[ Table Of Contents | Keyword Index ]

critcl_use(n) 3.1.16 doc "C Runtime In Tcl (CriTcl)"

Name

critcl_use - Using Critcl

Table Of Contents

Description

Welcome to the C Runtime In Tcl, CriTcl for short, a system to build C extension packages for Tcl on the fly, from C code embedded within Tcl scripts, for all who wish to make their code go faster.

This document is a (hopefully) gentle introduction to Critcl by way of a series of small examples.

Readers which came directly to this document through a search or similar, and are thus in need of an overview of the whole system, are advised to read the Introduction To CriTcl first.

The examples here cover both how to embed C into Tcl with it, and how to build the distributable packages. As such the intended audience are both writers of packages with embedded C, and people building such packages. To make things easier the two themes each have their own section in this document, enabling all readers to quickly skip the part they are not interested in.

The sources of Critcl, should you have gotten them, contain several larger examples show-casing various aspects of the system. These demonstration packages can all be found in the sub-directory "examples/" of the sources.

Embedding C

This is the section for developers writing, or wishing to write, a package embedding C into Tcl via critcl.

I guess that we are allowed to asssume that you, gentle reader, are here because you have written some Tcl code which is not fast enough (any more) and you wish to make it "go faster" by replacing parts (or all) of it with speedy C.

Another, and I believe reasonable assumption to make would be that you have already investigated and ruled out or done things like changes to data structures and algorithms which reduce O(n*n) work to O (n*log n), O(n), or even O(1). Of course, nothing prevents you from forging ahead here even if you have not done such. Still, even in that case I would recommend that you consider investigating this line of making your code go faster as well.

Now, with these introductory words out of the way, lets jump into the meat of things.

A Simple Procedure

Starting simple, let us assume that the Tcl code in question is something like

    proc math {x y z} {
        return [expr {(sin($x)*rand())/$y**log($z)}]
    }

with the expression pretending to be something very complex and slow. Converting this to C we get:

    critcl::cproc math {double x double y double z} double {
        double up   = rand () * sin (x);
        double down = pow(y, log (z));
        return up/down;
    }

Notable about this translation:

  1. All the arguments got type information added to them, here "double". Like in C the type precedes the argument name. Other than that it is pretty much a Tcl dictionary, with keys and values swapped.

  2. We now also have to declare the type of the result, here "double", again.

  3. The reference manpage lists all the legal C types supported as arguments and results.

Custom Types, Introduction

When writing bindings to external libraries critcl::cproc is usually the most convenient way of writing the lower layers. This is however hampered by the fact that critcl on its own only supports a few standard (arguably the most import) standard types, whereas the functions we wish to bind most certainly will use much more, specific to the library's function.

The critcl commands argtype, resulttype and their adjuncts are provided to help here, by allowing a developer to extend critcl's type system with custom conversions.

This and the three following sections will demonstrate this, from trivial to complex.

The most trivial use is to create types which are aliases of existing types, standard or other. As an alias it simply copies and uses the conversion code from the referenced types.

Our example is pulled from an incomplete project of mine, a binding to Jeffrey Kegler's libmarpa library managing Earley parsers. Several custom types simply reflect the typedef's done by the library, to make the cprocs as self-documenting as the underlying library functions themselves.

    critcl::argtype Marpa_Symbol_ID     = int
    critcl::argtype Marpa_Rule_ID       = int
    critcl::argtype Marpa_Rule_Int      = int
    critcl::argtype Marpa_Rank          = int
    critcl::argtype Marpa_Earleme       = int
    critcl::argtype Marpa_Earley_Set_ID = int
    ...
    method sym-rank: proc {
        Marpa_Symbol_ID sym
        Marpa_Rank      rank
    } Marpa_Rank {
        return marpa_g_symbol_rank_set (instance->grammar, sym, rank);
    }
    ...

Custom Types, Semi-trivial

A more involved custom argument type would be to map from Tcl stringas to some internal representation, like an integer code.

The first example is taken from the tclyaml package, a binding to the libyaml library. In a few places we have to map readable names for block styles, scalar styles, etc. to the internal enumeration.

    critcl::argtype yaml_sequence_style_t {
        if (!encode_sequence_style (interp, @@, &@A)) return TCL_ERROR;
    }
    ...
    critcl::ccode {
        static const char* ty_block_style_names [] = {
            "any", "block", "flow", NULL
        };
        static int
        encode_sequence_style (Tcl_Interp* interp, Tcl_Obj* style,
                               yaml_sequence_style_t* estyle)
        {
            int value;
            if (Tcl_GetIndexFromObj (interp, style, ty_block_style_names,
                                     "sequence style", 0, &value) != TCL_OK) {
                return 0;
            }
            *estyle = value;
            return 1;
        }
    }
    ...
    method sequence_start proc {
        pstring anchor
        pstring tag
        int implicit
        yaml_sequence_style_t style
    } ok {
        /* Syntax: <instance> seq_start <anchor> <tag> <implicit> <style> */
        ...
    }
    ...

