On 2018-10-16 16:36:12 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > Attached is a patch that shrinks fmgr_builtins by 25%. That seems
> > worthwhile, it's pretty frequently accessed, making it more dense is
> > helpful. Unless somebody protests soon, I'm going to apply that...
>
> Hah. I'm pretty sure that struct *was* set up with an eye to padding ...
> on 32-bit machines.
Possible, the new layout should work just as well there, luckily.
> This does make it shorter on 64-bit, but also
> makes the size not a power of 2, which might add a few cycles to
> array indexing calculations. Might be worth checking whether that's
> going to be an issue anywhere.
I can't imagine that that outweight the cost of additional cache misses
on any platform where performance matters. On x86 I assume indexing
into an array with 24byte stride, will normally be just two leas (lea
eax, [eax + eax * 2]; lea eax, [ebx + eax * 8]; where eax initially is
the index, and ebx the array base). Indexing also plays less of a role
than in the past, because previously we did a binary search, but now we
normally look up the index via fmgr_builtin_oid_index.
> What's the point of the extra const decoration on funcName? ISTM
> the whole struct should be const, or not.
The whole array is const. There's cases where that allows the compiler
more freedom - but I guess it's really a bit redundant here.
Greetings,
Andres Freund