
273 | |
274 | // Control flow |
275 | ControlData WARN_UNUSED_RETURN addTopLevel(BlockSignature); |
276 | PartialResult WARN_UNUSED_RETURN addBlock(BlockSignature, Stack& enclosingStack, ControlType& newBlock, Stack& newStack); |
277 | PartialResult WARN_UNUSED_RETURN addLoop(BlockSignature, Stack& enclosingStack, ControlType& block, Stack& newStack, uint32_t loopIndex); |
278 | PartialResult WARN_UNUSED_RETURN addIf(ExpressionType condition, BlockSignature, Stack& enclosingStack, ControlType& result, Stack& newStack); |
279 | PartialResult WARN_UNUSED_RETURN addElse(ControlData&, const Stack&); |
280 | PartialResult WARN_UNUSED_RETURN addElseToUnreachable(ControlData&); |
281 | |
282 | PartialResult WARN_UNUSED_RETURN addReturn(const ControlData&, const ExpressionList& returnValues); |
283 | PartialResult WARN_UNUSED_RETURN addBranch(ControlData&, ExpressionType condition, const Stack& returnValues); |
284 | PartialResult WARN_UNUSED_RETURN addSwitch(ExpressionType condition, const Vector<ControlData*>& targets, ControlData& defaultTargets, const Stack& expressionStack); |
285 | PartialResult WARN_UNUSED_RETURN endBlock(ControlEntry&, Stack& expressionStack); |
286 | PartialResult WARN_UNUSED_RETURN addEndToUnreachable(ControlEntry&, const Stack& expressionStack = { }); |
287 | |
288 | // Calls |
289 | PartialResult WARN_UNUSED_RETURN addCall(uint32_t calleeIndex, const Signature&, Vector<ExpressionType>& args, Vector<ExpressionType, 1>& results); |
290 | PartialResult WARN_UNUSED_RETURN addCallIndirect(unsigned tableIndex, const Signature&, Vector<ExpressionType>& args, Vector<ExpressionType, 1>& results); |
291 | PartialResult WARN_UNUSED_RETURN addUnreachable(); |
292 | B3::PatchpointValue* WARN_UNUSED_RETURN emitCallPatchpoint(BasicBlock*, const Signature&, const Vector<ExpressionType, 1>& results, const Vector<TypedTmp>& args, Vector<ConstrainedTmp>&& = { }); |
293 | |
294 | PartialResult addShift(Type, B3::Air::Opcode, ExpressionType value, ExpressionType shift, ExpressionType& result); |
295 | PartialResult addIntegerSub(B3::Air::Opcode, ExpressionType lhs, ExpressionType rhs, ExpressionType& result); |
296 | PartialResult addFloatingPointAbs(B3::Air::Opcode, ExpressionType value, ExpressionType& result); |
297 | PartialResult addFloatingPointBinOp(Type, B3::Air::Opcode, ExpressionType lhs, ExpressionType rhs, ExpressionType& result); |
298 | |
299 | void dump(const Vector<ControlEntry>& controlStack, const Stack* expressionStack); |
300 | void setParser(FunctionParser<AirIRGenerator>* parser) { m_parser = parser; }; |
301 | void didFinishParsingLocals() { } |
302 | |
303 | static Vector<Tmp> toTmpVector(const Vector<TypedTmp>& vector) |
304 | { |
305 | Vector<Tmp> result; |
306 | for (const auto& item : vector) |
307 | result.append(item.tmp()); |
308 | return result; |
309 | } |
310 | |
311 | const Bag<B3::PatchpointValue*>& patchpoints() const |
312 | { |
313 | return m_patchpoints; |
314 | } |
315 | |
316 | private: |
317 | B3::Type toB3ResultType(BlockSignature returnType); |
318 | ALWAYS_INLINE void validateInst(Inst& inst) |
319 | { |
320 | if (!ASSERT_DISABLED) { |
321 | if (!inst.isValidForm()) { |
322 | dataLogLn("Inst validation failed:" ); |
323 | dataLogLn(inst, "\n" ); |
324 | if (inst.origin) |
325 | dataLogLn(deepDump(inst.origin), "\n" ); |
326 | CRASH(); |
327 | } |
328 | } |
329 | } |
330 | |
331 | static Arg (const TypedTmp& tmp) { return tmp.tmp(); } |
332 | static Arg (const Tmp& tmp) { return Arg(tmp); } |
333 | static Arg (const Arg& arg) { return arg; } |
334 | |
335 | template<typename... Arguments> |
336 | void append(BasicBlock* block, Kind kind, Arguments&&... arguments) |
337 | { |
338 | // FIXME: Find a way to use origin here. |
339 | auto& inst = block->append(kind, nullptr, extractArg(arguments)...); |
340 | validateInst(inst); |
341 | } |
342 | |
343 | template<typename... Arguments> |
344 | void append(Kind kind, Arguments&&... arguments) |
345 | { |
346 | append(m_currentBlock, kind, std::forward<Arguments>(arguments)...); |
347 | } |
348 | |
349 | template<typename... Arguments> |
350 | void appendEffectful(B3::Air::Opcode op, Arguments&&... arguments) |
351 | { |
352 | Kind kind = op; |
353 | kind.effects = true; |
354 | append(m_currentBlock, kind, std::forward<Arguments>(arguments)...); |
355 | } |
356 | |
357 | Tmp newTmp(B3::Bank bank) |
358 | { |
359 | switch (bank) { |
360 | case B3::GP: |
361 | if (m_freeGPs.size()) |
362 | return m_freeGPs.takeLast(); |
363 | break; |
364 | case B3::FP: |
365 | if (m_freeFPs.size()) |
366 | return m_freeFPs.takeLast(); |
367 | break; |
368 | } |
369 | return m_code.newTmp(bank); |
370 | } |
371 | |
372 | TypedTmp g32() { return { newTmp(B3::GP), Type::I32 }; } |
373 | TypedTmp g64() { return { newTmp(B3::GP), Type::I64 }; } |
374 | TypedTmp gAnyref() { return { newTmp(B3::GP), Type::Anyref }; } |
375 | TypedTmp gFuncref() { return { newTmp(B3::GP), Type::Funcref }; } |
376 | TypedTmp f32() { return { newTmp(B3::FP), Type::F32 }; } |
377 | TypedTmp f64() { return { newTmp(B3::FP), Type::F64 }; } |
378 | |
379 | TypedTmp tmpForType(Type type) |
380 | { |
381 | switch (type) { |
382 | case Type::I32: |
383 | return g32(); |
384 | case Type::I64: |
385 | return g64(); |
386 | case Type::Funcref: |
387 | return gFuncref(); |
388 | case Type::Anyref: |
389 | return gAnyref(); |
390 | case Type::F32: |
391 | return f32(); |
392 | case Type::F64: |
393 | return f64(); |
394 | case Type::Void: |
395 | return { }; |
396 | default: |
397 | RELEASE_ASSERT_NOT_REACHED(); |
398 | } |
399 | } |
400 | |
401 | TypedTmps tmpsForSignature(BlockSignature signature) |
402 | { |
403 | TypedTmps result(signature->returnCount()); |
404 | for (unsigned i = 0; i < signature->returnCount(); ++i) |
405 | result[i] = tmpForType(signature->returnType(i)); |
406 | return result; |
407 | } |
408 | |
409 | B3::PatchpointValue* addPatchpoint(B3::Type type) |
410 | { |
411 | auto* result = m_proc.add<B3::PatchpointValue>(type, B3::Origin()); |
412 | if (UNLIKELY(shouldDumpIRAtEachPhase(B3::AirMode))) |
413 | m_patchpoints.add(result); |
414 | return result; |
415 | } |
416 | |
417 | template <typename ...Args> |
418 | void emitPatchpoint(B3::PatchpointValue* patch, Tmp result, Args... theArgs) |
419 | { |
420 | emitPatchpoint(m_currentBlock, patch, result, std::forward<Args>(theArgs)...); |
421 | } |
422 | |
423 | template <typename ...Args> |
424 | void emitPatchpoint(BasicBlock* basicBlock, B3::PatchpointValue* patch, Tmp result, Args... theArgs) |
425 | { |
426 | emitPatchpoint(basicBlock, patch, Vector<Tmp, 1> { result }, Vector<ConstrainedTmp, sizeof...(Args)>::from(theArgs...)); |
427 | } |
428 | |
429 | void emitPatchpoint(BasicBlock* basicBlock, B3::PatchpointValue* patch, Tmp result) |
430 | { |
431 | emitPatchpoint(basicBlock, patch, Vector<Tmp, 1> { result }, Vector<ConstrainedTmp>()); |
432 | } |
433 | |
434 | template <typename ResultTmpType, size_t inlineSize> |
435 | void emitPatchpoint(BasicBlock* basicBlock, B3::PatchpointValue* patch, const Vector<ResultTmpType, 1>& results, Vector<ConstrainedTmp, inlineSize>&& args) |
436 | { |
437 | if (!m_patchpointSpecial) |
438 | m_patchpointSpecial = static_cast<B3::PatchpointSpecial*>(m_code.addSpecial(makeUnique<B3::PatchpointSpecial>())); |
439 | |
440 | auto toTmp = [&] (ResultTmpType tmp) { |
441 | if constexpr (std::is_same_v<ResultTmpType, Tmp>) |
442 | return tmp; |
443 | else |
444 | return tmp.tmp(); |
445 | }; |
446 | |
447 | Inst inst(Patch, patch, Arg::special(m_patchpointSpecial)); |
448 | Vector<Inst, 1> resultMovs; |
449 | switch (patch->type().kind()) { |
450 | case B3::Void: |
451 | break; |
452 | default: { |
453 | ASSERT(results.size()); |
454 | for (unsigned i = 0; i < results.size(); ++i) { |
455 | switch (patch->resultConstraints[i].kind()) { |
456 | case B3::ValueRep::StackArgument: { |
457 | Arg arg = Arg::callArg(patch->resultConstraints[i].offsetFromSP()); |
458 | inst.args.append(arg); |
459 | resultMovs.append(Inst(B3::Air::moveForType(m_proc.typeAtOffset(patch->type(), i)), nullptr, arg, toTmp(results[i]))); |
460 | break; |
461 | } |
462 | case B3::ValueRep::Register: { |
463 | inst.args.append(Tmp(patch->resultConstraints[i].reg())); |
464 | resultMovs.append(Inst(B3::Air::relaxedMoveForType(m_proc.typeAtOffset(patch->type(), i)), nullptr, Tmp(patch->resultConstraints[i].reg()), toTmp(results[i]))); |
465 | break; |
466 | } |
467 | case B3::ValueRep::SomeRegister: { |
468 | inst.args.append(toTmp(results[i])); |
469 | break; |
470 | } |
471 | default: |
472 | RELEASE_ASSERT_NOT_REACHED(); |
473 | } |
474 | } |
475 | } |
476 | } |
477 | |
478 | for (unsigned i = 0; i < args.size(); ++i) { |
479 | ConstrainedTmp& tmp = args[i]; |
480 | // FIXME: This is less than ideal to create dummy values just to satisfy Air's |
481 | // validation. We should abstrcat Patch enough so ValueRep's don't need to be |
482 | // backed by Values. |
483 | // https://bugs.webkit.org/show_bug.cgi?id=194040 |
484 | B3::Value* dummyValue = m_proc.addConstant(B3::Origin(), tmp.tmp.isGP() ? B3::Int64 : B3::Double, 0); |
485 | patch->append(dummyValue, tmp.rep); |
486 | switch (tmp.rep.kind()) { |
487 | case B3::ValueRep::ColdAny: // B3::Value propagates ColdAny information and later Air will allocate appropriate stack. |
488 | case B3::ValueRep::SomeRegister: |
489 | inst.args.append(tmp.tmp); |
490 | break; |
491 | case B3::ValueRep::Register: |
492 | patch->earlyClobbered().clear(tmp.rep.reg()); |
493 | append(basicBlock, tmp.tmp.isGP() ? Move : MoveDouble, tmp.tmp, tmp.rep.reg()); |
494 | inst.args.append(Tmp(tmp.rep.reg())); |
495 | break; |
496 | case B3::ValueRep::StackArgument: { |
497 | ASSERT(!patch->effects.terminal); |
498 | Arg arg = Arg::callArg(tmp.rep.offsetFromSP()); |
499 | append(basicBlock, tmp.tmp.isGP() ? Move : MoveDouble, tmp.tmp, arg); |
500 | ASSERT(arg.canRepresent(patch->child(i)->type())); |
501 | inst.args.append(arg); |
502 | break; |
503 | } |
504 | default: |
505 | RELEASE_ASSERT_NOT_REACHED(); |
506 | } |
507 | } |
508 | |
509 | for (auto valueRep : patch->resultConstraints) { |
510 | if (valueRep.isReg()) |
511 | patch->lateClobbered().clear(valueRep.reg()); |
512 | } |
513 | for (unsigned i = patch->numGPScratchRegisters; i--;) |
514 | inst.args.append(g64().tmp()); |
515 | for (unsigned i = patch->numFPScratchRegisters; i--;) |
516 | inst.args.append(f64().tmp()); |
517 | |
518 | validateInst(inst); |
519 | basicBlock->append(WTFMove(inst)); |
520 | for (Inst result : resultMovs) { |
521 | validateInst(result); |
522 | basicBlock->append(WTFMove(result)); |
523 | } |
524 | } |
525 | |
526 | template <typename Branch, typename Generator> |
527 | void emitCheck(const Branch& makeBranch, const Generator& generator) |
528 | { |
529 | // We fail along the truthy edge of 'branch'. |
530 | Inst branch = makeBranch(); |
531 | |
532 | // FIXME: Make a hashmap of these. |
533 | B3::CheckSpecial::Key key(branch); |
534 | B3::CheckSpecial* special = static_cast<B3::CheckSpecial*>(m_code.addSpecial(makeUnique<B3::CheckSpecial>(key))); |
535 | |
536 | // FIXME: Remove the need for dummy values |
537 | // https://bugs.webkit.org/show_bug.cgi?id=194040 |
538 | B3::Value* dummyPredicate = m_proc.addConstant(B3::Origin(), B3::Int32, 42); |
539 | B3::CheckValue* checkValue = m_proc.add<B3::CheckValue>(B3::Check, B3::Origin(), dummyPredicate); |
540 | checkValue->setGenerator(generator); |
541 | |
542 | Inst inst(Patch, checkValue, Arg::special(special)); |
543 | inst.args.appendVector(branch.args); |
544 | m_currentBlock->append(WTFMove(inst)); |
545 | } |
546 | |
547 | template <typename Func, typename ...Args> |
548 | void emitCCall(Func func, TypedTmp result, Args... args) |
549 | { |
550 | emitCCall(m_currentBlock, func, result, std::forward<Args>(args)...); |
551 | } |
552 | template <typename Func, typename ...Args> |
553 | void emitCCall(BasicBlock* block, Func func, TypedTmp result, Args... theArgs) |
554 | { |
555 | B3::Type resultType = B3::Void; |
556 | if (result) { |
557 | switch (result.type()) { |
558 | case Type::I32: |
559 | resultType = B3::Int32; |
560 | break; |
561 | case Type::I64: |
562 | case Type::Anyref: |
563 | case Type::Funcref: |
564 | resultType = B3::Int64; |
565 | break; |
566 | case Type::F32: |
567 | resultType = B3::Float; |
568 | break; |
569 | case Type::F64: |
570 | resultType = B3::Double; |
571 | break; |
572 | default: |
573 | RELEASE_ASSERT_NOT_REACHED(); |
574 | } |
575 | } |
576 | |
577 | auto makeDummyValue = [&] (Tmp tmp) { |
578 | // FIXME: This is less than ideal to create dummy values just to satisfy Air's |
579 | // validation. We should abstrcat CCall enough so we're not reliant on arguments |
580 | // to the B3::CCallValue. |
581 | // https://bugs.webkit.org/show_bug.cgi?id=194040 |
582 | if (tmp.isGP()) |
583 | return m_proc.addConstant(B3::Origin(), B3::Int64, 0); |
584 | return m_proc.addConstant(B3::Origin(), B3::Double, 0); |
585 | }; |
586 | |
587 | B3::Value* dummyFunc = m_proc.addConstant(B3::Origin(), B3::Int64, bitwise_cast<uintptr_t>(func)); |
588 | B3::Value* origin = m_proc.add<B3::CCallValue>(resultType, B3::Origin(), B3::Effects::none(), dummyFunc, makeDummyValue(theArgs)...); |
589 | |
590 | Inst inst(CCall, origin); |
591 | |
592 | Tmp callee = g64(); |
593 | append(block, Move, Arg::immPtr(tagCFunctionPtr<void*>(func, B3CCallPtrTag)), callee); |
594 | inst.args.append(callee); |
595 | |
596 | if (result) |
597 | inst.args.append(result.tmp()); |
598 | |
599 | for (Tmp tmp : Vector<Tmp, sizeof...(Args)>::from(theArgs.tmp()...)) |
600 | inst.args.append(tmp); |
601 | |
602 | block->append(WTFMove(inst)); |
603 | } |
604 | |
605 | static B3::Air::Opcode moveOpForValueType(Type type) |
606 | { |
607 | switch (type) { |
608 | case Type::I32: |
609 | return Move32; |
610 | case Type::I64: |
611 | case Type::Anyref: |
612 | case Type::Funcref: |
613 | return Move; |
614 | case Type::F32: |
615 | return MoveFloat; |
616 | case Type::F64: |
617 | return MoveDouble; |
618 | default: |
619 | RELEASE_ASSERT_NOT_REACHED(); |
620 | } |
621 | } |
622 | |
623 | void emitThrowException(CCallHelpers&, ExceptionType); |
624 | |
625 | void emitEntryTierUpCheck(); |
626 | void emitLoopTierUpCheck(uint32_t loopIndex); |
627 | |
628 | void emitWriteBarrierForJSWrapper(); |
629 | ExpressionType emitCheckAndPreparePointer(ExpressionType pointer, uint32_t offset, uint32_t sizeOfOp); |
630 | ExpressionType emitLoadOp(LoadOpType, ExpressionType pointer, uint32_t offset); |
631 | void emitStoreOp(StoreOpType, ExpressionType pointer, ExpressionType value, uint32_t offset); |
632 | |
633 | void unify(const ExpressionType& dst, const ExpressionType& source); |
634 | void unifyValuesWithBlock(const Stack& resultStack, const ResultList& stack); |
635 | |
636 | template <typename IntType> |
637 | void emitChecksForModOrDiv(bool isSignedDiv, ExpressionType left, ExpressionType right); |
638 | |
639 | template <typename IntType> |
640 | void emitModOrDiv(bool isDiv, ExpressionType lhs, ExpressionType rhs, ExpressionType& result); |
641 | |
642 | enum class MinOrMax { Min, Max }; |
643 | |
644 | PartialResult addFloatingPointMinOrMax(Type, MinOrMax, ExpressionType lhs, ExpressionType rhs, ExpressionType& result); |
645 | |
646 | int32_t WARN_UNUSED_RETURN fixupPointerPlusOffset(ExpressionType&, uint32_t); |
647 | |
648 | void restoreWasmContextInstance(BasicBlock*, TypedTmp); |
649 | enum class RestoreCachedStackLimit { No, Yes }; |
650 | void restoreWebAssemblyGlobalState(RestoreCachedStackLimit, const MemoryInformation&, TypedTmp instance, BasicBlock*); |
651 | |
652 | B3::Origin origin(); |
653 | |
654 | uint32_t outerLoopIndex() const |
655 | { |
656 | if (m_outerLoops.isEmpty()) |
657 | return UINT32_MAX; |
658 | return m_outerLoops.last(); |
659 | } |
660 | |
661 | FunctionParser<AirIRGenerator>* m_parser { nullptr }; |
662 | const ModuleInformation& m_info; |
663 | const MemoryMode m_mode { MemoryMode::BoundsChecking }; |
664 | const unsigned m_functionIndex { UINT_MAX }; |
665 | TierUpCount* m_tierUp { nullptr }; |
666 | |
667 | B3::Procedure& m_proc; |
668 | Code& m_code; |
669 | Vector<uint32_t> m_outerLoops; |
670 | BasicBlock* m_currentBlock { nullptr }; |
671 | BasicBlock* m_rootBlock { nullptr }; |
672 | Vector<TypedTmp> m_locals; |
673 | Vector<UnlinkedWasmToWasmCall>& m_unlinkedWasmToWasmCalls; // List each call site and the function index whose address it should be patched with. |
674 | GPRReg m_memoryBaseGPR { InvalidGPRReg }; |
675 | GPRReg m_memorySizeGPR { InvalidGPRReg }; |
676 | GPRReg m_wasmContextInstanceGPR { InvalidGPRReg }; |
677 | bool m_makesCalls { false }; |
678 | |
679 | Vector<Tmp, 8> m_freeGPs; |
680 | Vector<Tmp, 8> m_freeFPs; |
681 | |
682 | HashMap<BlockSignature, B3::Type> m_tupleMap; |
683 | // This is only filled if we are dumping IR. |
684 | Bag<B3::PatchpointValue*> m_patchpoints; |
685 | |
686 | TypedTmp m_instanceValue; // Always use the accessor below to ensure the instance value is materialized when used. |
687 | bool m_usesInstanceValue { false }; |
688 | TypedTmp instanceValue() |
689 | { |
690 | m_usesInstanceValue = true; |
691 | return m_instanceValue; |
692 | } |
693 | |
694 | uint32_t m_maxNumJSCallArguments { 0 }; |
695 | unsigned m_numImportFunctions; |
696 | |
697 | B3::PatchpointSpecial* m_patchpointSpecial { nullptr }; |
698 | }; |
699 | |
700 | // Memory accesses in WebAssembly have unsigned 32-bit offsets, whereas they have signed 32-bit offsets in B3. |
701 | int32_t AirIRGenerator::fixupPointerPlusOffset(ExpressionType& ptr, uint32_t offset) |
702 | { |
703 | if (static_cast<uint64_t>(offset) > static_cast<uint64_t>(std::numeric_limits<int32_t>::max())) { |
704 | auto previousPtr = ptr; |
705 | ptr = g64(); |
706 | auto constant = g64(); |
707 | append(Move, Arg::bigImm(offset), constant); |
708 | append(Add64, constant, previousPtr, ptr); |
709 | return 0; |
710 | } |
711 | return offset; |
712 | } |
713 | |
714 | void AirIRGenerator::restoreWasmContextInstance(BasicBlock* block, TypedTmp instance) |
715 | { |
716 | if (Context::useFastTLS()) { |
717 | auto* patchpoint = addPatchpoint(B3::Void); |
718 | if (CCallHelpers::storeWasmContextInstanceNeedsMacroScratchRegister()) |
719 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
720 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
721 | AllowMacroScratchRegisterUsageIf allowScratch(jit, CCallHelpers::storeWasmContextInstanceNeedsMacroScratchRegister()); |
722 | jit.storeWasmContextInstance(params[0].gpr()); |
723 | }); |
