Re: patch: Re: rfc: auto-casted vector types

[Aldy Hernandez <aldyh at redhat dot com>]

Yes, indeed I was confusing intrinsics with builtins.  The builtins are 
V2SI, the intrinsics are opaque types.

> Changing the intrinsic definitions in spe.h to cast the result to the 
> right
> type does sound like an easy solution to the problem.  The ABI defines 
> the
> behaviour of the intrinsics, but does not define the behaviour of the 
> gcc
> builtins, so all that really matters is that we get the intrinsics 
> right.

Easy is good.

> Graham is suggesting that we change the builtin function expanders in 
> rs6000.c
> to use the correct type.  Maybe we can change the builtin function 
> types
> without having to change the RTL modes.  This would mean using an 
> opaque
> vector type instead of V2SI_type_node in the rs6000.c file.  This 
> should have
> the same effect as adding casts to the intrinsics.  This means defining
> the opaque vector type as a builtin type instead of in spe.h.

Now I see.  Yes, I looked at this approach as well, but this one is 
more error prone.  V2SI_type_node is all over the place.  We'd add a 
new type, change the builtins accordingly, and then change 
is_ev64_opaque_type to recognize this new type as well.  What I am 
trying to avoid with my approach is changing the back end at all, and 
just change the user header file.  I also don't want to touch too much 
what ain't broke (in the back end).

One place where Graham's approach would be superior would be for the 
load intrinsics which currently cast to (void *).  With Graham's 
approach we could remove the cast, and then accept only vector types 
that are compatible with the opaque type, not just *any* pointer like 
it is now.  But I'm willing to live with that.

Does this sound like a reasonable approach Jim?

Aldy