[llvm] [MachineCopyPropagation] Recognise and delete no-op moves produced after forwarded uses (PR #129889)

[Steven Michaud via llvm-commits]

steven-michaud wrote:

I'm working on it. But LLVM is a complicated beast, and I'm not very familiar with it or with compilers in general.

This bug is important. ARM64 Firefox alone has two or three distinct crash bugs when built with LLVM 21. I'm sure many other apps are also effected. As LLVM 21 is more widely adopted, the shit will really hit the fan. So *someone* needs to work on this, and I'm probably the best suited among all my acquaintances. Among other things I'm retired, and can afford to spend much more time on "interesting" bugs than an employee would.

I suspect you won't find any way to really fix this commit, and that the part concerning 32-bit registers will have to be backed out (as per my second patch [here](https://bugzilla.mozilla.org/show_bug.cgi?id=1995582#c20)). But that's getting ahead of the game. In the meantime I'll be trying to create one or more reduced testcases.

https://github.com/llvm/llvm-project/pull/129889