1 | /* |
2 | * Copyright (C) 2011-2019 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 "DFGOSREntry.h" |
28 | |
29 | #if ENABLE(DFG_JIT) |
30 | |
31 | #include "CallFrame.h" |
32 | #include "CodeBlock.h" |
33 | #include "DFGJITCode.h" |
34 | #include "DFGNode.h" |
35 | #include "InterpreterInlines.h" |
36 | #include "JIT.h" |
37 | #include "JSCInlines.h" |
38 | #include "VMInlines.h" |
39 | #include <wtf/CommaPrinter.h> |
40 | |
41 | namespace JSC { namespace DFG { |
42 | |
43 | void OSREntryData::dumpInContext(PrintStream& out, DumpContext* context) const |
44 | { |
45 | out.print(m_bytecodeIndex, ", machine code = " , RawPointer(m_machineCode.executableAddress())); |
46 | out.print(", stack rules = [" ); |
47 | |
48 | auto printOperand = [&] (VirtualRegister reg) { |
49 | out.print(inContext(m_expectedValues.operand(reg), context), " (" ); |
50 | VirtualRegister toReg; |
51 | bool overwritten = false; |
52 | for (OSREntryReshuffling reshuffling : m_reshufflings) { |
53 | if (reg == VirtualRegister(reshuffling.fromOffset)) { |
54 | toReg = VirtualRegister(reshuffling.toOffset); |
55 | break; |
56 | } |
57 | if (reg == VirtualRegister(reshuffling.toOffset)) |
58 | overwritten = true; |
59 | } |
60 | if (!overwritten && !toReg.isValid()) |
61 | toReg = reg; |
62 | if (toReg.isValid()) { |
63 | if (toReg.isLocal() && !m_machineStackUsed.get(toReg.toLocal())) |
64 | out.print("ignored" ); |
65 | else |
66 | out.print("maps to " , toReg); |
67 | } else |
68 | out.print("overwritten" ); |
69 | if (reg.isLocal() && m_localsForcedDouble.get(reg.toLocal())) |
70 | out.print(", forced double" ); |
71 | if (reg.isLocal() && m_localsForcedAnyInt.get(reg.toLocal())) |
72 | out.print(", forced machine int" ); |
73 | out.print(")" ); |
74 | }; |
75 | |
76 | CommaPrinter comma; |
77 | for (size_t argumentIndex = m_expectedValues.numberOfArguments(); argumentIndex--;) { |
78 | out.print(comma, "arg" , argumentIndex, ":" ); |
79 | printOperand(virtualRegisterForArgument(argumentIndex)); |
80 | } |
81 | for (size_t localIndex = 0; localIndex < m_expectedValues.numberOfLocals(); ++localIndex) { |
82 | out.print(comma, "loc" , localIndex, ":" ); |
83 | printOperand(virtualRegisterForLocal(localIndex)); |
84 | } |
85 | |
86 | out.print("], machine stack used = " , m_machineStackUsed); |
87 | } |
88 | |
89 | void OSREntryData::dump(PrintStream& out) const |
90 | { |
91 | dumpInContext(out, nullptr); |
92 | } |
93 | |
94 | SUPPRESS_ASAN |
95 | void* prepareOSREntry(VM& vm, CallFrame* callFrame, CodeBlock* codeBlock, BytecodeIndex bytecodeIndex) |
96 | { |
97 | ASSERT(JITCode::isOptimizingJIT(codeBlock->jitType())); |
98 | ASSERT(codeBlock->alternative()); |
99 | ASSERT(codeBlock->alternative()->jitType() == JITType::BaselineJIT); |
100 | ASSERT(!codeBlock->jitCodeMap()); |
101 | ASSERT(codeBlock->jitCode()->dfgCommon()->isStillValid); |
102 | |
103 | if (!Options::useOSREntryToDFG()) |
104 | return nullptr; |
105 | |
106 | if (Options::verboseOSR()) { |
107 | dataLog( |
108 | "DFG OSR in " , *codeBlock->alternative(), " -> " , *codeBlock, |
109 | " from " , bytecodeIndex, "\n" ); |
110 | } |
111 | |
112 | sanitizeStackForVM(vm); |
113 | |
114 | if (bytecodeIndex) |
115 | codeBlock->ownerExecutable()->setDidTryToEnterInLoop(true); |
116 | |
117 | if (codeBlock->jitType() != JITType::DFGJIT) { |
118 | RELEASE_ASSERT(codeBlock->jitType() == JITType::FTLJIT); |
119 | |
120 | // When will this happen? We could have: |
121 | // |
