[llvm] [MachineLICM] Correctly Apply Register Masks (PR #95746)

[Martin Storsjö via llvm-commits]

mstorsjo wrote:

This commit causes miscompilations for aarch64 (where the generated code does the wrong thing). This can be observed in openh264 unit tests, that suddenly are failing where they previously were passing.

To reproduce, run the following sequence on aarch64 linux:
```
$ git clone https://github.com/cisco/openh264
$ cd openh264
$ make gtest-bootstrap
$ make CC=clang CXX=clang++ -j$(nproc) codec_unittest
$ ./codec_unittest --gtest_filter=McHorVer.16x8_c
```

For closer investigation of what's happening, the culprit object file for this test is `test/encoder/EncUT_MotionCompensation.o`. (There are also test failures triggered in the test `DecoderDecodeMbAux.IdctResAddPred8x8_c`, due to similar miscompilations in `test/decoder/DecUT_IdctResAddPred.o`.)

If there's no straightforward fix in sight, can we revert this change until the issue is resolved? (As this commit is a fix for a regression, do we need to revert that original change meanwhile as well?)

https://github.com/llvm/llvm-project/pull/95746