> Simon Marlow wrote:
> > [...] Just out of interest, do you know whether doubles
> *must* be aligned
> > on an 8-byte boundary on HP, or is it just a performance thing?
>
> Well, it depends on your definition of "performance": Failing to align
> doubles on 8-byte-boundaries on HPs results in highly efficient
> core-dump-producers: :-)
This is a particularly annoying thing. Every machine has a different policy
on misaligned doubles:
- x86: reduced performance
- sparcs: older (sun4c?) core dump
newer (sun4m, sun4u) reduced performance
- HP: core dump
- alpha: no core dump, but warning message on console :-)
GHC has to deal with misaligned doubles, because otherwise we'd have to
align every stack/heap object on an 8-byte boundary which would waste a load
of memory. We just need to know whether we can do it directly, or the
slightly slower way of casting back/forth using a union (actually, the C
compiler *ought* to generate the same code, but it doesn't always).
Cheers,
Simon
--
Simon Marlow
Microsoft Research Ltd., Cambridge, U.K.