It should be noted that this code precedes the advent of the supporting generator package critcl:emap. With it available we can simlply define the mapping via

    critcl::emap::def yaml_sequence_style_t {
        any   0
        block 1
        flow  2
    }

and the generator will not only provide the conversions, but also define the argument and result types needed for their use by cproc. Another example of such a semi-trivial argument type can be found in the CRIMP package, which defines a Tcl_ObjType for image values. This not only provides a basic argument type for any image, but also types which check that the image has a specific format. Here we see for the first time non-integer arguments, and the need to define the C types used for variables holding the C level value, and the type of function parameters (Due to C promotion rules we may need different types).

    critcl::argtype image {
        if (crimp_get_image_from_obj (interp, @@, &@A) != TCL_OK) {
            return TCL_ERROR;
        }
    } crimp_image* crimp_image*
    ...
        set map [list <<type>> $type]
        critcl::argtype image_$type [string map $map {
            if (crimp_get_image_from_obj (interp, @@, &@A) != TCL_OK) {
                return TCL_ERROR;
            }
            if (@A->itype != crimp_imagetype_find ("crimp::image::<<type>>")) {
                Tcl_SetObjResult (interp,
                                  Tcl_NewStringObj ("expected image type <<type>>",
                                                    -1));
                return TCL_ERROR;
            }
        }] crimp_image* crimp_image*
    ...

Custom Types, Support structures

The adjunct command critcl::argtypesupport is for when the conversion needs additional definitions, for example a helper structure.

An example of this can be found among the standard types of critcl itself, the pstring type. This type provides the C function with not only the string pointer, but also the string length, and the Tcl_Obj* this data came from. As cproc's calling conventions allow us only one argument for the data of the parameter a structure is needed to convey these three pieces of information.

Thus the argument type is defined as

    critcl::argtype pstring {
        @A.s = Tcl_GetStringFromObj(@@, &(@A.len));
        @A.o = @@;
    } critcl_pstring critcl_pstring
    critcl::argtypesupport pstring {
        typedef struct critcl_pstring {
            Tcl_Obj* o;
            char*    s;
            int      len;
        } critcl_pstring;
    }

In the case of such a structure being large we may wish to allocate it on the heap instead of having it taking space on the stack. If we do that we need another adjunct command, critcl::argtyperelease. This command specifies the code required to release dynamically allocated resources when the worker function returned, before the shim returns to the caller in Tcl. To keep things simple our example is synthetic, a modification of pstring above, to demonstrate the technique. An actual, but more complex example is the code to support the variadic args argument of cproc.

    critcl::argtype pstring {
        @A = (critcl_pstring*) ckalloc(sizeof(critcl_pstring));
        @A->s = Tcl_GetStringFromObj(@@, &(@A->len));
        @A->o = @@;
    } critcl_pstring* critcl_pstring*
    critcl::argtypesupport pstring {
        typedef struct critcl_pstring {
            Tcl_Obj* o;
            char*    s;
            int      len;
        } critcl_pstring;
    }
    critcl::argtyperelease pstring {
        ckfree ((char*)) @A);
    }

Note, the above example shows only the most simple case of an allocated argument, with a conversion that cannot fail (namely, string retrieval). If the conversion can fail then either allocation has to be defered to happen only on successful conversion, or the conversion code has to release the allocated memory itself in the failure path, because it will never reach the code defined via critcl::argtyperelease in that case.

Custom Types, Results

All of the previous sections dealt with argument conversions, i.e. going from Tcl into C. Custom result types are for the reverse direction, from C to Tcl. This is usually easier, as most of the time errors should not be possible. Supporting structures, or allocating them on the heap are not really required and therefore not supported.

The example of a result type shown below was pulled from KineTcl. It is a variant of the builtin result type Tcl_Obj*, aka object. The builtin conversion assumes that the object returned by the function has a refcount of 1 (or higher), with the function having held the reference, and releases that reference after placing the value into the interp result. The conversion below on the other hand assumes that the value has a refcount of 0 and thus that decrementing it is forbidden, lest it be released much to early, and crashing the system.

    critcl::resulttype KTcl_Obj* {
        if (rv == NULL) { return TCL_ERROR; }
        Tcl_SetObjResult(interp, rv);
        /* No refcount adjustment */
        return TCL_OK;
    } Tcl_Obj*

This type of definition is also found in Marpa and recent hacking hacking on CRIMP introduced it there as well. Which is why this definition became a builtin type starting with version 3.1.16, under the names Tcl_Obj*0 and object0.

Going back to errors and their handling, of course, if a function we are wrapping signals them in-band, then the conversion of such results has to deal with that. This happens for example in KineTcl, where we find

    critcl::resulttype XnStatus {
        if (rv != XN_STATUS_OK) {
            Tcl_AppendResult (interp, xnGetStatusString (rv), NULL);
            return TCL_ERROR;
        }
        return TCL_OK;
    }
    critcl::resulttype XnDepthPixel {
        if (rv == ((XnDepthPixel) -1)) {
            Tcl_AppendResult (interp,
                              "Inheritance error: Not a depth generator",
                              NULL);
            return TCL_ERROR;
        }
        Tcl_SetObjResult (interp, Tcl_NewIntObj (rv));
        return TCL_OK;
    }

Handling A Variable Number Of Arguments

In A Simple Procedure we demonstrated how easy a translation to C can be. Is it still as easy when we introduce something moderately complex like handling a variable number of arguments ? A feature which is needed to handle commands with options and optional arguments ?

