1868326 - Use aarch64 instrumented build for PGO profile generation for aarch64 macos

Status: RESOLVED FIXED

(Core :: Performance, enhancement)

Description

[Jamie Nicol [:jnicol]]

Similar to bug 1865886 on Android, we can get a 5.11% improvement on speedometer3 on aarch64 macos by using an aarch64 instrumented build and running the profile generation on an aarch64 mac.

https://treeherder.mozilla.org/perfherder/compare?originalProject=try&originalRevision=402549f3803e7da6d27b401f3b3d1f20b4123ad9&newProject=try&newRevision=ee9c1ce481846caef2ebfc34af1bf59875666c68&framework=13&page=1

Setting up the original PGO runs on x86_64 macos required adding a dedicated pool of hardware for only running level-3 PGO profile generation jobs (bug 1530732). So presumably we'll need to do the same with aarch64 hardware.

Comment 1

[Jamie Nicol [:jnicol]]

See https://mozilla-hub.atlassian.net/browse/RELOPS-789 for adding required workers to CI

Comment 2

[Jeff Muizelaar [:jrmuizel]]

Attachment: Bug 1868326. macOS aarch64 instrumented build

Comment 3

[Jeff Muizelaar [:jrmuizel]]

Attachment: Bug 1868326. macOS aarch64 profile generation.

Comment 4

[Jeff Muizelaar [:jrmuizel]]

Attachment: Bug 1868326. Use aarch64 pgo data for mac aarch64 builds.

Comment 5

[Markus Stange [:mstange]]

The performance difference does not appear to just be in aarch64-specific code. Instead, the difference seem to be spread over lots of different functions.

When using the x86_64 PGO training data, the arm64 build logs contain 16231 warnings of the type function control flow change detected (hash mismatch):

[task 2024-01-15T21:03:07.139Z] 21:03:07     INFO -  warning: Unified_cpp_parser_html1.cpp: function control flow change detected (hash mismatch) _ZN16nsHtml5Tokenizer9stateLoopI20nsHtml5FastestPolicyEEiiDsiPDsbii Hash = 1006606701270676095 up to 423331622 count discarded [-Wbackend-plugin]

With the aarch64 PGO training data, the arm64 build log only contains 1651 of those warnings (10x less).

Comment 6

[Markus Stange [:mstange]]

Attachment: Functions with mismatches when using x86_64 PGO data for aarch64 build

Comment 7

[Markus Stange [:mstange]]

Attachment: Functions with mismatches when using aarch64 PGO data for aarch64 build

Comment 8

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/2dabe7565934
macOS aarch64 instrumented build r=glandium
https://hg.mozilla.org/integration/autoland/rev/541a0ae99851
macOS aarch64 profile generation. r=glandium
https://hg.mozilla.org/integration/autoland/rev/38fbe387a1aa
Use aarch64 pgo data for mac aarch64 builds. r=glandium

Comment 9

[Natalia Csoregi [:nataliaCs]]

https://hg.mozilla.org/mozilla-central/rev/2dabe7565934
https://hg.mozilla.org/mozilla-central/rev/541a0ae99851
https://hg.mozilla.org/mozilla-central/rev/38fbe387a1aa

Comment 10

[Acasandrei Beatrice (needinfo me)]

(In reply to Pulsebot from comment #8)

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/2dabe7565934
macOS aarch64 instrumented build r=glandium
https://hg.mozilla.org/integration/autoland/rev/541a0ae99851
macOS aarch64 profile generation. r=glandium
https://hg.mozilla.org/integration/autoland/rev/38fbe387a1aa
Use aarch64 pgo data for mac aarch64 builds. r=glandium

== Change summary for alert #41083 (as of Fri, 19 Jan 2024 12:12:22 GMT) ==

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)	Performance Profiles
5%	speedometer3	macosx1300-64-shippable-qr	fission webrender	29.70 -> 31.19	Before/After

For up to date results, see: https://treeherder.mozilla.org/perfherder/alerts?id=41083