On 7/28/23 06:31, Robin Dapp via Gcc-patches wrote:
This is a draft patch. I would like to explain it's hard to make the
simplify generic and ask for some help.
There're 2 categories we need to optimize.
- The op in optab such as div / 1.
- The unspec operation such as mulh * 0, (vadc+vmadc) + 0.
Especially for the unspec operation, I found we need to write one by
one to match the special pattern. Seems there's no way to write a
generic pattern that will match mulh, (vadc+vmadc), sll... This way
is too complicated and not so elegant because need to write so much
md patterns.
Do you have any ideas?
Yes, it's cumbersome having to add the patterns individually
and it would be nicer to have the middle end optimize for us.
However, adding new rtl expressions, especially generic ones that
are useful for others and the respective optimizations is a tedious
process as well. Still, just recently Roger Sayle added bitreverse
and copysign. You can refer to his patch as well as the follow-up
ones to get an idea of what would need to be done.
("Add RTX codes for BITREVERSE and COPYSIGN")
So if we have few patterns that are really performance critical
(like for some benchmark) my take is to add them in a similar way you
were proposing but I would advise against using this excessively.
Is the mulh case somehow common or critical?
Well, I would actually back up even further. What were the
circumstances that led to the mulh with a zero operand? That would
tend to be an indicator of a problem earlier. Perhaps in the gimple
pipeline or the gimple->rtl conversion. I'd be a bit surprised to see a
const0_rtx propagate in during the RTL pipeline, I guess it's possible,
but I'd expect it to be relatively rare.
The one case I could see happening would be cases from the builtin
apis... Of course one might call that user error ;-)
jeff