Well, starting with version 3.1.16 critcl::cproc does have full support for optional arguments, args-style variadics, and default values, extending its range to everything covered by the builtin proc. The only thing not truly supported are options (i.e. flag arguments) of any kind.

For the moment, and the example, let us pretend that we can use critcl::cproc only if the number of arguments is fully known beforehand, i.e. at the time of declaration. Then we have to use critcl::ccommand to handle the translation of the procedure shown below:

    proc math {args} {
        set sum 0
        foreach y $args { set sum [expr {$sum + $y}] }
        return $sum
    }

Its advantage: Access to the low-level C arguments representing the Tcl arguments of the command. Full control over argument conversion, argument validation, etc.

Its disadvantage: Access to the low-level C arguments representing the Tcl arguments of the command. Assuming the burden of having to write argument conversion, argument validation, etc. Where critcl::cproc handles the task of converting from Tcl to C values (for arguments) and back (for the result), with critcl::command it is the developer who has to write all this code.

Under our restriction the translation of the example is:

    critcl::ccommand math {cd ip oc ov} {
        double sum = 0;
        double y;
        oc --;
        while (oc) {
            if (Tcl_GetDoubleFromObj (ip, ov[oc], &y) != TCL_OK) {
                return TCL_ERROR;
            }
            sum += y;
            oc --;
        }
        Tcl_SetObjResult (ip, Tcl_NewDoubleObj (sum));
        return TCL_OK:
    }

Notable about this translation:

  1. As promised/threatened, all the conversions between the Tcl and C domains are exposed, and the developer should know her way around Tcl's C API.

  2. The four arguments "cd ip oc ov" are our names for the low-level arguments holding

    1. ClientData (reference)

    2. Tcl_Interp (reference)

    3. Number of arguments, and

    4. Array of argument values, each a Tcl_Obj*.

    This list of arguments, while not optional in itself, is allowed to be empty, and/or to contain empty strings as argument names. If we do that critcl will supply standard names for the missing pieces, namely:

    1. clientdata

    2. interp

    3. objc

    4. objv

Now, letting go of our pretenses regarding the limitations of critcl::cproc, due to the support it does have for args-style variadics (since version 3.1.16) we can write a much simpler translation:

    critcl::cproc math {double args} double {
        double sum = 0;
        args.c --;
        while (args.c) {
            sum += args.v[args.c];
            args.c --;
        }
        return sum;
    }

Data As A Tcl Command

Here we assume that we have a Tcl procedure which returns a fixed string. In the final product we are going to C to hide this string from the casual user.

    proc somedata {} {
	return {... A large blob of characters ...}
    }

The translation of this is simple and easy:

    package require critcl
    critcl::cdata somedata {... A large blob of characters ...}

There is nothing really notable here.

Blocks of arbitrary C

Often just defining Tcl commands in C, as demonstrated in the sections A Simple Procedure, Handling A Variable Number Of Arguments, and Data As A Tcl Command is not really enough. For example we may have several of our new C commands using the same code over and over, and we wish avoid this duplication. Or we wish to pull in declarations and definitions from some external library.

In both cases we require the ability to embed an unstructured block of C code which can contain whatever we want, defines, functions, includes, etc. without being directly tied to Tcl commands. The command critcl::code provides us with exactly that. As our example now an excerpt taken from real code, the top of the "sha1c.tcl" critcl file in the sha1 module of Tcllib:

    critcl::ccode {
        #include "sha1.h"
        #include <stdlib.h>
        #include <assert.h>
        
        static
        Tcl_ObjType sha1_type; /* fast internal access representation */
        
        static void 
        sha1_free_rep(Tcl_Obj* obj)
        {
            SHA1_CTX* mp = (SHA1_CTX*) obj->internalRep.otherValuePtr;
            Tcl_Free(mp);
        }
        ...
    }

We see here the beginning of the C code defining a custom Tcl_ObjType holding the data of a SHA1 context used during the incremental calculation of sha1 hashes.

Constant values

While one might believe that there is no need for commands which returns constant values that is not true. Commands reporting on compile-time configuration, like version numbers, available features, etc. are at least one use case for such commands.

The reason for creating critcl commands to support them ? Convenience to the user, yes, but further than that, the ability to optimize the internals, i.e. the generated code.

A cproc would be easy to write, but incurs overhead due to a superfluous work function. A ccommand has no overhead, except that of the user having to write the argument checking and result conversion.

Using critcl::cconst is both convenient and without code overhead. Our example is a function found in package tcl-linenoise, that is, if cconst had existed at the time of writing. It returns a configuration value reporting to the policy layer if an extended mode for hidden input is available from the bound linenoise, or not.

    critcl::cconst linenoise::hidden_extended boolean 1

Lifting constants

When writing a critcl-based package to make a third-party library available to scripts we do not only have to make the relevant functions available as commands, often we also have to know all the various constants, flags, etc. these functions take.

