Replace non-performance-critical opcodes by call to methods of an internal object.

[iritkatriel]

There are a number of opcodes which are not performance critical and which seem to exist for the sole purpose of invoking a function or loading a non-picklable constant to the stack (which could be returned from a function). Examples:

LOAD_ASSERTION_ERROR, IMPORT_NAME, IMPORT_FROM, PRINT_EXPR, PREP_RERAISE_STAR.

Perhaps we could create a simple (picklable) object that has these functions as its methods, so that the compiler can load this object to the stack and invoke the methods via a series of generic opcodes. This object should ideally not be accessible from python, it should be stateless, it certainly should not be subclassable and the methods should be immutable.

This is an alternative to an idea that came up previously, to load AssertionError via a new LOAD_COMMON_CONST opcode that fetches it from an internal consts array. That approach could be good for performance critical consts, but for the slow ones we don’t actually need an opcode at all.

[ericsnowcurrently]

Seems worth exploring.

[markshannon]

Other candidates for lowering are:

LOAD_BUILD_CLASS can be replaced with LOAD_COMMON_CONST build_class
RAISE_VARARGS n can be replaced with LOAD_COMMON_CONST raise_varargs; PRECALL_FUNCTION; CALL n
BEFORE_ASYNC_WITH and BEFORE_WITH can be replaced with the lower level operations they are composed of. These will need a new LOAD_SPECIAL instruction.

BEFORE_WITH becomes

     COPY 1
     LOAD_SPECIAL "__exit__"
     SWAP 2
     LOAD_SPECIAL "__enter__"
     PRECALL_FUNCTION 
     CALL 0

[markshannon]

If we add a LOAD_LOCAL instruction (which would do the same as LOAD_NAME without falling back to the globals), then we can lower a couple more instructions, as try-except is now cheap.

LOAD_CLASSDEREF name becomes

try:
     LOAD_LOCAL name
except KeyError:
     LOAD_DEREF name

SETUP_ANNOTATION becomes

try:
     LOAD_LOCAL __annotations__
except:
    BUILD_MAP
    STORE_NAME __annotations__

[gvanrossum]

How cheap is raising an exception? 99.9% of LOAD_CLASSDEREF will raise NameError, which means a NameError object is created and immediately discarded.

[markshannon]

Raising an exception is expensive. Relative to the simple lookup, very expensive. So let's not do that.
A LBYL approach would work better; adding HAS_LOCAL instead of LOAD_LOCAL.

LOAD_CLASSDEREF name then becomes:

    HAS_LOCAL name
    POP_JUMP_IF_FALSE use_deref
    LOAD_LOCAL name
    JUMP exit
use_deref:
    LOAD_DEREF name
exit:

and SETUP_ANNOTATION becomes:

    HAS_LOCAL __annotations__
    POP_JUMP_IF_TRUE done
    BUILD_MAP
    STORE_NAME __annotations__
done:

[markshannon]

python/cpython#99005