docs/pdds/pdd19_pir.pod - Parrot Intermediate Representation


This document outlines the architecture and core syntax of the Parrot Intermediate Representation (PIR).

This document describes PIR, a stable, middle-level language for both compiler and human to target on.




PIR is a stable, middle-level language intended both as a target for the generated output from high-level language compilers, and for human use developing core features and extensions for Parrot.

Basic Syntax ^

A valid PIR program consists of a sequence of statements, directives, comments and empty lines.


A statement starts with an optional label, contains an instruction, and is terminated by a newline (<NL>). Each statement must be on its own line.

  [label:] [instruction] <NL>

An instruction may be either a low-level opcode or a higher-level PIR operation, such as a subroutine call, a method call, a directive, or PIR syntactic sugar.


A directive provides information for the PIR compiler that is outside the normal flow of executable statements. Directives are all prefixed with a ".", as in .local or .sub.


Comments start with # and last until the following newline. PIR also allows comments in Pod format. Comments, Pod content, and empty lines are ignored.


Identifiers start with a letter or underscore, then may contain additionally letters, digits, and underscores. Identifiers don't have any limit on length at the moment, but some sane-but-generous length limit may be imposed in the future (256 chars, 1024 chars?). The following examples are all valid identifiers.


Opcode names are not reserved words in PIR, and may be used as variable names. For example, you can define a local variable named print. [See RT #24251]

{{ NOTE: The use of :: in identifiers is deprecated. [See RT #48735] }}


A label declaration consists of a label name followed by a colon. A label name conforms to the standard requirements for identifiers. A label declaration may occur at the start of a statement, or stand alone on a line, but always within a compilation unit.

A reference to a label consists of only the label name, and is generally used as an argument to an instruction or directive.

A PIR label is accessible only in the compilation unit where it's defined. A label name must be unique within a compilation unit, but it can be reused in other compilation units.

  goto label1

Registers and Variables

There are three ways of referencing Parrot's registers. The first is direct access to a specific register by name In, Sn, Nn, Pn. The second is through a temporary register variable $In, $Sn, $Nn, $Pn. n consists of digit(s) only. There is no limit on the size of n.

The third syntax for accessing registers is through named local variables declared with .local.

  .local pmc foo

The type of a named variable can be int, num, string or pmc, corresponding to the types of registers. No other types are used. [See RT#42769]

The difference between direct register access and register variables or local variables is largely a matter of allocation. If you directly reference P99, Parrot will blindly allocate 100 registers for that compilation unit. If you reference $P99 or a named variable foo, on the other hand, Parrot will intelligently allocate a literal register in the background. So, $P99 may be stored in P0, if it is the only register in the compilation unit.

Constants ^

Constants may be used in place of registers or variables. A constant is not allowed on the left side of an assignment, or in any other context where the variable would be modified.

'single-quoted string constant'

Are delimited by single-quotes ('). They are taken to be ASCII encoded. No escape sequences are processed.

"double-quoted string constants"

Are delimited by double-quotes ("). A " inside a string must be escaped by \. Only 7-bit ASCII is accepted in string constants; to use characters outside that range, specify an encoding in the way below.

<<"heredoc", <<'heredoc'

Heredocs work like single or double quoted strings. All lines up to the terminating delimiter are slurped into the string. The delimiter has to be on its own line, at the beginning of the line and with no trailing whitespace.

Assignment of a heredoc:

  $S0 = <<"EOS"
A heredoc as an argument:

  function(<<"END_OF_HERE", arg)


You may have multiple heredocs within a single statement or directive:

   function(<<'INPUT', <<'OUTPUT', 'some test')
charset:"string constant"

Like above with a character set attached to the string. Valid character sets are currently: ascii (the default), binary, unicode (with UTF-8 as the default encoding), and iso-8859-1.

String escape sequences ^

Inside double-quoted strings the following escape sequences are processed.

