On 6/30/23 04:14, Robin Dapp wrote:
The explicit conversions I see are because we need the output of the
conversion in multiple vfmul instructions. That won't be helped by
the patch you've proposed.
FWIW on my local branch and the patch applied I see that the vfwmuls
are being generated (all of the vfmuls are replaced).
It'll need to be a define_insn_and_split as its a 3->3 splitter. The
split will emit the two extensions and the widening multiply as 3
distinct insns.
I tried this and while it worked for the first vfwmul the subsequent
ones are not being combined/optimized. Now I'm not a combine expert
at all but it looks as if the source float_extends are being deleted
deferring deletion of insn with uid = 39.
deferring deletion of insn with uid = 37.
with that pattern successfully matched, while they are only "rescanned"
with the synthetic "single widen" one. Them being deleted (or rather
absorbed by the vfwmul) no further combination is possible (until after
split?)
This seems to be a fundamental difference between the two approaches.
Maybe the "double widen" pattern can be adjusted to also handle this
or I did something wrong when writing the splitter?
With the "single widen" pattern, however, it works more or less
naturally therefore I'd still suggest going for it.
I'd hoped to have time to revisit all of this today, but I'm quickly
running out of time.
There has to be some kind of mismatch between the patch or testcase or
what we're looking at to judge success.
Monday and Tuesday are holidays in the US. Naturally that means the
rest of my work week is going to be busier than normal. I don't want to
hold things up unnecessarily.
While I really don't see the need to have the bridge pattern, I'm still
willing to believe that I've missed something, which is why I wanted to
dive into it myself. For example, we have heuristics to avoid trying
too many 4->n combine patterns and we might be tripping over that or who
knows what.
So my suggestion is that if both of you are getting the desired code,
then Robin handle the review side of the two patches that introduce the
helper patterns.
Jeff
