[llvm] [X86] Narrow BT/BTC/BTR/BTS compare + RMW patterns on very large integers (PR #165540)

[Simon Pilgrim via llvm-commits]

RKSimon wrote:

> > @durin42 Please do you have the test case yet? Are you sure this wasn't fixed by #165755 / #165850?
> 
> https://lab.llvm.org/buildbot/#/builders/66/builds/21507 should be reasonably easy to reproduce from build bot logs. However minimizing requires some effort. So I'd like to postpone that, with hope that you can guess what could go wrong just from your patch :)

I need to repro it before guessing as I don't want to recommit without seeing a local fix. But I can't seem to create compiler-rt sanitizer i386 unit tests - its been a while since I've had to work on sanitizers - I can see Sanitizer-x86_64-Test but not Sanitizer-i386-Test - any ideas?

> Maybe another easier than minimizing repro option is just to decompose #165540 into even smaller pieces? I am not sure how feasible it is.

Splitting this down causes codegen regressions which we really want to avoid - and I'm sort of expecting this to be causing a problem some place else tbh.

https://github.com/llvm/llvm-project/pull/165540