  \xhh        1..2 hex digits
  \ooo        1..3 oct digits
  \cX         control char X
  \x{h..h}    1..8 hex digits
  \uhhhh      4 hex digits
  \Uhhhhhhhh  8 hex digits
  \a, \b, \t, \n, \v, \f, \r, \e, \\
encoding:charset:"string constant"

Like above with an extra encoding attached to the string. For example:

  set S0, utf8:unicode:"«"
The encoding and charset gets attached to the string, no further processing is done, specifically escape sequences are not honored.

numeric constants

0x and 0b denote hex and binary constants respectively.

Directives ^

.local <type> <identifier> [:unique_reg]

Define a local name identifier for this compilation unit with the given type. You can define multiple identifiers of the same type by separating them with commas:

  .local int i, j
The optional :unique_reg modifier will force the register allocator to associate the identifier with a unique register for the duration of the compilation unit.

.lex <identifier>, <reg>

Declare a lexical variable that is an alias for a PMC register. For example, given this preamble:

    .lex "$a", $P0
    $P1 = new 'Integer'

    These two opcodes have an identical effect:

    $P0 = $P1
    store_lex "$a", $P1

    And these two opcodes also have an identical effect:

    $P1 = $P0
    $P1 = find_lex "$a"
.const <type> <identifier> = <const>

Define a constant named identifier of type type and assign value const to it. The constant is stored in the constant table of the current bytecode file.

.globalconst <type> <identifier> = <const>

As .const above, but the defined constant is globally accessible.

.namespace <identifier> [deprecated: See RT #48737]

Open a new scope block. This "namespace" is not the same as the .namespace [ <identifier> ] syntax, which is used for storing subroutines in a particular namespace in the global symbol table. This directive is useful in cases such as (pseudocode):

  local x = 1;
  print(x);       # prints 1
  do              # open a new namespace/scope block
    local x = 2;  # this x hides the previous x
    print(x);     # prints 2
  end             # close the current namespace
  print(x);       # prints 1 again
All types of common language constructs such as if, for, while, repeat and such that have nested scopes, can use this directive.

{{ NOTE: this variation of .namespace and .endnamespace are deprecated. They were a hackish attempt at implementing scopes in Parrot, but didn't actually turn out to be useful.}}

.endnamespace <identifier> [deprecated: See RT #48737]

Closes the scope block that was opened with .namespace <identifier>.

.namespace [ <identifier> ; <identifier> ]

Defines the namespace from this point onwards. By default the program is not in any namespace. If you specify more than one, separated by semicolons, it creates nested namespaces, by storing the inner namespace object in the outer namespace's global pad.

.pragma n_operators

Convert arithmethic infix operators to n_infix operations. The unary opcodes abs, not, bnot, bnots, and neg are also changed to use a n_ prefix.

 .pragma n_operators 1
 .sub foo
   $P0 = $P1 + $P2           # n_add $P0, $P1, $P2
   $P2 = abs $P0             # n_abs $P2, $P0
.loadlib "lib_name"

Load the given library at compile time, that is, as soon that line is parsed. See also the loadlib opcode, which does the same at run time.

A library loaded this way is also available at runtime, as if it has been loaded again in :load, so there is no need to call loadlib at runtime.

.HLL <hll_name>, <hll_lib>

Define the HLL for the current file. Takes two string constants. If the string hll_lib isn't empty this compile time pragma also loads the shared lib for the HLL, so that integer type constants are working for creating new PMCs.

.HLL_map <core_type>, <user_type>

Whenever Parrot has to create PMCs inside C code on behalf of the running user program it consults the current type mapping for the executing HLL and creates a PMC of type user_type instead of core_type, if such a mapping is defined. core_type and user_type may be any valid string constant.

For example, with this code snippet ...

  .loadlib 'dynlexpad'

  .HLL "Foo", ""
  .HLL_map 'LexPad', 'DynLexPad'

  .sub main :main
... all subroutines for language Foo would use a dynamic lexpad pmc.

{{ PROPOSAL: stop using integer constants for types RT#45453 }}


  .sub <identifier> [:<flag> ...]
  .sub <quoted string> [:<flag> ...]
Define a compilation unit. All code in a PIR source file must be defined in a compilation unit. See the section Subroutine flags for available flags. Optional flags are a list of flag, separated by empty spaces.

The name of the sub may be either a bare identifier or a quoted string constant. Bare identifiers must be valid PIR identifiers (see Identifiers above), but string sub names can contain any characters, including characters from different character sets (see Constants above).

