[llvm] [SLP]Initial support for (masked)loads + compress and (masked)interleaved (PR #132099)

[Alexander Kornienko via llvm-commits]

alexfh wrote:

> FWIW, it looks like rewriting this as a loop would make the code slightly simpler:
> ...
> And it also makes the crash disappear, though it may be due to the issue being fragile.

No, it only affects the crash when building with the same profile. I can still reproduce the issue after https://github.com/llvm/llvm-project/pull/136146, if I gather a new instrumented profile and build Clang with. 

And indeed, this looks like a miscompile by another Clang optimization. Neither `-fno-slp-vectorize` nor `-fno-vectorize` affect the crash, so it must be some other pass. I didn't have much success using `-opt-bisect-limit` with our LTO+iFDO build setup, probably I was holding it wrong. But it looks like @ZequanWu made some progress on this path. 

https://github.com/llvm/llvm-project/pull/132099