The Perl Toolchain Summit needs more sponsors. If your company depends on Perl, please support this very important event.

NAME

docs/pdds/pdd08_keys.pod - Indexing Aggregate PMCs

ABSTRACT

This PDD aims to clear up the confusion regarding the implementation of keyed access to PMCs in Parrot.

DESCRIPTION

First, let's define some terminology. An aggregate PMC is one which allows indexed access to sub-element that it stores or references, by supporting and implementing the _keyed variants of vtable methods. These variants are indexing operations, as they act on a specific indexed element of an aggregate PMC.

Non-aggregates also support _keyed variants of the vtable methods, but they may not do anything particularly clever - for instance, PMC types implementing Perl references will merely pass the index on to the referent. These aren't aggregates because they don't directly store or reference sub-elements.

Indexing operations take one or more aggregate-key atoms. At runtime, these operations will index the key into the aggregate returning a value. Here is a well-known indexing operation in Perl 6:

    @a[12] = $b;

The key here is the constant integer 12, and the aggregate is the PerlArray @a. In the process of this assignment, Parrot will have to extract the PMC in element 12 of the array, producing a value. $b is then assigned to this value.

Now, how does this all get implemented?

IMPLEMENTATION

The key structure

The key structure has to have the following features: it must bundle up multiple keys so that multidimensional aggregates can be indexed; keys may be specified as integer, string, number or PMC.

Hence, the following structure was produced. First, the individual keys as we think of them from a language level are each stored in a Key PMC with the key type encoded in the private flags of the PMC and the value stored in the PMCs cache.

So, for instance, indexing the multidimensional array @foo[$a;12;"hi"] produces three PMCs, one with a PMC type, one with an integer type and one with a string type.

As stated, the key type is encoded in the PMC flags using 5 bits based on the following scheme:

    typedef enum {
        KEY_integer_FLAG = PMC_private0_FLAG,
        KEY_number_FLAG = PMC_private1_FLAG,
        KEY_string_FLAG = PMC_private2_FLAG,
        KEY_pmc_FLAG = PMC_private3_FLAG,
        KEY_register_FLAG = PMC_private4_FLAG,

        KEY_type_FLAGS = KEY_integer_FLAG |
                         KEY_number_FLAG |
                         KEY_string_FLAG |
                         KEY_pmc_FLAG |
                         KEY_register_FLAG
    } KEY_flags;

The KEY_register_FLAG is used to indicate that the int_val component of the PMC's cache contains the number of a register of the appropriate type that itself contains the value of the key component.

The next issue is to combine these things into a single key. This is done by using the data member of the PMC to form a linked list so that each key points to the next key. A linked list is used so that partial keys can be easily generated as a multidimensional data structure is traversed.

These definitions, along with declarations of support routines used to manipulate keys, can be found in include/parrot/key.h

Aggregate and non-aggregate PMCs

We've already said that what separates the aggregate PMCs from the non-aggregates is their implementation of the _keyed vtable methods. So it is Hereby Decreed that the default vtable which everyone inherits from defines the _keyed forms to throw an exception.

Todo

Need discussion on whether OUT_OF_BOUNDS is a good exception for this, or whether something else should be used. It's really a compiler screw-up, since code which indexes a non-aggregate shouldn't be generated.

_keyed vtable methods

So what of these magical _keyed vtable methods? They are generated when you add the keyed tag to the appropriate entry in vtable.tbl. They are constructed by following every PMC argument with a second PMC argument which acts as the key for that argument; the name of the second PMC argument is formed by adding _key onto the end of the first PMC argument.

The reason why every PMC argument has an associated key is twofold. Firstly, it means that

    @a[$b] = $c

and

    $a = @b[$c]

use the same vtable method, reducing the multiplicity of methods. Secondly, a three-argument assign as suggested by the code above would be ambiguous - the code above uses 3 PMCs in different ways.

Also, operations which take an aggregate key for one of their arguments should take aggregate keys for all of their arguments. This is to avoid the following:

    void foo_keyed_i(PMC* x, PMC* x_key, INT a)
    void foo_keyed_n(PMC* x, PMC* x_key, NUM a)
    void foo_keyed_p(PMC* x, PMC* x_key, PMC a)
    void foo_keyed_p_keyed(PMC* x, PMC* x_key, PMC* a, PMC* a_key)

These are all replaced with the single entry

    void foo_keyed(PMC* x, PMC* a_key, PMC* a, PMC* a_key)

(Think how much worse it gets when there are three or more PMCs in an entry...)

