Ruby Hacking Guide

Chapter 12: Syntax tree construction



As I’ve already described, a Ruby program is first converted to a syntax tree. To be more precise, a syntax tree is a tree structure made of structs called “nodes”. In ruby, all nodes are of type NODE.


 128  typedef struct RNode {
 129      unsigned long flags;
 130      char *nd_file;
 131      union {
 132          struct RNode *node;
 133          ID id;
 134          VALUE value;
 135          VALUE (*cfunc)(ANYARGS);
 136          ID *tbl;
 137      } u1;
 138      union {
 139          struct RNode *node;
 140          ID id;
 141          int argc;
 142          VALUE value;
 143      } u2;
 144      union {
 145          struct RNode *node;
 146          ID id;
 147          long state;
 148          struct global_entry *entry;
 149          long cnt;
 150          VALUE value;
 151      } u3;
 152  } NODE;


Although you might be able to infer from the struct name RNode, nodes are Ruby objects. This means the creation and release of nodes are taken care of by the `ruby`’s garbage collector.

Therefore, flags naturally has the same role as basic.flags of the object struct. It means that T_NODE which is the type of a struct and flags such as FL_FREEZE are stored in it. As for NODE, in addition to these, its node type is stored in flags.

What does it mean? Since a program could contain various elements such as if and while and def and so on, there are also various corresponding node types. The three available union are complicated, but how these unions are used is decided to only one specific way for each node. For example, the below table shows the case when it is NODE_IF that is the node of if.

member union member role
u1 u1.node the condition expression
u2 u2.node the body of true
u3 u3.node the body of false

And, in node.h, the macros to access each union member are available.

▼ the macros to access NODE

 166  #define nd_head  u1.node
 167  #define nd_alen  u2.argc
 168  #define nd_next  u3.node
 170  #define nd_cond  u1.node
 171  #define nd_body  u2.node
 172  #define nd_else  u3.node
 174  #define nd_orig  u3.value


For example, these are used as follows:

NODE *head, *tail;
head->nd_next = tail;    /* head->u3.node = tail */

In the source code, it’s almost certain that these macros are used. A very few exceptions are only the two places where creating NODE in parse.y and where marking NODE in gc.c.

By the way, what is the reason why such macros are used? For one thing, it might be because it’s cumbersome to remember numbers like u1 that are not meaningful by just themselves. But what is more important than that is, there should be no problem if the corresponding number is changed and it’s possible that it will actually be changed. For example, since a condition clause of if does not have to be stored in u1, someone might want to change it to u2 for some reason. But if u1 is directly used, he needs to modify a lot of places all over the source codes, it is inconvenient. Since nodes are all declared as NODE, it’s hard to find nodes that represent if. By preparing the macros to access, this kind of trouble can be avoided and conversely we can determine the node types from the macros.

Node Type

I said that in the flags of a NODE struct its node type is stored. We’ll look at in what form this information is stored. A node type can be set by nd_set_type() and obtained by nd_type().

nd_type nd_set_type

 156  #define nd_type(n) (((RNODE(n))->flags>>FL_USHIFT)&0xff)
 157  #define nd_set_type(n,t) \
 158      RNODE(n)->flags = ((RNODE(n)->flags & ~FL_UMASK) \
                             | (((t)<<FL_USHIFT) & FL_UMASK))



 418  #define FL_USHIFT    11
 429  #define FL_UMASK  (0xff<<FL_USHIFT)


It won’t be so much trouble if we’ll keep focus on around nd_type. Fig.1 shows how it seems like.

Fig.1: The usage of RNode.flags

And, since macros cannot be used from debuggers, the nodetype() function is also available.


4247  static enum node_type
4248  nodetype(node)                  /* for debug */
4249      NODE *node;
4250  {
4251      return (enum node_type)nd_type(node);
4252  }


File Name and Line Number

The nd_file of a NODE holds (the pointer to) the name of the file where the text that corresponds to this node exists. Since there’s the file name, we naturally expect that there’s also the line number, but the corresponding member could not be found around here. Actually, the line number is being embedded to flags by the following macro:

nd_line nd_set_line

 160  #define NODE_LSHIFT (FL_USHIFT+8)
 161  #define NODE_LMASK  (((long)1<<(sizeof(NODE*)*CHAR_BIT-NODE_LSHIFT))-1)
 162  #define nd_line(n) \
          ((unsigned int)((RNODE(n)->flags >> NODE_LSHIFT) & NODE_LMASK))
 163  #define nd_set_line(n,l) \
 164      RNODE(n)->flags = ((RNODE(n)->flags & ~(-1 << NODE_LSHIFT)) \
                             | (((l)&NODE_LMASK) << NODE_LSHIFT))


nd_set_line() is fairly spectacular. However, as the names suggest, it is certain that nd_set_line() and nd_line works symmetrically. Thus, if we first examine the simpler nd_line() and grasp the relationship between the parameters, there’s no need to analyze nd_set_line() in the first place.

The first thing is NODE_LSHIFT, as you can guess from the description of the node types of the previous section, it is the number of used bits in flags. FL_USHIFT is reserved by system of ruby (11 bits, ruby.h), 8 bits are for its node type.

The next thing is NODE_LMASK.


This is the number of the rest of the bits. Let’s assume it is restbits. This makes the code a lot simpler.

#define NODE_LMASK  (((long)1 << restbits) - 1)

Fig.2 shows what the above code seems to be doing. Note that a borrow occurs when subtracting 1. We can eventually understand that NODE_LMASK is a sequence filled with 1 whose size is the number of the bits that are still available.


Now, let’s look at nd_line() again.


