[INTERP] Fix fpu_round

[ksco]

The nearbyint functions round their argument to an integer value in floating-point format, using the current rounding direction, therefore it's not "ROUND_Nearest". We should use round.

[ptitSeb]

I'm pretty sure nearbyint is used in more than 1 place!

[ksco]

Ah, it does! Let me change all of 'em then.

[ksco]

Ok, done.

[ptitSeb]

Tests are not happy with this changes

[ksco]

Seems 0.5 in cvtps2pd cvtps2dq should be rounded to 0 in "ROUND_Nearest" mode, that's really strange, am I missing something? 🤔️

[ptitSeb]

Seems 0.5 in cvtps2pd should be rounded to 0 in "ROUND_Nearest" mode, that's really strange, am I missing something? 🤔️

Mmm no, cvtps2pd 0.5 should still be 0.5. Value should not be round to integer at all on cvtps2pd, as it's just packed float to packet double.

[ksco]

Sorry I mean the cvtps2dq, these names confuses me :'(

[ksco]

Ah, it actually "Round to Nearest, ties to Even", so 0.5 is expected to round to 0.

[ksco]

@ptitSeb seems working; also tested test22 with BOX64_DYNAREC_TEST=1 locally, works too (this is how I found the issue).

[ptitSeb]

Ok, we don't want to set the rounding mode at each function call and not letting it reset to it's previous value

The rounding mode might needed to be tracked (might need to actually wrap fesetround to be able to restore old value after round is done (or do nothing if the current rounding mode is already ok).

[ksco]

@ptitSeb I pushed some commits to take care of library wrapping, please review again.