Re: rs6000: Prefer assigning the MMA vector operands to altivec registers [PR105556]

[Peter Bergner <bergner@linux.ibm.com>]

On 5/10/22 5:46 PM, Peter Bergner wrote:
>> Out of interest, did you try using v,?wa (so just two alternatives, not
>> four)?  Or did you think it wouldresult in  measurably worse code?  Or
>> did you decide it is not such bad backend code size explosion after
>> all :-)
> 
> I have not tried that, but I will and see what happens.
> I'm slightly worried that since 'v' is part of 'wa', then saying
> '?wa' might just cancel the 'v' usage, but I'll see how it works.
> If it works and simplifies the patch, great!  I'll report back.

So using 'v','?wa' for both vector operands does give us the code we want
for this loop.  Yea!

However, looking closer, that 2 alternative version isn't quite as "optimal"
as the version with 4 alternatives 'v','v','?d','?d' and 'v','?d','v','?d'.
That's because the 2 alt version is an all or nothing thing, you either get
both 'v' operands or you potentially get both 'd'.  The 4 alt version is
more fine grained and has alternatives where one gets a 'v' for sure and the
other a 'd'.

That said, if even one operand is forced into a 'd' register, then our chances
of good performance are lost, so the fact we can get one in a 'v' probably
doesn't matter and given it's cleaner code with fewer alternatives, let's go
with 'v','?wa'.  Ok with you?

I'll kick off a bootstrap and regtest with that change.

Peter