By the right shift, the unused space is shifted to the LSB. The bitwise AND leaves only the unused space. Fig.3 shows how flags is used. Since FL_USHIFT is 11, in 32-bit machine 32-(11+8)=13 bits are available for the line number.

Fig.3: How flags are used at NODE

… This means, if the line numbers becomes beyond 2^13=8192, the line numbers should wrongly be displayed. Let’s try.'overflow.rb', 'w') {|f|
    10000.times { f.puts }
    f.puts 'raise'

With my 686 machine, ruby overflow.rb properly displayed 1809 as a line number. I’ve succeeded. However, if you use 64-bit machine, you need to create a little bigger file in order to successfully fail.


Lastly let’s look at the function rb_node_newnode() that creates a node.


4228  NODE*
4229  rb_node_newnode(type, a0, a1, a2)
4230      enum node_type type;
4231      NODE *a0, *a1, *a2;
4232  {
4233      NODE *n = (NODE*)rb_newobj();
4235      n->flags |= T_NODE;
4236      nd_set_type(n, type);
4237      nd_set_line(n, ruby_sourceline);
4238      n->nd_file = ruby_sourcefile;
4240      n->u1.node = a0;
4241      n->u2.node = a1;
4242      n->u3.node = a2;
4244      return n;
4245  }


We’ve seen rb_newobj() in the Chapter 5: Garbage collection. It is the function to get a vacant RVALUE. By attaching the T_NODE struct-type flag to it, the initialization as a VALUE will complete. Of course, it’s possible that some values that are not of type NODE* are passed for u1 u2 u3, but received as NODE* for the time being. Since the syntax trees of ruby does not contain double and such, if the values are received as pointers, it will never be too small in size.

For the rest part, you can forget about the details you’ve learned so far, and assume NODE is

a struct type that has the above seven members.

Syntax Tree Construction

The role of the parser is to convert the source code that is a byte sequence to a syntax tree. Although the grammar passed, it does not finish even half of the task, so we have to assemble nodes and create a tree. In this section, we’ll look at the construction process of that syntax tree.


Essentially this chapter is about actions, thus YYSTYPE which is the type of $$ or $1 becomes important. Let’s look at the %union of ruby first.

%union declaration

 170  %union {
 171      NODE *node;
 172      ID id;
 173      int num;
 174      struct RVarmap *vars;
 175  }


struct RVarmap is a struct used by the evaluator and holds a block local variable. You can tell the rest. The most used one is of course node.

Landscape with Syntax Trees

I mentioned that looking at the fact first is a theory of code reading. Since what we want to know this time is how the generated syntax tree is, we should start with looking at the answer (the syntax tree).

It’s also nice using debuggers to observe every time, but you can visualize the syntax tree more handily by using the tool nodedump contained in the attached CD-ROM, This tool is originally the NodeDump made by Pragmatic Programmers and remodeled for this book. The original version shows quite explanatory output, but this remodeled version deeply and directly displays the appearance of the syntax tree.

For example, in order to dump the simple expression m(a), you can do as follows:

% ruby -rnodedump -e 'm(a)'
nd_file = "-e"
nd_nth  = 1
    nd_mid = 9617 (m)
        nd_alen = 1
            nd_mid = 9625 (a)
        nd_next = (null)

The -r option is used to specify the library to be load, and the -e is used to pass a program. Then, the syntax tree expression of the program will be dumped.

I’ll briefly explain about how to see the content. NODE_NEWLINE and NODE_FCALL and such are the node types. What are written at the same indent level of each node are the contents of its node members. For example, the root is NODE_NEWLINE, and it has the three members: nd_file nd_nth nd_next. nd_file points to the "-e" string of C, and ng_nth points to the 1 integer of C, and nd_next holds the next node NODE_CALL. But since these explanation in text are probably not intuitive, I recommend you to also check Fig.4 at the same time.

Fig.4: Syntax Tree

I’ll explain the meaning of each node. NODE_CALL is a Function CALL. NODE_ARRAY is as its name suggests the node of array, and here it expresses the list of arguments. NODE_VCALL is a Variable or CALL, a reference to undefined local variable will become this.

Then, what is NODE_NEWLINE ? This is the node to join the name of the currently executed file and the line number at runtime and is set for each stmt. Therefore, when only thinking about the meaning of the execution, this node can be ignored. When you require nodedump-short instead of nodedump, distractions like NODE_NEWLINE are left out in the first place. Since it is easier to see if it is simple, nodedump-short will be used later on except for when particularly written.

Now, we’ll look at the three type of composing elements in order to grasp how the whole syntax tree is. The first one is the leaves of a syntax tree. Next, we’ll look at expressions that are combinations of that leaves, this means they are branches of a syntax tree. The last one is the list to list up the statements that is the trunk of a syntax tree in other words.


First, let’s start with the edges that are the leaves of the syntax tree. Literals and variable references and so on, among the rules, they are what belong to primary and are particularly simple even among the primary rules.

% ruby -rnodedump-short -e '1'
nd_lit = 1:Fixnum

1 as a numeric value. There’s not any twist. However, notice that what is stored in the node is not 1 of C but 1 of Ruby (1 of Fixnum). This is because …

% ruby -rnodedump-short -e ':sym'
nd_lit = 9617:Symbol

This way, Symbol is represented by the same NODE_LIT when it becomes a syntax tree. As the above example, VALUE is always stored in nd_lit so it can be handled completely in the same way whether it is a Symbol or a Fixnum when executing. In this way, all we need to do when dealing with it are retrieving the value in nd_lit and returning it. Since we create a syntax tree in order to execute it, designing it so that it becomes convenient when executing is the right thing to do.

% ruby -rnodedump-short -e '"a"'
nd_lit = "a":String

A string. This is also a Ruby string. String literals are copied when actually used.

