Lexically-named vs. Globally-named Operators

At version 0.43, the way this module operates has changed. Since this version, the preferred method of operation for infix operators is that the module that registers them creates them with a fully-qualified name (i.e. including the name of the package that implements them). Other code which wishes to import the operator can then provide a local name for it visible within a scope. This acts as an alias for the fully-qualified registered one. The imported name needs not be the same in various scopes - this allows imported operators to be renamed in specific scopes to avoid name clashes.

Prior versions of this module only supported operators that have one global name that may or may not be enabled in lexical scopes. With this model, while the visiblity of an operator can be controlled per scope, if it is visible it must always have the same name and cannot be renamed to avoid a name clash.

At some future version of this module the older method may be removed, once all of the existing CPAN modules have been updated to the new model.

HAVE_PL_INFIX_PLUGIN

if( XS::Parse::Infix::HAVE_PL_INFIX_PLUGIN ) { ... }

This constant is true if built on a perl that supports the "PL_infix_plugin" extension mechanism, meaning that custom infix operators registered with this module will actually be recognised by the perl parser.

No actual production releases of perl yet support this feature, but see above for details of development versions which do.

apply_infix

$pkg->apply_infix( $on, $args, @infix );

Since version 0.43.

Intended to be called by the "import" and "unimport" methods of a module that implements infix operators. This sets up the various keys required in the lexical hints hash to create a lexical alias for the requested operator(s).

$on should be a true value for "import" or a false value for "unimport". $args should be an array reference to the arguments array, which will be modified to splice out recognised values, leaving behind anything else unrecognised. @infix should be a list of names of operators implemented by the importing package, named by $pkg.

After invoking this method, the caller should inspect for any remaining values in the arguments array, either to handle them in some other way, or just complain about them.

package Some::Infix::Module;
sub import
{
   my $pkg = shift;
   $pkg->XS::Parse::Infix::apply_infix( 1, \@_, qw( operator names here ) );
   die "Unrecognised symbols: @_" if @_;
}

When this module is "use"d, named operators are aliased from the fully-qualified names.

use Some::Infix::Module qw( operator );
my $z = $x operator $y;

At this point, the name "operator" becomes a lexical alias to "Some::Infix::Module::operator".

Each imported operator name can optionally be followed by import options, given in a hash reference. Currently the only named option recognised is "-as", which allows the importing scope to provide a different name for the imported operator, perhaps to avoid name clashes, or for neatness.

use Some::Infix::Module operator => { -as => "myname" };
my $z = $x myname $y;  # invokes Some::Infix::Module::operator

boot_xs_parse_infix

void boot_xs_parse_infix(double ver);

Call this function from your "BOOT" section in order to initialise the module and parsing hooks.

ver should either be 0 or a decimal number for the module version requirement; e.g.

boot_xs_parse_infix(0.14);

parse_infix

bool parse_infix(enum XSParseInfixSelection select, struct XSParseInfixInfo **infop);

Since version 0.27.

This function attempts to parse syntax for an infix operator from the current parser position. If it is successful, it fills in the variable pointed to by infop with a pointer to the actual information structure and returns "true". If no suitable operator is found, returns "false".

xs_parse_infix_new_op

OP *xs_parse_infix_new_op(const struct XSParseInfixInfo *info, U32 flags,
   OP *lhs, OP *rhs);

This function constructs a new optree fragment to represent invoking the infix operator with the given operands. It should be used much the same as core perl's "newBINOP" function.

The "info" structure pointer would be obtained from the "infix" field of the result of invoking the various "XPK_INFIX_*" token types from "XS::Parse::Keyword", or by calling "parse_infix" directly.

register_xs_parse_infix

void register_xs_parse_infix(const char *opname,
   const struct XSParseInfixHooks *hooks, void *hookdata);

This function installs a set of parsing hooks to be associated with the given operator name. This new operator will then be available via XS::Parse::Keyword by the various "XPK_INFIX_*" token types, "parse_infix", or to core perl's "PL_infix_plugin" if available.

These tokens will all yield an info structure, with the following fields:

struct XSParseInfixInfo {
   const char *opname;
   OPCODE opcode;  /* for built-in operators, or OP_CUSTOM for 
                      custom-registered ones */
   struct XSParseInfixHooks *hooks;
   void                     *hookdata;
   enum XSParseInfixClassification cls;  /* since version 0.28 */
};

If the operator name contains any non-ASCII characters they are presumed to be in UTF-8 encoding. This will matter for deparse purposes.

