[clang] [clang] constexpr atomic builtins (__c11_atomic_OP and __atomic_OP) (PR #98756)

[Aaron Ballman via cfe-commits]

Hana =?utf-8?q?Dusíková?= <hanicka at hanicka.net>,
Hana =?utf-8?q?Dusíková?= <hanicka at hanicka.net>
Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/98756 at github.com>


================
@@ -17893,4 +18005,425 @@ std::optional<bool> EvaluateBuiltinIsWithinLifetime(IntExprEvaluator &IEE,
   IsWithinLifetimeHandler handler{Info};
   return findSubobject(Info, E, CO, Val.getLValueDesignator(), handler);
 }
+
 } // namespace
+
+static bool EvaluateAtomicOrder(const AtomicExpr *E, EvalInfo &Info) {
+  // We need to evaluate Order argument(s), but we ignore it as constant
+  // evaluation is single threaded.
----------------
AaronBallman wrote:

Is this actually correct though? My reading of https://eel.is/c++draft/atomics keeps hitting preconditions that say "order is memory_order​::​relaxed, memory_order​::​consume, memory_order​::​ac-
quire, or memory_order​::​seq_cst", which means we should reject code like this, shouldn't we?
```
#include <atomic>

constexpr int foo(std::atomic<int> i) {
  return i.load(static_cast<std::memory_order>(1001));
}
```
and while we don't have to catch UB in the library functions, it seems more user-friendly to do so.

https://github.com/llvm/llvm-project/pull/98756