Name Description Size
ABIArgGenerator.h jit_ABIArgGenerator_h 2633
ABIFunctionList-inl.h 12758
ABIFunctions.h jit_VMFunctions_h 2678
AliasAnalysis.cpp 10458
AliasAnalysis.h jit_AliasAnalysis_h 1453
AlignmentMaskAnalysis.cpp 3205
AlignmentMaskAnalysis.h namespace jit 694
arm
arm64
Assembler.h jit_Assembler_h 1174
AtomicOp.h jit_AtomicOp_h 2983
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 GenerateAtomicOperations.py. 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 https://www.usenix.org/legacy/event/hotpar11/tech/final_files/Boehm.pdf for details.) 13065
AutoWritableJitCode.h jit_AutoWritableJitCode_h 2952
BacktrackingAllocator.cpp 169458
BacktrackingAllocator.h 30204
Bailouts.cpp 13221
Bailouts.h 8248
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. 71836
BaselineCacheIRCompiler.cpp 138731
BaselineCacheIRCompiler.h 5937
BaselineCodeGen.cpp HandlerArgs = 199010
BaselineCodeGen.h 17977
BaselineDebugModeOSR.cpp 19154
BaselineDebugModeOSR.h 962
BaselineFrame-inl.h jit_BaselineFrame_inl_h 3982
BaselineFrame.cpp 5415
BaselineFrame.h 13003
BaselineFrameInfo-inl.h jit_BaselineFrameInfo_inl_h 1372
BaselineFrameInfo.cpp 6453
BaselineFrameInfo.h 13008
BaselineIC.cpp 77642
BaselineIC.h 16674
BaselineICList.h jit_BaselineICList_h 2037
BaselineJIT.cpp 32312
BaselineJIT.h 20087
BitSet.cpp 2627
BitSet.h jit_BitSet_h 4250
BytecodeAnalysis.cpp stackDepth= 8035
BytecodeAnalysis.h jit_BytecodeAnalysis_h 2276
CacheIR.cpp 432915
CacheIR.h 16384
CacheIRCloner.h jit_CacheIRCloner_h 2071
CacheIRCompiler.cpp 316267
CacheIRCompiler.h 45392
CacheIRGenerator.h 35932
CacheIRHealth.cpp 13402
CacheIRHealth.h JS_CACHEIR_SPEW 4484
CacheIROps.yaml 52410
CacheIRReader.h isSpread = 4947
CacheIRSpewer.cpp 12267
CacheIRSpewer.h JS_CACHEIR_SPEW 3217
CacheIRWriter.h 22081
CalleeToken.h namespace js::jit 2242
CodeGenerator.cpp 626460
CodeGenerator.h 17415
CompactBuffer.h 7020
CompileInfo.h env chain and argument obj 14292
CompileWrappers.cpp static 6453
CompileWrappers.h 4166
Disassemble.cpp 3139
Disassemble.h jit_Disassemble_h 680
EdgeCaseAnalysis.cpp 1461
EdgeCaseAnalysis.h jit_EdgeCaseAnalysis_h 700
EffectiveAddressAnalysis.cpp 7121
EffectiveAddressAnalysis.h namespace jit 874
ExecutableAllocator.cpp willDestroy = 10476
ExecutableAllocator.h 6356
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 3672
FoldLinearArithConstants.h namespace jit 639
GenerateAtomicOperations.py INLINE_ATTR void %(fun_name)s() { asm volatile ("mfence\n\t" ::: "memory"); } 34831
GenerateCacheIRFiles.py \ /* 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 http://mozilla.org/MPL/2.0/. */ #ifndef %(includeguard)s #define %(includeguard)s /* This file is generated by jit/GenerateCacheIRFiles.py. Do not edit! */ %(contents)s #endif // %(includeguard)s 19838
GenerateLIRFiles.py \ /* 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 http://mozilla.org/MPL/2.0/. */ #ifndef %(includeguard)s #define %(includeguard)s /* This file is generated by jit/GenerateLIRFiles.py. Do not edit! */ %(contents)s #endif // %(includeguard)s 9678
GenerateMIRFiles.py \ /* 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 http://mozilla.org/MPL/2.0/. */ #ifndef %(includeguard)s #define %(includeguard)s /* This file is generated by jit/GenerateMIRFiles.py. Do not edit! */ %(contents)s #endif // %(includeguard)s 12977
ICState.h 7267
ICStubSpace.h jit_ICStubSpace_h 2013
InlinableNatives.cpp 12471
InlinableNatives.h 11526
InlineList.h 15541
InlineScriptTree-inl.h jit_InlineScriptTree_inl_h 2097
InlineScriptTree.h jit_InlineScriptTree_h 3338
InstructionReordering.cpp 8278
InstructionReordering.h 593
InterpreterEntryTrampoline.cpp countIncludesThis = 8621
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. 2382
Invalidation.h jit_Invalidation_h 1891
Ion.cpp 82246
Ion.h this 4414
IonAnalysis.cpp 162422
IonAnalysis.h 6561
IonCacheIRCompiler.cpp shouldDiscardStack = 73546
IonCacheIRCompiler.h jit_IonCacheIRCompiler_h 2404
IonCompileTask.cpp 6311
IonCompileTask.h jit_IonCompileTask_h 3097
IonIC.cpp static 21142
IonIC.h 20358
IonOptimizationLevels.cpp 4792
IonOptimizationLevels.h 5722
IonScript.h 19013
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. 37986
Jit.cpp osrFrame = 6769
Jit.h jit_Jit_h 1149
JitAllocPolicy.h 5279
JitCode.h 5673
JitcodeMap.cpp 38840
