[ruby-core:20426] Re: Leave my open classes alone (was Behavior: autoload calls rb_require() directly)

On Sat, Dec 06, 2008 at 10:17:52AM +0900, Brent Roman wrote:
> 
> Paul,
> 
> I understand your concern, but it may not be an tricky as you believe.
> 
> I think the MagLev folks addressed this by truncating the binary
> exponent to 8 bits in their immediate object format while always
> preserving the full mantissa.  There should thus be no loss of
> precision, but there will be a loss of dynamic range.  Those few
> floats whose absolute value is greater than approx. 2**127 (1.0e+38)
> could not be represented as SmallFloats, so they would be
> automatically promoted to non-immediate Float objects, much like 2**30
> gets silently promoted to a Bignum in 32-bit MRI today.

A few comments:
- My initial concern was that arithmetic would be done on objects that
  don't have the same bit representation, but after thinking about it I
  see now that would be impossible.  As long as the immediate object is
  converted to a double with the same bit representation as the original
  before arithmetic operations are performed, there should be little
  problem.
- While this scheme would reduce the number of objects created, without
  seeing benchmarks I'm not convinced that the arithmetic necessary to
  extract the exponent and create the immediate object would be faster
  to execute than removing a pre-allocated object from a freelist of
  Floats and populate it (likely there would be an overall performance
  gain, due to Ruby's non-generational GC).  This is how Python's float
  objects are implemented (at least as of 2.4; I'm not familiar with
  newer versions).
- A 96-bit or 128-bit object reference representation would be able to
  represent all possible Floats, and would also be able to represent 3-
  and 4-tuples as immediate objects.

Paul