[Bug target/43743] New: Dumb use of SSE regs for MMX operation

[ian at airs dot com]

On 32-bit x86 I compiled this test case with -msse2 -mmmx:

#include <mmintrin.h>
__m64 foo (int *p)
{
  __m64 m1 = _mm_cvtsi32_si64 (*p++);
  __m64 m2 = _mm_cvtsi32_si64 (*p);
  return _mm_unpacklo_pi8 (m1, m2);
}

Note that all operations are MMX functions defined in mmintrin.h.  The
resulting assembler code includes the following:

        movd    (%eax), %xmm0
        movq    %xmm0, -8(%ebp)
        movd    4(%eax), %xmm0
        movq    -8(%ebp), %mm0
        movq    %xmm0, -16(%ebp)
        punpcklbw       -16(%ebp), %mm0

If I take off -msse2, I get this:

        movd    (%eax), %mm0
        movd    4(%eax), %mm1
        punpcklbw       %mm1, %mm0

That is not optimal--the 4(%eax) can appear directly in the punpcklbw
instruction--but it is much better than the first result.  The mere presence of
-msse2 should not cause SSE registers to be used when only MMX operations are
requested.


-- 
           Summary: Dumb use of SSE regs for MMX operation
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: ian at airs dot com
GCC target triplet: i686-pc-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43743