Optimizations should know rounding up a pointer moves it by less than the alignment

[scottmcm]

If you round up a pointer and calculate how far it moved from the original pointer (((p + (a-1)) & -a) - p), that distance is always less than the alignment: https://alive2.llvm.org/ce/z/fL_uXq

define i1 @src(i64 noundef %p, i64 noundef %a) noundef zeroext {
%start:
  %0 = ctpop i64 noundef %a
  %_3 = icmp eq i64 %0, 1
  assume i1 %_3
  %_8 = add i64 noundef %p, -1
  %_7 = add i64 %_8, noundef %a
  %_9 = sub i64 0, noundef %a
  %_6 = and i64 %_7, %_9
  %offset = sub i64 %_6, noundef %p
  %1 = icmp ult i64 %offset, noundef %a
  ret i1 %1
}
=>
define i1 @tgt(i64 noundef %p, i64 noundef %a) noundef zeroext {
%start:
  %0 = ctpop i64 noundef %a
  %_3 = icmp eq i64 %0, 1
  assume i1 %_3
  ret i1 1
}
Transformation seems to be correct!

But it appears that optimizations can't take advantage of that today https://llvm.godbolt.org/z/Mcc7ads1j, not even for constant alignment.

---

(Found trying to use https://doc.rust-lang.org/nightly/std/primitive.pointer.html#method.align_offset in Rust, when I noticed unnecessary checks.)