Goodbye Livejournal
I've decided to discontinue using Livejournal. The posts on this blog will remain up but future posts will appear on Google+.
Initialisation dogma considered harmful
In C, variables aren't initialised to a value automatically. I've seen some people who seem to have adopted a rule of initialising every variable they use when they declare it. That is to say, they write code like this:
Even though they initialise the variables before using them in their code, they include an initialiser in the declaration to set the value to zero or NULL. Presumably the logic is something like: variables that aren't initialised can hold any value, which can cause bugs, so it's a good idea to always initialise every variable just in case.
In this situation, "obj" is "uninitialised", as there is a code path where it isn't set that leads to a crash. But technically, it has been initialised in the variable declaration, as part of this initialisation dogma. If it hadn't, the compiler would have been able to detect this bug, but it can't.
void frobnicate(void)
{
int num_frobs = 0;
void *frobs = NULL;
num_frobs = get_num_frobs();
frobs = malloc(num_frobs);
....
Even though they initialise the variables before using them in their code, they include an initialiser in the declaration to set the value to zero or NULL. Presumably the logic is something like: variables that aren't initialised can hold any value, which can cause bugs, so it's a good idea to always initialise every variable just in case.
I think this is an actively harmful practise and should be discouraged. The reason I think this is that modern C compilers can detect when variables might be used without being initialised first (and gcc is scarily good at it), and adopting a rule like this of initialising everything breaks that functionality. Consider an example like this:
void clever_function_name(void)
{
mystruct *obj = NULL;
int bar = 0;
if (somecondition) {
obj = new_mystruct();
bar = 1;
} else {
// oops, forgot to initialise obj here as well
bar = 2;
}
...
obj->member = bar;
}
In this situation, "obj" is "uninitialised", as there is a code path where it isn't set that leads to a crash. But technically, it has been initialised in the variable declaration, as part of this initialisation dogma. If it hadn't, the compiler would have been able to detect this bug, but it can't.
I think the key thing is that initialising to a value like zero or NULL doesn't actually solve anything or prevent any bugs because it's just an arbitrary value. In a general sense, having a variable that is set to those values isn't any more helpful than having some random garbage that got left on the stack. It doesn't prevent uninitialised variable bugs, but it does prevent the compiler from detecting them.
How to coerce gdb into giving a backtrace
When debugging with gdb, I sometimes encounter a problem in getting it to recognise a stack trace. Sometimes it only gives a few function calls, or none at all. Here's an example:
This obviously isn't a very useful backtrace. At this point it can be useful to have a look at the registers and the contents of the stack.
The values in bold are the frame pointers, forming a linked list back up the stack. If you compiled with -fomit-frame-pointer things will be harder to figure out. In this case it certainly doesn't look like the stack is corrupt. Perhaps gdb is just confused.
(gdb) bt #0 0xb7886424 in __kernel_vsyscall () #1 0xb7559163 in ?? () from /lib/i686/cmov/libc.so.6 #2 0xb74f1387 in ?? () from /lib/i686/cmov/libc.so.6 Backtrace stopped: previous frame identical to this frame (corrupt stack?)
This obviously isn't a very useful backtrace. At this point it can be useful to have a look at the registers and the contents of the stack.
