=encoding utf8

=head1 TITLE

Synopsis 6: Subroutines

=head1 AUTHORS

    Damian Conway <damian@conway.org>
    Allison Randal <al@shadowed.net>
    Larry Wall <larry@wall.org>
    Daniel Ruoso <daniel@ruoso.com>

=head1 VERSION

    Created: 21 Mar 2003

    Last Modified: 17 Nov 2010
    Version: 147

This document summarizes Apocalypse 6, which covers subroutines and the
new type system.

=head1 Subroutines and other code objects

B<Subroutines> (keyword: C<sub>) are non-inheritable routines with
parameter lists.

B<Methods> (keyword: C<method>) are inheritable routines which always
have an associated object (known as their invocant) and belong to a
particular kind or class.

B<Submethods> (keyword: C<submethod>) are non-inheritable methods, or
subroutines masquerading as methods. They have an invocant and belong to
a particular kind or class.

B<Regexes> (keyword: C<regex>) are methods (of a grammar) that perform
pattern matching. Their associated block has a special syntax (see
Synopsis 5).  (We also use the term "regex" for anonymous patterns
of the traditional form.)

B<Tokens> (keyword: C<token>) are regexes that perform low-level
non-backtracking (by default) pattern matching.

B<Rules> (keyword: C<rule>) are regexes that perform non-backtracking
(by default) pattern matching (and also enable rules to do whitespace

B<Macros> (keyword: C<macro>) are routines whose calls execute as soon
as they are parsed (i.e. at compile-time). Macros may return another
source code string or a parse-tree.

=head1 Routine modifiers

B<Multis> (keyword: C<multi>) are routines that can have multiple
variants that share the same name, selected by arity, types, or some
other constraints.

B<Prototypes> (keyword: C<proto>) specify the commonalities (such
as parameter names, fixity, and associativity) shared by all multis
of that name in the scope of the C<proto> declaration.  A C<proto>
also adds an implicit C<multi> to all routines of the same short
name within its scope, unless they have an explicit modifier.
(This is particularly useful when adding to rule sets or when attempting
to compose conflicting methods from roles.)  Abstractly, the C<proto>
is a generic wrapper around the dispatch to the C<multi>s.  Each C<proto>
is instantianted into an actual dispatcher for each scope that
needs a different candidate list.

B<Only> (keyword: C<only>) routines do not share their short names
with other routines.  This is the default modifier for all routines,
unless a C<proto> of the same name was already in scope.  (For subs,
the governing C<proto> must have been declared in the same file, so
C<proto> declarations from the setting or other modules don't have
this effect unless explicitly imported.)

A modifier keyword may occur before the routine keyword in a named routine:

    only sub foo {...}
    proto sub foo {...}
    dispatch sub foo {...}      # internal
    multi sub foo {...}

    only method bar {...}
    proto method bar {...}
    dispatch method bar {...}   # internal
    multi method bar {...}

If the routine keyword is omitted, it defaults to C<sub>.

Modifier keywords cannot apply to anonymous routines.

A C<proto> is a generic dispatcher, which any given scope with a unique
candidate list will instantiate into a C<dispatch> routine.  Hence
a C<proto> is never called directly, much like a C<role> can't be
used as an instantiated object.

When you call any routine (or method, or rule) that may have multiple
candidates, the basic dispatcher is really only calling an "only"
sub or method--but if there are multiple candidates, the "only" that
will be found is really a dispatcher.  This instantiated C<dispatch>
is always called first (at least in the abstract--this can often be
optimized away).  In essence, a C<dispatch> is dispatched exactly
like an C<only> sub, but the C<dispatch> itself may delegate to any
of the candidates it is "managing".

It is the C<dispatch>'s responsibility to first vet the arguments for all the
candidates; any call that does not successfully bind the C<dispatch>'s signature fails outright.
(Its signature is a copy of one belonging to the C<proto> from which it was instantiated.)
The C<dispatch> does not necessarily send the original capture to its candidates, however.
Named arguments that bind to positionals in the C<dispatch> sig will become positionals
for all subsequent calls to its managed multis.

The dispatch then considers its list of managed candidates from the
viewpoint of the caller or object, sorts them into some order, and
dispatches them according to the rules of multiple dispatch as defined
for each of the various dispatchers.  In the case of multi subs, the
candidate list is known at compile time.  In the case of multi methods,
it may be necessary to generate (or regenerate) the candidate list at
run time, depending on what is known when about the inheritance tree.

This default dispatch behavior is symbolized within the original
C<proto> by a block containing of a single C<*> (that is, a
"whatever").  Hence the typical C<proto> will simply have a body
of C<{*}>.

    proto method bar {*}

(We don't use C<...> for that because it would fail at run time,
and the proto's instantiated C<dispatch> blocks are not stubs, but
are intended to be executed.)

Other statements may be inserted before and after the C<{*}>
statement to capture control before or after the multi dispatch:

    proto foo ($a,$b) { say "Called with $a $b"; {*}; say "Returning"; }

(That C<proto> is only good for C<multi>s with side effects and no return
value, since it returns the result of C<say>, which might not be what
you want.  See below for how to fix that.)

The syntactic form C<&foo> (without a modifying signature) can never
refer to a C<multi> candidate or a generic C<proto>.  It may only
refer to the single C<only> or C<dispatch> routine that would first
be called by C<foo()>.  Individual C<multi>s may be named by appending
a signature to the noun form: C<&foo:($,$,*@)>.

We used the term "managed" loosely above to indicate the set of C<multi>s in
question; the "managed set" is more accurately defined as the intersection
of all the C<multi>s in the C<proto>'s downward scope with all the C<multi>s that
are visible to the caller's upward-looking scope.  For ordinary routines
this means looking down lexical scopes and looking up lexical scopes.  [This
is more or less how C<multi>s already behave.]

For methods this means looking down or up the inheritance tree; "managed set"
in this case translates to the intersection of all methods in the C<proto>'s
class or its subclasses with all C<multi> methods visible to the object in its
parent classes, that is, the parent classes of the object's actual type on
whose behalf the method was called.  [Note, this is a change from prior
multi method semantics, which restricted multimethods to a single class;
the old semantics is equivalent to defining a C<proto> in every class that has
multimethods.  The new way gives the user the ability to intermix C<multi>s at
different inheritance levels.  (The first version of Rakudo * is not expected
to implement this).]

Also, the old semantics of C<proto> providing the most-default C<multi> body
is hereby deprecated.  Default C<multi>s should be marked with "C<is default>".

It is still possible to provide default behavior in the C<proto>, however, by
using it as a wrapper:

    my proto sub foo (@args) {
        {*}       # call into the managed set, then come back

Note that this returns the value of do-something-after(), not the C<multi>.
There are two ways to get around that.  Here's one way:

    my proto sub foo (@args) {
        ENTER do-something-before(@args);
        LEAVE do-something-after(@args);

Alternately, you can spell out what C<{*}> is actually sugar for,
which would be some dispatcher macro such as:

    my proto sub foo (|$cap (@args)) {
        my |$retcap := MULTI-DISPATCH-CALLWITH($?ROUTINE,$cap);
        return |$retcap;

which optimizes (we hope) to an inlined multidispatcher to locate all
the candidates for these arguments (hopefully memoized), create the dynamic
scope of a dispatch, start the dispatch, manage C<callnext> and C<lastcall>
semantics, and return the result of whichever C<multi> succeeded, if any.

Which why we have C<{*}> instead.

Another common variant would be to propagate control to the
outer/higher routine that would have been found if this one didn't

    my proto method foo { {*}; UNDO nextsame; }  # failover to super foo

Note that, in addition to making C<multi>s work similarly to each other,
the new C<proto> semantics greatly simplify top-level dispatchers, which
never have to worry about C<multi>s, because C<multi>s are always in the
second half of the double dispatch (again, just in the abstract, since
the first dispatch can often be optimized away, as if the C<proto> were
inlined).  So in the abstract, C<foo()> only ever calls a single
C<only>/C<proto> routine, and We Know Which One It Is at compile time.

This is less of a shift for method dispatch, which already assumed that there
is something like a single proto in each class that redispatches inside
the class.  Here the change is that multi-method dispatcher needs to look
more widely for its candidates than the current class.  But note that our
semantics were inconsistent before, insofar as regex methods already had to
look for this larger managed set in order to do transitive LTM correctly.
Now the semantics of normal method C<proto>s and regex C<proto>s are nearly
identical, apart from the fact that regex candidate lists naturally have
fancier tiebreaking rules involving longest token matching.

A C<dispatch> must be generated for every scope that contains one or more C<multi>
declaration.  This is done by searching backwards and outwards (or up the
inheritance chain for methods) for a C<proto> to instantiate.  If no such
C<proto> is found, a "most generic" C<proto> will be generated, something like:

    proto sub foo (*@, *%) {*}
    proto method foo (*@, *%) {*}

Obviously, no named-to-positional remapping can be done in this case.

[Conjecture: we could instead autogen a more specific signature for
each such autogenerated C<dispatch> once we know its exact candidate
set, such that consistent use of positional parameter names is rewarded
with positional names in the generated signature, which could remap
named parameters.]

=head2 Named subroutines

The general syntax for named subroutines is any of:

     my RETTYPE sub NAME ( PARAMS ) TRAITS {...}    # lexical only
    our RETTYPE sub NAME ( PARAMS ) TRAITS {...}    # also package-scoped
                sub NAME ( PARAMS ) TRAITS {...}    # same as "our"

The return type may also be put inside the parentheses:

    sub NAME (PARAMS --> RETTYPE) {...}

Unlike in Perl 5, named subroutines are considered expressions,
so this is valid Perl 6:

    my @subs = (sub foo { ... }, sub bar { ... });

Another difference is that subroutines default to C<my> scope rather
than C<our> scope.  However, subroutine dispatch searches lexical
scopes outward, and subroutines are also allowed to be I<postdeclared>
after their use, so you won't notice this much.  A subroutine that is
not declared yet may be called using parentheses around the arguments,
in the absence of parentheses, the subroutine call is assumed to take
multiple arguments in the form of a list operator.

=head2 Anonymous subroutines

The general syntax for anonymous subroutines is:

    sub ( PARAMS ) TRAITS {...}

But one can also use the C<anon> scope modifier to introduce the return type first:

    anon RETTYPE sub ( PARAMS ) TRAITS {...}

When an anonymous subroutine will be assigned to a scalar variable,
the variable can be declared with the signature of the routines that
will be assigned to it:

    my $grammar_factory:(Str, int, int --> Grammar);
    $grammar_factory = sub (Str $name, int $n, int $x --> Grammar) { ... }; 

Covariance allows a routine (that has a more derived return type than what is
defined in the scalar's signature) to be assigned to that scalar.
Contravariance allows a routine (with parameter types that are less derived
than those in the scalar's signature) to be assigned to that scalar. The
compiler may choose to enforce (by type-checking) such assignments at
compile-time, if possible. Such type annotations are intended to help the
compiler optimize code to the extent such annotations are included and/or to
the extent they aid in type inference.

The same signature can be used to mark the type of a closure parameter to
another subroutine:

    sub (int $n, &g_fact:(Str, int, int --> Grammar) --> Str) { ... }

B<Trait> is the name for a compile-time (C<is>) property.
See L<"Properties and traits">.

=head2 Perl5ish subroutine declarations

You can declare a sub without parameter list, as in Perl 5:

    sub foo {...}

This is equivalent to one of:

    sub foo () {...}
    sub foo (*@_) {...}
    sub foo (*%_) {...}
    sub foo (*@_, *%_) {...}

depending on whether either or both of those variables are used in the body of the routine.

Positional arguments implicitly come in via the C<@_> array, but
unlike in Perl 5 they are C<readonly> aliases to actual arguments:

    sub say { print qq{"@_[]"\n}; }   # args appear in @_

    sub cap { $_ = uc $_ for @_ }   # Error: elements of @_ are read-only

Also unlike in Perl 5, Perl 6 has true named arguments, which come in
via C<%_> instead of C<@_>.  (To construct pseudo-named arguments that
come in via C<@_> as in Perl 5, the p5-to-p6 translator will define and use the ugly
C<< p5=> >> operator instead of Perl 6's C<< => >> Pair constructor.)

If you need to modify the elements of C<@_> or C<%_>, declare the
array or hash explicitly with the C<is rw> trait:

    sub swap (*@_ is rw, *%_ is rw) { @_[0,1] = @_[1,0]; %_<status> = "Q:S"; }

Note: the C<rw> container trait is automatically distributed to the
individual elements by the slurpy star even though there is no
actual array or hash passed in.  More precisely, the slurpy star
means the declared formal parameter is I<not> considered readonly; only
its elements are.  See L</Parameters and arguments> below.

Note also that if the sub's block contains placeholder variables
(such as C<$^foo> or C<$:bar>), those are considered to be formal
parameters already, so in that case C<@_> or C<%_> fill the role of
sopping up unmatched arguments.  That is, if those containers are
explicitly mentioned within the body, they are added as slurpy
parameters.  This allows you to easily customize your error message
on unrecognized parameters.  If they are not mentioned in the body,
they are not added to the signature, and normal dispatch rules will
simply fail if the signature cannot be bound.

=head2 Blocks

Raw blocks are also executable code structures in Perl 6.

Every block defines an object of type C<Block> (which C<does Callable>), which may either be
executed immediately or passed on as a C<Block> object.  How a block is
parsed is context dependent.

A bare block where an operator is expected terminates the current
expression and will presumably be parsed as a block by the current
statement-level construct, such as an C<if> or C<while>.  (If no
statement construct is looking for a block there, it's a syntax error.)
This form of bare block requires leading whitespace because a bare
block where a postfix is expected is treated as a hash subscript.

A bare block where a term is expected merely produces a C<Block> object.
If the term bare block occurs in a list, it is considered the final
element of that list unless followed immediately by a comma or colon
(intervening C<\h*> or "unspace" is allowed).

=head2 "Pointy blocks"

Semantically the arrow operator C<< -> >> is almost a synonym for the
C<sub> keyword as used to declare an anonymous subroutine, insofar as
it allows you to declare a signature for a block of code. However,
the parameter list of a pointy block does not require parentheses,
and a pointy block may not be given traits.  In most respects,
though, a pointy block is treated more like a bare block than like
an official subroutine.  Syntactically, a pointy block may be used
anywhere a bare block could be used:

    my $sq = -> $val { $val**2 };
    say $sq(10); # 100

    my @list = 1..3;
    for @list -> $elem {
        say $elem; # prints "1\n2\n3\n"

It also behaves like a block with respect to control exceptions.
If you C<return> from within a pointy block, the block is transparent
to the return; it will return from the innermost enclosing C<sub> or
C<method> (et al.), not from the block itself.  It is referenced by C<&?BLOCK>,
not C<&?ROUTINE>.

A normal pointy block's parameters default to C<readonly>, just like
parameters to a normal sub declaration.  However, the double-pointy variant
defaults parameters to C<rw>:

    for @list <-> $elem {

This form applies C<rw> to all the arguments:

    for @kv <-> $key, $value {
        $key ~= ".jpg";
        $value *= 2 if $key ~~ :e;

=head2 Stub declarations

To predeclare a subroutine without actually defining it, use a "stub block":

    sub foo {...}     # Yes, those three dots are part of the actual syntax

The old Perl 5 form:

    sub foo;

is a compile-time error in Perl 6 (because it would imply that the body of the
subroutine extends from that statement to the end of the file, as C<class> and
C<module> declarations do).  The only allowed use of the semicolon form is to
declare a C<MAIN> sub--see L</Declaring a MAIN subroutine> below.

Redefining a stub subroutine does not produce an error, but redefining
an already-defined subroutine does. If you wish to redefine a defined sub,
you must explicitly use the "C<supersede>" declarator.  (The compiler may
refuse to do this if it has already committed to the previous definition.)

The C<...> is the "yadayadayada" operator, which is executable but
returns a failure.  You can also use C<???> to produce a warning,
or C<!!!> to always die.  These also officially define stub blocks.
Any of these yada operators will be taken as a stub if used as the main
operator of the first statement in the block.  (Statement modifiers
are allowed on that statement.)  The yada operators differ from their
respective named functions in that they all default to a message
such as: "Unimplemented stub of sub foo was executed".

It has been argued that C<...> as literal syntax is confusing when
you might also want to use it for metasyntax within a document.
Generally this is not an issue in context; it's never an issue in the
program itself, and the few places where it could be an issue in the
documentation, a comment will serve to clarify the intent, as above.
The rest of the time, it doesn't really matter whether the reader
takes C<...> as literal or not, since the purpose of C<...> is to
indicate that something is missing whichever way you take it.

=head2 Globally scoped subroutines

Subroutines and variables can be declared in the global namespace
(or any package in the global namespace), and are thereafter visible
everywhere in the program via the GLOBAL package (or one of its
subpackages).  They may be made directly visible by importation,
but may not otherwise be called with a bare identifier, since subroutine
dispatch only looks in lexical scopes.

Global subroutines and variables are normally referred to by prefixing
their identifiers with the C<*> twigil, to allow dynamically scoped overrides.

    GLOBAL::<$next_id> = 0;
    sub GLOBAL::saith($text)  { say "Yea verily, $text" }

    module A {
        my $next_id = 2;     # hides any global or package $next_id
        &*saith($next_id);   # print the lexical $next_id;
        &*saith($*next_id);  # print the dynamic $next_id;

To disallow dynamic overrides, you must access the globals directly:


The fact that this is verbose is construed to be a feature.  Alternately,
you may play aliasing tricks like this:

    module B {
        import GLOBAL <&saith $next_id>;
        saith($next_id);    # Unambiguously the global definitions

Despite the fact that subroutine dispatch only looks in lexical scopes, you
can always call a package subroutine directly if there's a lexical alias
to it, as the C<our> declarator does:

    module C;
    our sub saith($text)  { say "Yea verily, $text" }
    saith("I do!")      # okay
    C::saith("I do!")   # also okay

=head2 Dynamically scoped subroutines

Similarly, you may define dynamically scoped subroutines:

    my sub myfunc ($x) is dynamic { ... }
    my sub &*myfunc ($x) { ... }        # same thing

This may then be invoked via the syntax for dynamic variables:


=head2 Lvalue subroutines

Lvalue subroutines return a "proxy" object that can be assigned to.
It's known as a proxy because the object usually represents the
purpose or outcome of the subroutine call.

Subroutines are specified as being lvalue using the C<is rw> trait.

An lvalue subroutine may return a variable:

    my $lastval;
    sub lastval () is rw { return $lastval }

or the result of some nested call to an lvalue subroutine:

    sub prevval () is rw { return lastval() }

or a specially tied proxy object, with suitably programmed
C<FETCH> and C<STORE> methods:

    sub checklastval ($passwd) is rw {
        return new Proxy:
                FETCH => method {
                            return lastval();
                STORE => method ($val) {
                            die unless check($passwd);
                            lastval() = $val;

Other methods may be defined for specialized purposes such as temporizing
the value of the proxy.

=head2 Operator overloading

Operators are just subroutines with special names and scoping.
An operator name consists of a grammatical category name followed by
a single colon followed by an operator name specified as if it were
a one or more strings.  So any of these indicates the same binary addition operator:


Use the C<&> sigil just as you would on ordinary subs.

Unary operators are defined as C<prefix> or C<postfix>:

    sub prefix:<OPNAME>  ($operand) {...}
    sub postfix:<OPNAME> ($operand) {...}

Binary operators are defined as C<infix>:

    sub infix:<OPNAME> ($leftop, $rightop) {...}

Bracketing operators are defined as C<circumfix> where a term is expected
or C<postcircumfix> where a postfix is expected.  A two-element slice
containing the leading and trailing delimiters is the name of the

    sub circumfix:<LEFTDELIM RIGHTDELIM> ($contents) {...}
    sub circumfix:['LEFTDELIM','RIGHTDELIM'] ($contents) {...}

Contrary to Apocalypse 6, there is no longer any rule about splitting an even
number of characters.  You must use a two-element slice.  Such names
are canonicalized to a single form within the symbol table, so you
must use the canonical name if you wish to subscript the symbol table
directly (as in C<< PKG::{'infix:<+>'} >>).  Otherwise any form will
do.  (Symbolic references do not count as direct subscripts since they
go through a parsing process.)  The canonical form always uses angle
brackets and a single space between slice elements.  The elements
are escaped on brackets, so C<< PKG::circumfix:['<','>'] >> is canonicalized
to C<<< PKG::{'circumfix:<\< \>>'} >>>, and decanonicalizing may always
be done left-to-right.

Operator names can be any sequence of non-whitespace characters
including Unicode characters. For example:

    sub infix:<(c)> ($text, $owner) { return $text but Copyright($owner) }
    method prefix:<±> (Num $x --> Num) { return +$x | -$x }
    multi sub postfix:<!> (Int $n) { $n < 2 ?? 1 !! $n*($n-1)! }
    macro circumfix:«<!-- -->» ($text) is parsed / .*? / { "" }

    my $document = $text (c) $me;

    my $tolerance = ±7!;

    <!-- This is now a comment -->

Whitespace may never be part of the name (except as separator
within a C<< <...> >> or C<«...»> slice subscript, as in the example above).

A null operator name does not define a null or whitespace operator, but
a default matching subrule for that syntactic category, which is useful when
there is no fixed string that can be recognized, such as tokens beginning
with digits.  Such an operator I<must> supply an C<is parsed> trait.
The Perl grammar uses a default subrule for the C<:1st>, C<:2nd>, C<:3rd>,
etc. regex modifiers, something like this:

    sub regex_mod_external:<> ($x) is parsed(token { \d+[st|nd|rd|th] }) {...}

Such default rules are attempted in the order declared.  (They always follow
any rules with a known prefix, by the longest-token-first rule.)

Although the name of an operator can be installed into any package or
lexical namespace, the syntactic effects of an operator declaration are
always lexically scoped.  Operators other than the standard ones should
not be installed into the C<*> namespace.  Always use exportation to make
non-standard syntax available to other scopes.

=head1 Parameters and arguments

Perl 6 subroutines may be declared with parameter lists.

By default, all parameters are readonly aliases to their corresponding
arguments--the parameter is just another name for the original
argument, but the argument can't be modified through it.  This is
vacuously true for value arguments, since they may not be modified in
any case.  However, the default forces any container argument to also
be treated as an immutable value.  This extends down only one level;
an immutable container may always return an element that is mutable if
it so chooses.  (For this purpose a scalar variable is not considered
a container of its singular object, though, so the top-level object
within a scalar variable is considered immutable by default.  Perl 6
does not have references in the same sense that Perl 5 does.)

To allow modification, use the C<is rw> trait.  This requires a mutable
object or container as an argument (or some kind of type object that
can be converted to a mutable object, such as might be returned
by an array or hash that knows how to autovivify new elements).
Otherwise the signature fails to bind, and this candidate routine
cannot be considered for servicing this particular call.  (Other multi
candidates, if any, may succeed if they don't require C<rw> for this
parameter.)  In any case, failure to bind does not by itself cause
an exception to be thrown; that is completely up to the dispatcher.

To pass-by-copy, use the C<is copy> trait.  An object container will
be cloned whether or not the original is mutable, while an (immutable)
value will be copied into a suitably mutable container.  The parameter
may bind to any argument that meets the other typological constraints
of the parameter.

If you have a readonly parameter C<$ro>, it may never be passed on to
a C<rw> parameter of a subcall, whether or not C<$ro> is currently
bound to a mutable object.  It may only be rebound to readonly or
copy parameters.  It may also be rebound to a parcel parameter (see
"C<is parcel>" below), but modification will fail as in the case where
an immutable value is bound to a C<parcel> parameter.

Aliases of C<$ro> are also readonly, whether generated explicitly with C<:=>
or implicitly within a C<Capture> object (which are themselves immutable).

Also, C<$ro> may not be returned from an lvalue subroutine or method.

Parameters may be required or optional. They may be passed by position,
or by name. Individual parameters may confer an item or list context
on their corresponding arguments, but unlike in Perl 5, this is decided
lazily at parameter binding time.

Arguments destined for required positional parameters must come before
those bound to optional positional parameters. Arguments destined
for named parameters may come before and/or after the positional
parameters.  (To avoid confusion it is highly recommended that all
positional parameters be kept contiguous in the call syntax, but
this is not enforced, and custom arg list processors are certainly
possible on those arguments that are bound to a final slurpy or
arglist variable.)

A signature containing a name collision is considered a compile time
error. A name collision can occur between positional parameters, between
named parameters, or between a positional parameter and a named one.
The sigil is not considered in such a comparison, except in the case of
two positional parameters -- in other words, a signature in which two
or more parameters are identical except for the sigil is still OK (but
you won't be able to pass values by that name).

    :($a, $a)       # wrong, two $a
    :($a, @a)       # OK (but don't do that)
    :($a, :a($b))   # wrong, one a, one a through renaming
    :($a, :a(@b))   # wrong 
    :(:$a, :@a)     # wrong

=head2 Named arguments

Named arguments are recognized syntactically at the "comma" level.
Since parameters are identified using identifiers, the recognized
syntaxes are those where the identifier in question is obvious.
You may use either the adverbial form, C<:name($value)>, or the
autoquoted arrow form, C<< name => $value >>.  These must occur at
the top "comma" level, and no other forms are taken as named pairs
by default.  Pairs intended as positional arguments rather than named
arguments may be indicated by extra parens or by explicitly quoting
the key to suppress autoquoting:

    doit :when<now>,1,2,3;      # always a named arg
    doit (:when<now>),1,2,3;    # always a positional arg

    doit when => 'now',1,2,3;   # always a named arg
    doit (when => 'now'),1,2,3; # always a positional arg
    doit 'when' => 'now',1,2,3; # always a positional arg

Only bare keys with valid identifier names are recognized as named arguments:

    doit when => 'now';         # always a named arg
    doit 'when' => 'now';       # always a positional arg
    doit 123  => 'now';         # always a positional arg
    doit :123<now>;             # always a positional arg

Going the other way, pairs intended as named arguments that don't look
like pairs must be introduced with the C<|> prefix operator:

    $pair = :when<now>;
    doit $pair,1,2,3;                # always a positional arg
    doit |$pair,1,2,3;               # always a named arg
    doit |get_pair(),1,2,3;          # always a named arg
    doit |('when' => 'now'),1,2,3;   # always a named arg

Note the parens are necessary on the last one due to precedence.

Likewise, if you wish to pass a hash and have its entries treated as
named arguments, you must dereference it with a C<|>:

    %pairs = (:when<now>, :what<any>);
    doit %pairs,1,2,3;               # always a positional arg
    doit |%pairs,1,2,3;              # always named args
    doit |%(get_pair()),1,2,3;       # always a named arg
    doit |%('when' => 'now'),1,2,3;  # always a named arg

Variables with a C<:> prefix in rvalue context autogenerate pairs, so you
can also say this:

    $when = 'now';
    doit $when,1,2,3;   # always a positional arg of 'now'
    doit :$when,1,2,3;  # always a named arg of :when<now>

In other words C<:$when> is shorthand for C<:when($when)>.  This works
for any sigil:

    :$what      :what($what)
    :@what      :what(@what)
    :%what      :what(%what)
    :&what      :what(&what)

Ordinary hash notation will just pass the value of the hash entry as a
positional argument regardless of whether it is a pair or not.
To pass both key and value out of hash as a positional pair, use C<:p>

    doit %hash<a>:p,1,2,3;
    doit %hash{'b'}:p,1,2,3;

The C<:p> stands for "pairs", not "positional"--the C<:p> adverb may be
placed on any C<Associative> access subscript to make it mean "pairs" instead of "values".
If you want the pair (or pairs) to be interpreted as named arguments,
you may do so by prefixing with the C<< prefix:<|> >> operator:

    doit |%hash<a>:p,1,2,3;
    doit |%hash{'b'}:p,1,2,3;

C<Pair> constructors are recognized syntactically at the call level and
put into the named slot of the C<Capture> structure.  Hence they may be
bound to positionals only by name, not as ordinary positional C<Pair>
objects.  Leftover named arguments can be slurped into a slurpy hash.

Because named and positional arguments can be freely mixed, the
programmer always needs to disambiguate pairs literals from named
arguments with parentheses or quotes:

    # Named argument "a"
    push @array, 1, 2, :a<b>;

    # Pair object (a=>'b')
    push @array, 1, 2, (:a<b>);
    push @array, 1, 2, 'a' => 'b';

Perl 6 allows multiple same-named arguments, and records the relative
order of arguments with the same name.  When there are more than one
argument, the C<@> sigil in the parameter list causes the arguments
to be concatenated:

    sub fun (Int @x) { ... }
    fun( x => 1, x => 2 );              # @x := (1, 2)
    fun( x => (1, 2), x => (3, 4) );    # @x := (1, 2, 3, 4)

Other sigils bind only to the I<last> argument with that name:

    sub fun (Int $x) { ... }
    fun( x => 1, x => 2 );              # $x := 2
    fun( x => (1, 2), x => (3, 4) );    # $x := (3, 4)

This means a hash holding default values must come I<before> known named
parameters, similar to how hash constructors work:

    # Allow "x" and "y" in %defaults to be overridden
    f( |%defaults, x => 1, y => 2 );

=head2 Invocant parameters

A method invocant may be specified as the first parameter in the parameter
list, with a colon (rather than a comma) immediately after it:

    method get_name ($self:) {...}
    method set_name ($_: $newname) {...}

The corresponding argument (the invocant) is evaluated in item context
and is passed as the left operand of the method call operator:

    print $obj.get_name();

The invocant is actually stored as the first positional argument of a C<Capture>
object.  It is special only to the dispatcher, otherwise it's just a normal
positional argument.

