performance regression for Core.Intrinsicts.rem_float in 1.8/1.9

[lmiq]

I noticed this in the use of mod, but it seems that the issue can be tracked down to Core.Intrinsics.rem_float:

On 1.8.0:

julia> @btime Core.Intrinsics.rem_float(x, 1.0) setup=(x=rand())
  7.422 ns (0 allocations: 0 bytes)
0.11718355488076748

On 1.7.3:

julia> @btime Core.Intrinsics.rem_float(x, 1.0) setup=(x=rand())
  2.208 ns (0 allocations: 0 bytes)
0.21209837479215266

This is causing a significant performance regression in a package of mine which uses mod many many times.

system info:

Linux m3g 5.15.0-46-generic #49-Ubuntu SMP Thu Aug 4 18:03:25 UTC 2022 x86_64 x86_64 x86_64 GNU/Linux

edit: on the nightly build:

julia> @btime Core.Intrinsics.rem_float(x, 1.0) setup=(x=rand())
  7.414 ns (0 allocations: 0 bytes)
0.2578476302582853

[oscardssmith]

Is this a difference in which libm we call?

[lmiq]

I noticed now that similar performance regressions can be felt in the floor, trunc and other functions of this family.

[oscardssmith]

that's unsurprising but quite bad.

[oscardssmith]

Also, this affects 1.9 as well.

[gbaraldi]

Was this a change in our side? Did we change from system libm to Openlibm or something like it?

[oscardssmith]

yeah we're hitting the openlibm one now. The ideal fix would be to move this to pure Julia but I doubt I'll get around to it.

[gbaraldi]

Could we port this algorithm. It doesn't seem too too complicated. And license wise it would be fine I guess.

[oscardssmith]

want to take a crack at it?

[gbaraldi]

Sure, I'll try it out, see how it compares :)