(gdb) info registers eax 0xfffffe00 -512 ecx 0x80 128 edx 0x2 2 ebx 0xb75c33a0 -1218694240 esp 0xbfe55018 0xbfe55018 <= stack pointer on x86 ebp 0xbfe55048 0xbfe55048 esi 0x0 0 edi 0x0 0 eip 0xb7886424 0xb7886424 <__kernel_vsyscall+16> eflags 0x202 [ IF ] cs 0x73 115 ss 0x7b 123 ds 0x7b 123 es 0x7b 123 fs 0x0 0 gs 0x33 51 (gdb) x/64x $sp 0xbfe55018: 0xbfe55048 0x00000002 0x00000080 0xb7559163 0xbfe55028: 0xb75c33a0 0xb75c1ff4 0x09ec0470 0xb74f1387 0xbfe55038: 0xb77c4afd 0xb782199c 0xb782199c 0x09ec0478 0xbfe55048: 0xbfe55078 0xb77c4bf4 0x09ec0478 0xb782199c 0xbfe55058: 0x0000ffff 0xb75c3438 0xbfe55098 0xb789a240 0xbfe55068: 0x00000000 0xb782199c 0x0000ffff 0xb75c3438 0xbfe55078: 0xbfe55098 0xb7818736 0x09ec0478 0x00000000 0xbfe55088: 0xb77c50cd 0xb782199c 0xb7818709 0xb782199c 0xbfe55098: 0xbfe550b8 0xb77c54ad 0x00000000 0x00000000 0xbfe550a8: 0x0000000b 0xb75c3438 0xb77c5449 0xb782199c 0xbfe550b8: 0xbfe550c8 0xb77bb7dd 0xb782199c 0x0000000b 0xbfe550c8: 0xbfe550e8 0xb77bb84e 0x0000ffff 0xb789a240 0xbfe550d8: 0x00000000 0xb77bb830 0xb77bb839 0xb782199c 0xbfe550e8: 0xbfe55108 0xb77bc08f 0x0000000b 0x00000000 0xbfe550f8: 0x00000000 0x00000010 0xb75c1ff4 0x3efafafb 0xbfe55108: 0xbfe556c8 0xb7886400 0x0000000b 0x00000033
The values in bold are the frame pointers, forming a linked list back up the stack. If you compiled with -fomit-frame-pointer things will be harder to figure out. In this case it certainly doesn't look like the stack is corrupt. Perhaps gdb is just confused.
Pick a location slightly further up the stack and set the $sp variable, and suddenly backtrace works!
(gdb) set $sp=0xbfe55048 (gdb) bt #0 0xb7886424 in __kernel_vsyscall () #1 0xb782199c in ?? () from /usr/lib/libSDL-1.2.so.0 #2 0xb7818736 in ?? () from /usr/lib/libSDL-1.2.so.0 #3 0xb77c54ad in ?? () from /usr/lib/libSDL-1.2.so.0 #4 0xb77bb7dd in SDL_QuitSubSystem () from /usr/lib/libSDL-1.2.so.0 #5 0xb77bb84e in SDL_Quit () from /usr/lib/libSDL-1.2.so.0 #6 0xb77bc08f in ?? () from /usr/lib/libSDL-1.2.so.0 #7#8 0xb74ede80 in ?? () from /lib/i686/cmov/libc.so.6 #9 0xb74efc8c in malloc () from /lib/i686/cmov/libc.so.6 #10 0xb7764eec in ?? () from /usr/lib/libSDL_mixer-1.2.so.0 #11 0xb7765b98 in Mix_SetPanning () from /usr/lib/libSDL_mixer-1.2.so.0 #12 0x080828b1 in I_SDL_StartSound (id=-1218694088, channel=-1218699276, vol=120, sep=38) at i_sdlsound.c:671 #13 0x08073e9b in S_StartSound (origin_p=0xb624f990, sfx_id=22) at s_sound.c:656 #14 0x08061459 in T_MoveFloor (floor=0xb63271b8) at p_floor.c:220 #15 0x0806d0ea in P_RunThinkers () at p_tick.c:119 #16 0x0806d161 in P_Ticker () at p_tick.c:153 #17 0x08052fb6 in G_Ticker () at g_game.c:1151 #18 0x0804dbf4 in D_DoomLoop () at d_main.c:437 #19 0x0804ecc0 in D_DoomMain () at d_main.c:1506 #20 0x08054d1c in main (argc=5, argv=0xbfe55d04) at i_main.c:152
Most absurd CMOS battery ever?
I have a Performa 6400 PowerMac, which I bought off a friend several years ago. It runs Linux and I use it for portability testing (as it's a big endian machine). Considering its age, it's not surprising that the battery for the onboard clock ran out years ago. It's not a huge problem, but it's certainly annoying to have make complain about modification dates in the future, and be confronted with "filesystem has not been checked for 19370 days" on boot-up. So I decided to replace the battery.
This is the battery that is supposed to go in it. Almost £15 for a battery! This seemed far too expensive to me, so I set out to find an alternative. Unfortunately 4.5 volt batteries are rather uncommon. However, I managed to find this, which was listed as a "lantern battery" used for torches, bike lamps and doorbells. I suspect that there are probably 3 AA batteries inside:
- Greencell 312G 4.5V lantern battery alongside the original (depleted) CMOS battery.
I did a quick sanity check of the old battery to make sure I had the polarity right. The voltage is down to 1V:
All that needs to be done now is to connect the connector from the old battery to the new one. I've always been hopeless at soldering, but fortunately this is a job that is simple enough for sellotape.
This is the logic board from the Mac. I've maxed out the RAM and added PCI Ethernet and USB cards. The black square at the bottom left is where the battery is supposed to be attached (it has a velcro strip at the back).
Obviously this is far too small for the new battery, and the card slides into the back of the machine, mounted vertically. It has to be attached to the board somehow. Where can it go?
Fortunately, the engineers at Apple were apparently smart enough to anticipate this very problem, and designed a convenient space on the riser card to put the battery in.
Interview with me
fragorama.se is a new website about Classic Doom, and they have just published an interview with me!
Southampton Test Hustings
I just attended the Hustings for Southampton Test. These are my thoughts on some of the candidates.
Alan Whitehead (incumber Labour MP): Seemed rather nervous at the start but gained confidence later in the debate. To his credit he made some good points; I was impressed that he was bold enough to state that the law of the land should trump religious beliefs. However I also got the impression that he was less up to speed with other candidates on local issues and perhaps hadn't been paying proper attention to his constituency. Surprisingly enough he opposes Trident in favour of a cruise missile system like the LibDems advocate.
Jeremy Moulton (Conservatives): Made some good points and seemed a confident speaker. Some of the answers he gave seemed to have been slighty evasive/misleading when audience members responded to his answers. Attacked Alan Whitehead for using the communications allowance and supposedly putting the Labour party name on it (?)
Dave Callaghan (Liberal Democrats): Seemed the most honest of the lot. He highlighted some of the local issues that he's been campaigning for, like the closure of the Millbrook library, which he attacked Jeremy on (Jeremy is responsible for finances on the local council?).
Pearline Hingston (UKIP): A UKIP candidate who is an immigrant (how's that for a brain-breaker?). She came across as completely clueless and in general contributed very little of note. The one time she really attempted to express an opinion on something (cyclists riding on the pavement) she got smacked down by a member of the audience in response for not having a clue what she was talking about.
Chris Bluemel (Green): Surprisingly clueful and well-spoken. He spoke out in favour of nuclear disarmament and did it well, even though I don't agree with his views.
During the debates I sat next to an older gentleman who spent the time scribbling down notes on the back of an envelope. When he asked a question to the panel, he made some strange comments about Halliburton and BP. He seemed to think that there were plans to site nuclear submarines in Southampton docks, and was worried they might blow up and destroy the city. Very odd.
There was an obvious large Christian presence in the audience, and I suspect that siting the debate in a church probably didn't help. The candidates were asked at one point why they had all refused to sign a petition (I forget the name of it) declaring their support for Christian beliefs, although it was then revealed that none of the candidates had even heard of it. Several questions were asked about Christian rights that were obviously homophobic (eg. anti-gay marriage), though the people posing the questions tried to veil this by speaking in vague terms that made it less obvious what they were talking about.
Alan Whitehead (incumber Labour MP): Seemed rather nervous at the start but gained confidence later in the debate. To his credit he made some good points; I was impressed that he was bold enough to state that the law of the land should trump religious beliefs. However I also got the impression that he was less up to speed with other candidates on local issues and perhaps hadn't been paying proper attention to his constituency. Surprisingly enough he opposes Trident in favour of a cruise missile system like the LibDems advocate.
Jeremy Moulton (Conservatives): Made some good points and seemed a confident speaker. Some of the answers he gave seemed to have been slighty evasive/misleading when audience members responded to his answers. Attacked Alan Whitehead for using the communications allowance and supposedly putting the Labour party name on it (?)
