[Bug rtl-optimization/24160] [4.1 Regression] ICE with -O1 -ftree-vectorize -msse

["uros at kss-loka dot si" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #9 from uros at kss-loka dot si  2005-10-04 12:45 -------
The problem here is that we are out of GENERAL_REGS at the point of. This can
be seen in code, produced with -fomit-frame-pointer:

...
.L4:
        movl    1052(%esp), %edx
        movl    %ebp, (%edx,%ecx,4)         <<<<< load to %ebp
        incl    %ecx
        cmpl    %ecx, %eax
        ja      .L4
        movl    $10, %edx
        movl    %edx, %esi
        subl    %ecx, %esi
        cmpl    $10, %eax
        je      .L13
        subl    %eax, %edx
        movl    %edx, 4(%esp)
        movl    %edx, %ebx
        shrl    $2, %ebx
        movl    %ebx, %edi
        sall    $2, %edi
        je      .L8
.L9:
        movl    %ebp, (%esp)          <<< here %ebp is spilled to stack
        movss   (%esp), %xmm1
        shufps  $0, %xmm1, %xmm1
        movaps  %xmm1, %xmm0
        movl    1052(%esp), %edx
...

It looks that register allocator is running out of registers. It allocates the
last available reg, %ebp, that is usually used as a frame pointer. x86_64
target doesn't have this problem, as it has more registers (and is
-fomit-frame-pointer by default).

This bug is unfortunatelly not fixable in the i386 backend without fixing
current register allocator.

BTW: Peerhaps a tree expert could look into an optimized tree dump, maybe
something can be done there.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24160