Simplify `urem` to `icmp+select` in `(x + 1) % n`

[scottmcm]

(Inspired by someone doing self.write_index = (self.write_index + 1) % self.container.len(); in a circular buffer.)

If x < n (such as because it was just used to index an array of that length), then there's no need to modulo in (x + 1) % n, since there's only one case where the modulo does anything, and thus we could detect that case instead.

Alive2 proof: https://alive2.llvm.org/ce/z/UNmz9j

define i64 @src(i64 noundef %x, i64 noundef %n) noundef {
%start:
  %_3 = icmp ult i64 noundef %x, noundef %n
  assume i1 %_3
  %_6 = add nuw i64 noundef %x, 1
  %0 = urem i64 %_6, noundef %n
  ret i64 %0
}
; =>
define i64 @tgt(i64 noundef %x, i64 noundef %n) noundef {
%start:
  %_3 = icmp ult i64 noundef %x, noundef %n
  assume i1 %_3
  %_6 = add nuw i64 noundef %x, 1
  %w = icmp eq i64 %_6, noundef %n
  %0 = select i1 %w, i64 0, i64 %_6
  ret i64 %0
}
; Transformation seems to be correct!

Repro in opt: https://llvm.godbolt.org/z/5vP7T8fE7

[vfdff]

candidate MR: https://reviews.llvm.org/D143883

[scottmcm]

Pardon my ignorance for how these checks work, @vfdff, but will that work where it's a dominating condition even if that's not an assume?

A less-simplified but more-realistic (based on this rust code) version of this would be something like this https://llvm.godbolt.org/z/hreKaK5dc where the x < n condition does dominate the urem, but where it's not literally an assume.

(If isValidAssumeForContext already handles this properly, sorry for the noise.)

[vfdff]

Thanks for your info , I will try to make enhancements your 2nd case if this direction is correct.

[vfdff]

https://reviews.llvm.org/D144248 is used to address the dominating condition

[vfdff]

reopen as https://reviews.llvm.org/D144248 need be review, too