commit ea4dd3eea4d21c29a5eb93aaf82e6bd71f880ecb
author Florian Leimgruber <fleimgruber@google.com> Thu Jan 25 09:12:59 2024
committer Chromium LUCI CQ <chromium-scoped@luci-project-accounts.iam.gserviceaccount.com> Thu Jan 25 09:12:59 2024
tree dc0acbac1a39e1b94eda2d0a05ecf4f4ddc8e4fa
parent f6258ec2ad18342c9eb1822fabbdd3c348f3852f

Don't recompute (shadow) predictions unnecessarily

For every active and non-active PatternSource, predictions are computed
in DetermineHeuristicTypes() by calling ParseFieldTypesWithPatterns(),
rerunning all the parsing logic.

If the patterns are very similar, the hope was that
AutofillEnableCacheForRegexMatching will prevent a lot of the duplicate
work. This seems true to some extend; based on offline measurements,
the overhead of computing shadow predictions for an additional
PatternSource with the same patterns is about a factor of 1.76, not 2.
(Presumably, since it's an LRU cache, the cache is not particularly
 effective)

This CL adds an additional optimisation: If all patterns of the shadow
PatternSource are the same as the (already computed) active
PatternSource for the given page language, simply reuse those
predictions. With that, there's  no measurable performance overhead
anymore.

Likely, we won't be rolling out new regexes all the time. So this saves
users CPU cycles on every page load.

Technically, this is implemented by precomputing the equality during
compile time, while transpiling the JSON files.

Bug: 1479353
Change-Id: I294cf9c1b119463c24df4752cfb966e010ecab1f
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/5200177
Reviewed-by: Christoph Schwering <schwering@google.com>
Commit-Queue: Florian Leimgruber <fleimgruber@google.com>
Cr-Commit-Position: refs/heads/main@{#1251958}