[Bug optimization/14559] New: casts between vector pointer types are very slow

[Bug optimization/14559] New: casts between vector pointer types are very slow

[michaelni at gmx dot at]

see attached trivial code (return *(mmxdw*)&w;) 
 
gcc  testyy.c -O3 -mtune=pentium3 -march=pentium3 -fomit-frame-pointer 
generates: 
test: 
        subl    $20, %esp 
        movl    24(%esp), %eax 
        movq    %mm0, 8(%esp) 
        movl    8(%esp), %ecx 
        movl    12(%esp), %edx 
        movl    %ecx, (%esp) 
        movl    %edx, 4(%esp) 
        movq    (%esp), %mm0 
        movq    %mm0, (%eax) 
        addl    $20, %esp 
        ret     $4 
 
human generator: 
        movl    4(%esp), %eax 
        movq    %mm0, (%eax) 
        ret     $4 
 
additionally note, that reading after writing from the same address with a 
different size may cause partial memory stalls

-- 
           Summary: casts between vector pointer types are very slow
           Product: gcc
           Version: 3.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: michaelni at gmx dot at
                CC: gcc-bugs at gcc dot gnu dot org
  GCC host triplet: pentium3-debian-linux
GCC target triplet: pentium3-debian-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14559

[Bug optimization/14559] casts between vector pointer types are very slow

[michaelni at gmx dot at]

------- Additional Comments From michaelni at gmx dot at  2004-03-12 17:39 -------
Created an attachment (id=5910)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=5910&action=view)
the source


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14559

[Bug optimization/14559] casts between vector pointer types are very slow

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-03-12 17:46 -------
This time this is invalid as you should be using an union instead of a cast as the vecotrs types are in 
two different aliasing sets.
The following code produces the code you want.
 typedef short mmxw  __attribute__ ((mode(V4HI)));
 typedef int   mmxdw __attribute__ ((mode(V2SI)));
union t
{
 mmxw t1;
 mmxdw t2;
};
mmxdw test(mmxw w){
    union t t3;
    t3.t1 = w;
    return t3.t2;
}

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID
            Summary|casts between vector pointer|casts between vector pointer
                   |types are very slow         |types are very slow


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14559