724 | emitPatchpoint(block, patchpoint, Tmp(), instance); |
725 | return; |
726 | } |
727 | |
728 | // FIXME: Because WasmToWasm call clobbers wasmContextInstance register and does not restore it, we need to restore it in the caller side. |
729 | // This prevents us from using ArgumentReg to this (logically) immutable pinned register. |
730 | auto* patchpoint = addPatchpoint(B3::Void); |
731 | B3::Effects effects = B3::Effects::none(); |
732 | effects.writesPinned = true; |
733 | effects.reads = B3::HeapRange::top(); |
734 | patchpoint->effects = effects; |
735 | patchpoint->clobberLate(RegisterSet(m_wasmContextInstanceGPR)); |
736 | GPRReg wasmContextInstanceGPR = m_wasmContextInstanceGPR; |
737 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& param) { |
738 | jit.move(param[0].gpr(), wasmContextInstanceGPR); |
739 | }); |
740 | emitPatchpoint(block, patchpoint, Tmp(), instance); |
741 | } |
742 | |
743 | AirIRGenerator::AirIRGenerator(const ModuleInformation& info, B3::Procedure& procedure, InternalFunction* compilation, Vector<UnlinkedWasmToWasmCall>& unlinkedWasmToWasmCalls, MemoryMode mode, unsigned functionIndex, TierUpCount* tierUp, ThrowWasmException throwWasmException, const Signature& signature) |
744 | : m_info(info) |
745 | , m_mode(mode) |
746 | , m_functionIndex(functionIndex) |
747 | , m_tierUp(tierUp) |
748 | , m_proc(procedure) |
749 | , m_code(m_proc.code()) |
750 | , m_unlinkedWasmToWasmCalls(unlinkedWasmToWasmCalls) |
751 | , m_numImportFunctions(info.importFunctionCount()) |
752 | { |
753 | m_currentBlock = m_code.addBlock(); |
754 | m_rootBlock = m_currentBlock; |
755 | |
756 | // 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 |
757 | const PinnedRegisterInfo& pinnedRegs = PinnedRegisterInfo::get(); |
758 | |
759 | m_memoryBaseGPR = pinnedRegs.baseMemoryPointer; |
760 | m_code.pinRegister(m_memoryBaseGPR); |
761 | |
762 | m_wasmContextInstanceGPR = pinnedRegs.wasmContextInstancePointer; |
763 | if (!Context::useFastTLS()) |
764 | m_code.pinRegister(m_wasmContextInstanceGPR); |
765 | |
766 | if (mode != MemoryMode::Signaling) { |
767 | m_memorySizeGPR = pinnedRegs.sizeRegister; |
768 | m_code.pinRegister(m_memorySizeGPR); |
769 | } |
770 | |
771 | if (throwWasmException) |
772 | Thunks::singleton().setThrowWasmException(throwWasmException); |
773 | |
774 | if (info.memory) { |
775 | switch (m_mode) { |
776 | case MemoryMode::BoundsChecking: |
777 | break; |
778 | case MemoryMode::Signaling: |
779 | // Most memory accesses in signaling mode don't do an explicit |
780 | // exception check because they can rely on fault handling to detect |
781 | // out-of-bounds accesses. FaultSignalHandler nonetheless needs the |
782 | // thunk to exist so that it can jump to that thunk. |
783 | if (UNLIKELY(!Thunks::singleton().stub(throwExceptionFromWasmThunkGenerator))) |
784 | CRASH(); |
785 | break; |
786 | } |
787 | } |
788 | |
789 | m_code.setNumEntrypoints(1); |
790 | |
791 | GPRReg contextInstance = Context::useFastTLS() ? wasmCallingConvention().prologueScratchGPRs[1] : m_wasmContextInstanceGPR; |
792 | |
793 | Ref<B3::Air::PrologueGenerator> prologueGenerator = createSharedTask<B3::Air::PrologueGeneratorFunction>([=] (CCallHelpers& jit, B3::Air::Code& code) { |
794 | AllowMacroScratchRegisterUsage allowScratch(jit); |
795 | code.emitDefaultPrologue(jit); |
796 | |
797 | { |
798 | GPRReg calleeGPR = wasmCallingConvention().prologueScratchGPRs[0]; |
799 | auto moveLocation = jit.moveWithPatch(MacroAssembler::TrustedImmPtr(nullptr), calleeGPR); |
800 | jit.addLinkTask([compilation, moveLocation] (LinkBuffer& linkBuffer) { |
801 | compilation->calleeMoveLocation = linkBuffer.locationOf<WasmEntryPtrTag>(moveLocation); |
802 | }); |
803 | jit.emitPutToCallFrameHeader(calleeGPR, CallFrameSlot::callee); |
804 | jit.emitPutToCallFrameHeader(nullptr, CallFrameSlot::codeBlock); |
805 | } |
806 | |
807 | { |
808 | const Checked<int32_t> wasmFrameSize = m_code.frameSize(); |
809 | const unsigned minimumParentCheckSize = WTF::roundUpToMultipleOf(stackAlignmentBytes(), 1024); |
810 | const unsigned = WTF::roundUpToMultipleOf(stackAlignmentBytes(), std::max<uint32_t>( |
811 | // This allows us to elide stack checks for functions that are terminal nodes in the call |
812 | // tree, (e.g they don't make any calls) and have a small enough frame size. This works by |
813 | // having any such terminal node have its parent caller include some extra size in its |
814 | // own check for it. The goal here is twofold: |
815 | // 1. Emit less code. |
816 | // 2. Try to speed things up by skipping stack checks. |
817 | minimumParentCheckSize, |
818 | // This allows us to elide stack checks in the Wasm -> Embedder call IC stub. Since these will |
819 | // spill all arguments to the stack, we ensure that a stack check here covers the |
820 | // stack that such a stub would use. |
821 | (Checked<uint32_t>(m_maxNumJSCallArguments) * sizeof(Register) + jsCallingConvention().headerSizeInBytes).unsafeGet() |
822 | )); |
823 | const int32_t checkSize = m_makesCalls ? (wasmFrameSize + extraFrameSize).unsafeGet() : wasmFrameSize.unsafeGet(); |
824 | bool needUnderflowCheck = static_cast<unsigned>(checkSize) > Options::reservedZoneSize(); |
825 | bool needsOverflowCheck = m_makesCalls || wasmFrameSize >= minimumParentCheckSize || needUnderflowCheck; |
826 | |
827 | // This allows leaf functions to not do stack checks if their frame size is within |
828 | // certain limits since their caller would have already done the check. |
829 | if (needsOverflowCheck) { |
830 | GPRReg scratch = wasmCallingConvention().prologueScratchGPRs[0]; |
831 | |
832 | if (Context::useFastTLS()) |
833 | jit.loadWasmContextInstance(contextInstance); |
834 | |
835 | jit.addPtr(CCallHelpers::TrustedImm32(-checkSize), GPRInfo::callFrameRegister, scratch); |
836 | MacroAssembler::JumpList overflow; |
837 | if (UNLIKELY(needUnderflowCheck)) |
838 | overflow.append(jit.branchPtr(CCallHelpers::Above, scratch, GPRInfo::callFrameRegister)); |
839 | overflow.append(jit.branchPtr(CCallHelpers::Below, scratch, CCallHelpers::Address(contextInstance, Instance::offsetOfCachedStackLimit()))); |
840 | jit.addLinkTask([overflow] (LinkBuffer& linkBuffer) { |
841 | linkBuffer.link(overflow, CodeLocationLabel<JITThunkPtrTag>(Thunks::singleton().stub(throwStackOverflowFromWasmThunkGenerator).code())); |
842 | }); |
843 | } else if (m_usesInstanceValue && Context::useFastTLS()) { |
844 | // No overflow check is needed, but the instance values still needs to be correct. |
845 | jit.loadWasmContextInstance(contextInstance); |
846 | } |
847 | } |
848 | }); |
849 | |
850 | m_code.setPrologueForEntrypoint(0, WTFMove(prologueGenerator)); |
851 | |
852 | if (Context::useFastTLS()) { |
853 | m_instanceValue = g64(); |
854 | // FIXME: Would be nice to only do this if we use instance value. |
855 | append(Move, Tmp(contextInstance), m_instanceValue); |
856 | } else |
857 | m_instanceValue = { Tmp(contextInstance), Type::I64 }; |
858 | |
859 | ASSERT(!m_locals.size()); |
860 | m_locals.grow(signature.argumentCount()); |
861 | for (unsigned i = 0; i < signature.argumentCount(); ++i) { |
862 | Type type = signature.argument(i); |
863 | m_locals[i] = tmpForType(type); |
864 | } |
865 | |
866 | CallInformation wasmCallInfo = wasmCallingConvention().callInformationFor(signature, CallRole::Callee); |
867 | |
868 | for (unsigned i = 0; i < wasmCallInfo.params.size(); ++i) { |
869 | B3::ValueRep location = wasmCallInfo.params[i]; |
870 | Arg arg = location.isReg() ? Arg(Tmp(location.reg())) : Arg::addr(Tmp(GPRInfo::callFrameRegister), location.offsetFromFP()); |
871 | switch (signature.argument(i)) { |
872 | case Type::I32: |
873 | append(Move32, arg, m_locals[i]); |
874 | break; |
875 | case Type::I64: |
876 | case Type::Anyref: |
877 | case Type::Funcref: |
878 | append(Move, arg, m_locals[i]); |
879 | break; |
880 | case Type::F32: |
881 | append(MoveFloat, arg, m_locals[i]); |
882 | break; |
883 | case Type::F64: |
884 | append(MoveDouble, arg, m_locals[i]); |
885 | break; |
886 | default: |
887 | RELEASE_ASSERT_NOT_REACHED(); |
888 | } |
889 | } |
890 | |
891 | emitEntryTierUpCheck(); |
892 | } |
893 | |
894 | B3::Type AirIRGenerator::toB3ResultType(BlockSignature returnType) |
895 | { |
896 | if (returnType->returnsVoid()) |
897 | return B3::Void; |
898 | |
899 | if (returnType->returnCount() == 1) |
900 | return toB3Type(returnType->returnType(0)); |
901 | |
902 | auto result = m_tupleMap.ensure(returnType, [&] { |
903 | Vector<B3::Type> result; |
904 | for (unsigned i = 0; i < returnType->returnCount(); ++i) |
905 | result.append(toB3Type(returnType->returnType(i))); |
906 | return m_proc.addTuple(WTFMove(result)); |
907 | }); |
908 | return result.iterator->value; |
909 | } |
910 | |
911 | void AirIRGenerator::restoreWebAssemblyGlobalState(RestoreCachedStackLimit restoreCachedStackLimit, const MemoryInformation& memory, TypedTmp instance, BasicBlock* block) |
912 | { |
913 | restoreWasmContextInstance(block, instance); |
914 | |
915 | if (restoreCachedStackLimit == RestoreCachedStackLimit::Yes) { |
916 | // The Instance caches the stack limit, but also knows where its canonical location is. |
917 | static_assert(sizeof(decltype(static_cast<Instance*>(nullptr)->cachedStackLimit())) == sizeof(uint64_t), "" ); |
918 | |
919 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfPointerToActualStackLimit(), B3::Width64)); |
920 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfCachedStackLimit(), B3::Width64)); |
921 | auto temp = g64(); |
922 | append(block, Move, Arg::addr(instanceValue(), Instance::offsetOfPointerToActualStackLimit()), temp); |
923 | append(block, Move, Arg::addr(temp), temp); |
924 | append(block, Move, temp, Arg::addr(instanceValue(), Instance::offsetOfCachedStackLimit())); |
925 | } |
926 | |
927 | if (!!memory) { |
928 | const PinnedRegisterInfo* pinnedRegs = &PinnedRegisterInfo::get(); |
929 | RegisterSet clobbers; |
930 | clobbers.set(pinnedRegs->baseMemoryPointer); |
931 | clobbers.set(pinnedRegs->sizeRegister); |
932 | if (!isARM64()) |
933 | clobbers.set(RegisterSet::macroScratchRegisters()); |
934 | |
935 | auto* patchpoint = addPatchpoint(B3::Void); |
936 | B3::Effects effects = B3::Effects::none(); |
937 | effects.writesPinned = true; |
938 | effects.reads = B3::HeapRange::top(); |
939 | patchpoint->effects = effects; |
940 | patchpoint->clobber(clobbers); |
941 | patchpoint->numGPScratchRegisters = Gigacage::isEnabled(Gigacage::Primitive) ? 1 : 0; |
942 | |
943 | patchpoint->setGenerator([pinnedRegs] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
944 | AllowMacroScratchRegisterUsage allowScratch(jit); |
945 | GPRReg baseMemory = pinnedRegs->baseMemoryPointer; |
946 | GPRReg scratchOrSize = Gigacage::isEnabled(Gigacage::Primitive) ? params.gpScratch(0) : pinnedRegs->sizeRegister; |
947 | |
948 | jit.loadPtr(CCallHelpers::Address(params[0].gpr(), Instance::offsetOfCachedMemorySize()), pinnedRegs->sizeRegister); |
949 | jit.loadPtr(CCallHelpers::Address(params[0].gpr(), Instance::offsetOfCachedMemory()), baseMemory); |
950 | |
951 | jit.cageConditionally(Gigacage::Primitive, baseMemory, pinnedRegs->sizeRegister, scratchOrSize); |
952 | }); |
953 | |
954 | emitPatchpoint(block, patchpoint, Tmp(), instance); |
955 | } |
956 | } |
957 | |
958 | void AirIRGenerator::emitThrowException(CCallHelpers& jit, ExceptionType type) |
959 | { |
960 | jit.move(CCallHelpers::TrustedImm32(static_cast<uint32_t>(type)), GPRInfo::argumentGPR1); |
961 | auto jumpToExceptionStub = jit.jump(); |
962 | |
963 | jit.addLinkTask([jumpToExceptionStub] (LinkBuffer& linkBuffer) { |
964 | linkBuffer.link(jumpToExceptionStub, CodeLocationLabel<JITThunkPtrTag>(Thunks::singleton().stub(throwExceptionFromWasmThunkGenerator).code())); |
965 | }); |
966 | } |
967 | |
968 | auto AirIRGenerator::addLocal(Type type, uint32_t count) -> PartialResult |
969 | { |
970 | size_t newSize = m_locals.size() + count; |
971 | ASSERT(!(CheckedUint32(count) + m_locals.size()).hasOverflowed()); |
972 | ASSERT(newSize <= maxFunctionLocals); |
973 | WASM_COMPILE_FAIL_IF(!m_locals.tryReserveCapacity(newSize), "can't allocate memory for " , newSize, " locals" ); |
974 | |
975 | for (uint32_t i = 0; i < count; ++i) { |
976 | auto local = tmpForType(type); |
977 | m_locals.uncheckedAppend(local); |
978 | switch (type) { |
979 | case Type::Anyref: |
980 | case Type::Funcref: |
981 | append(Move, Arg::imm(JSValue::encode(jsNull())), local); |
982 | break; |
983 | case Type::I32: |
984 | case Type::I64: { |
985 | append(Xor64, local, local); |
986 | break; |
987 | } |
988 | case Type::F32: |
989 | case Type::F64: { |
990 | auto temp = g64(); |
991 | // IEEE 754 "0" is just int32/64 zero. |
992 | append(Xor64, temp, temp); |
993 | append(type == Type::F32 ? Move32ToFloat : Move64ToDouble, temp, local); |
994 | break; |
995 | } |
996 | default: |
997 | RELEASE_ASSERT_NOT_REACHED(); |
998 | } |
999 | } |
1000 | return { }; |
1001 | } |
1002 | |
1003 | auto AirIRGenerator::addConstant(Type type, uint64_t value) -> ExpressionType |
1004 | { |
1005 | return addConstant(m_currentBlock, type, value); |
1006 | } |
1007 | |
1008 | auto AirIRGenerator::addConstant(BasicBlock* block, Type type, uint64_t value) -> ExpressionType |
1009 | { |
1010 | auto result = tmpForType(type); |
1011 | switch (type) { |
1012 | case Type::I32: |
1013 | case Type::I64: |
1014 | case Type::Anyref: |
1015 | case Type::Funcref: |
1016 | append(block, Move, Arg::bigImm(value), result); |
1017 | break; |
1018 | case Type::F32: |
1019 | case Type::F64: { |
1020 | auto tmp = g64(); |
1021 | append(block, Move, Arg::bigImm(value), tmp); |
1022 | append(block, type == Type::F32 ? Move32ToFloat : Move64ToDouble, tmp, result); |
1023 | break; |
1024 | } |
1025 | |
1026 | default: |
1027 | RELEASE_ASSERT_NOT_REACHED(); |
1028 | } |
1029 | |
1030 | return result; |
1031 | } |
1032 | |
1033 | auto AirIRGenerator::addBottom(BasicBlock* block, Type type) -> ExpressionType |
1034 | { |
1035 | append(block, B3::Air::Oops); |
1036 | return addConstant(type, 0); |
1037 | } |
1038 | |
1039 | auto AirIRGenerator::addArguments(const Signature& signature) -> PartialResult |
1040 | { |
1041 | RELEASE_ASSERT(m_locals.size() == signature.argumentCount()); // We handle arguments in the prologue |
1042 | return { }; |
1043 | } |
1044 | |
1045 | auto AirIRGenerator::addRefIsNull(ExpressionType& value, ExpressionType& result) -> PartialResult |
1046 | { |
1047 | ASSERT(value.tmp()); |
1048 | result = tmpForType(Type::I32); |
1049 | auto tmp = g64(); |
1050 | |
1051 | append(Move, Arg::bigImm(JSValue::encode(jsNull())), tmp); |
1052 | append(Compare64, Arg::relCond(MacroAssembler::Equal), value, tmp, result); |
1053 | |
1054 | return { }; |
1055 | } |
1056 | |
1057 | auto AirIRGenerator::addRefFunc(uint32_t index, ExpressionType& result) -> PartialResult |
1058 | { |
1059 | // FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>. |
1060 | result = tmpForType(Type::Funcref); |
1061 | emitCCall(&operationWasmRefFunc, result, instanceValue(), addConstant(Type::I32, index)); |
1062 | |
1063 | return { }; |
1064 | } |
1065 | |
1066 | auto AirIRGenerator::addTableGet(unsigned tableIndex, ExpressionType& index, ExpressionType& result) -> PartialResult |
1067 | { |
1068 | // FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>. |
1069 | ASSERT(index.tmp()); |
1070 | ASSERT(index.type() == Type::I32); |
1071 | result = tmpForType(m_info.tables[tableIndex].wasmType()); |
1072 | |
1073 | emitCCall(&operationGetWasmTableElement, result, instanceValue(), addConstant(Type::I32, tableIndex), index); |
1074 | emitCheck([&] { |
1075 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::Zero), result, result); |
1076 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1077 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTableAccess); |
1078 | }); |
1079 | |
1080 | return { }; |
1081 | } |
1082 | |
1083 | auto AirIRGenerator::addTableSet(unsigned tableIndex, ExpressionType& index, ExpressionType& value) -> PartialResult |
1084 | { |
1085 | // FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>. |
1086 | ASSERT(index.tmp()); |
1087 | ASSERT(index.type() == Type::I32); |
1088 | ASSERT(value.tmp()); |
1089 | |
1090 | auto shouldThrow = g32(); |
1091 | emitCCall(&operationSetWasmTableElement, shouldThrow, instanceValue(), addConstant(Type::I32, tableIndex), index, value); |
1092 | |
1093 | emitCheck([&] { |
1094 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::Zero), shouldThrow, shouldThrow); |
1095 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1096 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTableAccess); |
1097 | }); |
1098 | |
1099 | return { }; |
1100 | } |
1101 | |
1102 | auto AirIRGenerator::addTableSize(unsigned tableIndex, ExpressionType& result) -> PartialResult |
1103 | { |
1104 | // FIXME: Emit this inline <https://bugs.webkit.org/show_bug.cgi?id=198506>. |
1105 | result = tmpForType(Type::I32); |
1106 | |
1107 | emitCCall(&operationGetWasmTableSize, result, instanceValue(), addConstant(Type::I32, tableIndex)); |
1108 | |
1109 | return { }; |
1110 | } |
1111 | |
1112 | auto AirIRGenerator::addTableGrow(unsigned tableIndex, ExpressionType& fill, ExpressionType& delta, ExpressionType& result) -> PartialResult |
1113 | { |
1114 | ASSERT(fill.tmp()); |
1115 | ASSERT(isSubtype(fill.type(), m_info.tables[tableIndex].wasmType())); |
1116 | ASSERT(delta.tmp()); |
1117 | ASSERT(delta.type() == Type::I32); |
1118 | result = tmpForType(Type::I32); |
1119 | |
1120 | emitCCall(&operationWasmTableGrow, result, instanceValue(), addConstant(Type::I32, tableIndex), fill, delta); |
1121 | |
1122 | return { }; |
1123 | } |
1124 | |
1125 | auto AirIRGenerator::addTableFill(unsigned tableIndex, ExpressionType& offset, ExpressionType& fill, ExpressionType& count) -> PartialResult |
1126 | { |
1127 | ASSERT(fill.tmp()); |
1128 | ASSERT(isSubtype(fill.type(), m_info.tables[tableIndex].wasmType())); |
1129 | ASSERT(offset.tmp()); |
1130 | ASSERT(offset.type() == Type::I32); |
1131 | ASSERT(count.tmp()); |
1132 | ASSERT(count.type() == Type::I32); |
1133 | |
1134 | auto result = tmpForType(Type::I32); |
1135 | emitCCall(&operationWasmTableFill, result, instanceValue(), addConstant(Type::I32, tableIndex), offset, fill, count); |
1136 | |
1137 | emitCheck([&] { |
1138 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::Zero), result, result); |
1139 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1140 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTableAccess); |
1141 | }); |
1142 | |
1143 | return { }; |
