[clang] [Clang] Generate the GEP instead of adding AST nodes (PR #73730)

[Kees Cook via cfe-commits]

kees wrote:

> ```
> int foo(struct s *p, int index) {
>   return __builtin_dynamic_object_size((++p)->array[index], 1);
> }
> ```
> 
> This _shouldn't_ increment `p`, but we need to get the array size of the element _after_ `p`. I suspect that this is probably a horrible security violation in the making, but we at least need to handle such an eventuality gracefully. For a first pass, I think returning `-1` or `0` (depending on the default return value) for _any_ pointer arithmetic is probably okay...maybe even the best option?
> 
> @kees Thoughts?

Owch. That really shouldn't be legal: we can't have arrays of structs ending in a FAM, so the `++p` isn't sane. That said, I guess pointer arithmetic must follow `sizeof`, so `++p` isn't illegal; it's just awful. FWIW, GCC just returns `SIZE_MAX` for this sort of "aaah, where are you going?!" insanity for non-FAM structs. But Clang appears to just go along for the ride:
https://godbolt.org/z/Pss1oz7sW

So that needs to be fixed in Clang too.

But for doing a FAM struct like this ... my instinct is to say it should return 0, much like requesting an out of bounds array index. (Rather than `SIZE_MAX`.)

https://github.com/llvm/llvm-project/pull/73730