[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: NaN Propagation



There is plenty of disagreement on the bitwise ORing of the operands, in any
radix. Just what do you assume the payload is here?

Bitwise OR is useful if and only if the sole payload is a one-hot powerset
showing which kind of NaN this is. That is a reasonable payload - but if
that's all you are going to permit the implementation, you might as well go
the whole way and specify which hot is which (i.e. make a NaN transparent)
so that NaNs are wholly portable across machines. Bit 1 is underflow, bit 2
is overflow, etc. However, there seems no good reason to preclude
implementations from representing the same payload as an enumeration instead
of a powerset, except the Or of enumerations becomes garbage.

And how about those implementations that wish to keep line number and other
debugger information in the NaN? Let's see: (function f line 17) OR
(function g line 5) == (function withNoFloatingPoint line
largerThanWholeProgram), or similar nonsense.

The inside of NaNs is an area where implementations should be able to
compete with good ideas without the standard getting in the way. There is
some reasonable argument (NOT yet settled in the group, but reasonable) that
the composition of NaNs in a commutative operation should be commutative.
Bitwise OR is one such implementation, but far from the only one. There is
some reasonable argument (NOT yet settled in the group, but reasonable) that
the composition of NaNs in any operation should yield one of the input NaNs.
Bitwise OR is fails this. There will be much discussion ahead.

Ivan

----- Original Message ----- 
From: "Mark Erle" <mark@xxxxxxxxxxxxxx>
To: <stds-754@xxxxxxxx>
Sent: Sunday, December 14, 2003 11:57 AM
Subject: NaN Propagation

---clip---
> I propose the result of an operation involving more than one NaN be the
> bit-wise OR of the significand portions of the NaNs, for operands in
binary
> formats and operands in decimal formats.
>
> There seemed to be little disagreement on the ORing of binary operands, so
> I'll just speak to the issue of decimal operands.
---clip---
> Mark A. Erle
> merle@xxxxxxxxxx
>


754 | revision | FAQ | references | list archive