Do not remove boxing for struct methods which my expose references to their fields #121396
Conversation
github-actions
bot
added
the
needs-area-label
EgorBo
added
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
|
|
||
| result = (CORINFO_METHOD_HANDLE)pUnboxedMD; | ||
| requiresInstMTArg = !!pUnboxedMD->RequiresInstMethodTableArg(); | ||
| // Give up on methods which may potentially expose byrefs to struct's fields: |
There was a problem hiding this comment.
This makes assumptions about the kind of optimizations that the JIT may perform. Should this check be done by the JIT, or be explicitly requested in the JIT/EE API?
| // Give up on methods which may potentially expose byrefs to struct's fields: | ||
| // * byref return types | ||
| // * byref-like return types | ||
| // * byref to byref-like types in parameters |
There was a problem hiding this comment.
What about a native pointer to a byref like type? Or worse, a void pointer cast in the body to a pointer to a byref like type.
I'm curious what exposed the problem here. It reminds me of the "unboxing of nullables" case a little bit where we don't produce a heap pointer. I wonder if we have to spec ourselves into being allowed to do this optimization.
(More robust would probably be to remove the box when we can prove that it doesn't escape.)
3 participants
No description provided.