Hi,
On Sat, Nov 09, 2024 at 08:00:35AM +0900, Michael Paquier wrote:
> On Fri, Nov 08, 2024 at 11:18:09PM +1300, David Rowley wrote:
> > I'm slightly worried due to the current rate we're receiving cleanup
> > suggestions that someone might come along and think they'd be doing us
> > a favour by submitting a patch to "fixup the inefficient bitwise-ORs
> > and use boolean-OR". Maybe a comment like the following might prevent
> > that from happening.
>
> Not sure, but OK by me to tweak things more.
>
> > Can you share your test case for this? I tried with [1] and the
> > latest gcc does not seem to be smart enough to figure this out. I
> > tried adding some additional len checks that the compiler can use as a
> > cue and won't need to emit code for the checks providing the function
> > does get inlined. That was enough to get the compiler to not emit the
> > loops when they'll not be used. See the -DCHECK_LEN flag I'm passing
> > in the 2nd compiler window. I just don't know if putting something
> > like that into the code is a good idea as if the function wasn't
> > inlined for some reason, the extra len checks would have to be
> > compiled into the function.
>
> Feel free to use that (I hope it works), and see the difference once
> the aligned structure is 121 bytes or more:
> https://godbolt.org/z/94393nPGG
>
> At least, I can see that the SIMD loop is ignored.
What I see (with the godbolt you shared) is that with BLCKSZ of 120:
gcc: then no SIMD instructions are used (I think that's because sizeof(AlignedBlock)
is 120 which is not a multiple of 16 (SIMD xmm register size)).
with BLCKSZ of 121:
gcc: then SIMD instructions are used (I think that's because sizeof(AlignedBlock)
is 128 which is a multiple of 16 (SIMD xmm register size)).
While clang uses SIMD instructions in both cases (more complex code with more
branches at least in the 120 case).
Regards,
--
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com
