Re: Patch to decl2.c

[Geoffrey Keating <geoffk at apple dot com>]

On Friday, September 5, 2003, at 09:34  AM, Mark Mitchell wrote:

> Geoff, Richard --
>
> Shouldn't this patch be reverted?
>
> 2003-09-02  Geoffrey Keating  <geoffk@apple.com>
>
> 	PR 12161
> 	* decl2.c (mark_used): Use ggc_push_context/ggc_pop_context.
> 	* tree.c (cp_cannot_inline_tree_fn): Likewise.
>
> I objected to the idea of this patch, and Richard's patch to optimize.c
> was supposed to handle this problem, if I understand correctly.

Richard's patch didn't work; see 
<http://gcc.gnu.org/ml/gcc-regression/2003-09/msg00032.html> for the 
message the regression tester sent saying so.  I think Richard's patch 
is now redundant, but haven't gotten around to removing it and testing 
the result.

Reverting my patch will cause to GCC to again fail to bootstrap on 
powerpc-darwin, unless something else is done.

> This patch seems wrong both in concept and in implementation.  If this
> needs to happen at all, it should happen inside instantiate_decl at the
> point where it actually starts instantiating the function, not wrapping
> the calls.  Especially in the mark_used case, this change is 
> potentially
> expensive.
>
> What's going on here?

When I was writing this patch, I asked myself the question "Of the 
functions in the backtrace, which would I not expect to be calling 
ggc_collect?"  It seemed reasonable to me that instantiate_decl should 
call ggc_collect.  It seemed unreasonable that mark_used and 
cp_cannot_inline_tree_fn should call ggc_collect.  This, however, is 
just my instinctive judgement and I'm certainly willing to go along 
with any reasonable design for this; I was mostly concerned with fixing 
the Darwin bootstrap.

--
Geoff Keating <geoffk@apple.com>