Alignment of gfortran allocated arrays

[John Travers]

On 17/08/07, Tim Prince <tprince@computer.org> wrote:
> John Travers wrote:
> >
> > Yes, my two primary systems (including the one I posted from) are 64
> > bit. I thought it was clear what I was talking about as we were
> > discussing getting 16-byte alignment.
> >
> 16-byte alignments are needed for SSE support on both 32- and 64-bit
> OS.  Various compilers for 32-bit linux handle it in incompatible ways.
> You could get into trouble calling a vectorized gcc or gfortran
> function  from ifort, for example.  malloc() gives 4-byte alignment on
> 32-bit Windows (even scalar double precision performance may suffer),
> and 8-byte on linux.
> Windows and linux 64-bit both adopted 16-byte alignment as a default.
> If malloc() doesn't give a 16-byte aligned pointer, that's a bug.  The
> URLs you referenced indicate that gfortran already has resolved the
> ALLOCATABLE problem as far as 64-bit linux is concerned.  Solaris is the
> only 64-bit OS I know of which has trouble here.
>

I think we keep miss-interpreting each other :). My 64-bit systems do
not have any problem at all, so no bugs in malloc. I was simply
explaining that I was aware of 64 bit systems being common, but my
desire to sort out the 16 byte alignment (with respect to the default
8 byte mentioned in the posts I linked to), implied I was talking
about 32 bit systems. I should have been clearer on that, I'm sorry.

Anyway, somebody contacted me off list to suggest I use the shiny new
iso_c_bindings to call fftw_malloc. I knew of the bindings, but didn't
realise I could coerce a fortran array pointer to take data allocated
in c. I'm now enlightened and it works perfectly!

So I regard this issue (of mine) now solved, and think the c bindings
technique is better than fiddling with gfortran internals, so I won't
try and make any patches to wrap gfortran's malloc.

Sorry for creating so much noise on this list and thanks for all the
helpful replies!

Best regards,
John