basicblock_addop Assertion

[15r10nk]

Bug report

Bug description:

I found the following error with the latest cpython 3.12 from git (4a0c118), which I build myself with --with-pydebug enabled.

source:

def get_namespace_uri():
    while element and something:
        try:
            return something
        except:
            pass

output (Python 3.12.0rc3+):

python3: Python/flowgraph.c:114: basicblock_addop: Assertion `0 <= oparg && oparg < (1 << 30)' failed.

CPython versions tested on:

3.12

Operating systems tested on:

Linux

Linked PRs

• gh-109627: duplicated smalll exit blocks need to be assigned jump target labels #109630
• [3.12] gh-109627: duplicated small exit blocks need to be assigned jump target labels (#109630) #109632

[JelleZijlstra]

Confirmed on main, 3.11 is fine. cc @iritkatriel for flowgraph.c.

I'll try to bisect but not sure I'll have time to finish the bisect.

[iritkatriel]

I'm looking. I don't think bisecting will help - the jump targets labels are new so you'll probably get to the large PR that added them.

[iritkatriel]

In order to ensure that function exits have a line number, the compiler duplicates small exit blocks which (1) do not have a line number in any instruction and (2) have more than one predecessor. Since these blocks are jump targets (at least one predecessor reaches this block via a jump or exception), they need to have labels assigned to them. This was missing, and is added in the attached PR.

This begs the question "how did this ever work?"

The assertion that fails here is in a piece of code that transforms a conditional backwards jump into a conditional forward jump followed by an unconditional backward jump (because all backward jumps are unconditional). It is this code that asserted the label. In general the labels are not that useful this late in the compiler's process, they are mostly used in the code generation phase, before we have the CFG with a direct pointer from a jump to its target block.

The typical case of a small basic block is the "return None" of a function, and it is not a target of a conditional backward branch - this is what made this edge case rare.

[iritkatriel]

@15r10nk - Thank you for finding this and for the tiny repro.