Regression after removal of field lookup cache [265148324]

Fixed

Bug

Status Update

No update yet.

Description

ze...@google.com

created issue #1

Jan 11, 2023 03:59PM

The resolution of b/151804585 did not introduce a new field backing, thus lookups within a class are linear search in the field arrays. This has noticeable performance impact witnessed in b/264430795

Using the dump information provided in b/264430795, the initial regression came with CLs:

353672ms 78a3b2fe030b465c46c1b2f858cada93659a0e40 Minor cleanup prior to definitionFor(DexField) removal
335307ms 80ef5bef88b866f9dea82efc73b024ee751b42e2 Remove all uses of definitionFor(DexMember)
337779ms e7186a7da0d3edc84b389169614c60be4e208b14 Remove all uses of definitionFor(DexReference)
267364ms 225ddc66dc42ce0fdef363d17b61c88807b661ed Add CLI tests for various flag additions and fixes.

With the bottom CL being fairly equal in runtime to the original report at version 2.0.99.

Total regression is about ~1.5min on a previously 4.5min build (increase of about 33% in compiletime).

We should introduce a (potentially) hash-map based field collection backing mirroring the backing for methods.

Comments

ze...@google.com <ze...@google.com> #2Jan 11, 2023 04:20PM

Will be this fix also part of 4.2 please?

ap...@google.com <ap...@google.com> #3Jan 13, 2023 12:02PM

for the "out" bundle issue, this seems indeed an issue. for the second one reported adding "debugUpdateAPKArtifact" to the path, this WAI as there can more more than one transformer and we need to automatically separate them so they do not use the same output folder. for #3, I will look at it but it might be difficult to change it at this time.

ap...@google.com <ap...@google.com> #4Jan 16, 2023 11:14AM

I understand that if there is more transformer is necessary to make every folder different, but there is the main problem, that the final one (last transform) is inside of the intermediate folder. The last transformation should end in the standard location, right? Maybe also with the default name not?

ap...@google.com <ap...@google.com> #5Jan 16, 2023 11:14AM

Hello, this is a P1 S1 issue for a while now, and it's blocking us from upgrading to AGP 4.2 and further. What's the current status? And do you have plans to backport it to 4.2.x?

ze...@google.com <ze...@google.com> Jan 18, 2023 08:37AM

Marked as fixed.

Issue 265148324