Re: What's the right way to do host libraries required for gcc

[Daniel Berlin <dberlin at dberlin dot org>]

On Wednesday, November 13, 2002, at 02:11  PM, Joseph S. Myers wrote:

> On Wed, 13 Nov 2002, Daniel Berlin wrote:
>
> > this feature just disappears from your gcc.  I've done this mainly
> > because while libbanshee is portable, I'm sure that there will be some
> > odd platform on which it doesn't build at first (since it's ANSI C, 
> > not
> > K&R), and i don't want to break their bootstraps until it gets fixed.
> I'd suggest removing the K&R requirement (hopefully the SC will 
> consider
> this again after 3.3) and making libbanshee required before merging 
> this
> feature to mainline.
The whole idea of using a support library besides libiberty is likely 
to be controversial (though in some cases a good idea. For instance, 
the PCH file size can be greatly reduced by requiring and using zlib), 
so I wanted to avoid making it required until i got a feel for how 
people would react.
>  The existence of optimizations in some copies of a
> given GCC version but not others is likely to yield confusion with bug
> reports some people can reproduce but not others on the same platform
> (which then get wrongly closed).
Which is why it's on a branch right now.
It's also not an optimization in and of itself, it just a different 
points-to analysis with *much* better precision, which in turn, 
generally enables more optimizations.  Either of the points-to analysis 
algorithms implemented (the other one isn't using the constraint engine 
yet) should always be conservatively correct.  Since the results are  
also not used directly by any optimization pass, but instead used by 
the SSA builder in constructing the alias lists for variables (IE it's 
just queried if we can't determine two variables don't alias through 
TBAA or other methods), a broken optimization would have to be broken 
without the analysis turned on.
All of this said, it is my eventual goal (before merging to the 
mainline) that it always be on and the default, I'm just trying to be 
realistic and sensitive to concerns in getting there.

>  It's a good principle that a given
> version of GCC should produce identical output for given input code and
> options, whatever the host system (at least provided sufficiently 
> recent
> GNU binutils is installed).
Sure.
> (Your question is still relevant for people trying the branch for now,
> although I don't know if any such people are using K&R systems.)
Which is why i asked.  I'm afraid some HPUX'er will try the branch and 
scream bloody murder.
:)

> --
> Joseph S. Myers
> jsm28@cam.ac.uk