Name Description Size
ABIArgGenerator.h jit_ABIArgGenerator_h 2635
ABIFunctionList-inl.h 13786
ABIFunctions.h jit_VMFunctions_h 2888
ABIFunctionType.h jit_ABIFunctionType_h 1651
ABIFunctionType.yaml 5870
AliasAnalysis.cpp 11355
AliasAnalysis.h jit_AliasAnalysis_h 1479
AlignmentMaskAnalysis.cpp 3231
AlignmentMaskAnalysis.h namespace jit 694
Assembler.h jit_Assembler_h 1174
AtomicOp.h jit_AtomicOp_h 3041
AtomicOperations.h [SMDOC] Atomic Operations The atomic operations layer defines types and functions for JIT-compatible atomic operation. The fundamental constraints on the functions are: - That their realization here MUST be compatible with code the JIT generates for its Atomics operations, so that an atomic access from the interpreter or runtime - from any C++ code - really is atomic relative to a concurrent, compatible atomic access from jitted code. That is, these primitives expose JIT-compatible atomicity functionality to C++. - That accesses may race without creating C++ undefined behavior: atomic accesses (marked "SeqCst") may race with non-atomic accesses (marked "SafeWhenRacy"); overlapping but non-matching, and hence incompatible, atomic accesses may race; and non-atomic accesses may race. The effects of races need not be predictable, so garbage can be produced by a read or written by a write, but the effects must be benign: the program must continue to run, and only the memory in the union of addresses named in the racing accesses may be affected. The compatibility constraint means that if the JIT makes dynamic decisions about how to implement atomic operations then corresponding dynamic decisions MUST be made in the implementations of the functions below. The safe-for-races constraint means that by and large, it is hard to implement these primitives in C++. See "Implementation notes" below. The "SeqCst" suffix on operations means "sequentially consistent" and means such a function's operation must have "sequentially consistent" memory ordering. See mfbt/Atomics.h for an explanation of this memory ordering. Note that a "SafeWhenRacy" access does not provide the atomicity of a "relaxed atomic" access: it can read or write garbage if there's a race. Implementation notes. It's not a requirement that these functions be inlined; performance is not a great concern. On some platforms these functions may call functions that use inline assembly. See In principle these functions will not be written in C++, thus making races defined behavior if all racy accesses from C++ go via these functions. (Jitted code will always be safe for races and provides the same guarantees as these functions.) The appropriate implementations will be platform-specific and there are some obvious implementation strategies to choose from, sometimes a combination is appropriate: - generating the code at run-time with the JIT; - hand-written assembler (maybe inline); or - using special compiler intrinsics or directives. Trusting the compiler not to generate code that blows up on a race definitely won't work in the presence of TSan, or even of optimizing compilers in seemingly-"innocuous" conditions. (See for details.) 13065
AutoWritableJitCode.h jit_AutoWritableJitCode_h 3424
BacktrackingAllocator.cpp 169753
BacktrackingAllocator.h 30204
Bailouts.cpp 13258
Bailouts.h 8583
BaselineBailouts.cpp BaselineStackBuilder helps abstract the process of rebuilding the C stack on the heap. It takes a bailout iterator and keeps track of the point on the C stack from which the reconstructed frames will be written. It exposes methods to write data into the heap memory storing the reconstructed stack. It also exposes method to easily calculate addresses. This includes both the virtual address that a particular value will be at when it's eventually copied onto the stack, as well as the current actual address of that value (whether on the heap allocated portion being constructed or the existing stack). The abstraction handles transparent re-allocation of the heap memory when it needs to be enlarged to accommodate new data. Similarly to the C stack, the data that's written to the reconstructed stack grows from high to low in memory. The lowest region of the allocated memory contains a BaselineBailoutInfo structure that points to the start and end of the written data. 76516
