- 04 Feb, 2021 1 commit
-
-
Jakob Gruber authored
TranslationArrays (TA) are large and rarely used, thus could benefit from compression. This CL adds a --turbo-compress-translation-arrays flag (off by default) to experiment with that. Each optimized Code object has an associated translation array (Code->DeoptimizationData->TranslationArray). These translation arrays have roughly the same size as the Code object itself. They are used only rarely: when deoptimizing, and when traversing the stack and looking into optimized frames. Neither of these code paths are especially performance critical. TA's contain only immutable, untagged data. They are thus good candidates for compression. The trade-off is between TA memory consumption and time spent in decompression/compression. This CL keeps everything on the main thread, but it would also be possible to move compression (the more expensive operation by a factor of 5 to 10) to a worker thread. Numbers from a local Octane2 run: Sum of Code instructions sizes: 4.6MB Sum of uncompressed TA sizes: 4.1MB Sum of compressed TA sizes: 0.6MB Compression times depend on the selected compression quality, but roughly: Compression: 50ms (40us avg per compilation) Decompression: 7us avg per compilation Drive-by: Translation arrays currently use run-length encoding; I disabled this for when --turbo-compress-translation-arrays is enabled (no need to compress twice). Bug: v8:11354 Change-Id: I7828d7d91eb074816b383b02f883c5d7b7e318b7 Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2652497 Commit-Queue: Jakob Gruber <jgruber@chromium.org> Reviewed-by: Michael Stanton <mvstanton@chromium.org> Cr-Commit-Position: refs/heads/master@{#72533}
-