[Bug target/16872] invalid use of MMX instructions

[jbeulich at novell dot com]

------- Additional Comments From jbeulich at novell dot com  2004-08-05 06:47 -------
I disagree here. This is not a dup of 14801, and following the Intel model of
using MMX stuff 14801 is not even a bug (instead, it's a programming error in
that f2() in that example deals with both MMX intrinsics and FP types). See the
Intel C++ Compiler User's Guide, section EMMS Usage Guidelines.
Other than for XMM register usage, MMX registers must be used only when
explicitly indicated by the programmer, since making the compiler automatically
insert [f]emms everywhere needed (and also recognise FP and MMX registers as
aliases of one another) would be a rather difficult job. This also applies to
the comment on the autovectorizer - it'd have to be rather smart to be allowed
to use MMX stuff correctly (though I'm not saying it's not doable). In any case,
correct code generation is by far more important than advanced features.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |UNCONFIRMED
         Resolution|DUPLICATE                   |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16872