Re: whole number constants

dancerchris@[EMAIL PROTECTED]
 wrote:
(snip)

> Thanks RM.  I do understand that a parameter has a constant value.  My
> question was the relation****p of inline usage of a number like label
> 50 of my original post.  Do compilers use the same value for multiple
> use of an integer in an equation of specified kind

> for example if I specify a constant:

> real(dp), parameter :: one = 1.0_dp
> real(dp) :: x,xnew,error

> 100 error = one-xnew/x+one
> 200 error = 1.0_dp-xnew/x+1.0_dp
> 300 error = 1-xnew/x+1

> it seems to me that label 100 is using a true constant such that the
> numerical value out to the machine level (past kind dp) is consistant
> in the equation (i.e. is copied from the storage register completely,
> not just the significant digits).  I think that label 200 does not
> guarantee that, each instance of "1.0_dp" is assigned a value to the
> dp accuracy  (plus extraneous bits)  300 might but it depends on how
> integers are assigned  to real values and whether each instance of 1
> is consistant.

For mixed mode arithmetic, such as between different KINDs of
REAL values, the shorter is always extended to match the
longer type before the operation (+, -, *, or /) is performed.
(** is a different question).   I don't know that the standard
requires that zero bits be used in doing that, but all systems
I know do that.  With zero extending integer 1 or single precision
1. or 1.E0 will exactly equal double precision 1.D0.

> Sorry if I'm being thick.

There is an interesting feature that I only know of in the
IBM Stretch machine.  When doing post normalization on
floating point values, one can choose to ****ft in zeros
or ones.  One can run a program both ways, and compare
the results.  Using that option, 1000001.-1000000. would
be much different from 1.0, but I don't believe that
extending either integer or single precision would
be done differently.

There is also stochastic rounding, where values half way
in between two rounded values are randomly rounded one
way or the other.  One might also suggest using random
bits for post normalization.  I don't know any that
do, though.

-- glen