1144 | } |
1145 | |
1146 | auto AirIRGenerator::getLocal(uint32_t index, ExpressionType& result) -> PartialResult |
1147 | { |
1148 | ASSERT(m_locals[index].tmp()); |
1149 | result = tmpForType(m_locals[index].type()); |
1150 | append(moveOpForValueType(m_locals[index].type()), m_locals[index].tmp(), result); |
1151 | return { }; |
1152 | } |
1153 | |
1154 | auto AirIRGenerator::addUnreachable() -> PartialResult |
1155 | { |
1156 | B3::PatchpointValue* unreachable = addPatchpoint(B3::Void); |
1157 | unreachable->setGenerator([this] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1158 | this->emitThrowException(jit, ExceptionType::Unreachable); |
1159 | }); |
1160 | unreachable->effects.terminal = true; |
1161 | emitPatchpoint(unreachable, Tmp()); |
1162 | return { }; |
1163 | } |
1164 | |
1165 | auto AirIRGenerator::addGrowMemory(ExpressionType delta, ExpressionType& result) -> PartialResult |
1166 | { |
1167 | result = g32(); |
1168 | emitCCall(&operationGrowMemory, result, TypedTmp { Tmp(GPRInfo::callFrameRegister), Type::I64 }, instanceValue(), delta); |
1169 | restoreWebAssemblyGlobalState(RestoreCachedStackLimit::No, m_info.memory, instanceValue(), m_currentBlock); |
1170 | |
1171 | return { }; |
1172 | } |
1173 | |
1174 | auto AirIRGenerator::addCurrentMemory(ExpressionType& result) -> PartialResult |
1175 | { |
1176 | static_assert(sizeof(decltype(static_cast<Memory*>(nullptr)->size())) == sizeof(uint64_t), "codegen relies on this size" ); |
1177 | |
1178 | auto temp1 = g64(); |
1179 | auto temp2 = g64(); |
1180 | |
1181 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfCachedMemorySize(), B3::Width64)); |
1182 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfCachedMemorySize()), temp1); |
1183 | constexpr uint32_t shiftValue = 16; |
1184 | static_assert(PageCount::pageSize == 1ull << shiftValue, "This must hold for the code below to be correct." ); |
1185 | append(Move, Arg::imm(16), temp2); |
1186 | addShift(Type::I32, Urshift64, temp1, temp2, result); |
1187 | append(Move32, result, result); |
1188 | |
1189 | return { }; |
1190 | } |
1191 | |
1192 | auto AirIRGenerator::setLocal(uint32_t index, ExpressionType value) -> PartialResult |
1193 | { |
1194 | ASSERT(m_locals[index].tmp()); |
1195 | append(moveOpForValueType(m_locals[index].type()), value, m_locals[index].tmp()); |
1196 | return { }; |
1197 | } |
1198 | |
1199 | auto AirIRGenerator::getGlobal(uint32_t index, ExpressionType& result) -> PartialResult |
1200 | { |
1201 | Type type = m_info.globals[index].type; |
1202 | |
1203 | result = tmpForType(type); |
1204 | |
1205 | auto temp = g64(); |
1206 | |
1207 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfGlobals(), B3::Width64)); |
1208 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfGlobals()), temp); |
1209 | |
1210 | int32_t offset = safeCast<int32_t>(index * sizeof(Register)); |
1211 | if (Arg::isValidAddrForm(offset, B3::widthForType(toB3Type(type)))) |
1212 | append(moveOpForValueType(type), Arg::addr(temp, offset), result); |
1213 | else { |
1214 | auto temp2 = g64(); |
1215 | append(Move, Arg::bigImm(offset), temp2); |
1216 | append(Add64, temp2, temp, temp); |
1217 | append(moveOpForValueType(type), Arg::addr(temp), result); |
1218 | } |
1219 | return { }; |
1220 | } |
1221 | |
1222 | auto AirIRGenerator::setGlobal(uint32_t index, ExpressionType value) -> PartialResult |
1223 | { |
1224 | auto temp = g64(); |
1225 | |
1226 | RELEASE_ASSERT(Arg::isValidAddrForm(Instance::offsetOfGlobals(), B3::Width64)); |
1227 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfGlobals()), temp); |
1228 | |
1229 | Type type = m_info.globals[index].type; |
1230 | |
1231 | int32_t offset = safeCast<int32_t>(index * sizeof(Register)); |
1232 | if (Arg::isValidAddrForm(offset, B3::widthForType(toB3Type(type)))) |
1233 | append(moveOpForValueType(type), value, Arg::addr(temp, offset)); |
1234 | else { |
1235 | auto temp2 = g64(); |
1236 | append(Move, Arg::bigImm(offset), temp2); |
1237 | append(Add64, temp2, temp, temp); |
1238 | append(moveOpForValueType(type), value, Arg::addr(temp)); |
1239 | } |
1240 | |
1241 | if (isSubtype(type, Anyref)) |
1242 | emitWriteBarrierForJSWrapper(); |
1243 | |
1244 | return { }; |
1245 | } |
1246 | |
1247 | inline void AirIRGenerator::emitWriteBarrierForJSWrapper() |
1248 | { |
1249 | auto cell = g64(); |
1250 | auto vm = g64(); |
1251 | auto cellState = g32(); |
1252 | auto threshold = g32(); |
1253 | |
1254 | BasicBlock* fenceCheckPath = m_code.addBlock(); |
1255 | BasicBlock* fencePath = m_code.addBlock(); |
1256 | BasicBlock* doSlowPath = m_code.addBlock(); |
1257 | BasicBlock* continuation = m_code.addBlock(); |
1258 | |
1259 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfOwner()), cell); |
1260 | append(Move, Arg::addr(cell, JSWebAssemblyInstance::offsetOfVM()), vm); |
1261 | append(Load8, Arg::addr(cell, JSCell::cellStateOffset()), cellState); |
1262 | append(Move32, Arg::addr(vm, VM::offsetOfHeapBarrierThreshold()), threshold); |
1263 | |
1264 | append(Branch32, Arg::relCond(MacroAssembler::Above), cellState, threshold); |
1265 | m_currentBlock->setSuccessors(continuation, fenceCheckPath); |
1266 | m_currentBlock = fenceCheckPath; |
1267 | |
1268 | append(Load8, Arg::addr(vm, VM::offsetOfHeapMutatorShouldBeFenced()), threshold); |
1269 | append(BranchTest32, Arg::resCond(MacroAssembler::Zero), threshold, threshold); |
1270 | m_currentBlock->setSuccessors(doSlowPath, fencePath); |
1271 | m_currentBlock = fencePath; |
1272 | |
1273 | auto* doFence = addPatchpoint(B3::Void); |
1274 | doFence->setGenerator([] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1275 | jit.memoryFence(); |
1276 | }); |
1277 | emitPatchpoint(doFence, Tmp()); |
1278 | |
1279 | append(Load8, Arg::addr(cell, JSCell::cellStateOffset()), cellState); |
1280 | append(Branch32, Arg::relCond(MacroAssembler::Above), cellState, Arg::imm(blackThreshold)); |
1281 | m_currentBlock->setSuccessors(continuation, doSlowPath); |
1282 | m_currentBlock = doSlowPath; |
1283 | |
1284 | emitCCall(&operationWasmWriteBarrierSlowPath, TypedTmp(), cell, vm); |
1285 | append(Jump); |
1286 | m_currentBlock->setSuccessors(continuation); |
1287 | m_currentBlock = continuation; |
1288 | } |
1289 | |
1290 | inline AirIRGenerator::ExpressionType AirIRGenerator::emitCheckAndPreparePointer(ExpressionType pointer, uint32_t offset, uint32_t sizeOfOperation) |
1291 | { |
1292 | ASSERT(m_memoryBaseGPR); |
1293 | |
1294 | auto result = g64(); |
1295 | append(Move32, pointer, result); |
1296 | |
1297 | switch (m_mode) { |
1298 | case MemoryMode::BoundsChecking: { |
1299 | // We're not using signal handling at all, we must therefore check that no memory access exceeds the current memory size. |
1300 | ASSERT(m_memorySizeGPR); |
1301 | ASSERT(sizeOfOperation + offset > offset); |
1302 | auto temp = g64(); |
1303 | append(Move, Arg::bigImm(static_cast<uint64_t>(sizeOfOperation) + offset - 1), temp); |
1304 | append(Add64, result, temp); |
1305 | |
1306 | emitCheck([&] { |
1307 | return Inst(Branch64, nullptr, Arg::relCond(MacroAssembler::AboveOrEqual), temp, Tmp(m_memorySizeGPR)); |
1308 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1309 | this->emitThrowException(jit, ExceptionType::OutOfBoundsMemoryAccess); |
1310 | }); |
1311 | break; |
1312 | } |
1313 | |
1314 | case MemoryMode::Signaling: { |
1315 | // We've virtually mapped 4GiB+redzone for this memory. Only the user-allocated pages are addressable, contiguously in range [0, current], |
1316 | // and everything above is mapped PROT_NONE. We don't need to perform any explicit bounds check in the 4GiB range because WebAssembly register |
1317 | // memory accesses are 32-bit. However WebAssembly register + offset accesses perform the addition in 64-bit which can push an access above |
1318 | // the 32-bit limit (the offset is unsigned 32-bit). The redzone will catch most small offsets, and we'll explicitly bounds check any |
1319 | // register + large offset access. We don't think this will be generated frequently. |
1320 | // |
1321 | // We could check that register + large offset doesn't exceed 4GiB+redzone since that's technically the limit we need to avoid overflowing the |
1322 | // 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 |
1323 | // than the declared 'maximum' will trap, so we can compare against that number. If there was no declared 'maximum' then we still know that |
1324 | // any access equal to or greater than 4GiB will trap, no need to add the redzone. |
1325 | if (offset >= Memory::fastMappedRedzoneBytes()) { |
1326 | uint64_t maximum = m_info.memory.maximum() ? m_info.memory.maximum().bytes() : std::numeric_limits<uint32_t>::max(); |
1327 | auto temp = g64(); |
1328 | append(Move, Arg::bigImm(static_cast<uint64_t>(sizeOfOperation) + offset - 1), temp); |
1329 | append(Add64, result, temp); |
1330 | auto sizeMax = addConstant(Type::I64, maximum); |
1331 | |
1332 | emitCheck([&] { |
1333 | return Inst(Branch64, nullptr, Arg::relCond(MacroAssembler::AboveOrEqual), temp, sizeMax); |
1334 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1335 | this->emitThrowException(jit, ExceptionType::OutOfBoundsMemoryAccess); |
1336 | }); |
1337 | } |
1338 | break; |
1339 | } |
1340 | } |
1341 | |
1342 | append(Add64, Tmp(m_memoryBaseGPR), result); |
1343 | return result; |
1344 | } |
1345 | |
1346 | inline uint32_t sizeOfLoadOp(LoadOpType op) |
1347 | { |
1348 | switch (op) { |
1349 | case LoadOpType::I32Load8S: |
1350 | case LoadOpType::I32Load8U: |
1351 | case LoadOpType::I64Load8S: |
1352 | case LoadOpType::I64Load8U: |
1353 | return 1; |
1354 | case LoadOpType::I32Load16S: |
1355 | case LoadOpType::I64Load16S: |
1356 | case LoadOpType::I32Load16U: |
1357 | case LoadOpType::I64Load16U: |
1358 | return 2; |
1359 | case LoadOpType::I32Load: |
1360 | case LoadOpType::I64Load32S: |
1361 | case LoadOpType::I64Load32U: |
1362 | case LoadOpType::F32Load: |
1363 | return 4; |
1364 | case LoadOpType::I64Load: |
1365 | case LoadOpType::F64Load: |
1366 | return 8; |
1367 | } |
1368 | RELEASE_ASSERT_NOT_REACHED(); |
1369 | } |
1370 | |
1371 | inline TypedTmp AirIRGenerator::emitLoadOp(LoadOpType op, ExpressionType pointer, uint32_t uoffset) |
1372 | { |
1373 | uint32_t offset = fixupPointerPlusOffset(pointer, uoffset); |
1374 | |
1375 | TypedTmp immTmp; |
1376 | TypedTmp newPtr; |
1377 | TypedTmp result; |
1378 | |
1379 | Arg addrArg; |
1380 | if (Arg::isValidAddrForm(offset, B3::widthForBytes(sizeOfLoadOp(op)))) |
1381 | addrArg = Arg::addr(pointer, offset); |
1382 | else { |
1383 | immTmp = g64(); |
1384 | newPtr = g64(); |
1385 | append(Move, Arg::bigImm(offset), immTmp); |
1386 | append(Add64, immTmp, pointer, newPtr); |
1387 | addrArg = Arg::addr(newPtr); |
1388 | } |
1389 | |
1390 | switch (op) { |
1391 | case LoadOpType::I32Load8S: { |
1392 | result = g32(); |
1393 | appendEffectful(Load8SignedExtendTo32, addrArg, result); |
1394 | break; |
1395 | } |
1396 | |
1397 | case LoadOpType::I64Load8S: { |
1398 | result = g64(); |
1399 | appendEffectful(Load8SignedExtendTo32, addrArg, result); |
1400 | append(SignExtend32ToPtr, result, result); |
1401 | break; |
1402 | } |
1403 | |
1404 | case LoadOpType::I32Load8U: { |
1405 | result = g32(); |
1406 | appendEffectful(Load8, addrArg, result); |
1407 | break; |
1408 | } |
1409 | |
1410 | case LoadOpType::I64Load8U: { |
1411 | result = g64(); |
1412 | appendEffectful(Load8, addrArg, result); |
1413 | break; |
1414 | } |
1415 | |
1416 | case LoadOpType::I32Load16S: { |
1417 | result = g32(); |
1418 | appendEffectful(Load16SignedExtendTo32, addrArg, result); |
1419 | break; |
1420 | } |
1421 | |
1422 | case LoadOpType::I64Load16S: { |
1423 | result = g64(); |
1424 | appendEffectful(Load16SignedExtendTo32, addrArg, result); |
1425 | append(SignExtend32ToPtr, result, result); |
1426 | break; |
1427 | } |
1428 | |
1429 | case LoadOpType::I32Load16U: { |
1430 | result = g32(); |
1431 | appendEffectful(Load16, addrArg, result); |
1432 | break; |
1433 | } |
1434 | |
1435 | case LoadOpType::I64Load16U: { |
1436 | result = g64(); |
1437 | appendEffectful(Load16, addrArg, result); |
1438 | break; |
1439 | } |
1440 | |
1441 | case LoadOpType::I32Load: |
1442 | result = g32(); |
1443 | appendEffectful(Move32, addrArg, result); |
1444 | break; |
1445 | |
1446 | case LoadOpType::I64Load32U: { |
1447 | result = g64(); |
1448 | appendEffectful(Move32, addrArg, result); |
1449 | break; |
1450 | } |
1451 | |
1452 | case LoadOpType::I64Load32S: { |
1453 | result = g64(); |
1454 | appendEffectful(Move32, addrArg, result); |
1455 | append(SignExtend32ToPtr, result, result); |
1456 | break; |
1457 | } |
1458 | |
1459 | case LoadOpType::I64Load: { |
1460 | result = g64(); |
1461 | appendEffectful(Move, addrArg, result); |
1462 | break; |
1463 | } |
1464 | |
1465 | case LoadOpType::F32Load: { |
1466 | result = f32(); |
1467 | appendEffectful(MoveFloat, addrArg, result); |
1468 | break; |
1469 | } |
1470 | |
1471 | case LoadOpType::F64Load: { |
1472 | result = f64(); |
1473 | appendEffectful(MoveDouble, addrArg, result); |
1474 | break; |
1475 | } |
1476 | } |
1477 | |
1478 | return result; |
1479 | } |
1480 | |
1481 | auto AirIRGenerator::load(LoadOpType op, ExpressionType pointer, ExpressionType& result, uint32_t offset) -> PartialResult |
1482 | { |
1483 | ASSERT(pointer.tmp().isGP()); |
1484 | |
1485 | if (UNLIKELY(sumOverflows<uint32_t>(offset, sizeOfLoadOp(op)))) { |
1486 | // FIXME: Even though this is provably out of bounds, it's not a validation error, so we have to handle it |
1487 | // as a runtime exception. However, this may change: https://bugs.webkit.org/show_bug.cgi?id=166435 |
1488 | auto* patch = addPatchpoint(B3::Void); |
1489 | patch->setGenerator([this] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1490 | this->emitThrowException(jit, ExceptionType::OutOfBoundsMemoryAccess); |
1491 | }); |
1492 | emitPatchpoint(patch, Tmp()); |
1493 | |
1494 | // We won't reach here, so we just pick a random reg. |
1495 | switch (op) { |
1496 | case LoadOpType::I32Load8S: |
1497 | case LoadOpType::I32Load16S: |
1498 | case LoadOpType::I32Load: |
1499 | case LoadOpType::I32Load16U: |
1500 | case LoadOpType::I32Load8U: |
1501 | result = g32(); |
1502 | break; |
1503 | case LoadOpType::I64Load8S: |
1504 | case LoadOpType::I64Load8U: |
1505 | case LoadOpType::I64Load16S: |
1506 | case LoadOpType::I64Load32U: |
1507 | case LoadOpType::I64Load32S: |
1508 | case LoadOpType::I64Load: |
1509 | case LoadOpType::I64Load16U: |
1510 | result = g64(); |
1511 | break; |
1512 | case LoadOpType::F32Load: |
1513 | result = f32(); |
1514 | break; |
1515 | case LoadOpType::F64Load: |
1516 | result = f64(); |
1517 | break; |
1518 | } |
1519 | } else |
1520 | result = emitLoadOp(op, emitCheckAndPreparePointer(pointer, offset, sizeOfLoadOp(op)), offset); |
1521 | |
1522 | return { }; |
1523 | } |
1524 | |
1525 | inline uint32_t sizeOfStoreOp(StoreOpType op) |
1526 | { |
1527 | switch (op) { |
1528 | case StoreOpType::I32Store8: |
1529 | case StoreOpType::I64Store8: |
1530 | return 1; |
1531 | case StoreOpType::I32Store16: |
1532 | case StoreOpType::I64Store16: |
1533 | return 2; |
1534 | case StoreOpType::I32Store: |
1535 | case StoreOpType::I64Store32: |
1536 | case StoreOpType::F32Store: |
1537 | return 4; |
1538 | case StoreOpType::I64Store: |
1539 | case StoreOpType::F64Store: |
1540 | return 8; |
1541 | } |
1542 | RELEASE_ASSERT_NOT_REACHED(); |
1543 | } |
1544 | |
1545 | |
1546 | inline void AirIRGenerator::emitStoreOp(StoreOpType op, ExpressionType pointer, ExpressionType value, uint32_t uoffset) |
1547 | { |
1548 | uint32_t offset = fixupPointerPlusOffset(pointer, uoffset); |
1549 | |
1550 | TypedTmp immTmp; |
1551 | TypedTmp newPtr; |
1552 | |
1553 | Arg addrArg; |
1554 | if (Arg::isValidAddrForm(offset, B3::widthForBytes(sizeOfStoreOp(op)))) |
1555 | addrArg = Arg::addr(pointer, offset); |
1556 | else { |
1557 | immTmp = g64(); |
1558 | newPtr = g64(); |
1559 | append(Move, Arg::bigImm(offset), immTmp); |
1560 | append(Add64, immTmp, pointer, newPtr); |
1561 | addrArg = Arg::addr(newPtr); |
1562 | } |
1563 | |
1564 | switch (op) { |
1565 | case StoreOpType::I64Store8: |
1566 | case StoreOpType::I32Store8: |
1567 | append(Store8, value, addrArg); |
1568 | return; |
1569 | |
1570 | case StoreOpType::I64Store16: |
1571 | case StoreOpType::I32Store16: |
1572 | append(Store16, value, addrArg); |
1573 | return; |
1574 | |
1575 | case StoreOpType::I64Store32: |
1576 | case StoreOpType::I32Store: |
1577 | append(Move32, value, addrArg); |
1578 | return; |
1579 | |
1580 | case StoreOpType::I64Store: |
1581 | append(Move, value, addrArg); |
1582 | return; |
1583 | |
1584 | case StoreOpType::F32Store: |
1585 | append(MoveFloat, value, addrArg); |
1586 | return; |
1587 | |
1588 | case StoreOpType::F64Store: |
1589 | append(MoveDouble, value, addrArg); |
1590 | return; |
1591 | } |
1592 | |
1593 | RELEASE_ASSERT_NOT_REACHED(); |
1594 | } |
1595 | |
1596 | auto AirIRGenerator::store(StoreOpType op, ExpressionType pointer, ExpressionType value, uint32_t offset) -> PartialResult |
1597 | { |
1598 | ASSERT(pointer.tmp().isGP()); |
1599 | |
1600 | if (UNLIKELY(sumOverflows<uint32_t>(offset, sizeOfStoreOp(op)))) { |
1601 | // FIXME: Even though this is provably out of bounds, it's not a validation error, so we have to handle it |
1602 | // as a runtime exception. However, this may change: https://bugs.webkit.org/show_bug.cgi?id=166435 |
1603 | auto* throwException = addPatchpoint(B3::Void); |
1604 | throwException->setGenerator([this] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
1605 | this->emitThrowException(jit, ExceptionType::OutOfBoundsMemoryAccess); |
1606 | }); |
1607 | emitPatchpoint(throwException, Tmp()); |
1608 | } else |
1609 | emitStoreOp(op, emitCheckAndPreparePointer(pointer, offset, sizeOfStoreOp(op)), value, offset); |
1610 | |
1611 | return { }; |
1612 | } |
1613 | |
1614 | auto AirIRGenerator::addSelect(ExpressionType condition, ExpressionType nonZero, ExpressionType zero, ExpressionType& result) -> PartialResult |
1615 | { |
1616 | ASSERT(nonZero.type() == zero.type()); |
1617 | result = tmpForType(nonZero.type()); |
1618 | append(moveOpForValueType(nonZero.type()), nonZero, result); |
1619 | |
1620 | BasicBlock* isZero = m_code.addBlock(); |
1621 | BasicBlock* continuation = m_code.addBlock(); |
1622 | |
1623 | append(BranchTest32, Arg::resCond(MacroAssembler::Zero), condition, condition); |
1624 | m_currentBlock->setSuccessors(isZero, continuation); |
1625 | |
1626 | append(isZero, moveOpForValueType(zero.type()), zero, result); |
