[llvm] [CodeGen] MachineLICM: Do not consider "loop liveins" as loop defs (PR #121769)

[Gaëtan Bossu via llvm-commits]

gbossu wrote:

> Thanks for the patch. I could review this. I had a first quick look at the AArch64 tests that generally look good, but left one question. I will look at the source code changes later today, but a quick test question in the meantime. Did you test this? Did you e.g. do a stage2 build and run its regression tests (as a minimum)?

Thank you very much for the reply. At the moment I've done a full `ninja check` with all official targets enabled. I have to say I'm not really familiar with "stage 2 builds" or other regression test suites. This is also why I would be happy to get reviews from maintainers of the affected targets. As I mentioned in the discourse post, the post-RA MachineLICM does not really have a cost-model like the pre-RA one, and will greedily hoist instructions if it thinks it is safe. That works great for our [AMD-XDNA fork](https://github.com/Xilinx/llvm-aie), but might cause regressions for other use cases. I'd be happy to assist if that's the case.

https://github.com/llvm/llvm-project/pull/121769