Re: [PATCH] On-demand loading of shlib's debuginfo

[Jan Kratochvil <jan dot kratochvil at redhat dot com>]

On Thu, 21 Jul 2011 20:36:04 +0200, Pedro Alves wrote:
> Oh, well:-)  I thought the separation was good, as:
> 
>  - target sections are used for "set trust-readonly-sections on"
>    and similar fallbacks to reading memory from the exec, which
>    requires separate dso lazy load points (disassembly? printing
>    some address?).
> 
>  - its not clear to me the pc -> dso mapping from the link map is
>    faster than from the bfd in all scenarios.  E.g., on remote
>    targets, it may be faster to get at the bfd info on the host, than
>    to remote read memory from the target.

I agree the remote targets may give opposite performance benefits.

I think it would be good to have both parts at hand to play with pros and
cons.  So far this patch needs explicit enable anyway as it breaks most of the
non-backtrace commands.


>  - I was curious on how much lazing debug info alone was helping,
>    vs lazing the bfd reads.

I had some very brief benchmark:
	0m24.824s -> 0m12.885s (cold disk cache with warm cache for gdb)
	F15.x86_64, therefore with .gdb_index.

This is by a variant of this patch - that is still with mapping bfds for
target sections, on HDD (not SSD).  And sure everything with linux-nat.


Thanks,
Jan