maria-discuss team mailing list archive

[Marko Mäkelä <marko.makela@xxxxxxxxxxx>]

Hello Kevin,

It was my pleasure to meet you in Shanghai.

On my flight back, I worked on a micro-optimization, trying to make
sure that native loads or stores are being used instead of memcpy(),
memset(), memcmp(), when the data is known to be aligned. I filed a
ticket for it:
https://jira.mariadb.org/browse/MDEV-21133 Optimize access to InnoDB
page header fields

My colleague Eugene Kosov pointed out that such loads or stores are
undefined behaviour (and cmake -DWITH_UBSAN=ON would likely agree).
But, he showed that wrapping the arguments of <string.h> functions
with __builtin_assume_aligned() actually works:
https://godbolt.org/z/jCF_6q
Eugene also pointed out to some related work:
http://open-std.org/JTC1/SC22/WG21/docs/papers/2019/p1774r1.pdf

I found a claim that Aarch64 does support unaligned access in practice:
https://stackoverflow.com/questions/38535738/does-aarch64-support-unaligned-access
Can you provide a more authoritative answer? Is there some flag that
should be passed to gcc or clang to enable it to generate simpler
code?
I also found a claim that POWER8 supports unaligned access, and I seem
to remember that the latest version of the SPARC introduced support
for that as well. (IA-32 and AMD64 have always supported unaligned
access, except for some SIMD operations.)

Last, I believe that we could get some performance benefits if
include/byte_order_generic.h was rewritten in a suitable way. Ideally,
include/byte_order_generic_x86_64.h would be replaced with a portable
version of both, and compilers could simply perform the optimizations.
I have been told that replacing the + in the macros with | could
already be a good start. I would welcome patches in this area.

Related note: Maybe a year ago, I was positively surprised to learn
that the InnoDB monster function mach_read_from_4() is being
translated into a single 80486 BSWAP instruction, or an AMD64 MOVBE
instruction.

With best regards,

Marko