Re: [PATCH] gdb/dwarf: fix order of operations when reading .debug_names

[Simon Marchi <simon.marchi@polymtl.ca>]

On 2026-05-20 11:15, Tom Tromey wrote:
>>>>>> "Simon" == Simon Marchi <simon.marchi@efficios.com> writes:
> 
> Simon>  - calls to the build_and_check_* functions, which do some lookups in
> Simon>    the all_units vector, requiring it to be sorted (oops)
> 
> I wonder if either there's some way to assert that the vector must be
> sorted on entry, or if there's a way to have a self-sorting data type
> here, rendering this kind of bug impossible.

assert: in dwarf2_find_unit, we could check that the vector is sorted,
but that would replicate the check done by -D_GLIBCXX_DEBUG.  I'm not
sure we want to pay this price for every dwarf2_find_unit call in
production binaries.

Alternatively, we could maintain a flag that says "is all_units
currently sorted", clear that flag when doing any modification to the
vector, and set it in the sort_all_units method.  All modifications to
the vector should be done through methods to ensure the flag gets
cleared.  I think that would be cheap enough to implement, I'll give it
a try.

self-sorting data type: we could insert the new units at the right place
from the start, but that would be less efficient, because that would be
akin to an insertion sort.  It's more efficient to do one sort at the
end.

We could perhaps use std::set, but there is also the case in
fill_in_sig_entry_from_dwo_file, when the section of a unit becomes
known at a later time.  At this point, the key for the unit would
change, so we would still need to remove it and reinsert it in the set.
I don't see a way to enforce that automatically.

> Anyway I think this is ok.
> Approved-By: Tom Tromey <tom@tromey.com>

Thanks, I'll push it and work on a patch to add the assert.

Simon