[libc] Add and use 'cpp::launder' to guard placement new (#146123)

Summary:
In the GPU allocator we reinterpret cast from a void pointer. We know
that an actual object was constructed there according to the C++ object
model, but to make it fully standards compliant we need to 'launder' it
to forward that information to the compiler. Add this function and call
it as appropriate.

Author: jhuber6

libc/src/__support/CPP/new.h

@@ -29,6 +29,14 @@ enum class align_val_t : size_t {};
 
 namespace LIBC_NAMESPACE_DECL {
 
+namespace cpp {
+template <class T> [[nodiscard]] constexpr T *launder(T *p) {
+  static_assert(__has_builtin(__builtin_launder),
+                "cpp::launder requires __builtin_launder");
+  return __builtin_launder(p);
+}
+} // namespace cpp
+
 class AllocChecker {
   bool success = false;
 

libc/src/__support/GPU/allocator.cpp

@@ -544,8 +544,8 @@ void deallocate(void *ptr) {
     return impl::rpc_free(ptr);
 
   // The original slab pointer is the 2MiB boundary using the given pointer.
-  Slab *slab = reinterpret_cast<Slab *>(
-      (reinterpret_cast<uintptr_t>(ptr) & ~SLAB_ALIGNMENT));
+  Slab *slab = cpp::launder(reinterpret_cast<Slab *>(
+      (reinterpret_cast<uintptr_t>(ptr) & ~SLAB_ALIGNMENT)));
   slab->deallocate(ptr);
   release_slab(slab);
 }