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 https://github.com/torvalds/linux/blob/f2906aa863381afb0015a9eb7fefad885d4e5a56/tools/perf/Documentation/jitdump-specification.txt 1592
JitFrames-inl.h jit_JitFrames_inl_h 862
JitFrames.cpp 86116
JitFrames.h HASCACHEDSAVEDFRAME_BIT 23311
JitHints-inl.h jit_JitHints_inl_h 1572
JitHints.h The JitHintsMap implements a BitBloomFilter to track whether or not a script, identified by filename+sourceStart, has been baseline compiled before in the same process. This can occur frequently during navigations. The bloom filter allows us to have very efficient storage and lookup costs, at the expense of occasional false positives. The number of entries added to the bloom filter is monitored in order to try and keep the false positivity rate below 1%. If the entry count exceeds MaxEntries_, which indicates the false positivity rate may exceed 1.5%, then the filter is completely cleared to reset the cache. 2011
JitOptions.cpp 15699
JitOptions.h 5277
JitRealm.h 5878
JitRuntime.h 15588
JitScript-inl.h jit_JitScript_inl_h 1144
JitScript.cpp depth= 21850
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. 18427
JitSpewer.cpp 19029
JitSpewer.h Information during sinking 10100
JitZone.h 5780
JSJitFrameIter-inl.h jit_JSJitFrameIter_inl_h 1806
JSJitFrameIter.cpp 24648
JSJitFrameIter.h 26221
JSONSpewer.cpp 6788
JSONSpewer.h JS_JITSPEW 1192
KnownClass.cpp 3217
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 21304
LIR.h 62523
LIROps.yaml 75902
loong64
Lowering.cpp useAtStart = 236573
Lowering.h useAtStart = 2863
MachineState.h jit_MachineState_h 3636
MacroAssembler-inl.h 38728
MacroAssembler.cpp 233593
MacroAssembler.h 248189
mips-shared
mips32
mips64
MIR.cpp 208817
MIR.h Everything needed to build actual MIR instructions: the actual opcodes and instructions, the instruction interface, and use chains. 375958
MIRGenerator.h 5465
MIRGraph.cpp 42000
MIRGraph.h 30535
MIROps.yaml 63379
MoveEmitter.h jit_MoveEmitter_h 1172
MoveResolver.cpp 14054
MoveResolver.h 9924
moz.build 9782
none
PcScriptCache.h jit_PcScriptCache_h 2458
PerfSpewer.cpp 36361
PerfSpewer.h 6818
ProcessExecutableMemory.cpp Inspiration is V8's OS::Allocate in platform-win32.cc. 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. 30070
ProcessExecutableMemory.h 4524
RangeAnalysis.cpp 119199
RangeAnalysis.h 25935
ReciprocalMulConstants.cpp 4472
ReciprocalMulConstants.h jit_ReciprocalMulConstants_h 990
Recover.cpp 61429
Recover.h 28343
RegExpStubConstants.h jit_RegExpStubConstants_h 1202
RegisterAllocator.cpp 22357
RegisterAllocator.h 10958
Registers.h 9357
RegisterSets.h 40256
RematerializedFrame-inl.h 746
RematerializedFrame.cpp static 6430
RematerializedFrame.h 7432
riscv64
SafepointIndex-inl.h jit_SafepointIndex_inl_h 694
SafepointIndex.cpp 732
SafepointIndex.h namespace jit 2376
Safepoints.cpp 16627
Safepoints.h jit_Safepoints_h 3831
ScalarReplacement.cpp 96299
ScalarReplacement.h jit_ScalarReplacement_h 675
ScalarTypeUtils.h jit_ScalarTypeUtils_h 1328
ScriptFromCalleeToken.h namespace js::jit 1023
shared
SharedICHelpers-inl.h jit_SharedICHelpers_inl_h 1328
SharedICHelpers.h jit_SharedICHelpers_h 1280
SharedICRegisters.h jit_SharedICRegisters_h 1348
ShuffleAnalysis.cpp 25145
ShuffleAnalysis.h 4568
Simulator.h jit_Simulator_h 1037
Sink.cpp 9457
Sink.h jit_Sink_h 598
Snapshots.cpp 20031
Snapshots.h 16031
StackSlotAllocator.h jit_StackSlotAllocator_h 3426
TemplateObject-inl.h jit_TemplateObject_inl_h 3392
TemplateObject.h jit_TemplateObject_h 2583
Trampoline.cpp 9349
TrialInlining.cpp 30753
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. 5982
TypeData.h jit_TypeData_h 1314
TypePolicy.cpp 43105
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. 46781
ValueNumbering.h jit_ValueNumbering_h 4635
VMFunctionList-inl.h 26036
VMFunctions.cpp Unexpected return type for a VMFunction. 92890
VMFunctions.h 27518
WarpBuilder.cpp maybePred = 107475
WarpBuilder.h Intentionally not implemented 12876
WarpBuilderShared.cpp 3327
WarpBuilderShared.h 11760
WarpCacheIRTranspiler.cpp 182091
WarpCacheIRTranspiler.h jit_WarpCacheIRTranspiler_h 905
WarpOracle.cpp 41359
WarpOracle.h jit_WarpOracle_h 2479
WarpSnapshot.cpp 13274
WarpSnapshot.h 18946
wasm32
WasmBCE.cpp 4922
WasmBCE.h jit_wasmbce_h 964
x64
x86
x86-shared
XrayJitInfo.cpp 622