688744 - float modulo incorrectly produces the same result as double modulo

Status: VERIFIED WONTFIX

(Tamarin Graveyard :: Virtual Machine, defect)

Description

[Brent Baker]

f1%f2 == Number(f1)%Number(f2) this should not be true since the float modulo should happen in single precision, which should produce worse precision than the double modulo which should happen in double precision, so the results of the 2 should NOT be equal.

Comment 1

[Brent Baker]

Comments from Virgil:

modulo on float & number should THEORETICALLY yield different results. However, on gcc & VisualC (windows/x86), fmod seems to return identical results with fmodf if the input values are identical (and no, I didn't get fooled by the "fmod" version with float arguments - I explicitly casted the parameters to double). I suspect the runtime library implementations perform both operations on 64 bits. The thing is, there's no hardware instruction for modulo - except the x87 FPU (it has a "FPREM" instruction). Bottom line is - the C runtimelibrary makes no diffeernce (not on x86 and x64, Mac/Win, at least)

Comment 2

[Brent Baker]

Virgil -> QRB: We could theoretically fix it by implementing our own modulo, but....  probably not worth the effort

Comment 3

[Lars T Hansen]

Brent, what are the inputs you have asumed here (the values of f1 and f2)?

Comment 4

[Brent Baker]

(In reply to Lars T Hansen from comment #3)
> Brent, what are the inputs you have asumed here (the values of f1 and f2)?

var f1:float = 1.2345678e9;
var f2:float = 6.543210987;

Comment 5

[Lars T Hansen]

f1 % f2 must have run-time type "float" though, it must never have run-time type "Number".  I think we can live with the computation being slow (we already have a bug open on % being slow in general but that's worse for Number than for float) and being performed in the Number domain, as long as the result is of type float.

I should probably add appropriate language to the spec about this.  I will file a bug.

If we ever end up with a proper float modulo then this is a pretty easy thing to add a versioned fix for.

Comment 6

[Brent Baker]

Closing as NAB. ABC Extensions spec (draft 5) states that Float modulo can operate in double precision

6.1 Float:
Float modulo may operate with double precision.
Note: It appears that most C/C++ run-time libraries implement float modulo (fmodf) on top of the double 
modulo method (fmod), possibly because FPUs only support the latter.