[x64] Teach ArgumentsAdaptorTrampoline about is_safe_to_skip_adapt_arguments.
This adds support for dynamically checking the newly introduced is_safe_to_skip_adapt_arguments flag on the SharedFunctionInfo inside the ArgumentsAdaptorTrampoline builtin, so that if there's an arguments mismatch, but said bit is set, we can just massage the current stack frame according to the SharedFunctionInfo and don't need to create an arguments adaptor frame. This improves the general (baseline) performance of calls with arguments mismatches by up to 35%, as can be seen with the test case from the tracking bug, which goes from around callUnderApplication: 608 ms. callOverApplication1: 633 ms. callOverApplication2: 635 ms. to callUnderApplication: 390 ms. callOverApplication1: 439 ms. callOverApplication2: 443 ms. when the TurboFan fast-path (that was previously introduced) is forcibly disabled. That's a pretty solid baseline improvement and there's probably still some room for further improvement (i.e. avoid loading the SFI multiple times, etc.). Drive-by-fix: Refactor and cleanup the ArgumentsAdaptorTrampoline builtin a bit to make it more efficient and easier to read (it's still pretty messy, but I don't wanna change too many things at once). This is only the initial x64 port, since for ia32 we're really short on registers and it might not even pay off (initial prototype shows a lot less performance impact). The arm and arm64 port is in progress and will land once ready. Bug: v8:8834, v8:8895 Change-Id: I2e1c8e4767a808a43b932487fdedc0e44cdd35c5 Reviewed-on: https://chromium-review.googlesource.com/c/1488751 Commit-Queue: Benedikt Meurer <bmeurer@chromium.org> Reviewed-by: Jakob Gruber <jgruber@chromium.org> Cr-Commit-Position: refs/heads/master@{#59881}
Showing
Please
register
or
sign in
to comment