707635 - Number->Atom about 4x faster than float->atom

Status: VERIFIED INVALID

(Tamarin Graveyard :: Virtual Machine, defect, P3)

Description

[Lars T Hansen]

I'm attaching a simple benchmark.  On my Mac Pro, R-D 32-bit build with -Dnodebugger from TR-float 6999, the loop f1() runs in 476ms, the loop f2() in 102ms.  The core of each loop is a conversion from float and Number, respectively, to atom.  This suggests that floatToAtom needs more work.

Comment 1

[Lars T Hansen]

Virgil notes the following in email:

"I beg to differ: floatToAtom can’t be faster (if I remember it correctly).

If my memory serves me right, doubleToAtoom_SSE2 will be faster indeed – but only when the Atom turns out to be an integer. We don’t have the luxury to encode ANY float value as an atom, we allocate each time."

Comment 2

[Lars T Hansen]

Indeed when the data are non-integers, the two loops run the same speed.

Comment 3

[Lars T Hansen]

Closing as invalid - fact of life until we have different representations.