Re: [G95-develop] Re: Vectorising pass (was: Re: SIMPLE)

[Richard Henderson <rth at twiddle dot net>]

On Wed, Jan 09, 2002 at 09:53:13PM +0100, Steven Bosscher wrote:
> So far, there has been little discussion (unlike that other topics
> discussed on the gcc mailing list lately ;-) about vector trees in a new
> GCC IR. Maybe not everybody is as interested in them, and we should just 
> forget about it.

I think very few people actually know anything about them.  I'd say
that we should put vector IR on the back-burner until the scalar IR
works, then re-address the issue.

If done right, a vector IR will be of use to more than Fortran.  

Something that SGI's compiler does is to define nodes corresponding
to e.g. Fortran's do-loop (which has the property that the iterator
cannot be modified within the loop), and when compiling C would
attempt to prove that a particular for-loop has semantics equivalent
to a do-loop, and transform the tree accordingly.  Later optimizers
can apply particular transformations to the loop without having to
examine the loop in detail, since the semantics of do-loop assure
that the transformation is valid.

We ought to be able to do similar things with vector IR nodes.


r~