JavaScriptCore should discard baseline code after some time
https://bugs.webkit.org/show_bug.cgi?id=149220
Reviewed by Saam Barati.
This is a bit more complicated than discarding optimized code because
the engine previously assumed that we would never discard baseline code.
(JSC::CodeBlock::CodeBlock): Record creation time (and compute time since
creation) instead of install time because CodeBlocks can be installed
more than once, and we don't want to have to worry about edge cases
created by CodeBlocks seeming to get younger.
(JSC::CodeBlock::visitAggregate): Be explicit about only doing the
weak reference fixpoint for optimized CodeBlocks. We used to avoid the
fixpoint for baseline CodeBlocks implicitly, since they would always
visit themselves strongly right away. But now baseline CodeBlocks might
not visit themselves strongly, since they might choose to jettison due
to old age.
(JSC::CodeBlock::shouldVisitStrongly): Add old age as a reason not to
visit ourselves strongly, so that baseline CodeBlocks can jettison due
to old age.
(JSC::CodeBlock::shouldJettisonDueToWeakReference): Be explicit about
only jettisoning optimized CodeBlocks due to weak references so that we
don't confuse ourselves into thinking that we will jettison a baseline
CodeBlock due to weak references.
(JSC::CodeBlock::shouldJettisonDueToOldAge): Updated to use creation time.
(JSC::CodeBlock::visitOSRExitTargets): Clarify a comment and add an
ASSERT to help record some things I discovered while debugging.
(JSC::CodeBlock::jettison): Allow a baseline CodeBlock to jettison. Don't
assume that we have an alternative or a profiler.
(JSC::CodeBlock::install): Deleted.
(JSC::CodeBlock::releaseAlternative): Deleted.
(JSC::CodeBlock::setInstallTime): Deleted.
(JSC::CodeBlock::timeSinceInstall): Deleted.
- dfg/DFGOSRExitPreparation.cpp:
(JSC::DFG::prepareCodeOriginForOSRExit): Simplified the computation of
baseline CodeBlock.
(JSC::DFG::Plan::checkLivenessAndVisitChildren): Be sure to strongly
visit our inline callframes because we assume that an optimized CodeBlock
will keep its OSR exit targets alive, but the CodeBlock object won't be
able to mark them for itself until compilation has completed (since it
won't have a JITCode object yet).
- dfg/DFGToFTLDeferredCompilationCallback.cpp:
(JSC::DFG::ToFTLDeferredCompilationCallback::compilationDidComplete):
Updated for interface change.
(JSC::JITCode::timeToLive): Provide a time to live for interpreter and
baseline code, so they will jettison when old. Use seconds in our
code so that we don't need comments. Make DFG 2X interpreter+baseline,
and FTL 2X DFG+interpreter+baseline, also matching the time we allot
before throwing away all code.
- jit/JITToDFGDeferredCompilationCallback.cpp:
(JSC::JITToDFGDeferredCompilationCallback::compilationDidComplete):
- llint/LLIntSlowPaths.cpp:
(JSC::LLInt::jitCompileAndSetHeuristics): Updated for interface change.
(JSC::ScriptExecutable::installCode): Allow our caller to install nullptr,
since we need to do this when jettisoning a baseline CodeBlock. Require
our caller to specify the details of the installation because we can't
rely on a non-null CodeBlock in order to compute them.
(JSC::ScriptExecutable::newCodeBlockFor):
(JSC::ScriptExecutable::prepareForExecutionImpl):
(JSC::ScriptExecutable::recordParse): Updated for interface change.
- runtime/Options.h: Renamed the CodeBlock liveness option since it now
controls baseline and optimized code.