BaselineCacheIRCompiler.cpp 149566
BaselineCacheIRCompiler.h 7041
BaselineCodeGen.cpp HandlerArgs = 201650
BaselineCodeGen.h 17977
BaselineDebugModeOSR.cpp 19251
BaselineDebugModeOSR.h 962
BaselineFrame-inl.h jit_BaselineFrame_inl_h 4616
BaselineFrame.cpp 5415
BaselineFrame.h 14566
BaselineFrameInfo-inl.h jit_BaselineFrameInfo_inl_h 1372
BaselineFrameInfo.cpp 6453
BaselineFrameInfo.h 12999
BaselineIC.cpp 82007
BaselineIC.h 17363
BaselineICList.h jit_BaselineICList_h 2145
BaselineJIT.cpp 32276
BaselineJIT.h 20317
BitSet.cpp 2627
BitSet.h jit_BitSet_h 4250
BranchHinting.cpp 2236
BranchHinting.h jit_BranchHinting_h 632
BytecodeAnalysis.cpp stackDepth= 11016
BytecodeAnalysis.h jit_BytecodeAnalysis_h 2276
CacheIR.cpp 480459
CacheIR.h 17010
CacheIRCloner.h jit_CacheIRCloner_h 2181
CacheIRCompiler.cpp 356652
CacheIRCompiler.h 49808
CacheIRGenerator.h 38406
CacheIRHealth.cpp 13545
CacheIRHealth.h JS_CACHEIR_SPEW 4484
CacheIROps.yaml 58662
CacheIRReader.h 5212
CacheIRSpewer.cpp 13569
CacheIRSpewer.h JS_CACHEIR_SPEW 3391
CacheIRWriter.h 23643
CalleeToken.h namespace js::jit 2242
CodeGenerator.cpp 744864
CodeGenerator.h 22163
CompactBuffer.h 6485
CompileInfo.h env chain and argument obj 14868
CompileWrappers.cpp static 7151
CompileWrappers.h 4387
Disassemble.cpp 4064
Disassemble.h jit_Disassemble_h 680
EdgeCaseAnalysis.cpp 1487
EdgeCaseAnalysis.h jit_EdgeCaseAnalysis_h 700
EffectiveAddressAnalysis.cpp 7150
EffectiveAddressAnalysis.h namespace jit 874
ExecutableAllocator.cpp willDestroy = 10594
ExecutableAllocator.h 6370
FixedList.h jit_FixedList_h 2249
FlushICache.cpp 4642
FlushICache.h Flush the instruction cache of instructions in an address range. 3355
FoldLinearArithConstants.cpp namespace jit 3698
FoldLinearArithConstants.h namespace jit 639 \ /* This Source Code Form is subject to the terms of the Mozilla Public * License, v. 2.0. If a copy of the MPL was not distributed with this * file, You can obtain one at */ #ifndef %(includeguard)s #define %(includeguard)s /* This file is generated by jit/ Do not edit! */ %(contents)s #endif // %(includeguard)s 17384 INLINE_ATTR void %(fun_name)s() { asm volatile ("mfence\n\t" ::: "memory"); } 34532 \ /* This Source Code Form is subject to the terms of the Mozilla Public * License, v. 2.0. If a copy of the MPL was not distributed with this * file, You can obtain one at */ #ifndef %(includeguard)s #define %(includeguard)s /* This file is generated by jit/ Do not edit! */ %(contents)s #endif // %(includeguard)s 20635 \ /* This Source Code Form is subject to the terms of the Mozilla Public * License, v. 2.0. If a copy of the MPL was not distributed with this * file, You can obtain one at */ #ifndef %(includeguard)s #define %(includeguard)s /* This file is generated by jit/ Do not edit! */ %(contents)s #endif // %(includeguard)s 9144 \ /* This Source Code Form is subject to the terms of the Mozilla Public * License, v. 2.0. If a copy of the MPL was not distributed with this * file, You can obtain one at */ #ifndef %(includeguard)s #define %(includeguard)s /* This file is generated by jit/ Do not edit! */ %(contents)s #endif // %(includeguard)s 12570
ICState.h 7378
ICStubSpace.h jit_ICStubSpace_h 1158
InlinableNatives.cpp 13634
InlinableNatives.h 12321
InlineList.h 15541
InlineScriptTree-inl.h jit_InlineScriptTree_inl_h 2269
InlineScriptTree.h jit_InlineScriptTree_h 4163
InstructionReordering.cpp 8278
InstructionReordering.h 593
InterpreterEntryTrampoline.cpp countIncludesThis = 8571
