[PATCH] gdb: do not add const sections to the section map

[PATCH] gdb: do not add const sections to the section map

[Ilya Leoshkevich]

From: Ulrich Weigand <ulrich.weigand@de.ibm.com>

build_objfile_section_table () creates four synthetic sections, which
significantly slow down section map sorting.  This is especially
noticeable when debugging JITs that report a lot of objfiles.  Since
these sections are not useful for find_pc_section (), do not add them
to the section map.
---
 gdb/objfiles.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/gdb/objfiles.c b/gdb/objfiles.c
index 80f68fda1c1..259c19584a6 100644
--- a/gdb/objfiles.c
+++ b/gdb/objfiles.c
@@ -1005,6 +1005,9 @@ insert_section_p (const struct bfd *abfd,
   if ((bfd_section_flags (section) & SEC_THREAD_LOCAL) != 0)
     /* This is a TLS section.  */
     return 0;
+  if (bfd_is_const_section (section))
+    /* This is one of the global *ABS*, *UND*, *IND*, or *COM* sections.  */
+    return 0;
 
   return 1;
 }
-- 
2.35.1

Re: [PATCH] gdb: do not add const sections to the section map

[Andrew Burgess]

Ilya Leoshkevich via Binutils <binutils@sourceware.org> writes:

Thanks for this.  GDB patches need to be sent to
gdb-patches@sourceware.org for review.

> From: Ulrich Weigand <ulrich.weigand@de.ibm.com>
>
> build_objfile_section_table () creates four synthetic sections, which
> significantly slow down section map sorting.  This is especially
> noticeable when debugging JITs that report a lot of objfiles.  Since
> these sections are not useful for find_pc_section (), do not add them
> to the section map.
> ---
>  gdb/objfiles.c | 3 +++
>  1 file changed, 3 insertions(+)
>
> diff --git a/gdb/objfiles.c b/gdb/objfiles.c
> index 80f68fda1c1..259c19584a6 100644
> --- a/gdb/objfiles.c
> +++ b/gdb/objfiles.c
> @@ -1005,6 +1005,9 @@ insert_section_p (const struct bfd *abfd,
>    if ((bfd_section_flags (section) & SEC_THREAD_LOCAL) != 0)
>      /* This is a TLS section.  */
>      return 0;
> +  if (bfd_is_const_section (section))
> +    /* This is one of the global *ABS*, *UND*, *IND*, or *COM* sections.  */
> +    return 0;

You either need to a { ... } around the comment and the return, or move
the comment before the `if` here.

Thanks,
Andrew

>  
>    return 1;
>  }
> -- 
> 2.35.1

Re: [PATCH] gdb: do not add const sections to the section map

[Tom Tromey]

>>>>> Ilya Leoshkevich <iii@linux.ibm.com> writes:

> From: Ulrich Weigand <ulrich.weigand@de.ibm.com>
> build_objfile_section_table () creates four synthetic sections, which
> significantly slow down section map sorting.  This is especially
> noticeable when debugging JITs that report a lot of objfiles.  Since
> these sections are not useful for find_pc_section (), do not add them
> to the section map.

How does it slow down the sorting?

It seems a little strange at least to drop the *ABS* section, since IIUC
that's for symbols that specify some kind of absolute address.  Couldn't
these matter for lookup purposes?  It's probable I don't really know
what's going on here...

I wonder if we can drop all support for all these sections from gdb.

Tom

Re: [PATCH] gdb: do not add const sections to the section map

[Ilya Leoshkevich]

On Thu, 2022-05-19 at 09:10 -0600, Tom Tromey wrote:
> > > > > > Ilya Leoshkevich <iii@linux.ibm.com> writes:
> 
> > From: Ulrich Weigand <ulrich.weigand@de.ibm.com>
> > build_objfile_section_table () creates four synthetic sections,
> > which
> > significantly slow down section map sorting.  This is especially
> > noticeable when debugging JITs that report a lot of objfiles. 
> > Since
> > these sections are not useful for find_pc_section (), do not add
> > them
> > to the section map.
> 
> How does it slow down the sorting?

CoreCLR JIT reports each compiled function as a separate objfile with
a single .text section and a few debug sections [1].  Once you have a
few thousand of them - which is easily reachable - std::sort becomes a
bottleneck.  Ignoring synthetic sections from each objfile reduces the
amount of sections to sort significantly.

That's unfortunately not quite enough to debug bigger apps.  But it's
still a noticeable improvement, so I've decided to post it.

I'm currently trying to implement a more complex optimization in the
section map area ([2] - not working yet, and probably not reviewable
either, just FYI).

> It seems a little strange at least to drop the *ABS* section, since
> IIUC
> that's for symbols that specify some kind of absolute address. 
> Couldn't
> these matter for lookup purposes?  It's probable I don't really know
> what's going on here...

All four seem to have a size of 0, so they should never be returned
by find_pc_section ().

> I wonder if we can drop all support for all these sections from gdb.

They were added in 65cf35635977, which introduced
gdb_bfd_section_index ().  This function is used in quite a few places,
so I guess as long as bfd can report these sections, this should stay?

> Tom

[1]
https://github.com/dotnet/runtime/blob/v6.0.5/src/coreclr/vm/gdbjit.cpp#L2147
[2]
https://github.com/iii-i/binutils-gdb/releases/tag/section-map-20220519