Regression in InstCombine Pass: Missing Simplification of icmp with Bitwise AND and llvm.assume Condition

[RonDahan101]

In LLVM 18.1.0, instCombine considered the llvm.assume condition and managed to prove that the branch is always false.

https://godbolt.org/z/181GG7vvE

Now, instCombine fails to simplify the branch's condition.

https://godbolt.org/z/z4xMT533G

[lquinn2015]

Just finished bisecting this will look into it but it seems stuff around this patch getting apply/ remove and reapplied cause the regression still trying to understand the change made @nikic

05670b4

[nikic]

@RonDahan101 Did this pattern occur in real-world code?

[RonDahan101]

@nikic Yes, it happened on our custom OpenCL target.

[lquinn2015]

I was looking into this and am not to sure where to go it seems like the change i listed above allows for a simplification in the

%iters = icmp ult i32 %and, 8 ----> %iters = icmp ult i32 %add, 8

The bits you assume are on %and and not on %add so there likely needs to be changes made to to LazyValueInfo or CorrelatedValuePropogation passes. I haven't been able to figure out what should change there. Still pretty new to this :/ but that is kind of where i got to

[dtcxzyw]

https://alive2.llvm.org/ce/z/-hCC7E

[dtcxzyw]

Generalized: https://alive2.llvm.org/ce/z/rZgH-P