1559469 - Remove MOZ_JPROF

Status: RESOLVED WONTFIX

(Core :: DOM: Core & HTML, task, P3)

Description

[Andrew McCreight [:mccr8]]

In his review comment for bug 1559263, bz said: "We should really kill the non-working JProfFunctions thing and find a less sprinkly way of doing JS_DefineProfilingFunctions for all our globals." This is for the former part.

This code is not compiled by default and quite old, so I assume there's no chance it is actually working.

Comment 1

[Boris Zbarsky [:bzbarsky]]

Peter, am I right that this stuff is most likely dead now? And anyway, the Gecko profiler supersedes it...

Comment 2

[Boris Zbarsky [:bzbarsky]]

Randell, is this something we actually want to support, given bug 1399146? If so, should we just move this into the JS profiling functions stuff?

Comment 3

[Randell Jesup [:jesup] (needinfo me)]

jprof does work currently; I'll check on that bug. I'll chat with some people if there's any reason to keep it in the tree anymore. It does have a few functions not provided by the Gecko Profiler, and works closer to how 'perf' works (a sampling-during-running profiler). It has been largely superceded by Gecko Profiler (and even more so the ability to run 'perf' and show the results in Gecko Profiler). I've used it in the last year, but even I use it rarely. (I do build and run with it enabled frequently, however.)

Comment 4

[Peter Van der Beken [:peterv]]

Randell is a better candidate to decide about the jprof integration.

Comment 5

[Randell Jesup [:jesup] (needinfo me)]

So in bug 1399146 I've updated the patch(es) to switch from being able to output the old cleopatra format (which did work), to linux 'perf' format which the current Gecko Profiler can ingest. It works generally similarly to perf in practice (samples running code), but tends to produce more coherent stacks with less "roots" than perf. Nick Alexander was interested in it as a possible alternative to 'simpleperf' on android, which isn't all that simple. Also, unlike perf, we could in theory run the Gecko Profiler in markers-only mode (no sampling), and get much of the functionality of regular Gecko Profiles, except the JS frames in stacks (and if we switch to the LUL stackwalker, perhaps we can get those.

Still, removing it is certainly still on the table (and right now there's a occasional crash in glibc's backtrace() for some odd reason - though LUL would avoid that too). If we do keep it, the old HTML output may get dropped, as most everything in it is duplicated or better (more or less) in Gecko Profiler.

Comment 6

[Sylvestre Ledru [:Sylvestre]]

we are going to remove it - See bug 1920521
closing