Re: In PG12, query with float calculations is slower than PG11

Hi,

On 2020-02-06 14:25:03 +0900, keisuke kuroda wrote:
> That's because check_float8_val() (in PG 12) is a function
> whose arguments must be evaluated before
> it is called (it is inline, but that's irrelevant),
> whereas CHECKFLOATVAL() (in PG11) is a macro
> whose arguments are only substituted into its body.

Hm - it's not that clear to me that it is irrelevant that the function
gets inlined. The compiler should know that isinf is side-effect free,
and that it doesn't have to evaluate before necessary.

Normally isinf is implemented by a compiler intrisic within the system
headers. But not in your profile:
> ★ 5.41% postgres libc-2.17.so [.] __isinf

I checked, and I don't see any references to isinf from within float.c
(looking at the disassembly - there's some debug strings containing the
word, but that's it).

What compiler & compiler version on what kind of architecture is this?

Greetings,

Andres Freund