The RETURN_VALUE and RETURN_CONST instructions return the value on top of the stack and a constant, respectively.
RETURN_CONST is effectively a superinstruction: LOAD_CONST + RETURN_VALUE.
This worked well up to 3.13, but is likely to be an impediment to further optimizations:
- The JIT already breaks
RETURN_CONST into LOAD_CONST and RETURN_VALUE so the superinstruction has no value.
RETURN_VALUE has two different paths internally, one for normal functions and the other for generators. We know statically which sort of function we are in, so we should emit different bytecodes for the different behavior
- With the proposed changes to reference counting, it will help in the interpreter (we already do this in the JIT) to have a different instruction for immortal constants. (This might only be temporarily, as we expect to be able to defer reference counting on all constants)
Changes to the instruction set:
RETURN_CONST and INSTRUMENTED_RETURN_CONST will be removed.
RETURN_VALUE_GEN, INSTRUMENTED_RETURN_GEN and LOAD_CONST_IMMORTAL will be added.
We expect that LOAD_CONST_IMMORTAL will be removed again before 3.14, but that is not a requirement.
Linked PRs
The
RETURN_VALUEandRETURN_CONSTinstructions return the value on top of the stack and a constant, respectively.RETURN_CONSTis effectively a superinstruction:LOAD_CONST + RETURN_VALUE.This worked well up to 3.13, but is likely to be an impediment to further optimizations:
RETURN_CONSTintoLOAD_CONSTandRETURN_VALUEso the superinstruction has no value.RETURN_VALUEhas two different paths internally, one for normal functions and the other for generators. We know statically which sort of function we are in, so we should emit different bytecodes for the different behaviorChanges to the instruction set:
RETURN_CONSTandINSTRUMENTED_RETURN_CONSTwill be removed.RETURN_VALUE_GEN,INSTRUMENTED_RETURN_GENandLOAD_CONST_IMMORTALwill be added.We expect that
LOAD_CONST_IMMORTALwill be removed again before 3.14, but that is not a requirement.Linked PRs
LOAD_CONSTinto three. #125972