Single-dispatch semantics may also be requested by using the indirect object syntax, with a colon
after the invocant argument. The colon is just a special form of the comma, and has the
same precedence:

    set_name $obj: "Sam";
    $obj.set_name("Sam");   # same as the above

An invocant is the topic of the corresponding method if that formal
parameter is declared with the name C<$_>.  A method's invocant
always has the alias C<self>.  Other styles of self can be declared
with the C<self> pragma.

If you have a call of the form:


the compiler must defer the decision on whether to treat it as a method
or function dispatch based on whether the supplied C<Capture>'s first
argument is marked as an invocant.  For ordinary calls this can
always be determined at compile time, however.

=head2 Longname parameters

A routine marked with C<multi> can mark part of its parameters to
be considered in the multi dispatch.  These are called I<longnames>;
see S12 for more about the semantics of multiple dispatch.

You can choose part of a C<multi>'s parameters to be its longname,
by putting a double semicolon after the last one:

    multi sub handle_event ($window, $event;; $mode) {...}
    multi method set_name ($self: $name;; $nick) {...}

A parameter list may have at most one double semicolon; parameters
after it are never considered for multiple dispatch (except of course
that they can still "veto" if their number or types mismatch).

[Conjecture: It might be possible for a routine to advertise multiple
long names, delimited by single semicolons.  See S12 for details.]

If the parameter list for a C<multi> contains no semicolons to delimit
the list of important parameters, then all positional parameters are
considered important.  If it's a C<multi method> or C<multi submethod>,
an additional implicit unnamed C<self> invocant is added to the
signature list unless the first parameter is explicitly marked with a colon.

=head2 Required parameters

Required parameters are specified at the start of a subroutine's parameter

    sub numcmp ($x, $y) { return $x <=> $y }

Required parameters may optionally be declared with a trailing C<!>,
though that's already the default for positional parameters:

    sub numcmp ($x!, $y!) { return $x <=> $y }

The corresponding arguments are evaluated in item context and may be
passed positionally or by name. To pass an argument by name,
specify it as a pair: C<< I<parameter_name> => I<argument_value> >>.

    $comparison = numcmp(2,7);
    $comparison = numcmp(x=>2, y=>7);
    $comparison = numcmp(y=>7, x=>2);

Pairs may also be passed in adverbial pair notation:

    $comparison = numcmp(:x(2), :y(7));
    $comparison = numcmp(:y(7), :x(2));

Passing the wrong number of required arguments to a normal subroutine
is a fatal error.  Passing a named argument that cannot be bound to a normal
subroutine is also a fatal error.  (Methods are different.)

The number of required parameters a subroutine has can be determined by
calling its C<.arity> method:

    $args_required = &foo.arity;

=head2 Optional parameters

Optional positional parameters are specified after all the required
parameters and each is marked with a C<?> after the parameter:

    sub my_substr ($str, $from?, $len?) {...}

Alternately, optional fields may be marked by supplying a default value.
The C<=> sign introduces a default value:

    sub my_substr ($str, $from = 0, $len = Inf) {...}

Default values can be calculated at run-time. They may even use the values of
preceding parameters:

    sub xml_tag ($tag, $endtag = matching_tag($tag) ) {...}

Arguments that correspond to optional parameters are evaluated in
item context. They can be omitted, passed positionally, or passed by

    my_substr("foobar");            # $from is 0, $len is infinite
    my_substr("foobar",1);          # $from is 1, $len is infinite
    my_substr("foobar",1,3);        # $from is 1, $len is 3
    my_substr("foobar",len=>3);     # $from is 0, $len is 3

Missing optional arguments default to their default values, or to
an undefined value if they have no default.  (A supplied argument that is
undefined is not considered to be missing, and hence does not trigger
the default.  Use C<//=> within the body for that.)

You may check whether an optional parameter was bound to anything
by calling C<VAR($param).defined>.

=head2 Named parameters

Named-only parameters follow any required or optional parameters in the
signature. They are marked by a prefix C<:>:

    sub formalize($text, :$case, :$justify) {...}

This is actually shorthand for:

    sub formalize($text, :case($case), :justify($justify)) {...}

If the longhand form is used, the label name and variable name can be

    sub formalize($text, :case($required_case), :justify($justification)) {...}

so that you can use more descriptive internal parameter names without
imposing inconveniently long external labels on named arguments.
Multiple name wrappings may be given; this allows you to give both a
short and a long external name:

    sub globalize (:g(:global($gl))) {...}

Or equivalently:

    sub globalize (:g(:$global)) {...}

Arguments that correspond to named parameters are evaluated in item
context. They can only be passed by name, so it doesn't matter what
order you pass them in:

    $formal = formalize($title, case=>'upper');
    $formal = formalize($title, justify=>'left');
    $formal = formalize($title, :justify<right>, :case<title>);

See S02 for the correspondence between adverbial form and arrow notation.

While named and position arguments may be intermixed, it is suggested
that you keep all the positionals in one place for clarity unless you
have a good reason not to.  This is likely bad style:

    $formal = formalize(:justify<right>, $title, :case<title>, $date);

Named parameters are optional unless marked with a following C<!>.
Default values for optional named parameters are defined in the same
way as for positional parameters, but may depend only on existing
values, including the values of parameters that have already been
bound.  Named optional parameters default to C<Nil> if they have
no default.  Named required parameters fail unless an argument pair
of that name is supplied.

Bindings happen in declaration order, not call order, so any default
may reliably depend on formal parameters to its left in the signature.
In other words, if the first parameter is C<$a>, it will bind to
a C<:a()> argument in preference to the first positional argument.
It might seem that performance of binding would suffer by requiring
a named lookup before a positional lookup, but the compiler is able
to guarantee that subs with known fixed signatures (both C<only>s and
C<proto>s) translate named arguments to positional in the
first N positions.  Also, purely positional calls may obviously omit any
named lookups, as may bindings that have already used up all the named
arguments.  The compiler is also free to intuit proto signatures for
a given sub or method name as long as the candidate list is stable.

=head2 List parameters

List parameters capture a variable length list of data. They're used
in subroutines like C<print>, where the number of arguments needs to be
flexible. They're also called "variadic parameters", because they take a
I<variable> number of arguments.  But generally we call them "slurpy"
parameters because they slurp up arguments.

Slurpy parameters follow any required or optional parameters. They are
marked by a C<*> before the parameter:

    sub duplicate($n, *%flag, *@data) {...}

Named arguments are bound to the slurpy hash (C<*%flag>
in the above example). Such arguments are evaluated in item context.
Any remaining variadic arguments at the end of the argument list
are bound to the slurpy array (C<*@data> above) and are evaluated
in list context.

For example:

    duplicate(3, reverse => 1, collate => 0, 2, 3, 5, 7, 11, 14);
    duplicate(3, :reverse, :!collate, 2, 3, 5, 7, 11, 14);  # same

    # The @data parameter receives [2, 3, 5, 7, 11, 14]
    # The %flag parameter receives { reverse => 1, collate => 0 }

Slurpy scalar parameters capture what would otherwise be the first
elements of the variadic array:

    sub head(*$head, *@tail)         { return $head }
    sub neck(*$head, *$neck, *@tail) { return $neck }
    sub tail(*$head, *@tail)         { return @tail }

    head(1, 2, 3, 4, 5);        # $head parameter receives 1
                                # @tail parameter receives [2, 3, 4, 5]

    neck(1, 2, 3, 4, 5);        # $head parameter receives 1
                                # $neck parameter receives 2
                                # @tail parameter receives [3, 4, 5]

Slurpy scalars still impose list context on their arguments.

Single slurpy parameters are treated lazily -- the list is only flattened
into an array when individual elements are actually accessed:

    @fromtwo = tail(1..Inf);        # @fromtwo contains a lazy [2..Inf]

[Conjecture: However, if you use two or more slurpy arrays in a
signature, the list is instead evaluated in hyper context, and will
be asked to split itself into the number of lists corresponding to
the number of slurpies so declared.  A non-hyperable list will return
failure for this splitting operation, so the signature should only bind
on parallelizable list operations.  Likewise a list that "too short to
split" fails to bind, so a separate signature may match empty lists,
and perhaps singletons, if we define "too short" that way).]

You can't bind to the name of a slurpy parameter: the name is just there
so you can refer to it within the body.

    sub foo(*%flag, *@data) {...}

    foo(:flag{ a => 1 }, :data[ 1, 2, 3 ]);
        # %flag has elements (flag => (a => 1)) and (data => [1,2,3])
        # @data has nothing

=head2 Slurpy block

It's also possible to declare a slurpy block: C<*&block>.  It slurps
up any nameless block, specified by C<{...}>, at either the current positional
location or the end of the syntactic list.  Put it first if you want the
option of putting a block either first or last in the arguments.  Put it
last if you want to force it to come in as the last argument.

=head2 Argument list binding

The underlying C<Capture> object may be bound to a single scalar
parameter marked with a C<|>.

    sub bar ($a,$b,$c,:$mice) { say $mice }
    sub foo (|$args) { say $args.perl; &bar.callwith(|$args); }

This prints:

    foo 1,2,3,:mice<blind>;     # says "\(1,2,3,:mice<blind>)" then "blind"

As demonstrated above, the capture may be interpolated into another
call's arguments.  (The C<|> prefix is described in the next section.)
Use of C<callwith> allows the routine to be called without introducing
an official C<CALLER> frame.  For more see "Wrapping" below.

It is allowed to rebind the parameters within the signature, but
only as a subsignature of the capture argument:

    sub compare (|$args (Num $x, Num $y --> Bool)) { ... }

For all normal declarative purposes (invocants and multiple dispatch
types, for instance), the inner signature is treated as the entire

    method addto (|$args ($self: @x)) { trace($args); $self += [+] @x }

The inner signature is not required for non-C<multi>s since there can
only be one candidate, but for multiple dispatch the inner signature
is required at least for its types, or the declaration would not know
what signature to match against.

    multi foo (|$args (Int, Bool?, *@, *%)) { reallyintfoo($args) }
    multi foo (|$args (Str, Bool?, *@, *%)) { reallystrfoo($args) }

=head2 Parcel binding

It is possible to bind the argument list in an even more raw form
as a C<Parcel> object.  A signature to capture the parcel must be
declared as:

    sub foo (\|$parcel) { ... }

This single parcel parameter is the only thing allowed in such a signature
(apart from a return type).  An implementation must either be able to
reconstruct the original parcel object from a capture, or the compiler
must be smart enough to know when to save the original parcel object
based on proto declarations and/or call indirection.

=head2 Flattening argument lists

The unary C<|> operator casts its argument to a C<Capture>
object, then splices that capture into the argument list
it occurs in.  To get the same effect on multiple arguments you
can use the C<< |« >> hyperoperator.

C<Pair> and C<Hash> become named arguments:

    |(x=>1);          # Pair, becomes \(x=>1)
    |{x=>1, y=>2};    # Hash, becomes \(x=>1, y=>2)

Anything else that is C<Iterable> is simply turned into
positional arguments:

    |(1,2,3);         # Seq, becomes \(1,2,3)
    |(1..3);          # Range, becomes \(1,2,3)
    |(1..2, 3);       # Parcel, becomes \(1,2,3)
    |([x=>1, x=>2]);  # List (from an Array), becomes \((x=>1), (x=>2))

For example:

    sub foo($x, $y, $z) {...}    # expects three scalars
    @onetothree = 1..3;          # array stores three scalars

    foo(1,2,3);                  # okay:  three args found
    foo(@onetothree);            # error: only one arg
    foo(|@onetothree);           # okay:  @onetothree flattened to three args

The C<|> operator flattens lazily -- the array is flattened only if
flattening is actually required within the subroutine. To flatten before
the list is even passed into the subroutine, use the C<flat> list

    foo(|flat 1,2,3 Z 4,5,6);    # zip list flattened before interpolation
    foo |(1,2,3 Z 4,5,6).flat    # same thing

=head2 Multidimensional argument list binding

Some functions take more than one list of positional and/or named arguments,
that they wish not to be flattened into one list.  For instance, C<zip()> wants
to iterate several lists in parallel, while array and hash subscripts want to
process a multidimensional slice.  The set of underlying argument lists may be
bound to a single array parameter declared with a double C<**> marker:

    sub foo (**@slice) { ... }

Note that this is different from

    sub foo (|$slice) { ... }

insofar as C<|$slice> is bound to a single argument-list object that
makes no commitment to processing its structure (and maybe doesn't
even know its own structure yet), while C<**@slice> has to create
an array that binds the incoming dimensional lists to the array's
dimensions, and make that commitment visible to the rest of the scope
via the sigil so that constructs expecting multidimensional lists
know that multidimensionality is the intention.

It is allowed to specify a return type:

    sub foo (**@slice --> Num) { ... }

The invocant does not participate in multi-dimensional argument lists,
so C<self> is not present in the C<**@slice> below:

    method foo (**@slice) { ... }

The C<**> marker is just a variant of the C<*> marker that ends up
requesting parcels when binding (underlyingly calling C<.getarg>) rather
than requesting individual elements as the flatening C<*> does (underlyingly
calling C<.get>).

=head2 Zero-dimensional argument list

If you call a function without parens and supply no arguments, the
argument list becomes a zero-dimensional slice.  It differs from
C<\()> in several ways:

    sub foo (**@slice) {...}
    foo;        # +@slice == 0
    foo();      # +@slice == 1

    sub bar (|$args = \(1,2,3)) {...}
    bar;        # $args === \(1,2,3)
    bar();      # $args === \()

=head2 Feed operators

The variadic list of a subroutine call can be passed in separately from
the normal argument list, by using either of the I<feed> operators:
C<< <== >> or C<< ==> >>.  Syntactically, feed operators expect to find a
statement on either end.  Any statement can occur on the source end;
however not all statements are suitable for use on the sink end of a feed.

Each operator expects to find a call to a variadic receiver on its
"sharp" end, and a list of values on its "blunt" end:

    grep { $_ % 2 } <== @data;

    @data ==> grep { $_ % 2 };

It binds the (potentially lazy) list from the blunt end to the slurpy
parameter(s) of the receiver on the sharp end.  In the case of a receiver
that is a variadic function, the feed is received as part of its slurpy list.
So both of the calls above are equivalent to:

    grep { $_ % 2 }, @data;

Note that all such feeds (and indeed all lazy argument lists) supply
an implicit promise that the code producing the lists may execute
in parallel with the code receiving the lists.  (Feeds, hyperops,
and junctions all have this promise of parallelizability in common,
but differ in interface.  Code which violates these promises is
erroneous, and will produce undefined results when parallelized.)

