Internals of -readnow implementation

[Jan Vrany]

On Tue, 2025-11-11 at 14:18 -0700, Tom Tromey wrote:
> Jan> looking at how -readnow is implemented, I'm puzzled
> Jan> why implementation of readnow_functions inherits/user
> Jan> dwarf2_base_index_functions?
> 
> It's mostly a convenience, because that class implements some methods.

I see, thanks. 
> 
> Jan> My understanding is that if -readnow is specified, then
> Jan> all compunit symbtabs in that objfile are fully read.
> Jan> It seems to me that in theory one can implement it "directly" 
> Jan> by traversing compunit symtabs, its blockvector and symbols 
> Jan> (rather than using dwarf2_per_objfile and alike). 
> 
> Jan> Is that because it was just easier to implement it this way
> Jan> or is there a reason why it has to go through DWARF reader
> Jan> structures?
> 
> There's no deep reason.
> 
> For this bug:
> 
>     https://sourceware.org/bugzilla/show_bug.cgi?id=33554
> 
> ... I am planning to reimplement it as you say: just search things
> directly without regard to the DWARF data structures.

Yes! I encountered the same issue when working on Python JIT API
and came to the same realization.

In fact, I already started working on it and got relatively far
in replacing current implementation of readnow_functions but still 
have loose ends especially in ::search method. It seems to me that the 
implementations not always match its comment.

Jan

> 
> -readnow is weird because it's mostly there as a debugging aid, and
> sometimes a workaround for other bugs.  While I have recommended it to
> users, I don't really like doing so.  Occasionally I think about
> removing it.
> 
> Tom