• Dominik Inführ's avatar
    [codegen, heap] Improve TimeToSafepoint for concurrent compilation · d5416b99
    Dominik Inführ authored
    TimeToSafepoint is the time needed for all background threads to enter
    a safepoint after the GC was started on the main thread. This CL
    improves that metric during concurrent compilation to bytecode by doing:
    
    1) Park the LocalIsolate during
       InterpreterCompilationJob::ExecuteJobImpl. There are no concurrent
       heap accesses happening while generating bytecode for now. So instead
       of manually placing Safepoint() invocations in the code, simply park
       the local isolate.
    2) Destroy the LocalIsolate before the ReleaseParser operation. I've
       seen this take around 2ms, which regressed TimeToSafepoint a lot.
    3) Add explicit safepoints to concurrent allocations. This covers the
       rest of the code and from what I've seen so far this is good enough
       to keep TimeToSafepoint around a few microseconds.
    
    I've still seen TimeToSafepoint events with 20-80 microseconds but those
    were quite rare and always seemed to be related to Turbofan.
    
    AsLocalIsolate() is necessary in generic code to convert both Isolate
    and LocalIsolate to LocalIsolate.
    
    Bug: v8:10315
    Change-Id: Idaf9f04ffdf850d0ab0081ec372cc384a9fe7ef9
    Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2663159Reviewed-by: 's avatarJakob Gruber <jgruber@chromium.org>
    Reviewed-by: 's avatarLeszek Swirski <leszeks@chromium.org>
    Reviewed-by: 's avatarUlan Degenbaev <ulan@chromium.org>
    Commit-Queue: Dominik Inführ <dinfuehr@chromium.org>
    Cr-Commit-Position: refs/heads/master@{#72618}
    d5416b99
interpreter.h 4.06 KB