This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: gfortran test runs


On Wed, 2004-08-25 at 22:55, Joe Buck wrote:
> On Wed, Aug 25, 2004 at 11:29:09PM +0200, Toon Moene wrote:
> > Richard Earnshaw wrote:
> > 
> > > It really shouldn't be necessary for gcc developers to have to install
> > > components into /usr/lib (or any other system directory) in order to be
> > > able to develop and test the compiler.
> > 
> > While I agree with this sentiment (I got bitten by this situation on my 
> > Alpha machine that's not on a network and can't be easily apt-get update 
> > && apt-get dist-upgrade'd), but I sincerely ask the C crowd whether they 
> > think they can implement the whole of C99 without extensive support for 
> > compile time floating point manipulation.
> 

The issue isn't whether or not we should use GMP or even a specific
version of it.  It's about how we use it, and how we link to it, and how
we behave if the system doesn't provide it or puts it in a non-standard
place.

> I run tests for gcc on a large variety of machines.  For most of these
> machines, I don't have root.  If I can't do tests without modifying
> /usr/lib or /usr/local/lib, I can't do tests.
> 
> Requiring gmp, etc. is fine; it should suffice to install it in the same
> prefix as for gcc and have everything Just Work.  It's OK if the user has
> to put $prefix/lib into LD_LIBRARY_PATH or the equivalent, but no other
> special handling should be required.

No, I don't think it's acceptable for a user to have to set up
LD_LIBRARY_PATH every time they want to use gcc (especially since so
many things seem to clobber that variable).  Either we link to a
suitable libgmp that's already installed in a well-known place or we
should link it statically into the compiler.

What I don't understand, given the ample precedent for doing this with
other packages, is why we can't just add the gmp source to the compiler
as an imported package, in the same way we do intl and zlib (and
probably other things).  Then if the system doesn't provide a suitable
gmp we just build our own and link it in directly.

R>


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]