This is a common transformation for range-check style comparisons. It is always nice to see that LLVM reasons about it in a generalized fashion and applies it to vector operations, and it is what I ultimately ended up on when writing a similar to author's implementation.

I sometimes write intrinsic comparisons (e.g., something like "x < 'a'" instead of "_mm_cmplt_epi8(x, _mm_set1_epi8('a'))") just to make sure the compiler understands what's going on and can do that sort of transformations. (It also sometimes makes for more code sharing between e.g. x86 and NEON.)