Re: [PATCH] Remove bogosities from dwarf2cfi.c:cfi_pop_frame()

[Andrew Cagney <ac131313@ges.redhat.com>]

> Mark
>> > 
>> > Index: ChangeLog
>> > from  Mark Kettenis  <kettenis@gnu.org>
>> > 
>> > 	* dwarf2cfi.c (cfi_pop_frame): Use alloca() for regbuf.
>> > 	Don't call get_current_frame().
>> > 
> 
>> 
>> Heads up,
>> 
>> I think the write_register_bytes() call should be replaced by 
>> regcache_register_write(current_regcache, ).  This is new code so it 
>> doesn't need to go through any of that nasty write_register_bytes() stuff.
>> 
>> Andrew
>> 
>> 
> 
> 
> Why is this code trying to poke directly into the regcache at all?  AFAICT 
> it should be operating on the pseudo registers not the cache.

[cooked]

Hmm, yes, good point. It should at least use regcache_cpy() so that it 
is ``bug compatible'' with generic_pop_dummy_frame()(1).

At present GDB saves/restores registers in a confused sort of way. 
Firstly, the save/restore operation (see regcache_cpy):

- for old targets it iterates over 0..NUM_REGS (the full raw register 
space) using write_register_bytes() to restore the registers.

- for new targets it iterates over 0..NUM_REGS (the full raw register 
space) using regcache_write() to directly transfer the raw registers.

It then uses this single mechanism(2) to handle two orthogonal 
situtations, and that is the problem:

- saving / restoring a raw register cache before / after an inferior 
function call.  Here, a select list of raw registers should be transfered.

- unwinding a stack frame, storing those unwound register values, and 
hence having the effect of poping one or more frames.  Here, a select 
list of cooked registers should be stored.


Andrew

(1) I also suspect that some cleanups would let the function in question 
be eliminated from dwarf2cfi but that is another story.

(2) I believe the regcache changes were ``bug compatible''.