% ruby -rnodedump -e '[0,1]'
nd_file = "-e"
nd_nth  = 1
    nd_alen = 2
        nd_lit = 0:Fixnum
        nd_alen = 1
            nd_lit = 1:Fixnum
        nd_next = (null)

Array. I can’t say this is a leaf, but let’s allow this to be here because it’s also a literal. It seems like a list of NODE_ARRAY hung with each element node. The reason why only in this case I didn’t use nodedump-short is … you will understand after finishing to read this section.


Next, we’ll focus on “combinations” that are branches. if will be taken as an example.


I feel like if is always used as an example, that’s because its structure is simple and there’s not any reader who don’t know about if, so it is convenient for writers.

Anyway, this is an example of if. For example, let’s convert this code to a syntax tree.

▼The Source Program

if true
  'true expr'
  'false expr'

▼Its syntax tree expression

    nd_lit = "true expr":String
    nd_lit = "false expr":String

Here, the previously described nodedump-short is used, so NODE_NEWLINE disappeared. nd_cond is the condition, nd_body is the body of the true case, nd_else is the body of the false case.

Then, let’s look at the code to build this.

if rule

1373                  | kIF expr_value then
1374                    compstmt
1375                    if_tail
1376                    kEND
1377                      {
1378                          $$ = NEW_IF(cond($2), $4, $5);
1379                          fixpos($$, $2);
1380                      }


It seems that NEW_IF() is the macro to create NODE_IF. Among the values of the symbols, $2 $4 $5 are used, thus the correspondences between the symbols of the rule and $n are:

kIF    expr_value  then  compstmt  if_tail  kEND
 $1          $2      $3        $4       $5    $6
NEW_IF(expr_value,       compstmt, if_tail)

this way. In other words, expr_value is the condition expression, compstmt ($4) is the case of true, if_tail is the case of false.

On the other hand, the macros to create nodes are all named NEW_xxxx, and they are defined node.h. Let’s look at NEW_IF().


 243  #define NEW_IF(c,t,e) rb_node_newnode(NODE_IF,c,t,e)


As for the parameters, it seems that c represents condition, t represents then, and e represents else respectively. As described at the previous section, the order of members of a node is not so meaningful, so you don’t need to be careful about parameter names in this kind of place.

And, the code() which processes the node of the condition expression in the action is a semantic analysis function. This will be described later.

Additionally, fixpos() corrects the line number. NODE is initialized with the file name and the line number of the time when it is “created”. However, for instance, the code of if should already be parsed by end by the time when creating NODE_IF. Thus, the line number would go wrong if it remains untouched. Therefore, it needs to be corrected by fixpos().

fixpos(dest, src)

This way, the line number of the node dest is set to the one of the node src. As for if, the line number of the condition expression becomes the line number of the whole if expression.


Subsequently, let’s look at the rule of if_tail.


1543  if_tail         : opt_else
1544                  | kELSIF expr_value then
1545                    compstmt
1546                    if_tail
1547                      {
1548                          $$ = NEW_IF(cond($2), $4, $5);
1549                          fixpos($$, $2);
1550                      }

1553  opt_else        : none
1554                  | kELSE compstmt
1555                      {
1556                          $$ = $2;
1557                      }


First, this rule expresses “a list ends with opt_else after zero or more number of elsif clauses”. That’s because, if_tail appears again and again while elsif continues, it disappears when opt_else comes in. We can understand this by extracting arbitrary times.

if_tail: kELSIF .... if_tail
if_tail: kELSIF .... kELSIF .... if_tail
if_tail: kELSIF .... kELSIF .... kELSIF .... if_tail
if_tail: kELSIF .... kELSIF .... kELSIF .... opt_else
if_tail: kELSIF .... kELSIF .... kELSIF .... kELSE compstmt

Next, let’s focus on the actions, surprisingly, elsif uses the same NEW_IF() as if. It means, the below two programs will lose the difference after they become syntax trees.

if cond1                  if cond1
  body1                     body1
elsif cond2               else
  body2                     if cond2
elsif cond3                   body2
  body3                     else
else                          if cond3
  body4                         body3
end                           else

Come to think of it, in C language and such, there’s no distinction between the two also at the syntax level. Thus this might be a matter of course. Alternatively, the conditional operator (a?b:c) becomes indistinguishable from if statement after they become syntax trees.

The precedences was very meaningful when it was in the context of grammar, but they become unnecessary any more because the structure of a syntax tree contains that information. And, the difference in appearance such as if and the conditional operator become completely meaningless, its meaning (its behavior) only matters. Therefore, there’s perfectly no problem if if and the conditional operator are the same in its syntax tree expression.

I’ll introduce a few more examples. add and && become the same. or and || are also equal to each other. not and !, if and modifier if, and so on. These pairs also become equal to each other.

Left Recursive and Right Recursive

By the way, the symbol of a list was always written at the left side when expressing a list in Chapter 9: yacc crash course. However, have you noticed it becomes opposite in if_tail ? I’ll show only the crucial part again.

if_tail: opt_else
       | kELSIF ... if_tail

Surely, it is opposite of the previous examples. if_tail which is the symbol of a list is at the right side.

In fact, there’s another established way of expressing lists,

list: END_ITEM
    | ITEM list

when you write in this way, it becomes the list that contains continuous zero or more number of ITEM and ends with END_ITEM.

As an expression of a list, whichever is used it does not create a so much difference, but the way that the actions are executed is fatally different. With the form that list is written at the right, the actions are sequentially executed from the last ITEM. We’ve already learned about the behavior of the stack of when list is at the left, so let’s try the case that list is at the right. The input is 4 ITEM s and END_ITEM.

empty at first
ITEM ITEM ITEM ITEM list reduce END_ITEM to list
ITEM ITEM ITEM list reduce ITEM list to list
ITEM ITEM list reduce ITEM list to list
ITEM list reduce ITEM list to list
list reduce ITEM list to list

When list was at the left, shifts and reductions were done in turns. This time, as you see, there are continuous shifts and continuous reductions.

