Re: [PATCH 2/5] gdbsupport: Introduce interval_tree

[Ilya Leoshkevich <iii@linux.ibm.com>]

On Thu, 2022-06-02 at 15:12 +0100, Pedro Alves wrote:
> On 2022-06-02 14:35, Ilya Leoshkevich wrote:
> > 
> > Adding N JITed sections has the complexity O((N**2)*log(N)),
> > because
> > adding each section involves breakpoint handling, which needs to
> > resolve PCs and thus calls update_section_map().  When N is around
> > 10k,
> > this renders GDB unusable.
> 
> Does this adding of N JITed sections happen in batch?  Like, is this
> from
> jit_inferior_init, where we loop over JIT objects?  Or is it so that
> we
> get notified about JIT objects, one at a time?
> 
> In places where we add symbols in batch, we defer breakpoint
> re_setting exactly
> to avoid problems like this, via SYMFILE_DEFER_BP_RESET or something
> similar.
> Looks like jit.c doesn't try to do that.  Or is it not possible in
> the scenario
> in question?  Like, doesn't the JIT API let you register more than
> one object
> file at once?

In the scenario I'm running into, GDB is notified about them one by
one.  I don't think there is a batch API ([1] doesn't mention anything
like this), but even if there were, it wouldn't help in the general
case, because the decisions to JIT individual methods may be split in
time.

[1]
https://sourceware.org/gdb/onlinedocs/gdb/Registering-Code.html#Registering-Code