Eric Botcazou <[email protected]> writes:
>> In http://gcc.gnu.org/ml/gcc-patches/2012-11/msg00257.html I said:
>>
>> get_best_mode has various checks to decide what counts as an acceptable
>> bitfield mode. It actually has two copies of them, with slightly
>> different alignment checks:
>>
>> MIN (unit, BIGGEST_ALIGNMENT) > align
>>
>> vs.
>>
>> unit <= MIN (align, BIGGEST_ALIGNMENT)
>>
>> The second looks more correct, since we can't necessarily guarantee
>> larger alignments than BIGGEST_ALIGNMENT in all cases.
>>
>> PR 55438 shows why I was wrong. BIGGEST_ALIGNMENT is not (as I thought)
>> the biggest supported alignment, but the:
>>
>> Biggest alignment that any data type can require on this machine, in
>> bits. Note that this is not the biggest alignment that is supported,
>> just the biggest alignment that, when violated, may cause a fault.
>>
>> So it's perfectly possible for BIGGEST_ALIGNMENT to be 8 on 32-bit machines.
>
> That means that the Sequent check was flawed, doesn't it? It also seems that
> the entire business of alignment with size comparisons is dubious.
Yeah, I suppose the Sequent test didn't really have any effect on
BIGGEST_ALIGNMENT == BITS_PER_UNIT targets.
>> Also, in cases like these, the supposedly conservative:
>>
>> && GET_MODE_BITSIZE (mode) <= align
>>
>> doesn't preserve the cap in the original:
>>
>> MIN (unit, BIGGEST_ALIGNMENT) > align
>>
>> Fixed by using GET_MODE_ALIGNMENT instead.
>
> Note that the comment just above needs to be adjusted then.
OK, how about:
/* ??? For historical reasons, reject modes that would normally
receive greater alignment, even if unaligned accesses are
acceptable. This has both advantages and disadvantages.
> What about the similar check in next_mode?
>
> /* Stop if the mode requires too much alignment. */
> if (unit > align_ && SLOW_UNALIGNED_ACCESS (mode_, align_))
> break;
>
> It seems to me that we should change it as well.
I think here we really do want unit (i.e. the GET_MODE_BITSIZE).
We're dividing the bitfield into unit-sized chunks and want to know
whether those chunks are aligned or not. If they aren't aligned,
we need to know whether unaligned accesses are OK.
Richard