Always paired with .end.


End a compilation unit. Always paired with .sub.

.line <integer>, <string>

Set the line number and filename to the value specified. This is useful in case the PIR code is generated from some source file, and any error messages should print the source file, not the line number and filename of the generated file.

{{ DEPRECATION NOTE: was <#line <integer <string>>>. See [RT#45857], [RT#43269], and [RT#47141]. }}

Subroutine flags


Define "main" entry point to start execution. If multiple subroutines are marked as :main, the last marked subroutine is entered.


Run this subroutine during the load_bytecode opcode. If multiple subs have the :load pragma, the subs are run in source code order.


Run the subroutine when the program is run directly (that is, not loaded as a module). This is different from :load, which runs a subroutine when a library is being loaded. To get both behaviours, use :init :load.


Do not install this subroutine in the namespace. Allows the subroutine name to be reused.

:multi(Type1, Type2...)

Engage in multiple dispatch with the listed types. See "pdds/pdd27_multi_dispatch.pod" in docs for more information on the multiple dispatch system.


This subroutine is executed immediately after being compiled. (Analagous to BEGIN in perl5.)


Same as :immediate, except that the subroutine is not executed when the compilation of the file that contains the subroutine is triggered by a load_bytecode instruction in another file.

An example. File main.pir contains:

 .sub main
 load_bytecode "foo.pir"
The file foo.pir contains:

 .sub foo :immediate
   print "42"

 .sub bar :postcomp
   print "43"
When executing file foo.pir, it will execute both foo and bar. However, when executing the file main.pir, only foo will be executed.


The marked .sub is a method. In the method body, the object PMC can be referred to with self.


The marked .sub overrides a v-table method. By default, a sub with the same name as a v-table method does not override the v-table method. To specify that there should be no namespace entry (that is, it just overrides the v-table method but is callable as a normal method), use :vtable :anon. To give the v-table method a different name, use :vtable("..."). For example, to have the method ToString also be the v-table method get_string), use :vtable("get_string").


The marked .sub is lexically nested within the sub known by subname.

Directives used for Parrot calling conventions.

.begin_call and .end_call

Directives to start and end a subroutine invocation, respectively.

.begin_return and .end_return

Directives to start and end a statement to return values.

.begin_yield and .end_yield

Directives to start and end a statement to yield values.


Takes either 2 arguments: the sub and the return continuation, or the sub only. For the latter case an invokecc gets emitted. Providing an explicit return continuation is more efficient, if its created outside of a loop and the call is done inside a loop.


Directive to specify the object for a method call. Use it in combination with .meth_call.


Directive to do a method call. It calls the specified method on the object that was specified with the .invocant directive.


Directive to make a call through the Native Calling Interface (NCI). The specified subroutine must be loaded using the <dlfunc> op that takes the library, function name and function signature as arguments. See "pdds/pdd16_native_call" in docs for details.

.return <var> [:<flag>]*

Between .begin_return and .end_return, specify one or more of the return value(s) of the current subroutine. Available flags: :flat, :named.

.arg <var> [:<flag>]*

Between .begin_call and .call, specify an argument to be passed. Available flags: :flat, :named.

.result <var> [:<flag>]*

Between .call and .end_call, specify where one or more return value(s) should be stored. Available flags: :slurpy, :named, :optional, and :opt_flag.

Directives for subroutine parameters

.param <type> <identifier> [:<flag>]*

At the top of a subroutine, declare a local variable, in the manner of .local, into which parameter(s) of the current subroutine should be stored. Available flags: :slurpy, :named, :optional, :opt_flag and :unique_reg.

.param <type> "<identifier>" => <identifier> [:<flag>]*

Define a named parameter. This is syntactic sugar for:

 .param <type> <identifier> :named("<identifier>")

Parameter Passing and Getting Flags

See PDD03 for a description of the meaning of the flag bits SLURPY, OPTIONAL, OPT_FLAG, and FLAT, which correspond to the calling convention flags :slurpy, :optional, :opt_flag, and :flat.

