broken optimization in gen_close()

@markshannon

There is an optimization in gen_close which attempts to shortcut the case where a sub-generator is being closed, and it is wrapped in only one try block (which is presumed to be for the StopIteration so can be ignored).

However, there is a bug there where the exception depth is taken from frame->prev_instr[0].op.code, where it should be frame->prev_instr[0].op.arg.

At the moment this optimization is not doing anything because exception_handler_depth is never 1, it is either 118 (opcode of YIELD_VALUE) or 241 (opcode of INSTRUMENTED_YIELD_VALUE). If one of those opcodes will happen to be mapped to 1, then this code will have some impact but obviously not the intended one.

CC @markshannon

If I try to fix the typo and use op.arg instead of op.code, some tests fail: #111069

I think the return value is incorrect in some cases.

The failure in testcontextlib is fixed by setting the frame state in case we return early:

         if (exception_handler_depth == 1) {
+            gen->gi_frame_state = FRAME_COMPLETED;
             Py_RETURN_NONE;
         }

Probably related: #111058

Perf measurement for #111069:

https://github.com/faster-cpython/benchmarking/blob/main/results/bm-20231020-3.13.0a1%2B-90d6409/bm-20231020-linux-x86_64-iritkatriel-refleak-3.13.0a1%2B-90d6409-vs-base.md

Navigation Menu

Repository navigation

broken optimization in gen_close() #111107

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

broken optimization in gen_close() #111107

Description

Activity

iritkatriel commented on Oct 20, 2023

iritkatriel commented on Oct 20, 2023

Eclips4 commented on Oct 20, 2023

iritkatriel commented on Oct 23, 2023

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions