Solution for problem implicit conversion real -> double complex

[Caspar Terheggen]

I would like to thank everyone who responded to my article about real ->
double complex conversion. Since one of them asked me to summarize to the
net and this might be interesting for others as well, here goes:

In my previous article I mentioned that in Fortran 77

      complex*16 z
      z = 1.0/1.0

would give a wrong result on our Sun4, i.e. z wasn't equal to one.  The
problem is caused by a bug in version 1.1 of the Sun f77 compiler. Several
people told me that it is fixed in version 1.2.  Examining the assembly
code generated by the compiler Dan Mick narrowed the problem down to a
call to Ff_conv_z. According to Dan this routine starts with a SPARC fitod
instruction, which translates from *integer* to double precision floating
point.  Obviously this causes the conversion to fail.

Thank to Michael M. Fischbein and Daniel Ehrlich I found out which version
of f77 we have by looking at /usr/lib/lang_info (I didn't know about this
file). Roland Schneider and Dik Winter (many thanks) suggested a
(temporary) fix, which seems to work fine. Roland writes:

|        ... and they [Sun] suggested using inline floating point (hang
| "/usr/lib/libm.il" on the end of your COMPILE command.  This fixes the
| problem by avoiding a call to the conversion routine in the library.
| This is documented in the floating point programmers guide.

I find this 'fix' particularly useful, because it avoids the need to find
and change any source that might depend on implicit conversion. Still,
updating to the new version of the compiler and/or inserting explicit
conversion routines like CMPLX() (as Adam Mirowski suggested in his
article) may be more secure.

Also a few people mentioned that complex*16, both the existence and
notation, isn't standard Fortran 77, but is part of Sun's extension.

Regards,
 Caspar Terheggen  +31 80 612800  caspar at sci.kun.nl  U634008 at HNYKUN11.EARN