Probable Context-Aware Warnings for Floating-Point Equality Comparisons

[saadman19]

I am working at OpenRefactory, a static analysis tool company. Our CEO, @openrefactorymunawar previously had some interactions with you and we have been improving our tool by adding the capability to detect some of the bugs that you are describing.

I have a question for one such bug pattern.

Recently, we introduced a fixer that flags warnings when floating-point values are used in equality comparisons. This practice is known to be problematic, as it can lead to unexpected outcomes due to the inherent imprecision of floating-point arithmetic.
The idea for this improvement was inspired by the guidance provided in the
Secure Coding Guide for Python (CWE-664/CWE-197)

Upon running this fixer on various projects, we found a significant number of instances where equality comparisons with floating-point values were being used. For example:

PyTorch 67 instances
NVIDIA Nemo 28 instances

Examples of such occurrences:

1. Nemo
2. NeMo
3. pytorch

This raises a question: Could some of these cases be deliberate on the developers' part, where they knowingly use equality comparisons with floating-point values despite the risks, perhaps for a valid use case? If so, should we raise this warning universally, or should it be more context-dependent?

I'd love to hear your thoughts on this.

[david-a-wheeler]

If you could point to the specific files & line numbers that would be of interest. It's fine to compare "floats" if they're really always integers.

That said, a lot of developers don't really understand computer floating point at all. Using equality is a common mistake, for this reason.

[saadman19]

Here are some instances of such comparison:

1. https://github.com/pytorch/pytorch/blob/main/torch/ao/quantization/observer.py#L1055
2. https://github.com/pytorch/pytorch/blob/main/torch/nn/functional.py#L3686
3. https://github.com/pytorch/pytorch/blob/main/torch/onnx/symbolic_helper.py#L2067
4. https://github.com/pytorch/pytorch/blob/main/torch/_refs/linalg/__init__.py#L234
5. https://github.com/pytorch/pytorch/blob/main/torch/_refs/linalg/__init__.py#L153
6. https://github.com/pytorch/pytorch/blob/main/torch/_refs/__init__.py#L745
7. https://github.com/pytorch/pytorch/blob/main/torch/distributions/transforms.py#L761
8. https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/parts/utils/vad_utils.py#L597
9. https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/parts/utils/speaker_utils.py#L304
10. https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/parts/submodules/adapters/multi_head_attention_adapter_module.py#L71
11. https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/llm/t5/data/pre_training.py#L320
12. https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/vision/modules/common/megatron/vision_transformer.py#L73
13. https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/data/feature_to_text.py#L236

[gkunz]

cc @myteron @BartyBoi1128