Hi Robin,
On 30/05/2025 10:18, Robin Dapp wrote:
BTW it's the late-combine pass that performs the optimization, not the
combine pass. You might still want to fix this in the commit message.
I updated the commit message.
Actually, it depends whether or not the vec_duplicate is hoisted to the
loop preamble. The testcases in the patch are very simple so it's
late-combine; but in the SPEC example, where register pressure is high,
it is combine.
Please CC patchworks...@rivosinc.com for the next version so we have CI
coverage.
Hmm, that somehow didn't work. ISTM that it doesn't pick up "new"
patches in a thread. Could you re-send in a new thread please? It
should pick it up even without CC then. If the CI is green this is good
to go.
...
overall the patch looks reasonable to me now, provided the fr2vr followup.
I created a new thread with a patchset including both this one and the
fr2vr followup:
https://gcc.gnu.org/pipermail/gcc-patches/2025-May/685166.html
Checkstyle has a few minor complaints still, like tabs vs spaces in the
md changes, but we haven't been very strict there so I won't insist ;)
I fixed the tab/space error.
Starting with vfmn{add,sub}, my plan is to first check whether actual
gains can be achieved and if so I will be coming up with a patch
similar to this one. How does that sound?
The other patterns are definitely necessary, regardless of whether they
show up in SPEC or not.
Sure, will do.
Thanks,
--
PA
