amiga-libnix3/libnix.texi

\input texinfo
@setfilename libnix.info
@settitle libnix - a static library for GCC on the amiga
@setchapternewpage odd

@ifinfo
This is libnix.texi the documentation of the libnix library.
Since this document is part of the libnix package it stands
under the same copyright as the library itself, i.e. it's
public domain.
@end ifinfo
@titlepage
@title libnix
@subtitle a static library for GCC on the amiga
@author Matthias Fleischer & Gunther Nikl
@end titlepage

@node Top,Description,(dir),(dir)
@ifinfo
This is the documentation of libnix. The following is a list
of chapters. You need not read all of them - but reading
the chapter Features is recommended.
@end ifinfo

@menu
* Description::      What is libnix?
* Authors::	     Who did it?
* Disclaimer::	     Copyright and other legal stuff.
* Naming::	     Naming conventions.
* Usage::	     How to use it ;-)
* Features::	     Details of the implementation (and more).
* Special startups:: How to write shared libraries and devices with gcc.
* Set elements::     A nice feature of gnu ld.
* geta4::	     Some words on code and data models
* Library bases::    And how they work.
* Code size::	     How to write small programs with gcc
* FAQs::	     Frequently asked questions and answers.
@end menu
@node Description,Authors,Top,Top
@chapter What is libnix?

libnix is a static (i.e. link) library for usage on amiga computers
together with gcc 2.3.3 or above. It is very amigalike and contains
a lot of features you probably don't want to miss:

@itemize @bullet
@item
auto-open-library feature
@item
SAS compatible handling of WB startup message
@item
auto-detach startup code
@item
is very short
@item
does not require a shared library
@item
auto stack-extend
@item
and much more
@end itemize

So if you want to write amiga specific programs or if you only need ANSI support
instead of unix compatibility - and if you don't want to redistribute ixemul.library -
this can be your choice.

But be aware - libnix requires Amiga OS 2.0 or higher :(.

@node Authors,Disclaimer,Description,Top
@chapter Authors

If you want to change anything in the library please contact one of us first -
we know how to do it best since we wrote it. If you want to report
bugs please contact us, too. If we don't know about them we cannot fix them.

Matthias Fleischer
@*fleischr@@izfm.uni-stuttgart.de

Gunther Nikl
@*gnikl@@informatik.uni-rostock.de

Special thanks go to Gerhard M<>ller for writing gerlib, Philippe Brand for
finding a name for this beast and Kriton Kyrimis who contributed some
of the non-ANSI functions and a lot of ideas.

@node Disclaimer,Naming,Authors,Top
@chapter Disclaimer

This package is public domain. That means that you can copy, use and modify it
without any problems and that you can get it for free. If you actually
paid for getting it this is completely your fault - I didn't see a cent
of that money. It also means that I cannot be made responsible for any damage
resulting out of the use of it - you simply shouldn't trust anything you didn't
pay for :-).

@node Naming,Usage,Disclaimer,Top
@chapter Naming conventions

This library is not only for the end user but also for the library
programmer (if you want to write your own startup, etc).

If you want to write code for it you should be aware of the normal
naming conventions for ANSI libraries:

@itemize @bullet
@item
Names with no underscore @samp{foo} are ANSI or POSIX compliant - there is absolutely
no risk in using them. If you use only these you can write portable
programs.
@item
Names with a single underscore @samp{_foo} are ANSI extensions for the
end user. Usually they are very common on certain systems but not used on
others.
@item
Names with two underscores @samp{__foo} are for the library programmers only.
If they are not documented you cannot rely on them. And even if they are
you should use them only for writing library code.
@end itemize

There is only one exception of these conventions (@samp{__chkabort()})
and this is for compatibility reasons.

@node Usage,Features,Naming,Top
@chapter Usage

The usage of this library is like any other link library. The only
important thing is the right linkage order:

@enumerate
@item
The startup code has to be used first :-)
@item
The stubs-library has to be used last since it contains the library base pointers.
@item
The commandline parser should be used after your code but before most
other things or you will run into problems.
@end enumerate

Normally this is handled by the specs file of gcc.

@code{gcc (-fbaserel) (-resident) -noixemul YOUR_OBJECTS (-lnix_main) (-lm)}

If you use @samp{-lnix_main} you get a different commandline parser.
@samp{-lm} uses the math library.

Be aware that the formatted I/O-functions need the math library
to work correctly for floating point numbers. Without the math
library you get only floating point support for simple operators
like @samp{+}, @samp{*}, casts and that like.

If you don't use the assembler inline functions of gcc you will have
to use @samp{libamiga.a} if you want to use any Amiga OS function.
gcc comes with a free version of libamiga.a which is a subset
of the original one. You can build it yourself if you unpack the sources,
but be prepared that it may take some time.

For compiling a4-relative programs you should choose the
@samp{-fbaserel} option. You get resident (pure) programs if
you set the @samp{-resident} option. Anything else necessary for these options
is handled by the specs-file (choosing the right startups and libraries).

If you (for some reason) don't trust you specs-file you can call everything
by hand:

@code{gcc -nostdlib ncrt0.o YOUR_OBJECTS libnixmain.a (libm.a) libnix.a
libstubs.a}

But that's not the recommended way. Therefore I don't explain this in
detail here - use the @samp{-v} option of gcc for more details.

@node Features,Special startups,Usage,Top
@chapter Features - what you get

The following list contains the elements of this package in
the right linkage order. This means if you follow the list from
top to bottom and take one file of each menu entry you will get
a working configuration.

@menu
* Startup codes::      Does all work necessary for startup.
  Your objects	       Need I say anything about that?
* Commandline parser:: Calculates argc and argv.
  libm.a	       The math library (optional).
* libstack.a::	       Stack extension code (optional).
* libnix.a::	       The library itself.
* libamiga.a::	       If you have one.
* detach.o::	       Auto detaching
* libstubs.a::	       The library bases.
@end menu

@node Startup codes,Commandline parser,Features,Features
@chapter Startup codes

There is a lot of work to do before your @samp{main} function
can be called - open shared libraries, open stdin, stdout, etc.
Depending on the compiler options and the ANSI functions you use.
This work is done by the startup code.

@menu
* Startup interface::
* Startup usage::
@end menu

To get a short startup all the necessary modules are optional -
they get only linked in if you use them. There are 2 exceptions
from this (since the linker cannot check for it):

@itemize @bullet
@item
The commandline parser. It can be deactivated by declaring

@code{__nocommandline}

somewhere in your program (the type doesn't matter).
@item
The shared-library-opening module. You should not disable it unless
you know what you are doing since most library functions depend on it.
@end itemize

The startup codes itself are written in assembly to be as short as
possible.

Here is a little program to get the point:

@example
#include <inline/exec.h>
#include <dos/dos.h>
#include <inline/dos.h>
#include <workbench/workbench.h>

int __nocommandline=1; /* Disable commandline parsing  */
int __initlibraries=0; /* Disable auto-library-opening */

struct DosLibrary *DOSBase=NULL;

extern struct WBStartup *_WBenchMsg;

int main(void)
@{ if(_WBenchMsg==NULL)
  @{ if((DOSBase=(struct DosLibrary *)OpenLibrary("dos.library",37))!=NULL)
    @{ Write(Output(),"Hello world\n",12);
      CloseLibrary((struct Library *)DOSBase); @} @}
  return 0;
@}
@end example

compiled and linked with

@code{gcc -noixemul -s -O2 -fbaserel helloworld.c}

gives an executable of 492 bytes. And this with the normal @samp{main}
function!

So you never need to try to write a program without a startup code.

@node Startup interface,Startup usage,Startup codes,Startup codes
@chapter Startup code interface

The startup codes do the following:

@itemize @bullet
@item
They catch the workbench startup message and place it into the
variable

@code{extern struct WBStartup *_WBenchMsg}

you can simply look into this place (and test for a @samp{NULL} pointer) to
check if your program was started from WB. If this is a @samp{NULL} pointer

@code{extern char *__commandline}

contains the (@samp{\n} terminated) parameters of the commandline.
@item
They call all functions in the

@code{long __INIT_LIST__[];}

with ascending priority.
@item
They call the function

@code{int main(int argc,char *argv[])}

You can exit by simply falling through the end of @samp{main} or by calling

@code{__volatile void exit(int returncode)}

which does the cleanup:
@item
It calls all functions in the

@code{long __EXIT_LIST__[];}

with descending priority.
@item
It replys the WB startup message if necessary, resets the stackpointer
and returns to the shell.
@end itemize

@samp{__INIT_LIST__} and @samp{__EXIT_LIST__} are two set elements
which are a speciality of the gnu ld. Since everything that needs
initialization works over these two lists the bare startups are
very short. In fact they are even shorter then some low-level-startups

You can easily add your
own functions to the startup procedure by using the macros
in the file @samp{headers/stabs.h} - but keep in mind that this
is non-portable. Priority values <=0 are reserved for library implementors.

@node Startup usage,,Startup interface,Startup codes
@chapter Startup code usage

There are currently 3 startup codes in this package (maybe there will be more
in the future). Depending on the code and data model you use and some other
things you should choose one of them:

@table @samp
@item ncrt0.o
This is the normal (i.e. large code, large data model) startup.
It contains a @samp{geta4()} entry point to enable you to use one source
for two code models. There is no other need for this function.
@item nbcrt0.o
This one is for compiling small data model (a4 relative) programs.
There is a @samp{geta4()} entry that places the right information into a4.
Use this startup code if you compiled with @samp{-fbaserel}.
@item nrcrt0.o
This startup code allocates a new data area every time you call it.
Even if you don't call it at all the data are is there once.
This gives you multientrant and reentrant code.
Therefore this startup code is for compiling resident (pure) programs.
Resident programs are always small data model if you let the compiler
do the work.

There is no @samp{geta4()} entry - I just don't know how this could be done.
(If you start your code 10 times and want to access global data
out of a hook you cannot tell which one of the 10 data areas to use
because you want to access the data from a different task!)
@end table

@node Commandline parser,libstack.a,Startup codes,Features
@chapter Commandline parser

There are currently 2 commandline parser modules in the libnix package.
You can easily write your own by looking into the examples

@table @code
@item libnixmain.a
This is the normal one, i.e. it does all the work necessary for
ANSI compatibility and gives you the normal @samp{main} calling
convention. You can shut down the commandline parsing (if you want
to use the amiga OS commandline parser) by declaring

@code{__nocommandline}

somewhere in your code (the type of it actually doesn't matter). This
spares some bytes and is compatible to every other compiler.

And you can declare your own WB shell window by declaring a

@code{char __stdiowin[]}

variable somewhere in your code (but only if you parse the commandline -
without the commandline parser you get no window at all!).

@item libnix_main.a
This is a special version of a commandline parser - it doesn't call
the normal @samp{main} but

@code{int main(char *commandline)}

@samp{commandline} is the complete commandline - including the quoted
filename of your program (it's only quoted, not escaped - this is
for compatibility reasons :-( ). You might think the name of the game
should be @samp{_main} and not @samp{main} - and you are completely
right. You can use @samp{_main} for @samp{main} and @samp{_exit} for
@samp{exit} - there are symbol redirections for these and the linker
does the work.

This commandline parser is useful for compatibility.
You can use it as a second example or for recompiling PD programs that
use the single argument. You cannot use it for compiling ANSI code.
@end table

@node libstack.a,libnix.a,Commandline parser,Features
@chapter special stack handling facilities

The current Amiga OS (V3.1) has a very limited stack handling compared
to most other OSs: Every process has it's own fixed sized stack - and that's
all about it. The usual default for this stack is 4k, but that's not enough
for more complicated purposes (like for example compilers). Setting a higher
default is no real solution because it costs a lot of (widely unused) memory
and may be overrun, too :-(.

But fortunately you can get stack extension with a little help of the compiler ;-).
Starting with V0.9 of libnix and V2.7.0 of gcc you get a fully featured stack
extension facility. The old stack swap method is still provided (not only for
compatibility but also because it's simpler) but please don't try to mix it with
the newer check/extend methods.
@menu
* swapstack::		       Old method.
* stackextend implementation:: How stack extend works.
* stackextend usage::	       Usage. @strong{read}
* Advanced::		       Fine tuning.
* Costs::		       Some damned lies (Benchmarks ;-) ).
@end menu
@node swapstack,stackextend implementation,libstack.a,libstack.a
@chapter Minimum stack setting

Most large tools need more stack than the default 4096 bytes. If your tool is
one of them you can either rely on the user being able to raise the current stack
or you can let libnix raise the stack for you. At startup this module checks if the
current stack is large enough for your needs and switches to a new one if not.
All you have to do is to provide a variable

@samp{unsigned long __stack=@{required stacksize@};}

somewhere in your code and to link with the appropriate swapstack.o module.
@node stackextend implementation,stackextend usage,swapstack,libstack.a
@chapter Implementation of stack checking and extension

The basic principle of stack checking is that the compiler emits special code
to check if the stack is large enough whenever there's need for a bigger chunk
of stackspace, i.e. at function entry when local arrays are allocated, at
the start of blocks with local variable sized arrays and when calling @samp{alloca()}.
If the needed stackchunk is bigger than the left stackspace the program ends.

Since this special code costs memory and CPU time smaller stackneeds (e.g. when
calling library functions) are handled by not really checking against the hard border
of the stackframe but against one that leaves a certain amount of stackspace left
(@xref{Advanced}.). If you like to call functions with a lot of arguments
(more than 256 bytes) you should raise this value.

Stack extension builds on the same basic principle but allocates a new stackframe
whenever necessary. If this happens at the entry of a function with arguments
they have to be copied to the new stackframe so that the function may use them.
Since C allows for a variable number of arguments the compiler doesn't always know
how many arguments there are. Therefore only a fixed number of bytes is copied.
If your functions may have lots of arguments (again more than 256 bytes) you
should raise this number.

Since allocation and freeing of memory through OS functions costs a lot of
time (while a stack tends to be very dynamic) libnix caches once used stackframes
and utilizes them again if necessary. The memory needed for this doesn't accumulate
or such but just sticks to a maximum value raised once. This may look like a memory leak
(while in fact it isn't). Be prepared for it.
@node stackextend usage,Advanced,stackextend implementation,libstack.a
@chapter Using stack checking or extension
To utilize the stack checking or extension feature you need at least V2.7.0 of
gcc. With this compiler you get 2 new amiga specific options that emit special
code whenever necessary:

@itemize @bullet
@item
@samp{-mstackcheck} Emits code that checks if there is enough stack left.
The program exits if not.
@item
@samp{-mstackextend} Tries to extend the stack before exiting (this may
happen due to low or fragmented memory).
@end itemize

Always use those switches together with @samp{-lstack} to link with the stack
extension code or you will get a lot of undefined references ;-).
You can mix functions compiled with or without stack checking and extention
without problems.

@strong{Caution:}

Do not use stack checking and/or extension switches when
compiling hook or interrupt code. Both run in alien contexts with a different
stack and all stack magic must fail. Also don't try to do some other stack magic
on your own if you want to use stack extension.

Also note that a program compiled with stack extension/checking may @samp{exit()} at
@emph{any} function entry or when using alloca or variable sized arrays. Either prepare
your cleanup function accordingly (use @samp{atexit()}) or don't use this feature.

If you like to write or call functions with more than 256 bytes of arguments
(64 ints, longs or pointers) you should adjust the behaviour of the
stack extension code (@xref{Advanced}.).

@node Advanced,Costs,stackextend usage,libstack.a
@chapter Stack extension fine tuning

To adjust the behaviour of the stack extension code to your personal needs
you may set some of the following variables (or functions)

@samp{unsigned long __stk_minframe} (default: 32768)

Minimum amount of memory to allocate for a new stackframe. Setting a higher
value speeds the code up but costs more memory if it is unused.

@samp{unsigned long __stk_safezone} (default: 2048)

Size of the safety zone. Set this to a higher value
if you want to @emph{call} functions with lots of arguments.

@samp{unsigned long __stk_argbytes} (default: 256)

Number of bytes copied as arguments. Set this to a higher
value if @emph{your} functions may have lots of arguments.

@samp{void _CXOVF(void)}

Is a user replaceable stack overflow handler. The default one just pops
up a requester, then exits. This function is not allowed to return.

@node Costs,,Advanced,libstack.a
@chapter Overhead of stack extension

The additional code needed for stack extension (or checking) costs memory
and CPU power. Here are some numbers to give you a very rough idea for it.
(Times are in 1/60s, sizes in bytes):

@example
Test		  normal      checking	   extending	 extending
		(big stack)  (big stack)  (big stack)  (small stack)

Simple recursive
function runtime    152 	221	      225	     226
(function calling
overhead)

Variable sized	     52 	136	      398	     468
array runtime

alloca runtime	     31 	118	      118	     118

Own code size	   1040        1160	     1140

Library code size     0 	184	      788
@end example

@node libnix.a,libamiga.a,libstack.a,Features
@chapter Some ANSI (mis)features

I suppose you are familiar with C and especially ANSI C - if not
you should read a good book about it
@footnote{I recommend this one:

Brian W. Kernighan, Dennis M.Ritchie:
@*The C Programming Language (Second Edition)
@*Prentice Hall, Englewood Cliffs, 1988}.
This chapter only contains some special features of the
implementation - you should know these if you want to
use this library.

@menu
* Locale::
* Formatted I/O::
* atof strtod::
* Memory management::
* Standard I/O::
* Signal handling::
* setjmp longjmp::
* ctype::
* clock::
* Multibyte character functions::
@end menu

@node Locale,Formatted I/O,libnix.a,libnix.a
@chapter Locale

One feature of a complete ANSI compatible library is locale support. The
ANSI standard only knows of two locales:

@itemize @bullet
@item
"C" locale (normal C behaviour).
@item
Default locale.
@end itemize

Every other locale depends very heavily on the implementation.

To do locale support on the amiga I decided to use locale.library (what else).
This means that you normally have only these two locales - to have more than that
you must make some extra preferences files with the locale preferences editor
and give the path of these to the setlocale()-call. If you do not have locale.library
you will get only "C" locale. This is the default then :-(.

Another important point is that the ANSI standard requires the default locale to be
loaded at program startup. i.e. if you use german locale (for example) you
will just get it - printf and scanf will not work as expected but use the
decimal comma @samp{,} instead of the decimal point @samp{.} for their floating
point numbers, ctype functions will behave differently, too.

This can be very annoying if you don't want to use ANSI locale but rather
locale.library (which is not portable but IMHO much better) or if
you don't need locale support. And even dangerous if you don't
test your program under different locales.

To get around this problem I decided to do some nasty thing:
To get locale support you have to make up a reference to setlocale. You can do
this by just calling

setlocale(LC_ALL,"C");

immediately after program startup. (And get "C" locale then after program start which
is a much better choice). Or by just using setlocale anywhere in your program -
you will get default locale at program startup then.

@node Formatted I/O,atof strtod,Locale,libnix.a
@chapter Formatted I/O

The formatted I/O specifications are all there (remember: this library tries to
be ANSI compliant). But there are two things you should know about them:

@itemize @bullet
@item
The formatted I/O is affected by the setlocale() call - this is no bug,
just an ANSI feature.
@item
Half of the code of a full blown printf handles floating point numbers -
but not everybody needs them. So there are two functions for both
@samp{vfprintf} and @samp{vfscanf} - one in @samp{libnix.a} not including
floating point support and one in @samp{libm.a} including floating
point support.

So if you want to use one of the formatted I/O specifiers for floats
you should link with the math library @samp{-lm}.
@end itemize

@node atof strtod,Memory management,Formatted I/O,libnix.a
@chapter atof strtod

The two functions @samp{atof} and @samp{strtod} require a working
@samp{%f} specifier in @samp{vfscanf} - therefore they
require the formatted I/O functions in the math library.
Since @samp{libm.a} is linked before @samp{libnix.a} these
two functions have been gone into the math library.

@node Memory management,Standard I/O,atof strtod,libnix.a
@chapter Memory management

Most of the memory management of this library runs through malloc().
Only the commandline parser uses AllocVec() - so you can use it without
having the malloc function somewere in your program.

The memory management uses a local (to this task) memory pool to reduce
memory fragmentation. It uses the system functions to do so (not the
new pooled memory functions but just the older Allocate(), Deallocate() pair
which are the <3.0 fallback for libamiga.a's pooled memory functions, too)
so there should be no problems with it - these functions are tested very good.

The default blocksize for memory allocations is 16384 bytes - equivalent
to 4 MMU pages. Bigger allocations are blown up to a multiple of 4096 bytes.
So don't be alarmed if your program uses more memory then expected.

If you don't like this value (if you use bigger portions frequently or only
use very little memory) you can replace it by declaring

@code{unsigned long _MSTEP}

You should use a multiple of MMU pages. If you don't use a full MMU page
you gain nothing - malloc rounds up anyway.

@node Standard I/O,Signal handling,Memory management,libnix.a
@chapter Standard I/O - where stdin, stdout, stderr come from

2 of the 3 standard I/O streams are no real problem:

@itemize @bullet
@item
@samp{stdin} is set to the value the @samp{Input()} function of @samp{dos.library}
serves,
@item
@samp{stdout} is set to the @samp{Output()} value.
@end itemize

Both streams are managed by the OS and the library need not take much
care about them. But @samp{stderr} is a different thing since
there is no @samp{Errput()} ;-) function. So @samp{stderr} is handled
as follows:

@enumerate
@item
If @samp{process->pr_CES} is set, this value is taken. There are not
much shells that set this value so most of the time this leads to NULL.
@item
If this didn't work and your program was started from CLI the library
opens @samp{Open("*",MODE_NEWFILE)}.
This opens the last interactive terminal attached to stdout, i.e.
if you use the normal Amiga shell and redirect your output to a file you
get the terminal, if you redirect your output to @samp{NIL:} you get @samp{NIL:}.
@item
If this didn't work too (you never know) or your program was started from WB
you simply get the same stream as in @samp{stdout}.
@end enumerate

@node Signal handling,setjmp longjmp,Standard I/O,libnix.a
@chapter Signal handling

There is only support for the two signals SIGABRT and SIGINT. The library
knows of some other signals but cannot generate them. The support for
SIGABRT is simple - but SIGINT is a completely different thing:

You cannot use exec signal handlers since they are called at any time -
even in the middle of a library call. And if your library just blocked
a private semaphore and you jump out of the library code you will get a nice
deadlock :-(. (And for people who don't know: signal handlers are bogus
upto OS 2.0 (even there you need a good setpatch)).

So SIGINT (CTRL-C) is just polled at the start of most I/O-functions by
calling the function

@code{void __chkabort(void)}

Other signals are even more difficult to implement:

@itemize @bullet
@item
SIGSEGV simply doesn't exist - and if it does it's due
to a VM system and should not be generated.
@item
SIGFPE is not generated by the math libraries - so it would
be a bad thing to generate it by the mathematical coprocessor.
@item
SIGILL should never happen - your program must be faulty if you get one.
Most of the time this happens if you try to run a 68020+ compiled program
on a plain 68000.
@item
SIGTERM couldn't be disabled - even if it was there ;-).
@end itemize

You can disable CTRL-C handling by replacing @samp{__chkabort} with a do-nothing
stub function - but there is a better way. Just call

@code{signal(SIGINT,SIG_IGN)}

Replacing __chkabort is used very often by amiga-programs and if
your application does not need CTRL-C handling at all and is
amiga specific you can use this. The second method is the
ANSI standard method and works on all types of machines.

@node setjmp longjmp,ctype,Signal handling,libnix.a
@chapter setjmp, longjmp

This library is compatible to the header files that come with gcc -
and the jmp_buf in there is not large enough for the FPU registers.
So they are not restored! The ANSI standard doesn't even require to
restore any of the other local variables (they are restored :-) ),
so this is NO incompatibility to the ANSI standard.

@node ctype,clock,setjmp longjmp,libnix.a
@chapter ctype.h functions

If you look into ctype.h you will see that the functions in there are
just macros - and that they are duplicate in the library as functions.
This is NOT a mistake. The ANSI standard requires such macros to be
duplicate as functions.

And remember: These functions are affected by the setlocale() call.

@node clock,Multibyte character functions,ctype,libnix.a
@chapter The clock function

The clock() function's work is to measure processor time for the specific
task - but there is no information like this in the amiga OS :-(.
So it just measures the time from program start on - and is compatible with
this behaviour to all single tasking OSs around.

@node Multibyte character functions,,clock,libnix.a
@chapter Multibyte character functions

The multibyte character functions are all there - but since the
Amiga OS uses no other character set than ECMA Latin I they simulate
just "C" locale. This means they do nothing useful.

@node libstubs.a,detach.o,libamiga.a,Features
@chapter libstubs - automatic library opening

The Amiga OS shared libraries are a nice thing. All the tasks can
use them in parallel, they eat up memory only if you use them and
they are simple to use - and all this works even without a
memory management unit (MMU).

Another nice feature is the fact that you can open them under
program control, i.e. you can take some action if they do not
exist - warn the user, disable some features, etc. This nice
feature becomes a misfeature if you only need a certain list
of functions that are there all the time - exec, dos, intuition -
you still have to open the shared libraries.

So most Amiga compilers have a feature called automatic library
opening feature. This means that all libraries you reference
(by calling one of the functions) but don't open yourself get
opened for you by the compiler.

@menu
* Auto-library-opening usage::	   How to use it.
* Auto-library-opening interface:: How it works.
@end menu

@node Auto-library-opening usage,Auto-library-opening interface,libstubs.a,libstubs.a
@chapter Usage

To use this feature you have to do nothing (therefore it's called
automatic). But you can control the library version if you wish
by declaring

@code{long __oslibversion;}

somewhere in your program. But don't set this lower than 37 - most functions
of libnix (including the commandline parsers) need 37 or more.

@node Auto-library-opening interface,,Auto-library-opening usage,libstubs.a
@chapter Interface

Implementing such a feature is no hard work if you know how - this
implementation uses a (not so good known) feature of the gnu linker
called set elements:

@enumerate
@item
You write a library entry for every library base and link this library as the
last one. This means that the linker uses this library for every library base
pointer that is not defined but referenced somewhere.
@item
You tell the linker to collect these library bases together into a set element.
@item
You write a function that opens all libraries in the set element at
program start and cleans them up later.
@end enumerate

Some details:

There are two object files in the library for every library base pointer.
The first one is a

@example
struct lib
@{ struct Library *base;
  char *name; @};
@end example

containing the library base pointer (a @samp{NULL} pointer at program start)
and a pointer to the name of the library. This name

@samp{extern char name[]}

is the second object. All these structs are collected together into one
single set element called

@samp{extern struct lib *__LIB_LIST__[]}

To open and close the shared libraries there are two functions in
libstubs:

@code{void __initlibraries(void)}

and

@code{void __exitlibraries(void)}

Since it is still possible to open the shared libraries by hand I had
to take care about the library base pointers for libnix itself -
they are used in the commandline parsers - even before anybody could
open them. There exists a (library private) duplicate library base
pointer for each of these. They have normal names with
two underscores in front.

So don't be alarmed if some system monitor tells you that your program
opened dos.library twice - this is normal behaviour, most libraries
do this.

Opening libraries by hand works exactly the same way as on any other compiler:
@itemize @bullet
@item
You declare the library base variable somewhere globally:

@code{struct DosLibrary *DOSBase=NULL;}

The initialization @samp{=NULL} is necessary! Uninitialized variables
get overwritten by initialized ones in other object files - and
the library base pointers in @samp{libstubs.a} are initialized with @samp{NULL}.
This is a feature of the GNU ld and I cannot do much about it :(.
@item
You open the library before using it:

@code{DOSBase=(struct DosLibrary *)OpenLibrary("dos.library",37);}

and do some action if it fails :).
@end itemize

@node detach.o,,libstubs.a,Features
@chapter Detaching from the current CLI

Some people like multitasking that much that they tend to start everything
in the background. Some tools are able to do this automatically - and with
libnix you can write such tools, too. @footnote{ Please be aware that you
lose the ability to synchronize your tool with the calling CLI - this can be
very nasty if one needs to. Use this feature very sparse: Most of the time
it's better to just rely on the user being able to type @samp{Run >NIL: <NIL:}. }

To be able to detach from the current CLI the detach module has to know
how much stack your program needs, how to call the new process, etc.
Therefore you will have to provide some global variables that contain this
information. If you don't provide them you will get default values
(don't blame me for them - it's your own fault :-@} ).

Here is an example set of variables:

@example
char *__procname="My nifty tool";
long __priority=-1;	   /* We don't eat that much processor time */
unsigned long stack=50000; /* but need a large stack		    */
@end example

@node Special startups,Set elements,Features,Top
@chapter Special startups

As a serious Amiga programmer you may sooner or later want to write your
own shared library or device. This can be a very difficult task if you never
did it before. To make your life easier we did some of the work for you
if you decide to use one of our startups. Be aware that writing a shared library
is a task for the experienced programmer @footnote{ Most ANSI library functions
don't work out of a shared library. libnix makes no difference here - even simple
operations like multiplying two integers can fail - don't take anything for granted.}.

@menu
* Calling convention:: How to interface to the system.
* libinit.o::	       Shared library startup.
* libinitr.o::	       Shared library with a new data segment for each caller.
* devinit.o::	       Device startup.
@end menu

@node Calling convention,libinit.o,Special startups,Special startups
@chapter Calling convention

On the amiga almost all shared libraries are called with the library base
in a6 and other parameters in other registers. The result is placed in d0
usually.

If you want to write your own shared library you should stick to this model
to make it easier for others to interface with it - but unfortunately gcc
doesn't support registerized parameters. The solution to this problem is to
write an assembler wrapper for each function you need. Since all those
wrappers look the same it's easy to simplify their notation by using
preprocessor macros. You can find appropriate macros in the stabs.h file of
the libnix sources. Usage:

@example
/* Define some function that has 1 argument in d0 */
ADDTABL_1(__UserFunc,d0);
@end example

As a bonus these macros add your function to your library's jump vector.
All you have to do is to care for the right linkage order. And don't forget
to add a @samp{ADDTABL_END();} at the end of the vector.

Attention: Other programmers may decide to patch into your library's jump
vector - therefore it's a good idea to call even own functions over this
vector. To achieve this you have to provide some inline functions
(like those in the gnu:os-include/inline directory) or glue code and you have
to privatize the function's name by adding some '__'s or similar. You will
also have to set up your OWN library base.
@node libinit.o,libinitr.o,Calling convention,Special startups
@chapter Shared library startup

This startup gives you one data segment for all possible callers.
You will have to use semaphores to share special data between them.

To write a shared library you will have to provide some global variables

@example
const BYTE LibName[]="simple.library";
const BYTE LibIdString[]="version 1.0";
const UWORD LibVersion=1;
const UWORD LibRevision=0;
@end example

as well as some special functions
@footnote{It'll be possible to add an Open() and Close() function, too. But this
would be incompatible to libinitr.o and wouldn't give any advantages over this
method.}

@example
int __UserLibInit(struct Library *myLib);
void __UserLibCleanUp();
@end example

Please look into the examples directory for more details.

@node libinitr.o,devinit.o,libinit.o,Special startups
@chapter Shared library with different data segments

If you don't like the hassle with semaphores you can use this startup.
It provides a new data segment for each task that opens your library.
There are two disadvantages over libinit.o:

@itemize @bullet
@item
This method needs more memory.
@item
You cannot interact between your tasks.
@end itemize

The usage stays the same.

@node devinit.o,,libinitr.o,Special startups
@chapter Device startup

The device startup uses different names over the library startup
(though a device is always a shared library on the Amiga).

@example
const BYTE DevName[]="simple.device";
const BYTE DevIdString[]="version 1.0";
const UWORD DevVersion=1;
const UWORD DevRevision=0;
int __UserDevInit(struct Device *myDev);
void __UserDevCleanUp();
int __UserDevOpen(struct IORequest *iorq,ULONG unit,ULONG flags);
void __UserDevClose(struct IORequest *iorq);
(And some begin and abort function as well)
@end example

@node Set elements,geta4,Special startups,Top
@chapter Set elements - a nice feature of the gnu ld

Set elements are used very often by this library. Since most people don't
know them they are explained here a second time.

You can tell the linker to build up an array of pointers to every
global symbol in your program (functions or variables)
even if your symbols are scattered among some object files.
These arrays are called set elements.

You can take 4 Library base pointers

@samp{DOSBase}, @samp{IntuitionBase}, @samp{GfxBase}, @samp{IconBase}

tell the linker to put them together into a set element
called @samp{librarybases} by placing some assembler lines like

@code{asm(".stabs \"_librarybases\",24,0,0,_DOSBase")}

into your code (22 for text, 24 for data, 26 for bss - and don't
forget the single underscore) and get an array of pointers like this:

@example
void *librarybases[]=
@{ (void *)4,&DOSBase,&IntuitionBase,&GfxBase,&Iconbase,NULL @};
@end example

The first element contains the number of symbols. The last element
contains a NULL pointer. And remember: This are pointers to the
pointer variables.

This is the basis of global constructors and destructors in C++
and is very useful on the amiga to implement an auto-library-opening
feature :-). Set elements are used in this library for collecting
together library bases, initialization routines and cleanup routines.

@node geta4,Library bases,Set elements,Top
@chapter geta4 and other things - some words on code and data models

A program consists of two portions - code and data
(with the exception of self modifying code - you cannot get this
out of GCC and it is a bad thing to do - so forget about it).

A program usually accesses them in a unique style of addressing
modes for the machine instructions - called a code (or data) model.

On the amiga OS there exist two addressing styles for both of
them - code and data. With full 32 bit addresses - giving you access
to 4 GB of address space. And with reduced 16 bit addresses - giving
you access to only 32k of code and 64k of data. These styles are
called large (or normal) and small code and data models.
Usually small code comes together with small data - but that's not
necessary.

(Don't mix code and data models with the memory models of MS-DOS
machines: The memory in small data model is still flat 4 GB which
means by using pointers you can still address the whole memory.
Only the number of variables is limited.)

You may think that these limitations are a large disadvantage -
where are the benefits?

The benefits are simple: Code size and performance.

Every time you access a 16 bit address instead of 32 bit you spare
2 bytes in code size meaning 2 bytes in program size. And the
processor needs to load and process a smaller instruction that
needs less processor cycles. And since these 16 bit addresses are
relative the loader of the OS need not relocate them. Meaning
that you spare even 8 bytes more in executable size and some
loading time.

And there is another advantage: If all the data is addressed
relative it is simple to relocate it at program start - which
means that you can easily get multientrant and reentrant executables
(the code section is constant and need not be relocated).
You know these as pure=resident programs.

Some details:
@menu
* Data models::
* Code models::
@end menu

@node Data models,Code models,geta4,geta4
@chapter Data models - large and small

Let's take a simple C program:

@example
#include <stdio.h>

int max=100;
int count;
char string[]="Hello, world\n";

int main(void)
@{ int i;
  for(i=0;i<max;i++)
  @{ count++;
    printf("%s",string); @}
  return 0;
@}
@end example

If we look at it carefully we see 4 different types of data in it:

@enumerate
@item
The variable @samp{max} and the array @samp{string} - both are nonconstant
initialized global data.
@item
The variable @samp{count} - this is uninitialized (and therefore nonconstant)
global data.
@item
The string "%s" - this one is constant data.
@item
The variable @samp{i} - this one is local data.
@end enumerate

The compiler places these 4 types of data into 4 different places:

Data number 4 is local (and exists only in one function call).
The compiler places such data into registers if possible. On
the stack if this is not possible. If you do not want the compiler
to place data into registers declare it volatile - it will be
on the stack then all the time.

Data number 3 is constant - the compiler places it together with
all the other constant data into the code section (code is also
constant). Never change any constant data - the weirdest things
can happen.

Data number 2 is not initialized - it would be a bad thing to
put data without information into an executable. Therefore
such data goes into a special section - the BSS section. BSS data
does not increase executable size.

And the rest (number 1) goes into the data section :-).

To access the data section with machine instructions there are two
possible methods:

You can take the whole 32 bit address and store it into your machine
code. 32 bit means 4 byte every time you access a global variable.
This is know as the large (normal) data model because you can access
the whole bunch of 4 GB address space.

A lot of applications do not need such a large data section and it
would be a waste of memory to do so. So there exists a second
possibility:

You take one address register (a4) and use it as a pointer to your
data section. You access your data relative to this pointer with
16 bit references. This is known as small data model.

Since there are only 16 bit references you can access a total of
64k of data (32k in each direction from a4 on). And since you
use only one address register for this the data and BSS section get
merged together (BSS data still need not increase executable size -
there exist some tricks to prevent this - but not all linkers
support such tricks).

Beware: you should never lose the contents of your address register
(a4) or all the hell breaks lose.

If you ever lost them (this can only happen in certain cases when
using interrupts or hooks) you can restore them by calling
@samp{geta4()} or you can use no global data at all (and have no
problems then).

The second method is recommended - and it is possible sometimes
since the OS takes care of this and supports local data areas in
these nasty cases. But don't call any shared library - or you
will access (hidden) a library base pointer.

It is not possible to have a @samp{geta4()} function with
resident programs = multiple data sections
(which one would you choose? You access the data from a different task!)

It's in general not possible to mix objects compiled for
the two data models. There are some exceptions, but people that
know enough to prevent collosions need no explanation of when
it's possible ;-).

@node Code models,,Data models,geta4
@chapter Code models

All your constant data and all (constant) code form the code section.
To access the code section there exist two code models:

You can take the whole 32 bit address to call a function and can write
programs 4GB large. This is the large (normal) code model.

But you can even call your functions relative to the program counter
(pc) with 16 bit offsets. This is the small code model.
The advantages are the same as in the small data model - only the
disadvantages are different:

@itemize @bullet
@item
You can only have a total of 32k of code since you need to jump in
both directions. But even this is enough for a lot of programs.
@item
There is no address register that can be lost - the program counter
is valid the whole time.
@item
It is possible to mix large and small code model - but you can
hardly get more than 32k of code out of it.
@end itemize

@node Library bases,Code size,geta4,Top
@chapter Library base pointers - and how they work

The model of shared libraries on the amiga works as follows:

The library is not managed by the linker but by the application.
You open it through a system function. The advantage of this is clear:
You open libraries under program control, i.e. you can even check if
they are there and disable some features if not or take other action.

The result of this system function is a pointer to the upper end of a
jump table (a table of @samp{jmp} instructions to the different functions)
and the lower end of a library structure containing extra
information for every library you opened. These pointers are called
library base pointers. They are usually stored in normal global variables.

To call a system function you have to put the library base into
address register a6 and the parameters into certain other registers.
Then your program has to jump over the certain address of the jump table.
The function returns with the result in register d0 (and sometimes some more).

A compiler can handle this behaviour by two different methods:

@itemize @bullet
@item
It can just do the right thing and place everything into the desired
registers - gcc does this by declaring special assembler inline
functions that do the job.
@item
It can put the arguments on the stack (as in every normal function call)
and call a glue function that does nothing else then taking the arguments
from the stack and putting them into the right registers then calling
the function. This glue code is contained in @samp{amiga.lib}.
@end itemize

Both methods require to access the library base pointer
(and a valid value in it) so they make up a reference to
this variable.

@node libamiga.a,libstubs.a,libnix.a,Features
@chapter Glue code and some other things

It's not possible for me to redistribute amiga.lib - but you should
have one if you really want to use the possibilities of your amiga.
(You don't need one if you only want to use ANSI features or if you
use the inline headers of gcc)
If you want to compile resident programs you will need a baserelative
version too.

To solve this problem I decided to build a selfmade version of
libamiga.a. The gluecode of this library is built out of the
inline header files of gcc, some of the other functions are written
from scratch. This does not give you a complete version of libamiga.a
but a better that nothing version including sources.
To rebuild it unpack the sources, then type a 'make libamiga'.

If you want to have a fully functional version of libamiga.a you
can use the real one. To do this you will
have to convert normal amiga objectfile format to a.out format
(known by the linker). Type:

@example
cd <some empty directory>
stack 300000
sh
Hunk2GCC <path>amiga.lib
ar -q libamiga.a obj*
rm *.o
ranlib libamiga.a
exit
@end example

Doing this on @samp{RAM:} will improve performance a lot.

This doesn't give you the baserelative version blib/libamiga.a -
you will be unable to compile resident programs. To get a baserelative
version of amiga.lib try to get the @samp{libtos} program of the
@samp{DICE} compiler of M. Dillon (from fishdisk or somewhere else) -
it converts libraries to baserelative ones:

@example
cd <some empty directory>
lha x amigalibdisk491:dice/dice206_21.lzh #?/libtos
netdcc/bin/libtos <path>amiga.lib amigas.lib
@end example

Then do the same as above.

@node Code size,FAQs,Library bases,Top
@chapter Writing small programs

Writing very small programs is a trivial thing - if you know how to do it.
Here are some basic tips and tricks:

@table @asis
@item Do not use printf:
@samp{printf} is the most high-level function of
usual ANSI C libraries and builds on almost everything else. Often this function
can be replaced by simpler routines like e.g. @samp{puts}. The same is true
for the other formatted I/O functions.

If you write amiga-only programs you can use the @samp{sprintf} routine of
libamiga.a by linking with @samp{-lamiga}.

@item Try to avoid using buffered I/O:
A lot of programs can be written using the low-level I/O-functions @samp{read} and
@samp{write}. This saves the code responsible for buffered I/O.

@item Avoid -O3:
@samp{-O3} activates all those optimizations that sacrifice code size for speed.
An exception is @samp{-fstrength-reduce} which costs compile time but not
code size. @samp{-O2 -fstrength-reduce} will give small programs.

@item Strip your executables:
gcc doesn't strip executables by default. Setting @samp{-s} will give worse
debugable but small programs.

@item Use small code and data model:
@samp{-msmall-code -fbaserel} do this. Be aware that this won't work for
larger programs.

@item Don't use libgcc.a:
@samp{libgcc.a} which calls global C++ constructors and destructors builds
on @samp{atexit} which builds on @samp{malloc}, ... If your program doesn't
use them you can save some memory by linking by hand (@samp{-nostdlib}).

@item Set __nocommandline:
If you use @samp{ReadArgs} instead of @samp{argc, argv} you can save some
memory by setting a global variable named @samp{__nocommandline} to drop
parsing of commandline arguments.
@end table

@node FAQs,,Code size,Top
@chapter FAQs

@table @asis
@item Q:
I do not get a working executable out of it - my debugger
tells me the library bases are broken.
@item A:
The GNU ld that comes with GCC 2.5.8 (or lower) has some
serious bugs in conjunction with set elements.
Use the fixed version of ld that comes with gcc 2.6.0 (or above).

@item Q:
There are some prototypes missing in stdio.h.
@item A:
This stdio.h is only for internal use - use the normal GCC
stdio.h to compile your programs.

@item Q:
While printing floats printf prints weird characters.
@item A:
This problem should be fixed with the current release of libnix. It was caused
by a bug in the system math libraries which happens if you open them in the
wrong order. You can use the SetMathPatch program by Andreas Wolff to fix this
and another more serious bug with mc68040 processors.
@end table
@printindex cp
@contents
@bye