Rather than writing such magic numbers directly we would greatly prefer to use symbolic names instead. Instead of providing one or more commands to list and map the magic numbers to strings critcl only provides a single command which allows the export of C defines and enumeration values, mapping them to Tcl variables of the given names, whose values are the associated magic numbers.

This is good enough because the developers of the third-party library were very likely like us and wanted to use symbolic names instead of magic numbers. Which in C are declared as via defines and enumeration types. We just have to lift them up.

Our example comes from cryptkit, a Tcl binding to cryptlib, a cryptography library. The command

    critcl::cdefines CRYPT_* ::crypt

maps all Cryptlib specific #defines and enums into the namespace ::crypt, telling critcl to create aliases to the symbols.

Similarly

    critcl::cdefines {
        NULL
        TRUE
        FALSE
        TCL_OK
        TCL_ERROR
    } ::crypt

maps the listed defines into the namespace ::crypt.

An important thing to note: These commands do not create the defines in the C level. They only lift pre-existing material. Which can come from the headers of the third-party library, the usual case, but also from Blocks of arbitrary C.

A corrollary to the above: What is not where, cannot be lifted. All listed names and patterns which have no actual C code declaring them are ignored, i.e. not mapped.

Finding header files

A notable thing in the example shown in the section about Blocks of arbitrary C is the

    #include "sha1.h"

statement. Where does this header come from ? Looking at the Tcllib module we will find that the header is actually a sibling to the "sha1c.tcl" file containing the embedded C code. However, critcl does not know that. It has to be told. While without that knowledge it will invoke the compiler just fine, the compilation will fail because the header is not on the include paths used by the compiler, and therefore will not be found.

For this we have the critcl::cheaders command. It enables us to either tell the compiler the path(s) where the required headers can be found, using

    critcl::cheaders -I/path/to/headers/

or to tell it directly which headers we are using and where they live:

    critcl::cheaders sha1.h

And now critcl knows that "sha1.h" is important, and that it lives besides the ".critcl" file which referenced it (because of the relative path used). Note that this doesn't absolve us of the need to "#include" the header through a critcl::ccode block. This only tells critcl where it lives so that it can configure the compiler with the proper include paths to actually find it on use.

Further note that a C development environment is usually configured to find all the system headers, obviating the need for a critcl::cheaders declaration when such are used. For these a plain "#include" in a critcl::ccode block is good enough. In other words, the second form of invoking critcl::cheaders is pretty much only for headers which accompany the ".critcl" file.

Separate C sources

In all of the examples shown so far the C code was fully embedded in a ".critcl" file. However, if the C part is large it can make sense to break it out of the ".critcl" file into one or more separate proper ".c" file(s).

The critcl::csources command can then be used to make this code known to the original ".critcl" file again. This command accepts the paths to the ".c" files as arguments, and glob patterns as well. Our example comes from the struct::graph package in Tcllib. Its core C functions are in separate files, and the ".critcl" code then makes them known via:

namespace eval ::struct {
    # Supporting code for the main command.
    critcl::cheaders graph/*.h
    critcl::csources graph/*.c
    ...
}

which tells critcl that these files are in the subdirectory "graph" relative to the location of "graph_c.tcl", which is the relevant ".critcl" file.

This example also demonstrates again the use of critcl::cheaders, which we also saw in section Finding header files.

Finding external libraries

When creating a package exposing some third-party library to Tcl Finding header files is only the first part, to enable failure-free compilation. We also have to find the library/ies themselves so that they can be linked to our package. This is described here. The last issue, Lifting constants from C to Tcl for the use by scripts is handled in a separate section and example.

The relevant command is critcl::clibraries. Its basic semantics are like that of critcl::cheaders, i.e. It enables us to tell the linker the path(s) where the required libraries can be found, using

    critcl::clibraries -L/path/to/libraries/

name them

    critcl::clibraries -lfoo

or tell it directly which libraries we are using and where they live:

    critcl::clibraries /path/to/library/foo.so

This last way of using should be avoided however, as it intermingles searching and naming, plus the name is platform dependent.

For OS X we additionally have the critcl::framework command which enables us to name the frameworks used by our package. Note that this command can be used unconditionally. If the build target is not OS X it is ignored.

Customizing the compile and link steps

The commands critcl::cflags and critcl::ldflags enable you to provide custom options to the compile and link phases for a ".critcl" file.

This usually becomes necessary if the C code in question comes from an external library we are writing a Tcl binding for, with multiple configurations to select, non-standard header locations, and other things. Among the latter, especially platform-specific settings, for example byteorder.

This makes critcl::check an important adjunct command, as this is the API for Checking The Environment, and then selecting the compile & link flags to use.

I currently have no specific example to demonstrate these commands.

Having both C and Tcl functionality

Often enough only pieces of a package require recoding in C to boost the whole system. Or, alternatively, the package in question consists of a low-level layer C with a Tcl layer above encoding policies and routing to the proper low-level calls, creating a nicer (high-level) API to the low-level functionality, etc.