Dave Callaghan (Liberal Democrats): Seemed the most honest of the lot. He highlighted some of the local issues that he's been campaigning for, like the closure of the Millbrook library, which he attacked Jeremy on (Jeremy is responsible for finances on the local council?).
Pearline Hingston (UKIP): A UKIP candidate who is an immigrant (how's that for a brain-breaker?). She came across as completely clueless and in general contributed very little of note. The one time she really attempted to express an opinion on something (cyclists riding on the pavement) she got smacked down by a member of the audience in response for not having a clue what she was talking about.
Chris Bluemel (Green): Surprisingly clueful and well-spoken. He spoke out in favour of nuclear disarmament and did it well, even though I don't agree with his views.
During the debates I sat next to an older gentleman who spent the time scribbling down notes on the back of an envelope. When he asked a question to the panel, he made some strange comments about Halliburton and BP. He seemed to think that there were plans to site nuclear submarines in Southampton docks, and was worried they might blow up and destroy the city. Very odd.
There was an obvious large Christian presence in the audience, and I suspect that siting the debate in a church probably didn't help. The candidates were asked at one point why they had all refused to sign a petition (I forget the name of it) declaring their support for Christian beliefs, although it was then revealed that none of the candidates had even heard of it. Several questions were asked about Christian rights that were obviously homophobic (eg. anti-gay marriage), though the people posing the questions tried to veil this by speaking in vague terms that made it less obvious what they were talking about.
Chocolate Doom on OS X, and GNUstep
Chocolate Doom runs on Mac OS X and has done for several years; however, until now, getting it running has been overly complicated and required compiling the source code from scratch. Obviously this isn't really appropriate for a Mac; it certainly doesn't fit in with the Apple way of doing things. I recently set about trying to improve the situation.
I first investigated how things are installed on OS X. Generally speaking there are two ways that things are installed; the installer (.pkg files), and Application Bundles, typically contained inside a .dmg archive. The installer simply installs a bunch of files to your machine, while Application Bundles are a lot more fluid; to install, you simply drag an icon into the Applications folder.
Application Bundles seem obviously preferable, but there's the problem of how one should be structured. Chocolate Doom needs a Doom IWAD file that contains the data used by the game, so it's not sufficient to simply package the normal binary as a bundle. Then there's the setup tool as well - should that be in a separate bundle? Finally, people often like to load PWAD files containing extra levels and mods. How do you do that with a bundle?
In the end, I decided to write a minimalist launcher program. Everything is in a single bundle file which, when launched opens a launcher window. The launcher allows the locations of the IWAD files to be configured and extra command line parameters entered. There's also a button to open the setup tool.
The launcher also sets up file associations when installed, so that it is possible to double-click a WAD file in the Finder, and an appropriate command line is constructed to load it. The interface is not as fully-featured as other "launcher" programs are, but it's simple and I think fits with the philosophy of the project.
The interesting part is how I developed the launcher. I only have occasional use of a Mac, so I developed it on GNUstep. This is an earlier version of the launcher interface while it was under development:
GNUstep provides an implementation of the same Objective C API that OS X's Cocoa provides, albeit with a rather crufty-looking NeXTStep appearance. It also has Gorm, which works in a very similar way to OS X's Interface Builder application. Using GNUstep, I was able to mock up a working program relatively easy. Constructing interfaces is very straightforward: the controls are simply dragged-and-dropped onto a window. I was able to get the underlying code into a working state before porting to OS X.
Porting to OS X had some hassles. Firstly, Gorm/GNUstep uses its own native format for interface files, which are different to the .nibs used on OS X. Recent versions of Gorm can save .nibs, but I found that the program crashed when I tried to do this. I eventually just reconstructed the whole interface from scratch in Interface Builder. GNUstep can use .nibs, so I just threw the older Gorm interface away.
The other main annoyance was that the format for property lists is different on OS X. It seems that GNUstep uses the older NeXT format, which Apple have since replaced with a newer XML-based format. Finally, icon files on OS X are in a proprietary .icns format, while GNUstep simply uses PNGs.
