edge case when parsing complex numbers

[mattip]

Bug report

A clear and concise description of what the bug is.
Include a minimal, reproducible example (https://stackoverflow.com/help/minimal-reproducible-example), if possible.

Parsing -1.0j with no real part will assume the real part is -0, not +0. This is a change from Python2. Is it intentional? PyPy3 will assume +0. This is the root cause of this PyPy issue

$ python3.11
Python 3.11.3 (main, Apr  5 2023, 14:14:37) [GCC 11.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> -1.0j
(-0-1j)
>>> 0-1.0j
-1j
>>> 

Your environment

Debian-provided python3 all do this (python3.9, python3.10, python3.11).

• CPython versions tested on:
• Operating system and architecture:
  Debian, CPython 3.9, 3.10, 3.11

[arigo]

Note that the CPython3 behavior almost makes sense: the expression -1j really means -(1j), and if you do a=1j; -a then you get -0-1j. This is different from doing the subtraction 0-1j, which gives a result with a positive-zero real part. It's the same difference that also occurs without complex number, between -0.0 and 0-0.0.

It only "almost" makes sense, because the detail that confuses matter is that the result of the expression 0-1j is printed as -1j; but the result of the expression -1j is printed as (-0-1j)... (also, the result of the expression (-0-1j) is printed as -1j, because need the expression (-0.0-1j) to print (-0-1j)).

[mdickinson]

Yes, the Python 3 behaviour, and the change w.r.t. Python 2, is intentional.

1j is a complex number with real part +0.0 and imaginary part 1.0. The unary - negates both the real and imaginary part. In Python 2, there was a peculiar AST "optimization" that effectively absorbed the - into the literal.

@arigo: In 0 - 1j we're subtracting the complex number (0.0, 1.0) from the complex number (after coercing the int) (0.0, 0.0). The real part is therefore 0.0 - 0.0, which under the usual IEEE 754 round-ties-to-even mode is 0.0. So it's valid to drop the real part from the repr. In contrast, the real part of -1j is -0.0.

[mdickinson]

@arigo: Ah, sorry; I think I'm just repeating what you already said.

Yes indeed, eval-ing the repr "-1j" of the result of 0-1j doesn't give back precisely the same result. That's pretty much unavoidable, I'm afraid, and has been discussed in multiple other tracker issues.

For some history on the Python 2 to Python 3 change, see #53257.

For issues related to @arigo's comments, see #70026, #61538, #66738 (almost certainly an incomplete list).

[mdickinson]

Also FTR: #84450, and @serhiy-storchaka's PR #19593.