Right now this is hot on stylist rebuild times because the Rust's hashmap entry API forces an unconditional clone of the key. Most of the atoms in stylists are dynamic. I wonder if we could make it public so addrefing/releasing an atom is a function call, not an FFI call that calls a virtual method.

Nathan, does it sound reasonable to make DynamicAtom public? We could also entirely devirtualize nsIAtom::AddRef/Release, I guess, though that's harder because it inherits from nsISupports. But marking it |final| the compiler should be able to figure it out, I guess. The basic idea would be something like what stylo does already: nsrefcnt nsIAtom::AddRef() { if (mIsStatic) { return 2; } return static_cast<DynamicAtom*>(this)->AddRef(); } Where DynamicAtom::AddRef is an inline function. Similarly for Release(). This would avoid us exposing DynamicAtom around.

Xidorn already dup'd this, but I am confused about "something like what stylo does already"...does stylo glue code try to devirtualize things on its own?

(In reply to Nathan Froyd [:froydnj] from comment #3) > Xidorn already dup'd this, but I am confused about "something like what > stylo does already"...does stylo glue code try to devirtualize things on its > own? To some extent. We already avoid going through the FFI boundary if mIsStatic is non-zero.