Re: MPS: User GC customizations

[Pip Cet <pipcet@protonmail.com>]

On Thursday, July 4th, 2024 at 18:26, Ihor Radchenko <yantar92@posteo.net> wrote:
> Gerd Möllmann gerd.moellmann@gmail.com writes:
> 
> > Pip mentioned scan functions, so let's say we set a flag in_scan while
> > being in dflt_scan. MPS now calls dflt_scan, in the MPS thread, to do an
> > increment of its work. While in dflt_scan we get SIGPROF and land in the
> > signal handler in main thread, and the profiler sees in_scan == true.
> > 
> > Q: do we count that as part of GC work that the profiler should report,
> > although it happened in the MPS thread? I understood Ihor as saying that
> > he doesn't want that.
> 
> 
> I think that we should ideally have multiple flags like this:
> 1. when arena is locked
> 2. when our dflt_scan code is running
> 3. when we query MPS synchronously (e.g. memory allocation)
> 4. when Emacs thread is paused by MPS

5. when our dflt_scan code is running on the main thread

I think that's what Gerd meant to suggest, anyway?

> Then, we can report statistics for each flag.

Or each of the 32 combinations (okay, okay, 24).

> I think that it can give us some idea about how different aspects of MPS
> affect Emacs responsiveness.

I think it would help me understand what in the word MPS might be doing. AFAICT it Just Works.

Also, may I suggest we really need

6. whether Emacs is "idle"

Pip