Calling comprehensions always makes CALL_PY_EXACT_ARGS miss

まず、MAKE_FUNCTION によって作成された関数に対する CALL_PY_EXACT_ARGS の特殊化を追跡し、issue で説明されている DEOPT_IF func_version チェックに焦点を当てます。disassembly の例を reproducer として使用し、pyperformance の結果を比較します。これらの内包表記の呼び出しが静的に回避されるか、不要に miss しなくなり、報告されている benchmark に回帰がないことをもって完了とします。

Consider the function

def f():
    return [x for x in range(2)]
for i in range(20):
    f() # warmup
from dis import dis
dis(f, adaptive=True)

dis(f)
  1           0 RESUME_QUICK             0

  2           2 LOAD_CONST               1 (<code object <listcomp> at 0x000001DA71743920, file "<stdin>", line 2>)
              4 MAKE_FUNCTION            0
              6 LOAD_GLOBAL_BUILTIN      1 (NULL + range)
             18 LOAD_CONST               2 (2)
             20 CALL_BUILTIN_CLASS       1
             30 GET_ITER
             32 CALL_PY_EXACT_ARGS       0
             42 POP_TOP
             44 LOAD_CONST               0 (None)
             46 RETURN_VALUE

Disassembly of <code object <listcomp> at 0x000001DA71743920, file "<stdin>", line 2>:
  2           0 RESUME_QUICK             0
              2 BUILD_LIST               0
              4 LOAD_FAST                0 (.0)
        >>    6 FOR_ITER_RANGE           4 (to 18)
             10 STORE_FAST__LOAD_FAST     1 (x)
             12 LOAD_FAST                1 (x)
             14 LIST_APPEND              2
             16 JUMP_BACKWARD_QUICK      6 (to 6)
        >>   18 RETURN_VALUE

That CALL_PY_EXACT_ARGS always misses at

DEOPT_IF(func->func_version != read_u32(cache->func_version), CALL);

because the function is created by MAKE_FUNCION so it will never have the same func_version as the previous time.

CALL_PY_EXACT_ARGS misses roughly 5% of the time in pyperformance, and this might decrease that a hair.

Maybe this is insignificant enough that it won't matter, but it would theoretically be nice to statically decide not to specialize these calls.

cc @markshannon