B::Concise - Walk Perl syntax tree, printing concise info about ops


NAME

B::Concise - Walk Perl syntax tree, printing concise info about ops


SYNOPSIS

    perl -MO=Concise[,OPTIONS] foo.pl
    use B::Concise qw(set_style add_callback);


DESCRIPTION

This compiler backend prints the internal OPs of a Perl program's syntax tree in one of several space-efficient text formats suitable for debugging the inner workings of perl or other compiler backends. It can print OPs in the order they appear in the OP tree, in the order they will execute, or in a text approximation to their tree structure, and the format of the information displyed is customizable. Its function is similar to that of perl's -Dx debugging flag or the B::Terse module, but it is more sophisticated and flexible.


EXAMPLE

Here's is a short example of output (aka 'rendering'), using the default formatting conventions :

    % perl -MO=Concise -e '$a = $b + 42'
    8  <@> leave[1 ref] vKP/REFC ->(end)
    1     <0> enter ->2
    2     <;> nextstate(main 1 -e:1) v ->3
    7     <2> sassign vKS/2 ->8
    5        <2> add[t1] sK/2 ->6
    -           <1> ex-rv2sv sK/1 ->4
    3              <$> gvsv(*b) s ->4
    4           <$> const(IV 42) s ->5
    -        <1> ex-rv2sv sKRM*/1 ->7
    6           <$> gvsv(*a) s ->7

Each line corresponds to an opcode. Null ops appear as ex-opname, where opname is the op that has been optimized away by perl.

The number on the first row indicates the op's sequence number. It's given in base 36 by default.

The symbol between angle brackets indicates the op's type : for example, <2> is a BINOP, <@> a LISTOP, etc. (see OP class abbreviations).

The opname may be followed by op-specific information in parentheses (e.g. gvsv(*b)), and by targ information in brackets (e.g. leave[t1]).

Next come the op flags. The common flags are listed below (OP flags abbreviations). The private flags follow, separated by a slash. For example, vKP/REFC means that the leave op has public flags OPf_WANT_VOID, OPf_KIDS, and OPf_PARENS, and the private flag OPpREFCOUNTED.

Finally an arrow points to the sequence number of the next op.


OPTIONS

Arguments that don't start with a hyphen are taken to be the names of subroutines to print the OPs of; if no such functions are specified, the main body of the program (outside any subroutines, and not including use'd or require'd files) is printed. Passing BEGIN, CHECK, INIT, or END will cause all of the corresponding special blocks to be printed.

Options affect how things are rendered (ie printed). They're presented here by their visual effect, 1st being strongest. They're grouped according to how they interrelate; within each group the options are mutually exclusive (unless otherwise stated).

Options for Opcode Ordering

These options control the 'vertical display' of opcodes. The display 'order' is also called 'mode' elsewhere in this document.

-basic
Print OPs in the order they appear in the OP tree (a preorder traversal, starting at the root). The indentation of each OP shows its level in the tree. This mode is the default, so the flag is included simply for completeness.

-exec
Print OPs in the order they would normally execute (for the majority of constructs this is a postorder traversal of the tree, ending at the root). In most cases the OP that usually follows a given OP will appear directly below it; alternate paths are shown by indentation. In cases like loops when control jumps out of a linear path, a 'goto' line is generated.

-tree
Print OPs in a text approximation of a tree, with the root of the tree at the left and 'left-to-right' order of children transformed into 'top-to-bottom'. Because this mode grows both to the right and down, it isn't suitable for large programs (unless you have a very wide terminal).

Options for Line-Style

These options select the line-style (or just style) used to render each opcode, and dictates what info is actually printed into each line.

-concise
Use the author's favorite set of formatting conventions. This is the default, of course.

-terse
Use formatting conventions that emulate the output of B::Terse. The basic mode is almost indistinguishable from the real B::Terse, and the exec mode looks very similar, but is in a more logical order and lacks curly brackets. B::Terse doesn't have a tree mode, so the tree mode is only vaguely reminiscent of B::Terse.

-linenoise
Use formatting conventions in which the name of each OP, rather than being written out in full, is represented by a one- or two-character abbreviation. This is mainly a joke.

-debug
Use formatting conventions reminiscent of B::Debug; these aren't very concise at all.

-env
Use formatting conventions read from the environment variables B_CONCISE_FORMAT, B_CONCISE_GOTO_FORMAT, and B_CONCISE_TREE_FORMAT.

Options for tree-specific formatting

-compact
Use a tree format in which the minimum amount of space is used for the lines connecting nodes (one character in most cases). This squeezes out a few precious columns of screen real estate.

-loose
Use a tree format that uses longer edges to separate OP nodes. This format tends to look better than the compact one, especially in ASCII, and is the default.

-vt
Use tree connecting characters drawn from the VT100 line-drawing set. This looks better if your terminal supports it.