Catching Exceptions

Using the push_eh op you can install an exception handler. If an exception is thrown, Parrot will execute the installed exception handler. In order to retrieve the thrown exception, use the .get_results directive. This directive always takes 2 arguments: an exception object and a message string.

{{ Wouldn't it be more useful to make this flexible, or at least only the exception object? The message can be retrieved from the exception object. }}

   push_eh handler
   .local pmc exception
   .local string message
   .get_results (exception, message)

This is syntactic sugar for the get_results op, but any flags set on the targets will be handled automatically by the PIR compiler.

Syntactic Sugar ^

Any PASM opcode is a valid PIR instruction. In addition, PIR defines some syntactic shortcuts. These are provided for ease of use by humans producing and maintaing PIR code.

goto <identifier>

branch to identifier (label or subroutine name).


  goto END
if <var> goto <identifier>

If var evaluates as true, jump to the named identifier. Translate to if var, identifier.

unless <var> goto <identifier>

Unless var evaluates as true, jump to the named identifier. Translate to unless var, identifier.

if null <var> goto <identifier>

If var evaluates as null, jump to the named identifier. Translate to if_null var, identifier.

unless null <var> goto <identifier>

Unless var evaluates as null, jump to the named identifier. Translate to unless_null var, identifier.

if <var1> <relop> <var2> goto <identifier>

The relop can be: <, <=, ==, != >= > which translate to the PASM opcodes lt, le, eq, ne, ge or gt. If var1 relop var2 evaluates as true, jump to the named identifier.

unless <var1> <relop> <var2> goto <identifier>

The relop can be: <, <=, ==, != >= > which translate to the PASM opcodes lt, le, eq, ne, ge or gt. Unless var1 relop var2 evaluates as true, jump to the named identifier.

<var1> = <var2>

Assign a value. Translates to set var1, var2.

<var1> = <unary> <var2>

The unaries !, - and ~ generate not, neg and bnot ops.

<var1> = <var2> <binary> <var3>

The binaries +, -, *, /, % and ** generate add, sub, mul, div, mod and pow arithmetic ops. binary . is concat and only valid for string arguments.

<< and >> are arithmetic shifts shl and shr. >>> is the logical shift lsr.

&&, || and ~~ are logic and, or and xor.

&, | and ~ are binary band, bor and bxor.

{{PROPOSAL: Change description to support logic operators (comparisons) as implemented (and working) in imcc.y.}}

<var1> <op>= <var2>

This is equivalent to <var1> = <var1> <op> <var2>. Where op is called an assignment operator and can be any of the following binary operators described earlier: +, -, *, /, %, ., &, |, ~, <<, >> or >>>.

<var> = <var> [ <var> ]

This generates either a keyed set operation or substr var, var, var, 1 for string arguments and an integer key.

<var> = <var> [ <key> ]

{{ NOTE: keyed assignment is still valid in PIR, but the .. notation in keys is deprecated [See RT #48561], so this syntactic sugar for slices is also deprecated. See the (currently experimental) slice opcode instead. }}

where key is:

 <var1> .. <var2>
returns a slice defined starting at var1 and ending at var2.

 .. <var2>
returns a slice starting at the first element, and ending at var2.

 <var1> ..
returns a slice starting at var1 to the end of the array.

see src/pmc/slice.pmc and t/pmc/slice.t.

<var> [ <var> ] = <var>

A keyed set operation.

{{ DEPRECATION NOTE: this syntactic sugar will no longer be used for the assign substr op with a length of 1. }}

<var> = <opcode> <arguments>

All opcodes can use this PIR syntactic sugar. The first argument for the opcode is placed before the =, and all remaining arguments go after the opcode name. For example:

  new $P0, 'Type'

  $P0 = new 'Type'
global "string" = <var>

{{ DEPRECATED: op store_global was deprecated }}

<var> = global "string"

{{ DEPRECATED: op find_global was deprecated }}

([<var1> [:<flag1> ...], ...]) = <var2>([<arg1> [:<flag2> ...], ...])

This is short for:

  .arg <arg1> <flag2>
  .call <var2>
  .result <var1> <flag1>
<var> = <var>([arg [:<flag> ...], ...])

<var>([arg [:<flag> ...], ...])