Both OS X and GNUstep try to force you to use their build tools (Xcode, ProjectCenter) which seem to generate a whole load of junk. I wrote a Makefile instead. There are some conditional parts to handle the two systems - OS X and GNUstep application bundles have different internal structures, for example. On OS X, the Makefile will do the complete process of compiler code, constructing the application bundle and generating a .dmg archive.
One thing I did find interesting is how OS X handles libraries. The full paths to any .dylib libraries (which are like Linux .so files or Windows DLLs) are stored inside a program when it is compiled. In my case, my application bundle needs to include the SDL libraries that Chocolate Doom depends upon. There's a convenient program called install_name_tool that can be used to change these paths after the program has been compiled. A special macro called @executable_path can be used to mean "the path where this binary is". I wrote a script to copy a program along with any libraries it depends on, changing its library search paths appropriately.
As one last finishing touch, there's a "Command Line..." menu item that calls into Terminal.app to open a command line window, if you're of the type who prefers the command line. The window sets up the PATH to point to the directory within the package that contains the chocolate-doom binary, and also sets up DOOMWADPATH to point to all the IWAD files you've configured within the GUI.
GNUstep was certainly incredibly useful in this activity; the ability to develop the program on my usual (Linux) laptop was very convenient. From a technical perspective, GNUstep seems to be a very impressive project. There is great usefulness in having a Linux implementation of the OPENSTEP (ie. Cocoa) API, which is what GNUstep is. However, the NeXT-style interface clashes horribly with almost any desktop environment that you might want to run under Linux (Gnome/KDE/etc), which is a huge turn-off.
The main problems are (1) the mini-window icons (which represent the running application) and (2) the menus, which appear in a separate window to the other application windows. I expect these are things that I could get used to if I was running a full GNUstep desktop where everything was like this; however, I'm not and it wouldn't really be practical for me to do so. It is possible to theme GNUstep to look nicer than its default "ugly grey square" appearance, but these problem remain.
GNUstep is a frustrating project in this respect. I can't help wonder if the full potential of the project is limited by the short-sightedness of its developers. It seems like they're too hung up on their goal of recreating NeXTstep, when I doubt there are many people who would even want to use such a system nowadays. This entry from the developer FAQ gives a good example of what I'm talking about:
"Without adding any [advantages]" - except, of course, the ability to give GNUstep applications an appearance that is consistent with 99% of Linux desktops! If it was possible to use GNUstep to make applications that looked like Gtk+ apps, I bet it would be a lot more attractive to developers. The practical advantages of such a decision are dismissed completely in the face of architectural/technical advantages that probably have little practical use.
I first investigated how things are installed on OS X. Generally speaking there are two ways that things are installed; the installer (.pkg files), and Application Bundles, typically contained inside a .dmg archive. The installer simply installs a bunch of files to your machine, while Application Bundles are a lot more fluid; to install, you simply drag an icon into the Applications folder.
Application Bundles seem obviously preferable, but there's the problem of how one should be structured. Chocolate Doom needs a Doom IWAD file that contains the data used by the game, so it's not sufficient to simply package the normal binary as a bundle. Then there's the setup tool as well - should that be in a separate bundle? Finally, people often like to load PWAD files containing extra levels and mods. How do you do that with a bundle?
In the end, I decided to write a minimalist launcher program. Everything is in a single bundle file which, when launched opens a launcher window. The launcher allows the locations of the IWAD files to be configured and extra command line parameters entered. There's also a button to open the setup tool.
The launcher also sets up file associations when installed, so that it is possible to double-click a WAD file in the Finder, and an appropriate command line is constructed to load it. The interface is not as fully-featured as other "launcher" programs are, but it's simple and I think fits with the philosophy of the project.
Developing with GNUstep
The interesting part is how I developed the launcher. I only have occasional use of a Mac, so I developed it on GNUstep. This is an earlier version of the launcher interface while it was under development:
GNUstep provides an implementation of the same Objective C API that OS X's Cocoa provides, albeit with a rather crufty-looking NeXTStep appearance. It also has Gorm, which works in a very similar way to OS X's Interface Builder application. Using GNUstep, I was able to mock up a working program relatively easy. Constructing interfaces is very straightforward: the controls are simply dragged-and-dropped onto a window. I was able to get the underlying code into a working state before porting to OS X.
Porting to OS X
Porting to OS X had some hassles. Firstly, Gorm/GNUstep uses its own native format for interface files, which are different to the .nibs used on OS X. Recent versions of Gorm can save .nibs, but I found that the program crashed when I tried to do this. I eventually just reconstructed the whole interface from scratch in Interface Builder. GNUstep can use .nibs, so I just threw the older Gorm interface away.
The other main annoyance was that the format for property lists is different on OS X. It seems that GNUstep uses the older NeXT format, which Apple have since replaced with a newer XML-based format. Finally, icon files on OS X are in a proprietary .icns format, while GNUstep simply uses PNGs.
Both OS X and GNUstep try to force you to use their build tools (Xcode, ProjectCenter) which seem to generate a whole load of junk. I wrote a Makefile instead. There are some conditional parts to handle the two systems - OS X and GNUstep application bundles have different internal structures, for example. On OS X, the Makefile will do the complete process of compiler code, constructing the application bundle and generating a .dmg archive.
One thing I did find interesting is how OS X handles libraries. The full paths to any .dylib libraries (which are like Linux .so files or Windows DLLs) are stored inside a program when it is compiled. In my case, my application bundle needs to include the SDL libraries that Chocolate Doom depends upon. There's a convenient program called install_name_tool that can be used to change these paths after the program has been compiled. A special macro called @executable_path can be used to mean "the path where this binary is". I wrote a script to copy a program along with any libraries it depends on, changing its library search paths appropriately.
As one last finishing touch, there's a "Command Line..." menu item that calls into Terminal.app to open a command line window, if you're of the type who prefers the command line. The window sets up the PATH to point to the directory within the package that contains the chocolate-doom binary, and also sets up DOOMWADPATH to point to all the IWAD files you've configured within the GUI.
Thoughts on GNUstep
GNUstep was certainly incredibly useful in this activity; the ability to develop the program on my usual (Linux) laptop was very convenient. From a technical perspective, GNUstep seems to be a very impressive project. There is great usefulness in having a Linux implementation of the OPENSTEP (ie. Cocoa) API, which is what GNUstep is. However, the NeXT-style interface clashes horribly with almost any desktop environment that you might want to run under Linux (Gnome/KDE/etc), which is a huge turn-off.
The main problems are (1) the mini-window icons (which represent the running application) and (2) the menus, which appear in a separate window to the other application windows. I expect these are things that I could get used to if I was running a full GNUstep desktop where everything was like this; however, I'm not and it wouldn't really be practical for me to do so. It is possible to theme GNUstep to look nicer than its default "ugly grey square" appearance, but these problem remain.
GNUstep is a frustrating project in this respect. I can't help wonder if the full potential of the project is limited by the short-sightedness of its developers. It seems like they're too hung up on their goal of recreating NeXTstep, when I doubt there are many people who would even want to use such a system nowadays. This entry from the developer FAQ gives a good example of what I'm talking about:
How about implementing parts of the Application Kit with GTK?
Yes and No - The GNUstep architecture provides a single, platform-independent, API for handling all aspects of GUI interaction (implemented in the gstep-gui library), with a backend architecture that permits you to have different display models (display postscript, X-windows, win32, berlin ...) while letting you use the same code for printing as for displaying. Use of GTK in the frontend gui library would remove some of those advantages without adding any.
"Without adding any [advantages]" - except, of course, the ability to give GNUstep applications an appearance that is consistent with 99% of Linux desktops! If it was possible to use GNUstep to make applications that looked like Gtk+ apps, I bet it would be a lot more attractive to developers. The practical advantages of such a decision are dismissed completely in the face of architectural/technical advantages that probably have little practical use.
