Document the value shifting used in the serialized ATN

Fixes #1270

runtime/Java/src/org/antlr/v4/runtime/atn/ATNDeserializer.java

@@ -110,7 +110,22 @@ public class ATNDeserializer {
 	@SuppressWarnings("deprecation")
 	public ATN deserialize(char[] data) {
 		data = data.clone();
-		// don't adjust the first value since that's the version number
+
+		// Each char value in data is shifted by +2 at the entry to this method.
+		// This is an encoding optimization targeting the serialized values 0
+		// and -1 (serialized to 0xFFFF), each of which are very common in the
+		// serialized form of the ATN. In the modified UTF-8 that Java uses for
+		// compiled string literals, these two character values have multi-byte
+		// forms. By shifting each value by +2, they become characters 2 and 1
+		// prior to writing the string, each of which have single-byte
+		// representations. Since the shift occurs in the tool during ATN
+		// serialization, each target is responsible for adjusting the values
+		// during deserialization.
+		//
+		// As a special case, note that the first element of data is not
+		// adjusted because it contains the major version number of the
+		// serialized ATN, which was fixed at 3 at the time the value shifting
+		// was implemented.
 		for (int i = 1; i < data.length; i++) {
 			data[i] = (char)(data[i] - 2);
 		}

runtime/Java/src/org/antlr/v4/runtime/atn/ATNSerializer.java

@@ -340,6 +340,7 @@ public class ATNSerializer {
 			}
 		}
 
+		// Note: This value shifting loop is documented in ATNDeserializer.
 		// don't adjust the first value since that's the version number
 		for (int i = 1; i < data.size(); i++) {
 			if (data.get(i) < Character.MIN_VALUE || data.get(i) > Character.MAX_VALUE) {