This is the DJGPP Frequently-Asked Questions List.





Copyright © 1994, 1995, 1996, 1997, 1998, 2000 Eli Zaretskii.

This is the third edition of the FAQ list;
it is consistent with version 2.03 of DJGPP.

This FAQ list may be freely distributed with the DJGPP package or any part thereof, provided that this copyright notice is left intact on all copies.

Table of Contents

Node: Top, Next: 

DJGPP FAQ List

In DJGPP (see DJGPP overview), a 32-bit compiler and programming environment, originally written for Unix machines, meet a 16-bit MS-DOS operating system. Programmers who work in this environment have to master a large body of knowledge from both Unix and MS-DOS, especially if they want to use some advanced features, like interrupt handling, directly accessing peripheral devices, etc.

But because the DJGPP project is run by a group of volunteers on their free time, there isn't always enough time (or patience, or money ;-) to produce documentation which will describe all the subtle features and pitfalls a user should know about. The documentation of DJGPP-specific utilities and features is minimal at times, leaving wide space for confusion, in newcomers and veterans alike, and making the DJGPP learning curve steeper than it could be.

This FAQ list is an attempt to take the sting out of that learning curve, by supplying solutions for problems which are known to puzzle DJGPP users. (Another solution would be to pay to DJ Delorie and other people who develop DJGPP to produce more documentation ;-).

Some additional places to look for tutorials and other introductory material about DJGPP are listed below.

One good place to look for DJGPP features that are often overlooked is the DJGPP Knowledge Base. The Knowledge Base is also available in Info format; type info knowledge from the DOS prompt. In addition, a User's Guide is being written by several contributors; it is currently available from the DJGPP server.

You can browse the HTML version of this FAQ list on line at DJ Delorie's Web server. You can also download FAQ in several additional formats and browse it locally.

A previous version of this FAQ was translated into French, and is available via FTP, and also through the Web.

A Japanese translation is available via the Web.

This is Edition 2.30 of the FAQ, last updated 2 February 2000, for DJGPP Version 2.03.

The following master menu lists the major topics in this FAQ list, including all the indices.

Node: Urgent, Next: , Previous: Top, Up: Top

1 If You Are In a Hurry

Q: Do you really mean I have to read this looongish FAQ list to get my answers?

Q: I have this problem which I absolutely MUST solve NOW! What do I do?

A: No, you don't need to read all of the FAQ unless you want to (although this is by all means recommended). The questions in this document are listed, as much as possible, in the order they appear when one goes through getting DJGPP, installing it and using it. To quickly find an answer to your question, first look at the Table of Contents. If that doesn't help, try the indices at the end of this manual. You can look up your question either by program name, or by topic name.

If you don't find anything appropriate through the indices, search this FAQ for words which are pertinent to your problem1.

If searching the FAQ didn't help, try the DJGPP archives search, where you can find reports about similar problems and their solutions. If that doesn't help either, try asking the DJGPP gurus.

For those in a real hurry, here are some pointers to the most important topics in this FAQ list:

Node: DJGPP, Next: , Previous: Urgent, Up: Top

2 What is DJGPP?

Q: What is DJGPP?

A: DJGPP is a compiler and a set of tools that let you produce 32-bit protected-mode programs which run on MS-DOS/MS-Windows machines.

The originator and principal maintainer of DJGPP is DJ Delorie; that's where the "DJ" in "DJGPP" comes from. However, anybody is welcome and encouraged to contribute.

Programs compiled with DJGPP, and all development tools provided as part of DJGPP, look on the outside like normal DOS programs, and they rely on MS-DOS and BIOS for file I/O and other basic functions such as keyboard input, screen cursor position, etc. However, the bulk of the code in a DJGPP program is 32-bit protected-mode code; DJGPP programs use DPMI (the DOS Protected Mode Interface) to allow DOS/BIOS calls from protected mode. Therefore, any environment that can run DOS programs and provides DPMI services, will run DJGPP programs as well. Environments that are known to be compatible with DJGPP include MS-DOS, Caldera's DR-DOS, NWDOS, FreeDOS, Windows 3.X, 9X and NT, OS/2, and Linux DOSEmu. When DJGPP programs run on Windows 9X and Caldera's DR-DOS, they support long filenames.

It is important to understand that all these environments will treat DJGPP programs as DOS programs which call DPMI services. DJGPP cannot by itself create Win16 or Win32 applications; however, you can use the RSXNT package together with DJGPP to achieve this. See writing Windows applications with DJGPP.

Programs compiled with DJGPP can access all the physical memory on your machine and support virtual memory. All this memory presents a flat address space with no segmentation (you can say goodbye to far and huge pointers and to memory models), and is only limited by the amount of virtual memory supported by the DPMI server in use. A typical DPMI server can provide at least 64MB of virtual memory (if you have enough free disk space).

DJGPP is free: you don't have to pay anything to download and use it, even if you write commercial programs. DJGPP doesn't impose any restrictions on programs that you write and compile with it: you can make them commercial, shareware, freeware, or any other kind. (There are a few minor exceptions to that rule, see (un)restrictions on distribution of DJGPP apps.)

The core of DJGPP is the MS-DOS port of the GNU C/C++ compiler, GCC, and auxiliary utilities, such as assembler, linker, librarian, Make, and a hypertext docs browser. The DJGPP C library was written specifically for DJGPP, mainly by DJ Delorie himself, with help from a small group of volunteers. This core set of utilities and libraries is still actively developed and maintained.

DJGPP presents a set of tools which are remarkably ANSI- and Posix-compliant2. GCC complies to ANSI/ISO C Standard; the DJGPP C library is ANSI- and Posix-compliant (however, a small number of Posix features, like the fork system call, are unimplemented); the C++ libraries also comply to the latest standards; and the GNU development tools used by DJGPP are all Posix-compliant. As a result, DJGPP tools provide a complete and coherent Posix layer on top of Microsoft operating systems, to the degree that even the infamous limitations of DOS and incompatibilities between DOS/Windows and Unix are almost completely concealed from users and developers.

Here are some of the tasks that DJGPP is said to be good for:

DJGPP is also used as back-end for programming languages other than C/C++. ADA, Pascal and Fortran compilers have been ported to MS-DOS based on DJGPP; GNU Pascal (gpc) and GNU Fortran (g77) are available from the DJGPP archives. The latest GCC releases include front ends for Java and Chill languages.

Starting from v2.0, DJGPP programs do not need a separate extender program, only a DPMI server to run; DJGPP includes a free 32-bit DPMI server which allows for a 32-bit, 4 GByte flat address space and up to 512 MBytes of virtual memory on plain DOS machines taht lack a DPMI server of their own.

Node: Requirements, Next: , Previous: DJGPP, Up: Top

3 Hardware and Software Requirements

This chapter describes what are the hardware and software which will allow you to use DJGPP. Minimum, "reasonable" and optimal system configurations are listed.

Node: Minimum, Next: , Previous: Requirements, Up: Requirements

3.1 The minimum system requirements for using DJGPP

Q: What are the minimum system requirements for using DJGPP?

Q: Will DJGPP run on my brand-new Acme i986DX7/900 PC with a SCSI-III 10-Terabyte disk drive under MulticOS/42 v7.99 operating system?

A: DJGPP requires at least 386SX CPU and between 15 and 35 MB of free disk space (see more details on this below), including space for the software installation and some swap space. A minimum of 64K of free system memory is enough for DJGPP to run with CWSDPMI as your DPMI host (most other DPMI hosts will require much more), but at least 4MB of free extended RAM is recommended for reasonably fast compilation of large source files (8MB for compiling large C++ programs); you might see painfully slow compiles for large sources if you don't have at least that much. If your machine doesn't have a numeric co-processor, you will need to install an emulator to run floating-point code (DJGPP provides such an emulator) or link your applications with a special emulator library (also provided with DJGPP).

DJGPP requires MS-DOS version 3.1 or later; any other operating system is OK if it includes a DPMI server and supports some kind of "DOS box". Environments known to run DJGPP besides native DOS: Windows 3.1 & 3.11 DOS box, OS/2 (including Warp) DOS box, Windows 9X/DOS 7, Windows NT (on Intel CPUs), Novell NWDOS 7, FreeDOS and Caldera's DR-DOS (but several people have found the DPMI services of NWDOS and early versions of Caldera's DR-DOS to be incompatible with DJGPP, so they might need to be turned off and CWSDPMI used instead), and Linux DOSEmu environment.

Note that somebody reported that running Caldera's DR-DOS 7.03 with TaskManager enabled causes redirection of standard error stream to not work. In particular, RHIDE cannot display the compilation errors printed by the compiler. A work-around is to turn TaskManager off when compiling.

One user of Caldera's DR-DOS found that using their virtual disk drive, VDISK.SYS, caused strange crashes in DJGPP programs, unless the memory manager is Caldera's EMM386 (as opposed to HIMEM).

Other known problems with latest versions of Caldera's DR-DOS are that Ctrl-<C> and Ctrl-<BREAK> crash the system or require cold reboot; and there are some problems with programs that use the linear frame buffer under VESA 2, and with certain games.

Node: OS2, Next: , Previous: Minimum, Up: Requirements

3.2 Does it really work under OS/2?

Q: You tell me it will work under OS/2, but I'm experiencing strange crashes after several compilations ....

Q: DJGPP Make crashes when I run it on OS/2!

Q: When I press Ctrl-<C>, my program is aborted, even though I've set up a handler for SIGINT!

A: There was a bug in the DPMI server of the old OS/2 versions, which was triggered by spawning child processes (like GCC does when it invokes the various compiler passes). Current versions of OS/2 don't have that bug, so DJGPP programs should run fine under OS/2. If you can't make this happen, chances are that your setup is incorrect. One system parameter that can cause problems with DJGPP (reportedly, Make crashes if it isn't set correctly) is DPMI_DOS_API. Setting it to ENABLED instead of the default AUTO should solve the problem. I'm also told that experimenting with the value of DPMI_MEMORY_LIMIT sometimes solves problems on OS/2. Reportedly, version 4.0 of OS/2 solves problems with DPMI support, so the above is only required for OS/2 v3.0 or earlier.

One particular problem with OS/2 v3.0 is that RHIDE 1.4 and later exits after the compilation ends. This doesn't happen under OS/2 v4.0, so you should upgrade if you have such problems.

Problems with SIGINT are due to a known bug in OS/2 VDM: when you press Ctrl-<C>, VDM catches it ahead of the keyboard interrupt handler installed by the DJGPP startup code, and aborts the program. Thus, you cannot prevent your programs from being aborted by installing a SIGINT handler.

Programs that use the library function delay hang if the <Pause> key is pressed while inside the call to delay. It is possible that versions of OS/2 4.0 fixpack 13 or later will correct this. As a work-around, use usleep or write a loop which calls uclock.

Node: WindowsNT, Next: , Previous: OS2, Up: Requirements

3.3 Will DJGPP work on Windows/NT?

Q: What about Windows NT?

A: Current Windows NT versions support DPMI programs in the DOS box, so DJGPP programs should in general run fine under NT (but see the list of possible problems below).

The DPMI server built into NT (and Windows 9X) loses selectors with each child program that is invoked by a DJGPP program, so after about two thousand calls to functions from the spawnXX family you can see an error message like this: 

  Load error: no DPMI selectors

Some versions of NT lose DOS memory instead of selectors, so you might see another error message: 

  Load error: no DOS memory

These problems are likely to afflict only DJGPP ports of Unix shells (such as bash), since no other DJGPP program, not even Make, is likely to call so many child programs before it exits. The only known work-around is to exit the shell every now and then, because when all the available selectors are exhausted, the DOS box will crash. I'm told that Make sometimes fails on long Makefiles on Windows 9X, where the selectors are lost at even higher rate than on NT. If you ever run a very long Makefile and see Make crash, just run Make again, and it will pick up where the crashed session has left off.

The long filename API (the special functions of Int 21h which support file names longer than the DOS 8+3 limitation) for DOS box is not supported by current versions of Windows/NT, so you cannot have long filenames there from DJGPP programs. An alpha version of an LFN driver for NT which enables long file name support for DJGPP programs, written by Andrew Crabtree and significantly improved by Wojciech Galazka, can be downloaded the Web.

The popular DJGPP IDE RHIDE needs a -M switch to work on NT (to disable the mouse support which will otherwise crash RHIDE). Version 1.4.7b or RHIDE reportedly solves this problem and allows the mouse to be used on NT. Also, one user reported that he had to type rhide twice to enter RHIDE, because the first invocation immediately exits back to the command-line prompt with no message, if you don't disable the mouse with -M.

You might have problems with using the SVGA modes of your video card under Windows/NT; only standard VGA modes (including mode-X) work. That is because NT doesn't allow arbitrary direct access to the SVGA registers, without which it is impossible to recognize the type of the SVGA and employ its capabilities. For example, a user reported that GRX functions and the MODETEST.EXE program thought that only a standard VGA was installed, whereas he had an S3 card. There is nothing you can do about this feature of Windows/NT; that is the price you pay for the stability and protection you get under this OS (a runaway program that accesses hardware registers can wipe out your disk or wedge the entire system cold). However, I'm told that Windows/NT 4.0 supports DirectX which is a method of accessing screen, audio and other peripherals directly, and the Win32 ports of Allegro and other graphics packages can use it.

The Allegro library also has problems on NT. One user reports that even switching into the standard 640x480 video mode turns the screen black and kills the machine. Programs that use Allegro to switch into VESA modes usually don't work, since NT doesn't support SVGA graphics modes. In particular, the example programs provided with Allegro print an error message like this: 

 Error setting 24 bit graphics mode
 VESA not available

Programs that use the "nearptr" facility of DJGPP to access absolute memory addresses (e.g., for memory-mapped devices) won't work on NT, because its DPMI server silently ignores functions that set huge limits on selectors. Since the default algorithm which allocates memory from the DPMI server needs to set such huge limit in some rare cases, there's a small probability that a program will fail or crash even if it doesn't set selector limits in user code. It is best to use the Unix-style sbrk algorithm in programs that run on Windows/NT. See the library docs for the variable _crt0_startup_flags where the _CRT0_FLAG_UNIX_SBRK bit is explained, for more info on this issue. If you cannot switch to the Unixy sbrk (e.g., if you don't have access to the program's sources), I'm told that sometimes such problems can be worked around if you run DJGPP programs in a full-screen session; your mileage may vary.

Another problem on NT is that you cannot install a handler for the SIGFPE, SIGINT, or SIGALRM signals: if you do, your program will crash as soon as the signal is generated (in DJGPP v2.02 and later, FP exceptions are masked by default, so you will need to unmask them first, otherwise SIGFPE won't be generated). This is due to a bug in NT.

Windows/NT makes it impossible to use FP emulation on a machine that has the FP hardware. If you set 387=n on NT, the DJGPP startup code calls the DPMI function to switch on the FP emulation, but NT ignores it and continues to use the hardware FPU.

Yet another problem with NT is that interrupting some programs with Ctrl-<C> causes Dr. Watson to complain about "Access Violation" (that's NT'ese for GPF) and abort the program; careful inspection of Dr. Watson's logfile seems to indicate that the crash is inside NT's own code which handles the exception deliberately produced by the DJGPP's machinery that translates the Ctrl-<C> keypress into a signal. It seems NT uses the DJGPP stack for some of that processing, which is a no-no inside an exception handler. Sorry, no work-around.

The above-mentioned problems with signals are probably the cause for another type of calamities on Windows/NT: running a program compiled with the -pg option causes it to crash almost immediately due to--you guessed it--"Access Violation" in NTVDM (that's the NT DOS Emulator).

Windows/NT comes with its own version of redir.exe, which serves a different purpose. If you invoke redir, and the NT's winnt\system32 directory is before DJGPP's bin directory in your PATH, you will see a message saying "The VDM Redirector is already loaded". To solve this, rearrange your PATH or rename DJGPP's redir.exe to somethink like djredir.exe.

Programs that use the library function delay may hang if the <Pause> key is pressed while the program is inside the call to delay. The work-around is to use usleep or write a loop which calls uclock.

Another peculiarity of the NT DOS box is that beeping by printing the \007 character to stdout or stderr behaves strangely. Usually it beeps, but the beep is very long; sometimes, you get the Windows "ding" sound. It is recommended that you turn on the "visible bell" feature of the tools that support it, like Emacs and Less.

Accessing the serial communication ports on NT also has some problems. Anton Helm says that the first two invocations of a program that accesses the port behave abnormally; e.g., the data from the device on the other end of the link doesn't get fed into your program. After that, the third and subsequent invocations work correctly, but only if you use COMAND.COM as your shell. Using the default cmd.exe leaves the link in a state where you get the replies from the other device for the nth invocation of your program in the n+1st invocation.

In other words, to make the com port work on NT, you need to open the DOS box with COMMAND.COM, run your program and exit it two times, then invoke the program for the third time and start working.

Some people report that NT servers cause much more problems than NT workstations of the same version and build. It seems that these problems usually mean that NT installation was done incorrectly (maybe it is much harder to get it right with a server than with a workstation?). If you have such problems, try to install a workstation, or re-install the server, and see if that helps. And if you gain some insight as to why servers like DJGPP less than workstations, please tell what you've learned.

The Cygnus Win32 project is another (unrelated to DJGPP) port of GCC and development tools to Windows/NT and Windows 9X platforms, which specifically targets development of Windows programs. See description of the Cygwin project, for more details about the Cygnus ports.

Node: DOSEmu, Next: , Previous: WindowsNT, Up: Requirements

3.4 Can it run under Linux?

Q: You say it works on Linux, but I seem to be unable to run the compiler from within Make....

Q: I can run DJGPP on Linux, but Make crashes with SIGFPE on even the simplest Makefiles!

Q: When I run bash on Linux/DOSEmu, echoing of what I type is very slow.

A: Versions of Linux which were released before 13 March 1996 need a patch to be able to reliably run nested DJGPP programs. That patch was posted to the DJGPP mailing list and can be found by using the search capabilities of the DJGPP mail archives.

If you prefer to download that patch via ftp, you can find it on the DJGPP ftp server.

In general, upgrading to DOSEmu version 0.97.10 or later is recommended, at least with versions of Linux kernel earlier than 2.1; in particular, some users report that DJGPP programs sometimes crash on version 0.66.7 under Linux 2.0.35.

You might also need to edit the RAM section of the /etc/dosemu.conf file to make it comfortable for DJGPP. I suggest setting dpmi and xms to 16MB and ems to 4MB. For example, I'm told that building the Allegro library with the -O3 optimization switch fails in DOSEmu unless you allocate at least 16MB of DPMI memory to DOSEmu sessions, and building GCC needs 18MB.

If DJGPP programs crash with an error message like this4: 

 DPMI: Unhandled Execption 0d - Terminating Client
 It is likely that dosemu is unstable now and should be rebooted

then you should add a line saying secure off to your /etc/dosemu.conf file.

Some users reported that Make, and possibly other programs which use floating point computations, crash in DOSEmu environment on systems without an FPU, even if you set the 387 and EMU387 environment variables correctly (as explained in Setting up the FP emulator, below). The only known work-around is to not use floating point or to upgrade your machine hardware. DJGPP v2.03 corrected a few subtle bugs in the emulator code, so upgrading your DJGPP software might help. It is possible that newer versions of Linux might solve this problem too, so try upgrading your Linux software.

If your only problem is to run GNU Make, get the latest DJGPP port of Make, since ports of Make 3.75 or later can be configured to not issue FP instructions at all.

If you have problems running recursive Make's, or deeply nested DJGPP programs, edit src/dosext/dpmi/dpmi.h to enlarge the value of DPMI_MAX_CLIENTS (the default is 8) and then rebuild DOSEmu.

If DJGPP programs respond too slow to keyboard input, you might need to tune the HogThreshold parameter in the dosemu.conf file. Set it to zero and see if this helps; if so, further tune it until you get reasonable response time, but still leave Linux with enough cycles for the other programs that run.

Several users reported that DJGPP programs cannot get input from the keyboard if Caldera's DR-DOS is booted under DOSEmu. I'm told that adding rawkeyboard to dosemu.conf might solve this.

Some people complain that RHIDE crashes on DOSEmu whenever the mouse is moved. I'm told that using the -M switch when invoking RHIDE solves this problem. Alternatively, you could try giving DOSEmu access to the serial port to which the mouse is connected, and then using your DOS mouse driver. To this end, add the following to your dosemu.conf5: 

 serial { mouse com 2 device /dev/mouse }
 mouse {mousesystems device /dev/mouse emulate3buttons }

and then load a DOS mouse driver in the DOSEmu AUTOEXEC.BAT. Note that the example above assumes that the mouse is connected to the COM2 port; your mileage may vary.

If you have problems with mounting FAT32 partitions, upgrade the Linux kernel to version 2.0.34 or later.

I'm told that the problem with selectors being lost in nested DJGPP programs (see no DPMI selectors) exists in DOSEmu as well.

Node: i286, Next: , Previous: DOSEmu, Up: Requirements

3.5 Can I run it on a 286?

Q: Why can't I run DJGPP on my 286? It has protected mode also....

A: True, but the protected mode isn't an issue here. Gcc doesn't care much about memory protection, but it does care to run on a 32-bit processor, which the 286 isn't. A 386 or better CPU really is required.

Node: Windows apps, Next: , Previous: i286, Up: Requirements

3.6 MS-Windows applications and DJGPP

Q: Can I write MS-Windows applications with DJGPP?

A: Currently, you can only run DJGPP programs under Windows as DOS apps (i.e. inside the DOS Box). If you need to write true Windows apps, you will have to use auxiliary tools or another compiler. This section lists some of the possibilities.

RSXNTDJ is an add-on package for DJGPP which allows to develop Win32 programs using DJGPP development environment. It is essentially a cross-compiler targeted for Win32 (Windows 9X and NT) and Win32s (Windows 3.X + Win32s) platforms6; it supports DJGPP v2.x and includes debugging tools and an IDE. Beginning with version 1.60, RSXNTDJ is distributed under the terms of the GNU license (previous versions needed to be registered for a small fee if you wanted to develop commercial or shareware applications with it).

RSXNTDJ supports Win32 console, GUI, DLLs and bound programs (the latter can be run on DOS under the RSX extender, as well as on Windows). You can download RSXNTDJ via FTP. The latest version of the RSX IDE (called RSXIDE) is available from Rainer Schnitker's home page. Rainer's home page is also the place to look for the latest releases of RSXNTDJ, including beta releases.

RSXNTDJ version 1.31 was produced with GCC v2.7.2.1 and DJGPP v2.01. If you use it with later versions of GCC and DJGPP, it will need some tweaking; it is best to upgrade to RSXNTDJ v1.5 or later. Johan Venter wrote a HOWTO document that explains how to install and set up RSXNTDJ, and how to fix the various problems present in the current RSXNTDJ distribution. Make sure you read the RSXNTDJ HOWTO, before you try to install the package on your own; it will save you hours of banging your head against all kinds of weird problems, like missing files (due to truncated file names), linker error messages, inability to link C++ programs, etc.

In general, RSXNTDJ's support leaves a lot to be desired. Some problems take too long to fix, and response to user bug reports is slow. Even with the latest efforts by Johan Venter and others, you should expect to spend some time and effort installing and using the package. Another problem is that you depend on Microsoft SDK for the header files, and each new release of the SDK tends to break the patches to the header files required by RSXNTDJ.

These are disadvantages of RSXNTDJ with respect to other alternatives (see below); the most significant advantage is that you can use the entire DJGPP development environment. By contrast, other alternatives for free Win32 development usually provide less tools; in particular, Mingw32 provides only a few basic tools (like Make and GDB) beyond the compiler and Binutils. As a result, people who work with Mingw32 tend to use all kinds of different and subtly incompatible versions of shells, Make, etc. Cygwin has much more ports of GNU tools, but even Cygwin tool-chain doesn't have such a rich set of development tools all working together smoothly as DJGPP does. (Of course, if you don't mind developing with a minimal set of tools, this might not be a serious consideration for you.)

Here are some tips about RSXNTDJ:

If RSXNTDJ doesn't suit your needs, you can use a few other free compilers which target Win32 platforms:

Cygnus GNU-Win32 tools
This tool-chain includes native Win32 ports of GCC and of many GNU development tools. It requires you to comply to the GNU License, the GPL, when distributing programs built with these tools. The tools and the programs you build are native Win32 executables (won't run on DOS, Windows 3.X or Win32s platforms) and Posix-compliant, but you need to distribute a 4MB DLL file7, which implements the Posix layer, with all your programs. Also, GNU-Win32 is still in beta phase, and some bugs are still worked on. You can find GNU-Win32 on the Cygnus site, or via FTP.
Mingw32 (Minimal GNU-Win32)
This features native Win32 port of GCC, but it relies on the Windows C runtime (CRTDLL.DLL, which is standard on Windows 9X and Windows/NT) and doesn't require any additional DLLs like Cygnus ports do; however, you lose the Posix layer. The basic package includes, besides the compiler and Binutils, a few tools to compile resource files and convert DLLs into lib*.a wrappers. Since it doesn't use any GPL'ed stuff except GCC and its subprograms, the programs produced by Mingw32 are free.

A disadvantage of this package is a relative lack of development tools ported to Mingw32. The compiler, Binutils, Make, GDB, Textutils and Patch are available from the Mingw32 site, but ports of other utilities tend to be scattered around and not integrated together into a coherent package like what DJGPP or Cygwin present. This causes compatibility problems between the tools. It is possible to use the DJGPP tools where there are no equivalent Mingw32 ones, but you need to be aware of some incompatibilities, such as different methods of passing long command lines, lack of support for long file names on NT, etc.

More details, including ready binaries of ported utilities and source-level patches to build other utilities with Mingw32, are available on Mingw32 home page. For ports of additional developemnt tools, visit Jan-Jaap van der Heijden's site. Mingw32 has a mailing list.

Lcc-Win32 compiler and tools
This is a Win32 port of a freeware compiler Lcc, not related to GCC. It doesn't currently support C++ programs. The tool-chain includes some additional utilities such as a very good IDE, a resource compiler and a resource browser, a Make utility, and an icon maker. The package documentation is very good. For more information, visit the Lcc home page and the lcc-win32 home page.
Dev-C++ package
This is a freeware compiler and development environment for C and C++ programs which produces Win32 executables. Besides the compiler, the package includes a set of header files and libraries, a port of GDB, an IDE with a multi-window editor that supports syntax highlighting, and a project management tool. The package is maintained by Colin Laplace, and is available from SimTel.NET mirrors.

If you need on-line documentation of the Win32 API, you can find it as a Windows HLP file. Additional links to tutorials and other related information can be found on the bowman's home page.

The recommended book for learning Win32 programming seems to be Charles Petzold's Programming Windows: The Definitive Guide to the Win32 API, published by Microsoft Press.

Node: Optimal hardware, Next: , Previous: Windows apps, Up: Requirements

3.7 Machine you would like to buy...

Q: What is the optimal system configuration for running DJGPP?

A: Here is the description of your dream machine (at least for the next 6 months :-):

Node: Reasonable hardware, Next: , Previous: Optimal hardware, Up: Requirements

3.8 Machine most of us will actually buy ...

Q: OK, I don't have this much money. What is the reasonable configuration?

A: If you have the following machine, you should be able to stop worrying about memory and compilation performance:

This will leave you with about 19 MBytes of free extended RAM. Note that the RAM disk must be at least 4 MBytes to hold the output of the preprocessor for some exceedingly large source files (notably, some GCC source files). If you don't have that much RAM to spare and still want to compile very large source files, either reduce the disk cache so you can give more to RAM disk, or point TMPDIR to your hard disk and make the disk cache larger, if you can.

Node: Config, Next: , Previous: Reasonable hardware, Up: Requirements

3.9 How to configure your system for DJGPP

Q: How do I configure my system to get optimal performance under DJGPP?

A: That depends on the amount of RAM you have installed in your machine. Below are some guidelines to help you.

  1. If you have 2 MBytes or less RAM installed:

    With this configuration, GCC will run out of free physical RAM and start paging when compiling almost any C program and all C++ programs. If you are serious about DJGPP development, you need to buy more RAM urgently.

  2. If you have 2-4 MBytes of RAM installed:

    With this configuration, GCC will still run out of free physical RAM and start paging when compiling large C programs and most C++ programs. Plan to buy more RAM as soon as you can.

  3. If you have 5-8 MBytes of RAM installed:
  4. If you have more than 8 MBytes of RAM:

Some people disable the delayed-write feature for safety reasons, to avoid losing files due to system crashes. If you are worried about this, you can usually gain performance without sacrificing safety by enabling delayed-write together with an option that causes the cache to flush the write-behind data before the system returns to the DOS prompt. For a SmartDrv disk cache, this is achieved by specifying /N/F switches instead of /X.

Using a memory manager, such as EMM386 or QEMM, is not required (DJGPP will run without it), but highly recommended, since it has several advantages:

If your memory manager is EMM386, I recommend to put the NOEMS NOVCPI parameters on its command line. This will allow you to use UMBs and up to 128MB of physical memory (if you have that much installed). Without these parameters, many versions of EMM386 limit your physical memory to 32MB.

It is generally not recommended to install DJGPP on a networked drive, since this makes it slower, particularly when linking programs. If you do install DJGPP on a networked drive, you should consult your network administrator to configure the network for maximum performance. For Novell networks, a good place to look for advice is the Novell FAQ (search for a file called nov-faq.htm).

Node: More than 64MB, Previous: Config, Up: Requirements

3.10 How to get the most RAM for DJGPP programs?

Q: How do I set my system so that DJGPP programs could use all of my 256MB of installed physical RAM?

Q: I have 128MB of memory installed, but go32-v2 only reports 32MB, how can I get more?

Q: You say that CWSDPMI supports up to 512MB of total virtual memory, but I cannot get more than 128MB!

A: You can have as much as 256MB of physical memory in DJGPP programs, provided that you have at least that much installed, and that you observe the following guidelines:

Another possibility is to run your program from the Windows 9X DOS box, after changing the EMM386 line in your CONFIG.SYS like this: 

 DEVICE=C:\WINDOWS\EMM386.EXE NOEMS L=131072

I'm told that this line (here for 128MB of installed memory) together with an "Auto" setting of the DPMI memory for the DOS box allows DJGPP programs to use up to 117MB of memory when running from the DOS box under Windows 9X.

If you need to use more than 256MB of physical memory, upgrade to CWSDPMI r5 or later. Another possibility is to run under OS/2 which features a built-in DPMI 1.0 support which can be configured to support as much as 512MB of DPMI memory (the user who reported this didn't know how much of this can be physical RAM).

Node: Getting DJGPP, Next: , Previous: Requirements, Up: Top

4 Where and What to Download?

This chapter explains where and how can you get DJGPP, and recommends which parts of the archive you should download.

Node: SimTel, Next: , Previous: Getting DJGPP, Up: Getting DJGPP

4.1 Where can DJGPP be found?

Q: Where can I get DJGPP?

A: Look on any SimTel.NET mirror in the pub/simtelnet/gnu/djgpp/ subdirectory, world-wide.

The primary SimTel.NET site is:
ftp.simtel.net9
Here is a list of hosts by countries that offer mirror sites:


Australia:
mirror.aarnet.edu.au
Australia:
ftp.tas.gov.au
Australia:
sunsite.anu.edu.au
Vienna, Austria:
ftp.univie.ac.at
Brussels, Belgium:
ftp-public.linkline.be
Sao Paulo, Brazil:
ftp.unicamp.br
Brazil:
ftp.iis.com.br
Bulgaria:
ftp.eunet.bg
Alberta, Canada:
ftp.telusplanet.net
Ottawa, Canada:
ftp.crc.ca
Vancouver, Canada:
ftp.direct.ca
Chile:
sunsite.dcc.uchile.cl
Czech Republic:
ftp.eunet.cz
Prague, Czech Republic:
pub.vse.cz
Czech Republic:
ftp.zcu.cz
Denmark:
ftp.net.uni-c.dk
Espoo, Finland:
ftp.funet.fi
Neuilly, France:
ftp.grolier.fr
France:
ftp.lip6.fr
Germany:
ftp.mpi-sb.mpg.de
Bochum, Germany:
ftp.rz.ruhr-uni-bochum.de
Chemnitz, Germany:
ftp.tu-chemnitz.de
Heidelberg, Germany:
ftp.uni-heidelberg.de
Paderborn, Germany:
ftp.uni-paderborn.de
Trier, Germany:
ftp.uni-trier.de
Wuerzburg, Germany:
ftp.rz.uni-wuerzburg.de
Athens, Greece:
ftp.ntua.gr
Hong Kong:
ftp.comp.hkbu.edu.hk
Hong Kong:
ftp.cs.cuhk.hk
Hong Kong:
ftp.hkstar.com
Hong Kong:
sunsite.ust.hk
Hungary:
ftp.iif.hu
Dublin, Ireland:
ftp.heanet.ie
Dublin, Ireland:
ftp.iol.ie
Rome, Italy:
cis.uniroma2.it
Italy:
ftp.flashnet.it
Naples, Italy:
ftp.unina.it
Italy:
mcftp.mclink.it
Saitama, Japan:
ftp.saitama-u.ac.jp
Saitama, Japan:
ftp.riken.go.jp
Japan:
ftp.iij.ad.jp
Japan:
ftp.u-aizu.ac.jp
Japan:
ftp.web.ad.jp
Latvia:
ftp.lanet.lv
Malaysia:
ftp.jaring.my
Mexico:
ftp.gdl.iteso.mx
Netherlands:
ftp.euro.net
Utrecht, Netherlands:
ftp.nic.surfnet.nl
Bergen, Norway:
ftp.bitcon.no
Krakow, Poland:
ftp.cyf-kr.edu.pl
Warsaw, Poland:
ftp.icm.edu.pl
Poznan, Poland:
ftp.man.poznan.pl
Timisoara, Romania:
ftp.dnttm.ro
Singapore:
ftp.nus.edu.sg
Singapore:
ftp.singnet.com.sg
Slovakia:
ftp.uakom.sk
Slovenia:
ftp.arnes.si
Johannesburg, South Africa:
ftp.is.co.za
Stellenbosch, South Africa:
ftp.sun.ac.za
South Africa:
ftp.netactive.co.za
South Africa:
ftp.saix.net
South Korea:
ftp.sogang.ac.kr
South Korea:
sunsite.snu.ac.kr
Spain:
ftp.rediris.es
Stockholm, Sweden:
ftp.sunet.se
Zurich, Switzerland:
sunsite.cnlab-switch.ch
Chung-Li, Taiwan:
ftp.ncu.edu.tw
Taipei, Taiwan:
nctuccca.edu.tw
London, UK:
ftp.demon.co.uk
London, UK:
ftp.globalnet.co.uk
London, UK:
ftp.easynet.net
Lancaster, UK:
mic5.hensa.ac.uk
London, UK:
sunsite.doc.ic.ac.uk
Arizone, USA:
ftp.datacanyon.com
Concord, California, USA:
ftp.cdrom.com
California, USA:
ftp.digital.com
Georgia, USA:
ftp.peachnet.edu
Urbana, Illinois, USA:
uarchive.cso.uiuc.edu
Massachusets, USA
ftp.bu.edu
Rochester, Michigan, USA:
OAK.Oakland.Edu
Missouri, USA:
galileo.galilei.com
New York, NY, USA:
ftp.rge.com
Oklahoma, USA:
ftp.ou.edu
Corvallis, Oregon, USA:
ftp.orst.edu
Pennsylvania, USA:
ftp.epix.net
Pennsylvania, USA:
ftphost.simtel.net
Virginia, USA:
mirrors.aol.com

Node: How to download, Next: , Previous: SimTel, Up: Getting DJGPP

4.2 How do I download DJGPP?

Q: How do I download files from these sites?

A: FTP to the nearest site, log in as anonymous, give your full e-mail address as password, and chdir to the djgpp subdirectory (the exact path to it might be different on different mirrors, check out the DJGPP archive path, for your nearest mirror). Then issue the binary command and download the files you need (see the list of required files) with the get or mget commands.

Node: DJGPP by WWW, Next: , Previous: How to download, Up: Getting DJGPP

4.3 What if I don't know what FTP is?

Q: What is that FTP thing? I only use Netscape and IE4 for Internet access.

A: The SimTel.NET site is on the Web.

You can also convert any of the mirrors' addresses listed in the list of SimTel.NET mirrors, above to a valid URL by prepending ftp:// to it. For example, the URL for FTP from SimTel.NET is <ftp.simtel.net/pub/simtelnet/gnu/djgpp/>. Typing such a URL into your Web browser will cause it to display the directory contents, and you can then click on individual files to download them.

For those of you who only have an e-mail connection to the Internet, there is an ftp-mail server at <mailto:ftpmail@pub1.bryant.vix.com>. Send a message with a single word "help" in the body to the above address, to get instructions.

Walnut Creek, the company which maintains the SimTel.NET collection where the DJGPP archives are held, also sells a DJGPP Development System CDROM. It includes everything from the DJGPP sites on SimTel.NET (even the old version 1.12 of DJGPP), some example source code packages to get you started, and a ready-to-run feature, which allows you to use DJGPP directly from the CDROM; you can also use a provided install program to copy some or all of the packages to your hard disk. To order the CDROM, go to the Walnut Creek Web site.

The FSF, the organization behind the GNU project, sells a CD-ROM with a full DJGPP development environment and most of the DJGPP ports of GNU software. For details, see the FSF Web site. Salvador Eduardo Tropea (SET), himself a veteran DJGPP user and developer, sells a low-cost CDROM with all the DJGPP v2 files, plus a lot of related stuff downloaded from the net. For information, send email to set-soft@usa.net.

Node: What to download, Next: , Previous: DJGPP by WWW, Up: Getting DJGPP

4.4 What Files to Download?

Q: What's the minimum set of .zip files I need to download?

A: This depends on what you are planning to use DJGPP for.

The following table lists required and recommended files by category. An alternative method of choosing the files suitable for your needs is to use the DJGPP zip-picker feature which will guide you through the process.

Node: Disk space, Next: , Previous: What to download, Up: Getting DJGPP

4.5 How much disk space will I need?

Q: Wow, that's a lot of files. How much disk storage will I need?

A: The following lists the approximate disk space required for several major configurations, and additional storage required for some optional packages:

Execution-only environment 300 KBytes
Developing C programs 15 MBytes
Developing C++ programs 20 MBytes
Developing Objective-C programs 16 MBytes
Additional storage for RHIDE 4 MBytes
Additional storage for DJGPP sources 6 MBytes
Additional storage for GDB 1.1 MBytes
Additional storage for Emacs 30 MBytes
Additional storage for Flex 280 KBytes
Additional storage for Bison 310 KBytes
Additional storage for Diffutils 560 KBytes
Additional storage for Make 650 KBytes
Additional storage for Patch 180 KBytes
Additional storage for Sed 200 KBytes
Additional storage for Graphics libraries 4 MBytes

Note that the above lists only approximate numbers. In particular, the disk cluster size can significantly change the actual disk space required by some of the distributions (those with a large number of files). The numbers above are for disks which have 8KB or smaller clusters.

In addition to the space for installing the software, you will need some free disk space for the swap file. You should leave enough free disk space to make the total virtual memory at least 20 MBytes; that will be enough for most applications. Invoke the go32-v2.exe program without arguments to see how much DPMI memory and swap space DJGPP applications can use. Depending on your DPMI host, you might need to review its virtual memory settings in addition to leaving free disk space; CWSDPMI only requires that enough free disk space be available, but other DPMI hosts have special settings to specify how much virtual memory they let their clients use, as explained in how to set up memory, below.

Node: DJGPP Fatware, Next: , Previous: Disk space, Up: Getting DJGPP

4.6 Can I get away with less megabytes?

Q: The above table means that I need more about 20 MBytes for C/C++ development environment; that's about 7 1.44MB diskettes to hold even the compressed archive!! Seems to me DJGPP is afflicted by the fatware disease....

Q: Pulling that many megabytes through the net from my overloaded SimTel.NET mirror is almost impossible. Can't you prepare a ZIP archive which only includes stuff I can't do without?

A: There are a number of shareware/freeware programs floating around which allow formatting DOS diskettes to almost twice their usual capacity, so you can use less floppies. One such program is 2M.

Also, with the proliferation of drives that can burn CD-ROMs and the availability of high-capacity removable media, like Zip drives, using floppies is no longer the only solution for a movable DJGPP.

To make downloading DJGPP easier, download and compile the BatchFTP program. It allows you to submit a script of FTP commands and will repeatedly try to login into the FTP site you specify until the script is successfully completed. It is smart enough to understand the messages which the FTP server sends to you (like login refused etc.) and also is nice to the remote server by sleeping for some time between login attempts. BatchFTP is free software and can be found on many FTP sites.

BatchFTP is a Unix program; those who access the net from their PC (not by dialing into some Unix host with a shell account), can use one of the available programs for automated FTP access.

As for the minimal DJGPP installation, volunteers are welcome to prepare such an archive and make it publicly available, in the same spirit as EZ-GCC did for DJGPP v1.x.

Node: Uninstall, Previous: DJGPP Fatware, Up: Getting DJGPP

4.7 How to uninstall a DJGPP package.

Q: How can I uninstall a certain package?

Q: How can I install a newer version of some package without leaving traces of the older installation?

A: The *.mft files in the manifest subdirectory hold the lists of all the files included in every package you install. For example, when you unzip gcc2951b.zip, it puts a file called gcc2951b.mft into the manifest subdirectory. The easiest way to remove all those files is to use the *.mft files as response files to a command which deletes files. For example: 

  rm -f @manifest/gcc2951b.mft

The rm program is part of the GNU Fileutils package, available as v2gnu/fil316b.zip from the usual DJGPP FTP sites.

Some packages might not have the *.mft files. In general, you should complain about such cases; however, if a package installs entirely into its own directory tree, you can uninstall it by simply removing that tree: 

 rm -rf package-dir

(The -r option tells rm to recursively remove all subdirectories of the named directory package-dir).

When you install a new version of a package, it is best to uninstall the previous version first, like in the above example, and then install the new one. Otherwise, you might leave behind old files that the new version doesn't overwrite, and that will cause problems due to incompatibilities with the new version.

Node: Docs, Next: , Previous: Getting DJGPP, Up: Top

5 The DJGPP Documentation

This chapter explains where to find and how to read DJGPP documentation, and how to solve occasional problems with the docs system.

Node: Where is the docs, Next: , Previous: Docs, Up: Docs

5.1 Where are the documentation files?

Q: I don't see any documentation files....

A: The documentation files are in the info/ subdirectory of your main DJGPP installation directory. You will need a program to read these docs, which are hypertext structured files. You have several choices:

  1. Use the stand-alone Info reader. Texinfo includes INFO.EXE and its docs. Unzip it and type info <Enter>. It will bring up a (hopefully) self-explanatory online help system. Confused? Press <?> to see the list of all Info commands. Still confused? Press <h> to have Info take you on a guided tour through its commands and features.
  2. Use the Info command of your favorite editor.

    If you use Emacs, you already know about Info. (What's that? You don't? Type C-h <i> and you will get the top-level menu of all the Info topics.) RHIDE also has an integrated Info reader, which is the core of its help system.

  3. Use SET's InfView browser. This is a replacement for the stand-alone Info reader from the GNU Texinfo distribution, but with a better user interface. It is written and maintained by Salvador Eduardo Tropea (SET), and is the same Info viewer that is part of RHIDE and SETEdit.
  4. Get and install TkInfo, a graphical browser for Info documentation that runs on MS-Windows and uses a port of Tcl/Tk. TkInfo is free and available from the Web.

Node: No Info, Next: , Previous: Where is the docs, Up: Docs

5.2 How to read the docs without Info?

Q: I'm too old/lazy/busy to learn yet another browser, and I despise uGNUsable programs like Emacs. How in the world can I read the DJGPP docs??

A: Info files are almost plain ASCII files, so you should be able to view them with your favorite text file browser or editor. You will lose the hypertext structure and you might have a hard time finding the next chapter (hint: look up the name of the Next node at the beginning of this node, then use the search commands of the browser, or the Grep program, to find that name), but other than that, you should be able to read all the text.

You can also produce pure ASCII files yourself, if you have their Texinfo sources. These are usually called *.txi or *.texi and should be included with the source distribution of every package. (You can use the DJGPP server's downloading services, to download individual files.) To produce an ASCII file foo.txt from the Texinfo file foo.txi, invoke the Makeinfo program like this: 

 makeinfo --no-split --no-headers --output=foo.txt foo.txi

Makeinfo is one of the programs which come with the GNU Texinfo distribution.

If you prefer reading the docs through the Web, point your Web browser to the docs page of the DJGPP Web site.

The full documentation of the DJGPP C library in HTML format is available for downloading from the DJGPP server.

Node: Printed docs, Next: , Previous: No Info, Up: Docs

5.3 How to print the docs?

Q: I like my docs the old way: printed on paper, right near my workplace. How can I print the documentation files which come with DJGPP?

A: Most of the DJGPP packages already have their docs converted to a printable format, look for the files named *d.zip at the same place where you got the binary *b.zip distribution. For example, the ready-to-print docs of GCC 2.95.1 should be in the v2gnu/gcc2951d.zip archive. These archives include a .dvi and a .ps file. The latter can be printed directly on a PostScript printer. If you don't have access to such a printer, you can use the .dvi file in conjunction with a DVI driver for your printer to produce a printed copy of the docs. A DVI driver is a program that reads the .dvi file and translates it into commands for a particular printer device which cause it to print the document. DJGPP ports of DVI drivers for LaserJet series of printers are available on SimTel.NET mirrors in the v2apps/tex directory. Drivers for DeskJet series are also available from there, in the dvdjNNb.zip archive. For other devices, download and install the Ghostscript interpreter which supports a lot of popular printers.

You can also get the GNU documentation in DVI, PostScript, and two-up PostScript formats in .tar.gz format from the DJGPP server.

Note that some documentation files (notably, the one for GCC and Emacs) will produce voluminous print-outs. You have been warned!

If you cannot find a ready archive with printable files anywhere, you will need to get and install a program called TeX or its work-alike, like emTeX. A DJGPP port of TeX is available via FTP. Install TeX, then run the texi2dvi shell script12 on the docs' source files (called *.txi or *.texi) which you get with the source distribution of every package you download. TeX produces a .dvi file which you can then print using one of the available DVI drivers, as explained above. To convert a .dvi file into PostScript, use the DVIPS program; you can find it as dvps584.zip on the above-mentioned site, together with the TeX port.

If TeX won't run, check that you have the file texinfo.tex which defines the TeX macros specific to Texinfo files. If you don't, get the latest GNU or DJGPP Texinfo distribution which includes that file.

If you'd like to produce printed docs of the library reference, TeX might complain that it cannot find a file named libc2.tex. This file is generated from all the *.txh files in the DJGPP source distribution (djlsr203.zip) and is usually built as part of the library build procedure. In order to generate this file without building the entire library, you need to install djlsr203.zip and the C++ compiler, then go to the src/libc directory and type this from the DOS command prompt: 

  make misc.exe ../hostbin
  make -C mkdoc
  make -C libc info

DJGPP comes with a program called TEXI2PS which can convert *.txi files into a crude PostScript; try it if you don't care much about the appearance of the printed docs. Its advantage is that you don't need to install any additional packages, just to fetch the Texinfo sources of the docs.

Finally, if you don't mind paying for the printed documentation, the Free Software Foundation sells printed copies of manuals for GNU packages. You can contact the FSF for details.

For those who prefer reading docs with a Web browser, many GNU manuals in HTML (hypertext) format, suitable for reading with your Web browser, can be viewed at the DJGPP Web site. The *d.zip archives also include the docs converted to HTML, which you can browse locally on your machine.

Node: Cannot find docs, Next: , Previous: Printed docs, Up: Docs

5.4 Some docs are nowhere to be found...

Q: I looked in my info/ subdirectory, but I can't find docs for some of the utilities, like SED or GPROF.

Q: STL, the C++ Standard Template Library, seems to be undocumented....

A: SED and GPROF are documented in the latest GNU releases, v2gnu/sed302b.zip and v2gnu/bnu281b.zip. Download the latest releases, and you will get the missing docs.

The STL documentation is not included in the GNU GCC distribution (it appears that nobody has bothered to write a free documentation for it). But you can find the STL docs on the net; this includes the full documentation and a tutorial. Many books that describe C++ programming also include documentation of large parts of the STL. Another place to look for the reference material about C++ language and classes is the Dinkumware site. (Dinkumware is a company founded by P.J. Plauger, one of the world's leading experts on C++ and the STL, which produces the C++ library used by the MSVC compiler.)

The ANSI organization is selling the official ANSI C++ Standard (full document, about 800 pages) for $18 in PDF format.

If you have some other package without any docs, try downloading the source archive (*s.zip) for that package and look inside it, usually in the directory called man/ or doc/. Omitting documentation from the binary (*b.zip) distribution is generally considered a bug, so if you find the docs in source distribution only, please report these cases on the comp.os.msdos.djgpp news group, so that next binary release could fix this.

Node: Man pages, Next: , Previous: Cannot find docs, Up: Docs

5.5 What are these foo.1 files?

Q: Some docs files are called foo.1 or bar.man or baz.nroff, and they seem to be written in some weird format which is very difficult to read. How can I convert them to readable text files?

A: That weird format is the troff format which is used for writing Unix manual pages. The Unix command man converts them to formatted text files which are usually displayed with a program like more or less (and here less is considered to be more than more :-)). The formatted file includes bold and underlined letters produced by over-typing using Backspace characters.

DJGPP binary *b.zip distributions include such man pages already formatted and ready to be browsed. To browse formatted man pages, you will need to install a clone for the Unix man command. One such clone is available from the DJGPP sites.

man knows how to find a manual page file, and will format it if it isn't formatted already, but to browse these files you will need a program that can page through a text file and which understands how to show bold and underlined letters instead of backspace-over-typed characters. I suggest to download the DJGPP port of GNU Less, which uses colors to show bold and underlined letters.

Having installed man and Less, you should be able to view *.1 files like e.g. patch.1 with several alternative tools:

Note that all of these alternatives require man and Less to be installed.

The binary distribution of the DJGPP port of bash includes a simple SED script called man2txt.sh which will convert formatted man pages into plain text; you can then read them with any text browser or editor. To convert, invoke Sed like so: 

  sed -f man2txt.sh < file.1 > file.txt

If you want to be able to browse unformatted man pages, get and install the DJGPP port of Groff. man will automatically call Groff if it finds an unformatted page, so all the ways mentioned above to browse man pages will work with unformatted pages as well, once you install Groff.

Note that, for GNU packages, the man pages aren't always updated on a regular basis. If you need more up-to-date information, see the Info docs.

Node: Last resort, Previous: Man pages, Up: Docs

5.6 What if the docs don't say enough?

Q: OK, I've got the docs and have read them, but I still can't figure out some details.

A: Some ported packages include DJGPP-specific or MSDOS-specific README files (named README.dj, README.dos or some such), which explain DOS-specific issues; you should read them before any serious use of the package, or in case of any problems. If this doesn't help, download the sources and look there, or ask on the net--either the DJGPP News group or appropriate GNU News groups.

Node: Trouble, Next: , Previous: Docs, Up: Top

6 When the Compiler (or Make, or Info, or ...) Crashes...

This chapter explains how to deal with certain problems which may prevent DJGPP programs from running on your machine. The first 13 items on the next menu describe specific problems; if yours isn't solved with these techniques, read the description of the general debugging procedure.

Node: Programs hang, Next: , Previous: Trouble, Up: Trouble

6.1 GCC or some other DJGPP programs hang

Q: When I try to compile anything, GCC just hangs!

Q: Some programs, like Info and Less, hang on my Windows 9X machine after some time, requiring to close the DOS box. Help!!

Q: Bash hangs on Windows 9X after the first DJGPP program I run from it!

A: If you are using GCC 2.8.1, and if only GCC hangs, then chances are that you have set the DJGPP environment variable incorrectly, or didn't set it at all, or messed up your DJGPP.ENV file by editing it. Refer to what to do if GCC hangs, later in this FAQ, for details about possible problems with setting DJGPP. When DJGPP is not set, or points to a non-existent directory, the first release of GCC 2.8.1 would enter an endless loop (the NTDVM process on Windows/NT will go nuts allocating memory, as a result of this).

The latest uploads of the GCC binary (v2gnu/gccNNNb.zip) were modified to prevent them from hanging. They abort with an error message instead of hanging. You should upgrade to the latest binaries of GCC, and also set your DJGPP variable correctly.

Some people fail to read the release notes which come with GCC, and do not install it incorrectly. If you installed GCC recently and it began to hang, now is the time to read those instructions again (they are installed as gnu/gcc-X.YZ/problems.txt).

If interactive programs like Bash, Less, Info, Emacs, and RHIDE are those which hang, and if it only happens on Windows 9X after running another DJGPP program from within those programs, then your Windows 9X installation is afflicted by a subtle bug whereby programs which call function 1680h of the Interrupt 2Fh (to release the rest of their time slice when they are idle) hang after they spawn another DJGPP program. A modified version of the library function __dpmi_yield, which works around that bug in Windows, is available in DJGPP v2.02 and later, and latest uploads of the binaries for the affected programs should be free from this problem. If you cannot find a pre-compiled binary that works, get the sources and rebuild the program with the latest DJGPP release.

Node: No DPMI, Next: , Previous: Programs hang, Up: Trouble

6.2 GCC says "No DPMI"

Q: I'm trying to run gcc, but all I get is a message saying "Load error: no DPMI - Get csdpmi*.zip". What am I doing wrong?

A: You don't have a DPMI server installed, and DJGPP v2 requires it to run. You can either use one of the commercial DPMI servers (e.g., run GCC in a DOS box from Windows) or download and install CWSDPMI (v2misc/csdpmi4b.zip from SimTel.NET mirrors) which is a free DPMI server written for DJGPP.

If you already have CWSDPMI installed, and these messages still appear, it might be because of a messed up PATH setting. The DJGPP startup code looks for cwsdpmi.exe along the PATH, and, being optimized for size, it might not be robust enough to cope with all possible cases of weirdness in the value of PATH. Try to copy cwsdpmi.exe into the same directory as the program you are invoking, and if that helps, change your PATH as appropriate.

If you see the message "Load error: no DPMI - Get csdpmi*.zip" on Windows/NT, it most probably means that you have disabled the DPMI services built into NT. One way that this might happen is if you edit the autoexec.nt file and remove the line which loads dosx.exe, or change some of the parameters on that line. You cannot use CWSDPMI on NT, so your only bet is to restore NT's built-in DPMI services.

Node: Buggy DPMI, Next: , Previous: No DPMI, Up: Trouble

6.3 Buggy DPMI host or junk in DJGPP.ENV can crash v2.x programs

Q: I cannot run v2 applications: they all hang or reboot my system, while v1.x apps run OK. Is this what v2 is all about--getting me out of the DJGPP community?

A: No, believe it or not, we don't want to oust you. Your problems might be caused by a buggy DPMI (see DOS Protected Mode Interface) host installed on your machine. One DPMI host which is particularly known to be a source of trouble is the one which comes with Novell NWDOS (and also with early versions of Caldera's DR-DOS, a.k.a. OpenDOS, which is a derivative of NWDOS). Please see if DJGPP programs run when you disable DPMI services of your usual configuration (DJGPP programs will then use the CWSDPMI host supplied with DJGPP). To turn off the DPMI host built into Novell NWDOS and Caldera's DR-DOS, either remove the DPMI=TRUE parameter from the EMM386 command line, or type DPMI OFF from the DOS command prompt.

Version 7.03 and later of Caldera's DR-DOS reportedly don't have this bug in their DPMI server, so upgrade to a latest DR-DOS version if you can.

Another DPMI host which is known to cause problems in DJGPP is Quarterdeck's QDPMI which comes with QEMM 7.5. It was reported to cause Info and all DJGPP debuggers to crash. If you use QDPMI, upgrade to the version 7.53 or later (patches for that version are available from the Quarterdeck's ftp site), or disable QDPMI and use CWSDPMI.

Node: GCC optimizations, Next: , Previous: Buggy DPMI, Up: Trouble

6.4 GCC can crash during optimization

Q: When I compile my program, the compiler crashes, but the problem seems to go away if I compile without optimization.

Q: The compiler prints "Virtual memory exhausted" and dies while compiling some long functions with some optimization options, such as -funroll-loops or -fforce-addr.

A: For some programs, this can be caused by an insufficient stack. Some source files make cc1.exe or cc1plus.exe need preposterously large amounts of stack space, but only when you turn on optimizations. (One user reported that an innocent-looking C source file required 700KB of stack before cc1.exe was able to compile it with optimizations!) Try stubediting the compiler to enlarge its stack, as described elsewhere in this FAQ, how to enlarge the stack, before you try any other remedies in this section.

If GCC reports that it has exhausted virtual memory, you should first see if your DPMI memory plus the swap space is large enough (run go32-v2 with no arguments to display the available memory) and make more memory available, if the reported amount is too small. Some programs really need large amounts of memory to compile and/or link. For example, linking cc1.exe is known to consume more than 12MB of memory. On Windows 9X, be sure to set the amount of DPMI memory available to the DOS box to the maximum value of 65535K (64MB) in the DOS box property sheets, under Memory, as explained in how to enlarge memory in the DOS box.

Some users have reported that GCC seems to run out of virtual memory if TMPDIR environment variable points to a RAM disk which doesn't have enough free space. Changing TMPDIR to point to a hard disk would reportedly save the day in those cases.

Compiling with PGCC or EGCS variants of the GNU compiler, as well as GCC versions 2.95 and later (which are descendants of EGCS) can also sometimes run out of virtual memory. These versions of the compilers are memory hogs, especially when compiling C++ programs, and at high optimization levels. One particular case is when your program makes use of many STL classes. Try lowering the optimization level, or compile without optimizations.

With GCC 2.95 and later, using -pedantic or -Wreturn-type can cause an explosion in the amount of memory needed to compile template-heavy C++ code, such as code that uses the STL. Since -Wall includes -Wreturn-type, it can also cause massive memory consumption; try -Wall -Wno-return-type to work around the problem.

One user reported that optimization switches force GCC to use a math co-processor, which can cause it to crash on a machine that lacks a numeric processor. Be sure you didn't delete the emu387.dxe file from your bin subdirectory, when you compile on such machines, and that your emulation-related setup is right. See how to set up FP emulation, for details.

GCC can sometimes crash when optimizing, especially when compiling C++ programs, in particular if your code has some syntactic or semantic bug. (This is usually a genuine GCC bug, not something special to DJGPP.) Upgrade to the latest version of GCC. If that doesn't help, then narrow the offending code fragment using the #if 0 ... #endif paradigm. If this fragment includes an error, correct it and try again; if it is syntactically and semantically correct, then rewrite it as equivalent, but syntactically different one.

A GCC switch can sometimes help you zero in on the code fragment that causes GCC to crash. If you add -Q to the GCC command line, it will print the name of every function it compiles. The function that makes it crash is probably the one whose name is the last one printed, or the one after that.

As an extreme measure, don't optimize at all, if that's the only way to make your program compile.

Another reason for crashes could be some problem with your system hardware or the BIOS (like if you set an incorrect number of wait states when accessing memory, or overclock the CPU too much). To check, try running the same compilation on another machine, or review your BIOS settings and hardware configuration.

Yet another cause for such crashes can be connected with excess memory usage that GCC needs when compiling certain programs, which makes some DPMI hosts fail. For details about this, see CWSDPMI allocation problems, below.

Node: Missing subprograms, Next: , Previous: GCC optimizations, Up: Trouble

6.5 Why does GCC say "cannot exec as"?

Q: When I try compiling a program, GCC aborts saying "Installation problem, cannot exec `as': No such file or directory (ENOENT)". What does that mean?

Q: When I try compiling a program, GCC aborts saying "Installation problem, cannot exec `cpp': No such file". Huh?

A: This usually means that GCC couldn't find some program it needs to run to compile your source. Check the COMPILER_PATH environment variable or what the COMPILER_PATH line in the DJGPP.ENV file says, and make sure they point to the directory where DJGPP programs reside. Also check that the named directory has all the required programs: cpp.exe, cc1.exe, cc1plus.exe, cxxfilt.exe, gasp.exe, as.exe, ld.exe, and (for Objective-C) cc1obj.exe. A typical case is when people fail to install the Binutils package and GCC cannot find as.exe (the assembler) and ld.exe (the linker). You can use the -v switch to GCC to see what programs it invokes and which one of them causes the fatal error.

Beginning with version 2.8.0 of GCC, the place where the pre-processor, cpp.exe, and the C and C++ compilers, cc1.exe and cc1plus.exe, are installed, has changed: they are no more in the same directory as gcc.exe itself. If you are using GCC version 2.8.0 or later, and the compiler cannot find cpp.exe or cc1plus.exe, read the installation instructions carefully: the file problems.txt explains how to change the settings on DJGPP.ENV so that GCC will find the pre-processor. Also, be sure to remove all traces of the previous compiler installation, since mixing different compiler versions can be another cause for such problems. See uninstalling a package.

Node: Internal error, Next: , Previous: Missing subprograms, Up: Trouble

6.6 What does "Internal compiler error" mean?

Q: During compilation, GCC prints "Fatal: Error in DJGPP installation. Environment variable DJGPP is not defined" and then aborts....

Q: GCC aborts with "Internal compiler error" when compiling a large C++ program.

Q: GCC behaves erratically when compiling programs, sometimes crashes with register dump, sometimes compiles okay, sometimes reports "Internal compiler error". Why is this happening?

Q: When I try to compile any program, GCC prints "Abort!" and doesn't compile anything....

Q: The compiler crashes or dies with "Virtual memory exhausted" when I compile my simple program!

A: The fatal error message about DJGPP not being defined means just that--that your DJGPP environment variable is not defined. The other two messages you could see are: 

 Environment variable DJGPP point to file `XXYYZZ'
 which doesn't exist

or 

 Environment variable DJGPP points to wrong or corrupt file `ABCDE'

(In both cases, you will see an actual file name instead of XXYYZZ and ABCDE.) These messages mean that DJGPP points to a non-existing directory, or to a file whose contents are too messed up. Beginning with version 2.8.1, GCC refuses to work when the DJGPP variable doesn't point to the actual path name of a valid DJGPP.ENV file, because GCC uses the value of the DJGPP variable to find out where to look for its subprograms like cpp.exe, cc1.exe, etc. To solve this, set the DJGPP variable as the installation instructions (in the file readme.1st) describe. Also, make sure you didn't mess up the beginning of the DJGPP.ENV file, where the value of the DJDIR variable is computed (when in doubt, compare it with the stock DJGPP.ENV from the djdevNNN.zip distribution).

A possible cause for the "Abort!" message is that the TMPDIR environment variable points to a non-writable directory. If you don't set TMPDIR from your AUTOEXEC.BAT or from the DOS prompt, the DJGPP startup code sets it to the tmp subdirectory of the main DJGPP installation directory. If DJGPP is installed on a read-only drive, like CD-ROM or an unwritable networked drive, this default will not work. To solve this, set TMPDIR to point to a writable temporary directory. If TMPDIR is not set at all, GCC tries to use TEMP and TMP, in that order, so make sure these also point to a valid directory.

Internal compiler errors (a.k.a. bugs) can also cause GCC to print "Abort!". This FAQ describes a procedure that allows you to find the spot in the sources where the compiler aborts, see use of the -Q switch, above. Once you find the offending code, you could rewrite it and/or submit a bug report to the GCC maintainers.

When GCC aborts with a message such as "Internal compiler error" or "Exiting due to signal SIGSEGV", it might mean a genuine bug in GCC (which should be reported to FSF), but it can also happen when GCC requests additional chunk of memory, and the DPMI server fails to allocate it because it exhausts available memory for its internal tables. Old releases of CWSDPMI could fail like this if an application asks for a large number of small memory chunks. Beginning with release 2, CWSDPMI defines a larger (6KB) default heap that is configurable by CWSPARAM program to be anywhere between 3K and 40K bytes, without recompiling CWSDPMI. You should upgrade to the latest CWSDPMI if you experience such problems, and if that doesn't help, bump up the size of CWSDPMI heap using CWSPARAM.

Some innocent-looking programs are known to cause GCC to gobble preposterous amounts of memory, which could cause it to crash or abort after printing "Virtual memory exhausted". One particular case of such programs is when you initialize very large arrays. For example, to compile a source which initializes a char array of 300,000 elements requires more than 60MB(!) of memory. You should avoid such constructs in your programs.

Some programs require very large amounts of stack to compile. DJGPP programs have a fixed-size stack that is by default 256KB (512KB in DJGPP v2.02 and later). If the compiler, cc1.exe or cc1plus.exe, doesn't have enough stack to compile a program, it will overflow its stack and crash, or hang, or die with "Internal compiler error". You can enlarge the stack size of any DJGPP program by running the stubedit program, like this: 

  stubedit cc1.exe minstack=1024k

I recommend to enlarge the maximum stack size of cc1.exe to at least 1024K bytes and that of cc1plus.exe to at least 1.5MB. Some people report that they needed to enlarge both the heap of CWSDPMI and the stack of the C++ compiler to make such problems go away.

For a program that you wrote, another work-around for the cases where a program crashes due to failure of CWSDPMI to allocate more RAM is to use an alternative algorithm for sbrk, by putting the following somewhere in your program: 

  #include <crt0.h>
  int _crt0_startup_flags = _CRT0_FLAG_UNIX_SBRK;

Note that the Unix algorithm for sbrk might cause trouble in programs that install hardware interrupts handlers.

Note that the problems with insufficient stack size have nothing to do with the total available memory as reported by go32-v2. A compiler can crash because of insufficient stack size even though it has gobs of memory available to it. When in doubt, always enlarge the compiler stack size.

Sometimes, GCC can crash due to problems with your system hardware. In particular, bad memory chips can cause GCC to behave erratically, since the compiler is a memory-intensive program: it moves large buffers around alot, and uses lots of memory. One cause of problems with accessing memory is incorrect setting of the wait states in your BIOS setup, or too aggressive CPU cache mode that your motherboard cannot support reliably, or overclocking the CPU. Try to play with your BIOS setup and see if that helps. If you overclocked the CPU, try resetting it back to its normal speed.

One user reported that he had random crashes and seemingly-missing files due to a disk without proper cooling. So if your system sometimes cannot find files that you know are there, check whether your disk gets proper cooling and generally works okay.

Another rare case of crashes in GCC was reported on Windows 3.X. It seems to be related to the small probability of getting non-contiguous memory blocks from the Windows' DPMI server. In general, the DJGPP library handles these cases, so it is possible that the problem is actually somewhere in GCC (more accurately, in cc1, the C compiler). A tell-tale sign of this problem is that the CS and DS limit value printed in the crash message is very close to the end of the 4GB address space, like fffeffff. The only known cure for these cases is to patch or rebuild GCC with the Unix sbrk algorithm, see above.

Node: Unknown filetype, Next: , Previous: Internal error, Up: Trouble

6.7 What does "Unknown filetype" mean?

Q: I get error messages saying "Unknown filetype" from GCC.

Q: Since a few days ago, whenever I try to run most of the DJGPP programs, they print a message "C:\DJGPP\BIN\prog.exe: not COFF" and just terminate. Help!!!

A: It might be that your DJGPP programs and/or STUBIFY.EXE are infected by a virus. (This is not a joke! It did happen to a few of us and can happen even to you.) As the DOS stub prepended to the DJGPP programs is very short, many viruses cannot attach themselves to it without overwriting the beginning of the DJGPP COFF image which specifies vital information such as location and length of various program sections, therefore triggering this error from the code in the stub that loads the COFF image.

Another possible cause of the "Unknown filetype" message is that you mix a v2.0 gcc.exe driver with cc1plus.exe, cc1.exe or other programs from an old v1.x distribution.

Node: Make hangs, Next: , Previous: Unknown filetype, Up: Trouble

6.8 Compiler hangs, but only when invoked from Make

Q: My compiles run OK from the command line, but hang or crash when I invoke the compiler from Make.

Q: When I try to compile something, I get a message "16-bit DPMI not supported".

A: Be sure you aren't inadvertently invoking some 16-bit programs, such as Borland's Make or cpp.exe from Borland C. GCC cannot run them, and cannot run under Borland's Make (because Borland's programs are 16-bit DPMI clients, and the DPMI 0.9 spec doesn't allow mixing them with 32-bit DPMI clients such as the DJGPP programs). It might be that another program called make.exe or cpp.exe is found earlier on your PATH than the Make and the preprocessor which came with DJGPP. Moving DJGPP's bin directory to the beginning of your PATH will usually solve these problems.

Note that if you try to mix 16-bit and 32-bit DPMI clients in Windows DOS box (like use Borland's Make to run GCC, or invoking Borland's cpp.exe under GCC), the DOS box will usually crash. So this problem is not specific to CWSDPMI.

If you must use a non-DJGPP compiler or another utility together with DJGPP programs, one solution would be to find a version of that utility that doesn't use the 16-bit DPMI services. For example, many DOS compilers have a real-mode version that won't conflict with DJGPP.

If you use Make compiled under DJGPP v1.x, you will also experience all kinds of trouble when invoking programs compiled under DJGPP v2. That's because v1.x programs cannot spawn v2 programs directly (the v1.x program sees that the child is a DJGPP program and tries to call go32 to run it, but v1's go32 cannot run v2 programs). The result usually will be that the child either crashes or silently exits. If that's your problem, be sure to upgrade your Make to the port distributed with v2. (Note that v2.x programs generally know how to spawn both v1.x and v2.x programs.) You can use go32-v2 to work around this limitation (see description of go32-v2, below), but I suggest doing that only if you absolutely cannot upgrade to v2's Make.

Some users report that v1.x programs might sometimes hang or reboot the machine when invoked from v2.x Make. The reason for this is a known bug in the versions of go32-v2.exe program distributed with DJGPP v2.0 and 2.01: it would not restore the keyboard handler after the COFF image it runs exits. To work around that bug, don't touch the keyboard throughout the entire run of Make; to solve it, upgrade.

Node: Info cannot find Top, Next: , Previous: Make hangs, Up: Trouble

6.9 Info doesn't like some files

Q: When I run the Info browser, it tells me it cannot find the node "Top".

Q: Sometimes, when I mistype the name of the Info topic, info.exe seems to hang....

A: Check your installation of info files. The file DJGPP.ENV in the root of your DJGPP installation mentions the variable INFOPATH which should point to the directory where Info looks for its files. It must find there a file named DIR, the file you are trying to read, and other files with .iNN or .NN extension, where NN is a number. Also, make sure you didn't edit the beginning of DJGPP.ENV, where the value of %DJDIR% is computed; if you did, try the original DJGPP.ENV.

Also, the DJGPP environment variable should be set to point to the full pathname of the file DJGPP.ENV. See problems with DJGPP variable setting, for a description of some frequent problems with setting the DJGPP variable.

Assuming the above checks OK, and all the necessary info files are indeed installed in those directories (did you remember to give that -d switch to PKUNZIP when unzipping all DJGPP packages?), it might be that you have an old version of info.exe. Upgrading to version 3.12 or later should solve several problems which cause Info to complain about the "Top" node, or at least make its error messages more self-explaining.

Some people unzip the txi40b.zip file with WinZip and fail to disable its feature whereby by default each zip file is unzipped into a separate directory. This disrupts the DJGPP directory structure and break almost every DJGPP package, including Info. You need to unzip all DJGPP files under the same directory!

Another possibility is that the file DIR or the Info file that you want to browse is corrupted. For example, it might be a compressed file, but without the tell-tale .gz or similar extension that tells info.exe to decompress it. You could examine the offending file(s) with a text editor, and re-install them as needed.

If you invoke Info with a name of a topic that is non-existent or not installed on your system, and you don't have a man command or its clone installed, versions of Info before 3.12 would wait for 15 seconds before they print an error message and exit. If you think Info is wedged, wait for 15 seconds and see what happens then. The DJGPP port of Texinfo 3.12 removes this misfeature, so upgrade if you can.

Node: Info crashes, Next: , Previous: Info cannot find Top, Up: Trouble

6.10 Info Crashes During Startup

Q: Whenever I invoke info.exe, it immediately crashes! How can I use DJGPP without reading all those wonderful docs??

A: One possible reason for this is a bug in EMM386 shipped with some versions of DOS 6.x. This bug is only triggered if your system loads the DISPLAY.SYS driver, usually to display non-US characters supported by your national codepage. In addition, this bug causes Info to crash only if it is run in 35- or 40-line display mode; any other display size avoids the problem. (The default display mode is 40 lines, as set in the [info] section of DJGPP.ENV.) So either switching to another display size, or removing either EMM386 or DISPLAY.SYS from CONFIG.SYS, should prevent Info from crashing.

If you use DJGPP v2.0, yet another cause of crashes in Info might be trailing whitespace in the DJGPP.ENV file. The tell-tale signs of this failure are a stack dump that is bogus or doesn't start with your `main' function, or a series of SIGSEGV that won't stop. Actually, this is a bug in the DJGPP v2.0 startup code, so any v2.0 program could crash in this way, but since the last section of stock v2.0 DJGPP.ENV belongs to Info, it is the one which suffers most from this bug. Make sure your DJGPP.ENV doesn't have a ^Z character at the end (some DOS editors put it if you edit the file), and doesn't end with a blank line. Alternatively, upgrade to DJGPP v2.01 or later, where that bug is fixed.

Node: Bash crashes, Next: , Previous: Info crashes, Up: Trouble

6.11 Why does Bash crash?

Q: Bash crashes on me when I invoke shell scripts....

Q: Why does Bash say "pipe error: Access denied" when I try to run two programs via a pipe?

Q: When I run certain complex shell scripts, Bash sometimes prints a message saying "Cannot make a child for command substitution: (null)". What gives??

Q: What does it mean when Bash says "Can't make pipes for command substitution!"?

A: If Bash crashes when you invoke shell scripts, check whether those scripts have #!/bin/sh on their first line. If they do, the most probable reason for the crashes is that you don't have sh.exe anywhere on your PATH (it does not have to be in /bin!). Either copy bash.exe into sh.exe, or create a "symlink" to bash.exe, like this: 

 ln -s c:/djgpp/bin/bash.exe c:/djgpp/bin/sh.exe

(replace c:/djgpp with the actual pathname of your DJGPP installation).

Error messages about pipes and command substitution usually mean that your TMPDIR environment variable points to an invalid drive; make sure TMPDIR is set and points to an existing directory, and that the drive where it points is writable and not full. Old ports of Bash had problems with `command` substitution, so make sure you have the latest binaries.

Node: ThinkPad, Next: , Previous: Bash crashes, Up: Trouble

6.12 DJGPP programs crash on a ThinkPad

Q: I cannot run any DJGPP program on my ThinkPad! They all cause the Windows DOS box to crash with a message about a General Protection Exception at some cryptic addres like 0277:0044!

A: This is a known problem with the ThinkPad: it is supplied with a set of PCMCIA device drivers from CardWorks that conflict with many DOS-extended programs, not only DJGPP (the CardWorks README file mentions problems with PKZIP which also uses protected-mode instructions).

The only known solution is to uninstall the PCMCIA drivers. Obvioulsy, this is a solution only if you don't actually use any of the PCMCIA devices.

Node: Linker accesses other drives, Next: , Previous: ThinkPad, Up: Trouble

6.13 Why does the Linker Access my CD Drive or the network?

Q: Why is it that every time I link a program, the CD-ROM drive is accessed?

Q: Whenever I link programs, GCC invokes something called `collect2' which accesses my LAN when it runs. Why?

A: CD-ROMs or other drives being accessed during linking is due to a bug in Binutils 2.7 and in an early release of Binutils 2.8.1: the linker would always try to look for its script djgpp.djl in a certain directory on the D: or E: drive (the former in Binutils 2.7, the latter in 2.8.1), no matter which disk uses that letter (these accesses usually go unnoticed with hard disks, but are visible with CD-ROMs, Zip drives, or other slower devices). Download and install the latest bnuNNNb.zip archive you can find on SimTel.NET mirrors, and the problem should go away.

If collect2 seems to be accessing the network, it is due to a bug in the early ports of GCC 2.95: if a root directory of some drive appeared in your PATH setting, collect2 would try to access a file whose name has two slashes, like C:\/foo. This causes Windows 9X to treat this as a UNC (a.k.a. network share) name, and search the network for such a server which exports this share. The ports of GCC 2.95.1 and later don't have this bug.

You can see which directories on what drives does the linker try to access by passing the --verbose option to the linker. Here's an example: 

 gcc -o hello.exe hello.o -Xlinker --verbose > linker.log

This redirects the linker log to a file which you can then examine. Since the list of directories accessed by the linker doesn't depend on the program being linked, you can try this with any trivial program.

Sometimes, accesses to other drives come from some over-zealous anti-virus software. If you have one of these installed, check out its options: perhaps there are some superflous drive letters there.

Node: General trouble, Next: , Previous: Linker accesses other drives, Up: Trouble

6.14 Other kinds of trouble

Q: I've installed DJGPP just like explained in the README.* files, but when I run gcc, my machine crashes/hangs/needs cold boot.

Q: I get errors I can't figure out when I try to compile something.

A: Add the -v switch to the GCC command line and run it again. It will print all the subprograms (compiler passes) it is running. Then you can see which subprogram caused the error, or where your machine crashes. This might give you a hint on what's wrong.

One cause of such problems might be that your system is set up inefficiently. If GCC doesn't get enough free RAM, it will run very slowly, and you might think it crashed when in fact it didn't. (This kind of problem usually happens on memory-starved machines.) Invoking go32-v2 with no arguments will report the amount of memory available to DJGPP programs; large programs might require up to 20MBytes to compile, sometimes more. If you run from the DOS box on Windows 9X, set its DPMI memory property to 65535KB (64MB) and try again. Check out the system configuration advice, earlier in this FAQ list, for other suggestions about your system configuration.

Sometimes, if the TMPDIR environment variable points to a full disk, GCC may hang or crash as well. Make sure you have enough free disk space where TMPDIR points.

A similar case is when DJGPP programs are run off a CD-ROM: in this case, the default setting of TMPDIR points to the CD drive, which is unwritable. You need to point TMPDIR to a writable directory.

Node: Redirect, Next: , Previous: General trouble, Up: Trouble

6.15 I cannot keep up with the error messages

Q: I want to read all the error messages that GCC throws at me, but there are so many that I can't keep up. How can I redirect them to a file?

Q: When I add -v to the GCC command line, how can I put all the voluminous output into a file, so I don't miss anything when reporting a problem?

Q: I have this nifty graphics program which bombs from time to time, but the registers and traceback info are hidden by the graphics display. How can I see it?

A: Error messages are usually written to stderr, and stock COMMAND.COM cannot redirect it. There are several alternatives to do that:

  1. You can use a shell smarter then COMMAND.COM, such as 4DOS or bash, which knows how to redirect standard error stream to a file. 4DOS is shareware and can be found on SimTel.NET, while bash is available from the v2gnu directory of DJGPP archives on SimTel.NET.
  2. You can also run your program under any one of the programs which save all screen output of programs they spawn in a file. I suggest using a program called SCRIPT, which is similar to its Unix namesake. It has an advantage of saving everything which goes to screen and showing it on the screen at the same time. You can find SCRIPT on SimTel.NET.
  3. Or you can use the REDIR program which comes with DJGPP. It also redirects standard output and/or standard error to a file, but you don't get a chance to look at the output while the program runs. Here's how to run GCC with REDIR: 
      redir -o gcc.log -eo gcc -v ...

    (put the rest of the GCC command line instead of the dots). The messages printed by GCC will be written to the file gcc.log.

Windows/NT has its own program named redir.exe, so make sure the DJGPP's bin subdirectory is listed in the PATH variable before the NT directories.

Node: Archive search, Next: , Previous: Redirect, Up: Trouble

6.16 How to search DJGPP archives

Q: OK, I have all this voluminous output of gcc -v, but I still have no clue.

A: Your problem might be one which has already been posted and solved on the DJGPP News group. DJ Delorie has set up a searchable News group archive on his Web server. You can search the entire mailing list archives in just a few seconds. DJ's archives are always up to date, as they receive and store all posted messages automatically, but the index is updated every 24 hours, so the last day might not be searchable yet. To search the DJGPP archives, point your Web browser to the URL above and specify a list of keywords pertinent to your problem. You will get a list of messages which include those keywords; clicking on any of the messages will get the full text of that message.

Node: Totally lost, Previous: Archive search, Up: Trouble

6.17 How to ask DJGPP gurus for help

Q: I've read this monstrously-large FAQ, searched the news group archives, but didn't find anything helpful. I am totally lost. Help!!!

Q: I don't have time to download all these messages, not to mention looking through them. Can't you DJGPP gurus help me? Please??

A: DJGPP News group is famous for its outstanding user support. To get a fast and effective solution to your problem, you will have to supply the relevant info about your system, and describe exactly how things went wrong for you. To gather this info, do the following:

Be warned that you might get several dozen messages in reply to your request; this is not meant to overflow your mailbox or sabotage your relationship with your system manager, it's just the usual friendly response of fellow DJGPP'ers to your lonely cry for help. Some of the replies might suggest what you already checked and reported in your original message, or even miss the point altogether. Be ready for this and don't flame us for trying to help you as much as we can.

Node: Compiler performance, Next: , Previous: Trouble, Up: Top

7 Compiler and Linker Performance

This chapter deals with speed of compilation and linking under DJGPP, and how they could be improved.

If you already know whether the compiler or the linker is the slow part, go to the appropriate section; if not, add -v to your GCC command line and run it again. With the -v switch, GCC will print all the programs it invokes, and you will be able to tell which one is taking most of the time.

Node: Slow compiler, Next: , Previous: Compiler performance, Up: Compiler performance

7.1 Slow Compilation

Q: Why GCC is compiling sooo slooowww?

A: That depends on what you mean by "slow". The following table gives "normal" gcc compilation speed, in source lines per second, on a 166-MHz Pentium:

Source language Without optimizations With -O2
C++ 800 400
C 1800 1000

Note that the numbers for compilation with -O2 are about 30% slower for GCC 2.95 and later versions than for previous versions. This is because GCC now does much more optimizations under -O2 than previous versions did.

As another data point, compiling the Allegro library takes about 3 minutes on a P500 and about 50 minutes on a 486/DX2-66.

On machines faster or slower than P166, scale these numbers accordingly. For example, 486/DX2-66 is about 4 times slower than P166. When comparing to this table, don't forget to count header files your program #includes in the total line count. And don't check compilation speed on very short programs (like the classic Hello, world!), because the overhead of loading the multiple passes of the compiler will completely hide the compiler performance. It is also useful to run the compilation twice in succession, especially if you have a disk cache installed, to prevent the overhead of the first load from skewing the results.

If your results are close to these (deviations of a few percent are considered "close" here), then that's as fast as you can get with GCC. If they are significantly slower, you may indeed have a problem; read on.

First, check to see if GCC pages to disk when it compiles. This is manifested by a heavy disk traffic which won't go away even if you have a large write-back disk cache installed. To be sure, disable the virtual memory services for your DPMI host (for CWSDPMI, load it before your program with the -s- switch, or use the CWSPARAM program to point the swap file to a non-existent drive), or use CWSDPR0 or PMODE/DJ as the DPMI host, and then run the compilation again; if the compiler aborts with an error message saying there isn't enough memory, then it was paging in your original environment.

If paging does happen, you need to free more extended memory. If you have a RAM disk, make it smaller, or don't use it at all (it only makes compiles run about 20% faster), or make your disk cache smaller (but don't discard the disk cache altogether); if you have other programs which use extended RAM, make them use less of it. Failing all of the above, buy more RAM (see the description of reasonable configuration). Also see recommendations for optimal software configuration.

If GCC doesn't page, check the settings of your disk cache. If you don't use a cache, install one--this can slash your compilation times by as much as 40%, more so when compiling a large number of small files. If you already have a cache, enable its delayed-write (a.k.a. write-back, a.k.a. staggered-write) operation. Some people disable the delayed-write feature for safety reasons, to avoid losing files due to system crashes. If you are worried about this, you can usually gain performance without sacrificing safety by enabling delayed-write together with an option that causes the cache to flush the write-behind data before the system returns to the DOS prompt. (For SmartDrv disk cache, this is achieved by specifying /N/F switches instead of /X.) GCC usually gains a lot when you set up your cache in such a way, because each compiler pass (pre-processor, compiler, assembler) must write temporary files that are used by the following passes.

It is also worthwhile to check the settings of your system BIOS. In particular, the following items should be checked against your motherboard vendor recommendations:

Internal and external CPU cache set to Enable
CPU cache scheme set to Write-back, if possible
DRAM and SRAM wait states vendor-recommended optimal values

Incorrect or suboptimal settings of the above items can explain as much as 30% performance degradation on 486 machines, and as much as 500% (!) if you have a Pentium CPU.

Node: Slow linker, Previous: Slow compiler, Up: Compiler performance

7.2 Slow Linking

Q: The compiler finishes in a few seconds, but then the linker grinds away for more than a minute, even on a very short program....

A: Try linking the trivial Hello, world! program; it should take no more than 7-10 seconds on a 486, 3-5 seconds on a Pentium. If you see much slower linking on your system, then the following advice might help you.

If you use a disk cache, make sure you enable its write-back (a.k.a. delayed-write) operation. Some people disable the delayed-write feature for safety reasons, to avoid losing files due to system crashes. If you are worried about this, you can usually gain performance without sacrificing safety by enabling delayed-write together with an option that causes the cache to flush the write-behind data before the system returns to the DOS prompt. For SmartDrv disk cache, this is achieved by specifying /N/F switches instead of /X.

For very large (several MBytes) executables which are built from a large number of small source files, the link (as opposed to the compilation) stage might be the one which needs more RAM than you have free, and thus be the bottleneck of the time it takes to build your program. Check that the size of the executable isn't larger than the amount of your free RAM. If it is, then it might make sense to use a smaller (or even no) disk cache, and allow the linker as much physical RAM as it needs. Make sure that the linker wasn't stub-edited to make its transfer buffer too small.

The first release of GCC 2.95 ported to DJGPP had a bug in the collect2 program (used during the link stage) whereby if a root directory of any drive appeared in the PATH environment variable, collect2 would try to look for files with names like C:\/foo, which caused Windows 9X to search the network (because two slashes in a row would look like a network share name). This would create an illusion of a very slow link, when in fact collect2 simply waited for the network operation to time out.

Another reason for slow linking might be that the DJGPP.ENV file by default sets TMPDIR to a tmp/ subdirectory of the main DJGPP installation directory; if DJGPP is installed on a networked drive, this means all your temporary files go back and forth through the network (and networked disks are usually not cached on your PC). In such cases, setting TMPDIR to a directory on your local drive, or to a RAM disk, would probably make linking faster.

DJGPP can be slow if it is installed on a networked drive. Generally, this is not recommended. If you have to, you should configure your network interface for best performance. Consult your network administrator.

Node: Compiling, Next: , Previous: Compiler performance, Up: Top

8 Compile-time and Link-time Problems

Being of a Unix origin, GCC has a somewhat different flavor of command-line syntax and its peculiar compilation and link algorithms. It also has a plethora of optional switches, some of them obscure or semi-documented. These are known to confuse users, especially those who had previous experience with DOS-based C compilers.

This chapter explains how to solve some of those problems which tend to appear when compiling and linking your programs.

Node: No input files, Next: , Previous: Compiling, Up: Compiling

8.1 GCC doesn't find the source files

Q: I created a simple source file hello.c, but when I invoke the compiler, it says: "gcc.exe: hello.c: No such file or directory", and then exits with the message "No input files." But hello.c is there, so why won't the compiler find it??

A: One popular reason for this problem is that you use one of those Windows editors that think they know better how do you want them to name the files. For example, Notepad always attaches the .txt extension to the file name you provide, so when you type hello.c into the dialog box, Notepad actually creates hello.c.txt. In addition, the files listed by My Computer by default have their extensions not shown, which creates an illusion that hello.c really is there.

Use the DIR command in the DOS Box to see what files are in the directory where you run GCC. (If you have the GNU Fileutils installed, you can use ls as well.) This will always show the full names of the files, exactly like GCC sees them.

You are generally advised to stay away of such "helpful" editors. Notepad is not suited well for editing programs, anyway. If you must use it, a work-around is to type the file name in quotes: "hello.c"; then Notepad will leave it alone and not append the .txt extension.

Another reason for GCC to not be able to find the source file is because you use long file names on Windows/NT. Suppose you invoke GCC like this: 

 gcc -c file_name.c

The name file_name.c exceeds the DOS 8+3 limits, so if you have such a file, you probably created it with some Windows editor. However, DJGPP programs cannot access long file names on Windows/NT, so gcc doesn't find such a file and complains.

Type dir /x from the command line to see the short 8+3 alias name of your file (in the example above, it should be file_n~1.c or some such), and use that short name when you invoke GCC. In general, if you want to avoid such problems on Windows/NT, you should restrict yourself to file names that are valid DOS 8+3 names.

Node: Missing headers or libraries, Next: , Previous: No input files, Up: Compiling

8.2 GCC can't find headers or libraries

Q: Why does GCC complain that it cannot open -lstdcx?

Q: When I run the compiler it says it couldn't find header files and/or libraries. But the headers and libraries are all there, so why won't it find them?

Q: When I link my programs, ld.exe complains that it cannot open crt0.o, although that file exists in the lib subdirectory....

Q: I tried to compile a program, but GCC complained about missing header files netdb.h and socket.h. Can somebody please mail me those headers?

Q: Why does GCC complain that it "cannot open -lgcc: File format not recognized"?

A: An error message about missing -lstdcx usually means that the linker cannot find the standard C++ library, libstdcxx.a (it is truncated to libstdcx.a on DOS and NT systems). Look into your lib/ subdirectory to see if it's there; if not, unzip it from the gppNNNb.zip file.

If libstdcxx.a exists but the linker still complains, you most probably have a problem related to long file names on Windows 9X (libstdcxx.a exceeds the DOS 8+3 limits). For a quick fix, try to set LFN=y in the environment and see if that helps. If that doesn't help, make sure you unpacked gppNNNb.zip with an unzip program which supports long file names.

This issue is further complicated if you use RHIDE v1.4, and is described in full in the file gnu/gcc-2.95/readme.DJGPP which comes with the GCC distribution (and which you should have read before the installation). Bottom line is that you need to add a line either to rhide.env (the RHIDE distribution includes a file rhide_.env which you should rename) or to DJGPP.ENV which says this: 

 RHIDE_TYPED_LIBS_DJGPP.cc=stdcxx
 RHIDE_TYPED_LIBS_DJGPP.cxx=stdcxx
 RHIDE_TYPED_LIBS_DJGPP.cpp=stdcxx

When you add these lines, make sure neither they nor the [rhide] line have trailing whitespace, otherwise RHIDE will not recognize these lines.

DJGPP version 2.03 and later come with these lines in the DJGPP.ENV file right out of the box.

RHIDE v1.4.7 and later solves this bug, so upgrade to the latest version if you can.

See C++ headers not found, for similar problems specific to C++ header files.

In general, if the compiler complains about missing files, you need first to find out whether they at all exist on your system. For C header files, look in the include directory and its subdirectories; for C++ header files, look in the lang/cxx directory and its subdirectories; for libraries and the crt0.o file, look in the lib directory. Some header files and object files which are specific to a certain GCC version unzip into the lib/gcc-lib/djgpp/X.YZ directory (where X.YZ is the GCC version number, e.g. 2.95), so look there as well.

If a header file indeed is not there, and you cannot find it in the djdevNNN.zip and gppNNNb.zip distributions, it probably means that this header belongs to a package which isn't part of the basic DJGPP distribution. You need to find that package and install it. It is important to understand that if a package is missing, getting hold of the header files like socket.h is not enough: you need the library of the functions whose declarations and prototypes are in the header. For socket.h, you need a sockets library, such as libsock; see DJGPP packages. For graphics.h, you need GRX and the Borland-to-GRX interface, BCC2GRX (rename the file libbcc.h to graphics.h); see BCC2GRX interface package.

If the header or the library does in fact exist on your machine, then in order for the compiler to find them, you should have the following variable set in your environment13: 

 set DJGPP=c:/djgpp/djgpp.env

and it should point to the correct path of the file DJGPP.ENV on your system; the file itself is in djdev203.zip in the DJGPP distribution. In the above example it is assumed to be in the C:\DJGPP directory, but you should set it as appropriate for your installation.

Many of the problems with "missing" files, including the highly-confusing message about -lgcc ("File format not recognized"), are usually caused by having the DJGPP variable set incorrectly. The following describes some problems with defining DJGPP which pop up frequently on the DJGPP forum.

Sometimes, people make errors in their AUTOEXEC.BAT that cause the DJGPP variable to be defined incorrectly, or not defined at all (some of the more common errors are listed below). To check what is the actual setting, type from the DOS prompt: 

 set > env.dat

then examine the contents of the file env.dat. You should see there a line like this: 

 DJGPP=c:/djgpp/djgpp.env

If a line such as this isn't there, you should investigate the cause for this (see below for some of the possibilities).

Many problems with setting DJGPP happen when people put excess blanks around the = character, which has the effect of defining "DJGPP " (with the blank) which is not the same as "DJGPP" (without blanks). You should make sure there are no such excess blanks, or DJGPP won't find its files.

Another possible cause of DJGPP variable not being set is that you invoke another batch file from your AUTOEXEC.BAT before the line that sets DJGPP. Make sure such batch files are invoked with the CALL statement, because otherwise the subsidiary batch file will never return to process the rest of AUTOEXEC.BAT (that's a "feature" of DOS batch file processing).

The code that processes DJGPP.ENV assumes that this file resides in the main DJGPP installation directory. If that assumption is wrong, the compiler (and some other DJGPP programs) might fail to find some of the files or auxiliary programs they need. Do NOT move DJGPP.ENV to any other directory!

Note that if you run DJGPP on Windows/NT, you cannot use long names of the directories in the pathname of DJGPP.ENV when you set the above variable in the environment; you should use their 8+3 aliases instead. That's because Windows/NT doesn't support the LFN API for DOS programs, so the DJGPP startup code won't be able to find the DJGPP.ENV file using its long pathname. For example, the following setting won't work on Windows/NT because Development is longer than 8 characters: 

 set DJGPP=c:/Programs/Development/Djgpp/djgpp.env

If the DJGPP variable is set correctly, then check the following possible causes of this misbehavior:

Node: Missing C++ headers, Next: , Previous: Missing headers or libraries, Up: Compiling

8.3 GCC can't find C++ headers

Q: I installed all the packages, but GCC complains it can't find iostream.h, _string.h and other C++ headers. Where can I find those header files?

Q: GCC complains about being unable to find Complex.h, Regex.h and other header files which start with a capital letter, and I indeed don't see them in my lang/cxx/ directory. Where are they?

Q: My C++ program needs header files whose filenames exceed the 8+3 DOS filename restrictions, like stdiostream.h and streambuf.h, and GCC cannot find those files. How in the world can I write portable C++ programs??

A: All C++ include files are packaged as part of the GNU C++ compiler distribution zip file, so if you didn't install it, GCC won't find them. Files whose names usually start with a capital letter, on MS-DOS have an underscore _ prepended so they can be distinguished from complex.h, regex.h and the like under case-insensitive DOS. Change Complex.h to _Complex.h, and String.h to _String.h in your source, and GCC will find them. The same goes for the header iostreamP.h--you should use _iostreamP.h instead. If you don't have the underscore _ on your keyboard, you might find using strclass.h instead of _String.h easier.

Another possibility to handle header files like Complex.h in a portable way is to pass the -remap switch (supported by GCC 2.8.0 and later) to the pre-processor; see the cpp docs and the readme.DJGPP file in the GCC distribution, for more info about this feature.

The most probable cause of problems with header files whose names exceed the DOS 8+3 limits is that you are compiling on Windows 9X, but the Long File Names (a.k.a. LFN) support is disabled. DJGPP v2.01 comes with LFN disabled by default on the DJGPP.ENV file. To enable it, set the environment variable LFN to y, like this: 

  set LFN=y

If the problems with long names of header files aren't solved by this, it is possible that you unpacked the DJGPP distribution with a program which doesn't support long file names. The solution is to install DJGPP again using a different unzip program. unzip32.exe, available from the DJGPP sites, is one possibility.

Some users copy the DJGPP directories after unzipping to another place on their disk, or backup and restore them. If this is done by some program that doesn't support long file names, the compiler won't be able to find header files such as strambuf.h. Editing the directory with some disk-editing tool that doesn't support Windows 9X style long file names can also cause such loss of long file names: when Windows 9X starts up, it checks whether the long file names and their 8+3 aliases are in sync, and if they aren't, the long file names are deleted from the directory, leaving you only with the short file names such as stream~1.h. Type dir *.h to see what are the long file names in the directory; the long names are printed on the right side of the file listing, and the short aliases on the left side, like this: 

 stream   h           1,925  12-26-95  8:07p STREAM.H
 stream~1 h          17,020  01-24-96  2:11a streambuf.h

(The files' date, time, and size might be different in your case.) The easiest solution for these cases is to remove the entire DJGPP installation, and unzip everything again.

Another possible cause for lack of support for long file names is that you switch to the so-called "DOS Mode" when running DJGPP programs from Windows 9X. This unloads from memory most of Windows, including the VFAT Filesystem module that supports the LFN API used by DJGPP to access long file names. The solution is to make sure your DOS box's Properties don't force a switch to "DOS Mode".

If you have problems with header files with long filenames, and you run under Windows NT, it usually means that you used an unzip program which supports long file names on NT; unzip again using a DOS unzip program, such as unzip32.exe that is available from the DJGPP sites. Alternatively, you could install an LFN driver for Windows NT, see LFN driver for NT, earlier in this FAQ.

Another possible cause for problems with C++ include files is that your source file has a .c extension. GCC then thinks that this is a C program and doesn't instruct the pre-processor to search the include directories specific to C++. Rename your file to .cc or .cpp extension, or call GCC with the -x c++ switch, and the header files will be found. A full list of extension rules which GCC uses to determine the source language can be found in the list of language-specific suffixes, elsewhere in this FAQ.

Node: C++ comments, Next: , Previous: Missing C++ headers, Up: Compiling

8.4 GCC barfs on C++-style comments in C programs

Q: My C program compiles OK with Borland's C, but GCC complains about "parse error before `/' " at a line where I have a "//"-style comment.

A: That's because // isn't a comment neither in ANSI C nor in K&R C. Borland and Microsoft C compilers support it as an extension. GCC also supports this extension (beginning with version 2.7.0), but using the -ansi or -traditional switches to GCC disables this extension. In general, it's a bad practice to use this extension in a portable program until such time as the ANSI C standard includes it. If it's a C++ program, then rename it to have a suffix which will cause gcc to compile it as such (see list of language-specific suffixes), or use -x c++ switch. If it's a C program, but you want to compile it as C++ anyway, try -x c++; it can help, but can also get you in more trouble, because C++ has its own rules. For example, the following program will print 10 if compiled as a C program, but 5 if compiled as C++14: 

    #include <stdio.h>

    int
    main ()
    {
      printf ("%d \n", 10    //*
		     / 2    // */ 1
		       );
      return 0;
    }

If you must have both -ansi and C++-style comments, use -lang-c-c++-comments preprocessor switch. Gcc doesn't accept the -lang-XXX switches on its command line, so you will have to use the -Wp option, like this: 

 gcc -c -Wp,-lang-c-c++-comments myprog.c

Alternatively, you can add -lang-c-c++-comments to the *cpp: section of your lib/specs file (but that will cause gcc to pass it to cpp unconditionally).

Bottom line: until the future ANSI/ISO C standard includes this as part of the C language, it's best to change those // comments to C-style ones, if you really mean to write a C program. The following SED command will convert a C program with C++-style comments into a valid C source, provided you don't have the string "//" in a character string: 

 sed "s?//\(.*\)?/*\1 */?" file.c > newfile.c

SED can be found with the DJGPP archives on SimTel.NET, in the v2gnu directory.

If you want the compiler to print a warning message about usage of //-style comments in a C program, add the -ansi -pedantic options to the GCC command line. If you don't want to use -ansi for some reason (e.g., because it rejects some other code that you want to keep), try using -Wp,-lang-c89 instead; this tells the preprocessor to stick to the rules of the C89 standard.

Node: Which language, Next: , Previous: C++ comments, Up: Compiling

8.5 How does GCC recognize the source language?

Q: I type GCC PROG.CC and GCC complains that it can't recognize PROG.CC's file format. How come a C++ compiler doesn't recognize a C++ source??

Q: I type GCC PROG.C to compile a C program which I already remember to pass compilation without a single warning, and suddenly it gives all kinds of strange error messages and unresolved externals.

A: That's because you typed your source file extension in UPPER case. GCC is not case-insensitive about filenames like DOS is, and it uses the file's extension to determine how to compile a file. Valid extensions are:

.cc
.C
.cxx
.cpp
C++ source (passed through cpp).
.c
C source that must be passed through cpp first.
.i
Raw C source (no cpp pass).
.ii
Raw C++ source (not to be preprocessed).
.m
Objective-C source.
.S
Assembler that must be passed through cpp first.
.s
Raw assembler source (no cpp pass).

Any other file is passed to the linker, under the assumption that it's an object file.

In the examples above, PROG.C is taken as a C++ program, not a C one, and PROG.CC is passed to the linker as if it were an object file. You can see what GCC does by adding the -v switch to the GCC command line; if you see that it's invoking cc1plus.exe (the C++ compiler) instead of cc1.exe (the C compiler), or calling ld.exe (the linker) on a source file, then you'd know this is your problem. If you have problems keeping up with the verbose GCC output triggered by -v, see how to capture GCC output, earlier in this FAQ.

You can override the default rules gcc uses to decide how each input file should be treated, using the -x language switch. For instance, the command 

 gcc -x c++ prog.c

compiles prog.c as C++ source. See The GNU C Compiler Manual, for more info on -x options.

Node: Objective C, Next: , Previous: Which language, Up: Compiling

8.6 Problems with Objective C

Q: How do I tell gcc my .cc file is to be compiled as Objective-C source?

Q: I compile an Objective-C program, but get unresolved symbols.

Q: I can't compile the Objective-C test program which came with DJGPP.

A: Give your sources the .m extension, or use -x objective-c switch to GCC, so it will know you mean to compile with Objective C.

Node: DJGPP-specific, Next: , Previous: Objective C, Up: Compiling

8.7 Writing codes fragments which are specific to DJGPP

Q: I must put a DJGPP-specific code fragment into my program. What symbol should I use in the #ifdef directive to make it only visible under DJGPP?

A: Use __DJGPP__, like this: 

    #ifdef __DJGPP__
    ... DJGPP-specific code ...
    #else
    ... not seen under DJGPP ...
    #endif

__DJGPP__ has the value of the DJGPP major revision number, so you can write code fragments which have different behavior under different versions of DJGPP: 

    #ifdef __DJGPP__
    #if __DJGPP__ > 2
    .... will work only in DJGPP v3.x and later ...
    #else
    .... get here for DJGPP v2.x ...
    #endif
    #else
    .... get here in DJGPP v1.x or non-DJGPP environment
    #endif

If you need to distinguish between minor DJGPP revision numbers, use the symbol __DJGPP_MINOR__. For example: 

    #if defined(__DJGPP__) && __DJGPP__ == 2 && __DJGPP_MINOR__ == 1
    .... will work only in DJGPP v2.01 ....
    #endif

Another DJGPP-specific pre-processor symbol which DJGPP defines is __GO32__; but it is only provided for compatibility with previous versions of DJGPP (v1.x) and its use should be discouraged.

Node: Unresolved externals, Next: , Previous: DJGPP-specific, Up: Compiling

8.8 Undefined references when linking programs

Q: Why do I get so many undefined references when linking my programs?

Q: Why do I get "Undefined reference to yywrap" when linking programs produced by Flex?

Q: GCC complains that it cannot find -liostream. Where can I find this library?

A: By default, GCC instructs the linker to only look in two libraries: libgcc.a and libc.a. Some functions aren't included there, so the linker can't find them. If you need to link against some optional library, say libxy.a, put the library into the DJGPP lib/ subdirectory and append a -lxy to the link command line. The Standard C++ Template classes are in libstdcxx.a (it's called libstdc++.a on Unix); append -lstdcxx. To use the additional GNU C++ classes in the libgpp.a library (it's called libg++.a on Unix systems), append -lgpp. Flex-generated lexical analyzers call functions in the libfl.a library; you need to append -lfl when linking them. Append -lgrx if you are using GRX library, and -lalleg for linking with Allegro.

When linking C++ programs, you should use either the gpp or gxx commands instead of gcc; they will then instruct the linker to also scan the C++ libraries automatically, so you don't have to remember doing that yourself.

Another reason for undefined references when linking C++ programs is that you mix GCC and libstdcxx.a from different releases: they are usually incompatible. In particular, sometimes people install an additional compiler based on (old releases of) GCC, such as GNU Pascal or GNAT, the GNU Ada development environment, and these additional compilers overwrite some of the libraries, like libgcc.a, with older and incompatible versions. You should always make sure you don't mix different releases of the compiler and libraries; if you must install different releases, install them in separate directories and prepare some batch files or shortcuts to set up the environments for each of the compilers by pointing the DJGPP variable to different directories and changing the order of directories in PATH.

If your program uses a lot of floating-point math, or needs math functions beyond those specified in the ANSI/ISO standard, consider appending -lm to your link command line. The basic math functions required by ANSI/ISO standard are included in the libc.a library, but libm.a includes different versions of these functions which sometimes are more accurate or more compatible with widely-accepted standards for numeric computations; libm.a also includes some functions not included in the default library, like Gamma function and Bessel functions, support for different standards of behavior in case of errors, a matherr facility, etc.

Old C++ programs used to be built with the GNU iostream library, libiostream.a. The iostream classes are now part of the standard C++ library, libstdcxx.a, so if you come across a Makefile that passes the -liostream option to the compiler, change that to -lstdcxx instead.

Further problems which cause the linker to fail with C++ programs are discussed in listing libraries in the correct order, and in exceptions and inline functions.

Node: Which library, Next: , Previous: Unresolved externals, Up: Compiling

8.9 How not to lose your head with all these libraries

Q: I'm lost with all those different libraries. How in the world can I find out which functions are included in which library?

A: You can use the nm program to check what functions are included in a library. Run it with the -C option and with the library as its argument and look in the output for the name of your function (the -C, or --demangle option makes the function names look closer to what they are called in the source file). Functions which have their code included in the library have a capital T before their name. For example, the following is a fragment from the listing produced by nm: 

    c:\djgpp\lib> nm --demangle libc.a
    .
    .
    .
    stdio.o:
    000000e4 b .bss
    000000e4 d .data
    00000000 t .text
    00000098 t L12
    0000001e t L3
    00000042 t L6
    0000004d t L7
    0000006a t L9
    00000000 t __gnu_compiled_c
	     U _filbuf
	     U _flsbuf
    00000000 T clearerr
    000000ac T feof
    000000c2 T ferror
    000000d8 T fileno
    0000000c T getc
    00000052 T getchar
    0000002a T putc
    0000007c T putchar
    00000000 t gcc2_compiled.
    .
    .
    .

Here we see that the module stdio.o defines the functions clearerr, feof, ferror, fileno, getc, getchar, putc and putchar, and calls functions _filbuf and _flsbuf which aren't defined on this module.

Alternatively, you can call nm with the -s or --print-armap, which will print an index of what symbols are included in what modules. For instance, for libc.a, we will see: 

    c:\djgpp\lib> nm --print-armap libc.a
    .
    .
    .
    _feof in stdio.o
    _ferror in stdio.o
    _fileno in stdio.o
    .
    .
    .

which tells us that the functions feof, ferror and fileno are defined in the module stdio.o.

nm is fully described in the GNU docs. See GNU Binutils Manual.

Node: Libraries order, Next: , Previous: Which library, Up: Compiling

8.10 DJGPP uses a one-pass linker

Q: I give all the libraries to gcc, but I still get unresolved externals when I link. What gives?

A: Ld is a one-pass linker: it only scans each library once looking for unresolved externals it saw until that point. This means the relative position of object files and libraries' names on the command line is significant. You should put all the libraries after all the object files, and in this order: 

 -lstdcxx -lm

E.g., to link files main.o and sub.o into a C++ library, use the following command line: 

 gcc -o main.exe main.o sub.o -lstdcxx -lm

or, if you compile and link in one command: 

 gcc -o main.exe main.cc sub.cc -lstdcxx -lm

If you have any libraries of your own, put them before the above system libraries, like this: 

 gcc -o main.exe main.cc sub.cc -lmylib -lstdcxx -lm

When you use the gpp or the gxx drivers to compile a C++ program, it automatically names the C++ libraries in the correct order. (gpp and gxx are the alternative names for g++ on DOS, which doesn't allow the + character in file names.)

You can also force the linker to repeatedly scan a group of libraries until all externals are resolved. To this end, put the names of these libraries between the -( and the -) options (if you invoke GCC to link, use the -Wl or -Xlinker options to pass switches to the linker). Check out the linker docs for more info about -( ... -) groups.

If your installation tree is different from the default, i.e., if you keep the libraries not in the default lib/ subdirectory, then you should add that directory to the line in the [gcc] section of your DJGPP.ENV file which starts with LIBRARY_PATH, or put into your environment a variable called LIBRARY_PATH and point it to the directory where you keep the libraries. Note that if you invoke the linker by itself (not through the gcc driver), then LIBRARY_PATH will have no effect, because this variable is only known to the gcc driver. Invoking ld directly is not recommended, but if you must do it, use the -L option to tell it where to look for the libraries.

Node: Still unresolved, Next: , Previous: Libraries order, Up: Compiling

8.11 Some functions in C++ programs still not found

Q: I put all the libraries in the above order, but the linker still can't find some C++ functions from complex.h and iostream.h.

Q: I can't compile a program which uses the String class: the linker complains about undefined functions!

Q: I get many undefined references to symbols like __eh_pc, terminate, and __throw....

A: Some C++ functions are declared inline and defined on header files. (One example of this is the string constructor of the GNU String class.) However, GCC won't inline them unless you compile with optimizations enabled, so it tries to find the compiled version of the functions in the library. Workaround: compile with -O2.

Another cause of missing external symbols might be that your versions of libgcc.a and the compiler aren't in sync. These cases usually produce undefined references to symbols such as __throw and __eh_pc. You should only use libgcc.a from the same distribution where you got the compiler binaries. The reason for these problems is that the setup for supporting C++ exceptions is subtly different in each version of the compiler.

For C++ programs, be sure to compile all of your object files and libraries with the same version of the compiler. If you cannot recompile some of the old C++ object files or libraries, try using the -fno-exceptions -fno-rtti switches to GCC, it helps sometimes. However, note that -fno-rtti cannot be used with GCC version 2.95 and later: programs that use exceptions will crash if compiled with this option.

If you call C functions from a C++ program, you need to make sure the prototype of the C function is declared with the extern "C" qualifier. DJGPP header files take care about this, but headers you get with third-party libraries, or write yourself, might not. Failure to use extern "C" will cause the linker to look for a C++ function instead of a C function, which will fail because names of C++ functions are mangled by the compiler (to include the types of their arguments, since there can be many functions with the same name but different argument types).

Yet another possible cause for the linker to complain about undefined references is that you link files compiled using RSXNTDJ headers, or link RSXNTDJ-compiled object files with DJGPP libraries, or vice versa. DJGPP and RSXNTDJ are really incompatible as far as header files and libraries are concerned, so you cannot mix them. Add -v to the gcc command line and watch the order of searching the include directories and libraries printed by GCC: if you are compiling/linking a DJGPP program, but the RSXNTDJ directories appear first in the search order, you should change the order of the directories in the C_INCLUDE_PATH, CPLUS_INCLUDE_PATH and LIBRARY_PATH environment variables. Similarly, if you are compiling an RSXNTDJ program, but the DJGPP's include and lib subdirectories appear first in the list printed by gcc -v, you need to put the RSXNTDJ directories first in the environment variables mentioned above.

Node: djgpp_first_ctor, Next: , Previous: Still unresolved, Up: Compiling

8.12 Unresolved djgpp_first_ctor

Q: I do everything like your praised FAQ says, but the linker complains about unresolved symbols with strange names like djgpp_first_ctor, djgpp_last_dtor, etc. I looked in every library with nm, and I cannot find these creatures. Where in the world are they??

A: These symbols are defined by the djgpp.djl linker script that should be in your lib/ subdirectory. When you call gcc to link a program, it invokes ld.exe with the option -T djgpp.djl. If you invoke ld directly (this is generally not recommended), be sure to include that switch. If you did invoke it through gcc, maybe your linker is set up incorrectly. Add -v to the GCC switches and check that the command line that GCC gives to LD includes that switch, that your lib/ subdirectory includes that script file, and that the script file is intact and includes the definition of the above symbols.

Another reason might be that you have edited your DJGPP.ENV file in a way that prevents the linker from finding its djgpp.djl script.

Mixing an old v1.x installation with a v2.x one can also cause such problems. Be sure to delete the entire v1.x tree, or rename it, before installing the v2.x distribution.

Node: Large image, Next: , Previous: djgpp_first_ctor, Up: Compiling

8.13 C++ programs yield large .exe file

Q: It seems that declaring a large static array has the effect of bloating the program image on disk by that many bytes. Surely there is a more compact way of telling the loader to set the next N bytes of RAM to zero?

A: This only happens in C++ programs and is a (mis-)feature of GCC. You can use the -fconserve-space switch to GCC to prevent this from happening, but that switch also turns off the diagnostics of duplicate definitions, which, if uncaught, might cause your program to crash. Thus, this switch isn't recommended for programs which haven't been completely debugged (if there is such a creature). The -fconserve-space switch is described in the GCC docs, see GNU C Compiler Manual.

If the problems with using this switch doesn't deter you, you can even add this switch to your lib/specs file to make it permanent.

GCC versions 2.95.1 and later don't have this problem, even in C++ programs.

Node: Large executable, Next: , Previous: Large image, Up: Compiling

8.14 Why are DJGPP .exe files so large?

Q: I compiled a trivial "Hello world" program and got a 280KB executable file. That's ridiculously bloated!

Q: I switched to GCC 2.95.1, and my C++ executables are considerably larger than when compiled with GCC 2.7.2.1!

A: Did you link with -s switch to gcc, or run strip on the output of the linker? If not, the executable includes the debugging symbols, which makes it quite a lot larger. (It is not recommended to strip the symbols except when distributing production programs, because this makes debugging very hard indeed; that is why -s is not passed to gcc by default.)

A stripped "Hello world" program written in C should be about 42KB on disk; an analogous program written in C++ should be about 140KB on disk (the additional overhead is due to the C++ classes that are linked in to support cout).

C++ programs could be further bloated because the release of Binutils 2.8.1 was configured in a way that caused the assembler to put into the symbol table local labels generated when compiling code that uses exceptions. Later uploads of GNU Binutils should solve this problem, so consider upgrading to the latest bnuNNNb.zip.

Some other compilers with which people keep comparing the size of DJGPP programs use shared libraries or DLLs, so the size of the executable doesn't include the libraries. If you have an immediate question "Why won't DJGPP use DLLs as well", read the following section.

In general, judging code sizes by looking at the size of "Hello" programs is meaningless, because such programs consist mostly of the startup code. The DJGPP startup code does many things in preparation for running a protected-mode program in a Posix-compliant environment. This includes switching the processor to protected mode (which requires a lot of code), wildcard expansion, long command-line support, and loading the environment from a disk file; these usually aren't available with other DOS compilers. Exception and signal handling (not available at all in v1.x), FPU detection and emulator loading (which were part of go32 in v1.x), are now also part of the startup code.

Most of the power of all these features goes wasted in "Hello" programs. There is no point in running all that code just to print a 15-byte string and exit. However, the overhead induced by the code needed to set up the DJGPP run-time environment is additive; the larger the program, the smaller the overhead relative to the program size. For non-trivial programs, the code produced by DJGPP is usually smaller than what other compilers produce. For example, the DJGPP version of the Povray program is smaller by about 200KB than the same program compiled with the Watcom compiler.

If your program doesn't need parts of the startup code, it can be made smaller by defining certain functions with empty bodies. These functions are __crt0_glob_function, __crt0_load_environment_file, and __crt0_setup_arguments. If you define empty substitutes for all three of these, and compile with -O2 -s, you can make the "Hello" program be 31KB on disk. These functions are documented in the DJGPP libc reference, which see. Here's an example of definitions for these functions that will make the startup code as small as it gets15: 

  #include <crt0.h>

  char **__crt0_glob_function (char *arg) { return 0; }
  void   __crt0_load_environment_file (char *progname) { }
  void   __crt0_setup_arguments (void) { }

(To do this in a C++ program, prepend the extern "C" qualifier to each one of the three lines that define the substitute functions.)

Note that if you define an empty substitute for __crt0_setup_arguments, your program will not be able to access its command-line arguments via the argv[] array. So this is only recommended for programs which don't accept any arguments at all.

You can make your program image still smaller by compressing it with a compressor called DJP. DJP is a DJGPP-specific executable file compressor. It is fast and has no memory overhead. It also knows about DJGPP Dynamically Loaded Modules (DLM) technology. (Note that DJP before version 1.06 was incompatible with Binutils 2.8.1 and later16, so you should always use the latest DJP version available on SimTel.NET mirrors.)

DJP is not actively developed anymore; its successor is the UPX compressor, currently in beta testing. UPX is written by the same people who wrote DJP, compresses better, and supports a broader class of executable formats, including DOS .exe, .com and .sys, DJGPP's COFF, Watcom's LE, Win32 PE, and Linux's ELF. UPX is available via the Web.

Node: DJGPP and DLLs, Next: , Previous: Large executable, Up: Compiling

8.15 Why don't we use DLLs to make programs smaller?

Q: Many other compilers use shared libraries and DLLs to make the programs' size smaller; why don't you do the same?

A: DLLs are really a mixed blessing, because they introduce an additional dimension into configuration management and subtle system differences.

Consider a DJGPP developer who is maintaining a package. Let's say that this developer finds a bug in one of the library functions that adversely affects his/her package. The obvious solution is to fix the library bug, and then relink the application against the fixed library. In the current setup, the only thing that the developer needs to do is to upload a binary distribution with the new executables, which will ensure that all users can reliably use the fixed program.

Now imagine that we ditch the static linking and instead distribute the standard libraries as DLLs to be installed on every machine. Our developer will now need to put a fixed DLL into the binary distribution, otherwise the package will still exhibit the bug when it uses an old DLL on the end-user's machine. Installing the fixed package would then need to overwrite the DLLs on users' machines with the fixed version.

But now suppose that the bugfix in that single library function made it subtly incompatible with other library functions, which the program of our developer didn't use (and so these problems went unnoticed during testing). When users will install the fixed DLL, they will now have a broken system, where programs which are unrelated to the upgraded package, and which worked perfectly before, would now mysteriously fail or crash.

Moreover, since there are quite a few DJGPP packages, all maintained separately by different individuals, and since users are installing the new-and-improved versions all the time, after some time there's no way to know what versions of system DLLs are installed on any given machine. The result is that no developer can be sure that their programs would work on any particular system, because the precise mix of the system DLLs on that system cannot be predicted. What you have is an environment where major programs constantly crash.

Sounds familiar? Of course! this is precisely the reason that most Windows systems are so unstable: people are constantly installing the hottest new versions of Office, IE, etc., and are constantly overwriting their system DLLs with new and subtly incompatible versions. If you can afford it, try to install Windows 9X and run it for a year without installing any add-on packages which come with a replacement for system DLLs--you will see a rock-solid system that can be run for weeks without crashing. (Yes, I actually tried that; yes, it really didn't crash.)

In addition, DLLs in the DJGPP environment make much less sense than on Unix or Windows, because the DJGPP DLLs will only be used by DJGPP program, whereas Unix shared libraries and Windows DLLs are used by the OS itself as well. When the OS itself uses the same libraries, the libraries are most of the time in memory when the applications need them, and running several applications only loads the library once. DJGPP cannot take advantage of this, even on Windows, because each DJGPP program runs in a separate Virtual Machine with a separate address space. The only case where DJGPP program can benefit from shared libraries is when one DJGPP program invokes another.

So bottom line, I think wasting some disk space due to static linking is much cheaper than having to deal with the user frustration and outcry that would result from using the DLL approach. Perhaps a behemoth such as Microsoft can afford ignoring all the mess that DLLs bring to the end users, but around here, a good name of the product still counts.

Node: No EXE, Next: , Previous: DJGPP and DLLs, Up: Compiling

8.16 Linker fails to produce the EXE program

Q: When I link my program, it fails to produce the .EXE executable....

Q: I run STUBIFY on a networked drive under Novell, but it doesn't produce a .EXE file. How come?

A: One possible reason for this is that your disk is full, or there's no swap space available for the DOS box on Windows. Run go32-v2 with no arguments and see what it reports, then follow the advice in configuring your system, for the optimal configuration.

If you are running DJGPP on a networked drive, you might have another copy of the file with the same name that GCC is creating in another directory somewhere on that networked drive. If that other directory is on your PATH, it is searched by Novell when the linker and STUBIFY try to create the executable file, because that file doesn't exist in the current directory. So what might actually happen is that the linker and STUBIFY are overwriting the files they find on your PATH instead of creating new files in the current directory.

You can verify that this indeed is the problem by searching your networked disks for files with the same name as those you are trying to build, and looking at their time stamps. If that is indeed the problem, then you have several possible ways of solving it:

  1. You can remove the other files, rename them, or move them to another directory that isn't searched by Novell.
  2. You can rename the program you are trying to link.
  3. You can change the way Novell searches for files (a.k.a. the search mode), so that it won't look in the directories on your PATH.
  4. You can change your access rights to the directory on the PATH where the other files reside, so that you won't have write privileges to that directory.
  5. You can change the search mode for STUBIFY and the linker (or for any other program that gives you that trouble) by running commands like these: 
      SMODE stubify.exe 2
  SMODE ld.exe 2

Node: Allegro and GRX, Next: , Previous: No EXE, Up: Compiling

8.17 Building Allegro or GRX library fails

Q: When I try to build the Allegro library, liballeg.a, I get some cryptic message about register-opcode mismatch. What am I doing wrong?

Q: Why do I get these messages saying "fixed or forbidden register 0 (ax) was spilled" when I try to build Allegro?

Q: It seems I miss one of the source files from the Allegro distribution, because Make cannot find it when I try to build Allegro.

Q: I can't build Allegro: it keeps telling me that I "need to install gcc2721b.zip". But I already have GCC installed!

A: You should get the latest version of Allegro that is available either from SimTel.NET or from Shawn Hargreaves' site.

Versions of Allegro before 3.0 are known to have bugs which triggered register-opcode mismatch messages.

GCC 2.95 became more picky about some invalid use of clobber specifiers in Allegro's inline assembly, so what compiled with GCC 2.8.1 won't compile anymore; latest versions of Allegro (3.12 and above) correct that.

GRX versions 2.3 and older also have a few places where the newer GCC releases won't compile the inline assembly code. Ian Miller created two patch files that solve two different classes of problems with GRX 2.3 inline assembly, and made them available from his Web page: patches for the clobber list problem and patch for indirect calls. You will need to use the patch utility to apply these patch files, and then recompile the offending library. A DJGPP port of the GNU patch is available from SimTel.

For a general explanation of how to correct clobber list specifications in inline asm code so that they will compile with GCC 2.95 and later, see the GCC FAQ list.

Node: NULL redefined, Next: , Previous: Allegro and GRX, Up: Compiling

8.18 C++ compiler says "NULL redefined"

Q: When I compile a C++ program which includes some standard C header files, the compiler prints error messages about redefinition of NULL....

A: This is because GCC 2.8.1 comes with C++ header files which redefine NULL in a way that conflicts with the DJGPP headers. It's a bug in the GNU C++ headers, but until it is fixed, you will need to make sure you include the C++ headers after the C headers. If that doesn't help in your case, you will need to hack your headers to reconcile them.

The C header files that come with DJGPP v2.02 work around this problem, so upgrading to the latest DJGPP release should make these messages go away.

Node: C++ exceptions, Next: , Previous: NULL redefined, Up: Compiling

8.19 C++ exceptions support

Q: I've written a program that uses C++ exceptions, but instead of catching an exception, the program prints "Abort!" and dies....

Q: When linking C++ programs, I get messages about undefined references to __EH_FRAME_BEGIN__ and such likes. Huh?

Q: I cannot compile C++ programs that include the header math.h: the compiler complains about redefinition of class exception!

A: C++ exceptions were not fully supported in DJGPP before version 2.8.1 of GCC. Either upgrade to the latest version or compile with the -fsjlj-exceptions switch to GCC. GCC support of exceptions before v2.8.0, was minimal, so even this special switch won't work with previous versions. If you still use GCC 2.7.2.1 and cannot upgrade, you need to compile with the -frtti compiler switch and include the typeinfo header in your program.

Beginning with EGCS 1.1.2 and GCC 2.95, C++ exception support requires DJGPP v2.02 or later, and will not work with v2.01 or earlier, so you might need to upgrade your DJGPP library.

Note that exception support with -fsjlj-exceptions is very slow, since it has a significant runtime overhead, even if the exception doesn't occur.

If you already use GCC 2.8.1, these problems could happen if you failed to replace the specs file with the version which comes with the GCC 2.8.1 distribution. Read the file readme.DJGPP in the GCC distribution, for more details. GCC 2.95 and later should work with the specs file from djdev202.zip (or later) or with the specs file that comes with GCC itself.

Exception support in GCC is generally not stable enough yet, so you need to treat with some suspicion code produced by GCC 2.8.1 for programs that use exceptions. Latest versions of GCC support exceptions better, so upgrade to GCC 2.95 or later.

Undefined references to symbols like __EH_FRAME_BEGIN__ are a symptom of using an old linker script djgpp.djl. You should make sure that djgpp.djl in your lib subdirectory is from djdevNNN.zip file that belongs to DJGPP v2.02 or later. (GCC 2.8.1 distribution required to replace djgpp.djl with a version that came with the compiler, but the reason for that is no longer valid with newer GCC versions, and the compiler no longer comes with djgpp.djl. So you must restore djgpp.djl from djdevNNN.zip.) Again, readme.DJGPP in the GCC distribution has more on this.

If GCC complains about "Redefinition of class exception" when you compile C++ programs which include the header math.h, you need to replace that header. GCC 2.8.1 comes with a header exception that conflicts with math.h from DJGPP v2.01, which defines a struct exception. Version 2.02 of DJGPP corrects its math.h, but if you still use v2.01, a corrected version is included in the gcc281b.zip distribution. The corrected math.h is installed into the lib/gcc-lib/djgpp/2.81/include directory, so either delete or rename the old version in the include directory, or copy the corrected version into include. Another solution is to compile with the -posix or -ansi compiler switch, which cause math.h to not define struct exception.

Node: Assembly output, Next: , Previous: C++ exceptions, Up: Compiling

8.20 How to get GCC to generate assembly code

Q: How can I peek at the assembly code generated by GCC?

Q: How can I create a file where I can see the C code and its assembly translation together?

A: Use the -S (note: capital S) switch to GCC, and it will emit the assembly code to a file with a .s extension. For example, the following command: 

  gcc -O2 -S -c foo.c

will leave the generated assembly code on the file foo.s.

If you want to see the C code together with the assembly it was converted to, use a command line like this: 

 gcc -c -Wa,-a,-ad [other GCC options] foo.c > foo.lst

which will output the combined C/assembly listing to the file foo.lst.

If you need to both get the assembly code and to compile/link the program, you can either give the -save-temps option to GCC (which will leave all the temporary files including the .s file in the current directory), or use the -Wa,aln=foo.s option which instructs the assembler to output the assembly translation of the C code (together with the hex machine code and some additional info) to the file named after the =.

Node: movedata.h, Next: , Previous: Assembly output, Up: Compiling

8.21 What's wrong with sys/movedata.h?

Q: Whenever I try to compile a program that includes the sys/movedata.h header file, I get "parse error" messages from the compiler. Can't you guys make your system headers right?

A: This is a bug in the sys/movedata.h header file which comes with DJGPP v2.01. The bug is fixed in v2.02, but if you are stuck with v2.01, you should always include the sys/types.h header before sys/movedata.h in your programs.

Node: Libraries, Next: , Previous: movedata.h, Up: Compiling

8.22 How do I create a library of object files?

Q: I would like to distribute my package as a library that can be linked into programs, but I'm unsure how to go about it....

A: First, you need to compile all your sources into object .o files, like this: 

 gcc -c -Wall -O2 file1.c
 gcc -c -Wall -O2 file2.c
 gcc -c -Wall -O2 file3.c
 ...

The only GCC switch in this example that's required is -c, the rest are just recommended for better code generation and diagnostics.

Once you have the object files ready, use the ar ("Archiver") utility to create a library, let's call it libacme.a, like this: 

 ar rvs libacme.a file1.o file2.o file3.o ...

The rvs flags tell ar to put named files into the library, replacing any previous versions of these files if necessary, print the names of object files as it puts them into the library, and add an object-file index to the library, which makes it link faster.

If you use RHIDE, you can create a library by specifying a file with a .a extension as the main target in the project (choose Project | Main Target Name and enter a file name such as libacme.a).

The library is now ready to use. The simplest way to force the compiler to use it while linking is to mention its name in the link command line, like this: 

 gcc -o myprog.exe myprog.c libacme.a

This is better than just listing in the command line all the object files in the library, since the latter will cause the linker to link in all the object files, even those which aren't used by the program.

The name of the library which begins with a lib and ends with a .a extension is a convention used for convenience. When the link command line includes an argument -lXXYYZZ, GCC (and all Unix compilers) will look for a file libXXYYZZ.a in every directory they search by default. So, if your library libacme.a is installed in the DJGPP lib subdirectory, the user can instruct GCC to look into it by appending -lacme to the link command line. Other systems might be configured to look for different names when a switch such as -lfoo is mentioned. For example, Linux might look in /usr/lib for files libfoo.so.*, while Alpha/VMS will look for SYS$GNU:[LIBRARIES]FOO.LIB;*. Windows 98, of course, will look for something monstrously long like C:\Windows\Program Files\Vendors\GNU\gcc\libraries\foo.lib. If you don't follow this convention, you will need to type the full name of the library file.

If you need to update a certain object file in a library, use the same command ar rvs library-name object-name as above, but only with the name(s) of the object file(s) you need to replace.

ar is documented in the Binutils docs. To read, type this from the DOS prompt: 

 info binutils ar

Node: No stubify, Previous: Libraries, Up: Compiling

8.23 GCC Cannot find stubify.

Q: Whenever I try to compile something, GCC says "Installation problem, cannot exec stubify: No such file or directory (ENOENT)". The compiler came on a CD with one of those "Teach yourself C++" books....

A: Blame the vendor who created the CD: their installation program failed to copy the program stubify.exe to your hard disk. The compiler needs stubify.exe when it links your programs.

To solve the problem, find stubify.exe on the CD and manually copy it into the same directory where gcc.exe lives.

Node: Running, Next: , Previous: Compiling, Up: Top

9 Running Compiled Programs

This chapter discusses various problems which may happen when running DJGPP programs under different environments, and gives solutions to them.

Node: v2 crash, Next: , Previous: Running, Up: Running

9.1 My program crashes only in v2.0!

Q: My v2 program crashes, but only under CWSDPMI; it runs OK under other DPMI hosts like Windows, OS/2 or QDPMI. Is this a bug in CWSDPMI?

A: No, it probably is a bug in your program which just goes unnoticed on Windows. Unlike other DPMI hosts, CWSDPMI supports some DPMI 1.0 extensions which allow DJGPP to capture and disallow dereference of pointers which point to addresses less than 1000h (a.k.a. NULL pointer protection). The tell-tale sign of these problems is a message "Page fault at ..." that is printed when a program crashes, and an error code of 4 or 6. The NULL pointer protection feature can be disabled by setting the _CRT0_FLAG_NULLOK bit in _crt0_startup_flags and recompiling the program; if this makes SIGSEGV crashes go away, your program is using such invalid pointers; the stack trace printed when the program crashes should be a starting point to debug this. See how to debug SIGSEGV, for more details about debugging these problems.

To make spotting uninitialized memory simpler, you can set _crt0_startup_flags to _CRT0_FLAG_FILL_DEADBEEF (don't laugh!); this will cause the sbrk()'ed memory to be filled with the value 0xdeadbeef (-559038737 in signed decimal or 3735928559 in unsigned decimal) which should be easy to spot with a debugger. Any pointer variable which has this value was used without initializing it first.

An insufficient stack size can also be a cause of your program's demise, see setting the stack size, below.

Node: malloc crash, Next: , Previous: v2 crash, Up: Running

9.2 Programs that crash in malloc or free.

Q: Since I upgraded to DJGPP v2.02, my program started to crash, and the traceback points to library function free. This program worked flawlessly with v2.01, so I guess there's a bug in the new version of free, right?

A: Such problems are a tell-tale sign of programs that overwrite buffers allocated by malloc or calloc, or call free more than once with the same pointer, or pass to free a pointer that didn't originate from a call to malloc or calloc. If the program that crashes is a C++ program, you might have several objects that share the same data, and the object destructor crashes when it calls free several time with the same memory chunk.

These crashes happen inside the memory-allocation functions because these functions maintain some crucial information about the allocated and free memory blocks right before the beginning and beyond the end of the allocated buffers. For speed considerations, this information is not protected by any means like CRC or parity, so if you overwrite this information, malloc and free will become confused and eventually will blow up.

The version of malloc in DJGPP library before v2.02 left some slack space beyond the end of the allocated buffer (this was a side-effect of the algorithm it used, which was optimized for speed, but wasted some memory). Thus, a program could overrun the allocated buffer and still get away uncaught. The new version of malloc introduced with v2.02 doesn't waste memory, and because of this is much less tolerant to such bugs.

Bottom line: you should debug your program to find the offending code that overwrites the end of an allocated buffer. One way of doing that is to put a data breakpoint (a.k.a. watchpoint) inside a debugger at the address which gets overwritten; then, when the program overwrites it, the debugger will kick in and you will see whodunit.

Another possibility to debug such problems is to use the YAMD package, written and maintained by Nate Eldredge. YAMD is a malloc debugger which will catch and report many problems related to allocating, freeing, and using allocated memory. YAMD is available from Nate's home page.

Node: Crash traceback, Next: , Previous: malloc crash, Up: Running

9.3 The call stack traceback

Q: My program dies with a cryptic message like "SIGSEGV" or "Page Fault" or "General Protection Fault" and prints some funny-looking numbers. Can't I get some decent human-readable traceback information, so I could pinpoint where in the program did the problem happen?

A: Those "funny-looking numbers" are the traceback. They describe the sequence of function calls which led to the fatal error by giving you the addresses where each function was called. You can have these addresses translated to source line numbers by using the SYMIFY program. SYMIFY is included in the basic DJGPP development environment distribution, and should be in your bin/ subdirectory. To symify the traceback, make sure that your program was compiled with the -g switch, linked without the -s switch and not stripped of its debugging symbols by running the strip utility. Now invoke your program and do whatever it takes to make it crash. Then, with the traceback still on the screen, type this from the DOS command line: 

 symify program-name

(Note: program-name should include the .exe suffix.) SYMIFY then walks through the crash traceback by reading it from video memory, and matches the hex addresses to the source files and line numbers of the program. It then writes back the list of source files and line numbers right next to their hex addresses. Now you can start debugging. More info about this is available in how to analyze crash dumps.

One problem with this translation is that it relies on info generated by GCC that maps the instruction addresses to source line numbers. This usually works okay, but one notable exception is when you use inline assembly. In this case, GCC only records the last line of the inline assembly block, which might be way off if the block is large.

You can ask SYMIFY to put the stack trace into a file (so you can consult it later, e.g., from your editor while fixing the bug), by giving it an output file, like this: 

 symify -o problem.dmp program-name

You can also save the raw stack trace (without source info) to a disk file and submit it to SYMIFY later, like this: 

 symify -i problem.dmp program-name

This comes in handy when your program grabs the screen (e.g., for some graphics) and the stack trace can't be seen. You can then redirect the stack trace to a file, e.g., with the REDIR program which comes with DJGPP.

But what if you didn't compile your program with -g, and you aren't sure how to recreate the problem which crashed it, after you recompile? Well, you can submit the stack dump after you recompile your program. Just press that PrintScreen key or otherwise save the stack trace, then submit it to SYMIFY from a file as described above, after you've recompiled the program. Be sure to give gcc all the compilation switches (sans -s) that you gave it when you originally compiled your program (in addition to -g), including the optimization switches, or else the addresses shown in the stack trace might point to wrong places.

If all you have from the crash is the program counter, the eight-digit hex number after "eip=", you can still find out the corresponding source line using GDB. Assuming that the EIP value is NNNNNNNN, type this at the GDB prompt: 

 list *0xNNNNNNNN

Node: File data corrupted, Next: , Previous: Crash traceback, Up: Running

9.4 Reading and writing binary files

Q: I'm reading/writing data files, but the data gets corrupted.

Q: My program crashes when I read data files, but the same program on Unix works OK.

Q: When I read a file I get only a small portion of it.

Q: I'm trying to open an existing binary file for read/write using the fstream class, but no mater what I do, the file is always truncated after I write to it....

Q: I cannot read anything from a binary file using the ifstream class, even though I use ios::binary!!

A: Are your data files binary? The default file type in DOS is "text", even when you use the read and write library functions. Text files get their newlines converted to CR-LF pairs on write and vice versa on read; reading in "text" mode stops at the first ^Z character. Reading binary files as text will therefore corrupt the data and fail to read all the data you need. You must tell the system that a file is binary through the b flag in fopen, or O_BINARY in open, or use the setmode library function to switch the handle to binary mode (the latter method is handy when you didn't open the file in your code, like what happens with standard input and output).

Note that the above distinction between binary and text files is written into the ANSI/ISO C standard, so programs that rely on the Unix behavior whereby there's no such distinction, are strictly speaking not portable.

You can also use the low-level _read and _write library functions which give you the direct interface to the DOS file I/O; they always use binary I/O.

If you have problems with read/write access to binary files via the fstream class in C++ programs, then make sure you call the constructor with an explicit ios::in and/or ios::out parameter, like this: 

 ifstream object_name ("file", ios::binary | ios::in);

Likewise, if you want to write binary files, you need to mention the ios::out flag explicitly. (This is actually a bug in all versions of the GNU C++ iostreams library up to and including version 2.95.)

Versions of the GNU C++ library before 2.8.1 had a bug in the GNU iostream classes. This bug caused truncation of files, even if you never write to the file. If you still use such an old version and cannot upgrade, a workaround is to do something like this: 

 fstream inFile;
 int fd = open ("foobar", O_RDWR | O_BINARY);
 inFile.fstream (fd);

Node: Screen IO, Next: , Previous: File data corrupted, Up: Running

9.5 Buffered screen I/O surprises

Q: My program prompts the user to enter data from the keyboard, then reads its response. When compiled with a 16-bit compiler like BCC or MSC it works as expected, but with gcc the prompt doesn't show, or is printed much later in the program.

Q: My program prints text in a loop, but the text appears on the screen only after the loop is finished....

Q: Help! I cannot make `gotoxy' work! The text I print appears on the screen in incorrect locations after I use `gotoxy'!

Q: Why does the text appear in the default colors even though I call `textcolor' and `textbackground'?

A: Do you write to screen using buffered I/O (fprintf, fputs and the like) functions, or send your output to the C++ cout stream? Then what you see is the effect of the buffering of the standard output streams. The buffer is not written to screen until it's full, or until a newline is output, which might produce very unpleasant and unexpected behavior when used in interactive programs.

DJGPP library functions use more aggressive buffering than 16-bit real-mode compilers, because delivering the output to the screen requires an expensive switch from protected to real mode and back. DJGPP tries to minimize the amount of these mode switches for performance reasons.

It is usually a bad idea to use buffered I/O in interactive programs; you should instead use screen-oriented functions like cprintf and cputs. If you must use buffered I/O, you should be sure that both stdout and stderr are line-buffered or unbuffered (you can change the buffering by calling the setvbuf library function); another solution would be to fflush the output stream before calling any input function, which will ensure all pending output is written to the operating system. While this will generally work under DOS and DJGPP, note that in some cases the operating system might further buffer your output, so sometimes a call like fsync would be needed to actually cause the output be delivered to the screen.

The functions that set text attributes only affect the screen-oriented output (a.k.a. conio) functions (cputs, cprintf etc.), the text written by fprintf and other stdio functions doesn't change. This is unlike some 16-bit DOS compilers where stdio functions can also print colored text.

Node: Distributing, Next: , Previous: Screen IO, Up: Running

9.6 What do DJGPP programs need to run?

Q: When I copy my DJGPP application program to another PC where no DJGPP is installed, I can't run it. It complains that it cannot find DPMI (??). Do I really need all of your multi-megabyte installation to run compiled programs?

A: No, you don't. You can either (a) bring the CWSDPMI.EXE free DPMI host to the target machine and put it in the same directory as your compiled program or somewhere along the PATH, or (b) make sure there's another DPMI host (such as QDPMI, 386Max, Windows, etc.) installed on the target machine.

If your program could be run on a machine which lacks a floating-point processor, you should also distribute an emulator, or link your program with an emulator library. See floating-point emulation issues.

PMODE/DJ is an alternative DPMI host that can be bound with your program, so that you have a single self-sufficient executable, but remember that PMODE/DJ doesn't support virtual memory, so such programs will only run on machines with enough free physical RAM.

Node: File handles, Next: , Previous: Distributing, Up: Running

9.7 How many file handles can DJGPP use?

Q: The library reference tells me that DJGPP programs can use up to 255 file handles, but my program can only use much less, about 30....

Q: I put a FILES=60 directive in my CONFIG.SYS, but my programs cannot use more than 42 when they run on Windows. Why is that?

A: It's no wonder you are confused: this is one of the most complicated issues related to the DOS filesystem. I cannot discuss all the details here17, but I will try to explain at least those aspects which directly affect a typical DJGPP user.

It is true that the DJGPP library lets you open up to 255 handles--but only if the operating system allows it. The operating system further limits this number, depending on several factors.

First, if you create new handles by calling the dup library function (or the underlying function 45h of the DOS Interrupt 21h), you can always have up to 255 such handles (minus the 5 that are open by the system before the program starts), even if the FILES= directive sets a much smaller count. All such handles refer to the same file or device and moving the file pointer using one handle moves all the rest of them.

In nested programs (that is, programs that were invoked by other programs), this is a bit more complicated. By default, any handle that is open in the parent program is inherited by the child, unless the parent sets the special O_NOINHERIT bit when it opens the file. Thus, if the parent had 10 files open when it invoked the child, the child program will have 10 less available handles--245--to work with, even if it only calls dup18.

The FILES= directive comes into play when you call open or any of its brethren to create handles. Unlike the handles created by dup, open (and the underlying functions 3Dh or 6Ch of Interrupt 21h) create handles that are independent of each other, even if you open the same file over and over again. The operating system will not let you create more such handles than the limit set by the FILES= directive. This is because the FILES= directive sets the number of entries in the SFT, the System File Table maintained by DOS, where all the information about every open file is kept19. So, if your CONFIG.SYS specifies FILES=60, you cannot open more than 60 files. After that, a call to open will fail with ENFILE (Too many open files in system).

In practice, you won't even be able to get 60 handles if you have FILES=60 in your CONFIG.SYS, since several handles are always preconnected. On plain DOS, 5 handles are already open when a program starts. These correspond to standard input, standard output, and standard error streams, and the other 2 handles are connected to the AUX and PRN devices. So, if you have FILES=60, DOS will only let you open up to 55 independent handles. (If your program doesn't need some of the 5 standard handles, you can close them and gain some more handles to play with.)

The plot thickens even more if you run DJGPP programs on Windows. Since Windows itself uses up 10-15 handles in the System Virtual Machine (VM), it tries to make it up for the DOS programs by adding private file tables to each DOS box with additional handles, beyond those maintained in the system-wide SFT. The default is to add a private table with 10 handles to each DOS box, but the PerVMFiles= entry in the [386Enh] section of the SYSTEM.INI file can override that. So on Windows, you need to consider the PerVMFiles= setting as well, and the resulting limit on open handles is less predictable since the number of handles used by Windows isn't constant (for example, it depends on how many fonts are loaded by Windows programs at any given moment).

If you run DJGPP on Windows 3.X, and your system loads SHARE.EXE during bootstrap, things become even more complicated. SHARE.EXE prevents Windows from adding private file tables (because it couldn't spy on files open via those private handles), so you get 10-15 less handles than what the FILES= directive says, and sometimes even less than that. That is how somebody who has FILES=60 on their CONFIG.SYS could only get 42 handles on Windows. If you are looking for reasons not to load SHARE.EXE, here you have another one.

Node: Virus, Previous: File handles, Up: Running

9.8 DJGPP and Anti-Virus Software

Q: I upgraded my anti-virus software, and now it finds a virus in all DJGPP programs!!

A: Relax, this is most probably a false alarm. The DJGPP stub loader, a short 2KB DOS program prepended to each DJGPP program, is optimized for size, and employs some clever tricks to make its code smaller. A few over-zealous virus scanners take some of these tricks as tell-tale signs of a virus, and report that all DJGPP programs are infected.

The truth is that DJGPP is distributed via SimTel.NET mirrors, which are known to scan all binaries for viruses before the zip files are cleared for general use. In addition, many DJGPP packages are built on Unix systems, where a DOS/Windows virus cannot survive. So it is very unlikely that a real virus would get through, and infect all of the programs on top of that. A couple of such false alarms were seen in recent years, but all of them proved to be bugs in anti-virus programs.

Node: Graphics, Next: , Previous: Running, Up: Top

10 Writing and Running Graphics Programs

This chapter discusses some problems and explains some subtle points related to graphics programming under DJGPP.

Node: GRX driver, Next: , Previous: Graphics, Up: Graphics

10.1 What GRX driver to use with your SVGA

Q: Why won't GRX work with my SVGA adapter in any resolution but the standard VGA?

Q: How do I tell GRX which driver to use with my SVGA?

A: In order for GRX to work with your SVGA, you should set the GRX20DRV environment variable, like this: 

  set GRX20DRV=et4000 gw 1024 gh 768 nc 256

To set that variable, you need to know the chip-set on your adapter; refer to your SVGA documentation. Currently, GRX supports the following chip-sets:

ati28800
The ATI 28800 chip-set.
cl5426
Cirrus Logic CL-GD5426 or higher (like CL-GD5428) chip-set.
et4000
Tseng Labs ET4000 chip-set.
mach64
The ATI Mach-64 SVGA.
stdega
The standard EGA adapter.
stdvga
The standard VGA adapter.
VESA
For any VESA-compatible adapter.

After you set the GRX20DRV variable, run modetest.exe to see what modes you have available.

If your chip-set is not one of the above, try the VESA driver because many adapters support the VESA BIOS extensions. If yours doesn't, try installing a VESA BIOS emulator, like UNIVBE. The latest version of UNIVBE and related software is always available from SciTech Web site.

Node: Direct access, Next: , Previous: GRX driver, Up: Graphics

10.2 Accessing the video memory

Q: I try to access the video memory at 0xa0000, but my program crashes with SIGSEGV....

Q: How can I access the text-mode video memory of my VGA?

A: Absolute addresses of memory-mapped devices are mapped differently under DJGPP than what you might be used to under other DOS development environments. That's because DJGPP is a protected-mode environment, in which you can't just poke any address: that's what protected mode is all about! To access such absolute addresses, use the so-called "farptr" functions like _farpeekb and _farpokew; they are described in the C Library reference. See more details on using "farptr" functions to access absolute addresses in low memory, below.

For text-mode screen updates, you can use the ScreenUpdate and ScreenUpdateLine library functions to quickly update the screen from a buffer prepared in memory.

Using the _farpeekX/_farpokeX paradigm to access memory isn't much slower than direct access (they compile into 2 machine instructions when optimizations are enabled). But if you need even faster access (and don't want to write it in assembly), see using the "nearptr" access facilities, as described below.

Some examples of how to access video memory from DJGPP programs are available in Brennan Underwood's tutorial.

Node: Graphics and Windows, Next: , Previous: Direct access, Up: Graphics

10.3 Graphics screen restoring under Windows

Q: When I switch away from my DJGPP program under Windows 3.X, then switch back to it, graphics mode is down, or my screen is all messed up. Why?

Q: I cannot run my program which uses Allegro: Windows 9X says the program would work better in DOS Mode....

Q: When running a program that uses Allegro under Windows, I cannot switch away from it with Alt-<TAB>: instead of switching, the PC beeps at me.

A: Windows 3.X only saves the VGA screen in standard VGA modes (1..13h) when you task-switch away from a DOS application. In any other mode it only saves/restores the video mode number, but not the actual screen contents. Your application is most likely still in the proper video mode (if not, it's probably the fault of the Windows driver for your SVGA card), but the video memory is messed up. The beauty of all this is that your program has no way of knowing that the screen has been taken away and then returned to it.

The only reasonable thing to do is to dedicate a "hotkey" in your application (e.g., Alt-R) whose action is to redraw the entire screen. If you do that, it's best to start all the way from the beginning, e.g. with a call to GrSetMode (if you use GRX), as there are a few bad Windows video drivers which do not restore SVGA graphics modes properly upon the switch back.

Windows 9X does save and restore the SVGA state, but only if you task-switch with the Alt-<TAB> key. If the switch happens because of anything else, like a window popping up, or you pressing the Start button, there's nothing your application can do to ensure it restores correctly, because it just never gets moved back into focus. As soon as the user tries to restore it, Windows 9X comes up with this message: 

 This application cannot be restored and will be terminated.

If you cannot switch from a graphics program by pressing Alt-<TAB>, it usually means that some of the Windows drivers, most likely the graphics one, is faulty. Some SVGA drivers simply don't bother to implement the save- and restore-state functions which Windows needs to switch from a DOS program that uses SVGA graphics modes. The solution is to upgrade your driver, or replace the SVGA with one that is better supported.

To prevent Windows 9X from getting in your way when running graphics programs, like popping up messages that suggest to run the program in DOS Mode, just disable one or more of the relevant properties for that program. Here's a detailed procedure to disable them all:

Programs which use latest versions of Allegro should not usually trigger warning messages from Windows, so upgrade to the latest Allegro version if you keep getting such warnings.

Node: OpenGL, Previous: Graphics and Windows, Up: Graphics

10.4 OpenGL and related packages for DJGPP

Q: What is OpenGL? Where can I get a DJGPP-compatible version?

Q: Where can I find a version of MESA for DJGPP?

A: First, a little background. OpenGL is an abstract interface design specification for drawing individual polygons, originally designed by SGI. It is generally regarded as well-designed, relatively high-level, and easy to program. (In contrast, DirectX is low-level and notoriously hard to program, but much faster on machines without hardware acceleration.) There are many different libraries that implement the OpenGL interface.

MESA is a free implementation of OpenGL distributed under LGPL, the GNU Library License. MGL is a 2D graphics library written by SciTech, that includes a copy of MESA and is distributed under a license that is less restrictive than LGPL. (Unfortunately, SciTech failed to mention in their docs that MESA is under LGPL, and thus people who use MGL might inadvertently violate the LGPL because they don't know it applies to their code.) Latest versions of MGL are known to work on MS-DOS, Linux, OS/2, and QNX.

The DJGPP version of MESA doesn't support hardware acceleration, but it does support some 3D chipsets on other platforms, so it should be possible to support that on DOS also, given the motivation.

You can get MESA from <http://www.mesa3d.org>. Latest versions might not work with DJGPP out of the box, but version 2.6 is known to have a full DJGPP support.

Version 4.5 beta 4 of MGL is available from the SciTech FTP site; it includes version 3.0 of MESA that supports DJGPP. It also includes GLUT.

Node: Floating point, Next: , Previous: Graphics, Up: Top

11 Floating Point Issues and FP Emulation

This chapter deals with issues pertaining to floating-point code and floating-point emulation under DJGPP.

Node: Emulation, Next: , Previous: Floating point, Up: Floating point

11.1 Floating-point code without 80387

Q: I don't have an 80387. How do I compile and run floating point programs?

Q: What shall I install on a target machine which lacks hardware floating-point support?

A: Programs which use floating point computations and could be run on machines without an 80387 should either be linked with the libemu.a emulation library (add -lemu to your link command line) or be allowed to dynamically load the emu387.dxe emulator at run-time if needed. Linking with libemu makes distribution simpler at a price of adding about 20KB to the size of the program .exe file (the emulator functions will be used only if no hardware floating point support is detected at runtime). You should always do one of the above when you distribute floating-point programs.

A few users reported that the emulation won't work for them unless they explicitly tell DJGPP there is no x87 hardware, like this: 

  set 387=N
  set emu387=c:/djgpp/bin/emu387.dxe

This is probably due to some subtle bug in the emulator setup code. It is possible that it was fixed in the latest DJGPP version, so upgrade if you can. If the problem persists, please post the details to the comp.os.msdos.djgpp news group.

There is an alternative FP emulator called WMEMU (get the file v2misc/wmemu21b.zip). It mimics a real coprocessor more closely, but is larger in size and is distributed under the GNU General Public License (which generally means you need to distribute its source if you distribute wmemu387.dxe, or distribute the source or objects to your entire program, if you link it with libwmemu.a). Its advantage is that with WMEMU, you can debug FP apps on a non-FPU machine. (But you will need to get the latest binaries of WMEMU, since older distributions were compiled with a beta release of DJGPP v2.0 and will cause unresolved externals if you try linking against libwmemu.a without recompiling it.) Note, however, that even WMEMU doesn't solve all the problems of debugging FP programs on a non-FPU machine (e.g., emulating flags doesn't work).

Node: Emulator accuracy, Next: , Previous: Emulation, Up: Floating point

11.2 Floating point inaccuracies when using emulator

Q: I am experiencing inaccurate results in some floating point calculations, sometimes in the 2nd or 3rd significant digit (like getting 118.401 instead of 120.0). This is really unacceptable! (And no, I'm not using a buggy Pentium CPU.)

Q: I get some very inaccurate results when my program runs on a machine lacking an FPU....

A: Are you using the emu387.dxe emulator? If so, it might be that the emulator isn't as accurate as you expect. Versions of the emulator distributed with DJGPP 2.02 and earlier had a bug that affected addition, subtraction, and comparison of floating-point numbers with some specific bit patterns. This bug could produce inaccuracies in math functions such as sqrt, sin and tan for some specific argument values, and even cause a program to be trapped in an infinite loop. The emulation of the FPATAN instruction and functions based on it, like atan, asin and acos, also suffered loss of accuracy for some specific arguments. DJGPP v2.03 solves these problems, so upgrade and see if your problems go away.

However, even the emulator supplied with v2.03 and later suffers some accuracy degradation when computing trigonometric functions for arguments that are integral multiples of Pi/2 or Pi/4 (depending on the particular function you call), and when computing inverse trigonometric functions which should yield results that are such multiples. So, for example, if you use 4*atan(1.) to get the value of Pi, that might be your problem.

The reason for this accuracy degradation is that emu387.dxe does not store the value of Pi, with extra precision, like the real FPU does, and trig functions in libc.a rely on such extra accuracy to deliver accurate results.

For computing the value of Pi, the solution is simple: make it a constant, as God intended. The header file <math.h> includes the constant M_PI which you can use; or get the value of Pi from the net.

In many cases that involve trigonometric functions and yield inaccurate results, linking your program with the -lm switch might help. This switch causes the linker to use an alternative math library, libm.a, which doesn't rely on x87 instructions, and thus is more accurate when the emulator deviates from the actual x87.

The alternate emulator WMEMU is known to be accurate to 7 significant digits for float variables, and 15 digits for doubles. It also much more faithfully emulates the behavior of the x87 processor when abnormal arguments (Inf, NaN, etc.) are involved. So if emu387.dxe which comes with DJGPP v2.03 doesn't solve your problems, you might try using WMEMU as a solution.

Node: Emulation in Windows, Previous: Emulator accuracy, Up: Floating point

11.3 Problems with emulation on Windows

Q: My program which uses floating-point math hangs on Windows DOS box when I try to use FP emulation....

A: This is due to a bug in the emulator in DJGPP v2.02 and earlier. The bug affected those programs running on Windows 3.X and 9X which use the WAIT and FWAIT instructions (most of non-trivial FP programs do), both those which use the emulator emu387.dxe and those linked with the emulation library using the -lemu switch. The bug is solved in DJGPP versions 2.03 and later, so upgrade.

Node: Debugging, Next: , Previous: Floating point, Up: Top

12 Debugging DJGPP Programs

This chapter discusses the debuggers you can use with DJGPP and answers some of the questions you might have when debugging DJGPP programs.

Node: How to debug, Next: , Previous: Debugging, Up: Debugging

12.1 How to run a DJGPP program under debugger

Q: How do I debug my programs?

A: First, remember to use the -g switch when you compile and link. This puts debugging information into your executable. When linking, don't use the -s switch. Here are a few examples of compilation and link command lines when you intend to debug a program: 

 gcc -Wall -c -g -O myfile.c

 gcc -Wall -O2 -g -o myprog.exe mymain.c mysub1.c mysub2.c -lm

 gcc -g -o myprog myprog.o mysub.o

Note that with gcc, you can use optimization switches when compiling with -g. To use stabs debugging, compile with -gstabs3 or -gstabs+ instead of -g. (Stabs debugging info is more powerful than the default COFF debugging; if the debugger doesn't seem to support some feature, or behaves strangely, try compiling the program with -gstabs+ and see if that helps.) Stabs debugging is especially recommended for C++ programs, since the default format of debugging info is not powerful enough to record all the necessary information about C++ code.

If (or when) GCC supports the dwarf2 debugging info, compile the program with the -gdwarf2, since it is even better than stabs, especially with the new generation of GCC optimizations.

Then, to debug the program, use a command line like this (here for the GDB debugger): 

 gdb myprog.exe

Beginning with v2.01, DJGPP debuggers can debug both unstubbed COFF images and DOS-style .exe executables (v2.0 only supported COFF files). To debug a COFF file, name it without the .exe extension, like so: 

 gdb myprog

You can use one of several available debuggers with DJGPP:

  1. RHIDE, the DJGPP IDE by Robert Hoehne which is available from the DJGPP archives. It includes an integrated source-level debugger based on GDB code and presents a user interface like that of Borland's IDE or Turbo Debugger.
  2. GDB, the GNU Debugger. This is an extremely powerful source-level debugger, but it uses a line-oriented user interface. People who are familiar with using GDB on Unix should know about the following important differences in its operation on MS-DOS:
  3. FSDB, the full-screen debugger, from the djdev distribution. This presents a user interface like that of Borland's Turbo Debugger, but unlike TD, it isn't a source-level debugger (although it will show the source code together with the machine instructions). It also supports data-write breakpoints: a powerful feature for hunting down code which overwrites data it shouldn't touch. Another advantage of FSDB is that you can easily debug programs that grab the screen, because it can switch between the debugger screen and the application screen. Also, it allows to examine the FPU registers. The main disadvantage of FSDB is that you cannot easily examine the contents of complex data structures. Remember to prepend an underscore _ to the names of C identifiers when you use them with FSDB; for C++ programs you will have to find out the mangled names of static class variables and methods to make FSDB understand them.
  4. EDEBUG32 is the most basic debugger you can use with DJGPP.

You invoke any debugger like this: 

 <debugger-name> <program> <args...>

(except that with GDB, you need to pass the arguments from within the debugger).

Node: Crash dump, Next: , Previous: How to debug, Up: Debugging

12.2 How to begin debugging using the crash dump info

Q: My program crashed with SIGSEGV, but I'm unsure how to begin debugging it....

Q: Can you help me figure out all those funny numbers printed when my program crashes?

A: Debugging should always begin with examining the message printed when the program crashes. That message includes crucial information which usually proves invaluable during debugging. So the first thing you should do is carefully save the entire message. On plain DOS, use the <PrintScreen> key to get a hard copy of the message. On Windows, use the clipboard to copy the message to a text editor or the Notepad, and save it to a file. If you can easily reproduce the crash, try running the program after redirecting the standard error stream, where the crash dump is printed, to a file, e.g. like this: 

  redir -e crash.txt myprog [arguments to the program go here]

(here I used the redir program supplied with DJGPP; the -e switch tells it to redirect the standard error stream to the named file). Redirecting the standard error stream to a file has an additional advantage of printing the entire call frame traceback, even if it is very long, whereas when writing to the screen, the DJGPP exit code limits the number of printed stack frames so that the crash message won't scroll off the screen.

After you've saved the crash message, look at the name of the crashed program, usually printed on the 4th line. Knowing which program crashed is important when one program calls another, like if you run a program from RHIDE. Without this step, you might erroneously try to debug the wrong program. (If the program name is garbled, or if <??UNKNOWN??> is printed in its stead, it means the program crashed inside the startup code.)

The next step in the debugging is to find out where in the code did the program crash. The SYMIFY program will help you translate the call frame traceback, which is the last portion of the crash message, into a list of function names, source files and line numbers which describe the sequence of function calls that led to the crash. The top-most line in the call frame traceback is the place where the program crashed, the one below it is the place that called the function which crashed, etc. The last line will usually be in the startup code, in a function called __crt1_startup, but if the screen is too small to print the entire traceback without scrolling, the traceback will be truncated before it gets to the startup. See how to use SYMIFY, for more details about the call frame traceback and SYMIFY usage.

If you compiled your program without the -g switch, or if you stripped the debugging symbols (e.g., using the -s linker switch), running SYMIFY will just repeat the addresses instead of translating them to function names and source file info. You will have to rebuild the program with -g and without -s, before you continue.

Next, you need to get an idea about the cause of the crash. To this end, look at the first two lines of the crash message. There you will find a description of the type of the crash, like this: 

 Exiting due to signal SIGSEGV
 Page Fault at eip=00008e89, error=0004

(the actual text in your case will be different). The following table lists common causes for each type of crash:

Page Fault
This usually means the program tried to access some data via a NULL or an uninitialized pointer. A NULL pointer is a pointer which holds an address that is zero; it can come from a failed call to malloc (did your code check for that?). An uninitialized pointer holds some random garbage value; it can come from a missing call to malloc.

The error code (error=0004 in the example above) will usually be either 4 or 6. The former means that the program tried to read (take a value from) the invalid address, the latter means the program tried to write (change the stored value) there.

Sometimes, you might see a somehwat different format of a Page Fault message: 

 Page Fault cr2=10000000 at eip e75; flags=6
 eax=00000030 ebx=00000000 ecx=0000000c edx=00000000
 esi=0001a44a edi=00000000 ebp=00000000 esp=00002672
 cs=18 ds=38 es=af fs=0 gs=0 ss=20 error=0002

This message comes from CWSDPMI, which could happen when some crucial data structure in the low-level library code becomes trashed. The value in cr2 is the address which caused the Page Fault exception. In the example above, this address is 0x10000000, and since this is exactly the base address of the DJGPP program under CWSDPMI, it means the program dereferenced a NULL pointer.

If the message says Page Fault in RMCB, then it usually means that the program installed an interrupt handler or a real-mode callback (a.k.a. RMCB), but failed to lock all the memory accessed by the handler or functions it calls. See installing hardware interrupt handlers, for more about this. It also might mean that a program failed to unhook some interrupt before it exited.

General Protection Fault
This can be caused by a variety of reasons:

Overwriting the stack frame can usually be detected by looking at the values of the EBP and ESP registers, printed right below the first two lines. Normally, ESP is slightly smaller than EBP, smaller than the limit of the SS segment, and usually larger than EIP21; anything else is a clear sign of a stack being overrun or overwritten. In particular, if ESP is valid, but EBP is not, it usually means that the stack was overwritten. In some cases, EBP's value might look like a chunk of text, like 0x33313331 (the string 1313, after swapping the bytes due to the fact that x86 is a little-endian machine).

How do you know whether the values of ESP and EBP are valid? To help you, DJGPP v2.02 and later prints the valid limits of the application stack, like this: 

App stack: [000afb50..0002fb50]  Exceptn stack: [0002fa2c..0002daec]

(The second range of values, for the "Exceptn stack", shows the 8KB-long stack used by the library for processing hardware exceptions, because the normal application stack might be invalid when an exception happens.)

Another tell-tale sign of an overrun stack frame is that the symified traceback points to a line where the function returns, or to its closing brace. That's because, when a program overruns the stack, the return address saved there gets overwritten by a random value, and the program crashes when the offending function tries to return to an invalid address.

Suspect a stack overflow if the EBP and ESP values are close to one another, but both very low (the stack grows downwards) and outside the valid stack limits printed below the registers' dump, or if the call frame traceback includes many levels, which is a sign of a deep recursion.

Another sign of a stack overflow is when the traceback points to some internal library structure, like __djgpp_exception_table, or if the SS selector is marked as invalid in the crash message.

Stubediting the program to enlarge its stack size might solve problems with stack overflow (but not when the stack is being overwritten as described above). See changing stack size, for a description of how to enlarge the stack. If you use large automatic arrays, an alternative to stubediting is to make the array dimensions smaller, or make the array global, or allocate it at run time using malloc.

Note that, unlike in the cases, described above, where the stack was overwritten, stack overflow usually manifests itself by both ESP and EBP being invalid (outside the valid limits printed by the crashed program).

Stack Fault
Usually means a stack overflow, but can also happen if your code overruns the stack frame (see above).
Floating Point exception
Coprocessor overrun
Overflow
Division by Zero
These (and some additional) messages, printed when the program crashes due to signal SIGFPE, mean some error in floating-point computations, like division by zero or overflow. Sometimes such errors happen when an int is passed to a function that expects a float or a double.
Cannot continue from exception, exiting due to signal 0123
Cannot continue from exception, exiting due to signal SIGSEGV
This message is printed if your program installed a handler for a fatal signal such as SIGSEGV (0123 in hex is the numeric code of SIGSEGV; see the header signal.h for the other codes), and that handler attempted to return. This is not allowed, since returning to the locus of the exception will just trigger the same exception again and again, so the DJGPP signal-handling machinery aborts the program after printing this message.

If you indeed wanted SIGSEGV to be generated in that case, the way to solve such problems is to modify your signal handler so that it calls either exit or longjmp. If SIGSEGV should not have been triggered, debug this as described below.

Invalid TSS in RMCB
This usually means that a program failed to uninstall its interrupt handler or RMCB when it exited. If you are using DJGPP v2.0, one case where this happens is when a nested program exits by calling abort: v2.0 had a bug in its library whereby calling abort would bypass the cleanup code that restored the keyboard interrupt hooked by the DJGPP startup code; v2.01 solves this bug.

Using the itimer facility in v2.01 programs can also cause such crashes if the program exits abnormally, or doesn't disable the timer before it exits. The exit code in DJGPP v2.02 and later makes sure the original timer interrupt is always restored.

Double Fault
If this message appears when you run your program under CWSDPR0 and press the Interrupt key (Ctrl-<C> or Ctrl-<BREAK>) or the QUIT key (Ctrl-<\>), then this is expected behavior (the SIGINT generation works by invalidating the DS/SS selector, but since CWSDPR0 doesn't switch stacks on exceptions, there's no place to put the exception frame for the exception this triggers, so the program double faults and bails out). Otherwise, treat this as Page Fault.
Control-Break Pressed
Control-C Pressed
INTR key Pressed
QUIT key Pressed
These are not real crashes, but are listed here for completeness. They are printed when Ctrl-<BREAK> or the Interrupt key (by default, Ctrl-<C>) is pressed, which by default abort the program due to signal SIGINT. The QUIT key (by default, Ctrl-<\>) generates the SIGQUIT signal which by default is ignored, but some programs set it to abort the program as well.

If you are lucky, and the crash happened inside your function (as opposed to some library function), then the above info and the symified call frame traceback should almost immediately suggest where's the bug. You need to analyze the source line corresponding to the top-most EIP in the call frame traceback, and look for the variable(s) that could provide one of the reasons listed above. If you cannot figure it out by looking at the source code, run the program under a debugger until it gets to the point of the crash, then examine the variables involved in the crashed computation, to find those which trigger the problem. Finally, use the debugger to find out how did those variables come to get those buggy values.

People which are less lucky have their programs crash inside library functions for which SYMIFY will only print their names, since the libraries are usually compiled without the -g switch. You have several possible ways to debug these cases:

Node: Debug graphics, Next: , Previous: Crash dump, Up: Debugging

12.3 How to debug a graphics program

Q: How can I debug a graphics program? The debugger runs my program fine, but when a breakpoint is hit with the screen in a graphics mode I can't read the text printed by the debugger.

A: You can redirect the debugger output to your printer, like this: 

 gdb myprog > prn

This will only work if the program itself doesn't write to stdout (graphics programs usually don't); otherwise the debugger output will get mixed up with your program's output. If you use GDB 4.18 or later, you can work around this by redirecting the standard output of the debugged program to a different file or device as part of the run command, like this: 

 run > foo.out

Beginning with version 4.18, the ported GDB writes its output to the screen in a way that works even in graphics modes, provided that the system BIOS knows about the specific graphics mode your program uses.

RHIDE and RHGDB support debugging graphics programs by switching between debugger's and program's screen, so you can use RHIDE's built-in debugger or the stand-alone RHGDB subset. This support doesn't work for all video modes, but the standard VGA modes and VESA modes are supported. If you debug with RHIDE on Windows, switch your RHIDE session to full-screen before starting the debug session, otherwise Windows will cause problems when RHIDE switches between the program's graphics screen and RHIDE's own text screen.

The FSDB debugger can switch between the application screen and the debugger screen, so you might use it, at a price of working with a low-level debugger. Press Alt-<F5> to switch between the two screens. Stock FSDB as distributed with DJGPP can only do this with text screens, but a modified version of FSDB with graphics support is available that knows about many graphics modes. The same distribution can also be found on the Oulu repository.

As yet another possibility, consider using the MSHELL program which will redirect I/O from any program to the monochrome monitor at the BIOS level, so you can use it even with GDB. MSHELL was written by DJ Delorie and is available from DJ's server. Be sure that you don't have some other TSR installed that catches screen writes and bypasses the BIOS functions, or else MSHELL won't help you. For example, changing the code page (with the DOS CHCP or MODE commands) might do this.

RHIDE also supports dual-monitor systems for debugging, it allows you to use the monochrome monitor for interface with the debugger, while leaving the color screen for your program's display, with no need to swap between them.

If you have any problems with dual-monitor support, in particular with RHIDE, make sure your memory manager doesn't grab the B000 segment for its own purposes. This region should be available for the mono adapter, or your system might crash when you try using it.

Another way to redirect the output of a program to a monochrome monitor is by using the MDA display driver from BinaryInfosys. It is a true DOS device driver, and so can be opened as a file--handy for sending debug info, for example. This driver is free and is available from BinaryInfosys' home page.

Node: GDB and C++ source, Next: , Previous: Debug graphics, Up: Debugging

12.4 GDB finds only .cc source

Q: When I try to debug my C++ programs, the debugger claims it can't find the source file: 

 file.cc: No such file or directory.

The source file is there, but it's called file.cpp, not file.cc. Why does this happen?

A: It's a bug in GCC 2.7.2.1 and earlier. It erroneously assumes that a C++ source always has a .cc extension. If you are using GCC 2.7.2.1 or earlier, you'd better call your C++ files *.cc. If this is unacceptable, you can work around this bug by invoking cc1plus and the assembler pass manually. The bug in GCC manifests itself in that cc1plus is called with the option -dumpbase file.cc. If you replace this with -dumpbase file.cpp (or whatever your extension is), the debugger will happily find your sources.

GCC 2.8.0 and later corrects this bug, so upgrading is also a solution.

Node: C++ classes in GDB, Next: , Previous: GDB and C++ source, Up: Debugging

12.5 Can GDB print class members?

Q: It seems that GDB doesn't recognize C++ class members by their original, unmangled names. Do I really need to figure out the mangled names of all my class variables and methods to be able to debug them?

A: No, you don't. GDB does allow you to use the original names, it's just that it usually treats the :: in their names as word delimiters. Include the name of the method or a class static variable in single quotes, and GDB will recognize them as a single word. For example, if your class CMPForward has a method named go which you need to put a breakpoint in, use the following command: 

  breakpoint 'CMPForward::go'

Other GDB features that might be useful in this context are the various demangling options, like set print demangle, set demangle-style etc.; look them up in the GDB on-line docs.

However, there are some cases where you won't be able to get GDB to demangle C++ function names no matter how hard you try. This is due to a lack of sufficient debugging information in the COFF debug data format. There's simply not enough info there for GDB to detect the source language and support some C++-specific features. So, in some case, you will need to use mangled names. If you need a description of the GNU style of mangling C++ names (so you could mangle them yourself), look in the GDB or Libg++ source distribution, in the libiberty directory, for a file named cplus-demangle.c. You can also use the cxxfilt utility, supplied as part of the GNU Binutils package, to demangle the names and verify that your mangling is correct. Note that, as the debugger built into RHIDE uses GDB code, it will also sometimes have such problems with debugging C++ programs.

If you really need full C++ support in DJGPP, you will have to use the stabs debugging support. GCC 2.8.0 and later are distributed with built-in stabs support, so, if you need this, upgrade and compile your C++ programs with -gstabs+. Caveat emptor: FSDB, EDEBUG32 and SYMIFY don't understand the stabs format, so you will have to compile with -gcoff option to use these utilities (RHIDE distribution includes a utility called gsymify that can be used instead of SYMIFY with stabs debugging info).

Node: Included source, Next: , Previous: C++ classes in GDB, Up: Debugging

12.6 GDB cannot list source that was #include'd

Q: My source file #include's another source file, but I cannot set a breakpoint in that included code, because GDB says there is no such line, or no such source file....

Q: I cannot debug code produced by Flex, or Bison, or F2C, because GDB somehow messes up all the source file and line number info!

Q: Why can't I step with a debugger into an inline function defined in a header file?

Q: Why can't I trace into a function defined in an #included source file?

A: This is a genuine limitation of the COFF format used by DJGPP. It can only handle a single source file for a given object file. It does include correct line numbers, but the name of the source file is wrong, so setting breakpoints in such files or tracing into functions defined in such files just doesn't work. Using stabs debugging info (see the previous section) doesn't have this limitation, so upgrade to GCC 2.8.0 or later and recompile your program with the -gstabs+ switch.

For source files that include other source files, you can work around this even with COFF debugging, by just inserting the included source with your editor while you debug the program. For code produced by other programs, like F2C or Bison, Duncan Murdoch suggests a work-around: to copy the original source file (.y, .f, etc.) over the generated C file. For example, here's how you should go about debugging a Fortran program myprog.f using GCC, F2C and GDB:

  1. Run f2c with the -g option: 
     f2c -g myprog.f
  2. Compile using gcc with the -g option: 
     gcc -g myprog.c -o myprog.exe -lf2c -lm
  3. Copy the original Fortran source over the generated C: 
     copy myprog.f myprog.c
  4. Debug with GDB: 
     gdb myprog.exe

Node: Static vars, Next: , Previous: Included source, Up: Debugging

12.7 GDB cannot display or set static uninitialized variables

Q: Why can't I set or display the values of some static variables in my program?

A: This seems to be a limitation of the COFF debugging information emitted by GCC by default: the debuggers cannot display or set the value of an uninitialized static variables (those who are in the .bss section of the program). A work-around is to initialize these variables, which causes the linker to put them into the .data section. Another solution is to use the stabs debugging support; latest versions of GCC include this support, so upgrade and use -gstabs+ instead of -g.

Node: Bool vars, Next: , Previous: Static vars, Up: Debugging

12.8 Debugging bool data type

Q: How can I watch a bool variable with RHIDE or GDB? When I try, the debugger always displays void....

A: With the default COFF debugging format, you can't: it doesn't support the bool data type. You have to switch to stabs debugging format; see switching to stabs, for details.

Node: Complex vars, Next: , Previous: Bool vars, Up: Debugging

12.9 Debugging the complex data type

Q: I cannot display in GDB the values of my variables of type __complex__....

A: Current versions of GDB don't support __complex__ variables. A work-around is to manually cast them to a pair of numbers. For example, to access the real and imaginary part of a variable foo declared __complex__ double, do this: 

 (gdb) print *(double *)&foo
 $1 = 4
 (gdb) print *((double *)&foo + 1)
 $2 = 6

Node: Debugging woes, Previous: Complex vars, Up: Debugging

12.10 Debuggers choke on some programs ...

Q: I cannot debug Emacs (or any program that requests raw keyboard input): when I press Ctrl-C, any debugger I tried reported SIGINT. But I cannot operate the debugged program without Ctrl-C (in Emacs, it's necessary to exit the editor)!

Q: I cannot debug any program which catches signals!!??

Q: I compiled my program with -pg switch, and now I cannot debug it....

Q: When my program hits a breakpoint in GDB, the debugger reports SIGSEGV, but only under Windows....

A: Versions of DJGPP before v2.03 had a few grave limitations in debugging programs which use interrupts or exceptions. Programs compiled for profiling would crash under a debugger with SIGSEGV or a GPF, with no addresses that symify can identify; programs using alarm or setitimer couldn't be debugged, either. You couldn't hook the keyboard interrupt in a debugged program, and you couldn't debug a program which uses floating point on a machine without FP hardware (unless you use WMEMU as your emulator, but even WMEMU doesn't solve all the problems). The reason for all these problems was that any exceptions or signals that happen when your program runs under a debugger would be caught by the debugger instead, they won't get passed to the debuggee, and would usually terminate the debuggee.

This is no more the case. DJGPP v2.03 and later have a much better debug support, so all of the problems mentioned above are gone. The DJGPP port of GDB 4.18, released in August 1999, is based on the debugging support from DJGPP v2.03 and thus doesn't have most of these problems anymore. Latest versions of RHIDE also use this improved debugging support, as do the versions of edebug32 and fsdb from DJGPP v2.03 or later.

Some problems still remain, though, even in v2.03: if you use the stock emu387.dxe FP emulator while debugging floating-point programs or debug programs that call alarm or setitimer library functions, the program will sometimes crash with SIGSEGV. This is likely to change in the future.

Beginning with version 1.1, the debugger built into RHIDE supports debugging programs that hook keyboard and/or timer hardware interrupts, so if you need e.g. to debug programs built with the Allegro library or programs compiled for profiling, you can use RHIDE.

Another known problem is that GDB GP Faults when the program hits a breakpoint under Windows 3.X (Windows 9X doesn't have this problem). This is because the breakpoint instruction causes a software interrupt (as opposed to an exception) under Windows 3.X, and the DJGPP debug support currently only catches debug exceptions. The only work-around is to use the hardware breakpoints (which use the special debug registers of the i386 and higher CPUs, and which do work with DJGPP on Windows 3), and never have more than 4 of them active at the same time. FSDB will automatically use the hardware breakpoints for the first 4 breakpoints (so it works on Windows 3.X unless you set more than 4 breakpoints simultaneously), but with GDB, you will have to explicitly use the hbreak and thbreak (instead of break and tbreak) commands which set hardware breakpoints. This works with DJGPP ports of GDB 4.16 and later. Note that GDB and FSDB use ordinary breakpoints to implement single-stepping with the step, next, <F7>, <F8> and similar commands, so you can't use these on Windows 3.X; use temporary hardware breakpoints instead. The above is also true for watchpoints (which watch for variables to change value): you need to use hardware watchpoints with GDB (the total number of hardware breakpoints and watchpoints cannot exceed 4). Same considerations apply to the debugger built into RHIDE under Windows 3.X.

Node: Profiling, Next: , Previous: Debugging, Up: Top

13 Profiling DJGPP Programs

This chapter explains how to optimize your program for speed using the profiler, and discusses some problems you might have with it.

Node: How to profile, Next: , Previous: Profiling, Up: Profiling

13.1 How to profile a DJGPP program

Q: How can I profile my program to see where it spends most of its run time?

A: DJGPP includes a profiling facility. To use it, compile and link with -pg option, run your program as you usually would, then run a program called gprof: 

 gprof myprog.exe

(change myprog.exe to whatever name your program is). This will print an execution profile. You can now look at the profile and try to optimize the functions which take a large proportion of the execution time.

Gprof is further documented in the Binutils docs as part of the GNU Binutils distribution.

Node: Profiled crash, Next: , Previous: How to profile, Up: Profiling

13.2 Programs compiled with -pg crash when run

Q: I cannot profile my program: when I compile it with -pg, it crashes or wedges my machine!

Q: When I compile my program with -pg, it runs much slower. Does the profiling code have such a huge overhead?

Q: I profiled my program, but the profile contains an entry _mono_putc which I don't use, and which eats up about 70% of execution time!

Q: When I run a profiled program on my dual (VGA+MDA) display system, the mono screen shows loads of meaningless numbers. Is there a way to stop this behavior?

A: DJGPP v2.01 has a bug in one of its library functions which is linked into your program when it is compiled with the -pg option. The bug is that the profiled program tries to write to the secondary mono screen, which caused the profiled programs to crash in many environments, in particular when a memory manager remaps some of the high memory. On systems which actually have the additional mono display, the profiled programs won't crash, but would run significantly slower and print debugging info on the mono display.

A patch which corrects this bug was posted to the DJGPP News group; you can find it by searching the DJGPP mail archives. DJGPP v2.02 and later includes a fixed version of the offending function, so upgrade to the latest version. A work-around is to run the program compiled with -pg on vanilla DOS configuration (no memory managers such as EMM386 or QEMM, and no Windows). However, when you use this work-around, your program might run much slower, although the profile that you get should not be affected.

Node: Garbled profile, Next: , Previous: Profiled crash, Up: Profiling

13.3 Gprof produces garbled profile

Q: Whenever I compile my programs with -pg, the profile produced by Gprof shows that 100% of the run time is spent in a single function, and the rest of the code gets 0% of time. Huh??

A: This is due to a bug in the library shipped with DJGPP v2.02: the module which handles timers works incorrectly. (The same bug is responsible for problems with library functions setitimer and alarm.) The solution is to upgrade to DJGPP v2.03 where these bugs are solved. Relink your program with the v2.03 library, then rerun it, and Gprof will show reasonable results.

Node: IO bound programs, Next: , Previous: Garbled profile, Up: Profiling

13.4 Why is __dpmi_int so heavily used?

Q: I've profiled my program and found that the routine which takes 60% of the running time is some obscure library function called __dpmi_int. Can't you guys get your library right?

Q: What is the __dj_movedata function for, and why does it take such a large proportion of my program's running time?

A: Does your program use I/O or other real-mode services (like BIOS) extensively? All those services are invoked through a DPMI function call which is issued by __dpmi_int. The sibling function __dj_movedata moves data between the transfer buffer (see what is the transfer buffer) and your program, e.g., when it reads or writes disk files. So what the profile really says is that the running time of your program is governed by time-consuming operations such as disk I/O.

Node: No profile, Previous: IO bound programs, Up: Profiling

13.5 gprof doesn't produce output

Q: Every time I run the profiler it says "gmon.out: no such file or directory" and no profile is produced. What is this gmon.out file, and why won't gprof compute the profile?

A: gmon.out is the file with raw execution counts and timing info that gprof needs to produce the profile. The file is written by the profiled program when it exits. If the file isn't created, it might be because of one of the following reasons:

Node: Performance, Next: , Previous: Profiling, Up: Top

14 Run-time Performance of DJGPP Programs

This chapter deals with issues pertinent to run-time performance of DJGPP programs.

Node: How fast, Next: , Previous: Performance, Up: Performance

14.1 How efficient is DJGPP-generated code?

Q: How does DJGPP compare with other DOS-based C compilers in terms of efficiency of generated code?

Q: Won't my program run much slower when compiled by DJGPP, due to all those CPU cycles wasted in switches between protected and real mode?

A: The quality of code generated by GCC with optimization turned on (-O2 switch to the compiler) is generally at least as good as what you will get from top commercial products, like Borland, Microsoft and Watcom. Mode switches indeed impose a certain performance hit, but in most programs it is negligibly small, because only DOS and BIOS services require such a switch, and the majority of programs spend most of their time doing other things.

Up until version 2.95, MSVC was the only one of the commercial compilers that used to produce code which was better than what GCC generated (by about 25% on the average), when run on a Pentium. However, with the much-improved optimization technology that is now part of GCC 2.95 and later, this gap is all but closed. More details about this are available on the compiler comparison page, maintained by Salvador Eduardo Tropea.

Node: Older is faster, Next: , Previous: How fast, Up: Performance

14.2 Comparing newer versions with old ones

Q: I switched to v2 and my programs now run slower than when compiled with v1.x....

Q: I timed a test program and it seems that GCC 2.8.1 produces slower executables than GCC 2.7.2.1 was, which in turn was slower than DJGPP v1.x. Why are we giving up so much speed as we get newer versions?

Q: I installed Binutils 2.8.1, and my programs are now much slower than when they are linked with Binutils 2.7!

A: In general, newer versions of GCC generate tighter, faster code, than older versions. Comparison between different versions of GCC shows that they all optimize reasonably well, but it takes a different combination of the optimization-related options to achieve the greatest speed in each compiler version. The default optimization options can also change; for example, --force-mem is switched on by -O2 in 2.7.2.1; it wasn't before. GCC offers a plethora of optimization options which might make your code faster or slower (see the GCC docs for a complete list); the best way to find the correct combination for a given program is to profile and experiment. Here are some tips:

I'm told that the PGCC version of GCC has bugs in its optimizer which show when you use level 7 or higher. Until that is solved in some future version, you are advised to stick to -O6. Some programs actually run faster when compiled with -O2 or -O3, even when compiled with PGCC, so you might try that as well. Several users reported that PGCC v2.95.1 tends to crash a lot during compilation, especially with -O5, -O6 and -mpentium options. (In general, PGCC version 2.95 is deemed buggy; you are advised not to use it.)

Programs which manipulate multi-dimensional arrays inside their innermost loops can sometimes gain speed by switching from dynamically allocated arrays to static ones. This can speed up code because the size of a static array is known to GCC at compile time, which allows it to avoid dedicating a CPU register to computing offsets. This register is then available for general-purpose use.

Another problem that is related to C++ programs which manipulate arrays happens when you fail to qualify the methods used for array manipulation as inline. Each method or function that wasn't declared inline will not be inlined by GCC, and will incur an overhead of a function call at run time.

However, inlining only helps with small functions/methods; large inlined functions will overflow the CPU cache and typically slow down the code instead of speeding it up.

If your CPU is AMD's K6, try upgrading to GCC 2.96 or later and use the -mcpu=k6 switch. I'm told that K6-specific optimizations are much better in these versions of GCC.

A bug in the startup code distributed with DJGPP versions before v2.02 can also be a reason for slow-down. The problem is that the runtime stack of DJGPP programs was not guaranteed to be properly aligned. This usually only shows up on Windows (since CWSDPMI aligns the stack on its own), and even then only sometimes. But it has been reported that switching to Binutils 2.8.1 sometimes causes such slow-down, and switching to PGCC can reveal this problem as well. In some cases, restarting Windows would cause programs run at normal speed again. If you experience such problems too much, upgrade to v2.02.

Node: Pentium, Next: , Previous: Older is faster, Up: Performance

14.3 DJGPP programs on a Pentium

Q: Does DJGPP support Pentium-specific optimizations?

Q: I run the same program on a 486 and on a Pentium, and it's slower on a Pentium!!

A: Beginning with version 2.95, GCC includes Pentium-specific optimizations. Be sure to use the -mcpu=pentium switch when you optimize for Pentium or better CPUs.

A program might sometimes run slower on a Pentium due to alignment problems in DJGPP. GCC makes assumptions about how GAS (the assembler) handles alignment, but GAS from Binutils 2.8.1 and earlier was configured to treat alignment in a way that's different from what GCC assumes. The outcome of this is that longs are word-aligned, doubles are dword-aligned, etc. Depending on the DJGPP version, link order, library differences, you might get lucky (or unlucky) with a 50/50 chance to get an improper alignment. Different CPUs have different penalties for unaligned accesses, which may explain differences in speed.

DJGPP v2.01 had a bug in the startup code, whereby the runtime stack isn't aligned; this could also be a reason for slow-down, especially in programs compiled for Pentium.

You might consider adding some slack static variables to induce changes in alignment; if any of the changes suddenly cause a significant change in the runtime performance, then alignment might be the reason.

These alignment problems were finally solved in the DJGPP ports of Binutils 2.9.1 and GCC 2.95; so if you want to get rid of these problems, upgrade.

Node: IO speed, Next: , Previous: Pentium, Up: Performance

14.4 I/O speed in DJGPP programs

Q: I measured the time required to read a 2 MByte file in DJGPP and in Borland C. It took the DJGPP program 2.5 seconds to do it, while Borland did it in just under 2. This is horribly slow: it's 25% slower than Borland!

Q: I tried to improve DJGPP I/O throughput by defining a 64KB-large buffer for buffered I/O with a call to setvbuf, but that had no effect. Why is that?

Q: It is obvious that disk-bound programs compiled with DJGPP will run awfully slow, since FAT is such a lousy filesystem!

A: First, I would like to point out that waiting another 0.5sec for reading a 2 MByte file isn't that bad: it is indeed about 25% longer than you can do under DOS, but it's only half a second.... Besides, most programs read and write files which are only a few hundreds of kilobytes, and those will suffer only a negligible slow-down.

Doing I/O from protected-mode programs requires that low-level library functions move the data between the extended memory and low memory under the 1 MByte mark, where real-mode DOS can get at it. That area in the low memory is called the transfer buffer23. This data shuffling means that some I/O speed degradation is inevitable in any protected-mode program which runs on top of DOS (including, for example, Windows programs when Windows 3.X is set to 386-Enhanced mode).

By default, DJGPP moves data in chunks of 16 KB, so defining a buffer larger than that won't gain anything. The size of the transfer buffer is customizable up to a maximum of 64 KB24, so if your program really reads a lot of large files, you might be better off enlarging it (with the STUBEDIT program).

The DJGPP buffered I/O functions utilize a special algorithm to optimize both sequential and random reads. These two usually contradict, since sequential reads favor larger buffers, while random access favors small buffers. DJGPP solves this contradiction by doubling the buffer size on each sequential read, up to the size of the transfer buffer, and resetting the buffer size back to the minimum of 512 bytes each time the program calls fseek. Experience shows that programs which use both sequential and random access to files, like ld.exe, the linker, run significantly faster when linked with these optimized I/O functions (introduced with version 2.02 of DJGPP).

Some people think that FAT is such a lousy filesystem, that programs which do a lot of disk I/O must run terribly slow when compiled with DJGPP. This is a common misconception. The speed of disk I/O is determined primarily by how efficient is the code in the operating system kernel that handles the filesystem, and the device drivers for the I/O-related devices like the hard disk, not by the disk layout. It is true that DOS and BIOS don't implement I/O too efficiently (they use too many tight loops waiting for low-level I/O to complete), but a large disk cache can help them tremendously. In addition, Windows 9X bypasses DOS and BIOS I/O code entirely, and uses much more efficient protected-mode code instead. Experience shows that DJGPP programs on plain DOS systems with a large (8MB and up) disk cache installed run about 30% slower than a Linux system on the same machine; and Windows 9X will run the same programs at roughly the same speed as Linux. If you get much slower performance on DOS/Windows, chances are that your system is not configured optimally.

Some programs which only copy data between two files might gain significantly if you write your custom low-level I/O functions that avoid moving data to extended memory (only to move them back to the transfer buffer). However, these cases are rare.

Node: Slow-down, Previous: IO speed, Up: Performance

14.5 My ported program runs much slower!

Q: How come my program, which I ported from Borland/MS C and which doesn't use much I/O, still runs much slower under DJGPP?

A: Explore the following possible causes for this:

  1. You compiled the problem without optimizations. You should use at least -O2 to produce optimized code.

    If your program spends most of its time in a certain innermost loop, you should try enabling some of the optimization options which aren't enabled by -O2. Some of these are described in this FAQ, see speed-related optimization options.

  2. Your program extensively calls services other than I/O which require mode switch (like BIOS Int 10h, mouse services, etc.).

    You can tell how much your program switches to real mode by profiling your program. In the profile, look at the proportion of time your program spends in low-level library functions called __dpmi_int (which calls real-mode DOS/BIOS services) and __dj_movedata (which moves data between the transfer buffer and your program). If this proportion is large, try rewriting your program to minimize use of those functions which require a mode switch, even at a price of more computation (a mode switch usually eats up hundreds of CPU cycles).

  3. Your program might be running out of available physical memory and paging to disk. Watch the disk activity to find out whether this is the reason. If it is, you will have to configure your system differently (see system configuration), or change the way your program allocates memory.

    Sometimes, some device driver that uses extended memory takes up a significant portion of it, and leaves less for DJGPP programs, which then begin to page and slow down. For example, Novell Netware's VLM redirector and client software can use up to 0.5 MB of extended memory, even if you don't log into the network. A solution is not to load such resident software, or to buy more memory.

  4. Your program uses a lot of floating-point math, and you run it on a machine without an FPU. A tell-tale sign of this is that a function called __djgpp_exception_processor is high on the execution profile printed by Gprof. Due to the way FP emulation is implemented in DJGPP25, it might be significantly slower than the way real-mode DOS compilers handle it. The solution is either to rewrite your code so that it doesn't use floating-point code in its inner loops, or buy an FPU.
  5. Your program uses library functions/classes which are implemented less efficiently by DJGPP libc and the GNU C++ libraries. Or you might be a heavy user of functions which other compilers convert to inline code, while GCC doesn't inline most library functions. If this is the case, you will see those functions as "hot spots" on the program histogram produced by the Gprof profiler. If you find this to be the problem, write your own, optimized versions of those functions. It's best to write them as inline assembly functions, for maximum performance. If you find library functions which are inefficient, please inform the DJGPP news group by posting to the comp.os.msdos.djgpp news group, so this could be fixed by people who maintain the library.
  6. The size of the code/data in the innermost loop might be close to the size of the CPU cache (either L1 on-chip cache, or L2 cache on the motherboard). Compiling your program with a different compiler or a different combination of optimization options can cause the code to overflow the cache, which will dramatically affect the performance (usually, by a factor of 2). Running your program with the cache disabled will be instrumental to see whether this is your problem. If it is, try to rearrange your data/code, or use a different combination of optimization options.
  7. If the slow program was compiled for profiling (with the -pg switch), the slow-down might be due to a bug in the DJGPP library. See slow-down in profiled programs, for more about this.

Node: Memory, Next: , Previous: Performance, Up: Top

15 Run-Time Memory Issues

This chapter answers questions which are related to DJGPP run-time memory allocation.

Node: How much memory, Next: , Previous: Memory, Up: Memory

15.1 How much virtual memory do you have?

Q: How much virtual memory can I use in DJGPP programs?

A: That depends on the DPMI host you are using. The latest version r5 of CWSDPMI (the free DPMI host which comes with DJGPP) lets you use all the available extended memory, plus the available hard disk storage up to a total of 2GB. (Version r4 of CWSDPMI supports up to 256MB of physical memory and up to 256MB of disk space, for a grand total of 512MB of virtual memory, but has bugs when the total memory is more than 255MB.). Try a malloc(50*1024*1024) some day.

With other DPMI hosts, your mileage may vary. Quarterdeck's QDPMI, for instance, has a bug in some of its versions which effectively disables virtual memory under DJGPP (described in QDPMI VM bug, below), so you only have whatever free physical RAM is left. On Windows 3.X, the amount of virtual memory you get depends on various virtual memory settings in the Control Panel and on the .pif file settings for the program you run (see Windows allocation subtleties, below). On Windows 9X, the memory settings of the DOS Box Property Sheets define how much virtual memory a DJGPP program will get (see Windows 9X allocation details, below). OS/2 reportedly can be configured to support up to 512MB of DPMI memory.

Node: Confusing alloc, Next: , Previous: How much memory, Up: Memory

15.2 It seems malloc/free don't affect virtual memory...

Q: I did malloc(50*1024*1024), but didn't see any paging happen, and I only have 8 MBytes of RAM on my machine. Is this virtual memory thing for real?

Q: I malloc'ed a large chunk of memory, but when I check values returned by _go32_remaining_physical_memory or __dpmi_get_memory_information, I don't see any change!

Q: When I free allocated RAM, _go32_remaining_physical_memory reports there was no change in the available RAM....

Q: I'm looking for a way to tell how much memory is available, something like coreleft in Borland C?

A: CWSDPMI (and, possibly, other DPMI hosts) only pages in memory when it is actually accessed. If you only malloc it, but don't actually access it, it won't grab those pages. Try calloc and see the big difference.

When you call free, DJGPP library doesn't return memory to the system, it just adds it to its internal pool of free pages. So, from the point of view of the DPMI server, these pages are not "free".

In addition, several widely-used DPMI servers, such as those built into Windows, have their own quirks related to memory allocation. For example, some of them won't let you allocate more than half the available memory in a single chunk. As another example, on OS/2 _go32_remaining_physical_memory reports a constant very large value that doesn't change in the course of the program.

Note that the distinction between the physical and virtual memory is only meaningful on DOS. More sophisticated operating systems usually conceal the difference entirely, so only the sum of these two types of memory usually (but not always) gives an approximation of the total free memory.

Because of these peculiarities, there's no convenient and easy way to return the amount of free memory available at any given moment, even in plain DOS. Some programs only care about available physical RAM (they don't want to page to disk, since that causes a considerable slow-down); for these, I recommend to call the _go32_remaining_physical_memory library function at program startup, and then track memory usage with sbrk(0);. Alternatively, disabling virtual memory altogether (by using CWSDPR0 or by loading CWSDPMI with -s- parameter), and checking values returned by malloc against NULL, might be all you need to know when you are about to run out of free physical memory. Programs that need to know when they are about to run out of virtual memory should call _go32_remaining_virtual_memory instead. Once again, these methods will only work reasonably well in plain DOS.

Node: QDPMI VM, Next: , Previous: Confusing alloc, Up: Memory

15.3 Failure to get more memory than is physically installed

Q: When I try to access more memory than the free physical RAM, malloc returns a NULL pointer, or I get some cryptic error message, like "Memory Paging Violation" or "Unrecoverable Exception: 000Eh".

A: This is typical of Quarterdeck's DPMI host called QDPMI which comes with QEMM386 version 7.53 and earlier. Some versions of QDPMI (those which come with QEMM v6.x) fail to resize memory blocks when the new size is more than the available physical RAM, even though virtual memory services are enabled; other versions (those which come with QEMM v7.x) just don't let you allocate more memory than is physically available. If you must use more RAM than is physically available, disable QDPMI by going to the QEMM directory and typing this: 

 qdpmi off

DJGPP programs will then use CWSDPMI instead.

This bug was corrected in QDPMI version 1.10 or later, distributed with QEMM beginning with version 8.0, so upgrading to the latest version of QEMM might also be a solution. With QEMM 6.x, make sure your programs don't set _crt0_startup_flags to _CRT0_FLAG_UNIX_SBRK, which overrides the default type of sbrk (QEMM 8.0 and later can allocate virtual memory with both types of sbrk algorithm).

If you use another DPMI host, make sure that virtual memory is enabled. E.g., for 386Max, include the swapfile= parameter to establish a virtual memory swap file; you can make it permanent (this will speed up DJGPP start-up) with the /p option.

Node: QDPMI alloc, Next: , Previous: QDPMI VM, Up: Memory

15.4 Memory allocation fails before all memory is used

Q: OK, I've changed my program to never allocate more memory than is physically available, to work around that QDPMI VM bug, described in previous section, but my program still gets a NULL pointer from malloc/calloc!

Q: Why is my program dying with SIGSEGV under CWSDPMI when allocating a chunk of memory?

A: Another peculiarity of QDPMI which came with QEMM before version 8.0: it will never let you allocate a chunk which is larger than half of what's available. Windows 3.X behaves in the same way, and several people reported the same to be true under Windows 9X.

If your program asks for memory in lots of small allocations, then it might crash when you use CWSDPMI as your DPMI host. This is because CWSDPMI runs out of its tables, allocated in the heap, where it tracks memory allocations. Beginning with release 2, CWSDPMI defines a 6KB-large default heap that is configurable by CWSPARAM program to be anywhere between 3K and 40K bytes, without recompiling CWSDPMI. The default heap size is enough for about 21MBytes in small chunks. You should upgrade to the latest CWSDPMI if you experience such problems, and bump up its heap size as needed.

Node: Windows alloc, Next: , Previous: QDPMI alloc, Up: Memory

15.5 Memory allocation fails under Windows

Q: I'm running in Windows 3.X DOS box, but DJGPP complains about there not being enough DPMI memory, although virtual memory is enabled.

A: You must make sure the size of your Windows swap file can be at least 2 times the largest virtual memory size you need. Check if you have enough free disk space; if you do, run a defragger (Windows needs the swap file to be contiguous). The size of the swap file is normally limited by the "virtual = 4 times free physical" rule, but you can change that by inserting the line 

 PageOverCommit=n

in the [386Enh] section of your SYSTEM.INI file. The parameter n is 4 by default, but can be set to be as large as 20.

Node: Windows9X alloc, Next: , Previous: Windows alloc, Up: Memory

15.6 Memory allocation peculiarities under Windows 9X

Q: I seem to be unable to get more than 16 MBytes of virtual memory under Windows 9X, even though I have 32 MBytes of RAM installed on my machine, and a lot of disk space....

A: If your machine has 64MB or less of main memory, you must set the maximum amount of DPMI memory to 65535K in the DOS box Property Sheets. Click on the Properties button of the DOS box toolbar, then click on the Memory tab, and type the number 65535 in the box under DOS Protected Mode Memory. If you leave that setting at the default "Auto", your programs are at Windows' mercy, and in many cases will get only 16 MBytes. You must actually type 65535 inside the dialog box, as checking out the values from the list Windows offers will never get you past 16384 (i.e., 16MB).

Machines that have more than 64MB of physical memory should always leave the DPMI memory setting at "Auto", since the manual setting cannot be larger than 65535K. Setting the DPMI memory property to "Auto" usually leaves the DOS box with whatever is physically installed, minus about 5MB, assuming the disk used by the Windows swap file has enough free space.

Some users report that they need to edit their EMM386 command line in the CONFIG.SYS file to say this: 

 DEVICE=C:\WINDOWS\EMM386.EXE NOEMS L=131072

The L=NNN parameter (here for 128MB of installed memory) forces EMM386 to use all of the installed memory. (This should work by default with Windows 9X, but if it doesn't, try the above line.)

Note that you cannot allocate more than half the available memory in one chunk under Windows 9X, exactly as the things are under Windows 3.X, and you cannot have more than 64 MBytes of virtual memory available to DJGPP programs running on Windows, unless you have more than 64MB physical memory installed.

Node: EMM386 alloc, Next: , Previous: Windows9X alloc, Up: Memory

15.7 Memory allocation fails under EMM386 or HIMEM

Q: My machine has 48 MBytes of RAM, but when I run DJGPP programs, they start paging after 32 MBytes have been used....

Q: I have 5 MBytes of free RAM on my machine, but DJGPP programs start paging after only 256KBytes of memory were used??

A: This might be caused by some old versions of the memory manager installed in your machine (like HIMEM or EMM386 from an old version of DOS), which were limited to 32 MBytes of extended memory. Try running without them (CWSDPMI can use raw extended memory), or upgrade to a newer version of DOS.

If your programs start paging after only 256KBytes of memory were used, most probably you are using EMM386 and CWSDPMI, and your CONFIG.SYS specifies no amount of memory when it installs EMM386. Some old versions of EMM386 default to 256K in this case; you should tell EMM386 explicitly how much memory it should take over. You can use the go32-v2 program to see what amount of extended memory your DJGPP programs will get.

Node: Swap out, Next: , Previous: EMM386 alloc, Up: Memory

15.8 How much memory do parent DJGPP programs leave for their child?

Q: How much memory is available when I invoke other programs from my DJGPP program?

A: In the conventional (below 640K mark) memory, you are left with everything which was free before your program started, except what the DPMI host uses. The amount of conventional memory required by the DPMI host depends heavily on the host you use. For the first DJGPP program, CWSDPMI uses about 130KB (including 41KB to load CWSDPMI itself), QDPMI uses about 55KB, and Windows only 18 KBytes. Each subsidiary call to system or spawn (like in recursive invocation of Make) eats up about 18K (16K for the transfer buffer and 2K for the PSP and environment) for most DPMI servers; a notable exception is QDPMI which needs 104K bytes of low memory for the subsequent calls. If you change the size of the transfer buffer (with STUBEDIT), the amount of free conventional RAM left when shelling out of it will change accordingly.

In addition, most DPMI servers (with the notable exception of Windows) take up 16KB of expanded memory when they first load.

Extended memory management for the spawned programs is left to the DPMI server; DJGPP does nothing special about XMS when system or spawn is called. This means that all the extended memory used by the parent program is not freed when the child program starts; if the child requests more memory than is physically free, the DPMI server is expected to page some of the parent out to honor the request. (This is unlike DJGPP v1.x, where the go32 extender would completely page out the parent before starting the child.) The advantage of this is that spawning a child or shelling out is much faster in v2 than it used to be with v1.x, except on machines with low amounts of installed RAM. A disadvantage is that if you spawn a real-mode program that uses XMS, the extended memory used up by your DJGPP program will be unavailable to it, unless you use a memory manager (as opposed to when CWSDPMI uses raw XMS or HIMEM).

Note that if you use a memory manager such as EMM386 or QEMM386 with the NOEMS and NOVCPI parameters, CWSDPMI will use the XMS (as opposed to VCPI) services to allocate extended memory, and will allocate all of the available XMS memory for itself. So if, while your DJGPP program runs, some resident software such as device driver or TSR will try to allocate XMS, the allocation will fail.

Node: Stack size, Next: , Previous: Swap out, Up: Memory

15.9 How much stack can I have in DJGPP programs?

Q: My program bombs when I use very large automatic arrays.

Q: How much stack space do I have in my program?

Q: My program seems to overflow the stack, but only when I run it under a debugger....

Q: My program crashes with SIGSEGV, but the traceback makes no sense: it points to something called ___djgpp_exception_table... When I try to debug this, the traceback mysteriously changes to some innocent library function, like getc(). What is going on??

A: DJGPP v2 programs get fixed-size stack which is allocated by the startup code and then stays fixed for the entire lifetime of the program; this is due to a bug/feature of the DPMI 0.9 specification26. By default, you have a 512KB-long stack (DJGPP v2.01 and earlier used 256KB stack), but some programs which use large automatic arrays, or are deeply recursive, might need more. If the default stack size is not enough, you can change it with the STUBEDIT program (change the parameter "Minimum amount of stack space"), or by setting the global variable _stklen in your program. Example: 

 unsigned _stklen = 1048576;  /* need a 1MB stack */

The DJGPP startup code checks both the value in the stub (that can be changed by STUBEDIT) and the value of _stklen, and uses the larger of these two. Therefore, programs that are known to require large stack size should set _stklen to make sure they will always work, even if somebody stub-edits them to a lower value. Setting _stklen is also safer to ensure sufficient stack size during debugging (see below). However, you might be left with STUBEDIT as your only option of enlarging the stack with programs for which you don't have the sources handy, or rebuilding which is not practical.

Alternatively, you could rewrite your code to declare large arrays with the static qualifier, or put their declaration outside any function (thus making them static by default). Static arrays don't use stack space at all.

Programs that need an unusually large stack might crash with bogus stack traces, because part of the static data gets overwritten by the overflowing stack. To see if that is the cause of such crashes, run STUBEDIT on your program and crank up the stack size to a large value (like 4 MBytes). If that makes the problem go away, tune the stack limit to the minimum value your program can live with, then set _stklen to an appropriate value as explained above and recompile the program. (Some DPMI hosts will actually allocate the entire stack, even if not all of it is used, so leaving it at unnecessarily large value will hurt the program on low-memory machines.)

Some users have reported that they needed to enlarge the stack size of the C++ compiler, cc1plus.exe, to prevent it from crashing when compiling some exceedingly large and complex C++ programs. Another program that was reported to need a stack larger than the default is bccbgi.exe from the BCC2GRX package.

After you've used STUBEDIT to change the stack size, run it again to make sure it displays as default the value you thought you entered. This is because STUBEDIT will sometimes silently set the stack size to 0 (and then you will get the default 512K stack) if it doesn't like the value you type (e.g. if it has a wrong syntax).

When you run a raw COFF image under a debugger, the stack size is taken from the debugger's stack size, which might not be appropriate for your program . So the only way to change the default stack size in these cases is to set _stklen. You can also stubedit the debugger, to achieve the same effect, albeit at a price of more memory used by the debugger.

Under Windows 3.X, be sure you've allocated a sufficiently large swap file (let's say, 40MBytes) from the Windows' Control Panel, and make sure the .PIF file for your program doesn't have too low limit on EMS/XMS usage (better make them both -1). What's that? You don't have a .PIF file for this program? Then Windows uses the default file DOSPRMPT.PIF, which almost surely defines very low limits on these two, and your program might have problems getting the memory it needs for its stack.

DJGPP v2.0 has a subtle bug in its startup code that is seen very rarely, and that manifests itself by a program crashing with Page Fault or SIGSEGV. If you are using v2.0 and enlarging the stack and the CWSDPMI heap size didn't help, try adding some (e.g., 4KB) static data to your program and see if that helps. But the best way to overcome this is to upgrade to DJGPP v2.01 or later.

Node: Windows 98, Previous: Stack size, Up: Memory

15.10 Memory-related problems in Windows 98

Q: Whenever I run a batch file that launches a DJGPP program, Windows 98 pops up a message saying that "This program has accessed memory in use" and threatens me with all kinds of trouble unless I terminate the program. What's wrong with DJGPP programs??

A: Nothing's wrong with DJGPP programs, it's some weird Windows bug, and it is not limited to DJGPP programs, either. I'm told that to get rid of these messages, you need to change the settings in the DOS box Properties, under Memory. Specifically:

After doing this, click OK, close the DOS box and reopen it again. The problem should now go away.

Btw, the message popped up by Windows is a mere nuisance, since if you answer YES to Windows' question whether you want to continue, everything works just fine, and the message is never popped up again until you reboot the machine. Seems like just one more of those Windows teasers....

Node: Command line, Next: , Previous: Memory, Up: Top

16 Command-line Arguments Handling in DJGPP

DJGPP handles command-line arguments differently from most DOS-based compilers, to make it closer to Unix platforms (so that porting of Unix programs will be easier). This chapter answers some questions about this aspect of DJGPP.

Node: Filename globbing, Next: , Previous: Command line, Up: Command line

16.1 Filename wildcards expansion under DJGPP

Q: Can I do filename globbing with DJGPP?

Q: I call my program with an argument x*y and it complains about something called xyzzy??....

Q: I cannot find a way to use the /? switch with my programs!

A: The filename globbing in DJGPP is done by the start-up code, before your main function is called. Unlike other DOS-based compilers, where you must link your program with a special object module if you want the program to get expanded filenames, in DJGPP this is considered normal behavior and is performed by default on behalf of every DJGPP program. The x*y above was expanded to a file called xyzzy which was probably present in the current working directory; and /? is by default expanded to the list of one-letter files/directories you happen to have in the root directory of the current drive. (If you don't want the default expansion, refer to how to disable globbing.)

In DJGPP, filename globbing works like in Unix, which is more general than the usual wildcard expansion, both in DOS and even in Windows. The DJGPP wildcard expansion understands the following constructs with special meta-characters:

?
any single character.
*
zero or more arbitrary characters, including a dot `.'
[aA_]
any one of characters `a', `A', or `_'.
[a-d]
any one of characters `a', `b', `c', or `d'.
[!a-z]
anything but a lowercase letter.
...
all the subdirectories, recursively (VMS aficionados, rejoice!).
.../*
all the files in all subdirectories, recursively.

Unlike DOS, the * and ? meta-characters can appear anywhere in the filename pattern, like in [a-z]*[0-9].*pp. You can also use * instead of directories, like in */*/*.c, but not on drive letters (e.g., [a-c]:/ won't work).

Note that *.* only picks up files that actually have an extension. This is contrary to the usual DOS practice where it means all the files, with or without extension. Use * to get files with and without extensions.

An argument which cannot be expanded (no filenames matching that particular pattern) will be passed to the program verbatim. This is different from what you might see under Unix, where some shells (like csh) would say something like "No match" and won't call your program at all. DJGPP's behavior in this case is like shells of the Bourne legacy (sh, ksh, and bash).

If the wildcards include upper-case or mixed upper- and lower-case letters, the letter-case of the files is not ignored on Windows 9X when expanding the wildcards. For example, [A-D]* will not match a file called aFileName. Upper-case letters in wildcards also disable automatic down-casing of short 8+3 file names returned by the code that expand wildcards (even on plain DOS). By contrast, if the wildcards include only lower-case letters, the letter-case is ignored during expansion, and short 8+3 file names are automatically down-cased, unless the environment variable FNCASE is set to y. The effect of setting FNCASE is fully described in the DJGPP C Library reference, under the _preserve_fncase function; type info libc alphabetical _preserve_fncase from the DOS prompt.

Node: Disable globbing, Next: , Previous: Filename globbing, Up: Command line

16.2 How to disable filename wildcards expansion

Q: I don't want my program to glob its arguments (they aren't files at all, but they include characters like * and ?). What should I do?

A: You have these alternatives:

Node: Special chars, Next: , Previous: Disable globbing, Up: Command line

16.3 How to pass command-line arguments with quotes or @

Q: I have a file with a single quote in its name, but the quote seems to be stripped away when I pass it to my program ....

Q: How do I pass a command-line argument which contains double quotes?

Q: How do I pass an argument which begins with the @ character?

A: These special characters on the command-line arguments are handled by the filename expansion ("globbing") code before they are passed to the main function (see description of filename expansion), and the quote characters serve to protect the arguments from expansion. You should escape-protect the quote characters with a backslash in order for them to be treated as literal characters. For example, if you have a file called myfile.c'v, type it as myfile.c\'v when you call your program. If you have single quotes in your program arguments and you don't want those arguments to be expanded, then surround them by double quotes, like this: "*.c'v". The program will get the string *.c'v with the double quotes stripped away.

Note that backslashes are only special if they are in front of a quote or a backslash (they also serve as DOS directory separators, remember?). This is also different from what you get under a Unix shell, where a backslash quotes any character.

The @ character serves to signal a response file (see the description of response file method), so it's also special. To pass an argument whose first character is @, surround that argument with single or double quotes, otherwise it will be taken as a name of a response file which holds the actual command line.

You can quote only some parts of the wildcard to protect only those parts from expansion; the unquoted parts will still be expanded. This allows to use wildcards with embedded whitespace and expand file names with special characters which need to be quoted, like in c:/Prog*' 'F* (which should expand into c:/Program Files) and *.c"'"v (which should expand into all files with the *.c'v extension).

Node: Long commands, Next: , Previous: Special chars, Up: Command line

16.4 How to pass command lines longer than 126 characters

Q: Can I invoke my program with a command line longer than the infamous DOS 126-character limit?

Q: I have a Makefile of Unix origin which contains some very long command lines. Will it work with DJGPP?

A: Yes and yes. DJGPP supports several methods of passing command-line arguments which allow it to work around the DOS 126-character limit. These are:

Of course, if the DJGPP start-up code doesn't see any of the above methods, it will use the DOS command line by default.

Since the long command lines are a very important feature, DJGPP's version of the system library function avoids calling the DOS command processor, COMMAND.COM, unless it needs to run a batch file or an internal command of COMMAND.COM. Other features of the command processor, like redirection and pipes, are emulated internally by system. See the library reference for system, for more details about its extended functionality.

Node: How long, Next: , Previous: Long commands, Up: Command line

16.5 What is the maximum length of command line under DJGPP

Q: What is the longest command line I can pass to gcc when it is invoked by Make?

A: The arguments are passed to DOS Exec call (Int 21h function 4Bh) via the transfer buffer which is 16KB-long. Apart of the command line, it is also used to pass other info, such as the !proxy parameters and the copy of the environment for the child program (let's say, less than 2000 bytes in most cases, but your mileage may vary). This leaves at least 13K bytes for arguments (including a separating blank between any two arguments). So unless your arguments span more than 160 screen lines, you shouldn't worry. However, if your environment is very large (some people need as much as 6KB to accommodate for all the variables used by the various packages installed on their machine), be sure to stub-edit the programs that spawn other programs to a larger transfer buffer, or else they could fail.

The above limit depends on the size of the transfer buffer, so check the size of the value recorded in the stub header of the parent program before you pass extremely long command lines to its children. GCC is the first program you should worry about, because the linker (ld.exe) usually gets long command lines from GCC (they include the list of all the object files and libraries to be linked).

Node: Makefiles, Previous: How long, Up: Command line

16.6 Why Make passes only 126 characters to programs?

Q: I use Make to compile with GCC, but GCC gets only the first 126 characters of its command line. Didn't you just explain in so many words that invoking a DJGPP program (GCC) from another DJGPP program (Make) can safely pass up to 13K characters of command-line arguments using the !proxy method?

Q: I use RHIDE, but it only passes the first 126 characters of my long command lines to the compiler!

A: If you use Make 3.73, check your Makefile for SHELL = command.com statements, or for commands which include pipe or redirection characters like >, |, etc. If Make sees any such statements, it will invoke COMMAND.COM to run GCC, and COMMAND.COM can't pass more than 126 characters to GCC. To work around, comment-out the SHELL= line, and change your commands to work without redirection/pipe characters. One easy way to get rid of redirection characters without losing their effect is to use the redir program which comes with DJGPP. For example, the following command: 

  frobnicate foo.bar > myfile.tmp

can be re-written instead like this: 

  redir -o myfile.tmp frobnicate foo.bar

The ports of Make 3.75 and later don't call COMMAND.COM in the above cases, but rather emulate pipes and redirection internally, so upgrading to the latest version of Make will solve such problems. If you think about using Make 3.75 with DJGPP v2.0, don't: invoking v2.0 programs from v2.01 programs will cause subtle and hard-to-debug problems due to incompatibilities between these two versions regarding the methods of invoking child programs (in particular, v2.0 doesn't support the environment method of passing long command lines described above).

If you have problems with long command lines when using Make 3.75 and later, it might be caused by the environment variable SHELL that points to a Unix-style shell which is not a DJGPP program. This could happen, for example, if you use the shell from the MKS toolkit. When SHELL points to a program whose name looks like it's a Unix-style shell (sh.exe, bash.exe, etc.), library functions like system will invoke the shell to do everything, instead of using the internal shell emulator. If that shell isn't a DJGPP program, the long command lines will end up truncated. To solve these problems, either unset SHELL or override it by setting MAKESHELL to point to COMMAND.COM or to the DJGPP port of Bash. This is further explained in the GNU Make manual (see SHELL and MAKESHELL).

Problems with passing long commands from RHIDE are usually caused by invoking old programs compiled with DJGPP v2.0. Upgrade to the latest binaries.

Node: Converting, Next: , Previous: Command line, Up: Top

17 Converting DOS Programs/Libraries to DJGPP

If you have a program or a library developed under some other DOS-based compiler, which you want to port to DJGPP, read this chapter.

Node: Syntax, Next: , Previous: Converting, Up: Converting

17.1 GCC/Gas won't accept valid assembly code ...

Q: I have some code written in assembly which compiles under MASM and TASM, but GCC gives me a long list of error messages.

A: The GNU Assembler (as.exe), or Gas, called by GCC accepts the AT&T syntax, which is different from the Intel syntax. Notable differences between the two syntaxes are:

Examples: 

    Intel:  [ebp - 4]         AT&T:  -4(%ebp)
    Intel:  [foo + eax*4]     AT&T:  foo(,%eax,4)
    Intel:  [foo]             AT&T:  foo(,1)
    Intel:  gs:foo            AT&T:  %gs:foo

For a complete description of the differences, see GNU assembler documentation. If you don't read this FAQ with an Info browser, download GNU Binutils, unzip the files named as.iN (where N is a digit) from it, then type at the DOS prompt: 

 info as machine i386

You will see a menu of Gas features specific to x86 architecture.

A user guide for inline assembly was written by Brennan Underwood; it describes how to use inline assembly programming with DJGPP and includes a tutorial on the AT&T assembly syntax. Check out the DJGPP inline assembly tutorial.

Another useful tutorial about writing separate assembly-language modules for DJGPP was written by George Foot and is available from George's home page.

The DJGPP User's Guide also has a tutorial on writing assembly-language code. One of the sections there describes the CPU architecture, which is geared towards assembly-language programming.

Yet another tutorial on the subject of inline assembly is available at <http://www.castle.net/~avly/djasm.html>.

Many people who used Intel syntax and then converted to the AT&T style say that they like the AT&T variant more. However, if you prefer to stick with the Intel syntax, download and install NASM, which is a free portable assembler. It is compatible with DJGPP and accepts a syntax which is much more similar to the Intel style. A guide for using NASM with DJGPP was written by Matthew Mastracci and is available from Matthew's Web page.

Note that Binutils maintainers are working on adding an option to Gas which will cause it accept the Intel syntax as well, so it is most probable that beginning with Binutils 2.10, Gas will have this feature.

Node: Converting ASM, Next: , Previous: Syntax, Up: Converting

17.2 Converting between Intel ASM syntax and AT&T syntax

Q: Where can I find an automated conversion tool to convert my Intel-style assembly code into a code acceptable by Gas?

Q: Where can I find a converter from AT&T assembly to Intel style?

A: A SED script which should do most of the conversion was posted to the DJGPP news group.

A program called TA2AS which can convert TASM assembly source to AT&T style can be found on the DJGPP server and on Oulu. TA2AS was written by Frank van Dijk of the Spirit group; if you have questions about this tool, you may contact Jan Oonk. The authors say that the program is far from finished, but the sources are available with the package so you can fix whatever is broken for your needs.

Another similar converter is Intel2Gas, available from its Web page.

Beginning with Binutils 2.10, Gas has an option that causes it to accept the Intel syntax, so you can use Gas to assembly Intel-style code.

Alternatively, here is what you can do to make your code linkable with DJGPP programs:

Keep in mind that syntax is only one of the aspects of converting code written for DOS to DJGPP. You should also make sure your code doesn't violate any of the rules for protected-mode programming (see GPF in asm code).

If you need to perform the opposite conversion, from the AT&T style to the Intel style, try the Att2Intl converter written by Gregory Velichansky. Its output is intended for NASM or TASM. Att2Intl is available from Greg's home page.

Node: ASM GPF, Next: , Previous: Converting ASM, Up: Converting

17.3 Converted code GP Faults!

Q: OK, I've succeeded in converting and compiling my assembly-language program, but when I run it, I get "Segmentation Violation" and "General Protection Fault". This program works when compiled with MASM, so how can this be?

A: In DJGPP, your program runs in protected mode. There are certain things you can't do in protected-mode programs (that's why it is called protected mode). This issue is too complex to describe here, so only a few of the more important aspects will be briefly mentioned. If you are serious about writing assembly language protected-mode code, or have a large body of existing code to convert to protected mode, you should read any of the available books about protected-mode programming with 80x86 processors.

Here is a short list of some of the techniques found in many real-mode programs, which will trigger protection violation or erratic behavior in protected mode:

If your code uses one or more of these techniques, refer to low-level programming chapter, which describes the DJGPP facilities that will allow you to rewrite your code.

Node: ASM and C, Next: , Previous: ASM GPF, Up: Converting

17.4 Problems with combining assembly and C/C++ modules

Q: Which register can I safely change in my assembly code that is called from a C program?

Q: How come my program that calls assembly-language functions crashes with a GPF, but only if I compile it with -O2?

Q: When I try to link my assembly modules with a C++ program, the linker complains about the functions I wrote in assembly!

A: You can safely clobber EAX, ECX, EDX, FS and GS, as well as EFLAGS and the floating-point registers (including the FPU control and status words), but must save and restore all other registers at the end of your assembly function. Failure to preserve, e.g., ESI, EDI, EBX, ESP or DS in functions written in assembly can cause a C program linked with such functions to crash, since GCC expects those registers to be preserved across function calls. Special-purpose registers such as GDTR, LDTR, CR*, and DR*, although GCC and the DJGPP library don't use them, should probably not be touched at all, but if you do, it's a good idea to save and restore them.

Functions written in assembly or in C that are meant to be linked with C++ programs should be declared extern "C" in their prototype, like this: 

#ifdef __cplusplus
extern "C" {
#endif

int my_assembly_func (int);

#ifdef __cplusplus
}
#endif

This example shows how to produce a prototype that would work with both C and C++ programs; it is usually placed in a .h header file that is meant to be #included in the C or C++ programs.

This extern "C" declaration prevents the C++ compiler from mangling the names of external functions according to the usual C++ rules. (The name-mangling is needed because C++ allows several different functions to have the same name but different types of arguments.)

Node: OBJ and LIB, Next: , Previous: ASM and C, Up: Converting

17.5 I want to use a .obj or .lib code with DJGPP

Q: I have a set of useful functions in a .obj format, but no source code. Can I use them with my DJGPP program?

Q: I have this ACMELUXE.LIB library of functions which I want to use. I've extracted all the .obj files, but when I try to link them with my program, GCC complains: "File format not recognized". Can't I use these object files?

Q: I've got a bunch of .obj files I want to use. I've ran AR to make a GCC-style .a object library, but got an error message from GCC saying "couldn't read symbols: No symbols". How can I link them with my code?

A: Sorry, you probably can't. The GNU linker called by GCC doesn't understand the format of .obj files which other DOS-based compilers/assemblers emit. Unless you can get the source of those functions, convert it to protected-mode, flat-address model code and compile them with GCC, you most probably won't be able to use them29.

However, if you are really desperate, one conversion tool you might try is OBJ2BFD. It was written by Robert Hoehne based on the EMXAOUT utility from the emx/gcc package. OBJ2BFD requires the .obj files to be written for the flat-address memory model and will reportedly complain if you feed it with code written for segmented memory models. OBJ2BFD is available from the DJGPP sites.

Another automated conversion tool called OBJ2COFF was written by the SPiRiT team, and it can be used to convert .obj object files and .lib libraries to COFF format, provided that the original .obj files have been written for flat-address memory model.

OBJ2COFF is available via anonymous FTP transfer from the Oulu MSDOS repository. If you have any problems with it or questions about it, send them to its author Rico or to George van Venrooij. Note that the authors of OBJ2COFF have explicitly prohibited commercial use, so you shouldn't use OBJ2COFF for converting commercial object files.

You can also try using LIB32 librarian from Microsoft C8 to convert object files to COFF.

The main problem with these conversion methods is, of course, that most object files you'd want to converted were written for real-mode programs in memory models other than flat, and without extensive modifications would crash your program anyway.... (See previous question.)

Node: 16-bit code, Next: , Previous: OBJ and LIB, Up: Converting

17.6 I must use my 16-bit code with DJGPP!!

Q: If I cannot use 16-bit .obj files, then I would have to give up using DJGPP. I simply cannot live without these .obj files. Are you really sure there is nothing I can do??

A: If you need your old code that badly, then there might be a way, albeit a cumbersome one. You can write a 16-bit, real-mode program and link it with your precious functions you can't live without. Have this program spawn a DJGPP-compiled program and make the two communicate with each other via a buffer allocated in low memory, or via command-line parameters passed to the 32-bit program by the spawnXX function call. On the DJGPP side, you can directly call 16-bit functions from the real-mode program using the library function called __dpmi_simulate_real_mode_procedure_retf, provided the 16-bit program passes the CS:IP values of these functions to the 32-bit program. You can even put your 16-bit code as binary instructions into a buffer allocated in low memory and call it with __dpmi_simulate_real_mode_procedure_retf (but if you can do that, you can probably also disassemble the code into a source form and submit it to Gas).

Now will you consider sticking with DJGPP? Please??...

Node: NEAR and FAR, Next: , Previous: 16-bit code, Up: Converting

17.7 What should I do with those "near" and "far" declarations?

Q: I have this program that I need to port to DJGPP, but it is full of pointers and functions declared with the "near" and "far" keywords which GCC doesn't grok. What shall I do?

Q: A program written for a 16-bit compiler uses the MK_FP or _MK_FP macro, but DJGPP doesn't seem to have it. How should I port it?

Q: How do I compute a segment and an offset of a protected-mode address?

A: In DJGPP you use a flat address space with no segmentation (it is a kind of tiny model, since CS = DS = SS, but with a very large segment), so you don't need far pointers in the sense they are used in 16-bit code. Just define away those keywords and you will be fine: 

  #define far
  #define near
  #define huge
  #define _far
  #define _near
  #define _huge

Alternatively, you could add suitable -D switches to the GCC command line, like this: 

  gcc -Dfar= -Dnear= -Dhuge= -c myprog.c

Macros that create far pointers from the segment and offset (usually called MK_FP or _MK_FP) are mostly used in 16-bit code to access certain absolute addresses on memory-mapped peripheral devices, like the video RAM. These chores are done differently in DJGPP. Here's one possible way to express MK_FP in DJGPP (courtesy of Charles Sandmann): 

  #include <sys/nearptr.h>
  #include <crt0.h>

  void * MK_FP (unsigned short seg, unsigned short ofs)
  {
    if ( !(_crt0_startup_flags & _CRT0_FLAG_NEARPTR) )
      if (!__djgpp_nearptr_enable ())
        return (void *)0;
    return (void *) (seg*16 + ofs + __djgpp_conventional_base);
  }

The above uses the DJGPP nearptr facility, which effectively disables memory protection and doesn't work on some systems (e.g. NT); if you prefer to use farptr functions (which are safer and work with all known DPMI hosts), you will need to rewrite the code that uses these macros, so don't bother writing a replacement for the MK_FP macro itself. The details are described in Accessing absolute addresses, below.

Macros that extract the segment and the offset from a far pointer (called FP_SEG and FP_OFF) are required in 16-bit code to pass addresses in registers when calling real-mode DOS or BIOS services, like functions of interrupt 21h. See How to call real-mode interrupt functions, which describes how that should be done in DJGPP; here, too, you won't need to port the macros but instead rewrite the code that calls the DOS or BIOS service. In particular, you cannot compute a real-mode segment and offset of a protected-mode address, because real-mode addresses can only access the first 1MB of memory, whereas the variables of DJGPP programs all live above the 1MB mark.

Node: Pseudo-registers, Previous: NEAR and FAR, Up: Converting

17.8 How to convert _AX pseudo-registers?

Q: Since DJGPP doesn't recognize Borland-style pseudo-register variables like _AX, how should I port code which uses them to DJGPP?

A: These pseudo-variables are typically used in two different contexts:

Node: Low-level, Next: , Previous: Converting, Up: Top

18 Low-level DOS/BIOS and Hardware-oriented Programming

This chapter sheds some light on a few aspects of writing DJGPP programs which interact with hardware or use interrupts.

Node: int86, Next: , Previous: Low-level, Up: Low-level

18.1 Got "Unsupported INT 0xNN" calling int86

Q: Why does my program crash with "Unsupported DOS request 0xNN" or "Unsupported INT 0xNN" when I call int86 or intdos functions to invoke a software interrupt?

A: Calling real-mode DOS or BIOS services from protected-mode programs requires a switch to real mode, so the int86 family of functions in the DJGPP library should reissue the INT instruction after the mode switch. However, some services require pointers to memory buffers. Real-mode DOS/BIOS functions can only access buffers in conventional memory, so int86 has to move data between your program and low memory to transparently support these services. But this means int86 should know about all these services to perform these chores correctly, because each service has its own layout and size of the buffer(s). While int86 supports many of these services, it doesn't support all of them. The supported functions are listed in the library reference, see int86. For those services it doesn't support, you will have to call the __dpmi_int library function instead; it is also documented in the library reference. __dpmi_int requires that you set up all the data as required by the service you are calling, including moving the data to and from low memory (see how to use buffers with DOS/BIOS services).

Note that calling int86 and intdos can sometimes cause trouble due to size (16 bits as opposed to 32 bits) of the members in the union REGS structure. Do not assume that e.g. regs.x.ax is always 16 bit! This problem and the facilities available to specify the width of the registers are all described in the library reference; see int86.

In particular, programs which interface with the mouse via calls to the int86 library function, should mask off the high 16 bits of the registers which report mouse position and other values, since the high 16 bits aren't necessarily zeroed (which will wreak havoc in any program that interfaces to the mouse).

For these reasons, it is generally recommended to use __dpmi_int instead of int86 and intdos.

Node: Pointer segment, Next: , Previous: int86, Up: Low-level

18.2 How to use buffers with DOS/BIOS services

Q: I want to call a DOS/BIOS function which requires a pointer to a buffer in, e.g. ES:DI (or any other) register pair. How do I get the segment to put into the ES register?

Q: I have some real-mode code that calls the segread function. How can I make it work with DJGPP?

A: If you call __dpmi_int, then you must put into that register pair an address of some buffer in conventional memory (in the first 1 MByte). If the size of that buffer doesn't have to be larger than the size of transfer buffer used by DJGPP (at least 2KB, 16KB by default), then the easiest way is to use the transfer buffer. (Library functions don't assume the contents of the transfer buffer to be preserved across function calls, so you can use it freely.) That buffer is used for all DOS/BIOS services supported by DJGPP, it resides in conventional memory, and is allocated by the startup code. DJGPP makes the address and the size of the transfer buffer available for you in the _go32_info_block external variable, which is documented in the library reference. Check the size of the buffer (usually, 16K bytes, but it can be made as small as 2KB), and if it suits you, use its linear address this way: 

dpmi_regs.x.di =
 _go32_info_block.linear_address_of_transfer_buffer & 0x0f;
dpmi_regs.x.es =
 _go32_info_block.linear_address_of_transfer_buffer >> 4;

For your convenience, the header file go32.h defines a macro __tb which is an alias for _go32_info_block.linear_address_of_transfer_buffer.

Here's a simple example of calling a real-mode service. This function queries DOS about the country-specific information, by calling function 38h of the DOS Interrupt 21h, then returns the local currency symbol as a C-style null-terminated string in malloced storage. Note how the transfer buffer is used to retrieve the info: the address of the transfer buffer is passed to DOS, so it stores the data there, and the function then retrieves part of that data using dosmemget. 

 #include <sys/types.h>
 #include <sys/movedata.h>
 #include <dpmi.h>
 #include <go32.h>

 char * local_currency (void)
 {
   __dpmi_regs regs;

   regs.x.ax = 0x3800;        /* AH = 38h, AL = 00h  */
   regs.x.ds = __tb >> 4;     /* transfer buffer address in DS:DX  */
   regs.x.dx = __tb & 0x0f;
   __dpmi_int (0x21, &regs);  /* call DOS  */
   if (regs.x.flags & 1)      /* is carry flag set?  */
     /* The call failed; use the default symbol.  */
     return strdup ("$");
   else
     {
       /* The call succeeded.  The local currency symbol is stored
          as an ASCIIZ string at offset 2 in the transfer buffer.  */
       char *p = (char *)malloc (2);
       if (p != 0)
         dosmemget (__tb + 2, 2, p);
       return p;
     }
  }

If the size of the transfer buffer isn't enough, you will have to allocate your own buffer in conventional memory with a call to the __dpmi_allocate_dos_memory library function. It returns to you the segment of the allocated block (the offset is zero). If you only need a small number of such buffers which can be allocated once, then you don't have to worry about freeing them: they will be freed by DOS when your program calls exit. The only adverse effect of not freeing DOS memory until the program exits is that if you need to run subsidiary programs (via spawnXX or system library functions), those programs will have less conventional memory. Usually, this aspect should only be considered if a program allocates very large (like 100KB) buffers in conventional memory.

DOS memory can also be allocated by calling function 48h of Interrupt 21h via __dpmi_int and freed by calling function 49h. The only disadvantage of this method is that it doesn't create a protected-mode selector for the allocated block, so you must use the _dos_ds selector to reference the allocated memory, which is less safe: the _dos_ds selector spans the entire first megabyte of memory, whereas the selector created by __dpmi_allocate_dos_memory spans only the allocated block, and will therefore catch bugs that reference memory outside that block.

For bullet-proof code, you should test the size of the transfer buffer at runtime and act accordingly. This is because its size can be changed by the STUBEDIT program without your knowledge (however, it can never be less than 2KB, the size of the stub, because memory used by the stub is reused for the transfer buffer).

The function segread used by some real-mode compilers does not exist in DJGPP. It is used in real-mode code to store the values of the CS, DS, SS, and ES registers into a struct SREGS variable, when some service that needs one of these registers is called from code written for small and tiny memory models. DJGPP has the functions _my_cs, _my_ds, and _my_ss for that purpose (ES and DS always hold the same selector in code produced by GCC from a C or C++ source, so you don't need a fourth function). However, these will not be useful if the original real-mode code used the segment registers to invoke DOS/BIOS services. For these cases, you will need to rewrite the code so that it copies the data to/from the transfer buffer and passes the transfer buffer address via __dpmi_int, as described above.

If you use int86x or intdosx to call a DOS or BIOS function supported by them, then just put the address of your buffer into the register which expects the offset (regs.x.di), forget about the segment, and call int86 or intdos instead of int86x and intdosx. The DOS/BIOS functions supported by int86 and intdos are processed specially by the library, which will take care of the rest. Note that calling int86x and intdosx will usually crash your program, since they expect that you pass them a real-mode segment:offset address to a buffer in conventional memory; this is done more easily with __dpmi_int, as described above, so I don't recommend using int86x and intdosx.

Node: Zero SP, Next: , Previous: Pointer segment, Up: Low-level

18.3 How to call real-mode functions

Q: My program crashes/doesn't do what it should when I call __dpmi_simulate_real_mode_procedure_retf.

A: You should zero out some of the members of the __dpmi_regs structure before you call the DPMI function that invoke real-mode procedures. Random values in these members can cause your program to behave erratically. The members in point are .x.ss, .x.sp, and .x.flags. When .x.ss and .x.sp are zeroed, the DPMI host will provide a stack for the call. This stack is usually large enough, but sometimes you'll need to use your own, larger stack, e.g., if you expect interrupts to nest deeply, or if your handler needs a lot of stack space31. In these cases you should point .x.ss and .x.sp to a larger buffer which is in conventional memory (possibly part of the transfer buffer).

If SS:SP isn't zero, it will be used as the address of the stack for the interrupt handler, so if it points to a random location, your program will most certainly crash. A non-zero FLAGS member can also make the processor do all kinds of weird things (e.g., imagine that the single-step or the debug bit is set!).

Node: Xfer, Next: , Previous: Zero SP, Up: Low-level

18.4 How to move data between your program and conventional memory

Q: How can I move data between my program and the transfer buffer?

Q: How do I access my peripheral card which is memory-mapped to an address between 640K and 1M?

Q: How can I read or change a value of one of the variables in the BIOS data area?

Q: How can I peek at an address whose far pointer I get from an INT 21h call?

A: Usually, memory-mapped devices or absolute addresses below 1MB mark are outside your program's address space, so you cannot access them directly. "Direct access", when you just dereference a pointer, means in DJGPP that you use your program's DS selector, and all the addresses are offsets relative to the base of that selector. So first, you will need a special selector that will allow you to access your device or absolute address. There are several methods you can get such a selector:

Once you have a selector, you can use one of three methods to access your absolute addresses using that selector:

Node: Move structs, Next: , Previous: Xfer, Up: Low-level

18.5 How to move structs returned by real-mode services?

Q: My program uses the contents of a structure returned by a VBE function, but some of the struct members are garbled!

A: Most probably, this happens because of incorrect declaration of the structure in your program. Many people copy a declaration from some real-mode program, and that is exactly what gets them into trouble.

Here are some gotchas in this context:

For example, the following real-mode structure declaration: 

 struct ncb {
   unsigned ncb_command;
   int ncb_status;
   char far *ncb_buffer;  /* a far pointer to a buffer */
   char ncb_name[32];
   int far (*ncb_dispatch)();  /* a pointer to a far function */
 };

should be converted to this in a DJGPP program: 

 struct ncb {
   unsigned short ncb_command __attribute__((packed));
   short ncb_status __attribute__((packed));
   unsigned short ncb_buf_offset __attribute__((packed));
   unsigned short ncb_buf_segment __attribute__((packed));
   char ncb_name[32] __attribute__((packed));
   unsigned short ncb_dispatch_offset __attribute__((packed));
   unsigned short ncb_dispatch_segment __attribute__((packed));
 };

With the above declaration of struct ncb, the following real-mode code: 

 int status = ncb.ncb_status;
 int value  = *(int far *)ncb.buf[3];

should read in DJGPP: 

 short status, value;
 struct ncb ncb_struct

 /* Fetch the structure from the transfer buffer.  */
 dosmemget (__tb, sizeof (struct ncb), &ncb_struct);
 status = ncb_struct.ncb_status;
 value  = _farpeekw (_dos_ds,
                     ncb_struct.ncb_buf_segment*16
                     + ncb_buf_offset + 3);

In other words, you need to add code that moves the structure to and from the transfer buffer, and replace each pointer dereference with a call to an appropriate _farpeekX or _farpokeX function.

Node: Fat DS, Next: , Previous: Move structs, Up: Low-level

18.6 Fast access to absolute addresses

Q: The "farptr" functions are too slow for my application which MUST have direct access to a memory-mapped device under DPMI. How can I have this in DJGPP? My entire optimized graphics library is at stake if I can't! :-(

A: The following so-called Fat DS, or "nearptr" method was suggested by Junaid A. Walker (he also posted a program which uses this technique to access the video RAM; you can look it up by searching the mailing list archives). But first, a word of warning: the method I'm about to describe effectively disables memory protection, and so might do all kinds of damage if used by a program with a wild pointer. It is depressingly easy, e.g., to overwrite parts of DOS code or data with "Fat DS" on. Or, as Stephen Turnbull has put it when he read the description of this trick:

Surgeon General's WARNING: The description below uses the "Fat DS hack", a steroid derivative which gives your program great strength, a thick neck, baldness, and is known to be closely linked with the Alzheimer's disease.

In addition to the above warning, experience shows that many programs which use the safer "farptr" functions do not sacrifice performance at all. So, with the exception of a small number of programs, "nearptr" is really a convenience trick: it allows you to treat memory-mapped devices with usual C pointers, rather than with function calls. Therefore, I would generally advise against using "nearptr" due to speed considerations, unless your program absolutely needs the last percent of speed.

Having said that, here is the trick: you change the limit of the segment descriptor stored in DS to 0xffffffff (i.e., -1), using library function __djgpp_nearptr_enable. After that, you have access to all the memory which is currently mapped in. This works due to 32-bit wrap-around in the linear address space to access memory at, say, linear address 0xa0000 (which belongs to the VGA), or any other address on your memory-mapped device, by adding the value of the global variable __djgpp_conventional_base to the target address. __djgpp_conventional_base is the negated base address of the DS selector that you program is using to access its data. By adding the value of __djgpp_conventional_base, you effectively subtract the DS base address, which makes the result zero-based, exactly what you need to access absolute addresses.

You should know up front that this trick won't work with every DPMI host. Linux's DOSEmu and Windows/NT won't allow you to set such a huge limit on the memory segment, because these operating systems take memory protection seriously; in these cases __djgpp_nearptr_enable will return zero--a sign of a failure. CWSDPMI, QDPMI, Windows 3.X and Windows 9X all allow this technique (OS/2 Warp seems to allow it too, at least as of version 8.200), but some events break this scheme even for those DPMI hosts which will allow it. A call to malloc or any other library function which calls sbrk might sometimes change the base address of the DS selector and break this method unless the base address is recomputed after sbrk call. (The "nearptr" functions support this recomputation by providing you with the __djgpp_conventional_base variable, but it is your responsibility to recompute the pointers using it.) The same change can happen when you call system, and as a result of some other events external to the executing code thread, like multitasking or debugger execution.

You should also know that the __djgpp_nearptr_enable function in DJGPP v2.0 didn't verify that the limit was properly set. So if the DPMI server would fail the call silently, the function won't detect it and will not return a failure indication. DJGPP v2.01 corrects this omission by always verifying that the DPMI host has honored the request, and returns a failure indication if it hasn't.

If you are aware of these limitations, and don't need your code to run under all DPMI hosts, it might be the fix to your problems.

Confused about how exactly should you go about using this technique in your program? Look at the docs of the "nearptr" functions in the Info file libc.info (node __djgpp_nearptr_enable).

Another possibility is to use the DPMI function 0x508 (a wrapper function __dpmi_map_device_in_memory_block is available in the DJGPP library) that can map any range of physical memory addresses into a block that you allocate. Note that this is a DPMI 1.0 functionality which is not supported by most DPMI 0.9 hosts (CWSDPMI does support it). There is a convenience helper function __djgpp_map_physical_memory in the DJGPP C library that you can use to call these services.

If you need a nearptr-style access to a certain region of memory which is above the base address of the DS selector, you can enlarge the limit of the DS selector just enough to cover the highest address you need to access. To this end, use the library function __dpmi_set_segment_limit like this (thanks to Eric Rudd for posting this code): 

 unsigned long new_limit;

 if (__dpmi_set_segment_limit (_my_ds (), new_limit) == 0)
   {
     if (__dpmi_get_segment_limit (_my_ds ()) != new_limit)
       /* The DPMI host ignored the call.  Fail.  */
     else
       {
         __dpmi_set_segment_limit (__djgpp_ds_alias, new_limit);
         __dpmi_set_segment_limit (_my_cs (), new_limit);
         _crt0_startup_flags |= _CRT0_FLAG_NEARPTR;
       }
   }
 else
   /* The call failed.  */

Remember that new_limit should have all its lower 12 bits set, otherwise the above snippet will not work!

Node: Above 1MB, Next: , Previous: Fat DS, Up: Low-level

18.7 Accessing absolute address above 1MB

Q: How can I access memory-mapped peripheral devices (or any other absolute address) above 1 MByte mark?

A: You should use DPMI functions to allocate an LDT descriptor, and map it to an absolute physical address. This maps the physical address of the memory on the device to a linear address, and returns that linear address to you. You then create a selector to access the span of linear addresses on the device.

Here are the DPMI calls that you will have to use:

All of these DPMI calls have __dpmi_XXX wrappers in the DJGPP library. Here's a somewhat schematic example: 

 #include <dpmi.h>
 .
 .
 __dpmi_meminfo mi;
 int selector;
 .
 .
 /* Map the physical device address to linear memory.  */
 mi.address = physical_address;
 mi.size    = physical_address_size;
 __dpmi_physical_address_mapping (&mi);
 /* Now mi.address holds the linear address.  */
 .
 .
 /* Allocate an LDT descriptor and set it up to span the entire
    device on-board memory.  */
 selector = __dpmi_allocate_ldt_descriptor (1);
 __dpmi_set_segment_base_address (selector, mi.address);
 __dpmi_set_segment_limit (selector, mi.size - 1);

Note that the segment limit should be one less than the size. Also, segments over 1MB in length must be a multiple of 4KB, otherwise the DPMI server might fail the call, or silently change the limit.

You can then use the functions from the sys/farptr.h header file to access that device. See accessing memory-mapped devices, for more details about accessing memory-mapped devices given their linear address.

The DPMI function that is issued by __dpmi_physical_address_mapping only works reliably for addresses above 1MB mark. If you call it with a physical address in the first Megabyte, it might fail, depending on the DPMI server (e.g., CWSDPMI fails such calls). (The DPMI spec explicitly says that programs should not call this function to access memory below the 1MB boundary.) This failure usually means that the offending address is already mapped into the page tables, so you shouldn't worry about it; and most DPMI servers map the first Megabyte 1:1 anyway.

Node: RMCB, Next: , Previous: Above 1MB, Up: Low-level

18.8 How to make DOS/BIOS call your function

Q: How can I make any real-mode service call my function? E.g., the mouse driver has a provision (function 0Ch) to call a user-defined handler when certain events occur, which expects a far pointer to my function in the ES:DX register pair.

A: Those services expect a real-mode function, so you should wrap your protected-mode function with a real-mode wrapper. To this end, call either the _go32_dpmi_allocate_real_mode_callback_retf or the _go32_dpmi_allocate_real_mode_callback_iret library function, as required by the real-mode service you want to hook, and pass the segment and offset members it returns to the service you want (in the above example, Int 33h function 0Ch) by calling __dpmi_int.

Here's a code fragment that shows how to do this34: 

  #include <dpmi.h>
  #include <go32.h>

  static __dpmi_regs        callback_regs;
  static _go32_dpmi_seginfo callback_info;

  int install_mouse_handler (unsigned mask,
                             void (*func)(__dpmi_regs *))
  {
    __dpmi_regs r;

    callback_info.pm_offset = (long)func;
    if (_go32_dpmi_allocate_real_mode_callback_retf(&callback_info,
                                                    &callback_regs))
      return -1;  /* failure */

    r.x.ax = 0xc;
    r.x.cx = mask;
    r.x.es = callback_info.rm_segment;
    r.x.dx = callback_info.rm_offset;
    __dpmi_int (0x33, &r);
    return (r.x.flags & 1) ? -1 : 0;
  }

The handler (func in the above example) will be called with a pointer to a __dpmi_regs structure which is filled by values found in the CPU registers when the mouse driver calls the handler. See the docs in the library reference Info file for further details about allocating wrapper functions.

Additional considerations apply if your callback is a C++ class member function. First, you need to remember that member functions expect a hidden extra first parameter. Second, if the function is virtual, you will need to lock the class's virtual table. Third, you need to lock the object itself, not only the method you call on it.

Node: Hardware interrupts, Next: , Previous: RMCB, Up: Low-level

18.9 How to hook hardware interrupts

Q: How do I register my DJGPP function as a hardware interrupt handler?

A: The optimal setup depends on the interrupt frequency and on the amount of processing it requires. Therefore, only some basic considerations and techniques are listed below. What combination of these is best for your application is up to you to decide.

First, some background. Hardware interrupts can occur when the processor is either in real mode (like when your program calls some DOS service) or in protected mode. When your program runs under a DPMI host, hardware interrupts are caught by the DPMI host and passed to protected mode first; only if unhandled, they are then reflected to real mode. Therefore, in DPMI mode you can get away with installing only a protected-mode handler. However, if the interrupts happen at a high frequency (say, more than 10 KHz), and if your program spends lots of time calling real-mode DOS/BIOS functions, then the overhead of the interrupt reflection from real to protected mode might be too painful, and you should consider installing a real-mode interrupt handler in addition to the protected-mode one. Such a real-mode handler will be called before the interrupt gets to the DPMI host, and handle the interrupt entirely in real mode, so it must be written in assembly and located in conventional memory (below the 1MB mark). If you need to hook an interrupt with both PM and RM handlers, you must hook the PM interrupt first, then the RM one (because hooking the PM interrupt modifies the RM one). Also, you should know that some DPMI hosts don't allow you to hook the RM interrupt (CWSDPMI does), and some call both handlers, no matter in what mode the interrupt arrived (CWSDPMI will only call one of them); the only way to be sure is to try.

To install a protected-mode interrupt handler, you do this:

To install a real-mode interrupt handler, you do this:

Note that Windows 9X is reported to call both the RM and PM handlers if both are installed, at least for some interrupts (CWSDPMI only invokes one of them). So, if you want to play safe, you will need some kind of a semaphore variable that the two handlers could use so that only one of them actually handles the interrupt in any given case.

The DPMI spec says that 3 software interrupts are special, in that they also get reflected to a protected-mode handler. These interrupts are: 1Ch (the timer tick interrupt), 23h (Keyboard Break interrupt), and 24h (Critical Error interrupt). This means that, to catch these interrupts, you need to install a protected-mode handler only. Unlike hardware interrupts, it doesn't make sense to install dual RM and PM handlers for these software interrupts. In particular, Windows will call both RM and PM handlers if you install both, so you effectively wind up handling the same interrupt twice.

For examples of installing and using hardware interrupt handlers, see the sources of the Allegro library, the sample code written by Bill Currie, the Sound Blaster interrupt-driven functions, the mkkbd package, and the libhw library, described under sample DJGPP packages. Alaric B. Williams has written a tutorial on interrupt handling.

The DJGPP User's Guide includes a chapter on hardware interrupts, written by Peter Marinov, which includes sample code for hooking hardware interrupts.

The file src/libc/go32/dpmiexcp.c in the DJGPP library sources, djlsrNNN.zip, is one example of the subtleties involved with installing a real-mode interrupt handler. The handlers themselves are in the file src/libc/go32/exceptn.S.

Node: _go32 vs __dpmi, Next: , Previous: Hardware interrupts, Up: Low-level

18.10 Should I use _go32_XXX or __dpmi_YYY functions?

Q: In v1.x I was used to the _go32_... functions, but now comes v2 which also has __dpmi_... functions. Are there any differences between these two varieties?

Q: Do I need to convert my old v1.x code to use the new __dpmi_... functions?

A: These two groups of functions have different functionality, so don't just substitute the new ones for the older ones, because it usually won't work! The new __dpmi_... functions are just bare-bones wrappers of the DPMI API calls36 (see DPMI Specification), generally unsuitable for use with handlers written in C, whereas the old _go32_... functions are intelligent helper routines which only make sense if your interrupt handlers are C functions. They save all the registers on the stack (to be restored before return to caller), and set up DS, SS, and ES registers as GCC assumes in the code it produces for a C program. If these assumptions are wrong, the C functions called by an interrupt handler will crash miserably.

The problem with the _go32_... functions is that they don't lock all the code and data that your handlers use, so they can crash on memory-tight machines and thus aren't suitable for production-quality code. But they are certainly useful in the initial stages of writing and debugging code that hooks hardware interrupts, and for migrating existing v1.x code to v2. Some of the old names were just #defined to the new names where the functionality is identical.

The bottom line is that it shouldn't be necessary to convert your code for it to work at least as well as it did in v1.x; but if you want it to be more stable, you should rewrite your handlers in assembly and use the new __dpmi_... functions. See How to install a hardware interrupt handler.

Node: HW Int pitfalls, Next: , Previous: _go32 vs __dpmi, Up: Low-level

18.11 Hardware interrupt hooking has its subtleties

Q: I did everything you tell me to install the interrupt handler correctly, but my program occasionally still hangs....

Q: From time to time my program crashes with a message "Page Fault in RMCB". What's that?

A: Hooking interrupts in DJGPP (and in protected mode in general) has a few subtle aspects. In general, hardware interrupt handling in DJGPP v2.x is rock solid if you play by the rules. Unfortunately, the rules are a bit tricky.

One cause of your problems might be that your interrupt handler or some memory location it uses get paged out because of the virtual memory mechanism, or because your program spawned a child program. In that case, the interrupt might cause a call to a non-existent service routine, with the obvious results. You should lock all the memory pages that your handler accesses by calling the __dpmi_lock_linear_region library function. This also means in practice that you should write your handler in assembly, as described in how to set an interrupt handler, above. You can disable virtual memory, or put _CRT0_FLAG_LOCK_MEMORY into _crt0_startup_flags to make sure nothing is paged out (but then your program might not have enough memory to run, unless you run on memory-abundant systems).

When CWSDPMI detects that your handler accesses memory that is not locked, it aborts your program with a message saying "Page Fault in RMCB". This can happen if your program installs a callback for some real-mode service, like the mouse callback, as well as if you install a hardware interrupt handler; in both of these cases you need to lock all the memory touched by your handler or by functions it calls. CWSDPMI aborts your program if your program attempts to page while an interrupt handler or a real-mode callback are active, because paging uses DOS file I/O. Since DOS is non-reentrant, if the hardware interrupt handler was called in a middle of another DOS call, paging could badly damage your hard disk37. By refusing to page in these cases, CWSDPMI ensures the stability of your system and integrity of your files. You pay for that stability by having to lock all code and data touched by the handler.

Another problem might be that the hardware peripheral you use generates a lot of interrupts. Due to specifics of hardware interrupts handling in protected mode, there is a substantial overhead involved with reflection of interrupts between real and protected modes. For instance, on a 486DX/33 this reflection might consume up to 3000 clocks; on a 386SX/16, even a 1KHz clock might eat up 1/2 of available cycles. One user reported that a 120 MHz Pentium will be able to service up to 45-50K interrupts per second before exhausting its CPU resources, and a 486DX/50 is capable of about half that number. If your hardware fires too many interrupts, your CPU might not be able to keep up. A good rule of thumb is to consider 20KHz as the breaking point, if your program needs to do something non-trivial besides servicing interrupts. If you are beyond that interrupt rate, consider reducing the interrupt frequency, or move some of the processing done inside the interrupt handler to some other place. Use a ring-0 DPMI server such as CWSDPR0 or PMODE/DJ (of these two, the latter is the faster one) which don't swap interrupt stacks--this will reduce the overhead of the interrupt reflection to some degree. If your handler is written in C, write it in assembly and make sure it doesn't chain. And most important--make sure your program keeps the processor completely in protected mode while handling high-frequency interrupts: avoid unnecessary library calls, disk I/O, BIOS calls, and anything else that could generate a mode switch. For example, using BIOS services to wait a certain period of time while interrupts come in is clearly a bad idea when the interrupts come at high frequency.

Installing a good memory manager will usually also remove most of the mode switch overhead, since a memory manager runs the CPU in V86 mode, where hardware interrupts are delivered in protected mode by the processor, without any need for a mode switch.

Preventing the program from paging (by installing enough physical RAM and using memory efficiently) will also help keeping the CPU in protected mode, since paging is done by calling DOS in real mode. By keeping your processor in protected mode as much as you can, you avoid the expensive mode switches when the interrupts are reflected to your PM handler.

If all that still doesn't help, install a real-mode handler.

Some losing memory managers, notably EMM386, were reported to induce a high interrupt handling overhead. In one case, a user reported an increase in the maximum interrupt rate his program could support from 2 KHz to 6 KHz after uninstalling EMM386.

Still another possibility is that you use a non-default sbrk algorithm in your program. Check if the header file crt0.h is included anywhere in the program, and if so, if the _CRT0_FLAG_UNIX_SBRK bit in the _crt0_startup_flags variable is set by the program. If it is, then a hardware interrupt which happens at the wrong time could crash your machine, especially if you run under Windows 3.X.

You should also keep in mind that the DPMI server can decide to handle some of the interrupts itself and not pass them to your program, although this is rare. For example, Windows 9X won't pass the Ctrl-Alt-Del combination to your keyboard interrupt handler, but will rather act on it itself; QDPMI sometimes processes Ctrl-C keypresses so that your program never sees them, etc. Sometimes, but not always, you can change some configuration option to make some keys get to your handler (e.g., the Alt-TAB setting on the Windows3.X .PIF file).

If the above still doesn't explain your problem, then post your code on the DJGPP mailing list or the comp.os.msdos.djgpp news group, tell there how it fails and somebody will usually have a solution or a work-around for you.

Node: Inline Asm, Next: , Previous: HW Int pitfalls, Up: Low-level

18.12 Inline Assembly code with GCC

Q: I am used to writing inline assembly with Borland C, but can't figure out the way to do it with GCC....

Q: How can I reference C variables from my inline assembly code?

A: GCC has extensive inline assembly facilities. They allow you to specify everything other compilers let you (like the registers where GCC will put specific results), but in a way that doesn't interfere with the compiler's optimizations of the C code that includes inline assembly. Because of this flexibility, the syntax of the inline assembly code is very different from the other DOS-based compilers. The GCC on-line docs describe these facilities in detail; to read the relevant sections, type this from the DOS prompt: 

  info gcc "C Extensions" "Extended Asm"

(Note the quotes: they are important.) You will, of course, need the stand-alone Info reader to be installed on your system for the above command to work. If it is not already installed, get the file v2gnu/txi40b.zip from the DJGPP distribution and install it.

If you read this FAQ via WWW, you can also read about the GCC inline assembly extensions with your Web browser. Brennan Underwood has written a tutorial on using inline assembly, which is another valuable resource on this issue.

Node: DMA, Previous: Inline Asm, Up: Low-level

18.13 Using DMA with DJGPP

Q: How do I use DMA with DJGPP programs?

Q: I want to use DMA, but I don't know how to get the physical address of the buffer I allocate for that purpose.

A: The main problem in using DMA with DJGPP is how to get the physical address of a buffer, which is required to program the DMA controller. In protected-mode environment, memory addresses that your program manipulates are actually offsets from the base address of the data segment. You can obtain the base address of the data segment by calling the __dpmi_get_segment_base_address library function and add it to the address of your buffer, but the resulting address is a logical address, translated into a physical address by the memory-mapping unit which is part of the CPU.

You have several alternatives to get the physical address of your buffer:

Node: Legalese, Next: , Previous: Low-level, Up: Top

19 Legal Aspects

This chapter answers some questions about various legal aspects of writing programs with DJGPP.

Node: Application distribution, Next: , Previous: Legalese, Up: Legalese

19.1 Legal (un)restrictions on DJGPP applications

Q: Can you explain in plain English the legal restrictions of distributing programs compiled with DJGPP?

Q: Can I write commercial programs with DJGPP?

A: In most cases, you don't have to worry about any legal restrictions when you compile your programs with DJGPP. You only need to include information on how to get DJGPP, and a few other bits of information, as explained below, in the documentation of your software39.

Using the GNU C/C++ compiler doesn't make your programs subject to any restrictions. The C library which comes with DJGPP is free (unless you change the library sources, see below), which means you are free to use the stock libc.a in any way you like (but please try to comply with basic rules of courtesy.)

Some functions from the DJGPP C library are under the BSD copyright (their sources were taken from the Berkeley Software Distribution of Unix). These are time-related functions time, ctime, gmtime, localtime, mktime, and asctime, and also tzset and tzsetwall. random and related functions srandom, setstate and initstate are also from the BSD distribution.

The BSD copyright used to require that your binary distribution displays an acknowledgment of the BSD origin of these functions somewhere in the docs and in all the ads. However, as of July 1999, the University of California at Berkeley withdrew that requirement, and does not require to include that blurb anymore.

So, if you write C programs and link them with the stock version of the DJGPP library, you only need to tell your recipients how to get the latest versions of DJGPP, and have absolutely nothing else to worry about.

The basic C++ classes and the Standard Template Library (libstdcxx.a) which come with DJGPP allow you to use them binary-wise (i.e., without changing library sources) in your C++ programs without restrictions, unless you compile your programs with a compiler other than Gcc (which won't happen if you work with DJGPP). So C++ programs linked with the -lstdcxx switch are also free from any restrictions.

Only the library of additional GNU C++ classes (libgpp.a) requires that you provide your customers with source or object code of the application, so they could relink the application with future or modified versions of the C++ library. However, this library is deprecated and chances are most C++ programs won't use it. (If you intend to distribute commercial programs linked with the libgpp.a library, you are strongly advised to read the GNU Library General Public License which comes with the library, for rigorous definition of its terms.)

Two GNU packages, Flex and Bison, are also special in that using them to produce your programs doesn't place your programs under GPL or LGPL. In other words, lexers produced by Flex and parsers produced by Bison do not imply GPL/LGPL.

If you do use in your program any of the FSF sources that fall under GPL/LGPL (like some of the GCC's sources, or the GNU getopt or regex packages which come with many GNU programs), then you must comply with the terms of GNU licenses when distributing your programs; in this case your entire application becomes GPL. If that is unacceptable to you, consider using the versions of regex and getopt from the DJGPP C library, which are not as powerful, but are free from any restrictions.

You may ship any of the utilities developed specifically for DJGPP (e.g., the floating-point emulator emu387.dxe or the DPMI host cwsdpmi.exe) and the C library, as distributed by DJ Delorie, with your program with no other requirement besides telling your customers how to get DJGPP for themselves.

If you do change the sources of either the C library or the utilities distributed with the djdev package, they, and the programs developed with them, immediately fall under the GPL, the GNU License. In practice this means that you cannot distribute any binaries made with such a patched version of libc.a without offering the recipient full sources, including your own sources. However, if you find bugs in the library or the utilities and submit your patches to DJ Delorie, DJ allows to freely use and redistribute patched utilities and binaries made with the patched version of libc.a (even if no official DJGPP version was released with your patches yet).

For the precise legal terms of DJGPP distribution, see the file copying.dj via the Web. Latest versions of the djdevNNN.zip package also include that file, so look for it in you DJGPP installation directory.

Note that the above says nothing about the legal aspects of contributed packages, like GRX and others; you will need to read their docs to find out.

Node: DJGPP redistribution, Previous: Application distribution, Up: Legalese

19.2 Legal restrictions of DJGPP utilities and libraries

Q: Can I redistribute djgpp, and if so, how?

Q: I run a business that sells shareware for distribution costs. Can I include djgpp on my CD-ROM?

Q: I want to include djgpp in a product that happens to need a compiler provided with it. Can I do this?

Q: Is DJGPP public domain software?

Q: Is DJGPP shareware?

A: DJGPP is not public domain, neither is it shareware (you don't have to pay a license fee to use DJGPP). Parts of DJGPP (the compiler and some of the development tools) are GNU software, so you must comply with GNU GPL if you distribute those parts (usually, you won't need to distribute them, because they are freely available to everyone). A small part of the C library is taken from the Berkeley BSD sources, and is therefore in public domain.

Other parts of DJGPP, which include most of the C library, the free DPMI host CWSDPMI, and some of the utilities, are copyrighted, but in a way that allows you to use them freely and without restrictions. The copyright that covers these parts of DJGPP is GPL, the GNU License, but with a special exception: if you distribute the utilities unmodified, or build programs with the unmodified library, the GPL does not apply.

Compliance with GPL is therefore all you are legally required to consider when you redistribute DJGPP itself (as opposed to your programs compiled with DJGPP). However, based on many years of experience of DJGPP distribution, DJ Delorie requests vendors which distribute DJGPP to follow some additional rules. These rules are generally meant to provide a better service to the DJGPP user community:

In addition, it would be a courtesy to inform DJ that you are including DJGPP in your product, in case this information is obsolete. A token sample of your distribution would be nice also.

Note that the above are not legal restrictions (the latter are described in the file copying.dj mentioned in the previous section), they are recommended guidelines for redistributing DJGPP. These guidelines are based on many years of experience and are generally meant to make it easier for your clients to use DJGPP and get support from its developers. Vendors who do not follow these guidelines could risk public humiliation, verbal abuse, and boycott by the DJGPP community, but not legal action.

Note also that if you make source-level changes to DJGPP library or utilities, the changed software falls under the GNU License, GPL, unless these changes are made to fix bugs, and provided that you also submit all such bug-fixes to DJ Delorie for inclusion in a future DJGPP release.

Node: Help, Next: , Previous: Legalese, Up: Top

20 Getting Help

This chapter tells you how to get answers to questions you didn't find in this FAQ, by asking them on DJGPP-related forums.

Node: DJGPP is not GNU, Next: , Previous: Help, Up: Help

20.1 Don't post DJGPP-specific problems to GNU News groups

Q: I post my problem to the gnu.gcc.help news group, but don't get any answers....

A: Is your problem likely to be special to the DJGPP port or to the DOS environment? If so, don't post to GNU Usenet groups, but to the comp.os.msdos.djgpp news group or to the DJGPP mailing list. People who read GNU News groups usually neither know nor care about DOS-specific problems. Post there only if the problem seems to be generic to one of the GNU utilities on any platform. For most problems, this can be deduced only after either tracing a problem in the source code or testing it on some non-DOS platform. As a general rule, always post to the DJGPP forums first.

Node: How to post, Next: , Previous: DJGPP is not GNU, Up: Help

20.2 How to post to the DJGPP forum

Q: How do I post to the DJGPP mailing list?

A: Send mail to the list address as if it were a person. Please use the mailing list only if you cannot access the DJGPP news group, because reflecting the mail to and from the mailing lists incurs additional load on the DJGPP server. The DJGPP news group, <news:comp.os.msdos.djgpp>, is two-way gated to the mailing list. This means messages posted to either the mailing list or the news group will appear on both (once, let's hope ;-); you can read either one and post to either one, and everybody eventually sees everything. So please don't post to both the news group and the mailing list.

The gateway works on DJ's server, and has a very strict anti-spam filter which prevents spam from getting into the news group; it also has an additional keyword-based anti-crap filter that doesn't pass spam to the mailing list. The entire traffic ends up in the mail archives on the DJ's Web server within 24 hours, and is available for searching.

If you have a Usenet feed, now is the time to consider unsubscribing from the mailing list and switch to reading the news group instead, so that the load on the list server will get lower.

Node: Subscribing, Next: , Previous: How to post, Up: Help

20.3 How to become a subscriber to the mailing list

Q: How do I subscribe to the DJGPP mailing list?

A: Send mail to the list server (NOT to djgpp@!!), leave the subject line empty and in the body write:

subscribe <your e-mail address> djgpp

If you only want to receive announcements of new versions and ported software, but don't want to see any other DJGPP mail traffic, subscribe to the djgpp-announce by sending message to the list server which says so:

subscribe djgpp-announce

(Note: no email address when subscribing to the djgpp-announce list!)

The announcements which go to djgpp-announce get reflected to djgpp, so you don't need to subscribe to both these lists.

The DJGPP mailing list is available in the daily and weekly digest forms. To subscribe to one of these, send this one-line message to the above list server:

subscribe <your e-mail address> djgpp-digest-daily

or

subscribe <your e-mail address> djgpp-digest-weekly

Some mailers reject messages with too large size, so you might have trouble with the weekly digest. If you subscribe to it and don't get the digest, try the daily one instead, or switch to another mail software.

You can also subscribe to DJGPP-related mailing lists through DJ Delorie's WWW server.

Note that you don't have to subscribe to the djgpp mailing list if you don't want to get all the traffic in your mailbox (typically, about 30 messages per day). You can ask questions on the list even if you are not a subscriber, because people usually answer both to your e-mail address and to the list (well, actually, the mailer program does it automatically and most people don't bother to change that). If you want to be sure the mail gets to you directly, say in your message that you don't subscribe to the list, and ask people to answer directly. Be sure to provide a valid return address (remove any anti-spam, if you use one) when you ask for direct replies.

Node: Unsubscribing, Next: , Previous: Subscribing, Up: Help

20.4 How to unsubscribe from the mailing list

Q: Whew! There's too much traffic on the djgpp mailing list (at least the SysAdmin glaring over my shoulder thinks so... ;-). How do I unsubscribe myself?

Q: I've been trying for days to unsubscribe from the djgpp mailing list. What am I doing wrong?

A: You should send your unsubscribe messages to the list server (not djgpp@delorie.com!), with the contents being just this:

unsubscribe <your e-mail address> djgpp

When you unsubscribe, that stops new messages from being sent to you. Messages that are already in the mail queues of various mail programs between the DJGPP list server and the machine where you receive your mail--cannot be stopped. Therefore, allow some time before you decide that your unsubscribe message didn't work. In extreme cases, when one of the machines that are forwarding mail to you is down, you can get the messages up to 5 days after you've unsubscribed.

If you think you have waited enough and the messages still keep coming, write to listserv administrator and ask him to help you.

You can also unsubscribe yourself from any DJGPP-related mailing list through DJ Delorie's WWW server.

Recently, DJ has added a mail archive browser to his Web site. With this tool, you can list and read the messages by year, month and day, as well as search the last few days for something you might have missed. This service is available via World-Wide Web.

Node: Languages, Previous: Unsubscribing, Up: Help

20.5 Is it okay to post messages in languages other than English?

Q: Would you please switch to English in your messages instead of using some language nobody understands??

A: It is rude to require that people speak your language when you don't speak theirs. Therefore, any language is allowed on the DJGPP forum.

It is true that, since most people who read the DJGPP news group do speak English, posting in English will bring more answers and thus more efficient help. So, for the most efficient help, it is best to post in English. But this is not a requirement. If a DJGPP user has a question, but cannot express it in English, it is better to post it in some other language than not to be able to post it at all.

Some people can read English (and so will understand replies posted in English), but have difficulty writing in English. Therefore, if you see a message in a language other than English that you happen to know, please consider posting its translation, so that others could reply to it.

If you see a message in a language you don't understand well, and want to reply to it, try the Babelfish on-line translator. I'm told that it is not very good and sometimes returns an utter nonsense, but if its translations are treated creatively, it might help you understand the question (or the answer, as the case may be).

Node: New versions, Next: , Previous: Help, Up: Top

21 Version 2 vs v1.x

This chapter is for those who want to know where are the latest updates for DJGPP software, and what's new and improved in them. It also explains the differences between DJGPP v1.x and v2, for those who are still using DJGPP v1.x and want to know more about v2 while they consider switching.

Node: New and improved, Next: , Previous: New versions, Up: New versions

21.1 New features in DJGPP v2

Q: What exciting new features will I find in v2 as opposed to v1.x?

A: DJGPP v2.x is a DPMI-only environment, and it includes a free DPMI host for those who don't have another DPMI provider installed. In addition, v2 features the following major improvements upon v1.1x:

If you want to help in further v2 development, check out the list of features which have yet to be done and volunteer to implement some of them.

Node: Environment, Next: , Previous: New and improved, Up: New versions

21.2 DJGPP environment in v2.x

Q: There's been this talk about v2 and about go32 going away in that version, but I'm confused on what the new setup will be. Could you clarify the details of this change?

A: In v1.x of DJGPP, the go32 extender was responsible for the following:

In v2.x, a minority of these functions are done by a DPMI host, which is a memory-resident software required to run protected-mode programs under MS-DOS. There are a few commercial DPMI hosts (like Quarterdeck's QDPMI, Qualitas 386Max, MS-Windows 3.X and Windows 9X, OS/2, even Linux), but DJGPP v2 comes with a free DPMI host called CWSDPMI for those who don't have one already. Loading the application into protected-mode memory (a function done in v1.x by go32) is handled by a 2KB-long real-mode stub which runs at start-up, before the application's main functions is called (the stub will also load CWSDPMI if no other DPMI host is detected). All the other custom code required to process BIOS- and DOS-related calls from protected-mode is now built into the library functions which your program calls, so there is no need for a special extender, because the application just issues DPMI calls serviced by the DPMI host.

CWSDPMI can be loaded as a TSR, even loaded HIGH into the HMA/UMB, which will make applications load much faster.

Node: Slow development, Next: , Previous: Environment, Up: New versions

21.3 Why are new DJGPP versions released so slowly?

Q: It's more than a year since the last DJGPP version was released. Why won't you guys upload a new version in all this time?

Q: I've been suffering from this bug for months on end! You know there's a bug, 'cause I told you so ages ago. So why in the world didn't you fix that in a new version??

A: DJGPP is developed by volunteers on their free time. This sets both the schedule of new DJGPP releases and the DJGPP development agenda: they are determined by what those volunteers think is important and doable, and by the amount of free time they can devote to DJGPP-related work. Since the work of the development team is given away for free, you cannot demand that they do something they've decided not to do just yet. The only way to influence DJGPP development is to make your own contribution, by fixing a bug or adding a feature, and then submit your patches to DJ Delorie.

You don't need to look for a large project to make your contribution. The best way to start being involved with DJGPP development is to fix any small and minor problems you see, right when and where you see them. Even bugs and inaccuracies in the DJGPP documentation, like the libc.info Info file, are a good place to begin with. DJ Delorie says that if everybody corrects every small bug they see, we would run out of bugs very fast.

When you submit a bug report or code that implements a new feature that you'd like to add to DJGPP, be prepared to withstand some scrutiny and peer review from the other participants of the DJGPP development team. You might hear various comments, from critique of your code and design decisions to questions why your changes are at all needed, and even requests to submit the changes in certain unified format (see instructions for submitting changes). Please be ready for that review process and don't take it as a rebuttal.

Node: Latest DJGPP, Previous: Slow development, Up: New versions

21.4 Where to find the best C library for DJGPP

Q: There's been quite some time since the last DJGPP release. Where can I get all the latest stuff where the bugs are fixed?

Q: I have heard rumors that there's a better C library for DJGPP available from the net. Where is it?

A: New versions of most parts of DJGPP are released quite regularly. For example, DJGPP ports of most GNU packages are generally released short time after a new version of every package becomes available from the GNU FTP sites. So for many DJGPP packages, a new release should usually be available real soon, just stick around a bit.

The single most important component of DJGPP tool-chain that might suffer from long release schedule is the djdev package. This includes the C library, libc.a, which was written specifically for DJGPP, and some DJGPP-specific development tools, like redir and symify. However, the latter are usually stable and don't need too much fixing.

A full release of djdev is a lot of work, so DJ Delorie decided not to make interim releases (experience from v1.x development shows that such interim releases also generate confusion and are hard to maintain).

In general, you are advised to constantly improve your C library by fixing any bugs in the library sources and replacing old modules with fixed ones. All this takes is to edit the relevant source file, compile it, and put it into the library. For example, assuming you have made a source-level change in a file called foo.c, here's how you update your library: 

  gcc -c -O2 foo.c
  ar rvs c:/djgpp/lib/libc.a foo.o

(This example assumes that DJGPP is installed in the C:\DJGPP directory; if not, you will need to change the pathname of libc.a accordingly.)

Patching the library like that requires that you download the DJGPP library sources, djlsrNNN.zip (where NNN is the version number). That file includes sources to all the DJGPP functions and utilities, and you can extract them as the need to edit them arises.

Bug reports regarding the library and patches to fix them are posted from time to time to the comp.os.msdos.djgpp news group. In addition, the DJGPP bug-tracking system, stores many known bugs and the patches required to solve them. You can use these resources to find solutions to known bugs. Patches are applied using the patch utility which is available from the DJGPP sites.

Users who need the cutting edge of the DJGPP development sources can access the latest development versions of sources of the DJGPP library and utilities via the net. All the bugfixes and new features that are accepted for inclusion in DJGPP are checked into the development source tree using CVS, a free network-based software configuration management package. DJ Delorie has set up anonymous read-only access to the DJGPP CVS tree, whereby anybody who has a CVS client installed can check out the sources via the Internet. You can check out individual source files, specific subdirectories, or the entire DJGPP tree; read the instructions for using CVS for the details. Binaries of CVS clients for Windows platforms are available from the CVS download page. If you want to install the CVS client on Unix or Linux, download the CVS sources and build it on your machine.

Note that development sources are not always as stable as the official release; some of the changes might not be tested by anyone except the person who submitted those changes. I recommend to check out only those sources which fix problems that you cannot work around.

Node: Miscellany, Next: , Previous: New versions, Up: Top

22 Miscellany

This chapter is a hodgepodge of questions which don't belong to any of the other chapters.

Node: RHIDE, Next: , Previous: Miscellany, Up: Miscellany

22.1 Problems with using RHIDE.

Q: Why does RHIDE put all my source files in the C:\Windows\Desktop directory?

Q: RHIDE doesn't remember the path names of the files I edited in my previous session....

Q: Why does RHIDE show the "Compile" option as disabled, although there's a source file loaded?

Q: Sometimes, especially during debugging, RHIDE seems to screw up the display, or crash, or blank the screen and hang. How can I avoid this?

A: To prevent problems with source files not being found by RHIDE or being put into strange directories, use a simple two-step recipe:

Alternatively, you could double-click on the project file for your project; Windows will then invoke RHIDE in the directory of the project file. You might need to associate the .gpr extensions with RHIDE, before you can use this feature.

By default, your program's .exe executable file goes to the same directory where you keep the project file, but the Project dialog lets you specify a different directory, in case you need that.

Display-related problems with RHIDE are usually caused by a faulty video driver (on Windows 9X) or bugs in the SVGA firmware. RHIDE uses advanced VESA functions to save and restore the screen contents and mode-specific settings, and some SVGAs and video drivers don't implement these functions very well. One particularly problematic SVGA card (which will remain unnamed) has bugs even in the standard VGA modes. Downloading the latest video drivers from the vendor's site and upgrading the video BIOS usually helps; if not, the only solution is to replace the video adapter.

Node: g++.exe, Next: , Previous: RHIDE, Up: Miscellany

22.2 Unzipping complains about duplicate/invalid files.

Q: When I unzip the C++ compiler distribution in gppNNb.zip, the unzip program complains about something called g++.exe. What should I do?

Q: I installed DJGPP, and found two different versions of gxx.exe and two versions of cxxfilt.exe. Which one shall I keep?

A: If you install DJGPP on anything but Windows 9X, just ignore that error message about g++.exe and use gpp.exe or gxx.exe to compile C++ programs. g++.exe is an invalid file name on DOS, but is allowed on Windows 9X. It is included for compatibility with Unix, where the C++ compiler is called g++. All g++.exe does is just to run gxx or gpp. So you don't lose much by not having it.

If you are installing DJGPP on Windows 9X, find an unzip program which supports long file names and unzip the files again. Make sure that the DJGPP long file names support (a.k.a. LFN) is enabled, otherwise DJGPP programs such as Make won't be able to invoke g++. To enable LFN support, set LFN=y in the environment.

Duplicate versions of some programs come from different distributions that include the same programs. For example, the cxxfilt is part of two GNU distributions: GCC and Binutils. You should generally kep the latest version, judging by the time stamp of the executable file.

Node: Changing, Next: , Previous: g++.exe, Up: Miscellany

22.3 How to change a DJGPP package?

Q: I want to change cc1. How do I do this?

Q: How do I fix a bug/add a feature to one of the DJGPP programs?

Q: How should I produce patches for DJGPP programs I want to submit, and to whom should I submit them?

A: First, get the sources. These are called *s.zip in the DJGPP distribution. The C Library sources are in djlsr203.zip. Some sources are too big, and might be split into multiple zips, all of which must be unzipped to get a complete source distribution, like this: 

 em1934s1.zip
 em1934s2.zip
 em1934s3.zip

All sources are shipped in ready-to-build form. Any diffs that come with the source distribution, like the files called DIFFS, have already been applied, and any configuration scripts and/or batch files have been run already; you don't need to run them again.

Next, try to build the program without changing it. Look for a file called README.dos or README.djgpp: it should explain the build procedure and list any optional packages you need to install for that.

If such a README file is unavailable, you will have to poke around and figure things out for yourself; here are some hints to help you out:

The reason for an apparent lack of a standard here is that different packages were ported to DJGPP by different people, as best as they saw fit.

After you've successfully built the program, make your fixes and build the program the same way you did before.

Note that generally to build these programs, you must have the GNU Make program, installed, and some makefiles require that you install additional development utilities, like the SED editor. Sometimes the makefiles won't even run under COMMAND.COM (they require a smarter shell). In that case, either get a better shell, or convert the makefile to be runnable by COMMAND.COM, or do the required steps manually. If the Makefile is too complex for you and you can't figure out what are the necessary commands, invoke make with -n switch and see what it would have done.

If your machine lacks floating-point hardware (like a 386 without a 387, or a 486SX), then you should know that current versions of GNU Sed and GNU Make issue floating point instructions, so you will have to make provisions for loading an emulator, see above, FP Emulation. The port of Make 3.75 and later can be built so that it doesn't issue FP instructions, but you will have to get the sources and recompile Make first, as the stock version wasn't configured in that way.

If you think that you found a bug in one of the programs or libraries written for DJGPP (e.g. the C library, CWSDPMI, symify, etc.) be sure to check the list of known bugs. If your bug is not there, you can later submit it to the bug-tracking system.

Before you submit a bug report, please make every effort to verify that your bug is not caused by incorrect usage, or by problems in your DJGPP installation. Reports such as "All DJGPP programs crash" or "I cannot compile any program" are clearly not bugs, because these things work for many hundreds of DJGPP users every day; so either your system setup is messed up or you invoke programs incorrectly.

If you can investigate the cause of the bug and find a solution that makes it go away, submit a bug report with all the details. If you cannot find the cause(s), I suggest posting your problem description to the news group and asking people to verify that it is indeed a bug, before you submit a bug report. The bug-tracking system includes a list of all known bugs, many of them with solutions or work-arounds; please check them before creating a new bug report.

Patches to DJGPP programs and ports should be sent to the person who maintains the relevant package. Patches for the C library, utilities and other software which comes with the djdevNNN.zip distribution should be sent to DJ Delorie. If you don't know who maintains a particular package or port, post the patches to the comp.os.msdos.djgpp news group, since the maintainer is most probably reading that group.

To generate a patch, run the diff program (from GNU Diffutils, v2gnu/dif272b.zip) on the old and the new version of a source file. For example: 

 diff -c src/libc/dos/dos/int86.old src/libc/dos/dos/int86.c >int86.dif

The file int86.dif created this way should be sent to the maintainer, with a short description of the problem it solves. It is a good idea to run the patch file through DTOU (a utility which comes with DJGPP and converts DOS-style CR-LF pairs into Unix-style newlines), since this makes the patch work on Unix as well, in case the maintainer of the package in question does that on Unix. (The DJGPP port of GNU patch accept both Unix-style and DOS-style patch files.)

Observing the following guidelines when creating the patch will make your patches easy to apply:

In addition, it would be mighty nice if every change in functionality were accompanied by a suitable change in the relevant docs (e.g., for a patch to a library function, edit the corresponding .txh file with its docs), although you are under no obligation to do that.

Node: Packages, Next: , Previous: Changing, Up: Miscellany

22.4 Where to find DJGPP packages?

Q: Where can I find an example of XXXX / a package doing YYYY ?

A: Salvador Eduardo Tropea (SET) maintains a DJGPP Web Ring page. DJ Delorie offers another very large collection of DJGPP-related links.

Here is a list of places you might look into for examples of frequently needed code fragments, or for packages people keep asking about:

Node: Symlinks, Next: , Previous: Packages, Up: Miscellany

22.5 How to create symbolic links to programs

Q: How do I create symbolic links?

Q: I have this program that behaves differently depending on the name it's called. Under Unix, I just create symbolic links to achieve that, but DOS doesn't support links. Do I have to put several identical programs under different names on my disk??

A: DJGPP allows you to simulate symbolic links to programs. Generate a stub (which is a small DOS program attached to every DJGPP program by the stubify.exe program), call it by the name of the link you want, then edit its header to run another program. For example, let's say the real program is dj1.exe and we want to make a link called dj2.exe that really calls dj1.exe. First, generate a stub under the name dj2.exe. Next, run STUBEDIT to modify the new program's stub info block and change the name of the executable it runs. In this case, we'd change it to dj1: 

 C:\USR\BIN> stubify -g dj2.exe
 C:\USR\BIN> stubedit dj2.exe runfile=dj1

Voila! Now, when you run dj2, it tells the stub to load the image of dj1, but pass "dj2" in argv[0].

If you use the DJGPP port of GNU Fileutils 3.13 or later, the ln program there can do the above steps for you if you say this (like on Unix): 

 ln -s dj1.exe dj2.exe

Node: DPMI Spec, Next: , Previous: Symlinks, Up: Miscellany

22.6 Where to find the DPMI specification?

Q: What is this DPMI thing you are all talking about?

Q: Where can I find the specifications for the DPMI functions?

A: DPMI, the DOS Protected-Mode Interface, is an API that allows protected-mode programs to run on top of DOS, which is a real-mode operating system, and still be able call real-mode DOS and BIOS services. A special API is required because DOS code cannot be run in protected mode: if you try, your system will immediately crash, since the real-mode DOS and BIOS code violates many restrictions of protected-mode programming. See restrictions of protected mode, for more details. Another place to look for explanations why DPMI is necessary is in the Overview of DJGPP.

The DPMI API is implemented as a set of functions of Interrupt 31h which allow such chores as switch from real to protected mode and back (generally done upon startup and at exit), memory allocation, calling real-mode services, etc. DPMI is by far the most portable way of running protected-mode programs on MS-DOS, MS-Windows and compatible systems.

You can find the DPMI 0.9 spec by anonymous ftp to one of the following sites:

Node: WWW, Next: , Previous: DPMI Spec, Up: Miscellany

22.7 The DJGPP Web site.

Q: Where is the DJGPP Web site?

A: Yes, DJGPP has its own home on the Internet, set up and maintained by (who else?) DJ Delorie. It has an HTML version of this FAQ list with search capabilities, the entire set of DJGPP distribution files, a searchable archive of the DJGPP mailing list and news group traffic, plus other useful and interesting information about DJGPP. For instance, did you ever wonder how DJGPP got started and what DJ's original goals were? Rejoice: the Web site includes the story of DJGPP genesis.

To visit, point your browser to the DJGPP Web site.

Node: Upload, Next: , Previous: WWW, Up: Miscellany

22.8 Where to upload your contributions to DJGPP

Q: I wrote a program using DJGPP. How can I make it available to others?

Q: I found and corrected a bug in one of the programs distributed with DJGPP. Where should I put it?

Q: What should I do to upload my DJGPP package to SImTel.NET?

A: If your program/patches are small enough, consider posting it to the mailing list or the the comp.os.msdos.djgpp news group. Please do not post binaries to the news group; only post source code as plain text, and only if it is not too large. Many people who read the news group cannot save the MIME attachments, and don't like to be forced to download a large message; you are requested to respect that.

If the program is larger than, say, 50K bytes, it's best to upload it to a public site where everybody can get it. You can upload your contribution to a special directory on the DJ Delorie's FTP server. This directory is write-only, and it gets purged every couple of days, so be sure to write to DJ Delorie about your upload; he will then move it to the /pub/djgpp/contrib directory.

If you decide to upload, please send mail to the djgpp-announce list with a brief description of your program/patch. (The message will get reflected to both the news group and the DJGPP mailing list, so you don't have to cross-post there, but it also goes to people who only subscribe to djgpp-announce list because they want to get such announcements and nothing else.)

If your program is more than a patch or a beta version, you might consider uploading it to the DJGPP archives on SimTel.NET. Material uploaded there gets automatically distributed to all of the SimTel.NET mirrors throughout the world, which makes it easier to get.

DJ Delorie requests that all contributed packages uploaded to his server be source-only distributions, if at all possible. This is so there will be no danger of distributing programs infected by a virus. Please avoid uploading self-extracting archives because DJ extracts them on a Unix machine which can't run DOS executables. Detailed instructions for packaging and uploading DJGPP-related files to SimTel.NET are available in the DJGPP HOWTO repository. When the package appears on SimTel.NET mirrors, send an announcement to the djgpp-announce mailing list.

Node: Cross-DJGPP, Next: , Previous: Upload, Up: Miscellany

22.9 DJGPP as cross-compiler

Q: I want to use DJGPP as a cross-compiler for Motorola 68K targets. How should I proceed about this?

Q: I want to build GCC as a Unix-to-DOS cross-compiler. What should I do?

A: If you want a cross-compiler for m68k on a DOS machine, you need DJGPP configured as host=i386-go32, and target=m68k-coff. Such a package is already available. The binaries there are based on GCC 2.7.2. This package is reportedly no longer supported, but if you have questions about it, you can send them to Jim Karpinski. You can also try to contact Kai Ruottu, who is the provider of DOS-hosted gcc-m68k. Note that this package has only basic support for C++: the compiler is included, but the libraries, including libstc++ and libg++, and the C++ headers are missing, so even cin and cout don't work.

A Win32-hosted gcc-m68k is another possibility. It was created by David Fiddes. The Cygwin port of GCC can also be configured as a cross-compiler with m68k as the target. See the description of the Cygwin project, for more details about the Cygwin port. Object Software Inc. is a company that supports cross-builds based on GCC and DJGPP. Pre-built binaries of the compiler and some minimal development tools are available from this site for m68k, PowerPC, and Hitachi's SH-3 targets, all based on DJGPP v2. The CrossGCC FAQ, available from the same site, includes detailed instructions for building a cross-compiler for any target.

DJGPP can be built and installed as a cross-compiler running on a Unix machine and targeting DOS/Windows platforms. Detailed instructions for doing this on Linux can be found in the DJGPP HOWTO Repository. Here is the summary of the necessary steps to do that:

  1. Download the cross-compiler toolkit v2/djcrx203.zip from the usual DJGPP sites.
  2. Unpack djcrx203.zip on the Unix machine using unzip -a.

    You must use the -a switch, to force unzip to convert any DOS-style text file with CR-LF pairs at the end of each line to Unix-style text files. If you don't, things will break for you.

    unzip is available in source form from many FTP sites. Linux systems generally come with unzip, but if your system doesn't, download the sources and build it.

  3. Link or move the files in the cross directory to the top-level directory, where you unzipped djcrxNNN.zip.
  4. Download the latest GCC and Binutils distributions from your favorite GNU FTP site.

    The main GNU site is at ftp.gnu.org.

  5. Unpack GCC and Binutils from the same directory where you unzipped djcrxNNN.zip.
  6. You might need to edit the Makefile, either to choose an installation directory as appropriate for your machine, or to change the names of the directories where the GCC and Binutils distributions are unpacked (e.g., because their version numbers are different from those for which the Makefile was prepared).
  7. Follow instructions in the file cross/install that was unpacked from djcrx203.zip to build and install the cross compiler.

The cross-compiler you build is installed as dos-gcc, so to compile programs with it, use dos-gcc rather than cc or gcc, or set CC=dos-gcc when invoking Make.

The file cross/readme has some usage info for dos-gcc. It is generally correct, except that the version numbers for the various packages might not be up to date. You should always use the latest releases of every package.

Another alternative is the build.cross script included in the source distribution of the DJGPP port of GCC, gccNNNs.zip. This script was tested on GNU/Linux systems, and it allows to build GCC as a cross compiler hosted on a GNU/Linux system whose target is DJGPP-supported platforms. To use this script "out of the box", you will need to take care of some preliminaries:

Here, prefix is the root of your cross-development installation, usually /usr or /usr/local.

Node: 0xfe+0x20, Next: , Previous: Cross-DJGPP, Up: Miscellany

22.10 GCC says "garbage at end of number"

Q: There is a severe bug in GCC: it says "garbage at end of number" for this line: 

 i = 0xfe+0x20;

Ain't it silly that such a great compiler would fail so miserably?

A: That's not a bug, that's a feature of the ANSI C language definition. By ANSI rules, the above expression is a single preprocessing token, unless you place whitespace in front of the plus sign. The reason for this seemingly counterintuitive feature is the syntax of floating-point constants in which letters `e' and `E' followed immediately by a sign signal a decimal exponent. You can use the -traditional compiler switch to turn this feature off (however, it will also turn off a plethora of other ANSI features; see the GCC docs for details).

Judging by the published draft, this is unchanged even in the forthcoming C9X standard.

Node: Struct size, Next: , Previous: 0xfe+0x20, Up: Miscellany

22.11 What should sizeof (struct xyzzy) return?

Q: When I call sizeof on a struct, I sometimes get values which are larger than the sum of the sizes of the struct members, whereas in Borland C++ I always get the correct result. Is it a bug in GCC?

Q: I have a program that reads struct contents from a binary file. It works OK when compiled with BC, but reads garbage when compiled with DJGPP. This must be a bug in DJGPP, right?

A: No, it's not a compiler bug. GCC generates 32-bit code, and in that mode, there is a significant penalty (in terms of run-time performance) for unaligned accesses, like accessing a 16-bit short which isn't aligned on a 16-bit word boundary, or accessing a 32-bit int which isn't aligned on a 32-bit dword boundary. To produce faster code, GCC pads struct members so that each one can be accessed without delays; this sometimes produces struct size which is larger than the sum of the sizes of its members. If you need to minimize this padding (e.g., if your program uses large arrays of such structs, where padding will waste a lot of memory), lay out your structures so that the longer members are before the shorter ones. For example, let's say that you have a struct defined thus: 

  struct my_struct {
    char name[7];
    unsigned long offset;
    double quality;
  };

To make such a struct use the least number of bytes, rearrange the members, like this40: 

  struct my_struct {
    double quality;
    unsigned long offset;
    char name[7];
  };

If the layout of the structure cannot be changed (e.g., when it must match some external specification, like a block of data returned by a system call), you can use the __attribute__((packed)) extension of GCC (see GNU C/C++ Manual.) to prevent GCC from padding the structure members; this will make accesses to some of the members significantly slower.

Beginning with version 2.7.0, GCC has a command-line option -fpack-struct which causes GCC to pack all members of all structs together without any holes, just as if you used __attribute__((packed)) on every struct declaration in the source file you compile with that switch. If you use this switch, be sure that source files which you compile with it don't use any of the structures defined by library functions, or you will get some members garbled (because the library functions weren't compiled with that switch). Also, GCC 2.95.1 and 2.95.2 had bugs in their support of -fpack-struct (the bug is corrected in v2.96 and later).

Alternatively, you could declare a particular structure to be packed, like so: 

  struct my_struct {
    char name[7];
    unsigned long offset;
    double quality;
  } __attribute__ ((packed));

However, note that the latter will only work when you compile it as a C source; C++ doesn't allow such syntax, and you will have to fall back to declaring each struct member with the packed attribute. Therefore, it's best to only use declarations such as above if you are certain it won't be ever compiled as a C++ source.

The padding of struct members should be considered when you read or write struct contents from or to a disk file. In general, this should only be done if the file is read and written by the same program, because the exact layout of the struct members depends on some subtle aspects of code generation and the compiler switches used, and these may differ between programs, even if they were compiled by the same compiler on the same system. If you do need this method, be aware of the struct member padding and don't assume that the number of the file bytes that the structure uses is equal to the sum of the members' sizes, even if you instructed the compiler to pack structs: GCC still can add some padding after the last member. So always use sizeof struct foo to read and write a structure.

Another problem with porting programs that read structs from binary files is that the size of some data types might be different under different compilers. Specifically, an int is 16-bit wide in most DOS-based compilers, but in DJGPP it's 32-bit wide.

You should never read whole structures if they were written by other programs. Instead, read the struct members one by one, and make sure the member declarations are consistent with their definitions in the program that wrote the struct. For example, if a struct member was declared int in a 16-bit program, you need to declare it short in a DJGPP program.

The best, most robust and portable way to read and write structs is through a char buffer, which your code then uses to move the contents into or out of the struct members, one by one. This way, you always know what you are doing and your program will not break down if the padding rules change one day, or if you port it to another OS/compiler. The ANSI-standard offsetof macro comes in handy in many such cases. If you need to change the byte order in struct members that occupy more than a single byte, use special library functions such as ntohl and htons.

Node: Struct packing, Next: , Previous: Struct size, Up: Miscellany

22.12 C++ doesn't pack structs!

Q: When I use struct ffblk from the header dir.h in a C++ program, I get garbage in some members of the structure!

A: There is a known bug in GCC 2.7.2: the C++ compiler effectively ignores the __attribute__((packed)) directives, so the structures end up being not packed. GCC versions 2.7.2.1 and later corrected that bug, so upgrade. As a work-around, surround the declaration of the structure that needs to be packed with #pragma pack, like this: 

  #ifdef __cplusplus
  #pragma pack(1)
  #endif
  .
  .
  .
  #ifdef __cplusplus
  #pragma pack()
  #endif

Node: Int 24h, Next: , Previous: Struct packing, Up: Miscellany

22.13 How to avoid "Abort, Retry, Fail" messages

Q: How do I write a program that accesses floppy and CD-ROM drives, but avoids popping that "Abort, Retry, Fail?" message from DOS?

Q: Other DOS compilers supply a function named harderr or _harderr to hook the critical-error interrupt 24h, but DJGPP doesn't seem to have these...

A: Under DPMI, Int 24h is always hooked by the DPMI server, since Int 24h is issued by the real-mode DOS code, and it is not possible to terminate a DPMI client (like DJGPP programs) from real mode, if you press A in response to that prompt. The default handler under most DPMI servers will just set AL register to 3 and do an IRET, thus silently failing the DOS call that triggered Int 24h. The DJGPP startup code also hooks the protected-mode Int 24h with a handler that fails the DOS call as described above. So in most circumstances you won't see that DOS prompt at all; your program will just see a failed DOS call.

However, some DPMI hosts (notably, QDPMI), will sometimes crash your program if it generates Int 24h, for instance when you access an empty floppy drive. In such cases, or when the default action of failing the DOS call is not good enough, you will have to hook Int 24h with your handler. This should be done in exactly the same manner as hooking hardware interrupts (see how to set an interrupt handler), because Int 24h is one of the few software interrupts that, like all hardware interrupts, are always reflected to protected-mode. Note that CWSDPMI currently doesn't support hooking Int 24h; if you set an interrupt handler, it won't be called.

There are ways to avoid program crashes due to Int 24h (under those DPMI hosts that exhibit this buggy behavior) other than to install a handler for it. For instance, you can test if the floppy drive is empty with a BIOS call before accessing it with DOS functions; there are also similar ways to check if a CD-ROM drive is empty. The library function getmntent (see getmntent.) can be used to detect all the drives that can be safely accessed by DOS; or you can borrow some of the internal functions used by getmntent from the library source distribution, or from the zip picker.

Node: go32-v2, Next: , Previous: Int 24h, Up: Miscellany

22.14 What is that go32-v2.exe program?

Q: What is go32-v2 for?

A: The go32-v2 program does the following:

Node: DXE, Next: , Previous: go32-v2, Up: Miscellany

22.15 What is DXE?

Q: What is a DXE?

Q: Can I make a DLL using the DXE support?

Q: Where can I find information or examples about writing/loading the DXE files?

Q: Why do I get undefined references when I run dxegen?

A: DXE is a limited facility to dynamically load code which is rarely needed in DJGPP. An example is the floating-point emulator code (see the details of DJGPP FP emulator) which is only used on those few machines that lack an FPU. The DXE design is intentionally limited to keep it as simple as possible, so that the code that loads a DXE could be small (it's a few hundreds bytes). Because of this, there are a number of limitations in the DXE mechanism that prevent using it for full-fledged dynamic linking (i.e., a DLL). For instance, the DXE module cannot access variables or functions in the main module.

A DXE cannot link in any library functions which reference static variables (or which call other routines which reference static variables); this effectively prohibits linking in I/O functions, allocating memory, and many other useful things. If you do call any of these, you'll get unresolved externals from dxegen. To work around this limitation, introduce an array of function addresses, or a structure with pointers to functions as its members, which will be used from the DXE at run time to call the "special" routines you cannot link in. Then arrange for the address of this array to be returned by _dxe_load when it loads the DXE, and make the init routine fill the array with the actual addresses of those "special" functions.

Unloading a DXE is also not supported (but I'm told you can add this by making a few simple changes in the C library).

The only place you can find some docs and examples of writing and using a DXE is in the "tests" archive. The example there is exceedingly simplistic, but then so is the entire DXE mechanism....

Node: LFN, Next: , Previous: DXE, Up: Miscellany

22.16 Long Filenames Don't Work!

Q: I cannot make Info find some of its files under Windows 9X...

Q: Why does Make behave as if some of the files were not there?

A: Are you running DJGPP v2.0 on Windows 9X with long filename support enabled (LFN=y in the environment)? If so, set LFN=n from the DOS prompt and try again. If the problems go away, they are probably due to known bugs in some v2.0 programs wrt the LFN support. Make and Info which came with DJGPP v2.0 are two programs which are known to reveal these bugs. Before you decide that you are a victim of these bugs, though, make sure that all the files that your programs need to access have been renamed to their long names. For example, if Make needs to find a file called ALongFileName.cc (because the Makefile tells it to build ALongFileName.o), make sure there indeed is such a file in the directory. Sometimes people use archive tools (like PKZIP) that truncate long names, even on Windows 9X, when they unpack an archive, which leaves you with names like alongfil.cc, which is not the same as the original name when LFN is supported. Be sure to use archivers that support long filenames, e.g. use DJTAR, or rename all the files to their original long names after you unpack the archive.

If the problems persist even though the filenames are correct, upgrade to DJGPP v2.01 or later, where all programs should support long filenames properly. If you cannot upgrade, you will have to disable LFN support (set LFN=n from the DOS prompt, setting it in DJGPP.ENV does not always work in DJGPP v2.0).

Node: Missing separator, Next: , Previous: LFN, Up: Miscellany

22.17 Make says "missing separator"

Q: When I invoke Make, it refuses to do anything and prints a cryptic message: "makefile:10: *** missing separator. Stop." Now what kind of excuse is that?

A: Unlike most other DOS Make programs which accept any whitespace character at the beginning of a command in a rule, GNU Make insists that every such line begins with a TAB. (Most other Unix Make programs also require TABs, and the Posix standard requires it as well.) Make sure that the line whose number is printed in the error message (in this case, line 10) begins with a TAB.

Beginning with version 3.78, GNU Make prints a message that hints at a possible SPACEs-vs-TAB problem, like this: 

*** missing separator (did you mean TAB instead of 8 spaces?).  Stop.

If you need to repair a Makefile that uses spaces, one way of converting leading spaces into TABs is to use the unexpand program from the GNU Textutils package (v2gnu/txt20b.zip from the DJGPP sites). Another possibility is to open the Makefile in RHIDE and choose the Edit->Compact Text option from the menu bar.

Note that there are editors that automatically replace TABs with spaces, so even a Makefile that used to work can become unworkable if you edit them with such an editor. Don't use such editors.

Another, more rare, cause of the above error message is if you use static pattern rules (with the % character) incorrectly. Read the documentation that comes with Make carefully and try to find the error.

Node: Modification time, Next: , Previous: Missing separator, Up: Miscellany

22.18 Make says "foo has modification time in the future"

Q: I keep getting messages about file modification time in the future when I compile my programs using Make....

A: This happens on fast machines running Windows/NT and Windows 9X. (However, somebody even reported such a problem on a FAT32 drive under plain DOS.) It is evidently due to a misfeature in the way Windows reports the time a file was last modified. Current versions of Windows 9X store this time with 2-second granularity, but the file creation time is stored with a 100-nanosecond granularity. It seems that Windows blindly adds 2 seconds to the system clock reading when it calculates the file modification time, apparently to prevent it from being older than the file creation time, which could happen because of the greater accuracy used to store the creation time. On a fast machine, this 2-second add-on can very easily make the file modification time be ahead of the system clock when Make checks the time stamp of a file it has just created/updated.

GNU Make reports such cases because inconsistencies in file times could easily defeat Make's decisions about which files need to be rebuilt. In particular, if some of the files reside on a networked drive, and there's a clock skew between the machine where Make runs and the one which exports the drive, Make could really fail to rebuild some files.

DJGPP ports of GNU Make v3.77 and later allow for up to 3 seconds of positive difference between the file timestamp and the system clock (that is, the file is allowed to be up to 3 seconds into the future), before the above warning is printed. So upgrading to the latest version of Make should eliminate such bogus warnings and leave you only with messages due to real clock skews.

Node: Dual DOS/Windows, Next: , Previous: Modification time, Up: Miscellany

22.19 How to Set Up a Dual DOS/Windows Installation

Q: I want to be able to run DJGPP both under Windows 9X with long file names, and when I boot into plain DOS, where long file names aren't supported. How can I set that up?

A: Such a setup is possible, but it involves a few special actions, and some vigilance on your part during routine operations.

First, you must set the option in the Windows registry which prevents it from using numeric tails when it invents short 8+3 aliases for long file names. When numeric tails are enabled, and a file with a long name is created, Windows generates a short 8+3 alias for that long name by attaching to the first 6 characters of the basename a numeric tail ~n, where n is a digit. For example, a file called ALongFileName.LongExtension will get a short alias alongf~1.lon. When you then reboot into plain DOS, your programs will see this short version only, which will almost certainly break them, since, when a program running under DOS asks for a file with the above long name, DOS transparently truncates it to alongfil.lon, and such a file does not exist. Disabling the numeric tails forces Windows not to use numeric tails unless there is another file in the same directory whose short alias clashes with that of the new file. If no such clash happens, Windows will simply truncate the long name as DOS would, which is exactly what you want.

Here is how you disable the numeric tails on Windows 9X:

As an alternative to running regedit, you can create a file named, say notail.reg (the name is arbitrary, but it must have a .reg extension), with this content: 

REGEDIT4

[HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\FileSystem]
"NameNumericTail"=hex:00

then double-click on the name of this file in Explorer or My Computer. Windows will then run regedit for you. You still need to reboot the machine afterwards.

If setting NameNumericTail to 0 breaks some programs, you can restore its original setting temporarily, while you run those programs. NameNumericTail only affects the short names of new files being created, it has no effect on the files that already exist.41

Besides the numeric tails, you need to make sure any files and directories you create have unique 8+3 aliases which are true truncations of the long names to 8+3 limits. This means that you should avoid file names with leading dots, such as .emacs and .bashrc, file names with more than a single dot, like make-3.77.tar.gz, or file names which include characters not allowed by DOS, like libg++.a.

One other problem is to avoid using programs which create numeric tails even if they are disabled in Windows. One such program is pkunzip version 2.50. Don't use it, if you want to keep your dual DOS/Windows installation in working order.

The most simple method of deciding at boot time which configuration (DOS or Windows) to start is to edit the (hidden) file MSDOS.SYS, which is a text file in Windows 9X, and force the Windows boot process to present a menu where one menu item, called "Command Prompt Only", allows you to start DOS 7 without the Windows GUI. To this end, change the line of MSDOS.SYS that reads "BootMenu=0" to say "BootMenu=1" instead, and reboot. Since MSDOS.SYS is a hidden file, you will need to remove the hidden attribute from it before you can edit it; use the ATTRIB command for that.

Node: Zoneinfo, Next: , Previous: Dual DOS/Windows, Up: Miscellany

22.20 What is in that zoneinfo directory?

Q: When I installed DJGPP v2, it created a directory named zoneinfo with a lot of small files that take up 3.5MB of my disk space. What are they for? Can I delete them?

A: These files exist so that time-related library functions can correctly calculate the offset between the local time and the UTC (Universal Coordinated Time). This offset is required when you get files from another time-zone, like from the Internet, or when you download an archive that was compressed in another time-zone.

One case where the time stamps might be very important is when you need to rebuild some package with Make. Make uses file time stamps to decide which files need to be rebuilt. Another case is if you distribute some files compressed with Zip and want your recipients to be able to restore the correct time stamps of your files when they unzip them.

If you don't care about file time stamps being incorrect in such cases, you can delete all those files and never look back.

You might wonder why we need all these zoneinfo files when the UTC offset is required. Well, the simplest way to tell programs what the UTC offset is, is to have the user specify a single number which is the offset; but then this number needs to be changed twice a year, to accommodate for the daylight saving time. Another, not-quite-so-simple way is to have the user specify the current UTC offset and the DST rules; but this is a tedious and error-prone process, and many users get it wrong. Both of these methods have the drawback that if the rules change, programs misinterpret old time-stamps, since they treat them according to new rules. Using a table that is read from a file and includes the offset calculation rules for every year avoids all these problems and requires the user to point the TZ environment variable to the file that is pertinent to his/her time zone, which is easy: 

 set TZ=c:/djgpp/zoneinfo/israel

or 

 set TZ=c:/djgpp/zoneinfo/us/alaska

To find the rule suitable for your location, look into the src subdirectory of zoneinfo and browse the file whose name is your continent/part of the world. If no binary file exists with the name of your zone, you can create one with using the time-zone compiler zic which comes with the v2/djtzn203.zip file (it unzips into the etc subdirectory of the main DJGPP installation directory).

A public domain time-zone database exists, and is updated from time to time with the latest world-wide changes to the offset calculation rules. (The rules change because politicians in different countries make laws that change the local clock settings.) The contents of the zoneinfo directory which comes with DJGPP is based on this database, but if you want the latest rules, you can download them from the net as tzdata*.tar.gz; tzcode*.tar.gz in the same directory includes the programs that can be used to generate the offset tables from their source in tzdata*.tar.gz, the latest implementations of POSIX library functions that use time-zone information, and the manual pages that document the rules and the software. The last update as of this writing was in September 1999.

On any single machine, you don't need more than a single file from that directory, which is the file for your time zone; once you find that file, you can safely delete the rest. But if you distribute a program that uses the TZ setting, you will have to include all of the files, or tell your users how to get and install them.

Node: dev directory, Next: , Previous: Zoneinfo, Up: Miscellany

22.21 The dark secrets of the /dev/ directory...

Q: All DJGPP programs cannot find files in the d:\dev directory, but work okay in other directories. What is going on here??

Q: I installed DJGPP in the e:/dev/djgpp, and it doesn't work!

A: This is an unfortunate side-effect of the special treatment given to the \dev directory on each drive. DJGPP transparently supports Unix-style devices such as /dev/null and /dev/tty, so that programs ported from Unix that refer to these devices will work. Unfortunately, due to a half-hearted way DOS and Windows support devices, the DJGPP library must treat the \dev directory specially. The net effect is that if you have a real directory by that name, you will get erratic behavior.

A work-around is either to rename the \dev directory to some other name, like \devel, or move it down the directory hierarchy, for example make it d:\software\dev. (The special treatment is only reserved to the \dev directories immediately under the root of every drive.)

Node: ELF vs COFF, Next: , Previous: dev directory, Up: Miscellany

22.22 How about switching to ELF as DJGPP's object file format?

Q: I hear all that stuff about the limitations of the COFF format, and I don't understand why won't DJGPP switch to a modern standard such as ELF?

A: DJGPP uses COFF for historical reasons: at the time it was developed ELF was not available yet. There are several grave reasons why DJGPP didn't switch to ELF yet:

None of the above is a show-stopper, so such a switch is possible. But it is a large project, and without several devoted volunteers, chances are it will never happen.

Node: Random numbers, Next: , Previous: ELF vs COFF, Up: Miscellany

22.23 How to produce random numbers?

Q: How do I produce random numbers with DJGPP?

Q: I keep getting the same random numbers each time I run my program. How do I get a different series on every run?

Q: How do I get random numbers between 20 and 200?

A: DJGPP has in its library several functions to produce series of pseudo-random43 numbers. One of them, rand, is part of the ANSI C Standard, and is therefore very portable to other environments. Other random-number functions, random and the rand48 family of functions, are available on almost every Unix platform, but are usually unsupported by DOS/Windows compilers. On the other hand, series produced by random and rand48 have better qualities than those produced by rand. In particular, the least-significant bits in the numbers produced by random are much more random than those you get from rand, so if you need, say, a random number between 0 and 4, and portability is not an issue, you will get better results with random () % 5. However, the DJGPP implementation of rand is quite good, so when portability is important, you should use rand.

Both rand and random return a pseudo-random integer in the range [0..RAND_MAX), where RAND_MAX is defined in the stdlib.h header. Within the rand48 family, some functions return integers, either in the range [0..RAND_MAX) or in [-RAND_MAX..RAND_MAX), while others return a double value between 0.0 and 1.0.

By default, every time you restart a program, you get the same series of pseudo-random numbers. This is important in some applications, because it allows to reproduce exactly the results of running a program which used random series, and thus makes debugging easier. But sometimes, e.g. in a game, you will want a different series every time. To achieve that, you need to initialize the random series with a different seed. Every random-generating function has its own seed function provided for this purpose: rand has srand, random has srandom, the rand48 family can be seeded with either srand48 or seed48. You seed the series with a single call to the appropriate seed function, and then proceed by calling rand, random, etc. as usual.

A popular way of getting a different seed every run is to use the current system clock as the seed, like this: 

  srand (time (NULL));

If the 1-second granularity of the values returned by time is not enough for you (e.g., if you need to generate more than one series every second), use gettimeofday or uclock, or use the values returned by rand as an argument to srandom (or vice versa).

To produce random integers from the inclusive interval [low..high] (where low and high are two integer numbers), use code like this: 

 #include <stdlib.h>

 int random_number =
       low + (double)rand () * (high - low + 1) / RAND_MAX;

This produces a more random sequence than if you use the % operator, but for the price of producing slower code (since it involves floating-point math).

If you want to know more about random number generation, I suggest reading the article Random Number Generators: Good Ones Are Hard To Find, by Stephen K. Park and Keith W. Miller, in CACM, v31(10), 1988, pp. 1192-1201.

Node: Lexicon, Next: , Previous: Random numbers, Up: Miscellany

22.24 What are all these buzzwords I see?

Q: All your FAQs and tutorials seem to take for granted that I know what words like "compile", "link", "makefile" etc. mean. Where can I find all these buzzwords explained??

A: The DJGPP docs use some basic computer lexicon, without which it would be impossible to communicate. If you find yourself completely lost in terminology you don't understand, and this FAQ doesn't explain it, try looking up the words in the DJGPP Lexicon page. Another excellent resource for computer-related terminology is the Jargon File, also available from the GNU FTP site as an Info file and in several other formats.

Node: void main, Next: , Previous: Lexicon, Up: Miscellany

22.25 What should the main function return in a C/C++ program?

Q: Why does everybody tell me that void main is bad?

Q: If void main is incorrect, how come the compiler lets it compile?

A: The ANSI/ISO C Standard specifies that the main function be declared in one of the following two ways: 

 int main (void);

or 

 int main (int argc, char **argv);

In both cases the return type is an int, and your main function should therefore either return an int or call the library function exit, when the program ends. The C++ standard includes a similar requirements for C++ programs.

Since the runtime environment assumes that main returns an int, declaring main with any other return type, including void, invites trouble. The compiler might compile such a program, since the ANSI Standard doesn't require it to fail, but the behavior of such a program is, in the Standard's parlance, "undefined" (read: anything can happen). That is why GCC will print a warning in these cases if you use the -Wall switch.

To summarize, using void main is unsafe and can potentially do evil things to your program. It is best to avoid it.

Note that the C++ standard, in contrast to the C standard, explicitly prohibits void main(), and explicitly says that if the controls reaches the end of main without encountering a return statement, the effect is that of executing return 0;. When compiling a C++ program, GCC automatically generates the code to return zero from the main function, in case the programmer leaves that out.

Node: Reboot the PC, Next: , Previous: void main, Up: Miscellany

22.26 Rebooting the PC from a DJGPP program

Q: How can I reboot the PC from a DJGPP program?

A: There are several possible ways to achieve this:

Node: usleep, Next: , Previous: Reboot the PC, Up: Miscellany

22.27 Delaying execution for short periods of time

Q: The function usleep doesn't work for me! When its argument is less than 10000, there's no delay at all, and when the argument is larger than 10000, the delay is always the same....

Q: How can I delay the execution of my program for 2msec?

Q: I need to pause my program for 100 microseconds. Can I do it?

A: Most time-related facilities in DJGPP have the same 55msec granularity of the time intervals they measure. This is because the timer tick interrupt that updates the time has the frequency of 18.2Hz. This is why calling usleep with arguments less than 55000 produces strange effects: the resolution of the argument is 1usec (for compatibility with other compilers), but the granularity is still 55msec.

If you need to pause your program for periods of time shorter than 55msec, you have several alternatives:

  1. For delays longer than 1msec, use the library function delay. It is based on the CMOS clock chip whose frequency is 1024Hz.
  2. For periods shorter than 1msec, write your own wait loop that calls library function uclock to see when the pause time expires. uclock measures time with 840-nanosecond granularity. Here's a possible implementation of such a wait loop: 
     #include <time.h>

 uclock_t start;

 /* Wait for 200 microseconds.  */
 start = uclock ();
 while (uclock () < start + UCLOCKS_PER_SEC / 5000)
   ;

Node: CGI programs, Next: , Previous: usleep, Up: Miscellany

22.28 CGI programs and DJGPP

Q: I wrote a CGI program and compiled it with DJGPP, but it doesn't seem to work....

A: If you are using a Windows Web server, it probably won't work. The reason is that Windows programs cannot easily redirect standard input and output of DOS programs (because DOS programs are run by Windows in a different Virtual Machine), and many Windows Web servers don't consider the case of a DOS CGI program, and don't bother to include the machinery necessary to do the redirection in a way that would work for DOS programs. So the output of your program never gets to the server.

Consult the docs of the server: it might include some feature that enables redirection from DOS programs. If that doesn't help, rebuild your CGI program with a Windows compiler, such as Mingw32 or Lcc-Win32 (see free Win32 compilers), and it will work.

Node: Input EOF, Next: , Previous: CGI programs, Up: Miscellany

22.29 Why Do I Get EOF From stdin?

Q: Whenever I run my program, which reads stdin, from RHIDE or Bash, it gets EOF indication immediately, without any input being typed! This program runs okay from the DOS prompt, so something must be wrong with Bash and RHIDE, yes?

A: This is a known (mis-)behavior of DOS: to clear the EOF condition of the console device, you need to write something to it. If during some previous invocation of a program you typed Ctrl-<Z>, the resulting EOF condition will stick until something is written to the console device. When the program is run from the DOS prompt, that "something" is the prompt string printed by COMMAND.COM. However, RHIDE doesn't print any prompt, and Bash prints its prompt via the BIOS. So, as soon as you type Ctrl-<Z>, the console device is stuck in the EOF condition, and all subsequent invocations of programs that read standard input immediately get receive EOF and exit. (This problem is not specific to DJGPP: Norton Commander also exhibits it.)

A work-around is to output something to stdout or stderr (assuming they both are connected to the console device).

Node: FAQ format, Previous: Input EOF, Up: Miscellany

22.30 Generating the FAQ in your favorite format

Q: How can I generate the FAQ list in a format I'm used to?

A: First, I suggest to check whether the FAQ is already available in your format. The FAQ distribution includes the Info, plain-ASCII (text), and HTML (one large .html file) versions of the FAQ list. (Personally, I recommend to use the Info version, because Info readers generally have superior search facilities.) More formats will be available as the tools for their generation are developed/tested.

If you prefer to read the FAQ list as hard-copy, get faq230p.zip from the same place. It includes the FAQ in PostScript and PCL formats; the former is for printing on a PostScript printer, the latter is for the LaserJet series. Be warned: the FAQ is a large document, more than 200 printed pages. faqNNNp.zip also includes a .dvi file which you can print or view using one of the available DVI drivers and previewers, such as Ghostscript or dvivga.

If none of these formats is good enough for you, you will need to get the FAQ sources and convert them into the format of your liking. The sources of the latest version of this FAQ list are on SimTel. This includes the FAQ sources themselves (written in Texinfo), and all the auxiliary tools required to produce all the formats in faqNNNb.zip and faqNNNp.zip.

Once you download the sources, you will need one or more of the tools listed below to generate the FAQ list in other formats.

A program called Makertf can reportedly be used to convert a Texinfo sources of this FAQ to the Rich File Format which can then either be browsed by an RTF browser (such as Adobe Acrobat) or converted into a Windows Help file with a Windows Help compiler. The Windows Help Compiler is available via anonymous ftp from the Microsoft ftp site. I'm told that the Windows Help compiler, hcp, issues a lot of warnings when it runs on Makertfs output, but these warnings can be safely ignored.

A derivative of TeX called PDFTeX can be used to generate a PDF file from TeX sources. This means that a Texinfo source can also be submitted to PDFTeX, but I didn't try that. PDFTeX was ported to DJGPP and can be downloaded from its home site via FTP. Another option for generating a PDF version of the FAQ is to use the DviPDFm program which can convert the FAQ in the DVI format (included in the faqNNNp.zip distribution) into a PDF file. DviPDFm is available with several TeX distributions and can be compiled with DJGPP.

There's also a program called INFNG that can be used to convert the Info (not Texinfo) version of the FAQ to the Norton Guide format. INFNG is available from the DJGPP archives.

If you know about any format not mentioned above that can be generated using widely available tools, please drop me a note so I could update this list and consider that format or those tools for inclusion in a future release of the FAQ. If you develop any such tools, consider uploading them to a site where they will be publicly available, and tell me about that site.

Note that the FAQ sources are heavy users of the Texinfo macro facility, so any conversion program that doesn't support Texinfo macros will probably have hard time coping with the FAQ. When confronted with this problem, try feeding the converter with the macro-expanded version of the FAQ (the Makefile in the source distribution has a special target for such cases).

Node: About, Next: , Previous: Miscellany, Up: Top

23 About this FAQ

Maintainer: Eli Zaretskii.

Copyright © 1994, 1995, 1996, 1997, 1998, 2000 by Eli Zaretskii. This FAQ may be freely redistributed with the DJGPP package or any part thereof, provided that you don't prevent anybody else from redistributing it on the same terms, and that this copyright notice is left intact.

Comments about, suggestions for, or corrections to this FAQ list are welcome. Please make sure to include in your mail the version number of the document to which your comments apply (you can find the version at the beginning of this FAQ list).

Much of the info in this FAQ list was taken from the DJGPP mailing list/news group traffic, so many of you have (unbeknownst to you) contributed to this list. The following people deserve special credit for reading this list in its previous versions and providing useful feedback, comments, information and/or suggestions: 

John M. Aldrich
Anthony Appleyard
Gurunandan R Bhat
John Bodfish
Francois Charton
Alain CULOS
Bill Currie
Bill Davidson
DJ Delorie
Tom Demmer
Nate Eldredge
Juergen Erhard
Andy Eskilsson
Jeremy Filliben
Peter Gerwinski
Till Harbaum
James W. Haefner
Kris Heidenstrom
Koen Van Herck
Vik Heyndrickx
Robert Hoehne
Gordon Hogenson
Daniel Horchner
Harry Johnston
Jules
Martynas Kunigelis
Pieter Kunst
Y. Lazarovitch
Alexander Lehmann
Marty Leisner
Dave Love
Randy Maas
Cameron Mallory
Colin S. Miller
Duncan Murdoch
Rob Nader
Eric Nicolas
Adrian Oboroc
Jan Oonk
Elliott Oti
Bob Paddock
Esa A E Peuha
Prashant TR
Walter Prins
Steve Salter
Charles Sandmann
Terrel Shumway
Martin Strömberg
Andrew Szymkowiak
Launey Thomas
Chris Tilbury
Ned Ulbricht
Stephen Turnbull
Ronan Waide
Morten Welinder
Anthony Edward Wesley
K.B. Williams
Mark H. Wood

Node: Topic Index, Next: , Previous: About, Up: Top

24 Topic Index

This is an alphabetical list of all the topics covered in this FAQ. Use it to search for a description of your problem and follow the link to find the answer(s).

Node: Program Index, Previous: Topic Index, Up: Top

25 Program Index

This index lists the problems and solutions by the program/package to which they pertain. If you know what program or package gives you the trouble, look it up here.

Footnotes

  1. Please report any issues that couldn't be found via the indices to the maintainer of this FAQ, whose name and e-mail address can be found near the end of the FAQ.

  2. Posix is an international standard for a portable operating system. It specifies facilities of a compiler, its libraries, and the basic set of development tools. Posix was originally modeled on Unix systems, but is currently supported by most modern operating systems.

  3. For example, the DOS version of the well-known game Quake by id Software was compiled with DJGPP.

  4. The typo in the word Exception is in the actual message popped by Linux.

  5. This was reported for DOSEmu version 0.66.7; the format of dosemu.conf might be different in version 0.9x.

  6. But the development environment will only run on Windows 9X/NT.

  7. This DLL can be stripped off the debugging symbols using the strip utility, which leaves a much smaller--about 500KB--file.

  8. The BUFFERS=40,8 setting defines a primary cache of 40 512-byte blocks and a secondary cache of 8 blocks. The primary cache is used by DOS to store the data actually read by a program, in case it is re-read shortly afterwards; while the secondary cache is used to read data ahead of the requests, which is optimized towards sequential reads.

  9. ftp.simtel.net is actually several ftp sites arranged in a rotating pattern of IP addresses to help balance the load and to avoid access problems due to network outages and simultaneous user limits.

  10. The version numbers of the packages listed here might not be up to date by the time you read this. For the latest versions, check out the DJGPP Mini-FAQ posted weekly to the the comp.os.msdos.djgpp news group. The file LISTINGS.zip available on every DJGPP site includes one file for every zip that lists all the files inside that zip; download and unzip it on your machine, and you've got a set of files you can search for substrings. Another place where you can look for the latest versions of all files is on the DJGPP server, which is also a convenient way of downloading individual files if you have lost them somehow.

  11. Like its name suggests, texi2html converts a Texinfo source to HTML.

  12. You will need to install the port of Bash and some auxiliary utilities to be able to run shell scripts; texi2dvi itself is part of the GNU Texinfo distribution and comes with the v2gnu/txiNNNb.zip archive.

  13. The example uses Unix-style forward slashes, but DOS-style backslashes can also be used.

  14. While admittedly perverse, this little monstrosity was written with the sole purpose of demonstrating that C and C++ have quite different semantics under certain circumstances. Some people think that C is a proper subset of C++; the above example shows that this is not true.

  15. If you define an empty substitute for __crt0_setup_arguments, you don't need to define a substitute for __crt0_glob_function.

  16. In particular, running strip on a program and then compressing it with DJP would produce a program that crashes upon startup.

  17. Those who want all the details should consult a good book about DOS internals, such as Undocumented DOS, 2nd ed. by Andrew Schullman, or Geoff Chappel's DOS Internals.

  18. All DOS programs get the default 20-handle table when they start; DOS only copies the first 20 handles into the child, so it is not possible to inherit more than 20 handles. The expansion of the default 20-handle table to 255 handles is a special feature of the DJGPP library, and it only happens when the programs exhausts all of the 20 handles while it runs. Therefore, when all of the first 20 handles are taken up by files inherited from the parent program, the child program can fail to start because the DJGPP stub loader needs one free handle to open and read the COFF executable into memory. The stub cannot use the enlarged 255-handle table, since it cannot call the DJGPP library. Such problems indeed happen in programs compiled with DJGPP v2.01; v2.02 fixes this bug.

  19. Each handle created by a call to open uses up one slot in the SFT, whereas a handle created by dup just increments the use count of a slot that was already in use.

  20. That's the program being debugged, in case you didn't know.

  21. Programs that create machine code in malloced storage and then jump into it could have their EIP above EBP. The Allegro library utilizes this technique in some of its functions (specifically, compiled sprites and stretched blits are normally performed in this way).

  22. If you absolutely need to call ld.exe directly, invoke gcc once with a -v switch and you will see what are the arguments that you should pass to the linker in your case.

  23. Here's a more detailed explanation. DOS cannot access memory above 1MB mark, where your DJGPP program lives, since real-mode addresses are 20-bit wide, and 20-bit addresses cover only the first megabyte. So, each time a DJGPP program needs to call a DOS function (or any other real-mode service, like some BIOS interrupt) and needs to pass data to or from that service, we must use some buffer in conventional memory to communicate with DOS and BIOS. The transfer buffer is a block of conventional memory that the DJGPP startup code allocates for this purpose. When a real-mode service is called, the data that needs to be submitted to it is copied to the transfer buffer, and the address of the transfer buffer is passed to the real-mode service. If the service returns some data (e.g., if you want to read a portion of a file), data is copied from the transfer buffer when the service returns.

    The transfer buffer primarily exists for library functions, but it can also be used by an application, if it needs to invoke real-mode services.

  24. Actually, the maximum possible value is FEF0h, or 65254 in decimal, because the transfer buffer is created by the startup code by resizing the PSP memory block. Since the resized block needs to leave 256 bytes for the PSP, and needs to be aligned on a 16-byte boundary, you cannot have the entire 65535 bytes for the transfer buffer. In DJGPP v2.01, if you invoke stubedit with a bufsize=64k parameter, what you actually get is a 2KB buffer, since the combined size of the PSP and the transfer buffer will wrap around in a 16-bit variable when the startup code computes it. The versions of stubedit which will come with DJGPP v2.02 and later explicitly warn you about this case and will reset any value that is too large to the maximum allowed size of FE00h (65024 decimal) bytes--this is less than FEF0h because the latter is not aligned on the 512-byte DOS sector size, which could slow down disk I/O.

  25. Without a real x87 FPU, an exception is generated by the CPU each time a floating-point instruction is seen in the code. __djgpp_exception_processor is called for each such exception and services it by calling the emulator, emu387.dxe, or functions from the emulator library libemu.a (if the program was linked with -lemu), to emulate the instruction. Since exception processing incurs a lot of overhead, this emulation is slow.

  26. The DPMI 0.9 spec does not provide any means for the application to control where in the address space will the DPMI server allocate a particular chunk of memory. The application asks the DPMI server for whatever amount of memory it needs, and gets a chunk of that size whose address can be anywhere. Since the stack must be contiguous and expands downwards, growing it during program's run would require a very complicated code, unless it is pre-allocated at startup.

  27. In case you wonder, the name !proxy comes from the string which identifies the use of this method: instead of getting the actual command line, the program gets !proxy followed by the address of the actual command line.

  28. If you use MASM or LIB32, please post your experiences to the comp.os.msdos.djgpp news group, so that I can make the above instructions less vague.

  29. Note that mixing object files from different compilers generally doesn't work at all, even if all the object files are in .obj format.

  30. This function calls the video BIOS interrupt 10h to allow bright background colors to be used instead of blinking characters. DJGPP has a library function, called intensevideo, to do that, but for the purpose of this example, let's assume we have reasons not to use it.

  31. The DPMI spec indicates that you should not use the default stack if your procedure/interrupt handler uses more that 60 bytes, or 1/8 of the total stack space available by default.

  32. DJGPP v2.01 makes the limit of _dos_ds be 4GB, which effectively disables memory protection when you use that selector. However, since no memory outside the first 1MB is properly mapped into your program's address space without additional DPMI calls, and the DPMI host is then free to put memory-mapped devices, such as Weitek I/O space or the linear frame buffer of an SVGA, on any address it sees fit, that huge limit is an unjustified security hole. DJGPP v2.02 will really be limited by 1MB+64KB.

  33. If you want to decipher the 8-byte structure that is passed to __dpmi_set_descriptor in this example, read the documentation of the __dpmi_get_descriptor library function in the library reference. This structure is the descriptor maintained by the processor for each protected-mode segment, such as those loaded into the CS and DS registers.

  34. If you are using this example in your program, don't forget to disable the handler at program's exit by calling the same function 0Ch of Int 33h with a zero mask in the CX register, and then deallocate the callback by calling the _go32_dpmi_free_real_mode_callback library function. Also, remember that all code and data touched by the handler must be locked, otherwise it will crash under some DPMI servers, such as CWSDPMI.

  35. Locking a region of memory means that this region should be always present in RAM. Usually, the virtual-memory mechanism is allowed to page regions out of RAM when it needs to load another region that is not loaded. This happens if the program uses more memory than what is physically available to it. When a program needs to access an address that isn't currently in RAM, the operating system will look for some memory region that wasn't accessed for quite some time, and replace it with the block that needs to be accessed now. Locking a region prevents that region to be paged out, for as long as the program runs.

  36. This discussion does not pertain to the __dpmi_int function as opposed to the _go32_dpmi_simulate_int function. On the contrary, you will be much better off using __dpmi_int, since it automatically zeroes out some of the members of the real-mode registers structure, while with _go32_dpmi_simulate_int you need to do that by hand.

  37. Actually, it is possible to avoid reentrancy problems in interrupt-driven programs: programs known as TSRs (Terminate and Stay Resident) have been doing that for years. But doing so requires hooking and monitoring many DOS and BIOS interrupts, to know when it is safe to page. If CWSDPMI would use these techniques, it would take much more DOS memory to load and run. It would also need to be updated with every new DOS release, since some of the internal DOS structures it would need to monitor change their address and/or layout with new versions of DOS.

  38. CWSDPMI has an experimental VDS support in its sources, but the distributed binary was compiled without it. Contact Charles Sandmann if you want to try to enable VDS support in CWSDPMI.

  39. In case somebody thinks there is a contradiction here: I don't consider a requirement to provide information to be a restriction.

  40. Note that this still allows the struct to be padded at the end.

  41. For some reason, Microsoft doesn't like it when users disable numeric tails. Several Microsoft publications warn against doing that, and I'm told that Windows 98 has made it harder to disable them. I'm not sure why do they object so much. Presumably, some programs rely on certain directories to have numeric tails, so that they could be found even in plain DOS mode. Apparently, some of those programs have short aliases such as PROGRA~1 (the short version of the Program Files directory) hard-wired into them, and Microsoft is afraid you could reinstall or move those directories when numeric tails are disabled, and thus cause such programs not to find their "home". It is obvious that such programs are badly broken (e.g., the short alias could easily be PROGRA~2), and you have every right to yell at the vendor who sells them to you. But even if you have no other way than to live with them, my experience shows that you have nothing real to worry about. Remember: numeric tails only have effect when files are created or renamed. So, if you want to be on the safe side, re-enable them before installing Windows software, especially if the programs you install need to run in DOS mode as well (a typical example would be a disk-recovery package such as Norton Utilities). Then disable numeric tails again, once the installation is over.

    For what it's worth, I always run my system with numeric tails disabled, and I have yet to see a single real problem.

  42. If you need Binutils configured for elf32-i386 target that can be used with DJGPP, you can find it at http://www.multimania.com/~placr/.

  43. Since these series are computed using a deterministic algorithm, they are not really random. Real random numbers can only be a result of unpredictable physical processes such as radioactive decay etc. However, a good algorithm for pseudo-random numbers produces a series of numbers that pass many tests for randomality.