Perl 5 Internals - Part Four

The first three installments of this series covered Perl's core data types: scalars, arrays, and hashes. This final installment will cover something a bit different: the optree. Those of you who are familiar with compiler concepts are no doubt familiar with the notion of an abstract syntax tree (known as an AST for short). The optree is perl's take on the AST: it's something similar to, but not entirely the same as, an AST. Before we begin looking at the optree, I recommend reviewing the "Subroutines" and "Compiled code" sections of perlguts, as well as looking at perlcall. It is by no means required, but it might make digesting this content a little easier.

Inspecting the Optree

Just as we can use Devel::Peek to dig into the internals of Perl data structures, we can use B::Concise to dig into the optree:

# The 'O' module specifies an alternative compiler backend
$ perl -MO=Concise -E 'say q{hi!}'

Now that we know how to look at the optree, let's have a look at the output of that command:

6  <@> leave[1 ref] vKP/REFC ->(end)
1     <0> enter ->2
2     <;> nextstate(main 47 -e:1) v:%,{,469764096 ->3
5     <@> say vK ->6
3        <0> pushmark s ->4
4        <$> const(PV "hi!") s ->5
-e syntax OK

As you can see from this little snippet, the output from B::Concise is a little harder to interpret than that of Devel::Peek. The exact format is documented in the perldoc for B::Concise, but we can gloss over the details for the purposes of this introduction. First of all, we can see that the operations are organized in a tree-like structure (as you would expect from something called the optree). We can also pick a few things out of the individual lines. Take for example, the second line of output:

1     <0> enter ->2

enter is the OP name, and <nowiki>->2</nowiki> means that the OP to be executed after this one is OP #2. This is what perl calls the "execution order thread" (not to be confused with threads as in multithreading). Because of this, when perl executes an optree, it only needs to follow the execution order thread instead of having to perform a tree traversal.

While the functionality of some of the OPs are obvious from their names, the OP nextstate isn't as obvious. The purpose of the nextstate OP is simple: it's placed in between statements as a sort of "bookkeeping" operation; to clean up temporaries and let perl know where in the program it is for diagnostic reasons.

We can also tell B::Concise to dump the optree in execution order:

$ perl -MO=Concise,-exec -E 'say q{hi!}'
1  <0> enter
2  <;> nextstate(main 47 -e:1) v:%,{,469764096
3  <0> pushmark s
4  <$> const(PV "hi!") s
5  <@> say vK
6  <@> leave[1 ref] vKP/REFC
-e syntax OK

Here, you can see the the tree structure in the formatted output has gone away, but now we can see the instructions laid out in the order perl would execute them. B::Concise has many handy options; consult the documentation for more information. I personally like the -src and -base options; for the remainder of this post, I'll be using the -base option to make the output a little easier to read.

Looking at Conditionals

Now that we've covered a very basic program, let's look at a conditional. Conditionals work a little bit differently, considering that they need two execution order thread pointers, one of which is used depending on whether the condition is true or false at runtime. Observe:

$ perl -MO=Concise,-base10 -E 'my $value = 1; if($value) { say q{hi!} }'
12 <@> leave[1 ref] vKP/REFC ->(end)
1     <0> enter ->2
2     <;> nextstate(main 47 -e:1) v:%,{,469764096 ->3
5     <2> sassign vKS/2 ->6
3        <$> const(IV 1) s ->4
4        <0> padsv[$value:47,50] sRM*/LVINTRO ->5
6     <;> nextstate(main 50 -e:1) v:%,{,469764096 ->7
-     <1> null vK/1 ->12
8        <|> and(other->9) vK/1 ->12
7           <0> padsv[$value:47,50] s ->8
-           <@> scope vK ->-
-              <0> ex-nextstate v ->9
11             <@> say vK ->12
9                 <0> pushmark s ->10
10                <$> const(PV "hi!") s ->11
-e syntax OK

Here we can see a scalar assignment (the sassign, corresponding to the my $value = 1 statement) and an and OP, which is used to implement the if statement. If you look closely at the line containing the and OP, you can see it has two pointers to other OPs; the truthy path, and the falsy one.

Some optimizations

Let's see how perl handles an array:

$ perl -MO=Concise,-base10 -E 'my @array = (1); my $index = 0; say($array[0]); say($array[$index])'
21 <@> leave[1 ref] vKP/REFC ->(end)
1     <0> enter ->2
2     <;> nextstate(main 47 -e:1) v:%,{,469764096 ->3
6     <2> aassign[t2] vKS ->7
-        <1> ex-list lK ->5
3           <0> pushmark s ->4
4           <$> const(IV 1) sP ->5
-        <1> ex-list lK ->6
5           <0> padrange[@array:47,49] l/LVINTRO,1 ->6
-           <0> padav[@array:47,49] lRM*/LVINTRO ->-
7     <;> nextstate(main 48 -e:1) v:%,{,469764096 ->8
10    <2> sassign vKS/2 ->11
8        <$> const(IV 0) s ->9
9        <0> padsv[$index:48,49] sRM*/LVINTRO ->10
11    <;> nextstate(main 49 -e:1) v:%,{,469764096 ->12
14    <@> say vK ->15
12       <0> pushmark s ->13
-        <1> ex-aelem sK/2 ->14
13          <0> aelemfast_lex[@array:47,49] sR ->14
-           <0> ex-const s ->-
15    <;> nextstate(main 49 -e:1) v:%,{,469764096 ->16
20    <@> say vK ->21
16       <0> pushmark s ->17
19       <2> aelem sK/2 ->20
17          <0> padav[@array:47,49] sR ->18
18          <0> padsv[$index:48,49] s ->19
-e syntax OK

The size of the output that B::Concise is giving us is starting to look a bit daunting, but fortunately we don't have to look at all of it. Let's focus in on the array element accesses:

# $array[0]
-        <1> ex-aelem sK/2 ->14
13          <0> aelemfast_lex[@array:47,49] sR ->14
-           <0> ex-const s ->-
# $array[$index]
19       <2> aelem sK/2 ->20
17          <0> padav[@array:47,49] sR ->18
18          <0> padsv[$index:48,49] s ->19

There are two main differences between these subtrees that I'd like to point out. The first is that when using a variable for the index, we see aelem, and when using a literal, we see aelemefast_lex. Perl has a special OP for accessing an array's elements quickly, since it does this so often. Benchmarking the two is left as an exercise for you, dear reader!

The second difference is that optree for $array[0] has a few of its OPs prefixed with "ex-". What this means is that perl has optimized those nodes out of the execution order thread, and will no longer execute them. You might be wondering why perl doesn't remove them outright; that's because that by the time the optimizer runs, the nodes have formed a complicated structure, making the task difficult. Having the extra nodes also provides extra information for people analyzing it (like now, when we're inspecting the optree by hand), and for programs (like when you use B::Deparse to reconstruct the program).

Playing with the optree further

As much as I'd like to keep digging into the optree, it's a fairly complex part of the Perl internals, as you can see. I'm hoping that now that you've seen some of the basics, you are eager to try using B::Concise on other code. As an exercise, I recommend looking at the following:

  1. Subroutine argument handling with @_ and shift;
  2. Hash accesses
  3. Subroutine/method calls

Please feel free to experiment and share your findings with the community!


My coworkers Steffen and Yves shared an important sentiment with us during the training: most of these little tricks aren't documented, nor are they immediately clear from reading the code. Only after studying the code for a bit can they be intuited.

I hope that these posts have engendered an interest in how the Perl interpreter works among those who read them; going through the training and writing these posts certainly has my interest piqued!

Published on 2013-10-02