For all of this we have to be able to write a package which contains both C and Tcl, nevermind the fact the C parts are embedded in Tcl.

The easiest way to structure such a package is to have several files, each with a different duty. First, a ".critcl" file containing the embedded C, and second one or more ".tcl" files providing the Tcl parts. Then use the critcl::tsources command in the ".critcl" file to link the two parts together, declaring the ".tcl" files as necessary companions of the C part.

    package require critcl
    critcl::tsources your-companion.tcl ; # Companion file to use
    ... embedded C via critcl commands ...

With a declaration as shown above the companion file will be automatically sourced when the C parts are made available, thus making the Tcl parts available as well.

Using C with Tcl functionality as fallback

There is one special case of Having both C and Tcl functionality which deserves its own section.

The possibility of not having the fast C code on some platform, and using a slower Tcl implementation of the functionality. In other words, a fallback which keeps the package working in the face of failure to build the C parts. A more concrete example of this would be a module implementing the SHA hash, in both C and Tcl, and using the latter if and only if the C implementation is not available.

There two major possibilities in handling such a situation.

  1. Keep all the pieces separated. In that scenario our concrete example would be spread over three packages. Two low-level packages sha::c and sha::tcl containing the two implementations of the algorithm, and, thirdly, a coordinator package sha which loads either of them, based on availability.

    The Tcllib bundle of packages contains a number of packages structured in this manner, mostly in the struct module.

    Writing the C and Tcl parts should be simple by now, with all the examples we had so far. The only non-trivial part is the coordinator, and even that if and only if we wish to make it easy to write a testsuite which can check both branches, C, and Tcl without gymnastics. So, the most basic coordinator would be

        set sha::version 1
        if {[catch {
            package require sha::c $sha::version
        }]} {
            package require sha::tcl $sha::version
        }
        package provide sha $sha::version
    

    It tries to load the C implementation first, and falls back to the Tcl implementation if that fails. The code as is assumes that both implementations create exactly the same command names, leaving the caller unaware of the choice of implementations.

    A concrete example of this scheme can be found in Tcllib's md5 package. While it actually uses ythe Trf as its accelerator, and not a critcl-based package the principle is the same. It also demonstrates the need for additional glue code when the C implementation doesn't exactly match the signature and semantics of the Tcl implementation.

    This basic coordinator can be easily extended to try more than two packages to get the needed implementation. for example, the C implementation may not just exist in a sha::c package, but also bundled somewhere else. Tcllib, for example, has a tcllibc package which bundles all the C parts of its packages which have them in a single binary.

    Another direction to take it in is to write code which allows the loading of multiple implementations at the same time, and then switching between them at runtime. Doing this requires effort to keep the implementations out of each others way, i.e. they cannot provide the same command names anymore, and a more complex coordinator as well, which is able to map from the public command names to whatever is provided by the implementation.

    The main benefit of this extension is that it makes testing the two different implementations easier, simply run through the same set of tests multiple times, each time with different implementation active. The disadvantage is the additional complexity of the coordinator's internals. As a larger example of this technique here is the coordinator "modules/struct/queue.tcl" handling the C and Tcl implementations of Tcllib's struct::queue package:

        # queue.tcl --
        #       Implementation of a queue data structure for Tcl.
        package require Tcl 8.4
        namespace eval ::struct::queue {}
        ## Management of queue implementations.
        # ::struct::queue::LoadAccelerator --
        #       Loads a named implementation, if possible.
        proc ::struct::queue::LoadAccelerator {key} {
            variable accel
            set r 0
            switch -exact -- $key {
                critcl {
                    # Critcl implementation of queue requires Tcl 8.4.
                    if {![package vsatisfies [package provide Tcl] 8.4]} {return 0}
                    if {[catch {package require tcllibc}]} {return 0}
                    set r [llength [info commands ::struct::queue_critcl]]
                }
                tcl {
                    variable selfdir
                    if {
                        [package vsatisfies [package provide Tcl] 8.5] &&
                        ![catch {package require TclOO}]
                    } {
                        source [file join $selfdir queue_oo.tcl]
                    } else {
                        source [file join $selfdir queue_tcl.tcl]
                    }
                    set r 1
                }
                default {
                    return -code error "invalid accelerator/impl. package $key: must be one of [join [KnownImplementations] {, }]"
                }
            }
            set accel($key) $r
            return $r
        }
        # ::struct::queue::SwitchTo --
        #       Activates a loaded named implementation.
        proc ::struct::queue::SwitchTo {key} {
            variable accel
            variable loaded
            if {[string equal $key $loaded]} {
                # No change, nothing to do.
                return
            } elseif {![string equal $key ""]} {
                # Validate the target implementation of the switch.
                if {![info exists accel($key)]} {
                    return -code error "Unable to activate unknown implementation \"$key\""
                } elseif {![info exists accel($key)] || !$accel($key)} {
                    return -code error "Unable to activate missing implementation \"$key\""
                }
            }
            # Deactivate the previous implementation, if there was any.
            if {![string equal $loaded ""]} {
                rename ::struct::queue ::struct::queue_$loaded
            }
            # Activate the new implementation, if there is any.
            if {![string equal $key ""]} {
                rename ::struct::queue_$key ::struct::queue
            }
            # Remember the active implementation, for deactivation by future
            # switches.
            set loaded $key
            return
        }
        # ::struct::queue::Implementations --
        #       Determines which implementations are
        #       present, i.e. loaded.
        proc ::struct::queue::Implementations {} {
            variable accel
            set res {}
            foreach n [array names accel] {
                if {!$accel($n)} continue
                lappend res $n
            }
            return $res
        }
        # ::struct::queue::KnownImplementations --
        #       Determines which implementations are known
        #       as possible implementations.
        proc ::struct::queue::KnownImplementations {} {
            return {critcl tcl}
        }
        proc ::struct::queue::Names {} {
            return {
                critcl {tcllibc based}
                tcl    {pure Tcl}
            }
        }
        ## Initialization: Data structures.
        namespace eval ::struct::queue {
            variable  selfdir [file dirname [info script]]
            variable  accel
            array set accel   {tcl 0 critcl 0}
            variable  loaded  {}
        }
        ## Initialization: Choose an implementation,
        ## most preferred first. Loads only one of the
        ## possible implementations. And activates it.
        namespace eval ::struct::queue {
            variable e
            foreach e [KnownImplementations] {
                if {[LoadAccelerator $e]} {
                    SwitchTo $e
                    break
                }
            }
            unset e
        }
        ## Ready
        namespace eval ::struct {
            # Export the constructor command.
            namespace export queue
        }
        package provide struct::queue 1.4.2
    

