We were incorrectly handling a few opcodes that leave their operands on the stack. Treat all of these conservatively; assume that they always leave operands on the stack.

FORMAT_SIMPLE leaves its operand on the stack if its a unicode object.
GET_YIELD_FROM_ITER leaves its operand on the stack if its a coroutine or generator object.
END_SEND leaves the TOS in place.
SET_FUNCTION_ATTRIBUTE leaves the TOS in place.
PUSH_EXC_INFO leaves the TOS in place.
LOAD_SPECIAL may leave the TOS (self) in place.

Issue:Python 3.14+:python: Objects/unicodeobject.c:10387: _PyUnicode_JoinArray: Assertionres_data == PyUnicode_1BYTE_DATA(res) + kind * PyUnicode_GET_LENGTH(res)' failed.` in sqlglot #134889

mpage added7 commits

May 30, 2025 10:09

Fix handling ofFORMAT_SIMPLE when optimizingLOAD_FAST

59b5886

`FORMAT_SIMPLE` leaves its operand on the stack if its a unicode object.Since we cannot determine that at analysis (compile) time, treat itconservatively and assume it always leaves the operand on the stack.

Handle SET_FUNCTION_ATTRIBUTE correctly

220bbf9

It leaves the function on the stack

HandleGET_YIELD_FROM_ITER andEND_SEND correctly

58db93d

`GET_YIELD_FROM_ITER` doesn't consume its operand if its a coroutineor generator, so treat it conservatively.`END_SEND` leaves the TOS.

HandlePUSH_EXC_INFO correctly

3e6aa3f

It leaves the TOS in place.

Fix handling ofLOAD_SPECIAL

a60551b

Load special may leave `self` on TOS. Treat it conservatively andassume it always does.

Add blurb

0383f88

Actually bump the magic number, don't just write about it

84ad796

bedevere-appbot mentioned this pull request

May 31, 2025

Python 3.14+:python: Objects/unicodeobject.c:10387: _PyUnicode_JoinArray: Assertionres_data == PyUnicode_1BYTE_DATA(res) + kind * PyUnicode_GET_LENGTH(res)' failed.` in sqlglot#134889

Open

Copy link

Contributor

mgorny commentedMay 31, 2025

I suppose this means Python 3.14 is going to see an ABI breakage too, is that correct?

Copy link

ContributorAuthor

mpage commentedMay 31, 2025

I suppose this means Python 3.14 is going to see an ABI breakage too, is that correct?

I don’t think so. This doesn’t change the binary layout of anything. What about this would cause an ABI breakage?

Copy link

Contributor

thesamesam commentedMay 31, 2025

It changedPYC_MAGIC_NUMBER (which is fine if it can't be avoided, it's just a pain if it can be).

Copy link

ContributorAuthor

mpage commentedMay 31, 2025

It changed PYC_MAGIC_NUMBER (which is fine if it can't be avoided, it's just a pain if it can be).

I think I’m missing something. The magic number changes with every minor release. Since we haven’t released 3.14 yet, bumping the magic number shouldn’t cause any additional pain beyond what’s felt with a normal release.

Copy link

Contributor

mgorny commentedMay 31, 2025

A few Linux distributions have already engaged in a lot of Python 3.14 testing (at least Fedora and Gentoo I'm aware of), and these systems will be affected by the magic number change.