[Bug fortran/45586] [4.6 Regression] ICE non-trivial conversion at assignment

["mikael at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45586

--- Comment #27 from Mikael Morin <mikael at gcc dot gnu.org> 2011-01-17 13:01:23 UTC ---
(In reply to comment #25)
> Maybe it would be better to set the "inherited" pointer and target
> attributes much earlier, in gfc_variable_attr.  With a bit of luck,
> things would then work automatically.
If you mean set the gfc_component.attr.target field, I don't think it's right. 
As gfc_component structs belong to the type's gfc_symbol, they are shared
between non-pointer and pointer variants. 

(In reply to comment #26)
> (In reply to comment #24)
> > Or maybe drop restrict all together on the type for correctness.
> 
> Indeed when we introduced the restrict handling to the Fortran frontend
> we thought that the above case would not happen, thus that the testcase
> in question would be not a valid Fortran program.  At the moment
> we probably even could generate a testcase that generates wrong-code
> (when the ICE silences itself with release-checking).  This means that
> simply trying to make it not ICE is probably wrong and simply hides
> the real issue (which the ICE shows).
> 
> It is, btw, a sign of bad Fortran language design and makes the point
> of the pointer/target attributes moot.
I think it makes some sense:
If there is more than one data path to a struct, there is more than one data
path to all of its components. Thus the inherited target attribute.
It's just it doesn't fit the C-minded middle-end well. :-(

> 
> What we could do is drop restrict qualifications for all aggregate
> types.  At least such funny games do not seem possible with mere
> toplevel arrays or scalars, right?
Well, i think so. 


> 
> > By the way, does the middle-end support creating a variant of a type with a
> > sub-component restrict (un)qualified ?
> 
> No.  For type variants all the fields are usually shared, you'd have to
> create a deep copy.

Which means they are not type-compatible anymore ?
Which means casts everywhere ?