On Tue, Jan 30, 2024 at 4:13 AM Ants Aasma <ants.aasma@cybertec.at> wrote:
>
> But given that we know the data length and we have it in a register
> already, it's easy enough to just mask out data past the end with a
> shift. See patch 1. Performance benefit is about 1.5x Measured on a
> small test harness that just hashes and finalizes an array of strings,
> with a data dependency between consecutive hashes (next address
> depends on the previous hash output).
I pushed this with a couple cosmetic adjustments, after fixing the
endianness issue. I'm not sure why valgrind is fine with this way, and
the other ways I tried forming the (little-endian) mask raised errors.
In addition to "zero_byte_low | (zero_byte_low - 1)", I tried
"~zero_byte_low & (zero_byte_low - 1)" and "zero_byte_low ^
(zero_byte_low - 1)" to no avail.
On Thu, Mar 28, 2024 at 12:37 PM Jeff Davis <pgsql@j-davis.com> wrote:
> 0001 looks good to me, thank you.
>
> > v21-0003 adds a new file hashfn_unstable.c for convenience functions
> > and converts all the duplicate frontend uses of hash_string_pointer.
>
> Why not make hash_string() inline, too? I'm fine with it either way,
> I'm just curious why you went to the trouble to create a new .c file so
> it didn't have to be inlined.
Thanks for looking! I pushed these, with hash_string() inlined.
I've attached (not reindented for clarity) an update of something
mentioned a few times already -- removing strlen calls for dynahash
and dshash string keys. I'm not quite sure how the comments should be
updated about string_hash being deprecated to call directly. This
patch goes further and semi-deprecates calling it at all, so these
comments seem a bit awkward now.