bpo-44626: Merge basic blocks earlier to enable better handling of exit blocks without line numbers. #27138
Conversation
…ithout line numbers.
| } | ||
| } | ||
| break; | ||
| case FOR_ITER: |
There was a problem hiding this comment.
You don't want to call extend_block for the default case?
There was a problem hiding this comment.
We've already extended the blocks. I don't think there is anything to be gained by doing it again.
| } | ||
| } | ||
|
|
||
| for (basicblock *b = c->u->u_blocks; b != NULL; b = b->b_list) { |
There was a problem hiding this comment.
I am confused on why this works. Aren't the blocks supposed to be topologically linked already?
There was a problem hiding this comment.
We haven't computed the entry block yet, so we use b_list.
It also happens that using the backwards order gives better results.
Jumps to exits tend to be forward, so going backwards will be able to combine chains of jumps that going forwards would not.
I'm not making any claims that it is complete, it may well miss some blocks, but it should be good enough.
|
Thanks @markshannon for the PR, and @pablogsal for merging it 🌮🎉.. I'm working now to backport this PR to: 3.10. |
|
Sorry @markshannon and @pablogsal, I had trouble checking out the |
|
@markshannon Can you do the manual backport? |
…it blocks without line numbers (pythonGH-27138) (cherry picked from commit a86f7da)
Improve propagation of line numbers to artificial blocks.
This is an incomplete, but hopefully good enough, solution.
The proper solution requires more extensive CFG modification, which I don't think the current compiler design can support.
One day, I will implement https://bugs.python.org/issue42926 🙂
https://bugs.python.org/issue44626