However, feeds go a bit further than ordinary lazy lists in enforcing
the parallel discipline: they explicitly treat the blunt end as a
cloned closure that starts a subthread (presumably cooperative).  The only variables shared
by the inner scope with the outer scope are those lexical variables
declared in the outer scope that are visible at the time the closure is
cloned and the subthread spawned.  Use of such shared variables will
automatically be subject to transactional protection (and associated
overhead).  Package variables are not cloned unless predeclared
as lexical names with C<our>.  Variables declared within the blunt
end are not visible outside, and in fact it is illegal to declare a
lexical on the blunt end that is not enclosed in curlies somehow.

Because feeds are defined as lazy pipes, a chain of feeds may not begin
and end with the same array without some kind of eager sequence point.
That is, this isn't guaranteed to work:

    @data <== grep { $_ % 2 } <== @data;

either of these do:

    @data <== grep { $_ % 2 } <== eager @data;
    @data <== eager grep { $_ % 2 } <== @data;

Conjecture: if the cloning process eagerly duplicates C<@data>, it could
be forced to work.  Not clear if this is desirable, since ordinary clones
just clone the container, not the value.

Leftward feeds are a convenient way of explicitly indicating the typical
right-to-left flow of data through a chain of operations:

    @oddsquares = map { $_**2 }, sort grep { $_ % 2 }, @nums;

    # perhaps more clearly written as...

    @oddsquares = do {
        map { $_**2 } <== sort <== grep { $_ % 2 } <== @nums;

Rightward feeds are a convenient way of reversing the normal data flow in a
chain of operations, to make it read left-to-right:

    @oddsquares = do {
        @nums ==> grep { $_ % 2 } ==> sort ==> map { $_**2 };

Note that something like the C<do> is necessary because feeds operate
at the statement level.  Parens would also work, since a statement is
expected inside:

    @oddsquares = (
        @nums ==> grep { $_ % 2 } ==> sort ==> map { $_**2 };

But as described below, you can also just write:

    @nums ==> grep { $_ % 2 } ==> sort ==> map { $_**2 } ==> @oddsquares;

If the operand on the sharp end of a feed is not a call to a variadic
operation, it must be something else that can be interpreted as a list
receiver, or a scalar expression that can be evaluated to produce an
object that does the C<KitchenSink> role, such as an C<IO> object.
Such an object provides C<.clear> and C<.push> methods that will
be called as appropriate to send data.  (Note that an C<IO> object
used as a sink will force eager evaluation on its pipeline, so the
next statement is guaranteed not to run till the file is closed.
In contrast, an C<Array> object used as a sink turns into a lazy

Any non-variadic object (such as an C<Array> or C<IO> object) used as a filter
between two feeds is treated specially as a I<tap> that merely captures
data I<en passant>.  You can safely install such a tap in an extended pipeline
without changing the semantics.  An C<IO> object used as a tap does not
force eager evaluation since the eagerness is controlled instead by the
downstream feed.

Any prefix list operator is considered a variadic operation, so ordinarily
a list operator adds any feed input to the end of its list.
But sometimes you want to interpolate elsewhere, so any contextualizer
with C<*> as an argument may be used to indicate the target of a
feed without the use of a temporary array:

    foo() ==> say @(*), " is what I meant";
    bar() ==> @(*).baz();

Likewise, an C<Array> used as a tap may be distinguished from an C<Array> used
as a translation function:

    numbers() ==> @array ==> bar()          # tap
    numbers() ==> @array[@(*)] ==> bar()    # translation

To append multiple sources to the next sink, double the angle:

    my $sink;
    0..*       ==>  $sink;
    'a'..*     ==>> $sink;
    pidigits() ==>> $sink;

    # outputs "(0, 'a', 3)\n"...
    for $sink.zip { .perl.say }

Each such append adds another slice element (that is, a parcel), to
the sink.  (The original feed also created a parcel.)

You may use a variable (or variable declaration) as a receiver, in
which case the list value is bound as the "todo" of the variable.
(The append form binds addition todos to the receiver's todo list.)
Do not think of it as an assignment, nor as an ordinary binding.
Think of it as iterator creation.  In the case of a scalar variable,
that variable contains the newly created iterator itself.  In the case
of an array, the new iterator is installed as the method for extending
the array.  As with assignment, the old todo list is clobbered; use the
append form to avoid that and get push semantics.  In any case, feeding
an array always flattens.  You must use the scalar form to preserve
slice information.

In general you can simply think of a receiver array as representing
the results of the chain, so you can equivalently write any of:

    my @oddsquares <== map { $_**2 } <== sort <== grep { $_ % 2 } <== @nums;

    my @oddsquares
        <== map { $_**2 }
        <== sort
        <== grep { $_ % 2 }
        <== @nums;

    @nums ==> grep { $_ % 2 } ==> sort ==> map { $_**2 } ==> my @oddsquares;

    ==> grep { $_ % 2 }
    ==> sort
    ==> map { $_**2 }
    ==> my @oddsquares;

Since the feed iterator is bound into the final variable, the variable
can be just as lazy as the feed that is producing the values.

When feeds are bound to arrays with "push" semantics, you can have
a receiver for multiple feeds:

    my @foo;
    0..2       ==>  @foo;
    'a'..'c'   ==>> @foo;
    say @foo;   # 0,1,2,'a','b','c'

Note how the feeds are concatenated in C<@foo> so that C<@foo>
is a list of 6 elements.  This is the default behavior.  However,
sometimes you want to capture the outputs as a list of two iterators,
namely the two iterators that represent the two input feeds.  You can
get at those two iterators by using a scalar instead, which
will preserve the slice structure, which can be fed to any operation
that knows how to deal with a list of parcels as a slice, such as C<zip>:

    0..*       ==>  $foo;
    'a'..*     ==>> $foo;
    pidigits() ==>> $foo;

    for $foo.zip { .say }


Here C<$foo> is a list of three parcels, so


is equivalent to

    my (@a,@b,@c) := |$foo;
    zip(@a; @b; @c)

A semicolon inside brackets is equivalent to stacked feeds.  The code above
could be rewritten as:

    (0..*; 'a'..*; pidigits()) ==> my $foo;
    for $foo.zip { .say }

which is in turn equivalent to

    for zip(0..*; 'a'..*; pidigits()) { .say }

A named receiver array is useful when you wish to feed into an
expression that is not an ordinary list operator, and you wish to be
clear where the feed's destination is supposed to be:

    picklist() ==> my @baz;
    my @foo = @bar[@baz];

Various contexts may or may not be expecting multi-dimensional slices
or feeds.  By default, ordinary arrays are flattened in slurpy context, that is, they
have "list" semantics.  If you say

    (0..2; 'a'..'c') ==> my @tmp;
    for @tmp { .say }

then you get 0,1,2,'a','b','c'.  If you have a multidim array, you
can ask for flattening semantics explicitly with C<flat>:

    (0..2; 'a'..'c') ==> my $tmp;
    for $tmp.flat { .say }

As we saw earlier, "zip" produces an interleaved result by taking one element
from each list in turn, so

    (0..2; 'a'..'c') ==> my $tmp;
    for $tmp.zip { .say }

produces 0,'a',1,'b',2,'c'.

If you want the zip's result as a list of subarrays, then you need to put
the zip itself into a "chunky" C<LoL> context instead:

    (0..2; 'a'..'c') ==> my $tmp;
    for $tmp.zip.lol { .say }

This produces two values on each line.  But usually you want the flat
form so you can just bind it directly to a signature:

    for $tmp.zip -> $i, $a { say "$i: $a" }

Otherwise you'd have to say this:

    for $tmp.zip.lol -> [$i, $a] { say "$i: $a" }

Conjectural: To interpolate a feed into a subscript such that the
slice turns into multiple subscripts, you need to use C<< prefix:<||> >>,
which works just like prefix:<|> except that it interpolates at the slice
level rather than the flat list level:

    (0..10; 0..10).zip
    zip(0..10; 0..10)
    (0..10; 0..10) ==> zip( ||@(*) )
    (0..10; 0..10) ==> $feed; zip( ||$feed )

Usually it's better to just use the method form.

Note that with the current definition, the order of feeds is preserved
left to right in general regardless of the position of the receiver.


    ('a'..*; 0..*) ==> $feed;
     for $feed.zip <== @foo) -> $a, $i, $x { ... }

is the same as

    'a'..* ==> $feed;
     0..*  ==>> $feed;
     for $feed.zip <== @foo) -> $a, $i, $x { ... }

which is the same as

    for zip('a'..*; 0..*; @foo) -> $a, $i, $x { ... }

Also note that these come out to be identical for ordinary arrays:


=head2 Closure parameters

Parameters declared with the C<&> sigil take blocks, closures, or
subroutines as their arguments. Closure parameters can be required,
optional, named, or slurpy.

    sub limited_grep (Int $count, &block, *@list) {...}

    # and later...

    @first_three = limited_grep 3, {$_<10}, @data;

(The comma is required after the closure.)

Within the subroutine, the closure parameter can be used like any other
lexically scoped subroutine:

    sub limited_grep (Int $count, &block, *@list) {
        if block($nextelem) {...}

The closure parameter can have its own signature in a type specification written
with C<:(...)>:

    sub limited_Dog_grep ($count, &block:(Dog), Dog *@list) {...}

and even a return type:

    sub limited_Dog_grep ($count, &block:(Dog --> Bool), Dog *@list) {...}

When an argument is passed to a closure parameter that has this kind of
signature, the argument must be a C<Code> object with a compatible
parameter list and return type.

=head2 En passant type capture

Unlike normal parameters, type parameters often come in piggybacked
on the actual value as "kind", and you'd like a way to capture both
the value and its kind at once.  (A "kind" is a storage type, that is, a class or type that
an object is allowed to be.  An object is not officially allowed
to take on a constrained or contravariant type.)  A type variable
can be used anywhere a type name can, but instead of asserting that
the value must conform to a particular type, it captures the
actual "kind" of the object and also declares a package/type name
by which you can refer to that kind later in the signature or body.
In addition, it captures the nominal typing of any associated nominal
For instance, if you wanted to match any two Dogs as long as they
were of the same kind, you can say:

    sub matchedset (Dog ::T $fido, T $spot) {...}

This actually turns into something more like

    sub matchedset (Dog ::T $fido, Dog $spot where T) {...}

Note that C<::T> is not required to contain C<Dog>, only
a type that is compatible with C<Dog>.  Note also that the nominal
type, C<Dog>, is also included in the meaning of C<T>, along
with the notion that the actual type must match the storage
type of C<$fido>.

The C<::> quasi-sigil is short for "subset" in much the same way that C<&> is
short for "sub".  Just as C<&> can be used to name any kind of code,
so too C<::> can be used to name any kind of type.  Both of them insert
a bare identifier into the symbol table, though they fill different syntactic

Note that it is not required to capture the object associated with the
class unless you want it.  The sub above could be written as

    sub matchedset (Dog ::T, T) {...}

if we're not interested in C<$fido> or C<$spot>.  Or just

    sub matchedset (::T, T) {...}

if we don't care about anything but the matching.  Note here that
the second parameter may be more derived than the first.  If you
need them to be identical, you must say something like

    sub matchedset (::T, $ where { $_.WHAT === T }

=head2 Unpacking array parameters

Instead of specifying an array parameter as an array:

    sub quicksort (@data, $reverse?, $inplace?) {
        my $pivot := shift @data;

it may be broken up into components in the signature, by
specifying the parameter as if it were an anonymous array of

    sub quicksort ([$pivot, *@data], $reverse?, $inplace?) {

This subroutine still expects an array as its first argument, just like
the first version.

=head2 Unpacking a single list argument

To match the first element of the slurpy list, use a "slurpy" scalar:

    sub quicksort (:$reverse, :$inplace, *$pivot, *@data)

=head2 Unpacking tree node parameters

You can unpack hash values and tree nodes in various dwimmy ways by enclosing the bindings
of child nodes and attributes in parentheses following the declaration of
the node itself:

    sub traverse ( BinTree $top ( $left, $right ) ) {

In this, C<$left> and C<$right> are automatically bound to the left
and right nodes of the tree.  If C<$top> is an ordinary object, it binds
the C<$top.left> and C<$top.right> attributes.  If it's a hash,
it binds C<< $top<left> >> and C<< $top<right> >>.  If C<BinTree> is a
signature type and $top is a C<Capture> (argument list) object, the child types
of the signature are applied to the actual arguments in the argument
list object.  (Signature types have the benefit that you can view
them inside-out as constructors with positional arguments, such that
the transformations can be reversible.)

However, the full power of signatures can be applied to pattern match
just about any argument or set of arguments, even though in some cases
the reverse transformation is not derivable.  For instance, to bind to
an array of children named C<.kids> or C<< .<kids> >>, use something

    proto traverse ($) {*}
    multi traverse ( NAry $top ( :kids [$eldest, *@siblings] ) ) {
        traverse(:kids(@siblings));  # (binds @siblings to $top)
    multi traverse ( $leaf ) {...}

The second candidate is called only if the parameter cannot be bound to
both C<$top> and to the "kids" parsing subparameter.

Likewise, to bind to a hash element of the node and then bind to
keys in that hash by name:

    sub traverse ( AttrNode $top ( :%attr{ :$vocalic, :$tense } ) ) {
        say "Has {+%attr} attributes, of which";
        say "vocalic = $vocalic";
        say "tense = $tense";

You may omit the top variable if you prefix the parentheses with a colon
to indicate a signature.  Otherwise you must at least put the sigil of
the variable, or we can't correctly differentiate:

    my Dog ($fido, $spot)   := twodogs();       # list of two dogs
    my Dog $ ($fido, $spot) := twodogs();       # one twodog object
    my Dog :($fido, $spot)  := twodogs();       # one twodog object

Sub signatures can be matched directly within regexes by using C<:(...)>

    push @a, "foo";
    push @a, \(1,2,3);
    push @a, "bar";
    my ($i, $j, $k);
    @a ~~ rx/
            <,>                         # match initial elem boundary
            :(Int $i,Int $j,Int? $k)    # match lists with 2 or 3 ints
            <,>                         # match final elem boundary
    say "i = $<i>";
    say "j = $<j>";
    say "k = $<k>" if defined $<k>;

If you want a parameter bound into C<$/>, you have to say C<< $<i> >>
within the signature.  Otherwise it will try to bind an external C<$i>
instead, and fail if no such variable is declared.