The reason why if_tail places “list at the right” is to create a syntax tree from the bottom up. When creating from the bottom up, the node of if will be left in hand in the end. But if defining if_tail by placing “list at the left”, in order to eventually leave the node of if in hand, it needs to traverse all links of the elsif and every time elsif is found add it to the end. This is cumbersome. And, slow. Thus, if_tail is constructed in the “list at the right” manner.

Finally, the meaning of the headline is, in grammar terms, “the left is list” is called left-recursive, “the right is list” is called right-recursive. These terms are used mainly when reading papers about processing grammars or writing a book of yacc.


Leaf, branch, and finally, it’s trunk. Let’s look at how the list of statements are joined.

▼The Source Program


The dump of the corresponding syntax tree is shown below. This is not nodedump-short but in the perfect form.

▼Its Syntax Tree

    nd_file = "multistmt"
    nd_nth  = 1
        nd_lit = 7:Fixnum
        nd_file = "multistmt"
        nd_nth  = 2
            nd_lit = 8:Fixnum
            nd_file = "multistmt"
            nd_nth  = 3
                nd_lit = 9:Fixnum
        nd_next = (null)

We can see the list of NODE_BLOCK is created and NODE_NEWLINE are attached as headers. (Fig.5)


It means, for each statement (stmt) NODE_NEWLINE is attached, and when they are multiple, it will be a list of NODE_BLOCK. Let’s also see the code.


 354  stmts           : none
 355                  | stmt
 356                      {
 357                          $$ = newline_node($1);
 358                      }
 359                  | stmts terms stmt
 360                      {
 361                          $$ = block_append($1, newline_node($3));
 362                      }


newline_node() caps NODE_NEWLINE, block_append() appends it to the list. It’s straightforward. Let’s look at the content only of the block_append().


It this function, the error checks are in the very middle and obstructive. Thus I’ll show the code without that part.

block_append() (omitted)

4285  static NODE*
4286  block_append(head, tail)
4287      NODE *head, *tail;
4288  {
4289      NODE *end;
4291      if (tail == 0) return head;
4292      if (head == 0) return tail;
4294      if (nd_type(head) != NODE_BLOCK) {
4295          end = NEW_BLOCK(head);
4296          end->nd_end = end;    /*(A-1)*/
4297          fixpos(end, head);
4298          head = end;
4299      }
4300      else {
4301          end = head->nd_end;   /*(A-2)*/
4302      }

          /* ……omitted…… */

4325      if (nd_type(tail) != NODE_BLOCK) {
4326          tail = NEW_BLOCK(tail);
4327          tail->nd_end = tail;
4328      }
4329      end->nd_next = tail;
4330      head->nd_end = tail->nd_end;   /*(A-3)*/
4331      return head;
4332  }


According to the previous syntax tree dump, NEW_BLOCK was a linked list uses nd_next. Being aware of it while reading, it can be read “if either head or tail is not NODE_BLOCK, wrap it with NODE_BLOCK and join the lists each other.”

Additionally, on (A-1~3), the nd_end of the NODE_BLOCK of the head of the list always points to the NODE_BLOCK of the tail of the list. This is probably because in this way we don’t have to traverse all elements when adding an element to the tail (Fig.6). Conversely speaking, when you need to add elements later, NODE_BLOCK is suitable.

Fig.6: Appending is easy.

The two types of lists

Now, I’ve explained the outline so far. Because the structure of syntax tree will also appear in Part 3 in large amounts, we won’t go further as long as we are in Part 2. But before ending, there’s one more thing I’d like to talk about. It is about the two general-purpose lists.

The two general-purpose lists mean BLOCK and LIST. BLOCK is, as previously described, a linked list of NODE_BLOCK to join the statements. LIST is, although it is called LIST, a list of NODE_ARRAY. This is what is used for array literals. LIST is used to store the arguments of a method or the list of multiple assignments.

As for the difference between the two lists, looking at the usage of the nodes is helpful to understand.

NODE_BLOCK nd_head holding an element
nd_end pointing to the NODE_BLOCK of the end of the list
nd_next pointing to the next NODE_BLOCK
NODE_ARRAY nd_head holding an element
nd_alen the length of the list that follows this node
nd_next pointing to the next NODE_ARRAY

The usage differs only in the second elements that are nd_end and nd_alen. And this is exactly the significance of the existence of each type of the two nodes. Since its size can be stored in NODE_ARRAY, we use an ARRAY list when the size of the list will frequently be required. Otherwise, we use a BLOCK list that is very fast to join. I don’t describe this topic in details because the codes that use them is necessary to understand the significance but not shown here, but when the codes appear in Part 3, I’d like you to recall this and think “Oh, this uses the length”.

Semantic Analysis

As I briefly mentioned at the beginning of Part 2, there are two types of analysis that are appearance analysis and semantic analysis. The appearance analysis is mostly done by yacc, the rest is doing the semantic analysis inside actions.

Errors inside actions

What does the semantic analysis precisely mean? For example, there are type checks in a language that has types. Alternatively, check if variables with the same name are not defined multiple times, and check if variables are not used before their definitions, and check if the procedure being used is defined, and check if return is not used outside of procedures, and so on. These are part of the semantic analysis.

