3.5.5 Actions in Mid-Rule

Occasionally it is useful to put an action in the middle of a rule. These actions are written just like usual end-of-rule actions, but they are executed before the parser even recognizes the following components.

A mid-rule action may refer to the components preceding it using $n, but it may not refer to subsequent components because it is run before they are parsed.

The mid-rule action itself counts as one of the components of the rule. This makes a difference when there is another action later in the same rule (and usually there is another at the end): you have to count the actions along with the symbols when working out which number n to use in $n.

The mid-rule action can also have a semantic value. The action can set its value with an assignment to $$, and actions later in the rule can refer to the value using $n. Since there is no symbol to name the action, there is no way to declare a data type for the value in advance, so you must use the ‘$<…>n’ construct to specify a data type each time you refer to this value.

There is no way to set the value of the entire rule with a mid-rule action, because assignments to $$ do not have that effect. The only way to set the value for the entire rule is with an ordinary action at the end of the rule.

Here is an example from a hypothetical compiler, handling a let statement that looks like ‘let (variable) statement’ and serves to create a variable named variable temporarily for the duration of statement. To parse this construct, we must put variable into the symbol table while statement is parsed, then remove it afterward. Here is how it is done:

stmt:   LET '(' var ')'
                { $<context>$ = push_context ();
                  declare_variable ($3); }
        stmt    { $$ = $6;
                  pop_context ($<context>5); }

As soon as ‘let (variable)’ has been recognized, the first action is run. It saves a copy of the current semantic context (the list of accessible variables) as its semantic value, using alternative context in the data-type union. Then it calls declare_variable to add the new variable to that list. Once the first action is finished, the embedded statement stmt can be parsed. Note that the mid-rule action is component number 5, so the ‘stmt’ is component number 6.

After the embedded statement is parsed, its semantic value becomes the value of the entire let-statement. Then the semantic value from the earlier action is used to restore the prior list of variables. This removes the temporary let-variable from the list so that it won't appear to exist while the rest of the program is parsed.

In the above example, if the parser initiates error recovery (see section Error Recovery) while parsing the tokens in the embedded statement stmt, it might discard the previous semantic context $<context>5 without restoring it. Thus, $<context>5 needs a destructor (see section Freeing Discarded Symbols). However, Bison currently provides no means to declare a destructor specific to a particular mid-rule action's semantic value.

One solution is to bury the mid-rule action inside a nonterminal symbol and to declare a destructor for that symbol:

%type <context> let
%destructor { pop_context ($$); } let

%%

stmt:  let stmt
               { $$ = $2;
                 pop_context ($1); }
       ;

let:   LET '(' var ')'
               { $$ = push_context ();
                 declare_variable ($3); }
       ;

Note that the action is now at the end of its rule. Any mid-rule action can be converted to an end-of-rule action in this way, and this is what Bison actually does to implement mid-rule actions.

Taking action before a rule is completely recognized often leads to conflicts since the parser must commit to a parse in order to execute the action. For example, the following two rules, without mid-rule actions, can coexist in a working parser because the parser can shift the open-brace token and look at what follows before deciding whether there is a declaration or not:

compound: '{' declarations statements '}'
        | '{' statements '}'
        ;

But when we add a mid-rule action as follows, the rules become nonfunctional:

compound: { prepare_for_local_variables (); }
          '{' declarations statements '}'
        | '{' statements '}'
        ;

Now the parser is forced to decide whether to run the mid-rule action when it has read no farther than the open-brace. In other words, it must commit to using one rule or the other, without sufficient information to do it correctly. (The open-brace token is what is called the lookahead token at this time, since the parser is still deciding what to do about it. See section Lookahead Tokens.)

You might think that you could correct the problem by putting identical actions into the two rules, like this:

compound: { prepare_for_local_variables (); }
          '{' declarations statements '}'
        | { prepare_for_local_variables (); }
          '{' statements '}'
        ;

But this does not help, because Bison does not realize that the two actions are identical. (Bison never tries to understand the C code in an action.)

If the grammar is such that a declaration can be distinguished from a statement by the first token (which is true in C), then one solution which does work is to put the action after the open-brace, like this:

compound: '{' { prepare_for_local_variables (); }
          declarations statements '}'
        | '{' statements '}'
        ;

Now the first token of the following declaration or statement, which would in any case tell Bison which rule to use, can still do so.

Another solution is to bury the action inside a nonterminal symbol which serves as a subroutine:

subroutine: /* empty */
          { prepare_for_local_variables (); }
        ;


compound: subroutine
          '{' declarations statements '}'
        | subroutine
          '{' statements '}'
        ;

Now Bison can execute the action in the rule for subroutine without deciding which rule for compound it will eventually use.