gh-145866: Convert END_[FOR/SEND] to leave its inputs on the stack to be cleaned up by _POP_TOP

[NekoAsakura]

If I (and opus...) am not mistaken, no code path actually jumps to END_FOR. It's only there for FOR_ITER's oparg calibration and as an instrumentation slot (gets swapped to INSTRUMENTED_END_FOR). So converting it to macro(END_FOR) = POP_TOP should be safe. PyStackRef_CLOSE or PyStackRef_XCLOSE doesn't matter.

ref:

1. Generic _FOR_ITER
   

   cpython/Python/bytecodes.c

   Line 3631 in 03d2f03

   JUMPBY(oparg + 1);

2. INSTRUMENTED_FOR_ITER
   

   cpython/Python/bytecodes.c

   Line 3661 in 03d2f03

   JUMPBY(oparg + 1);

3. _ITER_JUMP_LIST
   

   cpython/Python/bytecodes.c

   Line 3690 in 03d2f03

   JUMPBY(oparg + 1);

4. _ITER_NEXT_LIST (free-threaded)
   

   cpython/Python/bytecodes.c

   Line 3719 in 03d2f03

   JUMPBY(oparg + 1);

5. _ITER_JUMP_TUPLE
   

   cpython/Python/bytecodes.c

   Line 3767 in 03d2f03

   JUMPBY(oparg + 1);

6. _ITER_JUMP_RANGE
   

   cpython/Python/bytecodes.c

   Line 3811 in 03d2f03

   JUMPBY(oparg + 1);

7. _FOR_ITER_TIER_TWO
   

   cpython/Python/bytecodes.c

   Lines 3644 to 3645 in 03d2f03

   	/* The translator sets the deopt target just past the matching END_FOR */
   	EXIT_IF(true);

8. _FOR_ITER_GEN_FRAME
   

   cpython/Python/bytecodes.c

   Line 3855 in 03d2f03

   frame->return_offset = (uint16_t)(INSTRUCTION_SIZE + oparg);

9. codegen_unwind_fblock
   

   cpython/Python/codegen.c

   Lines 546 to 553 in 03d2f03

   	case COMPILE_FBLOCK_FOR_LOOP:
   	/* Pop the iterator */
   	if (preserve_tos) {
   	ADDOP_I(c, *ploc, SWAP, 3);
   	}
   	ADDOP(c, *ploc, POP_TOP);
   	ADDOP(c, *ploc, POP_TOP);
   	return SUCCESS;

• Issue: Convert more uops to leave their inputs on the stack to be cleaned up by _POP_TOP #145866

[NekoAsakura]

#146185 (comment)
I’ve done that lldb thing here as well. Everything looks good to me. Let me know if you’d like me to add any specific unit tests.

[cocolato]

Remove the no_save_ip will change the INSTRUMENTED_POP_ITER's prev_instr from FOR_ITER to END_FOR, this may affect the sys.monitoring callback. This may cause differences in behavior between for loops that use generators and other for loops, such as for x in [1,2,3].

[cocolato]

And END_FOR will be A executed. You can try:

def foo():
    yield 42

for x in foo():
    pass