exp is not fully compatible with PG

[lmatz]

For example:
select exp(-10000000); returns 0 in RW
but it errors underflow in PG

See comments at #7971 (comment)

PG has special logic for this: https://github.com/postgres/postgres/blob/REL_15_2/src/backend/utils/adt/float.c#L1649

[erichgess]

Just for clarification, this ticket to update RW to have the same behavior as PG when the exp instruction returns 0? In other words, the result of this ticket would be to recreate the special logic from PG within RW?

[lmatz]

Hi @erichgess, I think you are right, we intend to let RW have the same behavior as PG's, i.e. report underflow and overflow under precisely the same situation. cc: @xiangjinwu

[erichgess]

Your point on the other ticket about not artificially restricting RW to what PG can do makes a lot of sense, but in this case, certainly, I think what PG is doing is the correct path, because their goal is to have consistent behavior across all platforms.

Could I hop on this ticket?

[lmatz]

Sure @erichgess, I think your reasoning makes sense
please go ahead, thanks for your interest!

[erichgess]

Great, thank you.

[erichgess]

@lmatz PG has separate error codes for flow overflow and underflow, but RW only has NumericOutOfRange, which means that RW will return NumericOutOfRange when exp(x) == 0, so it will not perfectly mirror PG.

Is the long term plan to eventually have the error codes in RW have the same granularity as the error codes from PG?

[lmatz]

@erichgess
Yes, we also try to have error codes compatible with PG ones whenever possible.

If you'd like to use overflow and underflow instead of simply out of range, feel free to change it. This is a good point!

[erichgess]

@lmatz The PG implementation uses unlikely to hint to the compiler that some if branches are unlikely to be executed. Do you want to replicate the same optimizations in RW?

[lmatz]

@erichgess, sure if you'd like to,
I just took a look,
and it seems people have some best practice when using stable Rust:https://users.rust-lang.org/t/compiler-hint-for-unlikely-likely-for-if-branches/62102/4
without relying on the nightly feature

[erichgess]

Should this ticket be closed now that the PR is merged?

[xiangjinwu]

Thank you @erichgess