An opcode optimize error in Python 3.12.0

[tiwb]

Bug report

Bug description:

import dis

def _f(a):
	print(a)

dis.dis(_f)

Here is a test in Python 3.12 test_dis.py BytecodeTests.test_disassembled, It outputs:

  3           0 RESUME                   0

  4           2 LOAD_GLOBAL              1 (NULL + print)
             12 LOAD_FAST                0 (a)
             14 CALL                     1
             22 POP_TOP
             24 RETURN_CONST             0 (None)

source code for this optimization is at: https://github.com/python/cpython/blob/3.12/Python/flowgraph.c#L1556-L1560

            case PUSH_NULL:
                if (nextop == LOAD_GLOBAL && (inst[1].i_opcode & 1) == 0) {
                    INSTR_SET_OP0(inst, NOP);
                    inst[1].i_oparg |= 1;
                }
                break;

I feel the correct check should be:

if (nextop == LOAD_GLOBAL && (inst[1].i_oparg & 1) == 0)

I'm unable to write a unit test for this code to make some real error, but if opcode for LOAD_GLOBAL is changed to some odd number, test_dis will fail.

CPython versions tested on:

3.12

Operating systems tested on:

Windows

Linked PRs

• [3.12] gh-112356: LOAD_GLOBAL can only include one PUSH_NULL #112566

[neonene]

It seems to me that inst[1].i_oparg here is supposed to be an even number via compile.c. If you can hit some quantity of an odd number, the tuning for 3.11 and 3.12 might be attractive for CPython team.

[sweeneyde]

This line dates back to 3011a09. It was fixed in main by a9caf9c, but the inst[1].i_opcode & 1 still exists in 3.11 and 3.12.

I feel the correct check should be: if (nextop == LOAD_GLOBAL && (inst[1].i_oparg & 1) == 0)

I agree--this is likely the mistake.

I'm unable to write a unit test for this code to make some real error, but if opcode for LOAD_GLOBAL is changed to some odd number, test_dis will fail.

#112566 uses test_peepholer.DirectCfgOptimizerTests