[PATCH] Fix PR middle-end/26306

[Richard Kenner]

> So, what's that got to with aggegates that are, say, SImode?  

There's no problem with SImode aggregates because we can do the applicable
load directly.

> And, why is there no well-defined meaning?  (The obvious thing would be
> to all bytes that make up the aggregate.)

I meant no well-defined meaning in GIMPLE for such a naked decl and I
don't think the RTL generators do anything about it.

> I get a memory read for "i", why wouldn't I expect that for:
> 
>    volatile struct S { int a; } i;

I think you indeed might, but I'm not sure if we ever emitted such if struct
S was BLKmode.  I'm not sure it's worth spending much time making that do the
right thing, but it is worth making sure we don't ICE.