Note that unlike a sub declaration, a regex-embedded signature has no
associated "returns" syntactic slot, so you have to use C<< --> >>
within the signature to specify the C<of> type of the signature, or match as
an arglist:

    :(Num, Num --> Coord)
    :(\Coord(Num, Num))

A consequence of the latter form is that you can match the type of
an object with C<:(\Dog)> without actually breaking it into its components.
Note, however, that it's not equivalent to say

    :(--> Dog)

which would be equivalent to


that is, match a nullary function of type C<Dog>.  Nor is it equivalent to


which would be equivalent to


and match a function taking a single parameter of type Dog.

Note also that bare C<\(1,2,3)> is never legal in a regex since the
first (escaped) paren would try to match literally.

=head2 Attributive parameters

If a submethod's parameter is declared with a C<.> or C<!> after the
sigil (like an attribute):

    submethod initialize($.name, $!age) {}

then the argument is bound directly to the object's attribute of the
same name. This avoids the frequent need to write code like:

    submethod initialize($name, $age) {
        $.name = $name;
        $!age  = $age;

The initialization of attributes requires special care to preserve
encapsulation; therefore the default for attributive parameters is
value semantics, that is, as if specified with C<is copy>.  Hence,
the submethod above is really more like:

    submethod initialize($name is copy, $age is copy) {
        $.name := $name;      # or maybe = here, since it's a parent's attr
        $!age  := $age;       # or maybe only $! parameters work really

If you wish to allow the user to initialize an attribute by reference,
you may either write your own initializer submethod explicitly, or
simply mark the attributes you want to work that way with C<is ref>:

    has $!age is ref;   # BUILD will automatically use ref binding, not copy

To rename an attribute parameter you can use the explicit pair form:

    submethod initialize(:moniker($.name), :youth($!age)) {}

The C<:$name> shortcut may be combined with the C<$.name> shortcut,
but the twigil is ignored for the parameter name, so

    submethod initialize(:$.name, :$!age) {}

is the same as:

    submethod initialize(:name($.name), :age($!age)) {}

Note that C<$!age> actually refers to the private "C<has>" variable that
can be referred to as either C<$age> or C<$!age>.

=head2 Placeholder variables

Even though every bare block is a closure, bare blocks can't have
explicit parameter lists. Instead, they use "placeholder" variables,
marked by a caret (C<^>) or a colon (C<:>) after their sigils.
The caret marks positional placeholders, while the colon marks named

Using placeholders in a block defines an implicit parameter list. The
signature is the list of distinct positional placeholder names,
sorted in Unicode order, following by the named placeholder names in
any order.  So:

    { say "woof" if $:dog; $^y < $^z && $^x != 2 }

is a shorthand for:

    -> $x,$y,$z,:$dog { say "woof" if $dog; $y < $z && $x != 2 }

Note that placeholder variables syntactically cannot have type constraints.
Also, it is illegal to use placeholder variables in a block that already
has a signature, because the autogenerated signature would conflict with that.
Positional placeholder names consisting of a single uppercase letter are disallowed,
not because we're mean, but because it helps us catch references to
obsolete Perl 5 variables such as C<$^O>.

The C<$_> variable functions as a placeholder in a block without any
other placeholders or signature.  Any bare block without placeholders
really has a parameter like this:

    -> \$_ = OUTER::<$_> { .mumble }

(However, statement control C<if> notices this and passes no argument,
so C<$_> ends up being bound to the outer C<$_> anyway.)

A block may also refer to either C<@_> or C<%_> or both, each of
which will be added to generated signature as a normal readonly
slurpy pararmeter:

    { say $:what; warn "bad option: $_\n" for keys %_; }

turns into

    -> :$what, *%_ { say $what; warn "bad option: $_\n" for keys %_; }

If not used, they are not added, and a dispatch with mispatched parameters will fail.

The use of a P5ish C<@_> in a signatureless sub falls naturally out of this:

    sub sayall { .say for @_ }

Note that in this case, C<$_> is not treated as a placeholder because
there is already the C<@_> placeholder.  And C<@_> is a placeholder
only because the sub has no official signature.  Otherwise it would
be illegal (unless explicitly declared).

Placeholders may also be used in method bodies that have no
formal signature.  The invocant is always removed first, so the
first placeholder argument always refers to the first non-invocant
argument.  C<@_> will never contain the invocant.  The invocant is
always available via C<self>, of course.

Since the placeholder declares a parameter variable without the twigil,
the twigil is needed only on the first occurrence of the variable within
the block.  Subsequent mentions of that variable may omit the twigil.
Within an internal nested block the twigil I<must> be omitted, since
it would wrongly attach to the inner block.

Note that, unlike in Perl 5, C<@_> may not be used within an inner block to
refer to the outer block's arguments:

    sub say-or-print {
        if $SAYIT {
            say @_;     # WRONG
        else {
            print @_;   # WRONG

because this desugars to:

    sub say-or-print {
        if $SAYIT -> *@_ {
            say @_;
        else -> *@_ {
            print @_;

Translators of Perl 5 will need to bear this in mind.

=head1 Properties and traits

Compile-time properties are called "traits". The
C<is I<NAME> (I<DATA>)> syntax defines traits on containers and
subroutines, as part of their declaration:

    constant $pi is Approximated = 3;   # variable $pi has Approximated trait

    my $key is Persistent(:file<.key>);

    sub fib is cached {...}

The C<will I<NAME> I<BLOCK>> syntax is a synonym for C<is I<NAME> (I<BLOCK>)>:

    my $fh will undo { close $fh };    # Same as: my $fh is undo({ close $fh });

The C<but I<NAME> (I<DATA>)> syntax specifies run-time properties on values:

    constant $pi = 3 but Inexact;       # value 3 has Inexact property

    sub system {
        return $error but False if $error;
        return 0 but True;

Properties are predeclared as roles and implemented as mixins--see S12.

=head2 Subroutine traits

These traits may be declared on the subroutine as a whole (individual
parameters take other traits).  Trait syntax depends on the particular
auxiliary you use, but for C<is>, the subsequent syntax is identical to
adverbial syntax, except that that colon may be omitted or doubled depending
on the degree of ambiguity desired:

    sub x() is ::Foo[...]       # definitely a parameterized typename
    sub x() is :Foo[...]        # definitely a pair with a list
    sub x() is Foo[...]         # depends on whether Foo is predeclared as type


=item C<is signature>

The signature of a subroutine. Normally declared implicitly, by providing a
parameter list and/or return type.

=item C<as>/C<is as>

The C<inner> type constraint that a routine imposes on its return value.

=item C<of>/C<is of>

The C<of> type that is the official return type of the routine.  Or you
can think of "of" as outer/formal.  If there is no inner type, the outer
type also serves as the inner type to constrain the return value.

=item C<will do>

The block of code executed when the subroutine is called. Normally declared
implicitly, by providing a block after the subroutine's signature definition.

=item C<is rw>

Marks a subroutine as returning an lvalue.

=item C<is parsed>

Specifies the subrule by which a macro call is parsed.  The parse
always starts after the macro's initial token.  If the operator has
two parts (circumfix or postcircumfix), the final token is also automatically
matched, and should not be matched by the supplied regex.

=item C<is reparsed>

Also specifies the subrule by which a macro call is parsed, but restarts
the parse before the macro's initial token, usually because you want
to parse using an existing rule that expects to traverse the initial
token.  If the operator has two parts (circumfix or postcircumfix), the
final token must also be explicitly matched by the supplied regex.

=item C<is cached>

Marks a subroutine as being memoized, or at least memoizable.
In the abstract, this cache is just a hash where incoming argument
C<Capture>s are mapped to return values.  If the C<Capture> is found in
the hash, the return value need not be recalculated.  If you use
this trait, the compiler will assume two things:


=item *

A given C<Capture> would always calculate the same return value.  That is,
there is no state hidden within the dynamic scope of the call.

=item *

The cache lookup is likely to be more efficient than recalculating
the value in at least some cases, because either most uncached calls
would be slower (and reduce throughput), or you're trying to avoid a
significant number of pathological cases that are unacceptably slow
(and increase latency).


This trait is a suggestion to the compiler that caching is okay.  The
compiler is free to choose any kind of caching algorithm (including
non-expiring, random, lru, pseudo-lru, or adaptive algoritms, or
even no caching algorithm at all).  The run-time system is free to
choose any kind of maximum cache size depending on the availability
of memory and trends in usage patterns.  You may suggest a particular
cache size by passing a numeric argument (representing the maximum number
of unique C<Capture> values allowed), and some of the possible
algorithms may pay attention to it.  You may also pass C<*> for the
size to request a non-expiring cache (complete memoization).  The
compiler is free to ignore this too.

The intent of this trait is to specify performance hints without
mandating any exact behavior.  Proper use of this trait should not
change semantics of the program; it functions as a kind of "pragma".
This trait will not be extended to reinvent other existing ways of
achieving the same effect.  To gain more control, write your own
trait handler to allow the use of a more specific trait, such as
"C<is lru(42)>".  Alternately, just use a state hash keyed on the
sub's argument capture to write your own memoization with complete
control from within the subroutine itself, or from within a wrapper
around your subroutine.

=item C<is tighter>/C<is looser>/C<is equiv>

Specifies the precedence of an operator relative to an existing
operator.  C<tighter> and C<looser> precedence levels default to
being left associative.  They define a new precedence level slighty
tighter or looser than the precedence level on which they're based.
Both C<tighter> and C<looser> may be specified, in which case the new
precedence level is generated midway between the specified levels.

Two different declarations using the same precedence derivation end
up at the same precedence level, as if C<equiv> was specified instead
of C<tighter>/C<looser>, and the second will clone the associativity
of the first.  If the second explicitly specifies an associativity
that differs from the first, unexpected parsing conflicts may result.
(See S03.)

In addition to cloning the precedence level,
C<equiv> also clones other traits, so it specifies
the default associativity to be the same as the operator to which
the new operator is equivalent.  The following are the default
equivalents for various syntactic categories if neither C<equiv> nor
C<assoc> is specified.  (Many of these have no need of precedence
or associativity because they are parsed specially.  Nevertheless,
C<equiv> may be useful for cloning other traits of these operators.)

    circumfix:<( )>
    postcircumfix:<( )>

The existing operator may be specified either as a function object
or as a string argument equivalent to the one that would be used in
the complete function name.  In string form the syntactic
category will be assumed to be the same as the new declaration.
Therefore these all have the same effect:

    sub postfix:<!> ($x) is equiv(&postfix:<++>) {...}
    sub postfix:<!> ($x) is equiv<++> {...}
    sub postfix:<!> ($x) {...}      # since equiv<++> is the default

Prefix operators that are identifiers are handled specially.  The
form with one argument defaults to named unary precedence
instead of autoincrement precedence:

    sub prefix:<foo> ($x) {...}
    foo 1, 2, 3;        # means foo(1), 2, 3

Likewise postfix operators that look like method calls are forced to
default to the precedence of method calls.  Any prefix operator that
requires multiple arguments defaults to listop precedence, even if it
is not an identifier:

    sub prefix:<☎> ($x,$y) {...}
    ☎ 1;                # ERROR, too few arguments
    ☎ 1, 2;             # okay
    ☎ 1, 2, 3;          # ERROR, too many arguments

You must use the C<< prefix:<foo> >> form in order to mutate
the grammar to parse as a named unary operator.  Normal function
definitions never change the grammar, and when called always parse
as listops, even if defined with a single argument:

    sub foo ($x) {...}  # a listop
    foo(1), 2, 3;       # okay
    (foo 1), 2, 3;      # okay
    foo 1, 2, 3;        # ERROR, too many arguments

Likewise 0-ary functions parse as listops.  Use C<< term:<foo> >>
(or a constant or enum declaration) to declare a term that expects
no arguments.

=item C<is assoc>

Specifies the associativity of an operator explicitly.  Valid values are:

    Tag         Examples        Meaning of $a op $b op $c       Default equiv
    ===         ========        =========================       =============
    left        + - * / x       ($a op $b) op $c                +
    right       ** =            $a op ($b op $c)                **
    non         cmp <=> ..      ILLEGAL                         cmp
    chain       == eq ~~        ($a op $b) and ($b op $c)       eqv
    list        | & ^ Z         op($a; $b; $c)                  |

Note that operators "C<equiv>" to relationals are automatically considered
chaining operators.  When creating a new precedence level, the chaining
is determined by the presence or absence of "C<< is assoc<chain> >>",
and other operators defined at that level are required to be the same.

Specifying an C<assoc> without an explicit C<equiv> substitutes a default
C<equiv> consistent with the associativity, as shown in the final column above.

=item C<PRE>/C<POST>

These phasers declare statements or blocks that are to be unconditionally
executed before/after the subroutine's C<do> block. They must return
a true value, otherwise an exception is thrown.

When applied to a method, the semantics provide support for the
"Design by Contract" style of OO programming: a precondition of
a particular method is met if all the C<PRE> phasers associated
with that method return true.  Otherwise, the precondition is met
if C<all> of the parent classes' preconditions are met (which may
include the preconditions of I<their> parent classes if they fail,
and so on recursively.)

In contrast, a method's postcondition is met if all the method's C<POST> phasers
return true I<and> all its parents' postconditions are also met recursively.

C<POST> phasers (and "C<will post>" phaser traits) declared within a C<PRE>
or C<ENTER> block are automatically hoisted outward to be called at the
same time as other C<POST> phasers.  This conveniently gives "circum"
semantics by virtue of wrapping the post lexical scope within the pre
lexical scope.  That is, the C<POST> closes over its outer scope, even
if that scope is gone by the time the C<POST> is run.

    method push ($new_item) {
        ENTER {
            my $old_height = self.height;
            POST { self.height == $old_height + 1 }

        $new_item ==> push @.items;

    method pop () {
        ENTER {
            my $old_height = self.height;
            POST { self.height == $old_height - 1 }

        return pop @.items;

Note that C<self> is available in phasers defined within methods.

Class invariants are declared with C<PRE>/C<POST> submethods instead of phasers.

Module invariants are declared with C<PRE>/C<POST> subs or protos.

[Conjecture: class and module invariants can applied more selectively
by marking C<PRE>/C<POST> declarations with a C<selective> trait that
stops it from running on internal calls (which might allow temporary
violations of invariants), but enforces the invariants when any routine
of this module is called from "outside" the current module or type,
however that's defined.  There could be arguments to this trait that
could refine the concept of what is foreign.]

=item C<ENTER>/C<LEAVE>/C<KEEP>/C<UNDO>/etc.

These phasers supply code that is to be conditionally executed before or
after the subroutine's C<do> block (only if used at the outermost level
within the subroutine; technically, these are added to the block traits on the C<do>
block, not the subroutine object).  These phasers are generally used only
for their side effects, since most return values will be ignored.  (Phasers
that run before normal execution may be used for their values, however.)


