Missed optimization: fold (x & C1) == C2 to x == C2 when C1 & C2 == C2

[zxt5]

define i1 @src(ptr %arg0) {
  %v0 = getelementptr inbounds nuw i8, ptr %arg0, i64 8
  %v1 = load i32, ptr %v0, align 8
  %v2 = and i32 %v1, 255
  %v3 = icmp eq i32 %v2, 17
  ret i1 %v3
}

define i1 @tgt(ptr %arg0) {
  %v0 = getelementptr inbounds i8, ptr %arg0, i64 8
  %v1 = load i8, ptr %v0, align 8
  %v2 = icmp eq i8 %v1, 17
  ret i1 %v2
}

Alive2: https://alive2.llvm.org/ce/z/tjnF6d
Godbolt: https://godbolt.org/z/9rrPhKE5b

Pattern found in: https://github.com/dtcxzyw/llvm-opt-benchmark/blob/main/bench/llvm/optimized/AArch64TargetMachine.ll

[pinskia]

The real rule is x & mask == C2 to ((new_type) x) == C2 when mask is the mask of a smaller type. And then only load the new_type part of x which might be offset by a value for big-endian.

[pinskia]

Note it looks like Alive2 only tests for little-endian by default. This might be reason for the confusion about the above.

[pinskia]

Also it looks like LLVM already does this.