merl

Metaprogramming in Erlang. Merl is a more user friendly interface to the erl_syntax module, making it easy both to build new ASTs from scratch and to match and decompose existing ASTs. For details that are outside the scope of Merl itself, please see the documentation of erl_syntax.

Quick start

To enable the full power of Merl, your module needs to include the Merl header file:

     -include_lib("syntax_tools/include/merl.hrl").

Then, you can use the ?Q(Text) macros in your code to create ASTs or match on existing ASTs. For example:

     Tuple = ?Q("{foo, 42}"),
     ?Q("{foo, _@Number}") = Tuple,
     Call = ?Q("foo:bar(_@Number)")

Calling merl:print(Call) will then print the following code:

     foo:bar(42)

The ?Q macros turn the quoted code fragments into ASTs, and lifts metavariables such as _@Tuple and _@Number to the level of your Erlang code, so you can use the corresponding Erlang variables Tuple and Number directly. This is the most straightforward way to use Merl, and in many cases it's all you need.

You can even write case switches using ?Q macros as patterns. For example:

     case AST of
         ?Q("{foo, _@Foo}") -> handle(Foo);
         ?Q("{bar, _@Bar}") when erl_syntax:is_integer(Bar) -> handle(Bar);
         _ -> handle_default()
     end

These case switches only allow ?Q(...) or _ as clause patterns, and the guards may contain any expressions, not just Erlang guard expressions.

If the macro MERL_NO_TRANSFORM is defined before the merl.hrl header file is included, the parse transform used by Merl will be disabled, and in that case, the match expressions ?Q(...) = ..., case switches using ?Q(...) patterns, and automatic metavariables like _@Tuple cannot be used in your code, but the Merl macros and functions still work. To do metavariable substitution, you need to use the ?Q(Text, Map) macro, e.g.:

     Tuple = ?Q("{foo, _@bar, _@baz}", [{bar, Bar}, {baz,Baz}])

The text given to a ?Q(Text) macro can be either a single string, or a list of strings. The latter is useful when you need to split a long expression over multiple lines, e.g.:

     ?Q(["case _@Expr of",
         "  {foo, X} -> f(X);",
         "  {bar, X} -> g(X)",
         "  _ -> h(X)"
         "end"])

If there is a syntax error somewhere in the text (like the missing semicolon in the second clause above) this allows Merl to generate an error message pointing to the exact line in your source code. (Just remember to comma-separate the strings in the list, otherwise Erlang will concatenate the string fragments as if they were a single string.)

Metavariable syntax

There are several ways to write a metavariable in your quoted code:

• Atoms starting with @, for example '@foo' or '@Foo'

• Variables starting with _@, for example _@bar or _@Bar

• Strings starting with "'@, for example "'@File"

• Integers starting with 909, for example 9091 or 909123

Following the prefix, one or more _ or 0 characters may be used to indicate "lifting" of the variable one or more levels, and after that, a @ or 9 character indicates a glob metavariable (matching zero or more elements in a sequence) rather than a normal metavariable. For example:

• '@_foo' is lifted one level, and _@__foo is lifted two levels

• _@@bar is a glob variable, and _@_@bar is a lifted glob variable

• 90901 is a lifted variable,90991 is a glob variable, and 9090091 is a glob variable lifted two levels

(Note that the last character in the name is never considered to be a lift or glob marker, hence, _@__ and 90900 are only lifted one level, not two. Also note that globs only matter for matching; when doing substitutions, a non-glob variable can be used to inject a sequence of elements, and vice versa.)

If the name after the prefix and any lift and glob markers is _ or 0, the variable is treated as an anonymous catch-all pattern in matches. For example, _@_, _@@_, _@__, or even _@__@_.

Finally, if the name without any prefixes or lift/glob markers begins with an uppercase character, as in _@Foo or _@_@Foo, it will become a variable on the Erlang level, and can be used to easily deconstruct and construct syntax trees:

     case Input of
         ?Q("{foo, _@Number}") -> ?Q("foo:bar(_@Number)");
         ...

We refer to these as "automatic metavariables". If in addition the name ends with @, as in _@Foo@, the value of the variable as an Erlang term will be automatically converted to the corresponding abstract syntax tree when used to construct a larger tree. For example, in:

     Bar = {bar, 42},
     Foo = ?Q("{foo, _@Bar@}")

(where Bar is just some term, not a syntax tree) the result Foo will be a syntax tree representing {foo, {bar, 42}}. This avoids the need for temporary variables in order to inject data, as in

     TmpBar = erl_syntax:abstract(Bar),
     Foo = ?Q("{foo, _@TmpBar}")

If the context requires an integer rather than a variable, an atom, or a string, you cannot use the uppercase convention to mark an automatic metavariable. Instead, if the integer (without the 909-prefix and lift/glob markers) ends in a 9, the integer will become an Erlang-level variable prefixed with Q, and if it ends with 99 it will also be automatically abstracted. For example, the following will increment the arity of the exported function f:

     case Form of
         ?Q("-export([f/90919]).") ->
             Q2 = erl_syntax:concrete(Q1) + 1,
             ?Q("-export([f/909299]).");
         ...

