On 27/01/15 00:51, Andres Freund wrote:
> On 2015-01-26 15:35:44 -0800, Peter Geoghegan wrote:
>> On Mon, Jan 26, 2015 at 3:12 PM, Andrew Gierth
>> <andrew@tao11.riddles.org.uk> wrote:
>>> Obvious overheads in float8 comparison include having to check for NaN,
>>> and the fact that DatumGetFloat8 on 64bit doesn't get inlined and forces
>>> a store/load to memory rather than just using a register. Looking at
>>> those might be more beneficial than messing with abbreviations.
>>
>> Aren't there issues with the alignment of double precision floating
>> point numbers on x86, too? Maybe my information there is at least
>> partially obsolete. But it seems we'd have to control for this to be
>> sure.
>
> I think getting rid of the function call for DatumGetFloat8() would be
> quite the win. On x86-64 the conversion then should amount to mov
> %rd?,-0x8(%rsp);movsd -0x8(%rsp),%xmm0 - that's pretty cheap. Both
> instructions have a cycle count of 1 + L1 access latency (4) + 2 because
> they use the same exection port. So it's about 12 fully pipelineable
> cycles. 2 if the pipeline can kept busy otherwise. I doubt that'd be
> noticeable if the conversion were inlined.
>
IIRC the DatumGetFloat8 was quite visible in the perf when I was writing
the array version of width_bucket. It was one of the motivations for
making special float8 version since not having to call it had
significant effect. Sadly I don't remember if it was the function call
itself or the conversion anymore.
-- Petr Jelinek http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training &
Services