uweigand wrote:
Hmm, interesting. Not sure this difference is observable under normal
circumstances, but if the z/OS builtins are guaranteed to perform the store
unconditionally, I agree this should be followed.
That said, I guess it would still in principle be possible to achieve the same
effect, using the old-style GCC `__sync_val_compare_and_swap` builtin:
```
unsigned int old, mem, newval;
int test_atomic() {
unsigned int oldval = old;
unsigned int memval = __sync_val_compare_and_swap(&mem, oldval, newval);
old = memval;
return memval != oldval;
}
```
However, with the current back-end it seems you'll be getting slightly less
efficient assembler; this could likely be fixed in the back-end, but you'd
still have to rely on optimization being enabled.
In the end, I guess having this as separate builtins is fine with me.
https://github.com/llvm/llvm-project/pull/202362
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
