TGMPFunction

TGMPFunction

The type-erased callable the dispatch core stores. Deliberately not a vtable.

Declared in Plugins/GMP/Source/GMP/GMP/GMPFunction.h

#define GMP_FUNCTION_PREDEFINED_ALIGN_SIZE 16

static constexpr int32_t kInlineSize = GMP_FUNCTION_PREDEFINED_INLINE_SIZE;
static constexpr int32_t kAlignSize  = GMP_FUNCTION_PREDEFINED_ALIGN_SIZE;

Invocation (GMPFunction.h:568):

R operator()(TArgs... Args) const
{
    CheckCallable();
    return reinterpret_cast<R (*)(void*, TArgs...)>(GetCallable())(GetObjectAddress(), Args...);
}

The representation

Two things: a thunk function pointer and the callable's own address. That is the whole erasure — there is no virtual base, no vtable pointer, and no dynamic dispatch.

The thunk is generated per concrete callable type and simply casts the void* back:

// Private/GMPFlexBackend.h:39
static void FlexThunk(void* Self, TArgs... Args) { (*static_cast<Func*>(Self))(static_cast<TArgs>(Args)...); }

Why it matters for cost

Virtual call	TGMPFunction
load vptr → load slot → call	pointer already in hand → one indirect call

And because the cast-and-call sits in tail position, -O2 turns it into a sibling call: a plain jmp, no new stack frame.

Small buffer optimisation

A callable up to kInlineSize bytes lives inside the storage, aligned to GMP_FUNCTION_PREDEFINED_ALIGN_SIZE (16). No heap allocation, and the callable is contiguous with the dispatch data.

The practical rule: keep listener lambdas small. [this] and one or two pointers stay inline. Capturing a large struct by value spills to the heap and costs an allocation per registration.

See also

FGMPRawSig · FGMPTypedAddr · Key baking · Build switches