[clang] [Clang] Add attribute for consteval builtins; Declare constexpr builtins as constexpr in C++ (PR #91894)

[Aaron Ballman via cfe-commits]

================
@@ -14,13 +14,18 @@ void __builtin_va_copy(double d);
 // expected-error at +2 {{cannot redeclare builtin function '__builtin_va_end'}}
 // expected-note at +1 {{'__builtin_va_end' is a builtin with type}}
 void __builtin_va_end(__builtin_va_list);
-// RUN: %clang_cc1 %s -fsyntax-only -verify 
-// RUN: %clang_cc1 %s -fsyntax-only -verify -x c
 
 void __va_start(__builtin_va_list*, ...);
 
+          void *__builtin_assume_aligned(const void *, size_t, ...);
 #ifdef __cplusplus
-void *__builtin_assume_aligned(const void *, size_t, ...) noexcept;
-#else
-void *__builtin_assume_aligned(const void *, size_t, ...);
+constexpr void *__builtin_assume_aligned(const void *, size_t, ...);
+          void *__builtin_assume_aligned(const void *, size_t, ...) noexcept;
+constexpr void *__builtin_assume_aligned(const void *, size_t, ...) noexcept;
+          void *__builtin_assume_aligned(const void *, size_t, ...) throw();
+constexpr void *__builtin_assume_aligned(const void *, size_t, ...) throw();
+
----------------
AaronBallman wrote:

Does anyone know why we allow builtins to be redeclared at all? CC @zygoloid @rjmccall 

I wonder just how much code we would break by not aiming for backwards compatibility here and requiring users to update their declarations on the assumption that declaring these at all is undefined behavior (using a reserved name) and so it's up to the user to match the compiler's behavior rather than make the compiler handle this with weird non-conforming language extension behavior.

I'm not looking to punish users in this case, but the workaround to keep things backwards compatible is pretty magical and awkward.

https://github.com/llvm/llvm-project/pull/91894