[Bug c++/95726] ICE with aarch64 __Float32x4_t as template argument

[jason at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95726

--- Comment #6 from Jason Merrill <jason at gcc dot gnu.org> ---
(In reply to rsandifo@gcc.gnu.org from comment #4)
> (In reply to Jakub Jelinek from comment #3)
> > But if they mangle differently, then structural_comptypes shouldn't treat
> > them as same types.

Definitely.

> That certainly avoids the ICE, and makes GCC's behaviour consistent
> with Clang for things like:
> 
>   typedef float vecf __attribute__((vector_size(16)));
>   vecf x;
>   float32x4_t &y = x;
> 
> Previously we accepted this, with the struct_comptypes change
> we reject it (like Clang does).  But that might break existing
> code, so I'm not sure it would be backportable.

If necessary we could add a conversion between the pointer-to-vector types.

> I guess the question then is: what does TYPE_STRUCTURAL_EQUALITY_P
> mean for VECTOR_TYPEs in the context of structural_comptypes?

The same thing it means for any other type: setting TYPE_CANONICAL properly is
too hard, so use structural_comptypes.

> And (if this is a different question) what case is that function's
> VECTOR_TYPE handling for?  I.e. when do we want to return true for
> a pair of VECTOR_TYPEs whose TYPE_MAIN_VARIANTs are different?

We want to return true if they should be considered the same type.

Generally, TYPE_MAIN_VARIANT isn't sufficient for checking type identity, as it
only looks through variants of the outermost type:  If I have

typedef int myint;
typedef myint* myintptr;

taking TYPE_MAIN_VARIANT of myintptr gives myint*, not int*.  That's why
TYPE_CANONICAL was introduced, so we didn't need to do structural comparison
whenever we wanted to compare types.