Baby's First AldyVec/AltiVec patch

[Matt Austern]

On Dec 29, 2003, at 12:20 PM, Segher Boessenkool wrote:

> On 23-dec-03, at 19:56, Matt Austern wrote:
>> And remember the downside if we don't introduce the Motorola syntax:
>> Altivec code won't be portable between the various compilers that all
>> claim to support Altivec.  That's a very serious problem, and it's 
>> exactly
>> what standards are supposed to prevent.
>
> Hey, the Motorola AltiVec PIM does not support 64-bit PowerPC at all;
> in fact, it requires incompatible changes to sanely support it.
>
> There is a real need for a new, better, C binding for VMX; I for sure
> don't care for compatibility with the Motorola AltiVec PIM at all,
> as it is very very awkward to use (in fact, the _only_ thing that
> it buys over writing real assembly is register allocation, and for
> that you have to write casts almost everywhere -- I prefer to stay
> with handwriting assembly (or having the compiler auto-vectorize
> stuff of course)).
>
> All in all, I think we would do better to learn from the mistakes
> in the Motorola AltiVec PIM and start over.

"Start over" really isn't practical given that the existing syntax
in the Motorola Altivec PIM is widely used.  Moving to a new syntax
might be a good idea, but it'll need a transition strategy and it'll
have to be something that gets buy-in from *all* compilers that
target this architecture.

			--Matt