On 06/06/2017 11:48 PM, William D Clinger wrote:
 > The spec for fl+* says it computes its result "as if to infinite
 > precision and rounded only once."  That's fine as an aspirational
 > goal, but it's overly ambitious as a requirement.  To implement
 > it portably, or on hardware that doesn't have machine instructions
 > that satisfy the requirement, satisfying the requirement probably
 > involves conversion to an exact representation for the operations
 > followed by conversion back to a flonum.  That's going to be quite
 > slow.  If speed is one of this procedure's goals, as I gather from
 > the discussion, the SRFI shouldn't impose absolute requirements
 > that force implementors to make the procedure much slower (on some
 > platforms) than the obvious composition.

Will:

In a previous email to this list:

http://srfi-email.schemers.org/srfi-144/msg/5816099

I stated:

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
In reviewing my previous comments, I see I should say more about fl+*:

The "fma" operation means "*fused* multiply-add", if it isn't fused it
isn't fma.  If fl+* isn't fma, then  fl-fast-fl+* isn't FP_FAST_FMA.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

The language "as if to infinite precision and rounded only once" is from
the C99 standard (I believe).  I don't see any benefit to fma without
the "fused" part of the operation, which requires precisely the "as if
to infinite precision and rounded only once" part.

Brad