4.4/4.5 PATCH: PR middle-end/45678: [4.4/4.5/4.6 Regression] crash on vector code with -m32 -msse

[H.J. Lu]

On Tue, Sep 21, 2010 at 2:49 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Mon, Sep 20, 2010 at 04:45:07PM -0700, H.J. Lu wrote:
>> Here is the patch for 4.4/4.5. fold_builtin_memory_op is very different
>> in 4.4/4.5.  Simple backport doesn't work.  Does this patch make
>> any senses?
>
> Your builtins.c part doesn't make sense.  The code already handles
> the alignment, see the
>          srctype = build_qualified_type (desttype, 0);
>          if (src_align < (int) TYPE_ALIGN (srctype))
>            {
>              if (AGGREGATE_TYPE_P (srctype)
>                  || SLOW_UNALIGNED_ACCESS (TYPE_MODE (srctype), src_align))
>                return NULL_TREE;
>
>              srctype = build_variant_type_copy (srctype);
>              TYPE_ALIGN (srctype) = src_align;
>              TYPE_USER_ALIGN (srctype) = 1;
>              TYPE_PACKED (srctype) = 1;
>            }
> hunk and similar hunk for dsttype.  The difference is just that
> 4.6 creates a valid unaligned MEM_REF while 4.5 creates a valid unaligned
> VCE.  As you can see above, in 4.5 and earlier it was using a packed type
> properly no matter whether target was STRICT_ALIGNMENT or not.
> The problem is during expansion that the VCE isn't expanded properly.
>
> So, IMHO we need something like this (so far untested; 4.5 version):
>
> 2010-09-21  Jakub Jelinek  <jakub@redhat.com>
>
>        PR middle-end/45678
>        * expr.c (expand_expr_real_1) <case VIEW_CONVERT_EXPR>: If
>        op0 isn't sufficiently aligned and there is movmisalignM
>        insn for mode, use it to load op0 into a temporary register.
>
>        Backport from mainline
>        2010-09-20  Jakub Jelinek  <jakub@redhat.com>
>
>        PR middle-end/45678
>        * cfgexpand.c (expand_one_stack_var_at): Limit alignment to
>        crtl->max_used_stack_slot_alignment.
>

With your patch, gcc 4.5 generates:

	xorps	%xmm0, %xmm0
	movlps	(%esp), %xmm0
	movhps	8(%esp), %xmm0

on gcc.dg/torture/pr45678-2.c. Where does xorps come from?



-- 
H.J.