<var>."_method"([arg [:<flag> ...], ...])

<var>._method([arg [:<flag> ...], ...])

Function or method call. These notations are shorthand for a longer PCC function call. var can denote a global subroutine, a local identifier or a reg.

{{We should review the (currently inconsistent) specification of the method name. Currently it can be a bare word, a quoted string or a string register. See #45859.}}

.return ([<var> [:<flag> ...], ...])

Return from the current compilation unit with zero or more values.

The surrounded parentheses are mandatory. Besides making sequence break more conspicuous, this is necessary to distinguish this syntax from other uses of the .return directive that will be probably deprecated.

.return <var>(args)

.return <var>."somemethod"(args)

.return <var>.somemethod(args)

Tail call: call a function or method and return from the sub with the function or method call return values.

Internally, the call stack doesn't increase because of a tail call, so you can write recursive functions and not have stack overflows.

Assignment and Morphing ^

The = syntactic sugar in PIR, when used in the simple case of:

  <var1> = <var2>

directly corresponds to the set opcode. So, two low-level arguments (int, num, or string registers, variables, or constants) are a direct C assignment, or a C-level conversion (int cast, float cast, a string copy, or a call to one of the conversion functions like string_to_num).

A PMC source with a low-level destination, calls the get_integer, get_number, or get_string vtable function on the PMC. A low-level source with a PMC destination calls the set_integer_native, set_number_native, or set_string_native vtable function on the PMC (assign to value semantics). Two PMC arguments are a direct C assignment (assign to container semantics).

For assign to value semantics for two PMC arguments use assign, which calls the assign_pmc vtable function.

{{ NOTE: response to the question:

    <pmichaud>  I don't think that 'morph' as a method call is a good idea
    <pmichaud>  we need something that says "assign to value" versus "assign to container"
    <pmichaud>  we can't eliminate the existing 'morph' opcode until we have a replacement


Macros ^

This section describes the macro layer of the PIR language. The macro layer of the PIR compiler handles the following directives:

The macro layer is completely implemented in the lexical analysis phase. The parser does not know anything about what happens in the lexical analysis phase.

When the .include directive is encountered, the specified file is opened and the following tokens that are requested by the parser are read from that file.

A macro expansion is a dot-prefixed identifier. For instance, if a macro was defined as shown below:

 .macro foo(bar)

this macro can be expanded by writing .foo(42). The body of the macro will be inserted at the point where the macro expansion is written.

A .macro_const expansion is more or less the same as a .macro expansion, except that a constant expansion cannot take any arguments, and the substitution of a .macro_const contains no newlines, so it can be used within a line of code.

Macro parameter list

The parameter list for a macro is specified in parentheses after the name of the macro. Macro parameters are not typed.

 .macro foo(bar, baz, buz)

The number of arguments in the call to a macro must match the number of parameters in the macro's parameter list. Macros do not perform multidispatch, so you can't have two macros with the same name but different parameters. Calling a macro with the wrong number of arguments gives the user an error.

If a macro defines no parameter list, parentheses are optional on both the definition and the call. This means that a macro defined as:

 .macro foo

can be expanded by writing either .foo or .foo(). And a macro definition written as:

 .macro foo()

can also be expanded by writing either .foo or .foo().

{{ NOTE: this is a change from the current implementation, which requires the definition and call of a zero-parameter macro to match in the use of parentheses. }}

Unique local labels

Within the macro body, the user can declare a unique label identifier using the value of a macro parameter, like so:

  .macro foo(a)
 .label $a:

Unique local variables

Within the macro body, the user can declare a local variable with a unique name.

  .macro foo()
  .macro_local int b
  .b = 42
  print .b # prints the value of the unique variable (42)

The .macro_local directive declares a local variable with a unique name in the macro. When the macro .foo() is called, the resulting code that is given to the parser will read as follows:

  .sub main
    .local int local__foo__b__2
    local__foo__b__2 = 42
    print local__foo__b__2


The user can also declare a local variable with a unique name set to the symbolic value of one of the macro parameters.

  .macro foo(b)
  .macro_local int $b
  .$b = 42
  print .$b # prints the value of the unique variable (42)
  print .b  # prints the value of parameter "b", which is
            # also the name of the variable.