InterpreterEntryTrampoline.h The EntryTrampolineMap is used to cache the trampoline code for each script as they are created. These trampolines are created only under --emit-interpreter-entry and are used to identify which script is being interpeted when profiling with external profilers such as perf. The map owns the JitCode objects that are created for each script, and keeps them alive at least as long as the script associated with it in case we need to re-enter the trampoline again. As each script is finalized, the entry is manually removed from the table in BaseScript::finalize which will also release the trampoline code associated with it. During a moving GC, the table is rekeyed in case any scripts have relocated. 2388
Invalidation.h jit_Invalidation_h 1891
Ion.cpp 84450
Ion.h this 4371
IonAnalysis.cpp 180574
IonAnalysis.h 6627
IonCacheIRCompiler.cpp 79855
IonCacheIRCompiler.h jit_IonCacheIRCompiler_h 2900
IonCompileTask.cpp 6584
IonCompileTask.h jit_IonCompileTask_h 3174
IonGenericCallStub.h jit_IonGenericCallStub_h 1206
IonIC.cpp static 21677
IonIC.h 21327
IonOptimizationLevels.cpp 5400
IonOptimizationLevels.h 5811
IonScript.h 19482
IonTypes.h [SMDOC] Avoiding repeated bailouts / invalidations To avoid getting trapped in a "compilation -> bailout -> invalidation -> recompilation -> bailout -> invalidation -> ..." loop, every snapshot in Warp code is assigned a BailoutKind. If we bail out at that snapshot, FinishBailoutToBaseline will examine the BailoutKind and take appropriate action. In general: 1. If the bailing instruction comes from transpiled CacheIR, then when we bail out and continue execution in the baseline interpreter, the corresponding stub should fail a guard. As a result, we will either increment the enteredCount for a subsequent stub or attach a new stub, either of which will prevent WarpOracle from transpiling the failing stub when we recompile. Note: this means that every CacheIR op that can bail out in Warp must have an equivalent guard in the baseline CacheIR implementation. FirstExecution works according to the same principles: we have never hit this IC before, but after we bail to baseline we will attach a stub and recompile with better CacheIR information. 2. If the bailout occurs because an assumption we made in WarpBuilder was invalidated, then FinishBailoutToBaseline will set a flag on the script to avoid that assumption in the future: for example, UninitializedLexical. 3. Similarly, if the bailing instruction is generated or modified by a MIR optimization, then FinishBailoutToBaseline will set a flag on the script to make that optimization more conservative in the future. Examples include LICM, EagerTruncation, and HoistBoundsCheck. 4. Some bailouts can't be handled in Warp, even after a recompile. For example, Warp does not support catching exceptions. If this happens too often, then the cost of bailing out repeatedly outweighs the benefit of Warp compilation, so we invalidate the script and disable Warp compilation. 5. Some bailouts don't happen in performance-sensitive code: for example, the |debugger| statement. We just ignore those. 27176
Jit.cpp osrFrame = 8248
Jit.h jit_Jit_h 1149
JitAllocPolicy.h 5279
JitCode.h 5902
JitcodeMap.cpp 38874
JitcodeMap.h The jitcode map implements tables to allow mapping from addresses in jitcode to the list of (JSScript*, jsbytecode*) pairs that are implicitly active in the frame at that point in the native code. To represent this information efficiently, a multi-level table is used. At the top level, a global AVL-tree of JitcodeGlobalEntry describing the mapping for each individual JitCode generated by compiles. The entries are ordered by their nativeStartAddr. Every entry in the table is of fixed size, but there are different entry types, distinguished by the kind field. 26790
JitCommon.h 2292
JitContext.cpp 3757
JitContext.h jit_JitContext_h 4375
Jitdump.h This file provides the necessary data structures to meet the JitDump specification as of 1592
JitFrames-inl.h jit_JitFrames_inl_h 862
JitFrames.cpp 88688
JitHints-inl.h jit_JitHints_inl_h 1949
JitHints.cpp 4697