If opname contains a double-colon ("::") sequence, it is presumed to be a fully-qualified operator name in the new interface style. If not, it is presumed to be an older globally-named one.

Flags

The "flags" field gives details on how to handle the operator overall. It should be a bitmask of the following constants, or left as zero:

XPI_FLAG_LISTASSOC

Since version 0.40.

If set, the operator supports n-way list-associative syntax; written in the form

OPERAND op OPERAND op OPERAND op ...

In this case, the custom operator will be a LISTOP rather than a BINOP, and every operand of the entire chain will be stored as a child op of it. The op function will need to know how many operands it is working on. There are two ways this may be indicated, depending on whether it was known at compile-time.

If the number operands is known at compile-time, the "OPf_STACKED" flag is not set, and the "op_private" field indicates the number of operand expressions that were present.

If the number is not known (for example, the operator is being used as the body of a generated wrapper function), the "OPf_STACKED" flag is set. The number of arguments will be passed as the UV of an extra SV which is pushed last to the stack. The op function should pop this first to find out.

It is typical to begin a list-associative op function with code such as:

int n = (PL_op->op_flags & OPf_STACKED) ? POPu : PL_op->op_private;

If the operator is list-associative, then "lhs_flags" and "rhs_flags" must be equal.

The "lhs_flags" and "rhs_flags" fields give details on how to handle the left- and right-hand side operands, respectively.

It should be set to one of the following constants, or left as zero:

XPI_OPERAND_TERM_LIST

The operand will be foced into list context, preserving the "OP_PUSHMARK" at the beginning. This means that the ppfunc for this infix operator will have to "POPMARK" to find that.

XPI_OPERAND_LIST

The same as above.

Older versions used to provide constants named "XPI_OPERAND_ARITH" and "XPI_OPERAND_TERM" but they related to an older version of the core perl branch. These names are now aliases for zero, and can be removed from new code.

In addition the following extra bitflags are defined:

XPI_OPERAND_ONLY_LOOK

If set, the operator function promises that it will not mutate any of its passed values, nor allow leaking of direct alias pointers to them via return value or other locations.

This flag is optional; omitting it when applicable will not change any observed behaviour. Setting it may enable certain optimisations to be performed.

Currently, this flag simply enables an optimisation in the call-checker for infix operator wrapper functions that take list-shaped operands. This optimisation discards an "OP_ANONLIST" operation which would create a temporary anonymous array reference for its operand values, allowing a slight saving of memory use and CPU time. This optimisation is only safe to perform if the operator does not mutate or retain aliases of any of the arguments, as otherwise the caller might see unexpected modifications or value references to the values passed.

The Selection Stage

The "cls" field gives a "classification" of the operator, suggesting what sort of operation it provides. This is used as a filter by the various "XS::Parse::Keyword" selection macros.

The classification should be one of the "XPI_CLS_*" constants found and described further in the main XSParseInfix.h file.

The "permit" Stage

As a shortcut for the common case, the "permit_hintkey" may point to a string to look up from the hints hash. If the given key name is not found in the hints hash then the keyword is not permitted. If the key is present then the "permit" function is invoked as normal.

If not rejected by a hint key that was not found in the hints hash, the function part of the stage is called next and should inspect whether the keyword is permitted at this time perhaps by inspecting other lexical clues, and return true only if the keyword is permitted.

Both the string and the function are optional. Either or both may be present. If neither is present then the keyword is always permitted - which is likely not what you wanted to do.

The "parse" Stage

If the optional "parse" hook function is present, it is called immediately after the parser has recognised the presence of the named operator itself but before it attempts to consume the right-hand side term. This hook function can attempt further parsing, in order to implement more complex syntax such as hyper-operators.

When invoked, it is passed a pointer to an "SV *"-typed storage variable. It is free to use this variable it desires to store a result, which will then later be made available to the "new_op" function.

The Op Generation Stage

If the infix operator is going to be used, then one of the "new_op" or the "ppaddr" fields explain how to create a new optree fragment.