    In this implementation the coordinator renames the commands of the low-level packages to the public commands, making the future dispatch as fast as if the commands had these names anyway, but also forcing a spike of bytecode recompilation if switching is ever done at the runtime of an application, and not just used for testing, and possibly disrupting introspection by the commands, especially if they move between different namespaces.

    A different implementation would be to provide the public commands as procedures which consult a variable to determine which of the loaded implementations is active, and then call on its commands. This doesn't disrupt introspection, nor does it trigger bytecode recompilation on switching. But it takes more time to dispatch to the actual implementation, in every call of the public API for the package in question.

    A concrete example of this scheme can be found in Tcllib's crc32 package.

  2. Mix the pieces together. Please note that while I am describing how to make this work I strongly prefer and recommend to use the previously shown approach using separate files/packages. It is much easier to understand and maintain. With this warning done, lets go into the nuts and bolts.

    If we care only about mode "compile & run" things are easy:

        package require critcl
        if {![critcl::compiling]} {
            proc mycommand {...} {
                ...
            }
        } else {
            critcl::cproc mycommand {...} {
                ...
            }
        }
    

    The command critcl::compiling tells us whether we have a compiler available or not, and in the latter case we implement our command in Tcl.

    Now what happens when we invoke mode "generate package" ? ... compiler failure ... ... ok - C code - everything fine ... fail - no package ? or just no C code ? declare self as tsource, to be used ? ... platform-specific C/Tcl -- uuid.

Unlazy Packages

By default critcl is a bit inconsistent between modes "compile & run" and "generate package". The result of the latter is a standard Tcl package which loads and sources all of its files immediately when it is required. Whereas "compile & run" defers actual compilation, linking, and loading until the first time one of the declared commands is actually used, making this very lazy.

This behaviour can be quite unwanted if Tcl companion files, or other users of the C commands use introspection to determine the features they have available. Just using [info commands] doesn't cut it, the auto_index array has to be checked as well, making things quite inconvenient for the users.

To fix this issue at the source, instead of in each user, be it inside of the package itself, or other packages, we have the command critcl::load. Used as the last command in a ".critcl" file it forces the compile, link, and load trinity, ensuring that all C commands are available immediately.

    package require critcl
    ... Declare C procedures, commands, etc.
    critcl::load ; # Force build and loading.

Note that is not allowed, nor possible to use critcl commands declaring anything after critcl::load has been called. I.e., code like

    package require critcl
    ... Declare C procedures, commands, etc.
    critcl::load ; # Force build and loading.
    ... More declarations of C code, ...
    critcl::code { ... }

will result in an error. The only package-related commands still allowed are

  1. critcl::done

  2. critcl::failed

  3. critcl::load

as these only query information, namely the build status, and are protected against multiple calls.

Checking your C

As said several times, by default critcl defers the compile and link steps for a file until it is needed, i.e. the first command of the ".critcl" file in question is actually invoked.

This not only has the effect of lazily loading the package's functionality, but also, when developing using mode "compile & run", of us not seeing any errors in our code until we are actually trying to run some demonstration.

If we do not wish to have such a delay we have to be able to force at least the execution of the compile step.

The command critcl::failed is exactly that. When called it forcibly builds the C code for the ".critcl" file it is part of, and returns a boolean vlaue signaling failure (true), or success (false).

    package require critcl
    ... Declare C procedures, commands, etc.
    if {[critcl::failed]} {
        ... signal error
    }

It is related and similar to critcl::load, the command to overcome the lazy loading, as shown in section Unlazy Packages.

