[llvm] [MemCpyOpt] Merge memset and skip unrelated clobber in one scan (PR #90350)

[Nikita Popov via llvm-commits]

https://github.com/nikic commented:

@XChy I did two experiments in https://llvm-compile-time-tracker.com/index.php?config=Overview&branch=nikic/perf/memcpyopt-memset-formation-4, one was to use unique_ptr and the other to limit the number of ranges to 16/4/1. Limiting to 1 made some difference (but I think this is due to second-order effects, i.e. differences in optimization behavior), but the rest had no impact. I assumed AA would be the expensive part of this, but apparently not.

I haven't had a chance to actually profile this code to understand where the overhead comes from...

https://github.com/llvm/llvm-project/pull/90350