On Wed, Sep 2, 2015 at 4:12 PM, Andres Freund <andres@anarazel.de> wrote:
>> Well, still needs to work for tuplestore, which does not have a SortTuple.
>
> Isn't it even more trivial there? It's just an array of void*'s? So
> prefetch(state->memtuples + 3 + readptr->current)?
All I meant is that there couldn't be one centralized definition for
both. I don't mind if you want to bake it into a macro for
tupelstore.c and tuplesort.c.
>> Because of the way tuples are fetched across translation unit
>> boundaries in the cases addressed by the patch, it isn't hard to see
>> why the compiler does not do this automatically (prefetch instructions
>> added by the compiler are not common anyway, IIRC).
>
> Hardware prefetchers just have gotten to be rather good and obliterated
> most of the cases where it's beneficial.
Well, if hardware prefetchers are bright enough to do this perfectly
nowadays, then that implies a cost for useless prefetch instructions.
It might still be worth it even then, if the cost is very low for
these platforms. It's not as if they're required to respect the
prefetch hint in any way.
> I'd be interested to see a perf stat -ddd comparison to the patch
> with/without prefetches. It'll be interesting to see how the number of
> cache hits/misses and prefetches changes.
>
> Which microarchitecture did you test this on?
My laptop has an Intel Core i7-3520M, which is a mobile processor that
is a bit old. So, Ivy Bridge.
--
Peter Geoghegan