Arraysizes in g77 (gcc-2.95.2)

Arraysizes in g77 (gcc-2.95.2)

[Andre Albsmeier]

Hello all,

I am using GCC 2.95.2. I have a program for antenna simulation which
is mostly written in fortran. I am using an array of complex variables
(the size of one element is 16 bytes). I can allocate only 33554431 of
these elements. When trying to use more than that, the compiler stops
with an error:

cd .; /usr/local/bin/g77 -O2 -finit-local-zero -fomit-frame-pointer -s -c qustn.f
celem.inc: In subroutine `qust':
celem.inc:3:
         complex*16  celem(ndim)
                     ^
Array `celem' at (^) is too large to handle

33554431 * 16 is exactly 512 * 1024 * 1024 - 16 bytes. So I
assume it is not possible to declare an array that is bigger
than 512MB - sizeof(one_element).

Now my question: Do you know if I can simply bump some number
in some file of GCC to increase this limit a bit?

Thanks a lot in advance!

        -Andre

Re: Arraysizes in g77 (gcc-2.95.2)

[Martin v. Loewis]

> Now my question: Do you know if I can simply bump some number
> in some file of GCC to increase this limit a bit?

I don't know much about fortran, but you may want to check the release
branch. One problem in gcc is that the array size is represented in
the number of bits, so that value probably overflows if the array gets
larger than 512 MB. I found the same problem in g++, and the g++
mainline has fixed it. I'm not sure whether the fix was specific to
C++, or whether that was a back-end thing, though.

In any case, there is probably not much you can do about it, if the
cause is really that integer flows over.

Regards,
Martin

Re: Arraysizes in g77 (gcc-2.95.2)

[toon]

Martin v. Loewis wrote:

> Andre Albsmeier wrote:

> > Now my question: Do you know if I can simply bump some number
> > in some file of GCC to increase this limit a bit?

> I don't know much about fortran, but you may want to check the release
> branch. One problem in gcc is that the array size is represented in
> the number of bits, so that value probably overflows if the array gets
> larger than 512 MB. I found the same problem in g++, and the g++
> mainline has fixed it. I'm not sure whether the fix was specific to
> C++, or whether that was a back-end thing, though.

AFAIK the bitsize thingy is a backend limitation.  I know that shortly afte
the release of g77 (February '95) I run into all sorts of sub-optimisations
due to the fact that the *8 (bits/byte) / 32 (bits/real) weren't folded 
completely (next to the problem of incomplete induction variable elimination
in loops.

So I think that this is a "middle-end" problem - I would be interested if
this were solved by the current CVS trunk of GCC.

Cheers,
Toon.