=head2 Parameter traits

The following traits can be applied to many types of parameters.


=item C<is readonly>

Specifies that the parameter cannot be modified (e.g. assigned to,
incremented). It is the default for parameters.  On arguments which
are already immutable values it is a no-op at run time; on mutable
containers it may need to create an immutable alias to the mutable object
if the constraint cannot be enforced entirely at compile time.  Binding
to a readonly parameter never triggers autovivification.

=item C<is rw>

Specifies that the parameter can be modified (assigned to, incremented,
etc). Requires that the corresponding argument is an lvalue or can be
converted to one.  Since this option forces an argument to be required,
it cannot co-exist with the C<?> mark to make an argument optional.
(It may, however, be used with C<=> indicating a default, but only
if the default expression represents something that is nameable at
compile time and that can bind as an lvalue, such as
C<< CALLER::<$/> >> or C<< OUTER::<$_> >>.)

When applied to a variadic parameter, the C<rw> trait applies to each
element of the list:

    sub incr (*@vars is rw) { $_++ for @vars }

(The variadic array as a whole is always modifiable, but such
modifications have no effect on the original argument list.)

=item C<is parcel>

Specifies that the parameter is passed as a "parcel", that is, as a
raw reference to a "parenthetical cell" or "parse list", an argument
object that has not yet had a context imposed.  (It may or may not
actually be of the C<Parcel> type, since a C<Parcel> containing only
one object automatically unwraps itself to be that object directly.)
In other words, this provides for lazy contextualization even through
function calls.  This is important if you wish to pass the parameter
onward to something else that will determine its context later.

You may modify the argument, but only if argument is already a suitable
lvalue since, unlike C<rw>, no attempt at autovivification is made,
so unsuitable lvalues will throw an exception if you try to modify
them within the body of the routine.  That is, if autovivification
happens, it happens at the point of use, not at the point of binding.

For better visual distinction, such a parameter is declared by
prefixing with a backslash rather than by using C<is parcel> directly.
The backslash is also more succint; the trait is there primarily
for introspection.

=item C<is copy>

Specifies that the parameter receives a distinct, read-writable copy of the
original argument. This is commonly known as "pass-by-value".

    sub reprint ($text, $count is copy) {
        print $text while $count-- > 0;

Binding to a copy parameter never triggers autovivification.

=item C<is dynamic>

Specifies that the parameter is to be treated as an "environmental"
variable, that is, a lexical that is accessible from the dynamic
scope (see S02).


=head2 Signature Introspection

A C<Signature> object can be introspected to find out the details of
the parameters it is defined as expected. The C<.params> method will
return a C<Parcel> of C<Parameter> objects, which have the following
readonly properties:

  name           The name of the lexical variable to bind to, if any
  type           The main type (the one multi-dispatch sorts by)
  constraints    Any further type constraints
  type_captures  List of names the argument type is captured into
  readonly       True if the parameter has C<is readonly> trait
  rw             True if the parameter has C<is rw> trait
  copy           True if the parameter has C<is copy> trait
  named          True if the parameter is to be passed named
  named_names    List of names a named parameter can be passed as
  capture        True if the paramater binds the caller's Capture
  parcel         True if the parameter is too lazy to contextualize
  slurpy         True if the parameter is slurpy
  optional       True if the parameter is optional
  default        A closure returning the default value
  invocant       True if the parameter is a method invocant
  multi_invocant True if the parameter is a multi invocant
  signature      A nested signature to bind the argument against

Note that C<constraints> will be something that can be smart-matched
against if it is defined; if there are many constraints it may be a
C<Junction> of some sort, but if there is just one it may be simply
that one thing.

Further, various things that appear in an original written signature
will have been deconstructed a bit. For example, a signature like:


Will introspect the same way as:

  :(Int $ where 1)

And if we have:

  subset Odd of Int where { $^n % 2 };
  sub foo(Odd $x) { ... }

Then the signature of foo will be equivalent to something like:

  :(Int $x where { $^n % 2 })

That is, the refinement type will have been deconstructed into the
part that nominal type that multiple dispatch uses for sorting the
candidates and an additional constraint.

=head1 Advanced subroutine features

=head2 Processing of returned values

It is a general policy that lvalues should only be returned up the dynamic
call stack if specifically requested.  Therefore, by default the returned
arguments are processed to enforce this, both for the implicit return from
the last statement of any block, as well as the explicit return done by
operators such as:


Specifically, this processing involves examining the returned parcel's
arguments and dereferencing any container that could be used as an lvalue,
replacing it with the container's value.

To override this processing for a routine, it must be declared C<rw>, or
the form C<return-rw> must be used.

To override for a C<gather>, use C<gather-rw> instead, or C<take-rw> on the
individual take.

Since blocks don't generally have traits, you must use C<leave-rw> to pass
an lvalue out of a block.

=head2 The C<return> function

The C<return> function notionally throws a control exception that is
caught by the current lexically enclosing C<Routine> to force a return
through the control logic code of any intermediate block constructs.
(That is, it must unwind the stack of dynamic scopes to the proper
lexical scope belonging to this routine.)  With normal blocks
(those that are autoexecuted in place because they're known to the
compiler) this unwinding can likely be optimized away to a "goto".
All C<Routine> declarations have an explicit declarator such as C<sub>
or C<method>; bare blocks and "pointy" blocks are never considered
to be routines in that sense.  To return from a block, use C<leave>
instead--see below.

The C<return> function preserves its argument list as a C<Capture> object, and
responds to the left-hand C<Signature> in a binding.  This allows named return
values if the caller expects one:

    sub f () { return :x<1> }
    sub g ($x) { print $x }

    my $x := |(f);  # binds 1 to $x, via a named argument
    g(|(f));        # prints 1, via a named argument

To return a literal C<Pair> object, always put it in an additional set of

    return( (:x<1>), (:y<2>) ); # two positional Pair objects

Note that the postfix parentheses on the function call don't count as
being "additional".  However, as with any function, whitespace after the
C<return> keyword prevents that interpretation and turns it instead
into a list operator:

    return :x<1>, :y<2>; # two named arguments (if caller uses |)
    return ( :x<1>, :y<2> ); # two positional Pair objects

If the function ends with an expression without an explicit C<return>,
that expression is also taken to be a C<Capture>, just as if the expression
were the argument to a C<return> list operator (with whitespace):

    sub f { :x<1> } # named-argument binding (if caller uses |)
    sub f { (:x<1>) } # always just one positional Pair object

On the caller's end, the C<Capture> is interpolated into any new argument list
much like an array would be, that is, as an item in item context, and as a
list in list context.  This is the default behavior, but the
caller may use C<< prefix:<|> >> to inline the returned values as part of the
new argument list.  The caller may also bind the returned C<Capture> directly.

A function is called only once at the time the
C<Capture> object is generated, not when it is later bound (which
could happen more than once).

=head2 The C<callframe> and C<caller> functions

The C<callframe> function takes a list of matchers and interprets them
as a navigation path from the current call frame to a location in the
call stack, either the current call frame itself or some frame
from which the current frame was called.  It returns an object
that describes that particular call frame, or a false value if
there is no such scope.  Numeric arguments are interpreted as number
of frames to skip, while non-numeric arguments scan outward for a
frame matching the argument as a smartmatch.

The current frame is accessed with a null argument list.

    say " file ", callframe().file,
        " line ", callframe().line;

which is equivalent to:

    say " file ", DYNAMIC::<$?FILE>,
        " line ", DYNAMIC::<$?LINE>;

The immediate caller of this frame is accessed by skipping one level:

    say " file ", callframe(1).file,
        " line ", callframe(1).line;

You might think that that must be the current function's caller,
but that's not necessarily so.  This might return an outer block in
our own routine, or even some function elsewhere that implements a
control operator on behalf of our block.  To get outside your current
routine, see C<caller> below.

The C<callframe> function may be given arguments
telling it which higher scope to look for.  Each argument is processed
in order, left to right.  Note that C<Any> and C<0> are no-ops:

    $ctx = callframe();        # currently running frame for &?BLOCK
    $ctx = callframe(Any);     # currently running frame for &?BLOCK
    $ctx = callframe(Any,Any); # currently running frame for &?BLOCK
    $ctx = callframe(1);       # my frame's caller
    $ctx = callframe(2);       # my frame's caller's caller
    $ctx = callframe(3);       # my frame's caller's caller's caller
    $ctx = callframe(1,0,1,1); # my frames's caller's caller's caller
    $ctx = callframe($i);      # $i'th caller

Note also that negative numbers are allowed as long as you stay within
the existing call stack:

    $ctx = callframe(4,-1);    # my frames's caller's caller's caller

Repeating any smartmatch just matches the same frame again unless you
intersperse a 1 to skip the current level:

    $ctx = callframe(Method);              # nearest frame that is method
    $ctx = callframe(Method,Method);       # nearest frame that is method
    $ctx = callframe(Method,1,Method);     # 2nd nearest method frame
    $ctx = callframe(Method,1,Method,1)    # caller of that 2nd nearest method
    $ctx = callframe(1,Block);             # nearest outer frame that is block
    $ctx = callframe(Sub,1,Sub,1,Sub);     # 3rd nearest sub frame
    $ctx = callframe({ .labels.any eq 'Foo' }); # nearest frame labeled 'Foo'

Note that this last potentially differs from the answer returned by


which returns the frame of the innermost C<Foo> block in the lexical scope
rather than the dynamic scope.  A call frame also responds to the C<.callframe>
method, so a given frame may be used as the basis for further navigation:

    $ctx = callframe(Method,1,Method);
    $ctx = callframe(Method).callframe(1).callframe(Method);     # same

You must supply args to get anywhere else, since C<.callframe> is
the identity operator when called on something that is already
a C<Context>:

    $ctx = callframe;
    $ctx = callframe.callframe.callframe.callframe;                # same

The C<caller> function is special-cased to go outward just far enough
to escape from the current routine scope, after first ignoring any
inner blocks that are embedded, or are otherwise pretending to be "inline":

    &caller ::= &callframe.assuming({ !.inline }, 1);

Note that this is usually the same as C<callframe(&?ROUTINE,1)>,
but not always.  A call to a returned closure might not even have
C<&?ROUTINE> in its dynamic scope anymore, but it still has a caller.

So to find where the current routine was called you can say:

    say " file ", caller.file,
        " line ", caller.line;

which is equivalent to:

    say " file ", CALLER::<$?FILE>,
        " line ", CALLER::<$?LINE>;

Additional arguments to C<caller> are treated as navigational from the
calling frame.  One frame out from your current routine is I<not>
guaranteed to be a C<Routine> frame.  You must say C<caller(Routine)>
to get to the next-most-inner routine.

Note that C<caller(Routine).line> is not necessarily going to give you the
line number that your current routine was called from; you're rather
likely to get the line number of the topmost block that is executing
within that outer routine, where that block contains the call to
your routine.

For either C<callframe> or C<caller>,
the returned CallFrame object supports at least the following methods:


The C<.callframe> and C<.caller> methods work the same as the functions
except that they are relative to the frame supplied as invocant.
The C<.leave> method can force an immediate return from the
specified call frame.

The C<.inline> method says whether this block was entered implicitly
by some surrounding control structure.  Any time you invoke a block or
routine explicitly with C<.()> this is false.  However, it is defined
to be true for any block entered using dispatcher-level primitives
such as C<.callwith>, C<.callsame>, C<.nextwith>, or C<.nextsame>.

The C<.my> method provides access to the lexical namespace associated with
the given call frame's current position.  It may be used to look
up ordinary lexical variables in that lexical scope.  It must not be
used to change any lexical variable that is marked as readonly.

The C<.hints> method gives access to a snapshot of compiler symbols in
effect at the point of the call when the call was originally compiled.
(For instance, C<caller.hints('&?ROUTINE')> will give you the caller's
routine object.)  Such values are always read-only, though in the
case of some (like the caller's routine above) may return a fixed
object that is nevertheless mutable.

=head2 The C<want> function

The C<want> function is gone. If you want context specific behavior,
return an object instead that responds accordingly to the various contextual

(Conjecture: in future we might want to provide some syntactic sugar that
makes it easier to create such objects. Or maybe a type that takes values
or code references for the various contexts, so that you can write

   return ContextProxy.new:
        Int   => 3,
        item  => { @.list.join(', ') },
        list  => { ...               },

or something similar.)

=head2 The C<leave> function

As mentioned above, a C<return> call causes the innermost surrounding
subroutine, method, rule, token, regex (as a keyword) or macro
to return.  Only declarations with an explicit declarator keyword
(C<sub>, C<submethod>, C<method>, C<macro>, C<regex>, C<token>, and
C<rule>) may be returned from.  Statement prefixes such a C<do> and
C<try> do not fall into that category.
You cannot use C<return> to escape directly into the surrounding
context from loops, bare blocks, pointy blocks, or quotelike operators
such as C<rx//>; a C<return> within one of those constructs will
continue searching outward for a "proper" routine to return from.
Nor may you return from property blocks such as C<BEGIN> or C<CATCH>
(though blocks executing within the lexical and dynamic scope of a
routine can of course return from that outer routine, which means
you can always return from a C<CATCH> or a C<FIRST>, but never from
a C<BEGIN> or C<INIT>.)

To return from blocks that aren't routines, the C<leave> method is used
instead.  (It can be taken to mean either "go away from" or "bequeath
to your successor" as appropriate.)  The object specifies the scope to exit,
and the method's arguments specify the return value.  If the object
is omitted (by use of the function or listop forms), the innermost
block is exited.  Otherwise you must use something like C<callframe>
or C<&?BLOCK> or a dynamic variable to specify the scope you
want to exit.  A label (such as a loop label) previously seen in
the lexical scope also works as a kind of singleton dynamic object:
it names a statement that is serving both as an outer lexical scope
and as a frame in the current dynamic scope.

As with C<return>, the arguments are taken to be a C<Capture> holding the
return values.

    leave;                      # return from innermost block of any kind
    callframe(Method).leave;    # return from innermost calling method
    &?ROUTINE.leave(1,2,3);     # Return from current sub. Same as: return 1,2,3
    &?ROUTINE.leave <== 1,2,3;  # same thing, force return as feed
    OUTER.leave;                # Return from OUTER label in lexical scope
    &foo.leave: 1,2,3;          # Return from innermost surrounding call to &foo

Note that these are equivalent in terms of control flow:

    last COUNT;

However, the first form explicitly sets the return value for the
entire loop, while the second implicitly returns all the previous
successful loop iteration values as a list comprehension.  (It may,
in fact, be too late to set a return value for the loop if it is
being evaluated lazily!)  A C<leave>
from the inner loop block, however, merely specifies the return value for
that iteration:

    for 1..10 { leave $_ * 2 }   # 2..20

Note that this:

    leave COUNT;

will always be taken as the function, not the method, so it returns
the C<COUNT> object from the innermost block.  The indirect object form
of the method always requires a colon:

    leave COUNT: ;

=head2 Temporization

The C<temp> macro temporarily replaces the value of an existing
variable, subroutine, context of a function call, or other object in a given scope:

       temp $*foo = 'foo';      # Temporarily replace global $foo
       temp &bar := sub {...};  # Temporarily replace sub &bar
    } # Old values of $*foo and &bar reinstated at this point

C<temp> invokes its argument's C<.TEMP> method. The method is expected
to return a C<Callable> object that can later restore the current
value of the object. At the end of the lexical scope in which the
C<temp> was applied, the subroutine returned by the C<.TEMP> method is

The default C<.TEMP> method for variables simply creates
a closure that assigns the variable's pre-C<temp> value
back to the variable.