Yes. This means that you may need to turn some things into PMCs that you didn't want to. Since the alternative is mega pollution and duplication in the vtable table, and since the majority of things that you'll deal with in a real world situation are expected to be PMCs anyway, this shouldn't be too much of a problem.

So, if you have a PMC in a _keyed method which you don't want to index, pass in NULL instead of a real key. Code implementing these methods should understand PMC* foo, PMC* NULL as meaning the entirety of foo in some sense; this is trivial to understand if foo is non-aggregate, and implementation-defined if foo is aggregate. If you remember that a key PMC is really a linked list, you'll notice that after traversing down through the list, you'll reach a NULL which again means the entirety of whatever object you traversed to.

Similarly, non-_keyed methods on aggregates are implementation defined; for instance, a set_integer on a PerlArray may be understood as setting @array.length.

Historically, we first implemented keys as two separate keyed methods per applicable method - ..._index and ..._index_s for integer and string indexing respectively. However, this didn't give us the flexibility and scalability that key structures give us.

Input to the assembler

There are several different valid specifications of an aggregate key to the assembler. These are:

    op arg, P1[1234]  # Constant integer key
    op arg, P1[I1]    # Integer key

    op arg, P1[12.34] # Constant number key - handled as constant key
    op arg, P1["foo"] # Constant string key - handled as constant key
    op arg, P1[I1;I2] # Multi-level key - handled as constant key

    op arg, P1[P1]    # Register key

(Rationale: fits programmer's expectation, easier to understand at a glance than op P1, P2, P3. Also, is op P1, P2, P3 the same as op P1[P2], P3 or op P1, P2[P3], or are these three separate PMCs?)

In all there are four types of key. The first two are integer keys and constant integer keys which are optimisations for the common case of single level integer keys.

The other two are constant keys, which can handle any combination of constants and registers with any number of levels; and register keys which are represented by a single PMC register that is assumed to contain a PMC of the Key class.

What the assembler did next

When the assembler sees an aggregate key, it "detaches" the key to form a separate argument. It then decides on the type of key. For integer keys (both constant and register) the data is encoded in the same way as an ordinary integer argument. For register keys the data is encoded as for an ordinary PMC register argument, while for constant keys a key constant is generated that encodes the list of constants and registers that make up the key and an appropriate index into the constant table is encoded as the argument.

Next it selects the appropriate op. Register keys have the signature k and constant keys have the signature kc. Integer register and constant keys are encoded as ki and kic respectively.

    set P1["hi"], 1234

finds an op named set_p_kc_i. On the other hand,

    set P1[P1], 1234

produces an op named set_p_k_i. Likewise, this:

    set P1[1], 1234

produces an op named set_p_kic, and this:

    set P1[I1], 1234

produces an op named set_p_ki.

Bytecode representation

The bytecode representation of these keys are as follows: constant keys are treated just like another constant, and are an index into the packfile's constant table.

Each key in that constant table consists of one word specifying its length in terms of number of keys. For instance, ["hi"] has length 1; ["hi";P1;S1;123] has length 4. Next, each key is specified using two words. The first word is a type specifier:

    1 - Integer constant
    2 - Number constant
    4 - String constant
    7 - Integer register
    8 - Number register
    9 - PMC register
   10 - String register

and the second word is either a value (for integer constants), a register number (for registers) or an index into the appropriate constant table.

The type values shown above are actually the PARROT_ARG_* values taken from include/parrot/op.h.

VERSION

CURRENT

   Maintainer: Simon Cozens <simon@netthink.co.uk>
   Class: Internals
   PDD Number: 8
   Version: 1.3
   Status: Developing
   Last Modified: 25 August, 2002
   PDD Format: 1
   Language: English

HISTORY

Sun Aug 25 11:14:43 GMT 2002 : Version 1.3

Updated to reflect Dan's decision to change keys to use PMCs instead of a custom data structure. Also corrects documentation of multi-level keys and how they are compiled and work. tom@compton.nu.

Thu Apr 25 18:30:36 UTC 2002 : Version 1.2

Renamed KEY_PAIR to KEY_ATOM, updated to reflect changeover to linked list. - steve@fink.com

Fri Mar 8 18:47:34 GMT 2002 : Version 1.1

updated to reflect Dan's comments that non-aggregates also support _keyed variant vtable methods.