Re: GCC 3.5 Status (2004-08-29)

[Dorit Naishlos <DORIT@il.ibm.com>]

> If you disagree with my decisions, feel free to send me a message
> explaining why you think that the decision was inappropriate.  I will
> consider your position, but you will have to be convincing in order to
> change my mind.  To be convincing, make sure you quantify the win: it
> you want me to believe that your optimization is worthwhile tell me
> how much it improves a well-known benchmark.

> The following changes will be postponed until GCC 3.6.  These changes
> either provide too little benefit, are too risky, or will take too
> much time to complete.
>
> ...
> * Vectorizer misaligned-loads support [Naishlos]

Let me try to convince you to change this decision:

1. Alignment is the number one limiting factor for vectorization on
mainline right now. Without this we can't vectorize anything that uses
pointers, for example. Together with the patches for unknown loop bound and
peeling for aligning a store (proposals that were accepted for 3.5), we
will be able to vectorize loops of the following (very basic and very
common) form:
   "while (--n) *q++ = *a++ + *b++".

2. Risk-free. It is entirely encapsulated within -ftree-vectorize which is
disabled by default (so no risks if not explicitly enabled).

3. Chances of meeting deadline - yes, Sept. 19 is aggressive. But note that
we already had this functionality implemented completely on the
apple-ppc-branch a couple of months ago, but using target builtins instead
of new tree-codes. Introducing the new tree-codes indeed takes some time,
but it may be doable by September 19th, considering the fast progress made
together with Richard. Given the huge benefit and zero risk, I think it's
worth while to give it a chance. What do you say?

thanks,
dorit