[RyuJIT] inline CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPE

[EgorBo]

Currently, for e.g. typeof(int) we emit something like this:

       mov      rcx, 0x7FF96C7D9BA0
       call     CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPE

Can we just emit a raw pointer to m_hLoaderAllocatorObjectHandle with a jit-time known offset based on the typehandle and get rid of that call? (for non-collectible/non-shared types).

In my current jit-diff dump (PMI) I see 5898 call CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPEHANDLE

/cc @jkotas

[MichalStrehovsky]

If the RuntimeType hasn't been allocated yet, the JIT would still have to call the helper so that it gets created. Are you hoping to save size or improve speed?

[jkotas]

This was discussed in #5973 (comment).

We may want to look into changing string literals and RuntimeTypes to be allocated on Pinned Object heap (for non-collectible types at least) and emitting their addresses directly, without any indirections. The Pinned Object Heap is sort of similar to frozen objects in this regard.

[EgorBo]

Are you hoping to save size or improve speed?

both! 🙂

We may want to look into changing string literals and RuntimeTypes to be allocated on Pinned Object heap (for non-collectible types at least) and emitting their addresses directly, without any indirections. The Pinned Object Heap is sort of similar to frozen objects in this regard.

Ah nice! Will try to prototype both (type and literals) this weekend to estimate impact.