JitHints.h [SMDOC] JitHintsMap The Jit hints map is an in process cache used to collect Baseline and Ion JIT hints to try and skip as much of the warmup as possible and jump straight into those tiers. Whenever a script enters one of these tiers a hint is recorded in this cache using the script's filename+sourceStart value, and if we ever encounter this script again later, e.g. during a navigation, then we try to eagerly compile it into baseline and ion based on its previous execution history. 6623
JitOptions.cpp 16966
JitOptions.h 6014
JitRuntime.h 18580
JitScript-inl.h jit_JitScript_inl_h 1144
JitScript.cpp depth= 29116
JitScript.h [SMDOC] ICScript Lifetimes An ICScript owns an array of ICEntries, each of which owns a linked list of ICStubs. A JitScript contains an embedded ICScript. If it has done any trial inlining, it also owns an InliningRoot. The InliningRoot owns all of the ICScripts that have been created for inlining into the corresponding JitScript. This ties the lifetime of the inlined ICScripts to the lifetime of the JitScript itself. We store pointers to ICScripts in two other places: on the stack in BaselineFrame, and in IC stubs for CallInlinedFunction. The ICScript pointer in a BaselineFrame either points to the ICScript embedded in the JitScript for that frame, or to an inlined ICScript owned by a caller. In each case, there must be a frame on the stack corresponding to the JitScript that owns the current ICScript, which will keep the ICScript alive. Each ICStub is owned by an ICScript and, indirectly, a JitScript. An ICStub that uses CallInlinedFunction contains an ICScript for use by the callee. The ICStub and the callee ICScript are always owned by the same JitScript, so the callee ICScript will not be freed while the ICStub is alive. The lifetime of an ICScript is independent of the lifetimes of the BaselineScript and IonScript/WarpScript to which it corresponds. They can be destroyed and recreated, and the ICScript will remain valid. When we discard JIT code, we mark ICScripts that are active on the stack as active and then purge all of the inactive ICScripts. We also purge ICStubs, including the CallInlinedFunction stub at the trial inining call site, and reset the ICStates to allow trial inlining again later. If there's a BaselineFrame for an inlined ICScript, we'll preserve both this ICScript and the IC chain for the call site in the caller's ICScript. See ICScript::purgeStubs and ICScript::purgeInactiveICScripts. 19911
JitSpewer.cpp 19178
JitSpewer.h Information during sinking 10190
JitZone.h 13403
JSJitFrameIter-inl.h jit_JSJitFrameIter_inl_h 1806
JSJitFrameIter.cpp 25953
JSJitFrameIter.h 27442
JSONSpewer.cpp 6788
JSONSpewer.h JS_JITSPEW 1192
KnownClass.cpp 3243
KnownClass.h 879
Label.cpp 883
Label.h 3254
LICM.cpp OUT 13204
LICM.h jit_LICM_h 629
Linker.cpp 2354
Linker.h jit_Linker_h 1248
LIR.cpp 21412
LIR.h 65031
LIROps.yaml 85491
Lowering.cpp useAtStart = 259425
Lowering.h useAtStart = 2998
MachineState.h jit_MachineState_h 3636
MacroAssembler-inl.h 39227
MacroAssembler.cpp 327262
MacroAssembler.h 270304
MIR-wasm.cpp 33084
MIR-wasm.h Everything needed to build actual MIR instructions: the actual opcodes and instructions, the instruction interface, and use chains. 100095
MIR.cpp 198762
MIR.h Everything needed to build actual MIR instructions: the actual opcodes and instructions, the instruction interface, and use chains. 294421
MIRGenerator.h 5517
MIRGraph.cpp 42084
MIRGraph.h 30870
MIROps.yaml 72817
MoveEmitter.h jit_MoveEmitter_h 1172
MoveResolver.cpp 14054
MoveResolver.h 9924 10078
PerfSpewer.cpp 35857
PerfSpewer.h 6862
ProcessExecutableMemory.cpp Inspiration is V8's OS::Allocate in VirtualAlloc takes 64K chunks out of the virtual address space, so we keep 16b alignment. x86: V8 comments say that keeping addresses in the [64MiB, 1GiB) range tries to avoid system default DLL mapping space. In the end, we get 13 bits of randomness in our selection. x64: [2GiB, 4TiB), with 25 bits of randomness. 32459
ProcessExecutableMemory.h 5307
RangeAnalysis.cpp 120153
RangeAnalysis.h 25935
ReciprocalMulConstants.cpp 4472
ReciprocalMulConstants.h jit_ReciprocalMulConstants_h 990
Recover.cpp 62054
