Re: [patch] for PR 18040

[kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)]

    ((cast)var).field is not valid gimple.

    Casts must be in a statement by themselves, unless the argument is
    is_gimple_min_invariant.

The issue here is historical.  The point is to view a single nested component
reference as valid so we don't try to split it up since it can normally be
done in a single operation and spliting it up will make it more expensive.
The idea is that handled_component_p returns true for exactly those things
that will be stripped by get_inner_reference.

So we wanted to support conversions *inside* a nest of references.  In other
words, ((cast) var.field1).field2.  I thought that was specified in the
grammer; if not, it should be.

Your example above is simply a degenerate form of that and certainly needs to
be supported because otherwise we pessimize code quite badly (at best).
Suppose that cast is to a variable-size type.  Then we can't even gimplify
that expression because we can't have a variable-size temporary.  Such an
expression is certainly valid GENERIC, so saying we can't gimplify it is bad.
But if we say that it's the responsibility of the front-end to generate that
temporary, we now have horrible code: we might be copying a 1MB structure
into a temporary just to retreive one byte!

However, a side-effect of allowing this is that we allow the conversion as
the *outer* operation.  That hasn't be quite correct for a while, but
apparently never caused problems before.

So my suggestion is to modify handled_component_p to exclude a conversion at
the outer level but to have a recursive version of it that *does* allow the
conversion inside fields.  But I don't know if that addresses your issue.