3.12: tokenize adds a newline when it is not there

[asottile]

Bug report

tokenize adds a concrete newline character when it is not present. this breaks any sort of roundtrip source generation and pycodestyle's end-of-file checker

here's an example of a file containing a single byte (generated via echo -n 1 > t.py)

# hd -c t.py 
00000000  31                                                |1|
0000000   1                                                            
0000001
# python3.12 -m tokenize t.py
0,0-0,0:            ENCODING       'utf-8'        
1,0-1,1:            NUMBER         '1'            
1,1-1,2:            NEWLINE        '\n'           
2,0-2,0:            ENDMARKER      ''             
# python3.11 -m tokenize t.py
0,0-0,0:            ENCODING       'utf-8'        
1,0-1,1:            NUMBER         '1'            
1,1-1,2:            NEWLINE        ''             
2,0-2,0:            ENDMARKER      '' 

Your environment

• CPython versions tested on: 3.12 dbd7d7c
• Operating system and architecture: ubuntu 22.04 LTS x86_64

cc @pablogsal

Linked PRs

• gh-105259: Ensure we don't show newline characters for trailing NEWLINE tokens #105364
• [3.12] gh-105259: Ensure we don't show newline characters for trailing NEWLINE tokens (GH-105364) #105367

[pablogsal]

@lysnikolaou can you take a look at this?

I think this is due to this line:

cpython/Parser/tokenizer.c

Line 1309 in e6373c0

/* Last line does not end in \n, fake one */

in combination with this line:

cpython/Python/Python-tokenize.c

Line 249 in e6373c0

str = PyUnicode_FromString("\n");

But the real question is what is going to break if we cover the first line with if (!tok->extra_tokens) and how we can fix it.~

[pablogsal]

@lysnikolaou The problem is that we are always adding a new line at the end of the line buffer so we cannot distinguish if the last one is there for real or not (also, we always hardcode \n for NEWLINE tokens which we should not do for the last one).

[lysnikolaou]

The problem is that we are always adding a new line at the end of the line buffer so we cannot distinguish if the last one is there for real or not

Do we? I thought you had removed this in #104850, but I guess I misunderstood that PR. I'll have a look.

[pablogsal]

Do we? I thought you had removed this in #104850, but I guess I misunderstood that PR. I'll have a look.

That was "reverted" because apparently that was required to keep backward compatibility IIRC. But this is also a different kind of "new line": the tokenize module needs new lines at the end of every logical line to correctly emit the last tokens. Just comment out this part and see what fails (locations of last tokens are wrong as well as the some implicit dedents are missing):

cpython/Parser/tokenizer.c

Line 1309 in e6373c0

/* Last line does not end in \n, fake one */

Notice that just removing that is not enough to fix this: we would also need to emit the last artificial newline if the input doesn't contain it.

[pablogsal]

@Yhg1s can you wait until we fix this for the release?

[Yhg1s]

Yeah, I can wait.

[pablogsal]

@asottile can you confirm this fixes it?

[asottile]

yep looks fixed -- this is the only outstanding breaking change I'm encountering: #105390