Missed optimization: (x & -2^C) -1 & -2^C + 2^C -> x & -2^C

[zxt5]

define i32 @src(i32 %0) {
  %v0 = and i32 %0, -4
  %v1 = add i32 %v0, -1
  %v2 = and i32 %v1, -4
  %v3 = add nuw i32 %v2, 4
  ret i32 %v3
}

define i32 @tgt(i32 %0) {
  %v0 = and i32 %0, -4
  ret i32 %v0
}

alive2: https://alive2.llvm.org/ce/z/PVar64
godbolt: https://godbolt.org/z/h65Kneqzh

Found this pattern in: https://github.com/dtcxzyw/llvm-opt-benchmark/blob/main/bench/opencv/optimized/t1.ll

[macsencasaus]

I'd like to work on this

[dtcxzyw]

@macsencasaus Please read https://llvm.org/docs/InstCombineContributorGuide.html before submitting a patch. Good luck!

This pattern seems related to bitfield arithmetics. There are many missed optimization reports like this issue. I would expect a general solution, instead of handling specific patterns. Can you briefly describe your solution and provide a generalized proof before opening a PR? I am not sure if this issue is suitable for new contributors.

[Explorer09]

(x & -2^C) -1 & -2^C + 2^C -> x & -2^C

Perhaps the pattern as stated in the title can be reworded to this:

((x & -2^C) - 1) & -2^C -> (x & -2^C) - 2^C