64-bit integers on IA32

64-bit integers on IA32

[David Rasmussen]

I was told half a year ago that the reason that gcc performing as well with
64-bit integers on IA32 as some other compilers (Intel, MS)  was the the
register-pair used for holding the 64-bit integer, was thought of as a whole
by the compiler, which messed with gcc's register scheduling, or something
like that. Has that and/or related problems been solved in the meantime, or
will they be solved in any near future?

Thanks in advance,
David

Re: 64-bit integers on IA32

[Richard Henderson]

On Mon, Sep 02, 2002 at 07:46:07PM +0200, David Rasmussen wrote:
> Has that and/or related problems been solved in the meantime, or
> will they be solved in any near future?

No and no.


r~

Re: 64-bit integers on IA32

[David Rasmussen]

Richard Henderson wrote:
> On Mon, Sep 02, 2002 at 07:46:07PM +0200, David Rasmussen wrote:
>> Has that and/or related problems been solved in the meantime, or
>> will they be solved in any near future?
>
> No and no.
>

Would you care to elaborate on that?
For example: Why is it hard to fix, can a fix ever be expected etc. etc. ?
Of course, with the x86 world soon moving to 64-bit, it won't matter much.
But this is still an area where gcc is consistently 30% or more behind other
good compilers. That is a shame.

/David

Re: 64-bit integers on IA32

[Jan Hubicka]

> Richard Henderson wrote:
> > On Mon, Sep 02, 2002 at 07:46:07PM +0200, David Rasmussen wrote:
> >> Has that and/or related problems been solved in the meantime, or
> >> will they be solved in any near future?
> >
> > No and no.
> >
> 
> Would you care to elaborate on that?
> For example: Why is it hard to fix, can a fix ever be expected etc. etc. ?

It is dificult to fix as quite a lot of GCC code is built around
assumption that pseudo register (one before register allocation) will
end up in single hardware register.  It is easy to fake 64bit registers
by using consetuctive pairs, but it is dificult to put there extra
information needed for arbitary pairs.

Hopefully once new register allocator matures and gets usefull we will
be ready to address this issue, but now it is just too much mess around.

Honza
> Of course, with the x86 world soon moving to 64-bit, it won't matter much.
> But this is still an area where gcc is consistently 30% or more behind other
> good compilers. That is a shame.
> 
> /David