Taken from /org/antlr/tool/Grammar.java all allowed options are

Option	Description
`language`	The target language for code generation. Default is `Java`. See Code Generation Targets for list of currently supported target languages.
`tokenVocab`	Where ANTLR should get predefined tokens and token types. Tree grammars need it to get the token types from the parser that creates its trees. TODO: Default value? Example?
`output`	The type of output the generated parser should return. Valid values are `AST` and `template`. TODO: Briefly, what are the interpretations of these values? Default value?
`ASTLabelType`	Set the type of all tree labels and tree-valued expressions. Without this option, trees are of type `Object`. TODO: Cross-reference default impl (org.antlr.runtime.tree.CommonTree in Java)?
`TokenLabelType`	Set the type of all token-valued expressions. Without this option, tokens are of type `org.antlr.runtime.Token` in Java (`IToken` in C#).
`superClass`	Set the superclass of the generated recognizer. TODO: Default value (`org.antlr.runtime.Parser` in Java)?
`filter`	In the lexer, this allows you to try a list of lexer rules in order. The first one that matches, wins. This is the token that `nextToken()` returns. If nothing matches, the lexer consumes a single character and tries the list of rules again. See Lexical filters for more.
`rewrite`	Valid values are `true` and `false`. Default is `false`. Use this option when your translator output looks very much like the input. Your actions can modify the `TokenRewriteStream` to insert, delete, or replace ranges of tokens with another object. Used in conjunction with `output=template`, you can very easily build translators that tweak input files.
`k`	Limit the lookahead depth for the recognizer to at most `k` symbols. This prevents the decision from using acyclic `LL*` DFA.
`backtrack`	Valid values are `true` and `false`. Default is `false`. Taken from http://www.antlr.org:8080/pipermail/antlr-interest/2006-July/016818.html : The new feature (a big one) is the `backtrack=true` option for grammar, rule, and block that lets you type in any old crap and ANTLR will backtrack if it can't figure out what you meant. No errors are reported by antlr during analysis. It implicitly adds a syn pred in front of every production, using them only if static grammar `LL*` analysis fails. Syn pred code is not generated if the pred is not used in a decision. This is essentially a rapid prototyping mode. It is what I have used on the java.g. Oh, it doesn't memoize partial parses (i.e. rule parsing results) during backtracking automatically now. You must also say `memoize=true`. Can make a HUGE difference to turn on.
`memoize`	Valid values are `true` and `false`. When backtracking, remember whether or not rule references succeed so that the same input position cannot be parsed more than once by the same rule. This effectively guarantees linear parsing when backtracking at the cost of more memory. TODO: Default value (`false`)?