Some parts are similar to what we have

• There's already memcmp optimization, so the predicate classification can be reused
• The elements classification can be reused from min/max_element optimization
• The wrappers for intrinsics can also be reused from min/max_element optimization

Scope

• 1 byte integers (for strings, IP addresses, binary data, etc., try to beat memcmp)
• 2 byte integers (for wide strings)
• 4 or 8 byte integers (for arrays of integers)
• 4 or 8 byte floats (for arrays of floats, after vectorize min/max_element using SSE4.1 for floats #3928)
• Can support infinite ranges removed in Don't memcmp(x, y, -1) from ranges::lexicographical_compare #3989 again