Re: [RFC][PATCH] rcu: Use typeof(p) instead of typeof(*p) *

On Tue, 5 Oct 2021 15:49:58 -0400 (EDT)
Mathieu Desnoyers [off-list ref] wrote:

----- On Oct 5, 2021, at 3:40 PM, rostedt rostedt@goodmis.org wrote:

quoted

On Tue, 5 Oct 2021 21:06:36 +0200 (CEST)
Jan Engelhardt [off-list ref] wrote:
  

quoted

On Tuesday 2021-10-05 20:40, Steven Rostedt wrote:  

quoted

quoted

quoted

quoted

quoted

quoted

typeof(*p) *________p1 = (typeof(*p) *__force)READ_ONCE(p);  

#define static_cast(type, expr) ((struct { type x; }){(expr)}.x)
typeof(p) p1 = (typeof(p) __force)static_cast(void *, READ_ONCE(p));

Let the name not fool you; it's absolutely _not_ the same as C++'s
static_cast, but still: it does emit a warning when you do pass an
integer, which is better than no warning at all in that case.

 *flies away*  

Are you suggesting I should continue this exercise ;-)  

“After all, why not?”

typeof(p) p1 = (typeof(p) __force)READ_ONCE(p) +
               BUILD_BUG_ON_EXPR(__builtin_classify_type(p) != 5);  

I may try it, because exposing the structure I want to hide, is pulling out
a lot of other crap with it :-p  

I like the static_cast() approach above. It is neat way to validate that the
argument is a pointer without need to dereference the pointer.

I would also be open to consider this trick for liburcu's userspace API.

About the other proposed solution based on __builtin_classify_type, I am
reluctant to use something designed specifically for varargs in a context
where they are not used.

Unfortunately, it doesn't solve the Debian gcc 10 compiler failing when
passing the function name instead of a pointer to the function in
RCU_INIT_POINTER()  That alone makes me feel like I shouldn't touch that
macro :-(

And who knows what other version of gcc will fail on passing the address :-p

-- Steve