Re: [patch] Fix ICE on unaligned record field

[Richard Biener <richard.guenther@gmail.com>]

On Thu, Dec 11, 2014 at 10:52 PM, Eric Botcazou <ebotcazou@adacore.com> wrote:
>> Note that I think the place of the check is unfortunate as you for example
>> will not remove the argument if it is unused.  In fact I'm not yet sure
>> what transform exactly we are disabling.  I am guessing we are
>> passing an aggregate by value that resides at a bit-aligned offset
>> of some outer object:
>>
>>   foo (x.aggr);
>>
>> and the function then does
>>
>> foo (Aggr a)
>> {
>>   int i = a.foo;
>> ...
>> }
>>
>> thus use only a part of the aggregate.  Then IPA SRA would like to
>> pass x.aggr.foo instead of x.aggr and thus tries to materialize a
>> load from x.aggr.foo at all callers but fails to do that in a valid way.
>
> Right, it's the usual MEM_EXPR business creating ADDR_EXPRs out of nowhere and
> miserably failing on something not addressable.

Well, I call it a convenience that MEM_EXPR, unlike INDIRECT_REF, can
be used to encapsulate an arbitrary byte-offset and view-conversion.  Of course
it's still a dereference of an address so that convenience doesn't work on sth
non-addressable.

>> Erics fix did, at all callers
>>
>>   Aggr tem = x.aggr;
>>   foo (tem.foo);
>>
>> ?
>
> Yes, because the code wants to take &tem afterwards.
>
>> While we should be able to simply do
>>
>>   foo (BIT_FIELD_REF <x.aggr, .....>)
>>
>> with the appropriate bit offset and size?  (if that's of register type
>> you need to do the load in a separate stmt of couse).
>>
>> Thus similar to Erics fix but avoiding the aggregate copy.
>
> Yes, that should be doable, but I'm not sure it's worth the hassle.

I'll leave that to you two to decide - Martins patch is ok if you are fine
with disabling the optimization (also removing an unused parameter).

Thanks,
Richard.

> --
> Eric Botcazou