[clang] [CIR] Implement missing __sync_* atomic builtins (PR #214606)

[Cyrus Ding via cfe-commits]

dingcyrus wrote:

> Sorry for sitting on this, my "changes-requested" was the only thing left holding it.
> 
> On the alignment question you asked back: leave it as it is, no explicit alignment needed. `checkAtomicAlignment` is a faithful port of OG's, force-to-natural path included:
> 
> ```c++
>   // Force address to be at least naturally-aligned.
>   return ptr.withAlignment(CharUnits::fromQuantity(bytes));
> ```
> 
> so classic lands on natural alignment for these builtins whatever the pointer's declared alignment was. Setting it from `destAddr` would be the thing that could diverge, not the default. Your test pins it either way now, and since the third RUN line reuses the `LLVM` prefix over classic's output, both columns have to produce the same `align N`.
> 
> Two things in the description, which is what ends up in the log:
> 
>     * It still says `__sync_lock_test_and_set -> cir.atomic.xchg (acquire)`. The code emits `seq_cst` now, which is right, `EmitBinaryAtomic` uses seq_cst for it, but the summary did not follow the fix.
> 
>     * "covering 30 builtin variants" is 20 at this point, with the unsuffixed forms now `llvm_unreachable` and `_16` back on `errorBuiltinNYI`.

Thanks! Description updated — __sync_lock_test_and_set now reads seq_cst, and the summary now says 20 size-suffixed variants, with a note that the unsuffixed forms are rewritten by Sema and _16 falls back to errorBuiltinNYI.

https://github.com/llvm/llvm-project/pull/214606