What kind of semantic analysis is done in the current ruby ? Since the error checks occupies almost all of semantic analysis in ruby, searching the places where generating errors seems a good way. In a parser of yacc, yyerror() is supposed to be called when an error occurs. Conversely speaking, there’s an error where yyerror() exists. So, I made a list of the places where calling yyerror() inside the actions.

These checks can roughly be categorized by each purpose as follows:

For example, “return outside of a method” is a check in order not to make the rule too complex. Since this error is a problem of the structure, it can be dealt with by grammar. For example, it’s possible by defining the rules separately for both inside and outside of methods and making the list of all what are allowed and what are not allowed respectively. But this is in any way cumbersome and rejecting it in an action is far more concise.

And, “an assignment to self” seems a check for the better error message. In comparison to “return outside of methods”, rejecting it by grammar is much easier, but if it is rejected by the parser, the output would be just "parse error". Comparing to it, the current

% ruby -e 'self = 1'
-e:1: Can't change the value of self
self = 1

this error is much more friendly.

Of course, we can not always say that an arbitrary rule is exactly “for this purpose”. For example, as for “return outside of methods”, this can also be considered that this is a check “for the better error message”. The purposes are overlapping each other.

Now, the problem is “a pure semantic analysis”, in Ruby there are few things belong to this category. In the case of a typed language, the type analysis is a big event, but because variables are not typed in Ruby, it is meaningless. What is standing out instead is the cheek of an expression that has its value.

To put “having its value” precisely, it is “you can obtain a value as a result of evaluating it”. return and break do not have values by themselves. Of course, a value is passed to the place where return to, but not any values are left at the place where return is written. Therefore, for example, the next expression is odd,

i = return(1)

Since this kind of expressions are clearly due to misunderstanding or simple mistakes, it’s better to reject when compiling. Next, we’ll look at value_expr which is one of the functions to check if it takes a value.


value_expr() is the function to check if it is an expr that has a value.


4754  static int
4755  value_expr(node)
4756      NODE *node;
4757  {
4758      while (node) {
4759          switch (nd_type(node)) {
4760            case NODE_CLASS:
4761            case NODE_MODULE:
4762            case NODE_DEFN:
4763            case NODE_DEFS:
4764              rb_warning("void value expression");
4765              return Qfalse;
4767            case NODE_RETURN:
4768            case NODE_BREAK:
4769            case NODE_NEXT:
4770            case NODE_REDO:
4771            case NODE_RETRY:
4772              yyerror("void value expression");
4773              /* or "control never reach"? */
4774              return Qfalse;
4776            case NODE_BLOCK:
4777              while (node->nd_next) {
4778                  node = node->nd_next;
4779              }
4780              node = node->nd_head;
4781              break;
4783            case NODE_BEGIN:
4784              node = node->nd_body;
4785              break;
4787            case NODE_IF:
4788              if (!value_expr(node->nd_body)) return Qfalse;
4789              node = node->nd_else;
4790              break;
4792            case NODE_AND:
4793            case NODE_OR:
4794              node = node->nd_2nd;
4795              break;
4797            case NODE_NEWLINE:
4798              node = node->nd_next;
4799              break;
4801            default:
4802              return Qtrue;
4803          }
4804      }
4806      return Qtrue;
4807  }



Summary: It sequentially checks the nodes of the tree, if it hits “an expression certainly not having its value”, it means the tree does not have any value. Then it warns about that by using rb_warning() and return Qfalse. If it finishes to traverse the entire tree without hitting any “an expression not having its value”, it means the tree does have a value. Thus it returns Qtrue.

Here, notice that it does not always need to check the whole tree. For example, let’s assume value_expr() is called on the argument of a method. Here:

▼ check the value of arg by using value_expr()

1055  arg_value       : arg
1056                      {
1057                          value_expr($1);
1058                          $$ = $1;
1059                      }


Inside of this argument $1, there can also be other nesting method calls again. But, the argument of the inside method must have been already checked with value_expr(), so you don’t have to check it again.

Let’s think more generally. Assume an arbitrary grammar element A exists, and assume value_expr() is called against its all composing elements, the necessity to check the element A again would disappear.

Then, for example, how is if ? Is it possible to be handled as if value_expr() has already called for all elements? If I put only the bottom line, it isn’t. That is because, since if is a statement (which does not use a value), the main body should not have to return a value. For example, in the next case:

def method
  if true
    return 1
    return 2

This if statement does not need a value.
But in the next case, its value is necessary.

def method( arg )
  tmp = if arg
        then 3
        else 98
  tmp * tmp / 3.5

So, in this case, the if statement must be checked when checking the entire assignment expression. This kind of things are laid out in the switch statement of value_expr().

Removing Tail Recursion

By the way, when looking over the whole value_expr, we can see that there’s the following pattern appears frequently:

while (node) {
    switch (nd_type(node)) {
      case NODE_XXXX:
        node = node->nd_xxxx;

This expression will also carry the same meaning after being modified to the below:

return value_expr(node->nd_xxxx)

A code like this which does a recursive call just before return is called a tail recursion. It is known that this can generally be converted to goto. This method is often used when optimizing. As for Scheme, it is defined in specifications that tail recursions must be removed by language processors. This is because recursions are often used instead of loops in Lisp-like languages.

However, be careful that tail recursions are only when “calling just before return”. For example, take a look at the NODE_IF of value_expr(),

if (!value_expr(node->nd_body)) return Qfalse;
node = node->nd_else;

As shown above, the first time is a recursive call. Rewriting this to the form of using return,

return value_expr(node->nd_body) && value_expr(node->nd_else);

If the left value_expr() is false, the right value_expr() is also executed. In this case, the left value_expr() is not “just before” return. Therefore, it is not a tail recursion. Hence, it can’t be extracted to goto.