-ascii
Draw the tree with standard ASCII characters like + and |. These don't look as clean as the VT100 characters, but they'll work with almost any terminal (or the horizontal scrolling mode of less(1)) and are suitable for text documentation or email. This is the default.

These are pairwise exclusive, i.e. compact or loose, vt or ascii.

Options controlling sequence numbering

-basen
Print OP sequence numbers in base n. If n is greater than 10, the digit for 11 will be 'a', and so on. If n is greater than 36, the digit for 37 will be 'A', and so on until 62. Values greater than 62 are not currently supported. The default is 36.

-bigendian
Print sequence numbers with the most significant digit first. This is the usual convention for Arabic numerals, and the default.

-littleendian
Print seqence numbers with the least significant digit first. This is obviously mutually exclusive with bigendian.

Other options

-main
Include the main program in the output, even if subroutines were also specified. This is the only option that is not sticky (see below)

-banner
B::Concise::compile normally prints a banner line identifying the function name, or in case of a subref, a generic message including (unfortunately) the stringified coderef. This option suppresses the printing of the banner.

Option Stickiness

If you invoke Concise more than once in a program, you should know that the options are 'sticky'. This means that the options you provide in the first call will be remembered for the 2nd call, unless you re-specify or change them.


FORMATTING SPECIFICATIONS

For each line-style ('concise', 'terse', 'linenoise', etc.) there are 3 format-specs which control how OPs are rendered.

The first is the 'default' format, which is used in both basic and exec modes to print all opcodes. The 2nd, goto-format, is used in exec mode when branches are encountered. They're not real opcodes, and are inserted to look like a closing curly brace. The tree-format is tree specific.

When a line is rendered, the correct format string is scanned for the following items, and data is substituted in, or other manipulations, like basic indenting. Any text that doesn't match a special pattern (the items below) is copied verbatim. (Yes, it's a set of s///g steps.)

(x(exec_text;basic_text)x)
Generates exec_text in exec mode, or basic_text in basic mode.

(*(text)*)
Generates one copy of text for each indentation level.

(*(text1;text2)*)
Generates one fewer copies of text1 than the indentation level, followed by one copy of text2 if the indentation level is more than 0.

(?(text1#varText2)?)
If the value of var is true (not empty or zero), generates the value of var surrounded by text1 and Text2, otherwise nothing.

#var
Generates the value of the variable var.

#varN
Generates the value of var, left jutified to fill N spaces.

~
Any number of tildes and surrounding whitespace will be collapsed to a single space.

The following variables are recognized:

#addr
The address of the OP, in hexidecimal.

#arg
The OP-specific information of the OP (such as the SV for an SVOP, the non-local exit pointers for a LOOP, etc.) enclosed in paretheses.

#class
The B-determined class of the OP, in all caps.

#classsym
A single symbol abbreviating the class of the OP.

#coplabel
The label of the statement or block the OP is the start of, if any.

#exname
The name of the OP, or 'ex-foo' if the OP is a null that used to be a foo.

#extarg
The target of the OP, or nothing for a nulled OP.

#firstaddr
The address of the OP's first child, in hexidecimal.

#flags
The OP's flags, abbreviated as a series of symbols.

#flagval
The numeric value of the OP's flags.

#hyphseq
The sequence number of the OP, or a hyphen if it doesn't have one.

#label
'NEXT', 'LAST', or 'REDO' if the OP is a target of one of those in exec mode, or empty otherwise.

#lastaddr
The address of the OP's last child, in hexidecimal.

#name
The OP's name.

#NAME
The OP's name, in all caps.

#next
The sequence number of the OP's next OP.

#nextaddr
The address of the OP's next OP, in hexidecimal.

#noise
A one- or two-character abbreviation for the OP's name.

#private
The OP's private flags, rendered with abbreviated names if possible.

#privval
The numeric value of the OP's private flags.

#seq
The sequence number of the OP. Note that this is now a sequence number generated by B::Concise, rather than the real op_seq value (for which see #seqnum).

#seqnum
The real sequence number of the OP, as a regular number and not adjusted to be relative to the start of the real program. (This will generally be a fairly large number because all of B::Concise is compiled before your program is).

#sibaddr
The address of the OP's next youngest sibling, in hexidecimal.

#svaddr
The address of the OP's SV, if it has an SV, in hexidecimal.

#svclass
The class of the OP's SV, if it has one, in all caps (e.g., 'IV').

#svval
The value of the OP's SV, if it has one, in a short human-readable format.

#targ
The numeric value of the OP's targ.

#targarg
The name of the variable the OP's targ refers to, if any, otherwise the letter t followed by the OP's targ in decimal.

#targarglife
Same as #targarg, but followed by the COP sequence numbers that delimit the variable's lifetime (or 'end' for a variable in an open scope) for a variable.

#typenum
The numeric value of the OP's type, in decimal.


