[LLVM] Add support to generate llvm.assume

[quic-sanirudh]

We're adding support to generate llvm.assume from tir.assume as we currently see an error when lowering code with tir.assume in certain cases.

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• No users to tag found in teams: llvm _{See #10317 for details}

_{Generated by tvm-bot}

[quic-sanirudh]

cc @Lunderberg @kparzysz-quic

There is a comment in layout_transformation.cc which suggests that tir.assume might be removed during lowering, but it seems that the tir.remove_assume pass has not yet been added to the default list of passes and so I ran into an error while lowering with assume.

Instead of adding this pass as a default one, I've added support to generate llvm.assume since this info might be useful for LLVM optimizations as well, and users can still manually add the tir.remove_assume pass before lowering to remove it.

[kparzysz-quic]

This looks good, but I'll defer to @Lunderberg to comment on what should happen to the assume builtins in TIR.

[quic-sanirudh]

@Lunderberg Could you please take a look at this change and provide your inputs on handling tir.assume, thanks

[Lunderberg]

The changes look good to me, and I like the idea of exposing it to the LLVM backend. The initial plan was to restrict the usage of tir.assume, such that it did not need to be supported by all code generators, but this is definitely a good use for it.

Perhaps there should be some description of builtins supported by each backend. That way, the RemoveAssume() pass could selectively remove the tir.assume() builtin from backends that do not support it, while preserving it for backends that do.