The whole picture of the value check

As for value checks, we won’t read the functions further. You might think it’s too early, but all of the other functions are, as the same as value_expr(), step-by-step one-by-one only traversing and checking nodes, so they are completely not interesting. However, I’d like to cover the whole picture at least, so I finish this section by just showing the call graph of the relevant functions (Fig.7).

Fig.7: the call graph of the value check functions

Local Variables

Local Variable Definitions

The variable definitions in Ruby are really various. As for constants and class variables, these are defined on the first assignment. As for instance variables and global variables, as all names can be considered that they are already defined, you can refer them without assigning beforehand (although it produces warnings).

The definitions of local variables are again completely different from the above all. A local variable is defined when its assignment appears on the program. For example, as follows:

lvar = nil
p lvar      # being defined

In this case, as the assignment to lvar is written at the first line, in this moment lvar is defined. When it is undefined, it ends up with a runtime exception NameError as follows:

% ruby lvar.rb
lvar.rb:1: undefined local variable or method `lvar'
for #<Object:0x40163a9c> (NameError)

Why does it say "local variable or method"? As for methods, the parentheses of the arguments can be omitted when calling, so when there’s not any arguments, it can’t be distinguished from local variables. To resolve this situation, ruby tries to call it as a method when it finds an undefined local variable. Then if the corresponding method is not found, it generates an error such as the above one.

By the way, it is defined when “it appears”, this means it is defined even though it was not assigned. The initial value of a defined variable is nil.

if false
  lvar = "this assigment will never be executed"
p lvar   # shows nil

Moreover, since it is defined “when” it “appears”, the definition has to be before the reference in a symbol sequence. For example, in the next case, it is not defined.

p lvar       # not defined !
lvar = nil   # although appearing here ...

Be careful about the point of “in the symbol sequence”. It has completely nothing to do with the order of evaluations. For example, for the next code, naturally the condition expression is evaluated first, but in the symbol sequence, at the moment when p appears the assignment to lvar has not appeared yet. Therefore, this produces NameError.

p(lvar) if lvar = true

What we’ve learned by now is that the local variables are extremely influenced by the appearances. When a symbol sequence that expresses an assignment appears, it will be defined in the appearance order. Based on this information, we can infer that ruby seems to define local variables while parsing because the order of the symbol sequence does not exist after leaving the parser. And in fact, it is true. In ruby, the parser defines local variables.

Block Local Variables

The local variables newly defined in an iterator block are called block local variables or dynamic variables. Block local variables are, in language specifications, identical to local variables. However, these two differ in their implementations. We’ll look at how is the difference from now on.

The data structure

We’ll start with the local variable table struct local_vars.

struct local_vars

5174  static struct local_vars {
5175      ID *tbl;                    /* the table of local variable names */
5176      int nofree;                 /* whether it is used from outside */
5177      int cnt;                    /* the size of the tbl array */
5178      int dlev;                   /* the nesting level of dyna_vars */
5179      struct RVarmap* dyna_vars;  /* block local variable names */
5180      struct local_vars *prev;
5181  } *lvtbl;


The member name prev indicates that the struct local_vars is a opposite-direction linked list. … Based on this, we can expect a stack. The simultaneously declared global variable lvtbl points to local_vars that is the top of that stack.

And, struct RVarmap is defined in env.h, and is available to other files and is also used by the evaluator. This is used to store the block local variables.

struct RVarmap

  52  struct RVarmap {
  53      struct RBasic super;
  54      ID id;                  /* the variable name */
  55      VALUE val;              /* its value */
  56      struct RVarmap *next;
  57  };


Since there’s struct RBasic at the top, this is a Ruby object. It means it is managed by the garbage collector. And since it is joined by the next member, it is probably a linked list.

Based on the observation we’ve done and the information that will be explained, Fig.8 illustrates the image of both structs while executing the parser.

Fig.8: The image of local variable tables at runtime

Local Variable Scope

When looking over the list of function names of parse.y, we can find functions such as local_push() local_pop() local_cnt() are laid out. In whatever way of thinking, they appear to be relating to a local variable. Moreover, because the names are push pop, it is clearly a stack. So first, let’s find out the places where using these functions.

local_push() local_pop() used examples

1475                  | kDEF fname
1476                      {
1477                          $<id>$ = cur_mid;
1478                          cur_mid = $2;
1479                          in_def++;
1480                          local_push(0);
1481                      }
1482                    f_arglist
1483                    bodystmt
1484                    kEND
1485                      {
1486                          /* NOEX_PRIVATE for toplevel */
1487                          $$ = NEW_DEFN($2, $4, $5,
1488                          if (is_attrset_id($2))
                                  $$->nd_noex = NOEX_PUBLIC;
1489                          fixpos($$, $4);
1490                          local_pop();
1491                          in_def--;
1492                          cur_mid = $<id>3;
1493                      }


At def, I could find the place where it is used. It can also be found in class definitions and singleton class definitions, and module definitions. In other words, it is the place where the scope of local variables is cut. Moreover, as for how they are used, it does push where the method definition starts and does pop when the definition ends. This means, as we expected, it is almost certain that the functions start with local_ are relating to local variables. And it is also revealed that the part between push and pop is probably a local variable scope.

Moreover, I also searched local_cnt().


 269  #define NEW_LASGN(v,val) rb_node_newnode(NODE_LASGN,v,val,local_cnt(v))


This is found in node.h. Even though there are also the places where using in parse.y, I found it in the other file. Thus, probably I’m in desperation.