1627 | append(isZero, Jump); |
1628 | isZero->setSuccessors(continuation); |
1629 | |
1630 | m_currentBlock = continuation; |
1631 | |
1632 | return { }; |
1633 | } |
1634 | |
1635 | void AirIRGenerator::emitEntryTierUpCheck() |
1636 | { |
1637 | if (!m_tierUp) |
1638 | return; |
1639 | |
1640 | auto countdownPtr = g64(); |
1641 | |
1642 | append(Move, Arg::bigImm(reinterpret_cast<uint64_t>(&m_tierUp->m_counter)), countdownPtr); |
1643 | |
1644 | auto* patch = addPatchpoint(B3::Void); |
1645 | B3::Effects effects = B3::Effects::none(); |
1646 | effects.reads = B3::HeapRange::top(); |
1647 | effects.writes = B3::HeapRange::top(); |
1648 | patch->effects = effects; |
1649 | patch->clobber(RegisterSet::macroScratchRegisters()); |
1650 | |
1651 | patch->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
1652 | AllowMacroScratchRegisterUsage allowScratch(jit); |
1653 | |
1654 | CCallHelpers::Jump tierUp = jit.branchAdd32(CCallHelpers::PositiveOrZero, CCallHelpers::TrustedImm32(TierUpCount::functionEntryIncrement()), CCallHelpers::Address(params[0].gpr())); |
1655 | CCallHelpers::Label tierUpResume = jit.label(); |
1656 | |
1657 | params.addLatePath([=] (CCallHelpers& jit) { |
1658 | tierUp.link(&jit); |
1659 | |
1660 | const unsigned = 0; |
1661 | RegisterSet registersToSpill = { }; |
1662 | registersToSpill.add(GPRInfo::argumentGPR1); |
1663 | unsigned numberOfStackBytesUsedForRegisterPreservation = ScratchRegisterAllocator::preserveRegistersToStackForCall(jit, registersToSpill, extraPaddingBytes); |
1664 | |
1665 | jit.move(MacroAssembler::TrustedImm32(m_functionIndex), GPRInfo::argumentGPR1); |
1666 | MacroAssembler::Call call = jit.nearCall(); |
1667 | |
1668 | ScratchRegisterAllocator::restoreRegistersFromStackForCall(jit, registersToSpill, RegisterSet(), numberOfStackBytesUsedForRegisterPreservation, extraPaddingBytes); |
1669 | jit.jump(tierUpResume); |
1670 | |
1671 | jit.addLinkTask([=] (LinkBuffer& linkBuffer) { |
1672 | MacroAssembler::repatchNearCall(linkBuffer.locationOfNearCall<NoPtrTag>(call), CodeLocationLabel<JITThunkPtrTag>(Thunks::singleton().stub(triggerOMGEntryTierUpThunkGenerator).code())); |
1673 | }); |
1674 | }); |
1675 | }); |
1676 | |
1677 | emitPatchpoint(patch, Tmp(), countdownPtr); |
1678 | } |
1679 | |
1680 | void AirIRGenerator::emitLoopTierUpCheck(uint32_t loopIndex) |
1681 | { |
1682 | uint32_t outerLoopIndex = this->outerLoopIndex(); |
1683 | m_outerLoops.append(loopIndex); |
1684 | |
1685 | if (!m_tierUp) |
1686 | return; |
1687 | |
1688 | ASSERT(m_tierUp->osrEntryTriggers().size() == loopIndex); |
1689 | m_tierUp->osrEntryTriggers().append(TierUpCount::TriggerReason::DontTrigger); |
1690 | m_tierUp->outerLoops().append(outerLoopIndex); |
1691 | |
1692 | auto countdownPtr = g64(); |
1693 | |
1694 | append(Move, Arg::bigImm(reinterpret_cast<uint64_t>(&m_tierUp->m_counter)), countdownPtr); |
1695 | |
1696 | auto* patch = addPatchpoint(B3::Void); |
1697 | B3::Effects effects = B3::Effects::none(); |
1698 | effects.reads = B3::HeapRange::top(); |
1699 | effects.writes = B3::HeapRange::top(); |
1700 | effects.exitsSideways = true; |
1701 | patch->effects = effects; |
1702 | |
1703 | patch->clobber(RegisterSet::macroScratchRegisters()); |
1704 | RegisterSet clobberLate; |
1705 | clobberLate.add(GPRInfo::argumentGPR0); |
1706 | patch->clobberLate(clobberLate); |
1707 | |
1708 | Vector<ConstrainedTmp> patchArgs; |
1709 | patchArgs.append(countdownPtr); |
1710 | |
1711 | for (auto& local : m_locals) |
1712 | patchArgs.append(ConstrainedTmp(local, B3::ValueRep::ColdAny)); |
1713 | for (unsigned controlIndex = 0; controlIndex < m_parser->controlStack().size(); ++controlIndex) { |
1714 | ExpressionList& expressionStack = m_parser->controlStack()[controlIndex].enclosedExpressionStack; |
1715 | for (auto& value : expressionStack) |
1716 | patchArgs.append(ConstrainedTmp(value, B3::ValueRep::ColdAny)); |
1717 | } |
1718 | |
1719 | TierUpCount::TriggerReason* forceEntryTrigger = &(m_tierUp->osrEntryTriggers().last()); |
1720 | static_assert(!static_cast<uint8_t>(TierUpCount::TriggerReason::DontTrigger), "the JIT code assumes non-zero means 'enter'" ); |
1721 | static_assert(sizeof(TierUpCount::TriggerReason) == 1, "branchTest8 assumes this size" ); |
1722 | patch->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
1723 | AllowMacroScratchRegisterUsage allowScratch(jit); |
1724 | CCallHelpers::Jump forceOSREntry = jit.branchTest8(CCallHelpers::NonZero, CCallHelpers::AbsoluteAddress(forceEntryTrigger)); |
1725 | CCallHelpers::Jump tierUp = jit.branchAdd32(CCallHelpers::PositiveOrZero, CCallHelpers::TrustedImm32(TierUpCount::loopIncrement()), CCallHelpers::Address(params[0].gpr())); |
1726 | MacroAssembler::Label tierUpResume = jit.label(); |
1727 | |
1728 | OSREntryData& osrEntryData = m_tierUp->addOSREntryData(m_functionIndex, loopIndex); |
1729 | // First argument is the countdown location. |
1730 | for (unsigned index = 1; index < params.value()->numChildren(); ++index) |
1731 | osrEntryData.values().constructAndAppend(params[index], params.value()->child(index)->type()); |
1732 | OSREntryData* osrEntryDataPtr = &osrEntryData; |
1733 | |
1734 | params.addLatePath([=] (CCallHelpers& jit) { |
1735 | AllowMacroScratchRegisterUsage allowScratch(jit); |
1736 | forceOSREntry.link(&jit); |
1737 | tierUp.link(&jit); |
1738 | |
1739 | jit.probe(operationWasmTriggerOSREntryNow, osrEntryDataPtr); |
1740 | jit.branchTestPtr(CCallHelpers::Zero, GPRInfo::argumentGPR0).linkTo(tierUpResume, &jit); |
1741 | jit.farJump(GPRInfo::argumentGPR1, WasmEntryPtrTag); |
1742 | }); |
1743 | }); |
1744 | |
1745 | emitPatchpoint(m_currentBlock, patch, Vector<Tmp, 1>(), WTFMove(patchArgs)); |
1746 | } |
1747 | |
1748 | AirIRGenerator::ControlData AirIRGenerator::addTopLevel(BlockSignature signature) |
1749 | { |
1750 | return ControlData(B3::Origin(), signature, tmpsForSignature(signature), BlockType::TopLevel, m_code.addBlock()); |
1751 | } |
1752 | |
1753 | auto AirIRGenerator::addLoop(BlockSignature signature, Stack& enclosingStack, ControlType& block, Stack& newStack, uint32_t loopIndex) -> PartialResult |
1754 | { |
1755 | BasicBlock* body = m_code.addBlock(); |
1756 | BasicBlock* continuation = m_code.addBlock(); |
1757 | |
1758 | newStack = splitStack(signature, enclosingStack); |
1759 | block = ControlData(origin(), signature, newStack, BlockType::Loop, continuation, body); |
1760 | |
1761 | append(Jump); |
1762 | m_currentBlock->setSuccessors(body); |
1763 | |
1764 | m_currentBlock = body; |
1765 | emitLoopTierUpCheck(loopIndex); |
1766 | |
1767 | return { }; |
1768 | } |
1769 | |
1770 | auto AirIRGenerator::addBlock(BlockSignature signature, Stack& enclosingStack, ControlType& newBlock, Stack& newStack) -> PartialResult |
1771 | { |
1772 | newStack = splitStack(signature, enclosingStack); |
1773 | newBlock = ControlData(origin(), signature, tmpsForSignature(signature), BlockType::Block, m_code.addBlock()); |
1774 | return { }; |
1775 | } |
1776 | |
1777 | auto AirIRGenerator::addIf(ExpressionType condition, BlockSignature signature, Stack& enclosingStack, ControlType& result, Stack& newStack) -> PartialResult |
1778 | { |
1779 | BasicBlock* taken = m_code.addBlock(); |
1780 | BasicBlock* notTaken = m_code.addBlock(); |
1781 | BasicBlock* continuation = m_code.addBlock(); |
1782 | |
1783 | // Wasm bools are i32. |
1784 | append(BranchTest32, Arg::resCond(MacroAssembler::NonZero), condition, condition); |
1785 | m_currentBlock->setSuccessors(taken, notTaken); |
1786 | |
1787 | m_currentBlock = taken; |
1788 | newStack = splitStack(signature, enclosingStack); |
1789 | result = ControlData(origin(), signature, tmpsForSignature(signature), BlockType::If, continuation, notTaken); |
1790 | return { }; |
1791 | } |
1792 | |
1793 | auto AirIRGenerator::addElse(ControlData& data, const Stack& currentStack) -> PartialResult |
1794 | { |
1795 | unifyValuesWithBlock(currentStack, data.results); |
1796 | append(Jump); |
1797 | m_currentBlock->setSuccessors(data.continuation); |
1798 | return addElseToUnreachable(data); |
1799 | } |
1800 | |
1801 | auto AirIRGenerator::addElseToUnreachable(ControlData& data) -> PartialResult |
1802 | { |
1803 | ASSERT(data.blockType() == BlockType::If); |
1804 | m_currentBlock = data.special; |
1805 | data.convertIfToBlock(); |
1806 | return { }; |
1807 | } |
1808 | |
1809 | auto AirIRGenerator::addReturn(const ControlData& data, const ExpressionList& returnValues) -> PartialResult |
1810 | { |
1811 | CallInformation wasmCallInfo = wasmCallingConvention().callInformationFor(*data.signature(), CallRole::Callee); |
1812 | if (!wasmCallInfo.results.size()) { |
1813 | append(RetVoid); |
1814 | return { }; |
1815 | } |
1816 | |
1817 | B3::PatchpointValue* patch = addPatchpoint(B3::Void); |
1818 | patch->setGenerator([] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
1819 | auto calleeSaves = params.code().calleeSaveRegisterAtOffsetList(); |
1820 | |
1821 | for (RegisterAtOffset calleeSave : calleeSaves) |
1822 | jit.load64ToReg(CCallHelpers::Address(GPRInfo::callFrameRegister, calleeSave.offset()), calleeSave.reg()); |
1823 | |
1824 | jit.emitFunctionEpilogue(); |
1825 | jit.ret(); |
1826 | }); |
1827 | patch->effects.terminal = true; |
1828 | |
1829 | ASSERT(returnValues.size() >= wasmCallInfo.results.size()); |
1830 | unsigned offset = returnValues.size() - wasmCallInfo.results.size(); |
1831 | Vector<ConstrainedTmp, 8> returnConstraints; |
1832 | for (unsigned i = 0; i < wasmCallInfo.results.size(); ++i) { |
1833 | B3::ValueRep rep = wasmCallInfo.results[i]; |
1834 | TypedTmp tmp = returnValues[offset + i]; |
1835 | |
1836 | if (rep.isStack()) { |
1837 | append(moveForType(toB3Type(tmp.type())), tmp, Arg::addr(Tmp(GPRInfo::callFrameRegister), rep.offsetFromFP())); |
1838 | continue; |
1839 | } |
1840 | |
1841 | ASSERT(rep.isReg()); |
1842 | if (data.signature()->returnType(i) == I32) |
1843 | append(Move32, tmp, tmp); |
1844 | returnConstraints.append(ConstrainedTmp(tmp, wasmCallInfo.results[i])); |
1845 | } |
1846 | |
1847 | emitPatchpoint(m_currentBlock, patch, Vector<Tmp, 1>(), WTFMove(returnConstraints)); |
1848 | return { }; |
1849 | } |
1850 | |
1851 | // NOTE: All branches in Wasm are on 32-bit ints |
1852 | |
1853 | auto AirIRGenerator::addBranch(ControlData& data, ExpressionType condition, const Stack& returnValues) -> PartialResult |
1854 | { |
1855 | unifyValuesWithBlock(returnValues, data.results); |
1856 | |
1857 | BasicBlock* target = data.targetBlockForBranch(); |
1858 | if (condition) { |
1859 | BasicBlock* continuation = m_code.addBlock(); |
1860 | append(BranchTest32, Arg::resCond(MacroAssembler::NonZero), condition, condition); |
1861 | m_currentBlock->setSuccessors(target, continuation); |
1862 | m_currentBlock = continuation; |
1863 | } else { |
1864 | append(Jump); |
1865 | m_currentBlock->setSuccessors(target); |
1866 | } |
1867 | |
1868 | return { }; |
1869 | } |
1870 | |
1871 | auto AirIRGenerator::addSwitch(ExpressionType condition, const Vector<ControlData*>& targets, ControlData& defaultTarget, const Stack& expressionStack) -> PartialResult |
1872 | { |
1873 | auto& successors = m_currentBlock->successors(); |
1874 | ASSERT(successors.isEmpty()); |
1875 | for (const auto& target : targets) { |
1876 | unifyValuesWithBlock(expressionStack, target->results); |
1877 | successors.append(target->targetBlockForBranch()); |
1878 | } |
1879 | unifyValuesWithBlock(expressionStack, defaultTarget.results); |
1880 | successors.append(defaultTarget.targetBlockForBranch()); |
1881 | |
1882 | ASSERT(condition.type() == Type::I32); |
1883 | |
1884 | // FIXME: We should consider dynamically switching between a jump table |
1885 | // and a binary switch depending on the number of successors. |
1886 | // https://bugs.webkit.org/show_bug.cgi?id=194477 |
1887 | |
1888 | size_t numTargets = targets.size(); |
1889 | |
1890 | auto* patchpoint = addPatchpoint(B3::Void); |
1891 | patchpoint->effects = B3::Effects::none(); |
1892 | patchpoint->effects.terminal = true; |
1893 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
1894 | |
1895 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
1896 | AllowMacroScratchRegisterUsage allowScratch(jit); |
1897 | |
1898 | Vector<int64_t> cases; |
1899 | cases.reserveInitialCapacity(numTargets); |
1900 | for (size_t i = 0; i < numTargets; ++i) |
1901 | cases.uncheckedAppend(i); |
1902 | |
1903 | GPRReg valueReg = params[0].gpr(); |
1904 | BinarySwitch binarySwitch(valueReg, cases, BinarySwitch::Int32); |
1905 | |
1906 | Vector<CCallHelpers::Jump> caseJumps; |
1907 | caseJumps.resize(numTargets); |
1908 | |
1909 | while (binarySwitch.advance(jit)) { |
1910 | unsigned value = binarySwitch.caseValue(); |
1911 | unsigned index = binarySwitch.caseIndex(); |
1912 | ASSERT_UNUSED(value, value == index); |
1913 | ASSERT(index < numTargets); |
1914 | caseJumps[index] = jit.jump(); |
1915 | } |
1916 | |
1917 | CCallHelpers::JumpList fallThrough = binarySwitch.fallThrough(); |
1918 | |
1919 | Vector<Box<CCallHelpers::Label>> successorLabels = params.successorLabels(); |
1920 | ASSERT(successorLabels.size() == caseJumps.size() + 1); |
1921 | |
1922 | params.addLatePath([=, caseJumps = WTFMove(caseJumps), successorLabels = WTFMove(successorLabels)] (CCallHelpers& jit) { |
1923 | for (size_t i = 0; i < numTargets; ++i) |
1924 | caseJumps[i].linkTo(*successorLabels[i], &jit); |
1925 | fallThrough.linkTo(*successorLabels[numTargets], &jit); |
1926 | }); |
1927 | }); |
1928 | |
1929 | emitPatchpoint(patchpoint, TypedTmp(), condition); |
1930 | |
1931 | return { }; |
1932 | } |
1933 | |
1934 | auto AirIRGenerator::endBlock(ControlEntry& entry, Stack& expressionStack) -> PartialResult |
1935 | { |
1936 | ControlData& data = entry.controlData; |
1937 | |
1938 | if (data.blockType() != BlockType::Loop) |
1939 | unifyValuesWithBlock(expressionStack, data.results); |
1940 | append(Jump); |
1941 | m_currentBlock->setSuccessors(data.continuation); |
1942 | |
1943 | return addEndToUnreachable(entry, expressionStack); |
1944 | } |
1945 | |
1946 | |
1947 | auto AirIRGenerator::addEndToUnreachable(ControlEntry& entry, const Stack& expressionStack) -> PartialResult |
1948 | { |
1949 | ControlData& data = entry.controlData; |
1950 | m_currentBlock = data.continuation; |
1951 | |
1952 | if (data.blockType() == BlockType::If) { |
1953 | append(data.special, Jump); |
1954 | data.special->setSuccessors(m_currentBlock); |
1955 | } |
1956 | |
1957 | if (data.blockType() == BlockType::Loop) { |
1958 | m_outerLoops.removeLast(); |
1959 | for (unsigned i = 0; i < data.signature()->returnCount(); ++i) { |
1960 | if (i < expressionStack.size()) |
1961 | entry.enclosedExpressionStack.append(expressionStack[i]); |
1962 | else |
1963 | entry.enclosedExpressionStack.append(addBottom(m_currentBlock, data.signature()->returnType(i))); |
1964 | } |
1965 | } else { |
1966 | for (const auto& result : data.results) |
1967 | entry.enclosedExpressionStack.append(result); |
1968 | } |
1969 | |
1970 | // TopLevel does not have any code after this so we need to make sure we emit a return here. |
1971 | if (data.blockType() == BlockType::TopLevel) |
1972 | return addReturn(data, entry.enclosedExpressionStack); |
1973 | |
1974 | return { }; |
1975 | } |
1976 | |
1977 | B3::PatchpointValue* AirIRGenerator::emitCallPatchpoint(BasicBlock* block, const Signature& signature, const Vector<ExpressionType, 1>& results, const Vector<TypedTmp>& args, Vector<ConstrainedTmp>&& patchArgs) |
1978 | { |
1979 | auto* patchpoint = addPatchpoint(toB3ResultType(&signature)); |
1980 | patchpoint->effects.writesPinned = true; |
1981 | patchpoint->effects.readsPinned = true; |
1982 | patchpoint->clobberEarly(RegisterSet::macroScratchRegisters()); |
1983 | patchpoint->clobberLate(RegisterSet::volatileRegistersForJSCall()); |
1984 | |
1985 | CallInformation locations = wasmCallingConvention().callInformationFor(signature); |
1986 | m_code.requestCallArgAreaSizeInBytes(WTF::roundUpToMultipleOf(stackAlignmentBytes(), locations.headerAndArgumentStackSizeInBytes)); |
1987 | |
1988 | size_t offset = patchArgs.size(); |
1989 | Checked<size_t> newSize = checkedSum<size_t>(patchArgs.size(), args.size()); |
1990 | RELEASE_ASSERT(!newSize.hasOverflowed()); |
1991 | |
1992 | patchArgs.grow(newSize.unsafeGet()); |
1993 | for (unsigned i = 0; i < args.size(); ++i) |
1994 | patchArgs[i + offset] = ConstrainedTmp(args[i], locations.params[i]); |
1995 | |
1996 | if (patchpoint->type() != B3::Void) |
1997 | patchpoint->resultConstraints = WTFMove(locations.results); |
1998 | emitPatchpoint(block, patchpoint, results, WTFMove(patchArgs)); |
1999 | return patchpoint; |
2000 | } |
2001 | |
2002 | auto AirIRGenerator::addCall(uint32_t functionIndex, const Signature& signature, Vector<ExpressionType>& args, Vector<ExpressionType, 1>& results) -> PartialResult |
2003 | { |
2004 | ASSERT(signature.argumentCount() == args.size()); |
2005 | |
2006 | m_makesCalls = true; |
2007 | |
2008 | for (unsigned i = 0; i < signature.returnCount(); ++i) |
2009 | results.append(tmpForType(signature.returnType(i))); |
2010 | |
2011 | Vector<UnlinkedWasmToWasmCall>* unlinkedWasmToWasmCalls = &m_unlinkedWasmToWasmCalls; |
2012 | |
2013 | if (m_info.isImportedFunctionFromFunctionIndexSpace(functionIndex)) { |
2014 | m_maxNumJSCallArguments = std::max(m_maxNumJSCallArguments, static_cast<uint32_t>(args.size())); |
2015 | |
2016 | auto currentInstance = g64(); |
2017 | append(Move, instanceValue(), currentInstance); |
2018 | |
2019 | auto targetInstance = g64(); |
2020 | |
2021 | // FIXME: We should have better isel here. |
2022 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
2023 | append(Move, Arg::bigImm(Instance::offsetOfTargetInstance(functionIndex)), targetInstance); |
2024 | append(Add64, instanceValue(), targetInstance); |
2025 | append(Move, Arg::addr(targetInstance), targetInstance); |
2026 | |
2027 | BasicBlock* isWasmBlock = m_code.addBlock(); |
2028 | BasicBlock* isEmbedderBlock = m_code.addBlock(); |
2029 | BasicBlock* continuation = m_code.addBlock(); |
2030 | |
2031 | append(BranchTest64, Arg::resCond(MacroAssembler::NonZero), targetInstance, targetInstance); |
2032 | m_currentBlock->setSuccessors(isWasmBlock, isEmbedderBlock); |
2033 | |
2034 | { |
2035 | auto* patchpoint = emitCallPatchpoint(isWasmBlock, signature, results, args); |
2036 | // We need to clobber all potential pinned registers since we might be leaving the instance. |
2037 | // We pessimistically assume we could be calling to something that is bounds checking. |