When to use the various forms of metavariables

Merl can only parse a fragment of text if it follows the basic syntactical rules of Erlang. In most places, a normal Erlang variable can be used as metavariable, for example:

     ?Q("f(_@Arg)") = Expr

but if you want to match on something like the name of a function, you have to use an atom as metavariable:

     ?Q("'@Name'() -> _@@_." = Function

(note the anonymous glob variable _@@_ to ignore the function body).

In some contexts, only a string or an integer is allowed. For example, the directive -file(Name, Line) requires that Name is a string literal and Line an integer literal:

     ?Q("-file(\"'@File\", 9090).") = ?Q("-file(\"foo.erl\", 42).")).

This will extract the string literal "foo.erl" into the variable Foo. Note the use of the anonymous variable 9090 to ignore the line number. To match and also bind a metavariable that must be an integer literal, we can use the convention of ending the integer with a 9, turning it into a Q-prefixed variable on the Erlang level (see the previous section).

Globs

Whenever you want to match out a number of elements in a sequence (zero or more) rather than a fixed set of elements, you need to use a glob. For example:

     ?Q("{_@@Elements}") = ?Q({a, b, c})

will bind Elements to the list of individual syntax trees representing the atoms a, b, and c. This can also be used with static prefix and suffix elements in the sequence. For example:

     ?Q("{a, b, _@@Elements}") = ?Q({a, b, c, d})

will bind Elements to the list of the c and d subtrees, and

     ?Q("{_@@Elements, c, d}") = ?Q({a, b, c, d})

will bind Elements to the list of the a and b subtrees. You can even use plain metavariables in the prefix or suffix:

     ?Q("{_@First, _@@Rest}") = ?Q({a, b, c})

or

     ?Q("{_@@_, _@Last}") = ?Q({a, b, c})

(ignoring all but the last element). You cannot however have two globs as part of the same sequence.

Lifted metavariables

In some cases, the Erlang syntax rules make it impossible to place a metavariable directly where you would like it. For example, you cannot write:

     ?Q("-export([_@@Name]).")

to match out all name/arity pairs in the export list, or to insert a list of exports in a declaration, because the Erlang parser only allows elements on the form A/I (where A is an atom and I an integer) in the export list. A variable like the above is not allowed, but neither is a single atom or integer, so '@@Name' or 909919 wouldn't work either.

What you have to do in such cases is to write your metavariable in a syntactically valid position, and use lifting markers to denote where it should really apply, as in:

     ?Q("-export(['@_@Name'/0]).")

This causes the variable to be lifted (after parsing) to the next higher level in the syntax tree, replacing that entire subtree. In this case, the '@_@Name'/0 will be replaced with '@@Name', and the /0 part was just used as dummy notation and will be discarded.

You may even need to apply lifting more than once. To match the entire export list as a single syntax tree, you can write:

     ?Q("-export(['@__Name'/0]).")

using two underscores, but with no glob marker this time. This will make the entire ['@__Name'/0] part be replaced with '@Name'.

Sometimes, the tree structure of a code fragment isn't very obvious, and parts of the structure may be invisible when printed as source code. For instance, a simple function definition like the following:

     zero() -> 0.

consists of the name (the atom zero), and a list of clauses containing the single clause () -> 0. The clause consists of an argument list (empty), a guard (empty), and a body (which is always a list of expressions) containing the single expression 0. This means that to match out the name and the list of clauses of any function, you'll need to use a pattern like ?Q("'@Name'() -> _@_@Body."), using a dummy clause whose body is a glob lifted one level.

To visualize the structure of a syntax tree, you can use the function merl:show(T), which prints a summary. For example, entering

     merl:show(merl:quote("inc(X, Y) when Y > 0 -> X + Y."))

in the Erlang shell will print the following (where the + signs separate groups of subtrees on the same level):

     function: inc(X, Y) when ... -> X + Y.
       atom: inc
       +
       clause: (X, Y) when ... -> X + Y
         variable: X
         variable: Y
         +
         disjunction: Y > 0
           conjunction: Y > 0
             infix_expr: Y > 0
               variable: Y
               +
               operator: >
               +
               integer: 0
         +
         infix_expr: X + Y
           variable: X
           +
           operator: +
           +
           variable: Y

This shows another important non-obvious case: a clause guard, even if it's as simple as Y > 0, always consists of a single disjunction of one or more conjunctions of tests, much like a tuple of tuples. Thus:

• "when _@Guard ->" will only match a guard with exactly one test

• "when _@@Guard ->" will match a guard with one or more comma-separated tests (but no semicolons), binding Guard to the list of tests

• "when _@_Guard ->" will match just like the previous pattern, but binds Guard to the conjunction subtree

• "when _@_@Guard ->" will match an arbitrary nonempty guard, binding Guard to the list of conjunction subtrees

• "when _@__Guard ->" will match like the previous pattern, but binds Guard to the whole disjunction subtree

• and finally, "when _@__@Guard ->" will match any clause, binding Guard to [] if the guard is empty and to [Disjunction] otherwise

Thus, the following pattern matches all possible clauses:

     "(_@Args) when _@__@Guard -> _@Body"