Like it is not allowed, nor possible to use critcl commands declaring anything after critcl::failed has been called, making it pretty much the last critcl command in a ".critcl" file. Code like

    package require critcl
    ... Declare C procedures, commands, etc.
    if {[critcl::failed]} { ... }
    ... More declarations of C code, ...
    critcl::code { ... }

will result in an error. The only package-related commands still allowed are

  1. critcl::done

  2. critcl::failed

  3. critcl::load

as these only query information, namely the build status, and are protected against multiple calls.

Which Tcl ?

When building the shared library from the embedded C sources one of the things critcl does for us is to provide the Tcl headers, especially the stubs declarations.

By default these are the Tcl 8.4 headers and stubs, which covers 90% of the cases. What when the package in question is meant for use with Tcl 8.5 or higher, using C-level features of this version of Tcl.

Use the critcl::tcl command to declare to critcl the minimum version of Tcl required to operate the package. This can be either 8.4, 8.5, or 8.6, and critcl then supplies the proper headers and stubs.

    package require critcl
    critcl::tcl 8.5
    ... Declare your code ...

Making A Widget

... requires compiling against the Tk headers, and linking with Tk's stubs. For our convenience we have a simple, single command to activate all the necessary machinery, with critcl supplying the header files and stubs C code, instead of having to make it work on our own via critcl::cflags, critcl::ldflags, critcl::cheaders, critcl::csources.

This command is critcl::tk.

    package require critcl
    critcl::tk ; # And now critcl knows to put in the Tk headers and other support.
    ... Declare your code ...

Please note that this doesn't release you from the necessity of learning Tk's C API and how to use it to make a widget work. Sorry.

Checking The Environment

... may be necessary when creating a binding to some third-party library. The headers for this library may be found in non-standard locations, ditto for the library/ies itself. We may not have the headers and/or library on the build host. Types with platform-dependent sizes and definitions. Endianness issues. Any number of things.

TEA-based packages can use autoconf and various predefined macros to deal with all this. We have the Power Of Tcl (tm) and critcl::check.

This command takes a piece of C code as argument, like critcl::ccode. Instead of saving it for later it however tries to compile it immediately, using the current settings, and then returns a boolean value reporting on the success (true) or failure (false). From there we can then branch to different declarations.

As example let us check for the existence of some header "FOO.h":

    package require critcl
    if {[critcl::check {
        #include <FOO.h>
    }]} {
        ... Code for when FOO.h is present.
    } else {
        ... Code for when FOO.h is not present.
    }

Should we, on the other hand, wish to search for the header ourselves, in non-standard locations we have the full power of Tcl available, i.e. loops, the file and glob commands, etc., which can then be followed by a critcl::cheader command to declare the location we found (See also Finding header files).

A nice extension to critcl would be a package collecting pocedures for common tasks like that, sort of like an autoconf for Tcl. critcl::config seems to be nice name for such a package.

Obvious adjunct commands which can be driven by results from critcl::check are

critcl::cflags
critcl::cheaders
critcl::clibraries
critcl::framework
critcl::ldflags

Less obvious, yet still valid are also

critcl::ccode
critcl::ccommand
critcl::cdata
critcl::cproc
critcl::csources
critcl::ctsources

and pretty much everything else you can imagine.

License Invoked

When writing packages it is always good manners to provide prospective users with the license the package is under, so that they can decide whether they truly want to use the package, or not.

As critcl-based packages often consist of only a single file a nice way of doing that is to embed the license in that file. By using a critcl command, namely critcl::license this information is then also available to the critcl application, which can put it into a standard location, i.e. "license.terms", of the generated packages.

I currently have no specific example to demonstrate the command.

Building Critcl Packages

This is the section for developers having to generate packages from ".critcl" files, i.e binaries for deployment,

Getting Help ...

... Is easy. Running

    critcl -help

prints the basics of using the application to stdout.

Pre-Filling The Result Cache

The default mode of the critcl application is to take a series of ".critcl" files, build their binaries, and leave them behind in the result cache. When the files are later actually used the compile and link steps can be skipped, leading to shorter load times.

The command line for this is

    critcl foo.tcl

or, to process multiple files

    critcl foo.tcl bar.tcl ...

One thing to be aware of, should critcl find that the cache already contains the results for the input files, no building will be done. If you are sure that these results are outdated use the option -force to force(sic!) critcl to rebuild the binaries.

    critcl -force foo.tcl

For debugging purposes it may be handy to see the generated intermediate ".c" files as well. Their removal from the cache can be prevented by specifying the option -keep.

    critcl -keep foo.tcl

These can be combined, of course.

Building A Package

To build the binary package for a ".critcl" file, instead of Pre-Filling The Result Cache, simply specify the option -pkg.

    critcl -pkg foo.tcl

This will geneate a package named foo. A simpler alternative to the above is

    critcl -pkg foo

The application will automatically assume that the input file to look for is "foo.tcl".