2038 | // FIXME: We shouldn't have to do this: https://bugs.webkit.org/show_bug.cgi?id=172181 |
2039 | patchpoint->clobberLate(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking)); |
2040 | |
2041 | patchpoint->setGenerator([unlinkedWasmToWasmCalls, functionIndex] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2042 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2043 | CCallHelpers::Call call = jit.threadSafePatchableNearCall(); |
2044 | jit.addLinkTask([unlinkedWasmToWasmCalls, call, functionIndex] (LinkBuffer& linkBuffer) { |
2045 | unlinkedWasmToWasmCalls->append({ linkBuffer.locationOfNearCall<WasmEntryPtrTag>(call), functionIndex }); |
2046 | }); |
2047 | }); |
2048 | |
2049 | append(isWasmBlock, Jump); |
2050 | isWasmBlock->setSuccessors(continuation); |
2051 | } |
2052 | |
2053 | { |
2054 | auto jumpDestination = g64(); |
2055 | append(isEmbedderBlock, Move, Arg::bigImm(Instance::offsetOfWasmToEmbedderStub(functionIndex)), jumpDestination); |
2056 | append(isEmbedderBlock, Add64, instanceValue(), jumpDestination); |
2057 | append(isEmbedderBlock, Move, Arg::addr(jumpDestination), jumpDestination); |
2058 | |
2059 | Vector<ConstrainedTmp> jumpArgs; |
2060 | jumpArgs.append({ jumpDestination, B3::ValueRep::SomeRegister }); |
2061 | auto* patchpoint = emitCallPatchpoint(isEmbedderBlock, signature, results, args, WTFMove(jumpArgs)); |
2062 | // We need to clobber all potential pinned registers since we might be leaving the instance. |
2063 | // We pessimistically assume we could be calling to something that is bounds checking. |
2064 | // FIXME: We shouldn't have to do this: https://bugs.webkit.org/show_bug.cgi?id=172181 |
2065 | patchpoint->clobberLate(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking)); |
2066 | patchpoint->setGenerator([] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2067 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2068 | jit.call(params[params.proc().resultCount(params.value()->type())].gpr(), WasmEntryPtrTag); |
2069 | }); |
2070 | |
2071 | append(isEmbedderBlock, Jump); |
2072 | isEmbedderBlock->setSuccessors(continuation); |
2073 | } |
2074 | |
2075 | m_currentBlock = continuation; |
2076 | // The call could have been to another WebAssembly instance, and / or could have modified our Memory. |
2077 | restoreWebAssemblyGlobalState(RestoreCachedStackLimit::Yes, m_info.memory, currentInstance, continuation); |
2078 | } else { |
2079 | auto* patchpoint = emitCallPatchpoint(m_currentBlock, signature, results, args); |
2080 | patchpoint->setGenerator([unlinkedWasmToWasmCalls, functionIndex] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2081 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2082 | CCallHelpers::Call call = jit.threadSafePatchableNearCall(); |
2083 | jit.addLinkTask([unlinkedWasmToWasmCalls, call, functionIndex] (LinkBuffer& linkBuffer) { |
2084 | unlinkedWasmToWasmCalls->append({ linkBuffer.locationOfNearCall<WasmEntryPtrTag>(call), functionIndex }); |
2085 | }); |
2086 | }); |
2087 | } |
2088 | |
2089 | return { }; |
2090 | } |
2091 | |
2092 | auto AirIRGenerator::addCallIndirect(unsigned tableIndex, const Signature& signature, Vector<ExpressionType>& args, Vector<ExpressionType, 1>& results) -> PartialResult |
2093 | { |
2094 | ExpressionType calleeIndex = args.takeLast(); |
2095 | ASSERT(signature.argumentCount() == args.size()); |
2096 | ASSERT(m_info.tableCount() > tableIndex); |
2097 | ASSERT(m_info.tables[tableIndex].type() == TableElementType::Funcref); |
2098 | |
2099 | m_makesCalls = true; |
2100 | // Note: call indirect can call either WebAssemblyFunction or WebAssemblyWrapperFunction. Because |
2101 | // WebAssemblyWrapperFunction is like calling into the embedder, we conservatively assume all call indirects |
2102 | // can be to the embedder for our stack check calculation. |
2103 | m_maxNumJSCallArguments = std::max(m_maxNumJSCallArguments, static_cast<uint32_t>(args.size())); |
2104 | |
2105 | auto currentInstance = g64(); |
2106 | append(Move, instanceValue(), currentInstance); |
2107 | |
2108 | ExpressionType callableFunctionBuffer = g64(); |
2109 | ExpressionType instancesBuffer = g64(); |
2110 | ExpressionType callableFunctionBufferLength = g64(); |
2111 | { |
2112 | RELEASE_ASSERT(Arg::isValidAddrForm(FuncRefTable::offsetOfFunctions(), B3::Width64)); |
2113 | RELEASE_ASSERT(Arg::isValidAddrForm(FuncRefTable::offsetOfInstances(), B3::Width64)); |
2114 | RELEASE_ASSERT(Arg::isValidAddrForm(FuncRefTable::offsetOfLength(), B3::Width64)); |
2115 | |
2116 | if (UNLIKELY(!Arg::isValidAddrForm(Instance::offsetOfTablePtr(m_numImportFunctions, tableIndex), B3::Width64))) { |
2117 | append(Move, Arg::bigImm(Instance::offsetOfTablePtr(m_numImportFunctions, tableIndex)), callableFunctionBufferLength); |
2118 | append(Add64, instanceValue(), callableFunctionBufferLength); |
2119 | append(Move, Arg::addr(callableFunctionBufferLength), callableFunctionBufferLength); |
2120 | } else |
2121 | append(Move, Arg::addr(instanceValue(), Instance::offsetOfTablePtr(m_numImportFunctions, tableIndex)), callableFunctionBufferLength); |
2122 | append(Move, Arg::addr(callableFunctionBufferLength, FuncRefTable::offsetOfFunctions()), callableFunctionBuffer); |
2123 | append(Move, Arg::addr(callableFunctionBufferLength, FuncRefTable::offsetOfInstances()), instancesBuffer); |
2124 | append(Move32, Arg::addr(callableFunctionBufferLength, Table::offsetOfLength()), callableFunctionBufferLength); |
2125 | } |
2126 | |
2127 | append(Move32, calleeIndex, calleeIndex); |
2128 | |
2129 | // Check the index we are looking for is valid. |
2130 | emitCheck([&] { |
2131 | return Inst(Branch32, nullptr, Arg::relCond(MacroAssembler::AboveOrEqual), calleeIndex, callableFunctionBufferLength); |
2132 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2133 | this->emitThrowException(jit, ExceptionType::OutOfBoundsCallIndirect); |
2134 | }); |
2135 | |
2136 | ExpressionType calleeCode = g64(); |
2137 | { |
2138 | ExpressionType calleeSignatureIndex = g64(); |
2139 | // Compute the offset in the table index space we are looking for. |
2140 | append(Move, Arg::imm(sizeof(WasmToWasmImportableFunction)), calleeSignatureIndex); |
2141 | append(Mul64, calleeIndex, calleeSignatureIndex); |
2142 | append(Add64, callableFunctionBuffer, calleeSignatureIndex); |
2143 | |
2144 | append(Move, Arg::addr(calleeSignatureIndex, WasmToWasmImportableFunction::offsetOfEntrypointLoadLocation()), calleeCode); // Pointer to callee code. |
2145 | |
2146 | // Check that the WasmToWasmImportableFunction is initialized. We trap if it isn't. An "invalid" SignatureIndex indicates it's not initialized. |
2147 | // 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 |
2148 | static_assert(sizeof(WasmToWasmImportableFunction::signatureIndex) == sizeof(uint64_t), "Load codegen assumes i64" ); |
2149 | |
2150 | // FIXME: This seems dumb to do two checks just for a nicer error message. |
2151 | // We should move just to use a single branch and then figure out what |
2152 | // error to use in the exception handler. |
2153 | |
2154 | append(Move, Arg::addr(calleeSignatureIndex, WasmToWasmImportableFunction::offsetOfSignatureIndex()), calleeSignatureIndex); |
2155 | |
2156 | emitCheck([&] { |
2157 | static_assert(Signature::invalidIndex == 0, "" ); |
2158 | return Inst(BranchTest64, nullptr, Arg::resCond(MacroAssembler::Zero), calleeSignatureIndex, calleeSignatureIndex); |
2159 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2160 | this->emitThrowException(jit, ExceptionType::NullTableEntry); |
2161 | }); |
2162 | |
2163 | ExpressionType expectedSignatureIndex = g64(); |
2164 | append(Move, Arg::bigImm(SignatureInformation::get(signature)), expectedSignatureIndex); |
2165 | emitCheck([&] { |
2166 | return Inst(Branch64, nullptr, Arg::relCond(MacroAssembler::NotEqual), calleeSignatureIndex, expectedSignatureIndex); |
2167 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2168 | this->emitThrowException(jit, ExceptionType::BadSignature); |
2169 | }); |
2170 | } |
2171 | |
2172 | // Do a context switch if needed. |
2173 | { |
2174 | auto newContextInstance = g64(); |
2175 | append(Move, Arg::index(instancesBuffer, calleeIndex, 8, 0), newContextInstance); |
2176 | |
2177 | BasicBlock* doContextSwitch = m_code.addBlock(); |
2178 | BasicBlock* continuation = m_code.addBlock(); |
2179 | |
2180 | append(Branch64, Arg::relCond(MacroAssembler::Equal), newContextInstance, instanceValue()); |
2181 | m_currentBlock->setSuccessors(continuation, doContextSwitch); |
2182 | |
2183 | auto* patchpoint = addPatchpoint(B3::Void); |
2184 | patchpoint->effects.writesPinned = true; |
2185 | // We pessimistically assume we're calling something with BoundsChecking memory. |
2186 | // FIXME: We shouldn't have to do this: https://bugs.webkit.org/show_bug.cgi?id=172181 |
2187 | patchpoint->clobber(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking)); |
2188 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2189 | patchpoint->numGPScratchRegisters = Gigacage::isEnabled(Gigacage::Primitive) ? 1 : 0; |
2190 | |
2191 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2192 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2193 | GPRReg newContextInstance = params[0].gpr(); |
2194 | GPRReg oldContextInstance = params[1].gpr(); |
2195 | const PinnedRegisterInfo& pinnedRegs = PinnedRegisterInfo::get(); |
2196 | GPRReg baseMemory = pinnedRegs.baseMemoryPointer; |
2197 | ASSERT(newContextInstance != baseMemory); |
2198 | jit.loadPtr(CCallHelpers::Address(oldContextInstance, Instance::offsetOfCachedStackLimit()), baseMemory); |
2199 | jit.storePtr(baseMemory, CCallHelpers::Address(newContextInstance, Instance::offsetOfCachedStackLimit())); |
2200 | jit.storeWasmContextInstance(newContextInstance); |
2201 | // FIXME: We should support more than one memory size register |
2202 | // see: https://bugs.webkit.org/show_bug.cgi?id=162952 |
2203 | ASSERT(pinnedRegs.sizeRegister != newContextInstance); |
2204 | GPRReg scratchOrSize = Gigacage::isEnabled(Gigacage::Primitive) ? params.gpScratch(0) : pinnedRegs.sizeRegister; |
2205 | |
2206 | jit.loadPtr(CCallHelpers::Address(newContextInstance, Instance::offsetOfCachedMemorySize()), pinnedRegs.sizeRegister); // Memory size. |
2207 | jit.loadPtr(CCallHelpers::Address(newContextInstance, Instance::offsetOfCachedMemory()), baseMemory); // Memory::void*. |
2208 | |
2209 | jit.cageConditionally(Gigacage::Primitive, baseMemory, pinnedRegs.sizeRegister, scratchOrSize); |
2210 | }); |
2211 | |
2212 | emitPatchpoint(doContextSwitch, patchpoint, Tmp(), newContextInstance, instanceValue()); |
2213 | append(doContextSwitch, Jump); |
2214 | doContextSwitch->setSuccessors(continuation); |
2215 | |
2216 | m_currentBlock = continuation; |
2217 | } |
2218 | |
2219 | append(Move, Arg::addr(calleeCode), calleeCode); |
2220 | |
2221 | Vector<ConstrainedTmp> ; |
2222 | extraArgs.append(calleeCode); |
2223 | |
2224 | for (unsigned i = 0; i < signature.returnCount(); ++i) |
2225 | results.append(tmpForType(signature.returnType(i))); |
2226 | |
2227 | auto* patchpoint = emitCallPatchpoint(m_currentBlock, signature, results, args, WTFMove(extraArgs)); |
2228 | |
2229 | // We need to clobber all potential pinned registers since we might be leaving the instance. |
2230 | // We pessimistically assume we're always calling something that is bounds checking so |
2231 | // because the wasm->wasm thunk unconditionally overrides the size registers. |
2232 | // FIXME: We should not have to do this, but the wasm->wasm stub assumes it can |
2233 | // use all the pinned registers as scratch: https://bugs.webkit.org/show_bug.cgi?id=172181 |
2234 | |
2235 | patchpoint->clobberLate(PinnedRegisterInfo::get().toSave(MemoryMode::BoundsChecking)); |
2236 | |
2237 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2238 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2239 | jit.call(params[params.proc().resultCount(params.value()->type())].gpr(), WasmEntryPtrTag); |
2240 | }); |
2241 | |
2242 | // The call could have been to another WebAssembly instance, and / or could have modified our Memory. |
2243 | restoreWebAssemblyGlobalState(RestoreCachedStackLimit::Yes, m_info.memory, currentInstance, m_currentBlock); |
2244 | |
2245 | return { }; |
2246 | } |
2247 | |
2248 | void AirIRGenerator::unify(const ExpressionType& dst, const ExpressionType& source) |
2249 | { |
2250 | ASSERT(isSubtype(source.type(), dst.type())); |
2251 | append(moveOpForValueType(dst.type()), source, dst); |
2252 | } |
2253 | |
2254 | void AirIRGenerator::unifyValuesWithBlock(const Stack& resultStack, const ResultList& result) |
2255 | { |
2256 | ASSERT(result.size() <= resultStack.size()); |
2257 | |
2258 | for (size_t i = 0; i < result.size(); ++i) |
2259 | unify(result[result.size() - 1 - i], resultStack[resultStack.size() - 1 - i]); |
2260 | } |
2261 | |
2262 | static void dumpExpressionStack(const CommaPrinter& comma, const AirIRGenerator::Stack& expressionStack) |
2263 | { |
2264 | dataLog(comma, "ExpressionStack:" ); |
2265 | for (const auto& expression : expressionStack) |
2266 | dataLog(comma, expression); |
2267 | } |
2268 | |
2269 | void AirIRGenerator::dump(const Vector<ControlEntry>& controlStack, const Stack* stack) |
2270 | { |
2271 | dataLogLn("Processing Graph:" ); |
2272 | dataLog(m_code); |
2273 | dataLogLn("With current block:" , *m_currentBlock); |
2274 | dataLogLn("Control stack:" ); |
2275 | for (size_t i = controlStack.size(); i--;) { |
2276 | dataLog(" " , controlStack[i].controlData, ": " ); |
2277 | CommaPrinter comma(", " , "" ); |
2278 | dumpExpressionStack(comma, *stack); |
2279 | stack = &controlStack[i].enclosedExpressionStack; |
2280 | dataLogLn(); |
2281 | } |
2282 | dataLogLn("\n" ); |
2283 | } |
2284 | |
2285 | auto AirIRGenerator::origin() -> B3::Origin |
2286 | { |
2287 | // FIXME: We should implement a way to give Inst's an origin. |
2288 | return B3::Origin(); |
2289 | } |
2290 | |
2291 | Expected<std::unique_ptr<InternalFunction>, String> parseAndCompileAir(CompilationContext& compilationContext, const FunctionData& function, const Signature& signature, Vector<UnlinkedWasmToWasmCall>& unlinkedWasmToWasmCalls, const ModuleInformation& info, MemoryMode mode, uint32_t functionIndex, TierUpCount* tierUp, ThrowWasmException throwWasmException) |
2292 | { |
2293 | auto result = makeUnique<InternalFunction>(); |
2294 | |
2295 | compilationContext.embedderEntrypointJIT = makeUnique<CCallHelpers>(); |
2296 | compilationContext.wasmEntrypointJIT = makeUnique<CCallHelpers>(); |
2297 | |
2298 | B3::Procedure procedure; |
2299 | Code& code = procedure.code(); |
2300 | |
2301 | procedure.setOriginPrinter([] (PrintStream& out, B3::Origin origin) { |
2302 | if (origin.data()) |
2303 | out.print("Wasm: " , bitwise_cast<OpcodeOrigin>(origin)); |
2304 | }); |
2305 | |
2306 | // This means we cannot use either StackmapGenerationParams::usedRegisters() or |
2307 | // StackmapGenerationParams::unavailableRegisters(). In exchange for this concession, we |
2308 | // don't strictly need to run Air::reportUsedRegisters(), which saves a bit of CPU time at |
2309 | // optLevel=1. |
2310 | procedure.setNeedsUsedRegisters(false); |
2311 | |
2312 | procedure.setOptLevel(Options::webAssemblyBBQAirOptimizationLevel()); |
2313 | |
2314 | AirIRGenerator irGenerator(info, procedure, result.get(), unlinkedWasmToWasmCalls, mode, functionIndex, tierUp, throwWasmException, signature); |
2315 | FunctionParser<AirIRGenerator> parser(irGenerator, function.data.data(), function.data.size(), signature, info); |
2316 | WASM_FAIL_IF_HELPER_FAILS(parser.parse()); |
2317 | |
2318 | |
2319 | for (BasicBlock* block : code) { |
2320 | for (size_t i = 0; i < block->numSuccessors(); ++i) |
2321 | block->successorBlock(i)->addPredecessor(block); |
2322 | } |
2323 | |
2324 | { |
2325 | if (UNLIKELY(shouldDumpIRAtEachPhase(B3::AirMode))) { |
2326 | dataLogLn("Generated patchpoints" ); |
2327 | for (B3::PatchpointValue** patch : irGenerator.patchpoints()) |
2328 | dataLogLn(deepDump(procedure, *patch)); |
2329 | } |
2330 | |
2331 | B3::Air::prepareForGeneration(code); |
2332 | B3::Air::generate(code, *compilationContext.wasmEntrypointJIT); |
2333 | compilationContext.wasmEntrypointByproducts = procedure.releaseByproducts(); |
2334 | result->entrypoint.calleeSaveRegisters = code.calleeSaveRegisterAtOffsetList(); |
2335 | } |
2336 | |
2337 | return result; |
2338 | } |
2339 | |
2340 | template <typename IntType> |
2341 | void AirIRGenerator::emitChecksForModOrDiv(bool isSignedDiv, ExpressionType left, ExpressionType right) |
2342 | { |
2343 | static_assert(sizeof(IntType) == 4 || sizeof(IntType) == 8, "" ); |
2344 | |
2345 | emitCheck([&] { |
2346 | return Inst(sizeof(IntType) == 4 ? BranchTest32 : BranchTest64, nullptr, Arg::resCond(MacroAssembler::Zero), right, right); |
2347 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2348 | this->emitThrowException(jit, ExceptionType::DivisionByZero); |
2349 | }); |
2350 | |
2351 | if (isSignedDiv) { |
2352 | ASSERT(std::is_signed<IntType>::value); |
2353 | IntType min = std::numeric_limits<IntType>::min(); |
2354 | |
2355 | // FIXME: Better isel for compare with imms here. |
2356 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
2357 | auto minTmp = sizeof(IntType) == 4 ? g32() : g64(); |
2358 | auto negOne = sizeof(IntType) == 4 ? g32() : g64(); |
2359 | |
2360 | B3::Air::Opcode op = sizeof(IntType) == 4 ? Compare32 : Compare64; |
2361 | append(Move, Arg::bigImm(static_cast<uint64_t>(min)), minTmp); |
2362 | append(op, Arg::relCond(MacroAssembler::Equal), left, minTmp, minTmp); |
2363 | |
2364 | append(Move, Arg::imm(-1), negOne); |
2365 | append(op, Arg::relCond(MacroAssembler::Equal), right, negOne, negOne); |
2366 | |
2367 | emitCheck([&] { |
2368 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), minTmp, negOne); |
2369 | }, |
2370 | [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2371 | this->emitThrowException(jit, ExceptionType::IntegerOverflow); |
2372 | }); |
2373 | } |
2374 | } |
2375 | |
2376 | template <typename IntType> |
2377 | void AirIRGenerator::emitModOrDiv(bool isDiv, ExpressionType lhs, ExpressionType rhs, ExpressionType& result) |
2378 | { |
2379 | static_assert(sizeof(IntType) == 4 || sizeof(IntType) == 8, "" ); |
2380 | |
2381 | result = sizeof(IntType) == 4 ? g32() : g64(); |
2382 | |
2383 | bool isSigned = std::is_signed<IntType>::value; |
2384 | |
2385 | if (isARM64()) { |
2386 | B3::Air::Opcode div; |
2387 | switch (sizeof(IntType)) { |
2388 | case 4: |
2389 | div = isSigned ? Div32 : UDiv32; |
2390 | break; |
