On Tue, Jul 19, 2011 at 7:33 PM, Uros Bizjak <ubiz...@gmail.com> wrote:
>>>> Sometimes, the compiler is really creative in inventing instructions:
>>>>
>>>> (insn 47 46 49 7 (set (reg:SI 68 [ D.1686 ])
>>>> (subreg:SI (plus:SF (reg:SF 159 [ D.1685 ])
>>>> (reg:SF 159 [ D.1685 ])) 0)) omp_atomic1.f90:17 247
>>>> {*lea_2}
>>>> (expr_list:REG_DEAD (reg:SF 159 [ D.1685 ])
>>>> (nil)))
>>>>
>>>> Really funny.
>>>
>>> That's the job of combiner to try all kinds of stuff and it is the
>>> responsibility of the backend to reject those. I think it would be better
>>> to get back to testing Pmode in the legitimate address hook, perhaps
>>> allowing ptr_mode too in addition to Pmode (which for -m32/-m64 won't mean
>>> any change, just for -mx32).
>>
>> Actually, there is a bypass in ix86_decompose_address, and this RTX
>> squeezed through. IMO constructs like this should be rejected in
>> i_d_a, which effectively only moves Pmode/ptr_mode check here.
>>
>> I'm looking into it.
>
> The problem was in fact the declaration of no_seg_address_operand
> predicate that was defined as special predicate and this way ignoring
> the mode of the operand.
This change should be backported to 4.6 and 4.5.
Uros.