New kinds of temporization can be created by writing storage classes with
their own C<.TEMP> methods:

    class LoudArray is Array {
        method TEMP {
            print "Replacing $.WHICH() at {caller.location}\n";
            my $restorer = $.SUPER::TEMP();
            return {
                print "Restoring $.WHICH() at {caller.location}\n";

You can also modify the behaviour of temporized code structures, by
giving them a C<TEMP> block. As with C<.TEMP> methods, this block is
expected to return a closure, which will be executed at the end of
the temporizing scope to restore the subroutine to its pre-C<temp> state:

    my $next = 0;
    sub next {
        my $curr = $next++;
        TEMP {{ $next = $curr }}  # TEMP block returns the closure { $next = $curr }
        return $curr;

    # and later...

    say next();     # prints 0; $next == 1
    say next();     # prints 1; $next == 2
    say next();     # prints 2; $next == 3
    if ($hiccough) {
        say temp next();  # prints 3; closes $curr at 3; $next == 4
        say next();       # prints 4; $next == 5
        say next();       # prints 5; $next == 6
    }                     # $next = 3
    say next();     # prints 3; $next == 4
    say next();     # prints 4; $next == 5

Note that C<temp> must be a macro rather than a function because the
temporization must be arranged before the function causes any state
changes, and if it were a normal argument to a normal function, the state
change would be happen before C<temp> got control.

Hypothetical variables use the same mechanism, except that the restoring
closure is called only on failure.

Note that dynamic variables may be a better solution than temporized
globals in the face of multithreading.

=head2 Wrapping

Every C<Routine> object has a C<.wrap> method. This method expects a
single C<Callable> argument. Within the code, the special C<callsame>,
C<callwith>, C<nextsame> and C<nextwith> functions will invoke the
original routine, but do not introduce an official C<CALLER> frame:

    sub thermo ($t) {...}   # set temperature in Celsius, returns old value

    # Add a wrapper to convert from Fahrenheit...
    $handle = &thermo.wrap( { callwith( ($^t-32)/1.8 ) } );

The C<callwith> function lets you pass your own arguments to the wrapped
function.  The C<callsame> function takes no argument; it
implicitly passes the original argument list through unchanged.
The C<callsame> and C<nextsame> functions are really short for:

    callwith( |callframe(Routine).args )
    nextwith( |callframe(Routine).args )

The call to C<.wrap> replaces the original C<Routine>'s C<do> property with the C<Callable>
argument, and arranges that any call to C<callsame>, C<callwith>,
C<nextsame> or C<nextwith> invokes the previous version of the
routine. In other words, the call to C<.wrap> has more or less the
same effect as:

    my &old_thermo := &thermo;
    &thermo = sub ($t) { old_thermo( ($t-32)/1.8 ) }

Note that C<&thermo.WHICH> stays the same after the C<.wrap>,
as it does with the equivalent assignment shown since assignment
to a C<Routine> works like a container, changing the contained C<do>
property but not the container itself.

The call to C<.wrap> returns a unique handle that has a C<restore> method
that will undo the wrapping:


This does not affect any other wrappings placed to the routine.

A wrapping can also be restricted to a particular dynamic scope with

    # Add a wrapper to convert from Kelvin
    # wrapper self-unwraps at end of current scope
    temp &thermo.wrap( { callwith($^t + 273.16) } );

The entire argument list may be captured by binding to a C<Capture> parameter.
It can then be passed to C<callwith> using that name:

    # Double the return value for &thermo
    &thermo.wrap( -> |$args { callwith(|$args) * 2 } );

In this case only the return value is changed.

The wrapper is not required to call the original routine; it can call another
C<Callable> object by passing the C<Capture> to its C<callwith> method:

    # Transparently redirect all calls to &thermo to &other_thermo
    &thermo.wrap( sub (|$args) { &other_thermo.callwith(|$args) } );

or more briefly:

    &thermo.wrap( { &other_thermo.callsame } );

Since the method versions of C<callsame>, C<callwith>, C<nextsame>,
and C<nextwith> specify an explicit destination, their semantics do
not change outside of wrappers.  However, the corresponding functions
have no explicit destination, so instead they implicitly call the
next-most-likely method or multi-sub; see S12 for details.

As with any return value, you may capture the returned C<Capture> of C<call>
by binding:

    my |$retval := callwith(|$args);
    ... # postprocessing
    return |$retval;

Alternately, you may prevent any return at all by using the variants
C<nextsame> and C<nextwith>.  Arguments are passed just as with
C<callsame> and C<callwith>, but a tail call is explicitly enforced;
any code following the call will be unreached, as if a return had
been executed there before calling into the destination routine.

Within an ordinary method dispatch these functions treat the rest
of the dispatcher's candidate list as the wrapped function, which
generally works out to calling the same method in one of our parent
(or older sibling) classes.  Likewise within a multiple dispatch the
current routine may defer to candidates further down the candidate
list.  Although not necessarily related by a class hierarchy, such
later candidates are considered more generic and hence likelier
to be able to handle various unforeseen conditions (perhaps).

Note that all routines are (by default) considered to be candidates
for inlining and constant folding.  The optimizer is allowed to start
making these optimizations after the main program's C<CHECK> time,
but not before.  After any routine is "hard" inlined or constant
folded, it is explicitly retyped as immutable; any attempt to
wrap an immutable routine will result in failure of the wrap call.
An immutable routine is so marked by recasting to type C<HardRoutine>,
a subclass of C<Routine>.

On the other hand, it is also possible to explicitly mark a routine
as mutable, and then the ability to wrap it must be preserved even
after C<CHECK> time.  This is done by recasting to C<SoftRoutine>.
Explicitly marking a routine as either mutable or immutable should
be considered permanent.  It is still possible to inline soft
routines, but only if the possibility of indirection is detected
inline as well, and provision made (either inline or via external
rewriting) for dealing with any wrappers.

Hence, any routine marked as soft before C<CHECK> time is exempt
from hard inlining or folding.  There are several methods of
marking a routine as soft, however no method is provided for marking
routines as hard, since that is the job of the optimizer.
A routine may be marked as soft:


=item *

if it is declared using "our" explicitly

=item *

if it is mentioned in the argument list of a C<use soft> pragma,

=item *

if its name matches any wildcard pattern (TBD) in a C<use soft>,

=item *

if the module or class in which it is defined is mentioned in a C<use soft>,

=item *

or if there is a general C<use soft *;> declaration, which basically
turns on AOP for everything at run time.  Be aware that this may turn
your optimizer into more of a "pessimizer".


For any normal standalone application, any C<use soft> pragma applies
to the entire program in which it participates, provided it is
performed before C<CHECK> time.  The optimizer may then harden
anything that was not requested to remain soft.

A plug-in system, such as for a web server, may choose either to allow
individual plug-ins to behave as independent programs by letting the
optimizer harden individual plug-ins independently, or treat all
plug-ins as a part of the same program by softening all plug-ins.
(Similar considerations apply to optimizing classes to closed/final.)

Note that installing a wrapper before C<CHECK> time is specifically
I<not> one of the ways to mark a routine as soft.  Such a routine
may still be hardened at C<CHECK> time despite being wrapped during
compile time.

=head2 The C<&?ROUTINE> object

C<&?ROUTINE> is always an alias for the lexically innermost C<Routine>
(which may be a C<Sub>, C<Method>, or C<Submethod>), so you can specify
recursion on an anonymous sub:

    my $anonfactorial = sub (Int $n) {
                            return 1 if $n<2;
                            return $n * &?ROUTINE($n-1);

You can get the current routine name by calling C<&?ROUTINE.name>.
Outside of any sub declaration, this call returns failure.

Note that C<&?ROUTINE> refers to the current single sub, even if it is
declared C<multi>.  To redispatch to the entire suite under a given short
name, just use the named form to call the C<proto>, since there are no anonymous C<multi>s.

=head2 The C<&?BLOCK> object

C<&?BLOCK> is always an alias for the current block, so you can
specify recursion on an anonymous block:

    my $anonfactorial = -> Int $n { $n < 2
                                        ?? 1
                                        !! $n * &?BLOCK($n-1)

C<&?BLOCK.labels> contains a list of all labels of the current block.
This is typically matched by saying

    if &?BLOCK.labels.any eq 'Foo' {...}

If the innermost lexical block happens to be the main block of a C<Routine>,
then C<&?BLOCK> just returns the C<Block> object, not the C<Routine> object
that contains it.

[Note: to refer to any C<$?> or C<&?> variable at the time the sub or
block is being compiled, use the C<< COMPILING:: >> pseudopackage.]

=head2 Currying

Every C<Callable> object has a C<.assuming> method. This method does a partial
binding of a set of arguments to a signature and returns a new function
that takes only the remaining arguments.

    &textfrom := &substr.assuming(str=>$text, len=>Inf);

or equivalently:

    &textfrom := &substr.assuming(:str($text) :len(Inf));

or even:

    &textfrom := &substr.assuming:str($text):len(Inf);

It returns a C<Callable> object that implements the same behaviour
as the original subroutine, but has the values passed to C<.assuming>
already bound to the corresponding parameters:

    $all  = textfrom(0);   # same as: $all  = substr($text,0,Inf);
    $some = textfrom(50);  # same as: $some = substr($text,50,Inf);
    $last = textfrom(-1);  # same as: $last = substr($text,-1,Inf);

The result of a C<use> statement is a (compile-time) object that also has
a C<.assuming> method, allowing the user to bind parameters in all the
module's subroutines/methods/etc. simultaneously:

    (use IO::Logging).assuming(logfile => ".log");

This special form should generally be restricted to named parameters.

To curry a particular C<multi> variant, it may be necessary to specify the type
for one or more of its parameters to pick out a single function:

    &woof ::= &bark:(Dog).assuming :pitch<low>;
    &pine ::= &bark:(Tree).assuming :pitch<yes>;

=head2 Macros

Macros are functions or operators that are called by the compiler as
soon as their arguments are parsed (if not sooner).  The syntactic
effect of a macro declaration or importation is always lexically
scoped, even if the name of the macro is visible elsewhere.  As with
ordinary operators, macros may be classified by their grammatical
category.  For a given grammatical category, a default parsing rule or
set of rules is used, but those rules that have not yet been "used"
by the time the macro keyword or token is seen can be replaced by
use of "is parsed" trait.  (This means, for instance, that an infix
operator can change the parse rules for its right operand but not
its left operand.)

In the absence of a signature to the contrary, a macro is called as
if it were a method on the current match object returned from the
grammar rule being reduced; that is, all the current parse information
is available by treating C<self> as if it were a C<$/> object.
[Conjecture: alternate representations may be available if arguments
are declared with particular AST types.]

Macros may return either a string to be reparsed, or a syntax tree
that needs no further parsing.  The textual form is handy, but the
syntax tree form is generally preferred because it allows the parser
and debugger to give better error messages.  Textual substitution
on the other hand tends to yield error messages that are opaque to
the user.  Syntax trees are also better in general because they are
reversible, so things like syntax highlighters can get back to the
original language and know which parts of the derived program come
from which parts of the user's view of the program.  Nevertheless,
it's difficult to return a syntax tree for an unbalanced construct,
and in such cases a textual macro may be a clearer expression of the
evil thing you're trying to do.

If you call a macro at runtime, the result of the macro is automatically
evaluated again, so the two calls below print the same thing:

    macro f { '1 + 1' }
    say f();    # compile-time call to &f
    say &f();   # runtime call to &f

=head2 Quasiquoting

In aid of returning syntax tree, Perl provides a "quasiquoting"
mechanism using the quote C<quasi>, followed by a block intended to
represent an AST:

    return quasi { say "foo" };

Modifiers to the C<quasi> can modify the operation:

    :ast(MyAst)         # Default :ast(AST)
    :lang(Ruby)         # Default :lang($?PARSER)
    :unquote<[: :]>     # Default "triple rule"

Within a quasiquote, variable and function names resolve according
to the lexical scope of the macro definition.  Unrecognized symbols raise
errors when the macro is being compiled, I<not> when it's being used.

To make a symbol resolve to the (partially compiled) scope of the macro
call, use the C<COMPILING::> pseudo-package:

    macro moose () { quasi { $COMPILING::x } }

    moose(); # macro-call-time error
    my $x;
    moose(); # resolves to 'my $x'

If you want to mention symbols from the scope of the macro call, use the
import syntax as modifiers to C<quasi>:

    :COMPILING<$x>      # $x always refers to $x in caller's scope
    :COMPILING          # All free variables fallback to caller's scope

If those symbols do not exist in the scope of the compiling scope, a
compile-time exception is thrown at macro call time.

Similarly, in the macro body you may either refer to the C<$x> declared in the
scope of the macro call as C<$COMPILING::x>, or bind to them explicitly:

    my $x := $COMPILING::x;

You may also use an import list to bind multiple symbols into the
macro's lexical scope:

    require COMPILING <$x $y $z>;

Note that you need to use the run-time C<require> form, not
C<use>, because the macro caller's compile-time is the macro's runtime.

=head2 Splicing

Bare AST variables (such as the arguments to the macro) may not be
spliced directly into a quasiquote because they would be taken as
normal bindings.  Likewise, program text strings to be inserted need
to be specially marked or they will be bound normally.  To insert an
"unquoted" expression of either type within a quasiquote, use the
quasiquote delimiter tripled, typically a bracketing quote of some sort:

    return quasi { say $a + {{{ $ast }}} }
    return quasi [ say $a + [[[ $ast ]]] ]
    return quasi < say $a + <<< $ast >>> >
    return quasi ( say $a + ((( $ast ))) )

The delimiters don't have to be bracketing quotes, but the following
is probably to be construed as Bad Style:

    return quasi / say $a + /// $ast /// /

(Note to implementors: this must not be implemented by finding
the final closing delimiter and preprocessing, or we'll violate our
one-pass parsing rule.  Perl 6 parsing rules are parameterized to know
their closing delimiter, so adding the opening delimiter should not
be a hardship.  Alternately the opening delimiter can be deduced from
the closing delimiter.  Writing a rule that looks for three opening
delimiters in a row should not be a problem.  It has to be a special
grammar rule, though, not a fixed token, since we need to be able to
nest code blocks with different delimiters.  Likewise when parsing the
inner expression, the inner parser subrule is parameterized to know that
C<}}}> or whatever is its closing delimiter.)

