On Fri, Jul 15, 2011 at 8:25 AM, Uros Bizjak <ubiz...@gmail.com> wrote:
> On Fri, Jul 15, 2011 at 3:03 PM, H.J. Lu <hjl.to...@gmail.com> wrote:
>> On Fri, Jul 15, 2011 at 5:49 AM, Uros Bizjak <ubiz...@gmail.com> wrote:
>>> On Sun, Jul 10, 2011 at 12:20 AM, H.J. Lu <hongjiu...@intel.com> wrote:
>>>
>>>> TARGET_MEM_REF only works on ptr_mode. That means base and index parts
>>>> of x86 address operand in x32 mode may be in ptr_mode. This patch
>>>> supports 32bit base and index parts in x32 mode. OK for trunk?
>>>>
>>>> Thanks.
>>>>
>>>>
>>>> H.J.
>>>> ---
>>>> 2011-07-09 H.J. Lu <hongjiu...@intel.com>
>>>>
>>>> * config/i386/i386.c (ix86_simplify_base_index_disp): New.
>>>> (ix86_decompose_address): Support 32bit address in x32 mode.
>>>> (ix86_legitimate_address_p): Likewise.
>>>> (ix86_fixup_binary_operands): Likewise.
>>>
>>> Why don't you handle translations in TARGET_LEGITIMIZE_ADDRESS (or
>>> maybe also LEGITIMIZE_RELOAD_ADDRESS) ?
>>>
>>
>> It is because ix86_decompose_address is also called from:
>>
>> predicates.md: ok = ix86_decompose_address (op, &parts);
>> predicates.md: ok = ix86_decompose_address (op, &parts);
>> predicates.md: ok = ix86_decompose_address (XEXP (op, 0), &parts);
>> predicates.md: ok = ix86_decompose_address (XEXP (op, 0), &parts);
>> predicates.md: ok = ix86_decompose_address (XEXP (op, 0), &parts);
>
> Yes, but you should legitimize the address created by reload before it
> enters into predicates.
>
> So, the questions are:
>
> + (set (reg:SI 40 r11)
> + (plus:SI (plus:SI (mult:SI (reg:SI 1 dx)
> + (const_int 8))
> + (subreg:SI (plus:DI (reg/f:DI 7 sp)
> + (const_int CONST1)) 0))
> + (const_int CONST2)))
> +
> + We translate it into
> +
> + (set (reg:SI 40 r11)
> + (plus:SI (plus:SI (mult:SI (reg:SI 1 dx)
> + (const_int 8))
> + (reg/f:SI 7 sp))
> + (const_int [CONST1 + CONST2])))
>
> If the first form of the address is not OK (it does not represent the
> hardware operation), then it should not enter into the insn stream.
> This means, that it should be fixed ("legitimized") to second form by
> appropriate function (it looks that LEGITIMIZE_RELOAD_ADDRESS should
> fix it, since the incorrect address is generated by IRA/reload). After
> this operation, various predicates, based on ix86_decompose_address
> will start to work, since they will decompose valid memory addresses.
>
IRA/.RELOAD isn't prepared to deal with it and it just ICEs. I opened
a few GCC bugs on this.
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47744
is one of them. That is why I went this route.
--
H.J.
