[Feature Request] Add new MAX ops `bitwise_and`, `bitwise_or` and `bitwise_xor`

[gabrieldemarmiesse]

What is your request?

torch functions:

• https://docs.pytorch.org/docs/stable/generated/torch.bitwise_and.html
• https://docs.pytorch.org/docs/stable/generated/torch.bitwise_or.html
• https://docs.pytorch.org/docs/stable/generated/torch.bitwise_xor.html

Note that we could provide ops with those names but also if we want to follow torch's api, we'll want to implement it in TensorValue's __and__, __or__ and __xor__. The behavior in pytorch for those is to use the bitwise version:

In [3]: torch.tensor([1, 2], dtype=torch.uint8) |  torch.tensor([2, 4], dtype=torch.uint8)
Out[3]: tensor([3, 6], dtype=torch.uint8)

What is your motivation for this change?

I need it to implement the max backend for torch, and surprisingly it's used in gemma3 (or at least it's what torch aotautograd decomposes some ops into).

Any other details?

I could code them in torch-max-backend directly in Mojo, but then I end up with this bug: https://github.com/modular/modular/issues/5299

[ferasboulala]

Looking at other implementations, there are two types of operations:

• Bitwise: Straightforward bitwise operations
• Logical: Operations that assume 0 is False and the rest is True They work on all data types.

MAX currently does not support the logical kind of operations and it has a constrain on the data type of the bitwise kind (must be bool).

I propose that we remove the data type constraint on the existing operations. This should give us full bitwise support. Then, we can add logical operations later.