So, the special $ character indicates whether the symbol is interpreted as just the value of the parameter, or that the variable by that name is meant. Obviously, the value of b should be a string.

The automatic name munging on .macro_local variables allows for using multiple macros, like so:

  .macro foo(a)
  .macro_local int $a

  .macro bar(b)
  .macro_local int $b

  .sub main

This will result in code for the parser as follows:

  .sub main
    .local int local__foo__x__2
    .local int local__bar__x__4

Each expansion is associated with a unique number; for labels declared with .macro_label and locals declared with .macro_local expansions, this means that multiple expansions of a macro will not result in conflicting label or local names.

Ordinary local variables

Defining a non-unique variable can still be done, using the normal syntax:

  .macro foo(b)
  .local int b
  .macro_local int $b

When invoking the macro foo as follows:


there will be two variables: b and x. When the macro is invoked twice:

  .sub main

the resulting code that is given to the parser will read as follows:

  .sub main
    .local int b
    .local int local__foo__x
    .local int b
    .local int local__foo__y

Obviously, this will result in an error, as the variable b is defined twice. If you intend the macro to create unique variables names, use .macro_local instead of .local to take advantage of the name munging.


Subroutine Definition ^

  .sub _sub_label [<subflag>]*
   .param int a
   .param int b
   .param int c
   .return xy

Subroutine Call ^

  .const .Sub $P0 = "_sub_label"
  $P1 = new 'Continuation'
  set_addr $P1, ret_addr
  .local int x
  .local num y
  .local str z
  .arg x
  .arg y
  .arg z
  .call $P0, $P1    # r = _sub_label(x, y, z)
  .local int r  # optional - new result var
  .result r

NCI Call ^

  load_lib $P0, "libname"
  dlfunc $P1, $P0, "funcname", "signature"
  .arg x
  .arg y
  .arg z
  .nci_call $P1 # r = funcname(x, y, z)
  .local int r  # optional - new result var
  .result r

Subroutine Call Syntactic Sugar ^

  ...  # variable decls
  r = _sub_label(x, y, z)
  (r1[, r2 ...]) = _sub_label(x, y, z)
  _sub_label(x, y, z)

This also works for NCI calls, as the subroutine PMC will be a NCI sub, and on invocation will do the Right Thing. Instead of the label a subroutine object can be used too:

   find_global $P0, "_sub_label"

Methods ^

  .namespace [ "Foo" ]

  .sub _sub_label :method [,Subpragma, ...]
   .param int a
   .param int b
   .param int c
   .return xy

The variable "self" automatically refers to the invocating object, if the subroutine declaration contains "method".

Calling Methods ^

The syntax is very similar to subroutine calls. The call is done with meth_call which must immediately be preceded by the .invocant:

   .local pmc class
   .local pmc obj
   newclass class, "Foo"
   new obj, class
  .arg x
  .arg y
  .arg z
  .invocant obj
  .meth_call "_method" [, $P1 ] # r = obj."_method"(x, y, z)
  .local int r  # optional - new result var
  .result r

The return continuation is optional. The method can be a string constant or a string variable.

Returning and Yielding ^

  .return ( a, b )      # return the values of a and b

  .return ()            # return no value

  .return func_call()   # tail call function

  .return o."meth"()    # tail method call

Similarly, one can yield using the .yield directive

  .yield ( a, b )      # yield with the values of a and b

  .yield ()            # yield with no value

Stack calling conventions ^

Arguments are saved in reverse order onto the user stack:

   .arg y   # save args in reversed order
   .arg x
   call _foo    #(r, s) = _foo(x,y)
   .local int r
   .local int s
   .result r    # restore results in order
   .result s    #

and return values are restored in argument order from there.

 .sub _foo      # sub foo(int a, int b)
   .param int a         # receive arguments from left to right
   .param int b

   .return mi       # return (pl, mi), push results
   .return pl       # in reverse order

Pushing arguments in reversed order on the user stack makes the left most argument the top of stack entry. This allows for a variable number of function arguments (and return values), where the left most argument before a variable number of following arguments is the argument count.






See docs/imcc/macros.pod