(j3.2006) parameterized derived types

[Robert Corbett]

On 12/10/15 15:42, Bill Long wrote:
> I appreciate that users ask for ?slower, more accurate? versions of functions for various reasons.  If they later complain that then are getting different answers, then the response is ?that?s what you asked for, wasn?t it??   This seems like something the standard should not get involved with.  Maybe just a documentation notation reminding readers that the ?slower, more accurate? functions will return slightly different answer from the default functions.  If they didn?t, there would be no point to the ?slower? bit.
>

I should have been clearer.  My comments about the math libraries were intended 
as part of the discussion about parameterized derived types.  Malcolm suggested 
that a possible solution to the problem was to ensure that component 
specification expressions produce the same results regardless of whether they 
are evaluated at compile-time or run-time.  My point was that I do not have 
control over which math library will be used for a given program, so making 
guarantees about expressions that involve the intrinsic transcendental functions 
is difficult but not impossible.  Operations such as raising a value to a power 
are also issues.

I could implement the desired functionality by adding the functions the compiler 
uses to evaluate constant expressions to the Fortran run-time support library.  
The functions will need to given new names, and for the sake of backward 
compatibility, whenever one of the functions is changed in the compiler another 
routine with a new name will have to be added to the library.  This solution to 
the problem would be hard to maintain, especially after I am gone.

Bob Corbett