Using Go1.19

Consider this benchmark, which has three functions:

• lessUTF16 is a generic function that operates on either []byte or string. It relies on a type switch to call either utf8.DecodeRune or utf8.DecodeRuneInString.
• lessUTF16Bytes is identical to lessUTF16 but is manually instantiated for []byte.
• lessUTF16String is identical to lessUTF16 but is manually instantiated for string.

The diff from lessUTF16 to lessUTF16Bytes is:

-		var rx, ry rune
-		var nx, ny int
-		switch any(x).(type) {
-		case string:
-			rx, nx = utf8.DecodeRuneInString(string(x))
-			ry, ny = utf8.DecodeRuneInString(string(y))
-		case []byte:
-			rx, nx = utf8.DecodeRune([]byte(x))
-			ry, ny = utf8.DecodeRune([]byte(y))
-		}
+		rx, nx := utf8.DecodeRune(x)
+		ry, ny := utf8.DecodeRune(y)

Similarly, the diff from lessUTF16 to lessUTF16String is:

-		var rx, ry rune
-		var nx, ny int
-		switch any(x).(type) {
-		case string:
-			rx, nx = utf8.DecodeRuneInString(string(x))
-			ry, ny = utf8.DecodeRuneInString(string(y))
-		case []byte:
-			rx, nx = utf8.DecodeRune([]byte(x))
-			ry, ny = utf8.DecodeRune([]byte(y))
-		}
+		rx, nx := utf8.DecodeRuneInString(x)
+		ry, ny := utf8.DecodeRuneInString(y)

I would expect lessUTF16[[]byte] to perform identically to lessUTF16Bytes and lessUTF16[string] to perform identically to lessUTF16String since the compiler can prove that only one of the cases are possibly executable. However, that is not what the benchmarks show:

BenchmarkString           	 1752459	       686.0 ns/op	       0 B/op	       0 allocs/op
BenchmarkBytes            	 1573508	       765.1 ns/op	       0 B/op	       0 allocs/op
BenchmarkGenericString    	 1595708	       744.2 ns/op	       0 B/op	       0 allocs/op
BenchmarkGenericBytes     	 1527207	       793.7 ns/op	       0 B/op	       0 allocs/op

\cc @randall77