[Bug c/34170] New: wrong code: mmx operation changes double to nan

["tijl at ulyssis dot org" <gcc-bugzilla at gcc dot gnu dot org>]

I was working on a little routine to generate random vectors of two doubles,
but it produced nothing but {nan,nan}. It was a simple linear congruential
random number generator, so all it did was operate on a seed (a vector of two
integers) and then convert that to double using an SSE2 intrinsic. The result
was then multiplied by (1.0 / 2147483648.0) to give random numbers between -1
and 1. It was (1.0 / 2147483648.0) that turned out to be converted to nan.

I've reduced it to a minimal test case below. It prints out nan instead of 1.0.
You can compile it with: `cc -O0 -Wall -mmmx -o test test.c`

--- begin test.c ---
#include <stdio.h>
#include <mmintrin.h>

static __m64 s;
int main( int argc, char **argv ) {
    __m64 a;
    double d;
    d = 1.0;
    s = a + a;
    printf( "%e\n", d );
    return 0;
}
--- end test.c ---


-- 
           Summary: wrong code: mmx operation changes double to nan
           Product: gcc
           Version: 4.2.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: tijl at ulyssis dot org
 GCC build triplet: i386-unknown-freebsd7
  GCC host triplet: i386-unknown-freebsd7
GCC target triplet: i386-unknown-freebsd7


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34170