Re: Omitted conversion float -> double in x86 back-end

[Andrew Haley <aph at redhat dot com>]

Michael Matz writes:
 > Hi,
 > 
 > On Tue, 21 Nov 2006, Roberto COSTA wrote:
 > 
 > > Hi Michael,
 > > too late, I've filed the report a few minutes ago.
 > > By the way, if I want to bypass the problem without paying any performance
 > > penalty, it is sufficient that I introduce a temporary variable in the C
 > > source.
 > > As a matter of fact, this looks to work:
 > > 
 > > double foo(signed long long l)
 > > {
 > >     double d = (float)l;
 > >     return d;
 > > }
 > 
 > Pure luck.  Try with -O2 or another compiler version and it will stop 
 > working again.  IOW: it's no workaround for the x87 problem.  The only one 
 > there is is -ffloat-store,

or volatile, in the one place where you really need to do the conversion.

double foo(signed long long l)
{
  volatile float tmp = l;
  return tmp;
}

foo:
        subl    $16, %esp
        fildll  20(%esp)
        fstps   12(%esp)
        flds    12(%esp)
        addl    $16, %esp
        ret

Andrew.