Re: [Tinycc-devel] [PATCH] stdatomic: ld/st/xchg/cmpxchg on simple types

[Michael Matz]

Hello,

On Mon, 22 Mar 2021, Dmitry Selyutin wrote:

> > Is the above atomic_load supposed to be similar to the GCC __atomic_load
> > intrinsic?
> No, it's supposed to be similar to __atomic_load_N.

I see.  Those return integer types and hence aren't convertible to struct 
types either (not even with a cast).  This doesn't (and shouldn't) 
compile:

% cat x.c
struct S { short a, b; };
struct S load(struct S *mem)
{
  struct S s;
  s = __atomic_load_4(mem, __ATOMIC_SEQ_CST);
  return s;
}
% gcc -c x.c
x.c: In function ‘load’:
x.c:5:7: error: incompatible types when assigning to type ‘struct S’ from 
type ‘unsigned int’

> > hence structs aren't passed around
> It depends on the calling convention.

No, it depends on which types the atomic builtins are defined to support. 
I'm aware that functions that do happen to pass around (or return) structs 
by value can use registers for this.  But the atomic functions don't do 
that.  Are you perhaps talking about the case where the compiler rewrites 
a generic __atomic_load into a size-specific one, ala:

  __atomic_load(mem, &s, __ATOMIC_SEQ_CST);

into

  <view_convert(struct S)>__atomic_load_4(mem, __ATOMIC_SEQ_CST);

So, are you wanting to do the same for TCC as well?  The easier way would 
rather be to simply not do such rewriting and just keep calling the 
generic __atomic_load.

So, again, please provide some example code that shows what you want to be 
compilable (and perhaps also in which way).

> For the __atomic_##_N, it seems like we can pass things this way, as
> long as pointers are naturally-aligned.
> As I mentioned in one of the earlier patches, there's no proper align
> check yet, for now we have size check.


Ciao,
Michael.