If "new_op" is defined then it will be used, and is expected to return an optree fragment that consumes the LHS and RHS arguments to implement the semantics of the operator. If the optional "parse" stage had been present earlier, the "SV **" pointer passed here will point to the same storage that "parse" had previously had access to, so it can retrieve the results.

If "new_op" is not present, then the "ppaddr" will be used instead to construct a new BINOP or LISTOP of the "OP_CUSTOM" type. If an earlier "parse" stage had stored additional results into the "SV *" variable these will be lost here.

The Wrapper Function

Additionally, if the "wrapper_func_name" field is set to a string, this gives the (fully-qualified) name for a function to be generated as part of registering the operator. This newly-generated function will act as a wrapper for the operator.

For operators whose RHS is a scalar, the wrapper function is assumed to take two simple scalar arguments. The result of invoking the function on those arguments will be determined by using the operator code.

$result = $lhs OP $rhs;
$result = WRAPPERFUNC( $lhs, $rhs );

For operators whose RHS is a list, the wrapper function takes at least one argument, possibly more. The first argument is the scalar on the LHS, and the remaining arguments, however many there are, form the RHS:

$result = $lhs OP @rhs;
$result = WRAPPERFUNC( $lhs, @rhs );

For operators whose LHS and RHS is a list, the wrapper function takes two arguments which must be array references containing the lists.

$result = @lhs OP @rhs;
$result = WRAPPERFUNC( \@lhs, \@rhs );

This creates a convenience for accessing the operator from perls that do not support "PL_infix_plugin".

In the case of scalar infix operators, the wrapper function also includes a call-checker which attempts to inline the operator directly into the callsite. Thus, in simple cases where the function is called directly on exactly two scalar arguments (such as in the following), no "ENTERSUB" overhead will be incurred and the generated optree will be identical to that which would have been generated by using infix operator syntax directly:

WRAPPERFUNC( $lhs, $rhs );
WRAPPERFUNC( $lhs, CONSTANT );
WRAPPERFUNC( $args[0], $args[1] );
WRAPPERFUNC( $lhs, scalar otherfunc() );

The checker is very pessimistic and will only rewrite callsites where it determines this can be done safely. It will not rewrite any of the following forms:

WRAPPERFUNC( $onearg );            # not enough args
WRAPPERFUNC( $x, $y, $z );         # too many args
WRAPPERFUNC( @args[0,1] );         # not a scalar
WRAPPERFUNC( $lhs, otherfunc() );  # not a scalar

The wrapper function for infix operators which take lists on both sides also has a call-checker which will attempt to inline the operator in similar circumstances. In addition to the optimisations described above for scalar operators, this checker will also inline an array-reference operator and omit the resulting dereference behaviour. Thus, the two following lines emit the same optree, without an "OP_SREFGEN" or "OP_RV2AV":

@lhs OP @rhs;
WRAPPERFUNC( \@lhs, \@rhs );

Note that technically, this optimisation isn't strictly transparent in the odd cornercase that one of the referenced arrays is also the backing store for a blessed object reference, and that object class has a "@{}" overload.

my @arr;
package SomeClass {
   use overload '@{}' => sub { return ["values", "go", "here"]; };
}
bless \@arr, "SomeClass";
# this will not actually invoke the overload operator
WRAPPERFUNC( \@arr, [4, 5, 6] );

As this cornercase relates to taking duplicate references to the same blessed object's backing store variable, it should not matter to any real code; regular objects that are passed by reference into the wrapper function will run their overload methods as normal.

The callchecker for list operands can optionally also discard an op of the "OP_ANONLIST" type, which is used by anonymous array-ref construction:

($u, $v, $w) OP ($x, $y, $z);
WRAPPERFUNC( [$u, $v, $w], [$x, $y, $z] );

This optimisation is only performed if the operator declared it safe to do so, via the "XPI_OPERAND_ONLY_LOOK" flag.

If a function of the given name already exists at registration time it will be left undisturbed and no new wrapper will be created. This permits the same infix operator to have multiple spellings of its name; for example to allow both a real Unicode and a fallback ASCII transliteration of the same operator. The first registration will create the wrapper function; the subsequent one will skip it because it would otherwise be identical.

Note that when generating an optree for a wrapper function call, the "new_op" hook function will be invoked with a "NULL" pointer for the "SV *"-typed parse data storage, as there won't be an opporunity for the "parse" hook to run in this case.