RFA (mmix): FIx PR46739

[Joern Rennecke]

Quoting "Joseph S. Myers" <joseph@codesourcery.com>:

> I don't think the code is correct either before or after this patch;
> aren't the halves of a CONST_DOUBLE in units of HOST_WIDE_INT not long?

That is true; but this is really a separate issue.  Note, for this to
break, there have to be two distinct type sizes larger than that of long.
Also worrying is the cast of the low part to unsigned, that will even break
if we have int / long / long long (or int128) as 32 / 64 / 128 bit.

> The whole block
...
> should be substantially simplified given that 2.95 not 2.7.2 is the
> minimum version for building a cross compiler.

It looks as if this was a specific decision to support 2.7.2 by the port
maintainer, albeit that was a while ago.  I'll leave it to the port
maintainers to decide if this is now obsolete.