1) optimization and 2) rounding modes.

[Jeff Enderwick]

1) optimization: I think the goal here should be to maximize the
utility of the language while avoiding unnecessary performance
penalties. With respect to floating point operations, I think that a
programmer ought to be able to get the same answer from two different
compilers for the same CPU, if he/she wants to. Bit for bit. Also,
programmers who want speed over accuracy shouldn't pay extra for
nothing. Have y'all considered adding a type qualifier for floats such
as "serial", where an expression containing any two such operands may
not have its operands reorganized, etc. From what (little) I know,
this is a reason why some numerical types blow-off C.

2) rounding modes: I think we should keep this out of the language. I
agree that there are times when you'd like to control the rounds on a
per-operation basis, but how many modes will be enough ? For those of
you who write compilers: do you want to have to implement a full set
of floating point arithmetic for each mode that your target doesn't
support ? Maybe we should pick a few modes and have standard ways to
detect them at compile time. Something like a defined constant
NEAREST_EVEN and #pragmas or type qualifiers to specify where and/or
what. I really think that implementing multiple rounding modes should
be optional.

I'm new to this mailing list, so these ideas have probably been stated
before. If so, just hit the D key ...


	Jeff Enderwick
	jeffadsp.sps.mot.com