But what when the name of the input file is not the name of the package to build ? This we can handle as well:

    critcl -pkg foo bar.tcl

The argument foo specifies the name, and "bar.tcl" is the file to process.

Going back to the very first example, it is of course possible to use an absolute path to specify the file to process:

    critcl -pkg /path/to/foo.tcl

The package name derived from that is still foo.

Building And Installing A Package

Here we assume that you know the basics of how to build a package. If not, please read section Building A Package first.

By default critcl will place all newly-made packages in the subdirectory "lib" of the current working directory. I.e. running

    critcl -pkg foo

will create the directory "lib/foo" which contains all the files of the package.

When this behaviour is unwanted the option -libdir is available, allowing the explicit specification of the destination location to use.

    critcl -pkg -libdir /path/to/packages foo

A common use might be to not only build the package in question, but to also immediately install it directly in the path where the user's tclsh will be able to find it. Assuming, for example, that the tclsh in question is installed at "/path/to/bin/tclsh", with the packages searched for under "/path/to/lib" ([info library]), the command

    critcl -pkg -libdir /path/to/lib foo

will build the package and place it in the directory "/path/to/lib/foo".

Building For Debugging

Here we assume that you know the basics of how to build a package. If not, please read section Building A Package first.

An important issue, when there is trouble with the package, debugging becomes necessary a evil. Critcl supports this through the -debug option. Using it enables various build modes which help with that.

For example, to activate the Tcl core's built-in memory debugging subsystem build your package with

    critcl -pkg -debug memory foo

The resulting binary for package foo will use Tcl's debug-enabled (de)allocation functions, making them visible to Tcl's memory command. This of course assumes that the Tcl core used was also built for memory debugging.

Further, built your package with

    critcl -pkg -debug symbols foo

to see the foo's symbols (types, functions, variables, etc.) when inspecting a "core" file it is involved in with a symbolic debugger,

To activate both memory debugging and symbols use either

    critcl -pkg -debug all foo

or

    critcl -pkg -debug symbols -debug memory foo

Retargeting The Binaries

The configuration settings critcl uses to drive the compiler, linker, etc. are by default selected based on the platform it is run on, to generate binaries which properly work on this platform.

There is one main use-case for overriding this selection, which is done with the option -target:

  1. Cross-compilation. The building of binaries for a platform T while critcl actually runs on platform B. The standard configuration of critcl currently has settings for two cross-compilation targets. So, to build 32bit Windows binaries on a Linux host which has the Xmingw cross-compilation development environment installed use

        critcl -pkg -target mingw32 foo
    

    Similarly, building a package for use on ARM processors while critcl is running in an Intel environment use

        critcl -pkg -target linux-arm foo
    

    Note that both configurations assume that the cross-compiling compiler, linke, etc. are found first in the PATH.

Custom Configurations

The compiler configurations coming with critcl currently cover all hosts having gcc installed (the foremost among these are Linux and OS X), plus the native compilers of the more common unix-like operating systems, i.e. Solaris, HP-UX, and AIX, and, on the non-unix side, Windows.

Developers using operating systems and compilers outside of this range will either have to install a gcc-based development environment, i.e. get into the covered range, or write their own custom configuration and then tell critcl about it.

The latter is the easier part, given that critcl supports the option -config whose argument is the path to the file containing the custom configuration(s). I.e.

    critcl -config /path/to/config ...

will run critcl with the custom configuration in "/path/to/config", with the other options and arguments as explained in previous sections. Depending on the choice of name for the new configuration(s) this may or may not require a -target option to select the configuration needed.

For the former, the writing of the custom configuration, the reader is refered to the section "Configuration Internals" of the CriTcl Package Reference for the necessary details. This is an advanced topic pretty much out of scope for this tutorial beyond what was already said.

Custom Header Path

Sometimes the use of critcl::headers might not be enough for a package to find its headers. Maybe they are outside of the paths checked by the setup code. To help the application recognizes the option -I which allows the user to supply a single additional include path to use during the build phase of the package.

Simply use

    critcl -I /path/to/header ...

and the specified header will be handed to the package to be built.

Introspection of Targets And Configurations

To see a list containing the names of all the available configurations, run

    critcl -targets

The configuration settings for either the default or user-chosen target can be inspected on stdout with

    critcl -show

and

    critcl -show -target TARGET

The raw contents of the configuration file used by critcl are dumped to stdout with

    critcl -showall

All of the above can of course be combined with custom configuration files.

Authors

Jean Claude Wippler, Steve Landers, Andreas Kupries

Bugs, Ideas, Feedback

This document, and the package it describes, will undoubtedly contain bugs and other problems. Please report them at https://github.com/andreas-kupries/critcl/issues. Ideas for enhancements you may have for either package, application, and/or the documentation are also very welcome and should be reported at https://github.com/andreas-kupries/critcl/issues as well.

Keywords

C code, Embedded C Code, code generator, compile & run, compiler, dynamic code generation, dynamic compilation, generate package, linker, on demand compilation, on-the-fly compilation

Category

Glueing/Embedded C code