Unquoted expressions are inserted appropriately depending on the
type of the variable, which may be either a syntax tree or a string.
(Again, syntax tree is preferred.)  The case is similar to that of a
macro called from within the quasiquote, insofar as reparsing only
happens with the string version of interpolation, except that such
a reparse happens at macro call time rather than macro definition
time, so its result cannot change the parser's expectations about
what follows the interpolated variable.

Hence, while the quasiquote itself is being parsed, the syntactic
interpolation of a unquoted expression into the quasiquote always
results in the expectation of an operator following the unquote.
(You must use a call to a submacro if you want to expect something
else.)  Of course, the macro definition as a whole can expect
whatever it likes afterwards, according to its syntactic category.
(Generally, a term expects a following postfix or infix operator,
and an operator expects a following term or prefix operator.  This
does not matter for textual macros, however, since the reparse of
the text determines subsequent expectations.)

Quasiquotes default to hygienic lexical scoping, just like closures.
The visibility of lexical variables is limited to the quasi expression
by default.  A variable declaration can be made externally visible using
the C<COMPILING::> pseudo-package.  Individual variables can be made visible,
or all top-level variable declarations can be exposed using the
C<quasi :COMPILING> form.

Both examples below will add C<$new_variable> to the lexical scope of
the macro call:

  quasi {  my $COMPILING::new_variable;   my $private_var; ... }
  quasi :COMPILING { my $new_variable; { my $private_var; ... } }

(Note that C<:COMPILING> has additional effects described in L</Macros>.)

=head1 Other matters

=head2 Anonymous hashes vs blocks

C<{...}> is always a block.  However, if it is completely empty or
consists of a single list, the first element of which is either a hash
or a pair, it is executed immediately to compose a C<Hash> object.

The standard C<pair> list operator is equivalent to:

    sub pair (*@LIST) {
        my @pairs;
        for @LIST -> $key, $val {
            push @pairs, $key => $val;
        return @pairs;

or more succinctly (and lazily):

    sub pair (*@LIST) {
        gather for @LIST -> $key, $val {
            take $key => $val;

The standard C<hash> list operator is equivalent to:

    sub hash (*@LIST) {
        return { pair @LIST };

So you may use C<sub> or C<hash> or C<pair> to disambiguate:

    $obj =  sub { 1, 2, 3, 4, 5, 6 };   # Anonymous sub returning list
    $obj =      { 1, 2, 3, 4, 5, 6 };   # Anonymous sub returning list
    $obj =      { 1=>2, 3=>4, 5=>6 };   # Anonymous hash
    $obj =      { 1=>2, 3, 4, 5, 6 };   # Anonymous hash
    $obj =  hash( 1, 2, 3, 4, 5, 6 );   # Anonymous hash
    $obj =  hash  1, 2, 3, 4, 5, 6  ;   # Anonymous hash
    $obj = { pair 1, 2, 3, 4, 5, 6 };   # Anonymous hash

=head2 Pairs as lvalues

Since they are immutable, Pair objects may not be directly assigned:

    (key => $var) = "value";    # ERROR

However, when binding pairs, names can be used to "match up" lvalues
and rvalues, provided you write the left side as a signature using
C<:(...)> notation:

    :(:who($name), :why($reason)) := (why => $because, who => "me");

(Otherwise the parser doesn't know it should parse the insides as a
signature and not as an ordinary expression until it gets to the C<:=>,
and that would be bad.  Alternately, the C<my> declarator can also
force treatment of its argument as a signature.)

=head2 Out-of-scope names

C<< GLOBAL::<$varname> >> specifies the C<$varname> declared in the C<*>
namespace.  Or maybe it's the other way around...

C<< CALLER::<$varname> >> specifies the C<$varname> visible in
the dynamic scope from which the current block/closure/subroutine
was called, provided that variable carries the "C<dynamic>"
trait.  (All variables with a C<*> twigil are automatically marked with the trait.
Likewise certain implicit lexicals (C<$_>, C<$/>, and C<$!>) are so marked.)

C<< DYNAMIC::<$varname> >> specifies the C<$varname> visible in the
innermost dynamic scope that declares the variable with the "C<is dynamic>"
trait or with a name that has the C<*> twigil.

C<< MY::<$varname> >> specifies the lexical C<$varname> declared in the current
lexical scope.

C<< OUR::<$varname> >> specifies the C<$varname> declared in the current
package's namespace.

C<< COMPILING::<$varname> >> specifies the C<$varname> declared (or about
to be declared) in the lexical scope currently being compiled.

C<< OUTER::<$varname> >> specifies the C<$varname> declared in the lexical
scope surrounding the current lexical scope (i.e. the scope in which
the current block was defined).

=head2 Declaring a C<MAIN> subroutine

Ordinarily a top-level Perl "script" just evaluates its anonymous
mainline code and exits.  During the mainline code, the program's
arguments are available in raw form from the C<@*ARGS> array.  At the end of
the mainline code, however, a C<MAIN> subroutine will be called with
whatever command-line arguments remain in C<@*ARGS>.  This call is
performed if and only if:


=item a)

the compilation unit was directly invoked rather than
by being required by another compilation unit, and

=item b)

the compilation unit declares a C<Routine> named "C<MAIN>", and

=item c)

the mainline code is not terminated prematurely, such as with an explicit call
to C<exit>, or an uncaught exception.


The command line arguments (or what's left of them after mainline
processing) is magically converted into a C<Capture> and passed to
C<MAIN> as its arguments, so switches may be bound as named args and
other arguments to the program may be bound to positional parameters
or the slurpy array:

    sub MAIN ($directory, :$verbose, *%other, *@filenames) {
        for @filenames { ... }

If C<MAIN> is declared as a C<proto> with a set of C<multi> subs,
multi dispatch is performed by the C<proto>.

As with module and class declarations, a sub declaration
ending in semicolon is allowed at the outermost file scope if it is the
first such declaration, in which case the rest of the file is the body:

    sub MAIN ($directory, :$verbose, *%other, *@filenames);
    for @filenames { ... }

This form is allowed only for simple subs named C<MAIN> that are intended
to be run from the command line.
A C<proto> or C<multi> definition may not be written in semicolon form,
nor may C<MAIN> subs within a module or class be written in semicolon form.  (A C<MAIN> routine
is allowed in a module or class, but is not usually invoked unless
the file is run directly (see a above).  This corresponds to the
"unless caller" idiom of Perl 5.)  In general, you may have only one
semicolon-style declaration that controls the whole file.

If an attempted dispatch to C<MAIN> fails, the C<USAGE> routine is called.
If there is no C<USAGE> routine, a default message is printed to standard
error. If the C<MAIN> routine does not declare a named parameter C<:$help>,
and C<--help> is passed as a command line argument to the program, the usage
message is printed to standard output instead.

This usage message is automatically generated from the signature (or
signatures) of C<MAIN>.  This message is generated at compile time,
and hence is available at any later time as C<$?USAGE>.

Common Unix command-line conventions are mapped onto the capture
as follows:

    Assuming C<-n> is the short name for C<--name>,
    On command line...         $*ARGS capture gets...

    # Short names
    -n                         :name
    -n=value                   :name<value>
    -nvalue                    :name<value>     # only if not declared Bool
    -n="spacey value"          :name«'spacey value'»
    -n='spacey value'          :name«'spacey value'»
    -n=val1,'val 2',etc        :name«val1 'val 2' etc»

    # Long names
    --name                     :name            # only if declared Bool
    --name=value               :name<value>     # don't care
    --name value               :name<value>     # only if not declared Bool

    --name="spacey value"      :name«'spacey value'»
    --name "spacey value"      :name«'spacey value'»
    --name='spacey value'      :name«'spacey value'»
    --name 'spacey value'      :name«'spacey value'»
    --name=val1,'val 2',etc    :name«val1 'val 2' etc»
    --name val1 'val 2' etc    :name«val1 'val 2' etc» # only if declared @
    --                                          # end named argument processing

    # Negation
    --/name                    :!name
    --/name=value              :name<value> but False
    --/name="spacey value"     :name«'spacey value'» but False
    --/name='spacey value'     :name«'spacey value'» but False
    --/name=val1,'val 2',etc   :name«val1 'val 2' etc» but False

    # Native
    :name                      :name
    :/name                     :!name
    :name=value                :name<value>
    :name="spacey value"       :name«'spacey value'»
    :name='spacey value'       :name«'spacey value'»
    :name=val1,'val 2',etc     :name«val1 'val 2' etc»

Exact Perl 6 forms are okay if quoted from shell processing:

    ':name<value>'             :name<value>
    ':name(42)'                :name(42)

For security reasons, only constants are allowed as arguments, however.

The default C<Capture> mapper pays attention to declaration of C<MAIN>'s
parameters to resolve certain ambiguities.  A C<--foo> switch needs to
know whether to treat the next word from the command line as an argument.
(Allowing the spacey form gives the shell room to do various things to
the argument.)  The short C<-foo> form never assumes a separate argument,
and you must use C<=>.  For the C<--foo> form, if there is a named parameter
corresponding to the switch name, and it is of type C<Bool>, then no argument
is expected.  Otherwise an argument is expected.  If the parameter is of
a non-slurpy array type, all subsequent words up to the next command-line
switch (or the end of the list) are bound to that parameter.

As usual, switches are assumed to be first, and everything after
the first non-switch, or any switches after a C<-->, are treated
as positionals or go into the slurpy array (even if they look like
switches).  Other policies may easily be introduced by calling C<MAIN>
explicitly.  For instance, you can parse your arguments with a grammar
and pass the resulting C<Match> object as a C<Capture> to C<MAIN>:

    @*ARGS ~~ /<MyGrammar::top>/;

    sub MAIN ($frompart, $topart, *@rest) {
        if $frompart<foo> { ... }
        if $topart<bar><baz> { ... }

This will conveniently bind top-level named matches to named
parameters, but still give you access to nested matches through those
parameters, just as any C<Match> object would.  Of course, in this example,
there's no particular reason the sub has to be named C<MAIN>.

To give both a long and a short switch name, you may use the pair
notation to install several names for the same parameter.
If any of the names is a single character, it will be considered a
short switch name, while all other parameters names are considered
as long switch name.  So if the previous declaration had been:

    sub MAIN (:f(:$frompart), :t(:$topart), *@rest)

then you could invoke the program with either C<-f> or C<--frompart>
to specify the first parameter.  Likewise you could use either C<-t>
or C<--topart> for the second parameter.

If a switch of the form C<-abc> cannot be matched against any
particular parameter, an attempt will be made to match it as if it
had been written C<-a -b -c>.

=head2 Relationship of MAIN routine with lexical setting

The preceding section describes the use of C<MAIN> in the user's code.
There may also be an implicit C<MAIN> routine supplied by the setting
of the current compilation unit.  (The C<-n> and C<-p> command-line
switches are implemented this way.)  In this case the user's mainline
code is not automatically executed; instead, execution is controlled
by the setting's C<MAIN> routine.  That routine calls C<{YOU_ARE_HERE}>
at the point where the user's code is to be lexically inserted (in
the abstract).  A setting may also call C<{YOU_ARE_HERE}> outside of
a C<MAIN> routine, in which case it functions as a normal setting,
and the C<{YOU_ARE_HERE}> merely indicates where the user's code
goes logically.  (Or from the compiler's point of view, which the
lexical scope to dump a snapshot of for later use by the compiler
as the setting for a different compilation unit.)  In this case the
execution of the user code proceeds as normal.  In fact, the C<CORE>
setting ends with a C<{YOU_ARE_HERE}> to dump the C<CORE> lexical
scope as the standard setting.  In this sense, C<CORE> functions as
an ordinary prelude.

If a C<MAIN> routine is declared both in the setting and in the
user's code, the setting's C<MAIN> functions as the actual mainline
entry point.  The user's C<MAIN> functions in an embedded fashion;
the setting's invocation of C<{YOU_ARE_HERE}> functions as the main
invocation from the point of view of the user's code, and the
user's C<MAIN> routine will be invoked at the end of each call to

=head2 Implementation note on autothreading of only subs

The natural way to implement autothreading for C<multi> subs is to
simply have the junctional signatures (the ones that can accept
C<Mu> or junction as well as C<Any> parameters) match more loosely than
the non-autothreading versions, and let multiple dispatch find the
appropriate sub based on the signature.  Those generic routines
then end up redispatching to the more specific ones.

On the other hand, the natural implementation of C<only> subs is to
call the sub in question directly for efficiency (and maybe even
inline it in some cases).  That efficiency is, after all, the main
reason for not just making all subs C<multi>.  However, this direct
call conflicts with the desire to allow autothreading.  It might
be tempting to simply make everything multi dispatch underneath,
and then say that the C<only> declaration merely means that you get
an error if you redeclare.  And maybe that is a valid approach if
the multiple dispatch mechanism is fast enough.

However, a direct call still needs to bind its arguments to its
parameters correctly, and it has to handle the case of failure to
bind somehow.  So it is also possible to implement autothreading
of C<only> subs based on failover from the binding failure.  This could
either be a one-shot failover followed by a conversion to a C<multi> call,
or it could failover every time you try to autothread. If we assume
that junctional processing is likely to be fairly heavyweight most of
the time compared to the cost of failing to bind, that tends to argue
for failing over every time.  This is also more conducive to inlining,
since it's difficult to rewrite inlined calls.  In any case, nowadays
a C<proto> declaration is considered to be a kind of C<only> sub, 
and needs to handle autothreading similaraly if the signature of
the C<proto> excludes junctions.

=head2 Introspection

This section describes the methods implemented by the routine objects
that allow introspection of the inner works of that routine.

=head3 Routine


=item .candidates

This method returns a (potentially lazy) list of the candidates associated with the current
routine. An "only" routine should return a list with itself as the
single item.

=item .signature

This method returns the signature of the current routine.

=item .cando(\$capture)

This method returns a (potentially lazy) list of the candidates that match the given
capture, ordered by goodness of match, with the best match first.

=item .push($candidate)

Adds C<$candidate> to the list of candidates for this C<proto>, calling this
method in an C<only> routine should result in a failure. It is also
accepted for C<multi>s declared in the source code to finalize the list
of candidates and also return a failure here. But C<Proto>s created by
calling C<Proto.new()> should be able add candidates at run-time.


=head3 Signature

See section L</Signature Introspection>.

=for vim:set expandtab sw=4: