                  CMunge: A free alternative to CMHG
                  ----------------------------------

[IMPORTANT: Please read the Copyright and Warranty files!]

What is CMunge? And what is CMHG?
---------------------------------

This is software for RISC OS. If you don't have a RISC OS machine, or you
aren't developing software for one, then don't bother reading any further.

OK. The base functionality of a RISC OS machine can be extended by loading
new software components called 'relocatable modules'. These are loosely
equivalent to DLLs on Windows etc. If you don't know what a relocatable
module is, then you are out of your depth. Go read the PRMs or something.

Modules have traditionally been written in assembler for speed and
compactness, but there are cases where it is necessary or desirable to
write them in C.

Acorn developed a tool called 'CMHG' (the C Module Header Generator) that
generates a suitable veneer to allow modules to be developed in C. This was
supplied with the Desktop Development Environment (and probably earlier
versions of the C compiler too).

CMunge has been developed as free alternative to CMHG. It reads the same
input files, and (coupled with an assembler) performs the same task as
CMHG.


Differences between CMunge and CMHG:
------------------------------------

CMHG takes a textual input file, optionally preprocesses it, and produces an
AOF output file.

CMunge takes the same textual input file, again optionally preprocesses it,
but produces a textual output file suitable for feeding to an assembler. The
process can stop here, or CMunge can go ahead and invoke the assembler on this
file to produce a functionally identical output file.

The command line flags aren't identical, but should be close enough to merely
change the command name.

See the section on 'CMunge only features', later for more details of
differences.


Command Line Syntax/Flags/Options:
----------------------------------

Usage: cmunge [options] infile

Options:

-h              Output help about the command.
-o <file>       Name the output AOF file.
-s <file>       Name the output assembler file.
-d <file>       Name a C header file that CMunge will generate.
-x<type> <file> Name an export file to place SWI numbers in. Types can be :
                  hdr   assembler header
                  h     C header
-p              Pre-process the file with the C pre-processor before parsing.
                In this mode, all content within the included files will be
                ignored (only the pre-processor directives will be obeyed).
-px             Pre-process the file with the C pre-processor before parsing
                in extended mode. In this mode, all content within the
                included files will be processed by CMunge.
-I<path>        Include <path> on the #include search path.
-D<symbol>=<value>
-D<symbol>      Define <symbol> on entry to the preprocessor.
-depend <file>  Name a dependency file for AMU
-U<symbol>      Undefine <symbol> on entry to the preprocessor.
-throwback      Use throwback for error processing.
-zbase          Define symbol Image__RO_Base in generated header.
-zerrors        Define errors and veneers only (for use in applications).
-zoslib         Use '#include "os.h"' (OSLib) instead of '#include "kernel.h"'
                (Acorn) in generated C header file.
-blank          Generate an empty CMunge file instead of reading one.
                When in 'CMHG' mode this will generate a warning but will
                generate a header containing only CMHG-based features.
-cmhg           Give warnings about CMHG incompatibilities.
-tgcc           Use GCC tool chain to generate output.
-tlcc           Use LCC tool chain to generate output.
-tnorcroft      Use Norcroft tool chain to generate output (default).
-32bit          Generate 32bit code.
-26bit          Generate 26bit code (default).
-apcs 3/<flags> Specify APCS variant to use - see below for more details.



Using CMunge:
-------------

From here it gets technical. If you are to the stage of generating modules
then you should be competent enough to understand the following.

Input to cmhg is given in the form of a text file. This text file is
relatively free form, but the following (insane) set of rules was laid
down by Acorn.

  * The file consists of a sequence of logical lines.

  * Comments begin with ';' (a semicolon) and extend to the end of the
    physical line they start on. All comments are removed from the lines
    before the following rules are considered.
    
  * If a physical line ends with ',' (a comma) then the logical line is
    considered to extend onto the next physical line.
  
  * Adjacent quoted strings (even separated by whitespace/newlines) are
    concatenated before being considered.

  * Numbers are assumed to be unsigned decimal unless preceded with '&'
    (ampersand) or '0x' in which case they are taken as hexadecimal (but
    see the section on 'Evaluation of integer elements')
  
  * Keywords are a series of alphabetic characters and minus signs terminated
    by ':' (a colon). They may be followed by zero or more arguments.

  * Lists of arguments may be separated by ',' (commas) or spaces; commas are
    required when splitting lists across lines.
    
The simplest way to get a quick reminder of the parameters that CMunge can
take on each of its keyword lines in the input file is to use the command :

  CMunge -blank blank

which will create an example CMunge file. This example file is commented and
can be used to construct a new header file for your module.

An example input file therefore might look like this:

---START--->8----8<---START---
; CMunge example input file

module-is-runnable:

initialisation-code: initialise

finalisation-code: finalise

service-call-handler: service 1 2 3,
                      &80500, 0x9999

title-string: CMungeExample

help-string: CMungeExample 0.01 Copyright (C) 2000 Robin Watts

command-keyword-table: command

       CMungeExOne(min_args: 0, max-args: 1,
                   gstrans-map: 0 fs-command: status:,
                   configure:, help:,
                   invalid-syntax: "Syntax: CMungeExOne <arg>",
                   help-text: "CMungeExOne has no purpose. It's only for "
                              "example purposes.\n" "So there\n")

swi-chunk-base-number: 0x12340

swi-handler-code: swi_handler

swi-decoding-table: CMungeEx, One Two Three Four,
                              Five Six Seven

swi-decoding-code: swi_decoder

irq-handlers: irq1, irq2/irq2_thingy

event-handler: event/event_thiny 1 2 3 4

vector-handlers: tickerv, palettev

library-initialisation: lib_init

generic-veneers: fs_entry/fs_handler
---END----->8----8<-----END---

Broadly each entry in that file, sets up some way of the module being
called. The exceptions to this are the title and help strings which simply
define how the module declares itself to the operating system.

Each possible keyword is explained below, along with the C interfaces
the generated header will have. If the -zoslib flag is used, then every
occurrence of _kernel_swi_regs below is replaced by os_register_block, and
every occurrence of _kernel_oserror is replaced by os_error.

  * module-is-runnable:

    When the module is run, the C main function will be called in user mode
    as below:
    
    int main(int argc, char *argv[]);

    Normal C functions such as malloc can be called. Without this keyword
    the main function will never be called.
 
  * module-is-c-plus-plus:

    If a module includes C++ code it will be necessary to perform certain
    operations to set up the environment for the language correctly. When
    this keyword is specified, these actions take place.
    
    When the module is first loaded, after the C environment is initialised,
    but before the initialisation-code entry (if any) is called, it is
    necessary to ensure that global constructors are called. The function
    ______main will be called to perform this task. This function is
    provided within the C++ library supplied with Norcroft C (often called
    C:o.c++lib).
    
    This directive can be emulated by placing a call to ______main() at the
    very start of the initialisation-code entry, should you wish to do so
    for CMHG compatibility.

  * initialisation-code: <C-function-name>
  
    When the module is first loaded or reinitialised a function of the
    specified name is called. For instance, with the name 'init' this would
    be as follows.
    
    _kernel_oserror *init(char *tail, int podule_base, void *pw);
    
    This function should do any initialisation required (claiming memory etc).
    'tail' points to the string of arguments (if any) that the module was
    invoked with. podule_base is the base address of the podule from which
    the module was invoked, or 0 if not loaded from a podule.
    
    pw is the private word for the module. This should never be used, except
    where it is required in calls to the operating system (for instance to
    setup calls to veneers generated by CMunge using OS_Claim etc).
    
    Return NULL for a successful initialisation, or a pointer to an error
    block otherwise.
    
  * finalisation-code: <C-function-name>
  
    Called when the module is killed (just before the C library is killed). For
    instance, with the name 'final' this would be as follows:
    
    _kernel_oserror *final(int fatal, int podule, void *pw);
    
    'fatal' is the value of R10 on entry to the finalisation code. 'podule'
    and 'pw' are as above.
    
    Return NULL to die, or an error pointer to refuse to die.
  
  * service-call-handler: <C-function-name> [<number> <number> ...]
  
    Called when a service call matching one of the given numbers is passed
    around. The generated header filters out calls for service calls that
    have not been declared to it. For C-function-name service, this gives:

    void *service(int number, _kernel_swi_regs *r, void *pw);
    
    If no service calls are specified, then all service calls are passed in.
    This is *not* recommended.
    
    When dealing with service calls it may be useful to have access to a
    pointer to the base of the module in memory. A suitable section of
    assembly code can be generated by CMunge to provide such a variable; to
    enable this, use the -zbase command line option. Use of this option will
    define the following in the header:
    
    extern int Image__RO_Base;

  * help-string: <module-name> <version-number> <comment>
  
    This defines the help string in the module header. The name of the module
    should ideally be less than 16 characters, and should ideally match the
    name given in the title-string entry.
    
    The version-number field should be in the form d.dd where d is a decimal
    digit.

  * command-keyword-table: <C-function-name> <command-description>+

    At least one command-description must be given.
  
    This generates the *command and help table for the module. Calls to the
    module are passed the the given C function as follows:
    
    _kernel_oserror *command(const char *arg_string, int argc, int number,
                             void *pw);
    
    'number' denotes which command was called (numbering from 0).
    
    If it is a *Help entry, then arg_string points to a buffer to write into;
    you will need to cast the 'const' away). Fill this with the zero terminated
    string to output, and return a pointer to it (i.e. the initial value of
    arg_string) to print. The created header file defines a special value
    'help_PRINT_BUFFER' to be an appropriate cast of arg_string. Be careful
    when using this; it relies on the value of arg_string not changing during
    the function. Otherwise return NULL to print nothing, or a pointer to an
    error block.
    
    If it is an actual command to be run, then arg_string is the command
    string to be executed and argc is the argument count. Return NULL on
    success or an error pointer.

    If its a *Configure option, and arg_string is set to arg_CONFIGURE_SYNTAX
    (as defined in the generated header file to be a cast of 0) then print the
    syntax string. Return NULL.
    
    If its a *Configure option, and arg_string is set to arg_STATUS (as defined
    in the generated header file to be a cast of 1) then print the current
    status. Return NULL.

    Otherwise, if its a *configure option, use arg_string and argc to set
    the configure option. Return NULL on success, an error pointer, or
    configure_BAD_OPTION (defined in the header file to be a cast of -1),
    configure_NUMBER_NEEDED (defined to be a cast of 1), configure_TOO_LARGE
    (2) or configure_TOO_MANY_PARAMS (3).

    Command descriptions are given in the following form:
    
      command-name ( min-args:       <integer>,         ; (0 to 255, default 0)
                     max-args:       <integer>,         ; (0 to 255, default 0)
                     gstrans-map:    <integer>,         ; (default 0)
                     fs-command:,                       ; (flag)
                     status:,                           ; (flag)
                     configure:,                        ; (flag)
                     help:,                             ; (flag)
                     international:,                    ; (flag)
                     add-syntax:,                       ; (flag)
                     invalid-syntax: <string>,          ; (default empty string)
                     help-text:      <string>,          ; (default empty string)
                     handler:        <C-function-name>, ; (explicit handler)
                     no-handler:                        ; (flag)
                   )

    Each sub-argument (min-args:, max-args: etc) is optional.
    
    The flags set bits in the flag word for that entry in the command table,
    except for add-syntax (which adds the invalid syntax message to the end of
    the help information).
    
    The handler specification allows dispatch of command to specific
    routines without a higher-level switching function. Commands without
    an explicit handler name will be dispatched to base function named
    at the start of the parameters. The base function may be specified as
    '-' to indicate that no such function exists. If this is the case,
    all the handlers must be specified within the commands, or the
    no-handler flag set.
    
    The no-handler flag indicates that there is no code handler for the
    command. Such command declarations are useful for providing help-only
    commands definitions.
    
    The help flag is not supported by this version of CMunge.
  
  * error-chunk-base-number: <number>
    
    This sets the error base number for the module. This is merely exported
    as ERROR_BASE in the C header. Error blocks are allocated in blocks of
    256, but CMunge will accept any alignment so that you can re-use areas
    of the blocks. You should inform the allocations service of any use of
    'sub-blocks'.
  
  * error-identifiers: <error description> ...
    
    This sets up error identifiers which will be embedded in the module
    header. Descriptions are of the form :

      identifiers ( [<number>,] <string> )
    
    You can specify each of these individually. Each of the errors will be
    allocated a unique number, starting at error-chunk-base-number. Each of
    the errors is defined in the header as being a _kernel_oserror *. If a
    number is given this will be used instead of the next error number in
    the chunk.
    [Note: The header defines these through a macro so that errors are not
     copied to the workspace]
  
  * swi-chunk-base-number: <number>
    
    This sets the SWI base number for the module. Do not specify the X-bit
    in your SWI base.

  * swi-handler-code: <C-function>
  
    Calls to the SWIs in the SWI block are passed through here; for instance
    the name 'swi_handler' would result in:
    
    _kernel_oserror *swi_handler(int number, kernel_swi_regs *r, void *pw);
    
    'number' is the swi number offset within the block (i.e. 0 to 63).
    
    'r' are the registers on input. Alter them (r0-r9) to change the return
    values.
    
  * swi-decoding-table: <SWI-base-name> <swiname>[/<C-function-name>]
                         <swiname>[/<C-function-name>] ...
    
    This defines the names put into the SWI decoding block. zero or more
    SWI names may be given. Where C function names are gives, these functions
    are called instead of the base SWI handler, with the same arguments.
    
  * swi-decoding-code: <C-function-name>[/<C-function-name>]
  
    This command works in 2 different ways, according to how many function
    names are supplied. The original definition in CMHG was to give just
    one name, which captured all the functionality. The latest versions of
    CMHG provide an improved interface using 2 C functions, one for each
    direction.

    Operation with 1 function:
    ~~~~~~~~~~~~~~~~~~~~~~~~~~
    This code can be called to translate SWI numbers<->names. For instance for
    a name 'decoder':
    
    void decoder(int r[4], void *pw);
    
    If r[0] < 0 then:
      r[1] points to a *control* terminated string to convert to a number.
      If you recognise the name, set r[0] to the offset within the block. If
      not, set r[0] < 0.
    otherwise:
      r[0] is the offset within the SWI block
      r[1] is a pointer to an output buffer
      r[2] is the offset within the buffer to start filling from
      r[3] is the pointer to the byte beyond the end of the buffer.
      Either fill the buffer from the required point (updating r[2], *without*
      a terminator), and return NULL or return an error pointer.

    Operation with 2 functions:
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~
    The first function handles name->number conversions, and the second
    handles number->name ones. For example 'name2number' and 'number2name':

    int name2number(const char *string, void *pw);
    int number2name(int number, char *buffer, int offset, int end, void *pw);

    name2number should convert the control (NOT zero) terminated string to
    a number, and return the number of the SWI within the SWI chunk (0 to 63).
    If unrecognised return <0.

    number2name should write the (unterminated) name of the SWI corresponding
    to the number of the SWI within the SWI chunk into buffer starting at
    offset, and return offset updated by the length of the name written. If
    there is no name for the SWI, just return offset, unaltered.

  * vector-handlers: <C-function-name>[/<C-function-name>][(<params>)] ...
 
    Several pairs of function names can be given here; each corresponds to an
    entry point into the module (to be used for OS_Claim or similar) and
    the name of the C function that that entry point calls. If you omit the
    '/' and the second half of the pair, than that name is constructed by
    adding '_handler' on to the end of the first name. Where parameters are
    supplied, the following parameters are supported :
    
      error-capable:
         The return parameter may be specified as 'VENEER_ERROR(error)',
         rather than an error block. This will cause the vector to be
         claimed and for an error to be returned in R0 with the V flag set.
         All other registers will be as returned in the parameter block.

    So, for 'vector-handlers: irq1' the following function is expected:
    
    int wrch_handler(_kernel_swi_regs *r, void *pw);

    and a symbol is setup to point to an entrypoint (NOT a C function):

    extern void wrch(void);
    
    You must provide wrch_handler; the CMunge generated header creates wrch
    for you. *Never* call wrch directly, just use its address in vector
    claims. All such vector claims must be done in SVC mode, such as from
    initialisation code or in a SWI handler.
    
    When wrch_handler is called, r->r[0]-r[9] are the values on entry. These
    can be updated to change the values on return. Consult the relevant
    vector documentation to be sure that you comply with its interfaces.

    Return VECTOR_CLAIM if you handled the interrupt, or VECTOR_PASSON to
    pass it on to the handlers deeper down the stack of handlers.
    Alternatively, if the error-capable veneer is in use, the
    VECTOR_ERROR(error) macro should be used to return the error pointer to
    the caller. If the 'error' parameter is NULL, the return will be
    equivilent to a VECTOR_CLAIM.
  
  * irq-handlers: <C-function-name>[/<C-function-name>] ...
  
    Several pairs of function names can be given here, generating entry points
    that call handler functions under the same entry/exit conditions as above.
    
    The irq-handlers declaration is the obsolete name for vector-handlers.
    Clients should use vector-handlers where possible.
 
  * event-handler: <C-function-name>[/<C-function-name>] [<number> <number>...]
  
    Only one pair of names can be given per use of this handler, but they
    generate an entry point that calls a handler function under the same
    entry/exit conditions as above. If multiple event-handler entries are
    required, using the token again allows this. Multiple event-handlers
    are accepted in cmhg mode, but only the last is processed fully in
    order to follow cmhg's behaviour.
    
    Fast accept/reject code is generated for the given event numbers. The
    numbers are optional, but strongly recommended.

  * generic-veneers: <C-function-name>[/<C-function-name>][(<params>)] ...
  
    Several pairs of function names can be given here, generating entry
    points that call handler functions under the same entry/exit conditions
    as above. Where parameters are supplied, the following parameters
    are supported :
    
      private-word: <register>
         The register in which the private word value will be passed on
         entry to he generic-veneer. This defaults to r12. Normally r12
         is not preserved by a generic-veneer unless a private-word register
         is provided. To force it to be preserved, use r12 as the register
         name.
      
      carry-capable:
         The return parameter may be specified as 'VENEER_SETCARRY', rather
         than an error block. This will cause the carry flag to be set on
         return from the handler function. All registers will be as
         returned in the parameter block.

  * vector-traps: <C-function-name>[/<C-function-name>] ...
  
    Several pairs of function names can be given here, generating entry points
    that call handler functions under the similar entry/exit conditions as
    above. The entry point for vector traps is slightly different :

    _kernel_oserror *handler(_kernel_swi_regs *r,void *pw,
                             vectortrap_f trap,void *trappw);
    
    The register block passed in is that which the vector was called with.
    The trap function can be called to call the remainder of the vector
    chain :

    _kernel_oserror *trap(_kernel_swi_regs *r2,void *trappw);
    
    The register block passed to the trap routine will be used to call the
    remainder of the chain and updated with the values returned by the
    call. If may be the same as the input register block, or a different
    block. If the vector chain returns an error, you can process this and
    return it to your caller, or perform any other operation.
    
    If you return an error from the routine, this will claim the vector
    and return the error.
    If you return NULL from the routine, this will merely claim the vector.
    If you return VECTORTRAP_PASSON, the remainder of the vector chain will
    be called, as if you had used 'return trap(r,trappw);'. The advantage
    of this is that it reduces stack usage quite considerably and allows
    you to reject vectors calls which are of no interest to you.

  * library-initialisation-code: <symbol-name>
  
    Without this supplied, the veneer calls _clib_initialisemodule. In some
    circumstances you may wish to do some work before calling this function,
    however, so this call can be redirected to any symbol-name you wish.
    
    This routine will be called without a proper C environment having been
    setup, so will need to be written in assembler. For proper operation of
    the C library, it is essential that your routine call 
    _clib_initialisemodule before it returns.

  * international-help-file: <pathname>

    This option allows a literal pathname to be embedded in the header of the
    module such that any help messages tagged as 'international' returned from
    the module will be looked up in this file, if the OS version supports it.


Makefile:
---------

If you want to use CMunge in a Makefile, and have it generate the header
file each time then you will need to define two static dependency lines.
The following example can help you :

h.modheader:	o.modheader
o.modheader:	cmhg.modheader
	cmunge -throwback -zoslib -zbase -o o.modheader \
	       -d h.modheader cmhg.modheader
o32.modheader:	cmhg.modheader
	cmunge -32bit -throwback -zoslib -zbase -o o.modheader \
	       -d h.modheader cmhg.modheader

For non-OSLib development you would probably want :

h.modheader:	o.modheader
o.modheader:	cmhg.modheader
	cmunge -throwback -o o.modheader \
	       -d h.modheader cmhg.modheader
o32.modheader:	cmhg.modheader
	cmunge -32bit -throwback -o o32.modheader \
	       -d h.modheader cmhg.modheader

With a rule based makefile you might use :

.SUFFIXES:		.c .o .o32 .cmhg
.cmhg.o .cmhg.h:;	cmunge -throwback -o o.$* -d h.$* $<
.cmhg.o32 .cmhg.h:;	cmunge -32bit -throwback -o o32.$* -d h.$* $<


CMunge only features:
---------------------

Because CMunge and CMHG are different, it is important to know where the
differences lie in case you have to change back to a CMHG only environment
(though why that might be when the source is available for bug fixing, and
CMunge is bug free anyhow, I don't know :-) ).

CMunge will parse CMHG files. If you find a CMHG file that will not work with
CMunge, or generates the wrong code, then it is NOT a feature of CMunge; it
is a bug. CMunge also adds to the file format :

  * Error base handling (error-chunk-base-number)
  * Error identifiers (error-identifiers)
  * SWI specific handlers (<swi>/<handler> in swi-decoding-table)
  * Vector trap handlers (vector-traps)
  * Generic veneers can now have parameters (private-word, and carry-capable)
  * Vector handlers can now claim with errors (error-capable)
  * Simple additive evaluation of integers (see below)
  * Simpler C++ module creation (module-is-c-plus-plus)
  * Command keyword handlers (handler) and handler omission (no-handler)

The command line differs from CMHG in the following ways :
  * Exported header types (-x switch)
  * Extended pre-processing operation (-px switch)
  * Optional build specification using the -z switch (OSLib-specific,
    error block generation only, Image__RO_Base definition)
  * Blank template file generation (-blank switch)
  * Multiple toolchain suport (-t switch)
  * Explicit APCS variant selection (-apcs switch)
  * Backwards compatibility with CMHG (-cmhg switch)

Within the header produced, there are a couple of deviances :

  * CMUNGE_VERSION is defined to the version of CMunge that produced the
    file.
    CMHG_VERSION is defined to be 531 which is the closest equivalent
    version of CMHG to the current version of CMunge.
    CMHG defines CMHG_VERSION to its version but does not define
    CMUNGE_VERSION.
  * Module_VersionString is defined to the /full/ version string in CMunge,
    but just the version number in CMHG.

If these cause a problem, please let us know.


Evaluation of integer elements
------------------------------

Standard CMHG supports integers which can be specified in a couple of
standard forms :

  * Prefixed by & for RISC OS-style hexadecimal notation.
  * Prefixed by 0x for C-style hexadecimal notation.
  * Unprefixed for decimal notation.

CMHG will strip any surrounding parentheses from the integer value in order
to allow values to be used from C-style header files through #include
directives.

CMunge also supports very simple additive operations on integer values.
This means that expressions such as '(&800 + 1)' are acceptable within the
context of an integer value in a CMHG file parsed by CMunge. No other
arithmetic operations are supported by CMunge. This change allows the 'base'
values to be defined within a header and offsets to be specified, for
example defining a service base and then adding on offsets for particular
reasons.


32bit header generation:
------------------------

In order to support generic compilation, CMunge has been updated to generate
32bit header veneers. These veneers should function identically on 26bit and
32bit, so far as the operation of the module is involved. 32bit code has been
tested in both 32bit and 26bit environments with a large number of components
and been found to be reliable.

The header generation is as close to that of CMHG as possible without
introducing any faults; this may result in slightly different code sequences
(over and above those required for 32bit) to those which might be expected
of a pure 26bit to 32bit conversion.

For 32bit builds, all that should be necessary is to provide the APCS
variant specifier as :
   -apcs 3/32bit/fpe2/swstackcheck/fp/nofpregargs
For brevity (and because the defaults are suitable, this can be abbreviated
to :
   -apcs 3/32


Explicit APCS specification:
----------------------------

In line with the compiler and assembler (for ARM and consequently Norcroft
tools), the CMunge tool may take an explicit specification of the APCS
variant which should be used. CMunge does not support the full range of
APCS variants, but this option allows makefiles to be created which use
APCS specifications to describe the type of compilation required. It is
recommended that the build tools use the full APCS specifier as this then
means that the same specifier can be supplied to all tools (cc, c++, objasm,
cmunge) regardless of the version of the tool. All version 5 Norcroft
tools accept the -apcs switch.

The full syntax for the APCS declaration can be obtained through the
command :

  *CMunge -help


GCC, LCC and Norcroft Toolchains:
---------------------------------

Because CMunge is intended to be a free replacement for CMHG, it can be used
with the freely available GCC toolchain as well as the commercial Norcroft
toolchain.

The '-tnorcroft' option on the command line will ensure that the Norcroft
toolchain is used. This is the default and all testing of the CMunge tool
has been performed using the Norcroft toolchain rather than the other two.

The '-tgcc' option on the command line will ensure that the GCC toolchain is
used. John Tytgat has contributed significantly to ensure that the GCC
toolchain is able to be used with CMunge.

The '-tlcc' option on the command line will ensure that the LCC toolchain is
used. LCC does not support generation of relocatable modules.
Filenames given to the LCC version of cpp must be in unix-form for the input
and path names. CMunge deals only in native filenames so it is recommended
that the LCC toolchain not be used with CMunge. As such, the lcc toolchain
should be considered unusable with CMunge.

Norcroft building:
  cc     - Pre-processes the input if -p used
  objasm - Assembles the generated code
  
GCC building:
  gcc    - Pre-processes the input if -p used (-xc mode)
           Note that there is no dependency support.
  gcc    - Assembles the generated code (-xassembler mode)
  
LCC building:
  cpp    - Pre-processes the input if -p used
           Note that there is no dependency or throwback support with the
           LCC version of cpp.
  as     - Assembles the generated code


Possible Future Enhancements:
-----------------------------

'Short' or 'Fast' veneers; don't set up the C stack entrails, just do the
bare minimum necessary to get C functions to work. Handler functions called
in this way will not be able to access static data or call Shared C library
functions. Is this worth the effort? Probably not; if its that speed
critical, ARM code it!

Generate modules capable of benefitting from the RISC OS 4 (Ursula) service
call entry table with bit 0 set (so the number passed in is not the service
call, but really the number of the service within our enumeration).

Should we have a facility to output C veneers for the SWIs exported ? No, I
think we leave that function to DefMod - it can do it more type-safely than
we can.

The C header generation at present exports only the SWI numbers. This
results in lots of modules having two files - one for the SWIs and one for
the definitions of the module (constants, services, etc). A way of providing
a template would help here. This has been addressed locally by the perl
'MergeModeSWIs script'.

Events, Vectors, etc with their own handler routines for particular cases.
These would remove the requirement for the switch statement in a central
handling module. The impact on SWI decoding was four instructions (the same
as it would be for a switch) when I added that, so it should not be too
important.

I (Justin) don't like the name 'codesupplied' in the source. It should really
be 'somehandlerssupplied' but that's a bit long. It was 'special', but that's
even more vague. Similarly, I dislike the 'errors_special' I've used for the
flagging of error blocks that have no base as they all have their numbers
allocated. It also doesn't work properly. I think I need a set of flags for
that.

Add some extra types to the -x generation rule :
  -xasm : Assembler SWI number header
  -xc   : C SWI veneers (needs more 'knowledge' of the format)
  -xbas : Basic SWI numbers (it's easy, so why not?)

It might be nice to have a 'reformat cmhg file' option, which read in the
file and wrote it out again 'neatly'. I'm not sure it'd be such a great idea,
but it's worth a little thought.

It would be nice to have a private value passed to generic-veneer functions.
At present, the call is global to the module and the module must itself
decide who it is dispatching for. In the case of callbacks (the most common
use), this isn't really a problem because all callback activities can take
place at once, but where multiple private callbacks are vectored through the
same handler this may now be possible. Using a parameter block containing the
private parameters to pass to the routine would be handy. I've written veneers
for this by hand, but it would be better to do this in a more generic manner.

Reformat this manual to be more readable.


Possible faults:
----------------

The cost of a private-word is taken by *all* generic veneers in that an
extra register must be preserved for all entry points. Any definitions using
a private word will also gain an extra instruction in their entry sequence
to relocate the private-word. I (Justin) don't feel this is a serious
problem.

The cost of a error-capable vector handler is taken by all vector handlers,
although it should only be a few instructions (the code sequences vary quite
a bit between 26bit and 32bit versions).