122 | // - An exit from the FTL JIT into the baseline JIT followed by an attempt |
123 | // to reenter. We're fine with allowing this to fail. If it happens |
124 | // enough we'll just reoptimize. It basically means that the OSR exit cost |
125 | // us dearly and so reoptimizing is the right thing to do. |
126 | // |
127 | // - We have recursive code with hot loops. Consider that foo has a hot loop |
128 | // that calls itself. We have two foo's on the stack, lets call them foo1 |
129 | // and foo2, with foo1 having called foo2 from foo's hot loop. foo2 gets |
130 | // optimized all the way into the FTL. Then it returns into foo1, and then |
131 | // foo1 wants to get optimized. It might reach this conclusion from its |
132 | // hot loop and attempt to OSR enter. And we'll tell it that it can't. It |
133 | // might be worth addressing this case, but I just think this case will |
134 | // be super rare. For now, if it does happen, it'll cause some compilation |
135 | // thrashing. |
136 | |
137 | if (Options::verboseOSR()) |
138 | dataLog(" OSR failed because the target code block is not DFG.\n" ); |
139 | return nullptr; |
140 | } |
141 | |
142 | JITCode* jitCode = codeBlock->jitCode()->dfg(); |
143 | OSREntryData* entry = jitCode->osrEntryDataForBytecodeIndex(bytecodeIndex); |
144 | |
145 | if (!entry) { |
146 | if (Options::verboseOSR()) |
147 | dataLogF(" OSR failed because the entrypoint was optimized out.\n" ); |
148 | return nullptr; |
149 | } |
150 | |
151 | ASSERT(entry->m_bytecodeIndex == bytecodeIndex); |
152 | |
153 | // The code below checks if it is safe to perform OSR entry. It may find |
154 | // that it is unsafe to do so, for any number of reasons, which are documented |
155 | // below. If the code decides not to OSR then it returns 0, and it's the caller's |
156 | // responsibility to patch up the state in such a way as to ensure that it's |
157 | // both safe and efficient to continue executing baseline code for now. This |
158 | // should almost certainly include calling either codeBlock->optimizeAfterWarmUp() |
159 | // or codeBlock->dontOptimizeAnytimeSoon(). |
160 | |
161 | // 1) Verify predictions. If the predictions are inconsistent with the actual |
162 | // values, then OSR entry is not possible at this time. It's tempting to |
163 | // assume that we could somehow avoid this case. We can certainly avoid it |
164 | // for first-time loop OSR - that is, OSR into a CodeBlock that we have just |
165 | // compiled. Then we are almost guaranteed that all of the predictions will |
166 | // check out. It would be pretty easy to make that a hard guarantee. But |
167 | // then there would still be the case where two call frames with the same |
168 | // baseline CodeBlock are on the stack at the same time. The top one |
169 | // triggers compilation and OSR. In that case, we may no longer have |
170 | // accurate value profiles for the one deeper in the stack. Hence, when we |
171 | // pop into the CodeBlock that is deeper on the stack, we might OSR and |
172 | // realize that the predictions are wrong. Probably, in most cases, this is |
173 | // just an anomaly in the sense that the older CodeBlock simply went off |
174 | // into a less-likely path. So, the wisest course of action is to simply not |
175 | // OSR at this time. |
176 | |
177 | for (size_t argument = 0; argument < entry->m_expectedValues.numberOfArguments(); ++argument) { |
178 | JSValue value; |
179 | if (!argument) |
180 | value = callFrame->thisValue(); |
181 | else |
182 | value = callFrame->argument(argument - 1); |
183 | |
184 | if (!entry->m_expectedValues.argument(argument).validateOSREntryValue(value, FlushedJSValue)) { |
185 | if (Options::verboseOSR()) { |
186 | dataLog( |
187 | " OSR failed because argument " , argument, " is " , value, |
188 | ", expected " , entry->m_expectedValues.argument(argument), ".\n" ); |
189 | } |
190 | return nullptr; |
191 | } |
192 | } |
193 | |
194 | for (size_t local = 0; local < entry->m_expectedValues.numberOfLocals(); ++local) { |
195 | int localOffset = virtualRegisterForLocal(local).offset(); |
196 | JSValue value = callFrame->registers()[localOffset].asanUnsafeJSValue(); |
197 | FlushFormat format = FlushedJSValue; |
198 | |
199 | if (entry->m_localsForcedAnyInt.get(local)) { |
200 | if (!value.isAnyInt()) { |
201 | dataLogLnIf(Options::verboseOSR(), |
202 | " OSR failed because variable " , localOffset, " is " , |
203 | value, ", expected " , |
204 | "machine int." ); |
205 | return nullptr; |
206 | } |
207 | value = jsDoubleNumber(value.asAnyInt()); |
208 | format = FlushedInt52; |
209 | } |
210 | |
211 | if (entry->m_localsForcedDouble.get(local)) { |
212 | if (!value.isNumber()) { |
213 | dataLogLnIf(Options::verboseOSR(), |
214 | " OSR failed because variable " , localOffset, " is " , |
215 | value, ", expected number." ); |
216 | return nullptr; |
217 | } |
218 | value = jsDoubleNumber(value.asNumber()); |
219 | format = FlushedDouble; |
220 | } |
221 | |
222 | if (!entry->m_expectedValues.local(local).validateOSREntryValue(value, format)) { |
223 | dataLogLnIf(Options::verboseOSR(), |
224 | " OSR failed because variable " , VirtualRegister(localOffset), " is " , |
225 | value, ", expected " , |
226 | entry->m_expectedValues.local(local), "." ); |
227 | return nullptr; |
228 | } |
229 | } |
230 | |
231 | // 2) Check the stack height. The DFG JIT may require a taller stack than the |
232 | // baseline JIT, in some cases. If we can't grow the stack, then don't do |
233 | // OSR right now. That's the only option we have unless we want basic block |
234 | // boundaries to start throwing RangeErrors. Although that would be possible, |
235 | // it seems silly: you'd be diverting the program to error handling when it |
236 | // would have otherwise just kept running albeit less quickly. |
237 | |
238 | unsigned frameSizeForCheck = jitCode->common.requiredRegisterCountForExecutionAndExit(); |
239 | if (UNLIKELY(!vm.ensureStackCapacityFor(&callFrame->registers()[virtualRegisterForLocal(frameSizeForCheck - 1).offset()]))) { |
240 | if (Options::verboseOSR()) |
241 | dataLogF(" OSR failed because stack growth failed.\n" ); |
242 | return nullptr; |
243 | } |
244 | |
245 | if (Options::verboseOSR()) |
246 | dataLogF(" OSR should succeed.\n" ); |
247 | |
248 | // At this point we're committed to entering. We will do some work to set things up, |
249 | // but we also rely on our caller recognizing that when we return a non-null pointer, |
250 | // that means that we're already past the point of no return and we must succeed at |
251 | // entering. |
252 | |
253 | // 3) Set up the data in the scratch buffer and perform data format conversions. |
254 | |
255 | unsigned frameSize = jitCode->common.frameRegisterCount; |
256 | unsigned baselineFrameSize = entry->m_expectedValues.numberOfLocals(); |
257 | unsigned maxFrameSize = std::max(frameSize, baselineFrameSize); |
258 | |
259 | Register* scratch = bitwise_cast<Register*>(vm.scratchBufferForSize(sizeof(Register) * (2 + CallFrame::headerSizeInRegisters + maxFrameSize))->dataBuffer()); |
260 | |
261 | *bitwise_cast<size_t*>(scratch + 0) = frameSize; |
262 | |
263 | void* targetPC = entry->m_machineCode.executableAddress(); |
264 | RELEASE_ASSERT(codeBlock->jitCode()->contains(entry->m_machineCode.untaggedExecutableAddress())); |
265 | if (Options::verboseOSR()) |
266 | dataLogF(" OSR using target PC %p.\n" , targetPC); |
267 | RELEASE_ASSERT(targetPC); |
268 | *bitwise_cast<void**>(scratch + 1) = retagCodePtr(targetPC, OSREntryPtrTag, bitwise_cast<PtrTag>(callFrame)); |
269 | |
270 | Register* pivot = scratch + 2 + CallFrame::headerSizeInRegisters; |
271 | |
272 | for (int index = -CallFrame::headerSizeInRegisters; index < static_cast<int>(baselineFrameSize); ++index) { |
273 | VirtualRegister reg(-1 - index); |
274 | |
275 | if (reg.isLocal()) { |
276 | if (entry->m_localsForcedDouble.get(reg.toLocal())) { |
277 | *bitwise_cast<double*>(pivot + index) = callFrame->registers()[reg.offset()].asanUnsafeJSValue().asNumber(); |
278 | continue; |
279 | } |
280 | |
281 | if (entry->m_localsForcedAnyInt.get(reg.toLocal())) { |
282 | *bitwise_cast<int64_t*>(pivot + index) = callFrame->registers()[reg.offset()].asanUnsafeJSValue().asAnyInt() << JSValue::int52ShiftAmount; |
283 | continue; |
284 | } |
285 | } |
286 | |
287 | pivot[index] = callFrame->registers()[reg.offset()].asanUnsafeJSValue(); |
288 | } |
289 | |
290 | // 4) Reshuffle those registers that need reshuffling. |
291 | Vector<JSValue> temporaryLocals(entry->m_reshufflings.size()); |
292 | for (unsigned i = entry->m_reshufflings.size(); i--;) |
293 | temporaryLocals[i] = pivot[VirtualRegister(entry->m_reshufflings[i].fromOffset).toLocal()].asanUnsafeJSValue(); |
294 | for (unsigned i = entry->m_reshufflings.size(); i--;) |
295 | pivot[VirtualRegister(entry->m_reshufflings[i].toOffset).toLocal()] = temporaryLocals[i]; |
296 | |
297 | // 5) Clear those parts of the call frame that the DFG ain't using. This helps GC on |
298 | // some programs by eliminating some stale pointer pathologies. |
299 | for (unsigned i = frameSize; i--;) { |
300 | if (entry->m_machineStackUsed.get(i)) |
301 | continue; |
302 | pivot[i] = JSValue(); |
303 | } |
304 | |
305 | // 6) Copy our callee saves to buffer. |
306 | #if NUMBER_OF_CALLEE_SAVES_REGISTERS > 0 |
307 | const RegisterAtOffsetList* registerSaveLocations = codeBlock->calleeSaveRegisters(); |
308 | RegisterAtOffsetList* allCalleeSaves = RegisterSet::vmCalleeSaveRegisterOffsets(); |
309 | RegisterSet dontSaveRegisters = RegisterSet(RegisterSet::stackRegisters(), RegisterSet::allFPRs()); |
310 | |
311 | unsigned registerCount = registerSaveLocations->size(); |
312 | VMEntryRecord* record = vmEntryRecord(vm.topEntryFrame); |
313 | for (unsigned i = 0; i < registerCount; i++) { |
314 | RegisterAtOffset currentEntry = registerSaveLocations->at(i); |
315 | if (dontSaveRegisters.get(currentEntry.reg())) |
316 | continue; |
317 | RegisterAtOffset* calleeSavesEntry = allCalleeSaves->find(currentEntry.reg()); |
318 | |
319 | *(bitwise_cast<intptr_t*>(pivot - 1) - currentEntry.offsetAsIndex()) = record->calleeSaveRegistersBuffer[calleeSavesEntry->offsetAsIndex()]; |
320 | } |
321 | #endif |
322 | |
323 | // 7) Fix the call frame to have the right code block. |
324 | |
325 | *bitwise_cast<CodeBlock**>(pivot - 1 - CallFrameSlot::codeBlock) = codeBlock; |
326 | |
327 | if (Options::verboseOSR()) |
