Re: [PATCH 5/6] workqueue: introduce NR_WORKER_POOLS and for_each_worker_pool()

[Peter Seebach <peter.seebach@windriver.com>]

On Fri, 13 Jul 2012 22:00:10 -0700
Linus Torvalds <torvalds@linux-foundation.org> wrote:
> (*) Technically, "&(x)[0]" is actually a really confused way of saying
> "(x+0)" while making sure that "x" was a valid pointer.

But wait, there's more!

Should someone some day try to use an implementation with a fairly
ferocious bounds-checker, the bounds of &x[0] are the bounds of the
first member of x, while the bounds of x are... well, whatever they
were. (If x is an array, they're definitely the bounds of the whole
array. If x is a pointer to something, then it depends on how the
pointer was obtained.)

I'm not sure anyone actually has an implementation that bothers with
this level of granularity in pointers, but I am about 90% sure that an
implementation which did would be conforming.  e.g.:

  int a[2];
  a[1] = 3; /* ok */
  int *b = a;
  b[1] = 3; /* ok */
  int *c = &a[0];
  c[1] = 3; /* bounds violation */

Note that "conforming" does not imply "could compile and run most
existing code without surprising new errors". The world is full of code
which assumes absolute identity between (a+i) and &(*(a+i)).

If the code which inspired your rant was actually doing it on purpose
to obtain this result, I shall have to buy a hat so I can eat it.
(Disclaimer: Hat must be made of something delicious.)

-s
-- 
Listen, get this.  Nobody with a good compiler needs to be justified.