2391 | case 8: |
2392 | div = isSigned ? Div64 : UDiv64; |
2393 | break; |
2394 | } |
2395 | |
2396 | append(div, lhs, rhs, result); |
2397 | |
2398 | if (!isDiv) { |
2399 | append(sizeof(IntType) == 4 ? Mul32 : Mul64, result, rhs, result); |
2400 | append(sizeof(IntType) == 4 ? Sub32 : Sub64, lhs, result, result); |
2401 | } |
2402 | |
2403 | return; |
2404 | } |
2405 | |
2406 | #if CPU(X86_64) |
2407 | Tmp eax(X86Registers::eax); |
2408 | Tmp edx(X86Registers::edx); |
2409 | |
2410 | if (isSigned) { |
2411 | B3::Air::Opcode convertToDoubleWord; |
2412 | B3::Air::Opcode div; |
2413 | switch (sizeof(IntType)) { |
2414 | case 4: |
2415 | convertToDoubleWord = X86ConvertToDoubleWord32; |
2416 | div = X86Div32; |
2417 | break; |
2418 | case 8: |
2419 | convertToDoubleWord = X86ConvertToQuadWord64; |
2420 | div = X86Div64; |
2421 | break; |
2422 | default: |
2423 | RELEASE_ASSERT_NOT_REACHED(); |
2424 | } |
2425 | |
2426 | // We implement "res = Div<Chill>/Mod<Chill>(num, den)" as follows: |
2427 | // |
2428 | // if (den + 1 <=_unsigned 1) { |
2429 | // if (!den) { |
2430 | // res = 0; |
2431 | // goto done; |
2432 | // } |
2433 | // if (num == -2147483648) { |
2434 | // res = isDiv ? num : 0; |
2435 | // goto done; |
2436 | // } |
2437 | // } |
2438 | // res = num (/ or %) dev; |
2439 | // done: |
2440 | |
2441 | BasicBlock* denIsGood = m_code.addBlock(); |
2442 | BasicBlock* denMayBeBad = m_code.addBlock(); |
2443 | BasicBlock* denNotZero = m_code.addBlock(); |
2444 | BasicBlock* continuation = m_code.addBlock(); |
2445 | |
2446 | auto temp = sizeof(IntType) == 4 ? g32() : g64(); |
2447 | auto one = addConstant(sizeof(IntType) == 4 ? Type::I32 : Type::I64, 1); |
2448 | |
2449 | append(sizeof(IntType) == 4 ? Add32 : Add64, rhs, one, temp); |
2450 | append(sizeof(IntType) == 4 ? Branch32 : Branch64, Arg::relCond(MacroAssembler::Above), temp, one); |
2451 | m_currentBlock->setSuccessors(denIsGood, denMayBeBad); |
2452 | |
2453 | append(denMayBeBad, Xor64, result, result); |
2454 | append(denMayBeBad, sizeof(IntType) == 4 ? BranchTest32 : BranchTest64, Arg::resCond(MacroAssembler::Zero), rhs, rhs); |
2455 | denMayBeBad->setSuccessors(continuation, denNotZero); |
2456 | |
2457 | auto min = addConstant(denNotZero, sizeof(IntType) == 4 ? Type::I32 : Type::I64, std::numeric_limits<IntType>::min()); |
2458 | if (isDiv) |
2459 | append(denNotZero, sizeof(IntType) == 4 ? Move32 : Move, min, result); |
2460 | else { |
2461 | // Result is zero, as set above... |
2462 | } |
2463 | append(denNotZero, sizeof(IntType) == 4 ? Branch32 : Branch64, Arg::relCond(MacroAssembler::Equal), lhs, min); |
2464 | denNotZero->setSuccessors(continuation, denIsGood); |
2465 | |
2466 | auto divResult = isDiv ? eax : edx; |
2467 | append(denIsGood, Move, lhs, eax); |
2468 | append(denIsGood, convertToDoubleWord, eax, edx); |
2469 | append(denIsGood, div, eax, edx, rhs); |
2470 | append(denIsGood, sizeof(IntType) == 4 ? Move32 : Move, divResult, result); |
2471 | append(denIsGood, Jump); |
2472 | denIsGood->setSuccessors(continuation); |
2473 | |
2474 | m_currentBlock = continuation; |
2475 | return; |
2476 | } |
2477 | |
2478 | B3::Air::Opcode div = sizeof(IntType) == 4 ? X86UDiv32 : X86UDiv64; |
2479 | |
2480 | Tmp divResult = isDiv ? eax : edx; |
2481 | |
2482 | append(Move, lhs, eax); |
2483 | append(Xor64, edx, edx); |
2484 | append(div, eax, edx, rhs); |
2485 | append(sizeof(IntType) == 4 ? Move32 : Move, divResult, result); |
2486 | #else |
2487 | RELEASE_ASSERT_NOT_REACHED(); |
2488 | #endif |
2489 | } |
2490 | |
2491 | template<> |
2492 | auto AirIRGenerator::addOp<OpType::I32DivS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2493 | { |
2494 | emitChecksForModOrDiv<int32_t>(true, left, right); |
2495 | emitModOrDiv<int32_t>(true, left, right, result); |
2496 | return { }; |
2497 | } |
2498 | |
2499 | template<> |
2500 | auto AirIRGenerator::addOp<OpType::I32RemS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2501 | { |
2502 | emitChecksForModOrDiv<int32_t>(false, left, right); |
2503 | emitModOrDiv<int32_t>(false, left, right, result); |
2504 | return { }; |
2505 | } |
2506 | |
2507 | template<> |
2508 | auto AirIRGenerator::addOp<OpType::I32DivU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2509 | { |
2510 | emitChecksForModOrDiv<uint32_t>(false, left, right); |
2511 | emitModOrDiv<uint32_t>(true, left, right, result); |
2512 | return { }; |
2513 | } |
2514 | |
2515 | template<> |
2516 | auto AirIRGenerator::addOp<OpType::I32RemU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2517 | { |
2518 | emitChecksForModOrDiv<uint32_t>(false, left, right); |
2519 | emitModOrDiv<uint32_t>(false, left, right, result); |
2520 | return { }; |
2521 | } |
2522 | |
2523 | template<> |
2524 | auto AirIRGenerator::addOp<OpType::I64DivS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2525 | { |
2526 | emitChecksForModOrDiv<int64_t>(true, left, right); |
2527 | emitModOrDiv<int64_t>(true, left, right, result); |
2528 | return { }; |
2529 | } |
2530 | |
2531 | template<> |
2532 | auto AirIRGenerator::addOp<OpType::I64RemS>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2533 | { |
2534 | emitChecksForModOrDiv<int64_t>(false, left, right); |
2535 | emitModOrDiv<int64_t>(false, left, right, result); |
2536 | return { }; |
2537 | } |
2538 | |
2539 | template<> |
2540 | auto AirIRGenerator::addOp<OpType::I64DivU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2541 | { |
2542 | emitChecksForModOrDiv<uint64_t>(false, left, right); |
2543 | emitModOrDiv<uint64_t>(true, left, right, result); |
2544 | return { }; |
2545 | } |
2546 | |
2547 | template<> |
2548 | auto AirIRGenerator::addOp<OpType::I64RemU>(ExpressionType left, ExpressionType right, ExpressionType& result) -> PartialResult |
2549 | { |
2550 | emitChecksForModOrDiv<uint64_t>(false, left, right); |
2551 | emitModOrDiv<uint64_t>(false, left, right, result); |
2552 | return { }; |
2553 | } |
2554 | |
2555 | template<> |
2556 | auto AirIRGenerator::addOp<OpType::I32Ctz>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2557 | { |
2558 | auto* patchpoint = addPatchpoint(B3::Int32); |
2559 | patchpoint->effects = B3::Effects::none(); |
2560 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2561 | jit.countTrailingZeros32(params[1].gpr(), params[0].gpr()); |
2562 | }); |
2563 | result = g32(); |
2564 | emitPatchpoint(patchpoint, result, arg); |
2565 | return { }; |
2566 | } |
2567 | |
2568 | template<> |
2569 | auto AirIRGenerator::addOp<OpType::I64Ctz>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2570 | { |
2571 | auto* patchpoint = addPatchpoint(B3::Int64); |
2572 | patchpoint->effects = B3::Effects::none(); |
2573 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2574 | jit.countTrailingZeros64(params[1].gpr(), params[0].gpr()); |
2575 | }); |
2576 | result = g64(); |
2577 | emitPatchpoint(patchpoint, result, arg); |
2578 | return { }; |
2579 | } |
2580 | |
2581 | template<> |
2582 | auto AirIRGenerator::addOp<OpType::I32Popcnt>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2583 | { |
2584 | result = g32(); |
2585 | |
2586 | #if CPU(X86_64) |
2587 | if (MacroAssembler::supportsCountPopulation()) { |
2588 | auto* patchpoint = addPatchpoint(B3::Int32); |
2589 | patchpoint->effects = B3::Effects::none(); |
2590 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2591 | jit.countPopulation32(params[1].gpr(), params[0].gpr()); |
2592 | }); |
2593 | emitPatchpoint(patchpoint, result, arg); |
2594 | return { }; |
2595 | } |
2596 | #endif |
2597 | |
2598 | emitCCall(&operationPopcount32, result, arg); |
2599 | return { }; |
2600 | } |
2601 | |
2602 | template<> |
2603 | auto AirIRGenerator::addOp<OpType::I64Popcnt>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2604 | { |
2605 | result = g64(); |
2606 | |
2607 | #if CPU(X86_64) |
2608 | if (MacroAssembler::supportsCountPopulation()) { |
2609 | auto* patchpoint = addPatchpoint(B3::Int64); |
2610 | patchpoint->effects = B3::Effects::none(); |
2611 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2612 | jit.countPopulation64(params[1].gpr(), params[0].gpr()); |
2613 | }); |
2614 | emitPatchpoint(patchpoint, result, arg); |
2615 | return { }; |
2616 | } |
2617 | #endif |
2618 | |
2619 | emitCCall(&operationPopcount64, result, arg); |
2620 | return { }; |
2621 | } |
2622 | |
2623 | template<> |
2624 | auto AirIRGenerator::addOp<F64ConvertUI64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2625 | { |
2626 | auto* patchpoint = addPatchpoint(B3::Double); |
2627 | patchpoint->effects = B3::Effects::none(); |
2628 | if (isX86()) |
2629 | patchpoint->numGPScratchRegisters = 1; |
2630 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2631 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2632 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2633 | #if CPU(X86_64) |
2634 | jit.convertUInt64ToDouble(params[1].gpr(), params[0].fpr(), params.gpScratch(0)); |
2635 | #else |
2636 | jit.convertUInt64ToDouble(params[1].gpr(), params[0].fpr()); |
2637 | #endif |
2638 | }); |
2639 | result = f64(); |
2640 | emitPatchpoint(patchpoint, result, arg); |
2641 | return { }; |
2642 | } |
2643 | |
2644 | template<> |
2645 | auto AirIRGenerator::addOp<OpType::F32ConvertUI64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2646 | { |
2647 | auto* patchpoint = addPatchpoint(B3::Float); |
2648 | patchpoint->effects = B3::Effects::none(); |
2649 | if (isX86()) |
2650 | patchpoint->numGPScratchRegisters = 1; |
2651 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2652 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2653 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2654 | #if CPU(X86_64) |
2655 | jit.convertUInt64ToFloat(params[1].gpr(), params[0].fpr(), params.gpScratch(0)); |
2656 | #else |
2657 | jit.convertUInt64ToFloat(params[1].gpr(), params[0].fpr()); |
2658 | #endif |
2659 | }); |
2660 | result = f32(); |
2661 | emitPatchpoint(patchpoint, result, arg); |
2662 | return { }; |
2663 | } |
2664 | |
2665 | template<> |
2666 | auto AirIRGenerator::addOp<OpType::F64Nearest>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2667 | { |
2668 | auto* patchpoint = addPatchpoint(B3::Double); |
2669 | patchpoint->effects = B3::Effects::none(); |
2670 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2671 | jit.roundTowardNearestIntDouble(params[1].fpr(), params[0].fpr()); |
2672 | }); |
2673 | result = f64(); |
2674 | emitPatchpoint(patchpoint, result, arg); |
2675 | return { }; |
2676 | } |
2677 | |
2678 | template<> |
2679 | auto AirIRGenerator::addOp<OpType::F32Nearest>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2680 | { |
2681 | auto* patchpoint = addPatchpoint(B3::Float); |
2682 | patchpoint->effects = B3::Effects::none(); |
2683 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2684 | jit.roundTowardNearestIntFloat(params[1].fpr(), params[0].fpr()); |
2685 | }); |
2686 | result = f32(); |
2687 | emitPatchpoint(patchpoint, result, arg); |
2688 | return { }; |
2689 | } |
2690 | |
2691 | template<> |
2692 | auto AirIRGenerator::addOp<OpType::F64Trunc>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2693 | { |
2694 | auto* patchpoint = addPatchpoint(B3::Double); |
2695 | patchpoint->effects = B3::Effects::none(); |
2696 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2697 | jit.roundTowardZeroDouble(params[1].fpr(), params[0].fpr()); |
2698 | }); |
2699 | result = f64(); |
2700 | emitPatchpoint(patchpoint, result, arg); |
2701 | return { }; |
2702 | } |
2703 | |
2704 | template<> |
2705 | auto AirIRGenerator::addOp<OpType::F32Trunc>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2706 | { |
2707 | auto* patchpoint = addPatchpoint(B3::Float); |
2708 | patchpoint->effects = B3::Effects::none(); |
2709 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2710 | jit.roundTowardZeroFloat(params[1].fpr(), params[0].fpr()); |
2711 | }); |
2712 | result = f32(); |
2713 | emitPatchpoint(patchpoint, result, arg); |
2714 | return { }; |
2715 | } |
2716 | |
2717 | template<> |
2718 | auto AirIRGenerator::addOp<OpType::I32TruncSF64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2719 | { |
2720 | auto max = addConstant(Type::F64, bitwise_cast<uint64_t>(-static_cast<double>(std::numeric_limits<int32_t>::min()))); |
2721 | auto min = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int32_t>::min()))); |
2722 | |
2723 | auto temp1 = g32(); |
2724 | auto temp2 = g32(); |
2725 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrUnordered), arg, min, temp1); |
2726 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2727 | append(Or32, temp1, temp2); |
2728 | |
2729 | emitCheck([&] { |
2730 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2731 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2732 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2733 | }); |
2734 | |
2735 | auto* patchpoint = addPatchpoint(B3::Int32); |
2736 | patchpoint->effects = B3::Effects::none(); |
2737 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2738 | jit.truncateDoubleToInt32(params[1].fpr(), params[0].gpr()); |
2739 | }); |
2740 | result = g32(); |
2741 | emitPatchpoint(patchpoint, result, arg); |
2742 | |
2743 | return { }; |
2744 | } |
2745 | |
2746 | template<> |
2747 | auto AirIRGenerator::addOp<OpType::I32TruncSF32>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2748 | { |
2749 | auto max = addConstant(Type::F32, bitwise_cast<uint32_t>(-static_cast<float>(std::numeric_limits<int32_t>::min()))); |
2750 | auto min = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int32_t>::min()))); |
2751 | |
2752 | auto temp1 = g32(); |
2753 | auto temp2 = g32(); |
2754 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrUnordered), arg, min, temp1); |
2755 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2756 | append(Or32, temp1, temp2); |
2757 | |
2758 | emitCheck([&] { |
2759 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2760 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2761 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2762 | }); |
2763 | |
2764 | auto* patchpoint = addPatchpoint(B3::Int32); |
2765 | patchpoint->effects = B3::Effects::none(); |
2766 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2767 | jit.truncateFloatToInt32(params[1].fpr(), params[0].gpr()); |
2768 | }); |
2769 | result = g32(); |
2770 | emitPatchpoint(patchpoint, result, arg); |
2771 | return { }; |
2772 | } |
2773 | |
2774 | |
2775 | template<> |
2776 | auto AirIRGenerator::addOp<OpType::I32TruncUF64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2777 | { |
2778 | auto max = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int32_t>::min()) * -2.0)); |
2779 | auto min = addConstant(Type::F64, bitwise_cast<uint64_t>(-1.0)); |
2780 | |
2781 | auto temp1 = g32(); |
2782 | auto temp2 = g32(); |
2783 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqualOrUnordered), arg, min, temp1); |
2784 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2785 | append(Or32, temp1, temp2); |
2786 | |
2787 | emitCheck([&] { |
2788 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2789 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2790 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2791 | }); |
2792 | |
2793 | auto* patchpoint = addPatchpoint(B3::Int32); |
2794 | patchpoint->effects = B3::Effects::none(); |
2795 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2796 | jit.truncateDoubleToUint32(params[1].fpr(), params[0].gpr()); |
2797 | }); |
2798 | result = g32(); |
2799 | emitPatchpoint(patchpoint, result, arg); |
2800 | return { }; |
2801 | } |
2802 | |
2803 | template<> |
2804 | auto AirIRGenerator::addOp<OpType::I32TruncUF32>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2805 | { |
2806 | auto max = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int32_t>::min()) * static_cast<float>(-2.0))); |
2807 | auto min = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(-1.0))); |
2808 | |
2809 | auto temp1 = g32(); |
2810 | auto temp2 = g32(); |
2811 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqualOrUnordered), arg, min, temp1); |
2812 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2813 | append(Or32, temp1, temp2); |
2814 | |
2815 | emitCheck([&] { |
2816 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2817 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2818 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2819 | }); |
2820 | |
2821 | auto* patchpoint = addPatchpoint(B3::Int32); |
2822 | patchpoint->effects = B3::Effects::none(); |
2823 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2824 | jit.truncateFloatToUint32(params[1].fpr(), params[0].gpr()); |
2825 | }); |
2826 | result = g32(); |
2827 | emitPatchpoint(patchpoint, result, arg); |
2828 | return { }; |
2829 | } |
2830 | |
2831 | template<> |
2832 | auto AirIRGenerator::addOp<OpType::I64TruncSF64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2833 | { |
2834 | auto max = addConstant(Type::F64, bitwise_cast<uint64_t>(-static_cast<double>(std::numeric_limits<int64_t>::min()))); |
2835 | auto min = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int64_t>::min()))); |
2836 | |
2837 | auto temp1 = g32(); |
2838 | auto temp2 = g32(); |
2839 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrUnordered), arg, min, temp1); |
2840 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2841 | append(Or32, temp1, temp2); |
2842 | |
2843 | emitCheck([&] { |
2844 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2845 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2846 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2847 | }); |
2848 | |
2849 | auto* patchpoint = addPatchpoint(B3::Int64); |
2850 | patchpoint->effects = B3::Effects::none(); |
2851 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2852 | jit.truncateDoubleToInt64(params[1].fpr(), params[0].gpr()); |
2853 | }); |
2854 | |
2855 | result = g64(); |
2856 | emitPatchpoint(patchpoint, result, arg); |
2857 | return { }; |
2858 | } |
2859 | |
2860 | template<> |
2861 | auto AirIRGenerator::addOp<OpType::I64TruncUF64>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2862 | { |
2863 | auto max = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<int64_t>::min()) * -2.0)); |
2864 | auto min = addConstant(Type::F64, bitwise_cast<uint64_t>(-1.0)); |
2865 | |
2866 | auto temp1 = g32(); |
2867 | auto temp2 = g32(); |
2868 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqualOrUnordered), arg, min, temp1); |
2869 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2870 | append(Or32, temp1, temp2); |
2871 | |
2872 | emitCheck([&] { |