328 | dataLogF(" OSR returning data buffer %p.\n" , scratch); |
329 | return scratch; |
330 | } |
331 | |
332 | MacroAssemblerCodePtr<ExceptionHandlerPtrTag> prepareCatchOSREntry(VM& vm, CallFrame* callFrame, CodeBlock* baselineCodeBlock, CodeBlock* optimizedCodeBlock, BytecodeIndex bytecodeIndex) |
333 | { |
334 | ASSERT(optimizedCodeBlock->jitType() == JITType::DFGJIT || optimizedCodeBlock->jitType() == JITType::FTLJIT); |
335 | ASSERT(optimizedCodeBlock->jitCode()->dfgCommon()->isStillValid); |
336 | |
337 | if (!Options::useOSREntryToDFG() && optimizedCodeBlock->jitCode()->jitType() == JITType::DFGJIT) |
338 | return nullptr; |
339 | if (!Options::useOSREntryToFTL() && optimizedCodeBlock->jitCode()->jitType() == JITType::FTLJIT) |
340 | return nullptr; |
341 | |
342 | CommonData* dfgCommon = optimizedCodeBlock->jitCode()->dfgCommon(); |
343 | RELEASE_ASSERT(dfgCommon); |
344 | DFG::CatchEntrypointData* catchEntrypoint = dfgCommon->catchOSREntryDataForBytecodeIndex(bytecodeIndex); |
345 | if (!catchEntrypoint) { |
346 | // This can be null under some circumstances. The most common is that we didn't |
347 | // compile this op_catch as an entrypoint since it had never executed when starting |
348 | // the compilation. |
349 | return nullptr; |
350 | } |
351 | |
352 | // We're only allowed to OSR enter if we've proven we have compatible argument types. |
353 | for (unsigned argument = 0; argument < catchEntrypoint->argumentFormats.size(); ++argument) { |
354 | JSValue value = callFrame->uncheckedR(virtualRegisterForArgument(argument)).jsValue(); |
355 | switch (catchEntrypoint->argumentFormats[argument]) { |
356 | case DFG::FlushedInt32: |
357 | if (!value.isInt32()) |
358 | return nullptr; |
359 | break; |
360 | case DFG::FlushedCell: |
361 | if (!value.isCell()) |
362 | return nullptr; |
363 | break; |
364 | case DFG::FlushedBoolean: |
365 | if (!value.isBoolean()) |
366 | return nullptr; |
367 | break; |
368 | case DFG::DeadFlush: |
369 | // This means the argument is not alive. Therefore, it's allowed to be any type. |
370 | break; |
371 | case DFG::FlushedJSValue: |
372 | // An argument is trivially a JSValue. |
373 | break; |
374 | default: |
375 | RELEASE_ASSERT_NOT_REACHED(); |
376 | } |
377 | } |
378 | |
379 | unsigned frameSizeForCheck = dfgCommon->requiredRegisterCountForExecutionAndExit(); |
380 | if (UNLIKELY(!vm.ensureStackCapacityFor(&callFrame->registers()[virtualRegisterForLocal(frameSizeForCheck).offset()]))) |
381 | return nullptr; |
382 | |
383 | auto instruction = baselineCodeBlock->instructions().at(callFrame->bytecodeIndex()); |
384 | ASSERT(instruction->is<OpCatch>()); |
385 | ValueProfileAndOperandBuffer* buffer = instruction->as<OpCatch>().metadata(baselineCodeBlock).m_buffer; |
386 | JSValue* dataBuffer = reinterpret_cast<JSValue*>(dfgCommon->catchOSREntryBuffer->dataBuffer()); |
387 | unsigned index = 0; |
388 | buffer->forEach([&] (ValueProfileAndOperand& profile) { |
389 | if (!VirtualRegister(profile.m_operand).isLocal()) |
390 | return; |
391 | dataBuffer[index] = callFrame->uncheckedR(profile.m_operand).jsValue(); |
392 | ++index; |
393 | }); |
394 | |
395 | // The active length of catchOSREntryBuffer will be zeroed by ClearCatchLocals node. |
396 | dfgCommon->catchOSREntryBuffer->setActiveLength(sizeof(JSValue) * index); |
397 | return catchEntrypoint->machineCode; |
398 | } |
399 | |
400 | } } // namespace JSC::DFG |
401 | |
402 | #endif // ENABLE(DFG_JIT) |
403 | |