Re: [RFC, PATCH] shrink struct vectorlike_header #2

[Stefan Monnier]

>> I doubt (*(struct Lisp_Vector **)((char *) v + header_size))
>> is any better.  I guess it just defeats gcc's detection of the problem.
> Casting through char * is better, because the C standard
> says that a compiler cannot do type-based alias inferencing
> in the presence of char * pointers.

That's so backwards: rather than force people to use low-level fiddly
code, why can't they say "oh wait, you're doing some funny cast, let's
be more conservative with type-based aliasing, as if the code used
char*".

> I presume this is why GCC generates all those warnings when we don't
> use char * -- GCC is warning us that it may be doing optimizations
> that will crash our code.

IOW it's telling us "beware that even though I have here evidence that
type-based alias analysis is unsafe for your code, I'll go ahead and use
it anyway even though I know how not to use it since I have to do that
when you use char* casts!".  Nice.

> struct Lisp_Vector
>   {
>      struct vectorlike_header header;
>      union
>        {
>           Lisp_Object contents[1];
>           struct Lisp_Vector *next;
>        } u;
>   };

Yes, that'd be better.


        Stefan