This NEW_LASGN is “new local assignment”. This should mean the node of an assignment to a local variable. And also considering the place where using it, the parameter v is apparently the local variable name. val is probably (a syntax tree that represents). the right-hand side value

Based on the above observations, local_push() is at the beginning of the local variable, local_cnt() is used to add a local variable if there’s a local variable assignment in the halfway, local_pop() is used when ending the scope. This perfect scenario comes out. (Fig.9)

Fig.9: the flow of the local variable management

Then, let’s look at the content of the function.

push and pop


5183  static void
5184  local_push(top)
5185      int top;
5186  {
5187      struct local_vars *local;
5189      local = ALLOC(struct local_vars);
5190      local->prev = lvtbl;
5191      local->nofree = 0;
5192      local->cnt = 0;
5193      local->tbl = 0;
5194      local->dlev = 0;
5195      local->dyna_vars = ruby_dyna_vars;
5196      lvtbl = local;
5197      if (!top) {
5198          /* preserve the variable table of the previous scope into val  */
5199          rb_dvar_push(0, (VALUE)ruby_dyna_vars);
5200          ruby_dyna_vars->next = 0;
5201      }
5202  }


As we expected, it seems that struct local_vars is used as a stack. Also, we can see lvtbl is pointing to the top of the stack. The lines relates to rb_dvar_push() will be read later, so it is left untouched for now.

Subsequently, we’ll look at local_pop() and local_tbl() at the same time.

local_tbl local_pop

5218  static ID*
5219  local_tbl()
5220  {
5221      lvtbl->nofree = 1;
5222      return lvtbl->tbl;
5223  }

5204  static void
5205  local_pop()
5206  {
5207      struct local_vars *local = lvtbl->prev;
5209      if (lvtbl->tbl) {
5210          if (!lvtbl->nofree) free(lvtbl->tbl);
5211          else lvtbl->tbl[0] = lvtbl->cnt;
5212      }
5213      ruby_dyna_vars = lvtbl->dyna_vars;
5214      free(lvtbl);
5215      lvtbl = local;
5216  }


I’d like you to look at local_tbl(). This is the function to obtain the current local variable table (lvtbl->tbl). By calling this, the nofree of the current table becomes true. The meaning of nofree seems naturally “Don’t free()”. In other words, this is like reference counting, “this table will be used, so please don’t free()”. Conversely speaking, when local_tbl() was not called with a table even once, that table will be freed at the moment when being popped and be discarded. For example, this situation probably happens when a method without any local variables.

However, the “necessary table” here means lvtbl->tbl. As you can see, lvtbl itself will be freed at the same moment when being popped. It means only the generated lvtbl->tbl is used in the evaluator. Then, the structure of lvtbl->tbl is becoming important. Let’s look at the function local_cnt() (which seems) to add variables which is probably helpful to understand how the structure is.

And before that, I’d like you to remember that lvtbl->cnt is stored at the index 0 of the lvtbl->tbl.

Adding variables

The function (which seems) to add a local variable is local_cnt().


5246  static int
5247  local_cnt(id)
5248      ID id;
5249  {
5250      int cnt, max;
5252      if (id == 0) return lvtbl->cnt;
5254      for (cnt=1, max=lvtbl->cnt+1; cnt<max;cnt++) {
5255          if (lvtbl->tbl[cnt] == id) return cnt-1;
5256      }
5257      return local_append(id);
5258  }


This scans lvtbl->tbl and searches what is equals to id. If the searched one is found, it straightforwardly returns cnt-1. If nothing is found, it does local_append(). local_append() must be, as it is called append, the procedure to append. In other words, local_cnt() checks if the variable was already registered, if it was not, adds it by using local_append() and returns it.

What is the meaning of the return value of this function? lvtbl->tbl seems an array of the variables, so there’re one-to-one correspondences between the variable names and “their index – 1 (cnt-1)”. (Fig.10)

Fig.10: The correspondences between the variable names and the return values

Moreover, this return value is calculated so that the start point becomes 0, the local variable space is probably an array. And, this returns the index to access that array. If it is not, like the instance variables or constants, (the ID of) the variable name could have been used as a key in the first place.

You might want to know why it is avoiding index 0 (the loop start from cnt=1) for some reasons, it is probably to store a value at local_pop().

Based on the knowledge we’ve learned, we can understand the role of local_append() without actually looking at the content. It registers a local variable and returns “(the index of the variable in lvtbl->tbl) – 1”. It is shown below, let’s make sure.


5225  static int
5226  local_append(id)
5227      ID id;
5228  {
5229      if (lvtbl->tbl == 0) {
5230          lvtbl->tbl = ALLOC_N(ID, 4);
5231          lvtbl->tbl[0] = 0;
5232          lvtbl->tbl[1] = '_';
5233          lvtbl->tbl[2] = '~';
5234          lvtbl->cnt = 2;
5235          if (id == '_') return 0;
5236          if (id == '~') return 1;
5237      }
5238      else {
5239          REALLOC_N(lvtbl->tbl, ID, lvtbl->cnt+2);
5240      }
5242      lvtbl->tbl[lvtbl->cnt+1] = id;
5243      return lvtbl->cnt++;
5244  }


It seems definitely true. lvtbl->tbl is an array of the local variable names, and its index – 1 is the return value (local variable ID).

Note that it increases lvtbl->cnt. Since the code to increase lvtbl->cnt only exists here, from only this code its meaning can be decided. Then, what is the meaning? It is, since “lvtbl->cnt increases by 1 when a new variable is added”, “lvtbl->cnt holds the number of local variables in this scope”.

Finally, I’ll explain about tbl[1] and tbl[2]. These '_' and '~' are, as you can guess if you are familiar with Ruby, the special variables named $_ and $~. Though their appearances are identical to global variables, they are actually local variables. Even If you didn’t explicitly use it, when the methods such as Kernel#gets are called, these variables are implicitly assigned, thus it’s necessary that the spaces are always allocated.