Recover.h 28387
RegExpStubConstants.h jit_RegExpStubConstants_h 1202
RegisterAllocator.cpp 22457
RegisterAllocator.h 10945
Registers.h 9403
RegisterSets.h 40255
RematerializedFrame-inl.h 746
RematerializedFrame.cpp static 6430
RematerializedFrame.h 7432
SafepointIndex-inl.h jit_SafepointIndex_inl_h 694
SafepointIndex.cpp 732
SafepointIndex.h namespace jit 2376
Safepoints.cpp 17714
Safepoints.h jit_Safepoints_h 4253
ScalarReplacement.cpp 96346
ScalarReplacement.h jit_ScalarReplacement_h 675
ScalarTypeUtils.h jit_ScalarTypeUtils_h 1328
ScriptFromCalleeToken.h namespace js::jit 1023
SharedICHelpers-inl.h jit_SharedICHelpers_inl_h 1328
SharedICHelpers.h jit_SharedICHelpers_h 1280
SharedICRegisters.h jit_SharedICRegisters_h 1348
ShuffleAnalysis.cpp 28239
ShuffleAnalysis.h 4756
Simulator.h jit_Simulator_h 1037
Sink.cpp 9483
Sink.h jit_Sink_h 598
Snapshots.cpp 20057
Snapshots.h 16031
StackSlotAllocator.h jit_StackSlotAllocator_h 3500
TemplateObject-inl.h jit_TemplateObject_inl_h 3392
TemplateObject.h jit_TemplateObject_h 2583
Trampoline.cpp 10343
TrampolineNatives.cpp 11299
TrampolineNatives.h jit_TrampolineNatives_h 1925
TrialInlining.cpp 32558
TrialInlining.h [SMDOC] Trial Inlining WarpBuilder relies on transpiling CacheIR. When inlining scripted functions in WarpBuilder, we want our ICs to be as monomorphic as possible. Functions with multiple callers complicate this. An IC in such a function might be monomorphic for any given caller, but polymorphic overall. This make the input to WarpBuilder less precise. To solve this problem, we do trial inlining. During baseline execution, we identify call sites for which it would be useful to have more precise inlining data. For each such call site, we allocate a fresh ICScript and replace the existing call IC with a new specialized IC that invokes the callee using the new ICScript. Other callers of the callee will continue using the default ICScript. When we eventually Warp-compile the script, we can generate code for the callee using the IC information in our private ICScript, which is specialized for its caller. The same approach can be used to inline recursively. 6020
TypeData.h jit_TypeData_h 1314
TypePolicy.cpp 43498
TypePolicy.h 19548
ValueNumbering.cpp [SMDOC] IonMonkey Value Numbering Some notes on the main algorithm here: - The SSA identifier id() is the value number. We do replaceAllUsesWith as we go, so there's always at most one visible value with a given number. - Consequently, the GVN algorithm is effectively pessimistic. This means it is not as powerful as an optimistic GVN would be, but it is simpler and faster. - We iterate in RPO, so that when visiting a block, we've already optimized and hashed all values in dominating blocks. With occasional exceptions, this allows us to do everything in a single pass. - When we do use multiple passes, we just re-run the algorithm on the whole graph instead of doing sparse propagation. This is a tradeoff to keep the algorithm simpler and lighter on inputs that don't have a lot of interesting unreachable blocks or degenerate loop induction variables, at the expense of being slower on inputs that do. The loop for this always terminates, because it only iterates when code is or will be removed, so eventually it must stop iterating. - Values are not immediately removed from the hash set when they go out of scope. Instead, we check for dominance after a lookup. If the dominance check fails, the value is removed. 47043
ValueNumbering.h jit_ValueNumbering_h 4635
VMFunctionList-inl.h 26955
VMFunctions.cpp Unexpected return type for a VMFunction. 100443
VMFunctions.h 28069
WarpBuilder.cpp maybePred = 111087
WarpBuilder.h Intentionally not implemented 13561
WarpBuilderShared.cpp 4291
WarpBuilderShared.h 12609
WarpCacheIRTranspiler.cpp 203157
WarpCacheIRTranspiler.h jit_WarpCacheIRTranspiler_h 905
WarpOracle.cpp 46801
WarpOracle.h jit_WarpOracle_h 2508
WarpSnapshot.cpp 13635
WarpSnapshot.h 18778
WasmBCE.cpp 4949
WasmBCE.h jit_wasmbce_h 964
XrayJitInfo.cpp 622