Incorrect optimization in JIT build

[Eclips4]

Bug report

Bug description:

Reproducer:

def foo():
    a = [1, 2, 3]
    exhit = iter(a)
    for _ in exhit:
        pass
    a.append("this should'be in exhit")
    print(f"got {list(exhit)}, should be []")

foo()
foo()
foo()
foo()
foo()
foo()

Output:

got [], should be []
got [], should be []
got [], should be []
got [], should be []
got [], should be []
got ["this should'be in exhit"], should be []

Obviously, the last line is incorrect.

Output with a PYTHON_LLTRACE=2 env:

got [], should be []
got [], should be []
got [], should be []
got [], should be []
got [], should be []
Optimizing foo (/home/eclips4/programming-languages/cpython/example.py:1) at byte offset 42
   1 ADD_TO_TRACE: _START_EXECUTOR (0, target=21, operand=0x7f4646e59832)
21: JUMP_BACKWARD(5)
   2 ADD_TO_TRACE: _CHECK_VALIDITY_AND_SET_IP (0, target=21, operand=0x7f4646e59832)
   3 ADD_TO_TRACE: _TIER2_RESUME_CHECK (0, target=21, operand=0)
18: FOR_ITER_LIST(3)
   4 ADD_TO_TRACE: _CHECK_VALIDITY_AND_SET_IP (0, target=18, operand=0x7f4646e5982c)
   5 ADD_TO_TRACE: _ITER_CHECK_LIST (3, target=18, operand=0)
   6 ADD_TO_TRACE: _GUARD_NOT_EXHAUSTED_LIST (3, target=18, operand=0)
   7 ADD_TO_TRACE: _ITER_NEXT_LIST (3, target=18, operand=0)
20: STORE_FAST(2)
   8 ADD_TO_TRACE: _CHECK_VALIDITY_AND_SET_IP (0, target=20, operand=0x7f4646e59830)
   9 ADD_TO_TRACE: _STORE_FAST (2, target=20, operand=0)
21: JUMP_BACKWARD(5)
  10 ADD_TO_TRACE: _CHECK_VALIDITY_AND_SET_IP (0, target=21, operand=0x7f4646e59832)
  11 ADD_TO_TRACE: _JUMP_TO_TOP (0, target=0, operand=0)
Created a proto-trace for foo (/home/eclips4/programming-languages/cpython/example.py:1) at byte offset 36 -- length 11
Optimized trace (length 10):
   0 OPTIMIZED: _START_EXECUTOR (0, jump_target=7, operand=0x7f4646e59e80)
   1 OPTIMIZED: _TIER2_RESUME_CHECK (0, jump_target=7, operand=0)
   2 OPTIMIZED: _ITER_CHECK_LIST (3, jump_target=8, operand=0)
   3 OPTIMIZED: _GUARD_NOT_EXHAUSTED_LIST (3, jump_target=9, operand=0)
   4 OPTIMIZED: _ITER_NEXT_LIST (3, target=18, operand=0)
   5 OPTIMIZED: _STORE_FAST_2 (2, target=20, operand=0)
   6 OPTIMIZED: _JUMP_TO_TOP (0, target=0, operand=0)
   7 OPTIMIZED: _DEOPT (0, target=21, operand=0)
   8 OPTIMIZED: _EXIT_TRACE (0, exit_index=0, operand=0)
   9 OPTIMIZED: _EXIT_TRACE (0, exit_index=1, operand=0)
got ["this should'be in exhit"], should be []

It's definitely related to this part of code:

cpython/Python/optimizer.c

Lines 1027 to 1033 in e4a97a7

	if (is_for_iter_test[opcode]) {
	/* Target the POP_TOP immediately after the END_FOR,
	* leaving only the iterator on the stack. */
	int extended_arg = inst->oparg > 255;
	int32_t next_inst = target + 1 + INLINE_CACHE_ENTRIES_FOR_ITER + extended_arg;
	jump_target = next_inst + inst->oparg + 1;
	}

I guess the culprit is there. If we remove the _GUARD_NOT_EXHAUSTED_LIST from is_for_iter_test, the problem will go away (although it can still be reproduced in other ways using other (range, tuple) iterators).

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux, macOS

Linked PRs

• GH-121012: Set index to -1 when list iterators become exhausted in tier 2 #121483
• [3.13] GH-121012: Set index to -1 when list iterators become exhausted in tier 2 (GH-121483) #121494

[brandtbucher]

This is a tier two thing, not a JIT thing. Let's reserve the topic-JIT label for issues that reproduce with --enable-experimental-jit (but not --enable-experimental-jit=interpreter).