Summary of local variables

Since the description of local variables were complex in various ways, let’s summarize it.

First, It seems the local variables are different from the other variables because they are not managed with st_table. Then, where are they stored in? It seems the answer is an array. Moreover, it is stored in a different array for each scope.

The array is lvtbl->tbl, and the index 0 holds the lvtbl->cnt which is set at local_pop(). In other words, it holds the number of the local variables. The index 1 or more hold the local variable names defined in the scope. Fig.11 shows the final appearance we expect.

Fig.11: correspondences between local variable names and the return values

Block Local Variables

The rest is dyna_vars which is a member of struct local_vars. In other words, this is about the block local variables. I thought that there must be the functions to do something with this, looked over the list of the function names, and found them as expected. There are the suspicious functions named dyna_push() dyna_pop() dyna_in_block(). Moreover, here is the place where these are used.

▼ an example using dyna_push dyna_pop

1651  brace_block     : '{'
1652                      {
1653                          $<vars>$ = dyna_push();
1654                      }
1655                    opt_block_var
1656                    compstmt '}'
1657                      {
1658                          $$ = NEW_ITER($3, 0, $4);
1659                          fixpos($$, $4);
1660                          dyna_pop($<vars>2);
1661                      }


push at the beginning of an iterator block, pop at the end. This must be the process of block local variables.

Now, we are going to look at the functions.


5331  static struct RVarmap*
5332  dyna_push()
5333  {
5334      struct RVarmap* vars = ruby_dyna_vars;
5336      rb_dvar_push(0, 0);
5337      lvtbl->dlev++;
5338      return vars;
5339  }


Increasing lvtbl->dlev seems the mark indicates the existence of the block local variable scope. Meanwhile, rb_dvar_push() is …


 691  void
 692  rb_dvar_push(id, value)
 693      ID id;
 694      VALUE value;
 695  {
 696      ruby_dyna_vars = new_dvar(id, value, ruby_dyna_vars);
 697  }


It creates a struct RVarmap that has the variable name id and the value val as its members, adds it to the top of the global variable ruby_dyna_vars. This is again and again the form of cons. In dyna_push(), ruby_dyan_vars is not set aside, it seems it adds directly to the ruby_dyna_vars of the previous scope.

Moreover, the value of the id member of the RVarmap to be added here is 0. Although it was not seriously discussed in this book, the ID of ruby will never be 0 while it is normally created by rb_intern(). Thus, we can infer that this RVarmap, as it is like NUL or NULL, probably has a role as sentinel. If we think based on this assumption, we can describe the reason why the holder of a variable (RVarmap) is added even though not any variables are added.

Next, dyna_pop().


5341  static void
5342  dyna_pop(vars)
5343      struct RVarmap* vars;
5344  {
5345      lvtbl->dlev--;
5346      ruby_dyna_vars = vars;
5347  }


By reducing lvtbl->dlev, it writes down the fact that the block local variable scope ended. It seems that something is done by using the argument, let’s see this later at once.

The place to add a block local variable has not appeared yet. Something like local_cnt() of local variables is missing. So, I did plenty of grep with dvar and dyna, and this code was found.

assignable() (partial)

4599  static NODE*
4600  assignable(id, val)
4601      ID id;
4602      NODE *val;
4603  {
4634              rb_dvar_push(id, Qnil);
4635              return NEW_DASGN_CURR(id, val);


assignable() is the function to create a node relates to assignments, this citation is the fragment of that function only contains the part to deal with block local variables. It seems that it adds a new variable (to ruby_dyna_vars) by using rb_dvar_push() that we’ve just seen.

ruby_dyna_vars in the parser

Now, taking the above all into considerations, let’s imagine the appearance of ruby_dyna_vars at the moment when a local variable scope is finished to be parsed.

First, as I said previously, the RVarmap of id=0 which is added at the beginning of a block scope is a sentinel which represents a break between two block scopes. We’ll call this “the header of ruby_dyna_vars”.

Next, among the previously shown actions of the rule of the iterator block, I’d like you to focus on this part:

$<vars>$ = dyna_push();    /* what assigned into $<vars>$ is ... */
dyna_pop($<vars>2);        /* …… appears at $<vars>2 */

dyna_push() returns the ruby_dyna_vars at the moment. dyna_pop() put the argument into ruby_dyna_vars. This means ruby_dyna_vars would be saved and restored for each the block local variable scope. Therefore, when parsing the following program,

iter {
    a = nil
    iter {
        b = nil
        iter {
            c = nil
            # nesting level 3
        bb = nil
        # nesting level 2
        iter {
            e = nil
    # nesting level 1

Fig.12 shows the ruby_dyna_vars in this situation.

Fig.12: ruby_dyna_vars when all scopes are finished to be parsed

This structure is fairly smart. That’s because the variables of the higher levels can naturally be accessed by traversing over all of the list even if the nesting level is deep. This way has the simpler searching process than creating a different table for each level.

Plus, in the figure, it looks like bb is hung at a strange place, but this is correct. When a variable is found at the nest level which is decreased after increased once, it is attached to the subsequent of the list of the original level. Moreover, in this way, the specification of local variable that “only the variables which already exist in the symbol sequence are defined” is expressed in a natural form.

And finally, at each cut of local variable scopes (this is not of block local variable scopes), this link is entirely saved or restored to lvtbl->dyna_vars. I’d like you to go back a little and check local_push() and local_pop().

By the way, although creating the ruby_dyna_vars list was a huge task, it is by itself not used at the evaluator. This list is used only to check the existence of the variables and will be garbage collected at the same moment when parsing is finished. And after entering the evaluator, another chain is created again. There’s a quite deep reason for this, … we’ll see around this once again in Part 3.