ABBREVIATIONS

OP flags abbreviations

    v      OPf_WANT_VOID    Want nothing (void context)
    s      OPf_WANT_SCALAR  Want single value (scalar context)
    l      OPf_WANT_LIST    Want list of any length (list context)
    K      OPf_KIDS         There is a firstborn child.
    P      OPf_PARENS       This operator was parenthesized.
                             (Or block needs explicit scope entry.)
    R      OPf_REF          Certified reference.
                             (Return container, not containee).
    M      OPf_MOD          Will modify (lvalue).
    S      OPf_STACKED      Some arg is arriving on the stack.
    *      OPf_SPECIAL      Do something weird for this op (see op.h)

OP class abbreviations

    0      OP (aka BASEOP)  An OP with no children
    1      UNOP             An OP with one child
    2      BINOP            An OP with two children
    |      LOGOP            A control branch OP
    @      LISTOP           An OP that could have lots of children
    /      PMOP             An OP with a regular expression
    $      SVOP             An OP with an SV
    "      PVOP             An OP with a string
    {      LOOP             An OP that holds pointers for a loop
    ;      COP              An OP that marks the start of a statement
    #      PADOP            An OP with a GV on the pad


Using B::Concise outside of the O framework

You can use B::Concise, and call compile() directly, and repeatedly. By doing so, you can avoid the compile-time only operation of 'perl -MO=Concise ..'. For example, you can use the debugger to step through B::Concise::compile() itself.

When doing so, you can alter Concise output by providing new output styles, and optionally by adding callback routines which populate new variables that may be rendered as part of those styles. For all following sections, please review FORMATTING SPECIFICATIONS.

Example: Altering Concise Renderings

    use B::Concise qw(set_style add_callback);
    set_style($your_format, $your_gotofmt, $your_treefmt);
    add_callback
      ( sub {
            my ($h, $op, $format, $level, $stylename) = @_;
            $h->{variable} = some_func($op);
        }
      );
    B::Concise::compile(@options)->();

set_style()

set_style accepts 3 arguments, and updates the three format-specs comprising a line-style (basic-exec, goto, tree). It has one minor drawback though; it doesn't register the style under a new name. This can become an issue if you render more than once and switch styles. Thus you may prefer to use add_style() and/or set_style_standard() instead.

set_style_standard($name)

This restores one of the standard line-styles: terse, concise, linenoise, debug, env, into effect. It also accepts style names previously defined with add_style().

add_style()

This subroutine accepts a new style name and three style arguments as above, and creates, registers, and selects the newly named style. It is an error to re-add a style; call set_style_standard() to switch between several styles.

add_callback()

If your newly minted styles refer to any #variables, you'll need to define a callback subroutine that will populate (or modify) those variables. They are then available for use in the style you've chosen.

The callbacks are called for each opcode visited by Concise, in the same order as they are added. Each subroutine is passed five parameters.

  1. A hashref, containing the variable names and values which are
     populated into the report-line for the op
  2. the op, as a B<B::OP> object
  3. a reference to the format string
  4. the formatting (indent) level
  5. the selected stylename

To define your own variables, simply add them to the hash, or change existing values if you need to. The level and format are passed in as references to scalars, but it is unlikely that they will need to be changed or even used.

Running B::Concise::compile()

compile accepts options as described above in OPTIONS, and arguments, which are either coderefs, or subroutine names.

compile() constructs and returns a coderef, which when invoked, scans the optree, and prints the results to STDOUT. Once you have the coderef, you may change the output style; thereafter the coderef renders in the new style.

walk_output lets you change the print destination from STDOUT to another open filehandle, or into a string passed as a ref.

    walk_output(\my $buf);
    my $walker = B::Concise::compile('-concise','funcName', \&aSubRef);
    print "Concise Banner for Functions: $buf\n";
    $walker->();
    print "Concise Rendering(s)?: $buf\n";

For each subroutine visited by Concise, the $buf will contain a banner naming the function or coderef about to be traversed. Once $walker is invoked, it prints the actual renderings for each.

To switch back to one of the standard styles like concise or terse, call set_style_standard, or pass the style name into B::Concise::compile() (as done above).

B::Concise::reset_sequence()

This function (not exported) lets you reset the sequence numbers (note that they're numbered arbitrarily, their goal being to be human readable). Its purpose is mostly to support testing, i.e. to compare the concise output from two identical anonymous subroutines (but different instances). Without the reset, B::Concise, seeing that they're separate optrees, generates different sequence numbers in the output.

Errors

All detected errors, (invalid arguments, internal errors, etc.) are resolved with a die($message). Use an eval if you wish to catch these errors and continue processing.

In particular, compile will die if you've asked for a non-existent function-name, a non-existent coderef, or a non-CODE reference.


AUTHOR

Stephen McCamant, <smcc@CSUA.Berkeley.EDU>.

 B::Concise - Walk Perl syntax tree, printing concise info about ops