Missed optimization: (x + y) & (2^C) -> x & 2^C when y % 2^(C+1) == 0

[zxt5]

define i1 @src(ptr %0) {
  %v0 = getelementptr inbounds nuw i8, ptr %0, i64 4
  %v1 = load i32, ptr %v0, align 4
  %v2 = zext i32 %v1 to i64
  %v3 = ptrtoint ptr %v0 to i64
  %v4 = add i64 %v2, %v3
  %v5 = and i64 %v4, 2
  %v6 = icmp eq i64 %v5, 0
  ret i1 %v6
}

define i1 @tgt(ptr %0) {
  %v0 = getelementptr inbounds nuw i8, ptr %0, i64 4
  %v1 = load i32, ptr %v0, align 4
  %v2 = and i32 %v1, 2
  %v3 = icmp eq i32 %v2, 0
  ret i1 %v3
}

alive2: https://alive2.llvm.org/ce/z/h8HYTo
godbolt: https://godbolt.org/z/Mqzs9W8q4

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

[egorshamshura]

I would like to take on this task. Can you please assign it to me? This will be my first time working on instcombine, so I would appreciate any advice or guidance

[dtcxzyw]

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

[Explorer09]

I think I can generalize the optimization pattern as stated in the title a little bit:

The constant doesn't have to be a power of 2, (x + y) & C == x & C when (y & (stdc_bit_ceil(C + 1) - 1)) == 0. The stdc_bit_ceil is assumed to return 0 when the rounded-up result would overflow the integer type (a required behavior in C23; it would be undefined behavior in C++).

[egorshamshura]

I think I can generalize the optimization pattern as stated in the title a little bit:

The constant doesn't have to be a power of 2, (x + y) & C == x & C when (y & (stdc_bit_ceil(C + 1) - 1)) == 0. The stdc_bit_ceil is assumed to return 0 when the rounded-up result would overflow the integer type (a required behavior in C23; it would be undefined behavior in C++).

Thanks! I think we need to change the title for this issue

[egorshamshura]

So it seems that sometimes it handles some cases (https://godbolt.org/z/6xas9GqM7). I believe that there is a problem with

 %v0 = getelementptr inbounds nuw i8, ptr %0, i64 4
  %v1 = load i32, ptr %v0, align 4
  %v2 = zext i32 %v1 to i64
  %v3 = ptrtoint ptr %v0 to i64

We are assuming that %v0 has alignment 4 due to %v1 = load i32, ptr %v0, align 4 => %v0 % 2^(C+1) == 0 => we can optimize this case. Am I right?