2873 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2874 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2875 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2876 | }); |
2877 | |
2878 | TypedTmp signBitConstant; |
2879 | if (isX86()) |
2880 | signBitConstant = addConstant(Type::F64, bitwise_cast<uint64_t>(static_cast<double>(std::numeric_limits<uint64_t>::max() - std::numeric_limits<int64_t>::max()))); |
2881 | |
2882 | Vector<ConstrainedTmp> args; |
2883 | auto* patchpoint = addPatchpoint(B3::Int64); |
2884 | patchpoint->effects = B3::Effects::none(); |
2885 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2886 | args.append(arg); |
2887 | if (isX86()) { |
2888 | args.append(signBitConstant); |
2889 | patchpoint->numFPScratchRegisters = 1; |
2890 | } |
2891 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2892 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2893 | FPRReg scratch = InvalidFPRReg; |
2894 | FPRReg constant = InvalidFPRReg; |
2895 | if (isX86()) { |
2896 | scratch = params.fpScratch(0); |
2897 | constant = params[2].fpr(); |
2898 | } |
2899 | jit.truncateDoubleToUint64(params[1].fpr(), params[0].gpr(), scratch, constant); |
2900 | }); |
2901 | |
2902 | result = g64(); |
2903 | emitPatchpoint(m_currentBlock, patchpoint, Vector<Tmp, 1> { result }, WTFMove(args)); |
2904 | return { }; |
2905 | } |
2906 | |
2907 | template<> |
2908 | auto AirIRGenerator::addOp<OpType::I64TruncSF32>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2909 | { |
2910 | auto max = addConstant(Type::F32, bitwise_cast<uint32_t>(-static_cast<float>(std::numeric_limits<int64_t>::min()))); |
2911 | auto min = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int64_t>::min()))); |
2912 | |
2913 | auto temp1 = g32(); |
2914 | auto temp2 = g32(); |
2915 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrUnordered), arg, min, temp1); |
2916 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2917 | append(Or32, temp1, temp2); |
2918 | |
2919 | emitCheck([&] { |
2920 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2921 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2922 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2923 | }); |
2924 | |
2925 | auto* patchpoint = addPatchpoint(B3::Int64); |
2926 | patchpoint->effects = B3::Effects::none(); |
2927 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2928 | jit.truncateFloatToInt64(params[1].fpr(), params[0].gpr()); |
2929 | }); |
2930 | result = g64(); |
2931 | emitPatchpoint(patchpoint, result, arg); |
2932 | return { }; |
2933 | } |
2934 | |
2935 | template<> |
2936 | auto AirIRGenerator::addOp<OpType::I64TruncUF32>(ExpressionType arg, ExpressionType& result) -> PartialResult |
2937 | { |
2938 | auto max = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<int64_t>::min()) * static_cast<float>(-2.0))); |
2939 | auto min = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(-1.0))); |
2940 | |
2941 | auto temp1 = g32(); |
2942 | auto temp2 = g32(); |
2943 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqualOrUnordered), arg, min, temp1); |
2944 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqualOrUnordered), arg, max, temp2); |
2945 | append(Or32, temp1, temp2); |
2946 | |
2947 | emitCheck([&] { |
2948 | return Inst(BranchTest32, nullptr, Arg::resCond(MacroAssembler::NonZero), temp2, temp2); |
2949 | }, [=] (CCallHelpers& jit, const B3::StackmapGenerationParams&) { |
2950 | this->emitThrowException(jit, ExceptionType::OutOfBoundsTrunc); |
2951 | }); |
2952 | |
2953 | TypedTmp signBitConstant; |
2954 | if (isX86()) |
2955 | signBitConstant = addConstant(Type::F32, bitwise_cast<uint32_t>(static_cast<float>(std::numeric_limits<uint64_t>::max() - std::numeric_limits<int64_t>::max()))); |
2956 | |
2957 | auto* patchpoint = addPatchpoint(B3::Int64); |
2958 | patchpoint->effects = B3::Effects::none(); |
2959 | patchpoint->clobber(RegisterSet::macroScratchRegisters()); |
2960 | Vector<ConstrainedTmp, 2> args; |
2961 | args.append(arg); |
2962 | if (isX86()) { |
2963 | args.append(signBitConstant); |
2964 | patchpoint->numFPScratchRegisters = 1; |
2965 | } |
2966 | patchpoint->setGenerator([=] (CCallHelpers& jit, const B3::StackmapGenerationParams& params) { |
2967 | AllowMacroScratchRegisterUsage allowScratch(jit); |
2968 | FPRReg scratch = InvalidFPRReg; |
2969 | FPRReg constant = InvalidFPRReg; |
2970 | if (isX86()) { |
2971 | scratch = params.fpScratch(0); |
2972 | constant = params[2].fpr(); |
2973 | } |
2974 | jit.truncateFloatToUint64(params[1].fpr(), params[0].gpr(), scratch, constant); |
2975 | }); |
2976 | |
2977 | result = g64(); |
2978 | emitPatchpoint(m_currentBlock, patchpoint, Vector<Tmp, 1> { result }, WTFMove(args)); |
2979 | |
2980 | return { }; |
2981 | } |
2982 | |
2983 | auto AirIRGenerator::addShift(Type type, B3::Air::Opcode op, ExpressionType value, ExpressionType shift, ExpressionType& result) -> PartialResult |
2984 | { |
2985 | ASSERT(type == Type::I64 || type == Type::I32); |
2986 | result = tmpForType(type); |
2987 | |
2988 | if (isValidForm(op, Arg::Tmp, Arg::Tmp, Arg::Tmp)) { |
2989 | append(op, value, shift, result); |
2990 | return { }; |
2991 | } |
2992 | |
2993 | #if CPU(X86_64) |
2994 | Tmp ecx = Tmp(X86Registers::ecx); |
2995 | append(Move, value, result); |
2996 | append(Move, shift, ecx); |
2997 | append(op, ecx, result); |
2998 | #else |
2999 | RELEASE_ASSERT_NOT_REACHED(); |
3000 | #endif |
3001 | return { }; |
3002 | } |
3003 | |
3004 | auto AirIRGenerator::addIntegerSub(B3::Air::Opcode op, ExpressionType lhs, ExpressionType rhs, ExpressionType& result) -> PartialResult |
3005 | { |
3006 | ASSERT(op == Sub32 || op == Sub64); |
3007 | |
3008 | result = op == Sub32 ? g32() : g64(); |
3009 | |
3010 | if (isValidForm(op, Arg::Tmp, Arg::Tmp, Arg::Tmp)) { |
3011 | append(op, lhs, rhs, result); |
3012 | return { }; |
3013 | } |
3014 | |
3015 | RELEASE_ASSERT(isX86()); |
3016 | // Sub a, b |
3017 | // means |
3018 | // b = b Sub a |
3019 | append(Move, lhs, result); |
3020 | append(op, rhs, result); |
3021 | return { }; |
3022 | } |
3023 | |
3024 | auto AirIRGenerator::addFloatingPointAbs(B3::Air::Opcode op, ExpressionType value, ExpressionType& result) -> PartialResult |
3025 | { |
3026 | RELEASE_ASSERT(op == AbsFloat || op == AbsDouble); |
3027 | |
3028 | result = op == AbsFloat ? f32() : f64(); |
3029 | |
3030 | if (isValidForm(op, Arg::Tmp, Arg::Tmp)) { |
3031 | append(op, value, result); |
3032 | return { }; |
3033 | } |
3034 | |
3035 | RELEASE_ASSERT(isX86()); |
3036 | |
3037 | if (op == AbsFloat) { |
3038 | auto constant = g32(); |
3039 | append(Move, Arg::imm(static_cast<uint32_t>(~(1ull << 31))), constant); |
3040 | append(Move32ToFloat, constant, result); |
3041 | append(AndFloat, value, result); |
3042 | } else { |
3043 | auto constant = g64(); |
3044 | append(Move, Arg::bigImm(~(1ull << 63)), constant); |
3045 | append(Move64ToDouble, constant, result); |
3046 | append(AndDouble, value, result); |
3047 | } |
3048 | return { }; |
3049 | } |
3050 | |
3051 | auto AirIRGenerator::addFloatingPointBinOp(Type type, B3::Air::Opcode op, ExpressionType lhs, ExpressionType rhs, ExpressionType& result) -> PartialResult |
3052 | { |
3053 | ASSERT(type == Type::F32 || type == Type::F64); |
3054 | result = tmpForType(type); |
3055 | |
3056 | if (isValidForm(op, Arg::Tmp, Arg::Tmp, Arg::Tmp)) { |
3057 | append(op, lhs, rhs, result); |
3058 | return { }; |
3059 | } |
3060 | |
3061 | RELEASE_ASSERT(isX86()); |
3062 | |
3063 | // Op a, b |
3064 | // means |
3065 | // b = b Op a |
3066 | append(moveOpForValueType(type), lhs, result); |
3067 | append(op, rhs, result); |
3068 | return { }; |
3069 | } |
3070 | |
3071 | template<> auto AirIRGenerator::addOp<OpType::F32Ceil>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3072 | { |
3073 | result = f32(); |
3074 | append(CeilFloat, arg0, result); |
3075 | return { }; |
3076 | } |
3077 | |
3078 | template<> auto AirIRGenerator::addOp<OpType::I32Mul>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3079 | { |
3080 | result = g32(); |
3081 | append(Mul32, arg0, arg1, result); |
3082 | return { }; |
3083 | } |
3084 | |
3085 | template<> auto AirIRGenerator::addOp<OpType::I32Sub>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3086 | { |
3087 | return addIntegerSub(Sub32, arg0, arg1, result); |
3088 | } |
3089 | |
3090 | template<> auto AirIRGenerator::addOp<OpType::F64Le>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3091 | { |
3092 | result = g32(); |
3093 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqual), arg0, arg1, result); |
3094 | return { }; |
3095 | } |
3096 | |
3097 | template<> auto AirIRGenerator::addOp<OpType::F32DemoteF64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3098 | { |
3099 | result = f32(); |
3100 | append(ConvertDoubleToFloat, arg0, result); |
3101 | return { }; |
3102 | } |
3103 | |
3104 | template<> auto AirIRGenerator::addOp<OpType::F32Min>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3105 | { |
3106 | return addFloatingPointMinOrMax(F32, MinOrMax::Min, arg0, arg1, result); |
3107 | } |
3108 | |
3109 | template<> auto AirIRGenerator::addOp<OpType::F64Ne>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3110 | { |
3111 | result = g32(); |
3112 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleNotEqualOrUnordered), arg0, arg1, result); |
3113 | return { }; |
3114 | } |
3115 | |
3116 | template<> auto AirIRGenerator::addOp<OpType::F64Lt>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3117 | { |
3118 | result = g32(); |
3119 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleLessThan), arg0, arg1, result); |
3120 | return { }; |
3121 | } |
3122 | |
3123 | auto AirIRGenerator::addFloatingPointMinOrMax(Type floatType, MinOrMax minOrMax, ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3124 | { |
3125 | ASSERT(floatType == F32 || floatType == F64); |
3126 | result = tmpForType(floatType); |
3127 | |
3128 | BasicBlock* isEqual = m_code.addBlock(); |
3129 | BasicBlock* notEqual = m_code.addBlock(); |
3130 | BasicBlock* isLessThan = m_code.addBlock(); |
3131 | BasicBlock* notLessThan = m_code.addBlock(); |
3132 | BasicBlock* isGreaterThan = m_code.addBlock(); |
3133 | BasicBlock* isNaN = m_code.addBlock(); |
3134 | BasicBlock* continuation = m_code.addBlock(); |
3135 | |
3136 | auto branchOp = floatType == F32 ? BranchFloat : BranchDouble; |
3137 | append(m_currentBlock, branchOp, Arg::doubleCond(MacroAssembler::DoubleEqual), arg0, arg1); |
3138 | m_currentBlock->setSuccessors(isEqual, notEqual); |
3139 | |
3140 | append(notEqual, branchOp, Arg::doubleCond(MacroAssembler::DoubleLessThan), arg0, arg1); |
3141 | notEqual->setSuccessors(isLessThan, notLessThan); |
3142 | |
3143 | append(notLessThan, branchOp, Arg::doubleCond(MacroAssembler::DoubleGreaterThan), arg0, arg1); |
3144 | notLessThan->setSuccessors(isGreaterThan, isNaN); |
3145 | |
3146 | auto andOp = floatType == F32 ? AndFloat : AndDouble; |
3147 | auto orOp = floatType == F32 ? OrFloat : OrDouble; |
3148 | append(isEqual, minOrMax == MinOrMax::Max ? andOp : orOp, arg0, arg1, result); |
3149 | append(isEqual, Jump); |
3150 | isEqual->setSuccessors(continuation); |
3151 | |
3152 | auto isLessThanResult = minOrMax == MinOrMax::Max ? arg1 : arg0; |
3153 | append(isLessThan, moveOpForValueType(floatType), isLessThanResult, result); |
3154 | append(isLessThan, Jump); |
3155 | isLessThan->setSuccessors(continuation); |
3156 | |
3157 | auto isGreaterThanResult = minOrMax == MinOrMax::Max ? arg0 : arg1; |
3158 | append(isGreaterThan, moveOpForValueType(floatType), isGreaterThanResult, result); |
3159 | append(isGreaterThan, Jump); |
3160 | isGreaterThan->setSuccessors(continuation); |
3161 | |
3162 | auto addOp = floatType == F32 ? AddFloat : AddDouble; |
3163 | append(isNaN, addOp, arg0, arg1, result); |
3164 | append(isNaN, Jump); |
3165 | isNaN->setSuccessors(continuation); |
3166 | |
3167 | m_currentBlock = continuation; |
3168 | |
3169 | return { }; |
3170 | } |
3171 | |
3172 | template<> auto AirIRGenerator::addOp<OpType::F32Max>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3173 | { |
3174 | return addFloatingPointMinOrMax(F32, MinOrMax::Max, arg0, arg1, result); |
3175 | } |
3176 | |
3177 | template<> auto AirIRGenerator::addOp<OpType::F64Mul>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3178 | { |
3179 | return addFloatingPointBinOp(Type::F64, MulDouble, arg0, arg1, result); |
3180 | } |
3181 | |
3182 | template<> auto AirIRGenerator::addOp<OpType::F32Div>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3183 | { |
3184 | return addFloatingPointBinOp(Type::F32, DivFloat, arg0, arg1, result); |
3185 | } |
3186 | |
3187 | template<> auto AirIRGenerator::addOp<OpType::I32Clz>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3188 | { |
3189 | result = g32(); |
3190 | append(CountLeadingZeros32, arg0, result); |
3191 | return { }; |
3192 | } |
3193 | |
3194 | template<> auto AirIRGenerator::addOp<OpType::F32Copysign>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3195 | { |
3196 | // FIXME: We can have better codegen here for the imms and two operand forms on x86 |
3197 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
3198 | result = f32(); |
3199 | auto temp1 = g32(); |
3200 | auto sign = g32(); |
3201 | auto value = g32(); |
3202 | |
3203 | // FIXME: Try to use Imm where possible: |
3204 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
3205 | append(MoveFloatTo32, arg1, temp1); |
3206 | append(Move, Arg::bigImm(0x80000000), sign); |
3207 | append(And32, temp1, sign, sign); |
3208 | |
3209 | append(MoveDoubleTo64, arg0, temp1); |
3210 | append(Move, Arg::bigImm(0x7fffffff), value); |
3211 | append(And32, temp1, value, value); |
3212 | |
3213 | append(Or32, sign, value, value); |
3214 | append(Move32ToFloat, value, result); |
3215 | |
3216 | return { }; |
3217 | } |
3218 | |
3219 | template<> auto AirIRGenerator::addOp<OpType::F64ConvertUI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3220 | { |
3221 | result = f64(); |
3222 | auto temp = g64(); |
3223 | append(Move32, arg0, temp); |
3224 | append(ConvertInt64ToDouble, temp, result); |
3225 | return { }; |
3226 | } |
3227 | |
3228 | template<> auto AirIRGenerator::addOp<OpType::F32ReinterpretI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3229 | { |
3230 | result = f32(); |
3231 | append(Move32ToFloat, arg0, result); |
3232 | return { }; |
3233 | } |
3234 | |
3235 | template<> auto AirIRGenerator::addOp<OpType::I64And>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3236 | { |
3237 | result = g64(); |
3238 | append(And64, arg0, arg1, result); |
3239 | return { }; |
3240 | } |
3241 | |
3242 | template<> auto AirIRGenerator::addOp<OpType::F32Ne>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3243 | { |
3244 | result = g32(); |
3245 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleNotEqualOrUnordered), arg0, arg1, result); |
3246 | return { }; |
3247 | } |
3248 | |
3249 | template<> auto AirIRGenerator::addOp<OpType::F64Gt>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3250 | { |
3251 | result = g32(); |
3252 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThan), arg0, arg1, result); |
3253 | return { }; |
3254 | } |
3255 | |
3256 | template<> auto AirIRGenerator::addOp<OpType::F32Sqrt>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3257 | { |
3258 | result = f32(); |
3259 | append(SqrtFloat, arg0, result); |
3260 | return { }; |
3261 | } |
3262 | |
3263 | template<> auto AirIRGenerator::addOp<OpType::F64Ge>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3264 | { |
3265 | result = g32(); |
3266 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqual), arg0, arg1, result); |
3267 | return { }; |
3268 | } |
3269 | |
3270 | template<> auto AirIRGenerator::addOp<OpType::I64GtS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3271 | { |
3272 | result = g32(); |
3273 | append(Compare64, Arg::relCond(MacroAssembler::GreaterThan), arg0, arg1, result); |
3274 | return { }; |
3275 | } |
3276 | |
3277 | template<> auto AirIRGenerator::addOp<OpType::I64GtU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3278 | { |
3279 | result = g32(); |
3280 | append(Compare64, Arg::relCond(MacroAssembler::Above), arg0, arg1, result); |
3281 | return { }; |
3282 | } |
3283 | |
3284 | template<> auto AirIRGenerator::addOp<OpType::I64Eqz>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3285 | { |
3286 | result = g32(); |
3287 | append(Test64, Arg::resCond(MacroAssembler::Zero), arg0, arg0, result); |
3288 | return { }; |
3289 | } |
3290 | |
3291 | template<> auto AirIRGenerator::addOp<OpType::F64Div>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3292 | { |
3293 | return addFloatingPointBinOp(Type::F64, DivDouble, arg0, arg1, result); |
3294 | } |
3295 | |
3296 | template<> auto AirIRGenerator::addOp<OpType::F32Add>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3297 | { |
3298 | result = f32(); |
3299 | append(AddFloat, arg0, arg1, result); |
3300 | return { }; |
3301 | } |
3302 | |
3303 | template<> auto AirIRGenerator::addOp<OpType::I64Or>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3304 | { |
3305 | result = g64(); |
3306 | append(Or64, arg0, arg1, result); |
3307 | return { }; |
3308 | } |
3309 | |
3310 | template<> auto AirIRGenerator::addOp<OpType::I32LeU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3311 | { |
3312 | result = g32(); |
3313 | append(Compare32, Arg::relCond(MacroAssembler::BelowOrEqual), arg0, arg1, result); |
3314 | return { }; |
3315 | } |
3316 | |
3317 | template<> auto AirIRGenerator::addOp<OpType::I32LeS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3318 | { |
3319 | result = g32(); |
3320 | append(Compare32, Arg::relCond(MacroAssembler::LessThanOrEqual), arg0, arg1, result); |
3321 | return { }; |
3322 | } |
3323 | |
3324 | template<> auto AirIRGenerator::addOp<OpType::I64Ne>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3325 | { |
3326 | result = g32(); |
3327 | append(Compare64, Arg::relCond(MacroAssembler::NotEqual), arg0, arg1, result); |
3328 | return { }; |
3329 | } |
3330 | |
3331 | template<> auto AirIRGenerator::addOp<OpType::I64Clz>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3332 | { |
3333 | result = g64(); |
3334 | append(CountLeadingZeros64, arg0, result); |
3335 | return { }; |
3336 | } |
3337 | |
3338 | template<> auto AirIRGenerator::addOp<OpType::F32Neg>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3339 | { |
3340 | result = f32(); |
3341 | if (isValidForm(NegateFloat, Arg::Tmp, Arg::Tmp)) |
3342 | append(NegateFloat, arg0, result); |
