Re: PATCH [5/n]: Prepare x32: PR middle-end/48016: Inconsistency in non-local goto save area

["H.J. Lu" <hjl.tools@gmail.com>]

On Thu, Jun 16, 2011 at 12:56 AM, Richard Guenther
<richard.guenther@gmail.com> wrote:
> On Wed, Jun 15, 2011 at 9:55 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
>> On Wed, Jun 15, 2011 at 8:16 AM, Michael Matz <matz@suse.de> wrote:
>>> Hi,
>>>
>>> On Wed, 15 Jun 2011, H.J. Lu wrote:
>>>
>>>> >> +  /* FIXME: update_nonlocal_goto_save_area may pass SA in the wrong mode.  */
>>>> >> +  if (GET_MODE (sa) != mode)
>>>> >> +    {
>>>> >> +      gcc_assert (ptr_mode != Pmode
>>>> >> +               && GET_MODE (sa) == ptr_mode
>>>> >> +               && mode == Pmode);
>>>> >> +      sa = adjust_address (sa, mode, 0);
>>>> >> +    }
>>>> >
>>>> > That may be appropriate for a branch, but trunk shouldn't contain FIXMEs
>>>> > that explain how something should be fixed, instead that something should
>>>> > be carried out.  I.e. just fix update_nonlocal_goto_save_area.
>>>> >
>>>>
>>>> I don't know update_nonlocal_goto_save_area enough to fix it
>>>> without breaking other targets.  This patch is the lest invasive.
>>>> Any suggestions how to properly fix it is appreciated.
>>>
>>> Well, the most obvious variant would be to move the above code right
>>> before the call of emit_stack_save in update_nonlocal_goto_save_area
>>> (using r_save and STACK_SAVEAREA_MODE (SAVE_NONLOCAL)).  All other callers
>>> of emit_stack_save already make sure to pass an object of correct mode, so
>>> this one should too.
>>>
>>> But I think it's better to just produce a correct array_ref from the
>>> start.  get_nl_goto_field creates an array_type for the
>>> nonlocal_goto_save_area of correct type (ptr_type_node or
>>> lang_hooks.types.type_for_mode (Pmode, 1)), and we should use that.
>>>
>>> So something like this in update_nonlocal_goto_save_area:
>>>  t_save = build4 (ARRAY_REF,
>>>                   TREE_TYPE (TREE_TYPE (cfun->nonlocal_goto_save_area)),
>>>                   cfun->nonlocal_goto_save_area,
>>>                   integer_one_node, NULL_TREE, NULL_TREE);
>>>
>>> instead of the current building of t_save.  Then r_save also should get
>>> the correct mode automatically.
>>>
>>
>> Here is the updated patch.  OK for trunk?
>
> The explow.c change is ok.  For the function.c change I wonder why
> convert_memory_address doesn't do the right thing - from it's documentation
> it definitely should, so it should be fixed instead of being replaced by
> adjust_address with a zero offset.
>

convert_memory_address may return a pseudo register converted
to Pmode.  But here what we want is the same memory address
adjusted for Pmode.  I don't think the usage of convert_memory_address


-- 
H.J.