BUG: Fix np.abs(nan) returning negative NaN on wasm32/Pyodide

[AnkitAhlawat7742]

Fix: #31421

PR summary

This PR fixes a bug where np.signbit(np.abs(np.nan)) incorrectly returns True instead of False.

problem

The LONGDOUBLE_absolute function in numpy/_core/src/umath/loops.c.src used tmp + 0 to clear -0.0, but this operation corrupted the sign bit of NaN values, causing np.signbit(np.abs(np.nan)) to return True.

Testing

Manually run the new test case and run all the numpy/_core/tests/test_umath.py file test cases.

[charris]

I want to know why adding 0.0 corrupted the nan sign, any ideas? Was it possible that the nan was already negative?

[AnkitAhlawat7742]

HI @charris ,

The bug occurs because nan > 0 always returns falseso the code takes the else branch: tmp = -in1 as per the existing code
For positive NaN:

• tmp = -(+nan) = -nan (flips the sign bit)
• tmp + 0 = -nan (the +0 doesn't fix NaN sign bits)
• Result: negative NaN

So + 0 was meant to clear -0.0, but it doesn't correct NaN sign bits so i checked npy_fabsl() fixes this issue because it's designed to handle NaN
Hope i debug the issue in right direction

[juntyr]

#31421 (comment)

Hi @juntyr , Did you also observe this issue with float32 and float64 types on wasm32/Pyodide? I noticed the same tmp + 0 pattern exists in those implementations as well. Need to check with reviewer that , can we create separate PR for other types (float32,float64`)...if reviewer agree with this solution

I noticed the problem on WASM with float32, float64, and float128; but surprisingly float16 works fine. Could the PR be amended to fix those as well?

[seberg]

float16 uses a custom bit-wise implementation for all of these. But yea, I guess this code predates the abs being mandatory functions. If the float abs also uses comparison, I think we might as well change it.
(Not sure what to think if wasm abs might just doesn't flip the signbit for NaN...)

[AnkitAhlawat7742]

Thank you for the feedback! I have amended the PR to fix float32 and float64 as well.they had the samea > 0 ? a : -apattern in loops_unary_fp.dispatch.c.src and modified the test case accordingly
@seberg please review it once again and let me know if any modification needed

[AnkitAhlawat7742]

HI @seberg ,
Please review the PR ..Happy to make any changes if needed!

[tylerjereddy]

could parametrize for better feedback, although array test below would then make sense to split off

[AnkitAhlawat7742]

Thanks for the feedback ..Test case has been updated accordingly ,I have parametrized the test for better feedback, split the array test into a separate function.

[seberg]

Thanks, LGTM, let's put it in, since C99 guarantees abs being defined.

I am not sure that C99 actually guarantees the sign of NaN (or that we should), but in practice I don't see anything plausibly diverging.