3343 | else { |
3344 | auto constant = addConstant(Type::I32, bitwise_cast<uint32_t>(static_cast<float>(-0.0))); |
3345 | auto temp = g32(); |
3346 | append(MoveFloatTo32, arg0, temp); |
3347 | append(Xor32, constant, temp); |
3348 | append(Move32ToFloat, temp, result); |
3349 | } |
3350 | return { }; |
3351 | } |
3352 | |
3353 | template<> auto AirIRGenerator::addOp<OpType::I32And>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3354 | { |
3355 | result = g32(); |
3356 | append(And32, arg0, arg1, result); |
3357 | return { }; |
3358 | } |
3359 | |
3360 | template<> auto AirIRGenerator::addOp<OpType::I32LtU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3361 | { |
3362 | result = g32(); |
3363 | append(Compare32, Arg::relCond(MacroAssembler::Below), arg0, arg1, result); |
3364 | return { }; |
3365 | } |
3366 | |
3367 | template<> auto AirIRGenerator::addOp<OpType::I64Rotr>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3368 | { |
3369 | return addShift(Type::I64, RotateRight64, arg0, arg1, result); |
3370 | } |
3371 | |
3372 | template<> auto AirIRGenerator::addOp<OpType::F64Abs>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3373 | { |
3374 | return addFloatingPointAbs(AbsDouble, arg0, result); |
3375 | } |
3376 | |
3377 | template<> auto AirIRGenerator::addOp<OpType::I32LtS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3378 | { |
3379 | result = g32(); |
3380 | append(Compare32, Arg::relCond(MacroAssembler::LessThan), arg0, arg1, result); |
3381 | return { }; |
3382 | } |
3383 | |
3384 | template<> auto AirIRGenerator::addOp<OpType::I32Eq>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3385 | { |
3386 | result = g32(); |
3387 | append(Compare32, Arg::relCond(MacroAssembler::Equal), arg0, arg1, result); |
3388 | return { }; |
3389 | } |
3390 | |
3391 | template<> auto AirIRGenerator::addOp<OpType::F64Copysign>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3392 | { |
3393 | // FIXME: We can have better codegen here for the imms and two operand forms on x86 |
3394 | // https://bugs.webkit.org/show_bug.cgi?id=193999 |
3395 | result = f64(); |
3396 | auto temp1 = g64(); |
3397 | auto sign = g64(); |
3398 | auto value = g64(); |
3399 | |
3400 | append(MoveDoubleTo64, arg1, temp1); |
3401 | append(Move, Arg::bigImm(0x8000000000000000), sign); |
3402 | append(And64, temp1, sign, sign); |
3403 | |
3404 | append(MoveDoubleTo64, arg0, temp1); |
3405 | append(Move, Arg::bigImm(0x7fffffffffffffff), value); |
3406 | append(And64, temp1, value, value); |
3407 | |
3408 | append(Or64, sign, value, value); |
3409 | append(Move64ToDouble, value, result); |
3410 | |
3411 | return { }; |
3412 | } |
3413 | |
3414 | template<> auto AirIRGenerator::addOp<OpType::F32ConvertSI64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3415 | { |
3416 | result = f32(); |
3417 | append(ConvertInt64ToFloat, arg0, result); |
3418 | return { }; |
3419 | } |
3420 | |
3421 | template<> auto AirIRGenerator::addOp<OpType::I64Rotl>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3422 | { |
3423 | if (isARM64()) { |
3424 | // ARM64 doesn't have a rotate left. |
3425 | auto newShift = g64(); |
3426 | append(Move, arg1, newShift); |
3427 | append(Neg64, newShift); |
3428 | return addShift(Type::I64, RotateRight64, arg0, newShift, result); |
3429 | } else |
3430 | return addShift(Type::I64, RotateLeft64, arg0, arg1, result); |
3431 | } |
3432 | |
3433 | template<> auto AirIRGenerator::addOp<OpType::F32Lt>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3434 | { |
3435 | result = g32(); |
3436 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThan), arg0, arg1, result); |
3437 | return { }; |
3438 | } |
3439 | |
3440 | template<> auto AirIRGenerator::addOp<OpType::F64ConvertSI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3441 | { |
3442 | result = f64(); |
3443 | append(ConvertInt32ToDouble, arg0, result); |
3444 | return { }; |
3445 | } |
3446 | |
3447 | template<> auto AirIRGenerator::addOp<OpType::F64Eq>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3448 | { |
3449 | result = g32(); |
3450 | append(CompareDouble, Arg::doubleCond(MacroAssembler::DoubleEqual), arg0, arg1, result); |
3451 | return { }; |
3452 | } |
3453 | |
3454 | template<> auto AirIRGenerator::addOp<OpType::F32Le>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3455 | { |
3456 | result = g32(); |
3457 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleLessThanOrEqual), arg0, arg1, result); |
3458 | return { }; |
3459 | } |
3460 | |
3461 | template<> auto AirIRGenerator::addOp<OpType::F32Ge>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3462 | { |
3463 | result = g32(); |
3464 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThanOrEqual), arg0, arg1, result); |
3465 | return { }; |
3466 | } |
3467 | |
3468 | template<> auto AirIRGenerator::addOp<OpType::I32ShrU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3469 | { |
3470 | return addShift(Type::I32, Urshift32, arg0, arg1, result); |
3471 | } |
3472 | |
3473 | template<> auto AirIRGenerator::addOp<OpType::F32ConvertUI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3474 | { |
3475 | result = f32(); |
3476 | auto temp = g64(); |
3477 | append(Move32, arg0, temp); |
3478 | append(ConvertInt64ToFloat, temp, result); |
3479 | return { }; |
3480 | } |
3481 | |
3482 | template<> auto AirIRGenerator::addOp<OpType::I32ShrS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3483 | { |
3484 | return addShift(Type::I32, Rshift32, arg0, arg1, result); |
3485 | } |
3486 | |
3487 | template<> auto AirIRGenerator::addOp<OpType::I32GeU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3488 | { |
3489 | result = g32(); |
3490 | append(Compare32, Arg::relCond(MacroAssembler::AboveOrEqual), arg0, arg1, result); |
3491 | return { }; |
3492 | } |
3493 | |
3494 | template<> auto AirIRGenerator::addOp<OpType::F64Ceil>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3495 | { |
3496 | result = f64(); |
3497 | append(CeilDouble, arg0, result); |
3498 | return { }; |
3499 | } |
3500 | |
3501 | template<> auto AirIRGenerator::addOp<OpType::I32GeS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3502 | { |
3503 | result = g32(); |
3504 | append(Compare32, Arg::relCond(MacroAssembler::GreaterThanOrEqual), arg0, arg1, result); |
3505 | return { }; |
3506 | } |
3507 | |
3508 | template<> auto AirIRGenerator::addOp<OpType::I32Shl>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3509 | { |
3510 | return addShift(Type::I32, Lshift32, arg0, arg1, result); |
3511 | } |
3512 | |
3513 | template<> auto AirIRGenerator::addOp<OpType::F64Floor>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3514 | { |
3515 | result = f64(); |
3516 | append(FloorDouble, arg0, result); |
3517 | return { }; |
3518 | } |
3519 | |
3520 | template<> auto AirIRGenerator::addOp<OpType::I32Xor>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3521 | { |
3522 | result = g32(); |
3523 | append(Xor32, arg0, arg1, result); |
3524 | return { }; |
3525 | } |
3526 | |
3527 | template<> auto AirIRGenerator::addOp<OpType::F32Abs>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3528 | { |
3529 | return addFloatingPointAbs(AbsFloat, arg0, result); |
3530 | } |
3531 | |
3532 | template<> auto AirIRGenerator::addOp<OpType::F64Min>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3533 | { |
3534 | return addFloatingPointMinOrMax(F64, MinOrMax::Min, arg0, arg1, result); |
3535 | } |
3536 | |
3537 | template<> auto AirIRGenerator::addOp<OpType::F32Mul>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3538 | { |
3539 | result = f32(); |
3540 | append(MulFloat, arg0, arg1, result); |
3541 | return { }; |
3542 | } |
3543 | |
3544 | template<> auto AirIRGenerator::addOp<OpType::I64Sub>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3545 | { |
3546 | return addIntegerSub(Sub64, arg0, arg1, result); |
3547 | } |
3548 | |
3549 | template<> auto AirIRGenerator::addOp<OpType::I32ReinterpretF32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3550 | { |
3551 | result = g32(); |
3552 | append(MoveFloatTo32, arg0, result); |
3553 | return { }; |
3554 | } |
3555 | |
3556 | template<> auto AirIRGenerator::addOp<OpType::I32Add>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3557 | { |
3558 | result = g32(); |
3559 | append(Add32, arg0, arg1, result); |
3560 | return { }; |
3561 | } |
3562 | |
3563 | template<> auto AirIRGenerator::addOp<OpType::F64Sub>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3564 | { |
3565 | return addFloatingPointBinOp(Type::F64, SubDouble, arg0, arg1, result); |
3566 | } |
3567 | |
3568 | template<> auto AirIRGenerator::addOp<OpType::I32Or>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3569 | { |
3570 | result = g32(); |
3571 | append(Or32, arg0, arg1, result); |
3572 | return { }; |
3573 | } |
3574 | |
3575 | template<> auto AirIRGenerator::addOp<OpType::I64LtU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3576 | { |
3577 | result = g32(); |
3578 | append(Compare64, Arg::relCond(MacroAssembler::Below), arg0, arg1, result); |
3579 | return { }; |
3580 | } |
3581 | |
3582 | template<> auto AirIRGenerator::addOp<OpType::I64LtS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3583 | { |
3584 | result = g32(); |
3585 | append(Compare64, Arg::relCond(MacroAssembler::LessThan), arg0, arg1, result); |
3586 | return { }; |
3587 | } |
3588 | |
3589 | template<> auto AirIRGenerator::addOp<OpType::F64ConvertSI64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3590 | { |
3591 | result = f64(); |
3592 | append(ConvertInt64ToDouble, arg0, result); |
3593 | return { }; |
3594 | } |
3595 | |
3596 | template<> auto AirIRGenerator::addOp<OpType::I64Xor>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3597 | { |
3598 | result = g64(); |
3599 | append(Xor64, arg0, arg1, result); |
3600 | return { }; |
3601 | } |
3602 | |
3603 | template<> auto AirIRGenerator::addOp<OpType::I64GeU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3604 | { |
3605 | result = g32(); |
3606 | append(Compare64, Arg::relCond(MacroAssembler::AboveOrEqual), arg0, arg1, result); |
3607 | return { }; |
3608 | } |
3609 | |
3610 | template<> auto AirIRGenerator::addOp<OpType::I64Mul>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3611 | { |
3612 | result = g64(); |
3613 | append(Mul64, arg0, arg1, result); |
3614 | return { }; |
3615 | } |
3616 | |
3617 | template<> auto AirIRGenerator::addOp<OpType::F32Sub>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3618 | { |
3619 | result = f32(); |
3620 | if (isValidForm(SubFloat, Arg::Tmp, Arg::Tmp, Arg::Tmp)) |
3621 | append(SubFloat, arg0, arg1, result); |
3622 | else { |
3623 | RELEASE_ASSERT(isX86()); |
3624 | append(MoveFloat, arg0, result); |
3625 | append(SubFloat, arg1, result); |
3626 | } |
3627 | return { }; |
3628 | } |
3629 | |
3630 | template<> auto AirIRGenerator::addOp<OpType::F64PromoteF32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3631 | { |
3632 | result = f64(); |
3633 | append(ConvertFloatToDouble, arg0, result); |
3634 | return { }; |
3635 | } |
3636 | |
3637 | template<> auto AirIRGenerator::addOp<OpType::F64Add>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3638 | { |
3639 | result = f64(); |
3640 | append(AddDouble, arg0, arg1, result); |
3641 | return { }; |
3642 | } |
3643 | |
3644 | template<> auto AirIRGenerator::addOp<OpType::I64GeS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3645 | { |
3646 | result = g32(); |
3647 | append(Compare64, Arg::relCond(MacroAssembler::GreaterThanOrEqual), arg0, arg1, result); |
3648 | return { }; |
3649 | } |
3650 | |
3651 | template<> auto AirIRGenerator::addOp<OpType::I64ExtendUI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3652 | { |
3653 | result = g64(); |
3654 | append(Move32, arg0, result); |
3655 | return { }; |
3656 | } |
3657 | |
3658 | template<> auto AirIRGenerator::addOp<OpType::I32Ne>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3659 | { |
3660 | result = g32(); |
3661 | RELEASE_ASSERT(arg0 && arg1); |
3662 | append(Compare32, Arg::relCond(MacroAssembler::NotEqual), arg0, arg1, result); |
3663 | return { }; |
3664 | } |
3665 | |
3666 | template<> auto AirIRGenerator::addOp<OpType::F64ReinterpretI64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3667 | { |
3668 | result = f64(); |
3669 | append(Move64ToDouble, arg0, result); |
3670 | return { }; |
3671 | } |
3672 | |
3673 | template<> auto AirIRGenerator::addOp<OpType::F32Eq>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3674 | { |
3675 | result = g32(); |
3676 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleEqual), arg0, arg1, result); |
3677 | return { }; |
3678 | } |
3679 | |
3680 | template<> auto AirIRGenerator::addOp<OpType::I64Eq>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3681 | { |
3682 | result = g32(); |
3683 | append(Compare64, Arg::relCond(MacroAssembler::Equal), arg0, arg1, result); |
3684 | return { }; |
3685 | } |
3686 | |
3687 | template<> auto AirIRGenerator::addOp<OpType::F32Floor>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3688 | { |
3689 | result = f32(); |
3690 | append(FloorFloat, arg0, result); |
3691 | return { }; |
3692 | } |
3693 | |
3694 | template<> auto AirIRGenerator::addOp<OpType::F32ConvertSI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3695 | { |
3696 | result = f32(); |
3697 | append(ConvertInt32ToFloat, arg0, result); |
3698 | return { }; |
3699 | } |
3700 | |
3701 | template<> auto AirIRGenerator::addOp<OpType::I32Eqz>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3702 | { |
3703 | result = g32(); |
3704 | append(Test32, Arg::resCond(MacroAssembler::Zero), arg0, arg0, result); |
3705 | return { }; |
3706 | } |
3707 | |
3708 | template<> auto AirIRGenerator::addOp<OpType::I64ReinterpretF64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3709 | { |
3710 | result = g64(); |
3711 | append(MoveDoubleTo64, arg0, result); |
3712 | return { }; |
3713 | } |
3714 | |
3715 | template<> auto AirIRGenerator::addOp<OpType::I64ShrS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3716 | { |
3717 | return addShift(Type::I64, Rshift64, arg0, arg1, result); |
3718 | } |
3719 | |
3720 | template<> auto AirIRGenerator::addOp<OpType::I64ShrU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3721 | { |
3722 | return addShift(Type::I64, Urshift64, arg0, arg1, result); |
3723 | } |
3724 | |
3725 | template<> auto AirIRGenerator::addOp<OpType::F64Sqrt>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3726 | { |
3727 | result = f64(); |
3728 | append(SqrtDouble, arg0, result); |
3729 | return { }; |
3730 | } |
3731 | |
3732 | template<> auto AirIRGenerator::addOp<OpType::I64Shl>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3733 | { |
3734 | return addShift(Type::I64, Lshift64, arg0, arg1, result); |
3735 | } |
3736 | |
3737 | template<> auto AirIRGenerator::addOp<OpType::F32Gt>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3738 | { |
3739 | result = g32(); |
3740 | append(CompareFloat, Arg::doubleCond(MacroAssembler::DoubleGreaterThan), arg0, arg1, result); |
3741 | return { }; |
3742 | } |
3743 | |
3744 | template<> auto AirIRGenerator::addOp<OpType::I32WrapI64>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3745 | { |
3746 | result = g32(); |
3747 | append(Move32, arg0, result); |
3748 | return { }; |
3749 | } |
3750 | |
3751 | template<> auto AirIRGenerator::addOp<OpType::I32Rotl>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3752 | { |
3753 | if (isARM64()) { |
3754 | // ARM64 doesn't have a rotate left. |
3755 | auto newShift = g64(); |
3756 | append(Move, arg1, newShift); |
3757 | append(Neg64, newShift); |
3758 | return addShift(Type::I32, RotateRight32, arg0, newShift, result); |
3759 | } else |
3760 | return addShift(Type::I32, RotateLeft32, arg0, arg1, result); |
3761 | } |
3762 | |
3763 | template<> auto AirIRGenerator::addOp<OpType::I32Rotr>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3764 | { |
3765 | return addShift(Type::I32, RotateRight32, arg0, arg1, result); |
3766 | } |
3767 | |
3768 | template<> auto AirIRGenerator::addOp<OpType::I32GtU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3769 | { |
3770 | result = g32(); |
3771 | append(Compare32, Arg::relCond(MacroAssembler::Above), arg0, arg1, result); |
3772 | return { }; |
3773 | } |
3774 | |
3775 | template<> auto AirIRGenerator::addOp<OpType::I64ExtendSI32>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3776 | { |
3777 | result = g64(); |
3778 | append(SignExtend32ToPtr, arg0, result); |
3779 | return { }; |
3780 | } |
3781 | |
3782 | template<> auto AirIRGenerator::addOp<OpType::I32GtS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3783 | { |
3784 | result = g32(); |
3785 | append(Compare32, Arg::relCond(MacroAssembler::GreaterThan), arg0, arg1, result); |
3786 | return { }; |
3787 | } |
3788 | |
3789 | template<> auto AirIRGenerator::addOp<OpType::F64Neg>(ExpressionType arg0, ExpressionType& result) -> PartialResult |
3790 | { |
3791 | result = f64(); |
3792 | if (isValidForm(NegateDouble, Arg::Tmp, Arg::Tmp)) |
3793 | append(NegateDouble, arg0, result); |
3794 | else { |
3795 | auto constant = addConstant(Type::I64, bitwise_cast<uint64_t>(static_cast<double>(-0.0))); |
3796 | auto temp = g64(); |
3797 | append(MoveDoubleTo64, arg0, temp); |
3798 | append(Xor64, constant, temp); |
3799 | append(Move64ToDouble, temp, result); |
3800 | } |
3801 | return { }; |
3802 | } |
3803 | |
3804 | template<> auto AirIRGenerator::addOp<OpType::F64Max>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3805 | { |
3806 | return addFloatingPointMinOrMax(F64, MinOrMax::Max, arg0, arg1, result); |
3807 | } |
3808 | |
3809 | template<> auto AirIRGenerator::addOp<OpType::I64LeU>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3810 | { |
3811 | result = g32(); |
3812 | append(Compare64, Arg::relCond(MacroAssembler::BelowOrEqual), arg0, arg1, result); |
3813 | return { }; |
3814 | } |
3815 | |
3816 | template<> auto AirIRGenerator::addOp<OpType::I64LeS>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3817 | { |
3818 | result = g32(); |
3819 | append(Compare64, Arg::relCond(MacroAssembler::LessThanOrEqual), arg0, arg1, result); |
3820 | return { }; |
3821 | } |
3822 | |
3823 | template<> auto AirIRGenerator::addOp<OpType::I64Add>(ExpressionType arg0, ExpressionType arg1, ExpressionType& result) -> PartialResult |
3824 | { |
3825 | result = g64(); |
3826 | append(Add64, arg0, arg1, result); |
3827 | return { }; |
3828 | } |
3829 | |
3830 | } } // namespace JSC::Wasm |
3831 | |
3832 | #endif // ENABLE(WEBASSEMBLY) |
3833 | |