Commit Graph

3 Commits

Author SHA1 Message Date
Jakob Linke
7086e988e7 [maglev] Fix osr-from-ml-to-tf flake again
We've previously reset profiler ticks on ML finalization to avoid
unexpected early tierups to TF. This CL additionally avoid early TF
tierup through the `small function` reason by disabling small-function
optimization.

Bug: v8:7700
Change-Id: I57ba294af0d1d189f76c2cb1ffc31af0837b1e42
Fixed: v8:13242
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3865550
Auto-Submit: Jakob Linke <jgruber@chromium.org>
Reviewed-by: Victor Gomes <victorgomes@chromium.org>
Commit-Queue: Victor Gomes <victorgomes@chromium.org>
Commit-Queue: Jakob Linke <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/main@{#82886}
2022-09-01 08:03:22 +00:00
Jakob Linke
6565685935 [maglev] Reset profiler ticks on ML finalization
.. since ticks collected in lower tiers use a different budget than
ticks collected in ML code. Keeping ticks from a lower tier would lead
to earlier-than-expected TF compilation requests.

Also, remove the runtime call between loops in osr-from-ml-to-tf.js
to not give TF too much time to tier up. It should no longer happen
now that ticks are cleared, but there's also no reason for the runtime
function to between the loops.

Bug: v8:7700
Fixed: v8:13242
Change-Id: I901c636079de05cb62d2aae1c92f0285a0f8c1ad
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3867310
Auto-Submit: Jakob Linke <jgruber@chromium.org>
Reviewed-by: Leszek Swirski <leszeks@chromium.org>
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Cr-Commit-Position: refs/heads/main@{#82859}
2022-08-31 10:46:41 +00:00
Jakob Linke
ed90ea5cf7 [maglev] Implement Maglev-to-Turbofan OSR
This implementation sticks closely to what Ignition-to-Turbofan (and now
Sparkplug-to-TF) does. OSR is detected in the TieringManager by having
optimized code available, without having entered it. The osr_urgency is
increased to enable OSR for increasing loop depths. When a candidate
JumpLoop backedge is reached, we call into runtime to trigger OSR
compilation.

JumpLoop also detects the availability of cached OSR'd code. When a
matching OSR code object is available, Maglev 1) deoptimizes s.t. the
unoptimized frame layout is reconstructed, and 2) delegates the actual
OSR tierup to the unoptimized tier. For purposes of 1), we add a new
DeoptimizeReason that causes a one-time eager deopt without invalidating
any code.

Drive-by: Annotate OSR for more --trace-opt output.

Todo: Refactor non-Sparkplug-specific bits of the BaselineAssembler
into a generic spot that both SP and ML can use.

Bug: v8:7700
Change-Id: I6ebab2df8b87f9f70ffb78162a3c1226ec545468
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3859850
Reviewed-by: Leszek Swirski <leszeks@chromium.org>
Commit-Queue: Jakob Linke <jgruber@chromium.org>
Cr-Commit-Position: refs/heads/main@{#82816}
2022-08-30 11:51:28 +00:00