__c11_atomic_add_fetch is not implemented

[pakluba]

Could you please implement the missing C11 variants of the atomic builtins?

I was using stdatomic.h in my code, sometimes directly with the __atomic_* builtins, but got a quite confusing compilation error:

address argument to atomic operation must be a pointer to integer, pointer or supported floating point type ('atomic_uint *' (aka '_Atomic(unsigned int) *') invalid

After doing some search it turned out that instead of the __atomic_* builtins the __c11_atomic_* builtins shall be used. It fixed most of my issues except where a __c11_* variant is not available.

While those are not available, it it OK to make the compilation error go away by casting the atomic pointer to ordinary pointer?

__atomic_add_fetch((unsigned int *)(ptr), val, memorder)

[frederick-vs-ja]

We've implemented __c11_atomic_fetch_add, not __c11_atomic_add_fetch. Would it be suitable to use __c11_atomic_fetch_add instead?

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (pakluba)

Details

Could you please implement the missing C11 variants of the atomic builtins?

I was using stdatomic.h in my code, sometimes directly with the __atomic_* builtins, but got a quite confusing compilation error:

address argument to atomic operation must be a pointer to integer, pointer or supported floating point type ('atomic_uint *' (aka '_Atomic(unsigned int) *') invalid

After doing some search it turned out that instead of the __atomic_* builtins the __c11_atomic_* builtins shall be used. It fixed most of my issues except where a __c11_* variant is not available.

While those are not available, it it OK to make the compilation error go away by casting the atomic pointer to ordinary pointer?

__atomic_add_fetch((unsigned int *)(ptr), val, memorder)

[pakluba]

@frederick-vs-ja Do you mean the atomic_fetch_add(&x, y) + y hack? I think that's quite ugly.

[shafik]

CC @AaronBallman

[AaronBallman]

Clang supports __atomic_fetch_add and __atomic_add_fetch; the C11 atomic intrinsics exist only to support stdatomic.h, and C11 has no semantics for __atomic_add_fetch.

Given that Clang supports __atomic_add_fetch already, what need is there for __c11_atomic_add_fetch?

[zygoloid]

Given that Clang supports __atomic_add_fetch already, what need is there for __c11_atomic_add_fetch?

They don't have the same semantics. The GNU atomic_add operations on pointers add the integer to the address stored by the pointer directly, whereas the C11 semantics are to add sizeof(T) multiplied by the integer. Example: note that f4 adds 4, and the other three add 1.

[pakluba]

How about because I'm using stdatomic in the rest of my code. And __atomic_add_fetch gives a BUILD ERROR if used with _Atomic.

[AaronBallman]

Given that Clang supports __atomic_add_fetch already, what need is there for __c11_atomic_add_fetch?

They don't have the same semantics. The GNU atomic_add operations on pointers add the integer to the address stored by the pointer directly, whereas the C11 semantics are to add sizeof(T) multiplied by the integer. Example: note that f4 adds 4, and the other three add 1.

Ah, today I learned. :-)

Okay, so this isn't really a bug (we never intended to support the interface), but it is a reasonable feature request (there's some utility to supporting it). I've adjusted the issue accordingly.

[zygoloid]

@frederick-vs-ja Do you mean the atomic_fetch_add(&x, y) + y hack? I think that's quite ugly.

So define a macro to hide this ugliness away somewhere. This seems like a normal programming problem. Internally we convert add_fetch into fetch_add plus an add anyway, so this wouldn't be losing you anything.

Fundamentally, the __c11_* builtins are specifically for implementing exactly the interface in C11's <stdatomic.h> header, which includes atomic_fetch_add and not atomic_add_fetch, so adding another __c11_* function doesn't really match the design philosophy here, especially given that you can produce the functionality yourself by using the provided builtins.