How to make a program just run
Starting with Windows Vista, Windows limits the privileges that are given to normal users, running programs as the Administrator user only when necessary. To smooth over the fact that install programs for most software need to run as Administrator, it uses heuristics to detect whether a program is an installer. One of these is to look at the file name - if it contains "setup" in the name (among others), it is treated as an installer.
This is a problem if you develop a program that is not an installer but has "setup" in the name, because Windows treats it as though it is an installer and prompts you for administrator privileges.
The first problem is that it prompts the user for administrator privileges. This is part of the User Account Control system. Fortunately, there's a way around this - it's possible to embed a special "manifest" XML file inside the EXE that tells Windows that Administrator privileges aren't necessary.
Here's the magic manifest file to do this:
The important part here is the "requestedExecutionLevel" statement, that specifies to run the program as the invoker. I think the "uiAccess" element is necessary as well. I'm not entirely sure what this control does, and there are some people who say it should be set to true. However, it seems that if set to true, the executable has to be digitally signed with a certificate, which all looks like a massive hassle, so I've just left it turned off.
The "assemblyIdentity" tag here matches the executable name, but I'm not sure it's actually necessary. The version number is a dummy value.
Embedding it inside an executable is a matter of writing a resource file containing a statement to include the manifest file. Here's the magic statement for that:
The resource file is then compiled to a .o (using windres) and incorporated into the build.
So far, so good. If the above is done properly, Windows won't prompt to run the program with administrator privileges any more. However, that's not the end of the story. Windows still thinks the program is an installer, just an installer that doesn't need administrator privileges. The next problem is the "Program Compatibility Assistant".
If your program exits without writing any files to disk (in Chocolate Setup, it's possible to quit without saving configuration file changes, for example), the compatibility assistant appears. Because Windows thinks the program is an installer, and it hasn't written any files to disk, it assumes that something must have gone wrong with installation, and it might be a compatibility problem with a program designed for an older version of Windows. The assistant is supposed to help you resolve the problems you've encountered.
To work around this requires an addition to the manifest file to state that Vista (and Windows 7) are supported OSes; therefore, if no files are written, it's no problem. Here's the new version of the manifest:
This is a problem if you develop a program that is not an installer but has "setup" in the name, because Windows treats it as though it is an installer and prompts you for administrator privileges.
User Account Control
The first problem is that it prompts the user for administrator privileges. This is part of the User Account Control system. Fortunately, there's a way around this - it's possible to embed a special "manifest" XML file inside the EXE that tells Windows that Administrator privileges aren't necessary.
Here's the magic manifest file to do this:
<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
<assembly xmlns="urn:schemas-microsoft-com:asm.v1" manifestVersion="1.0">
<assemblyIdentity version="0.0.0.0" processorArchitecture="X86"
name="foo.exe" type="win32"/>
<trustInfo xmlns="urn:schemas-microsoft-com:asm.v3">
<security>
<requestedPrivileges>
<requestedExecutionLevel level="asInvoker" uiAccess="false" />
</requestedPrivileges>
</security>
</trustInfo>
</assembly>
The important part here is the "requestedExecutionLevel" statement, that specifies to run the program as the invoker. I think the "uiAccess" element is necessary as well. I'm not entirely sure what this control does, and there are some people who say it should be set to true. However, it seems that if set to true, the executable has to be digitally signed with a certificate, which all looks like a massive hassle, so I've just left it turned off.
The "assemblyIdentity" tag here matches the executable name, but I'm not sure it's actually necessary. The version number is a dummy value.
Embedding it inside an executable is a matter of writing a resource file containing a statement to include the manifest file. Here's the magic statement for that:
1 24 MOVEABLE PURE "setup-manifest.xml"The resource file is then compiled to a .o (using windres) and incorporated into the build.
Compatibility Assistant
So far, so good. If the above is done properly, Windows won't prompt to run the program with administrator privileges any more. However, that's not the end of the story. Windows still thinks the program is an installer, just an installer that doesn't need administrator privileges. The next problem is the "Program Compatibility Assistant".
If your program exits without writing any files to disk (in Chocolate Setup, it's possible to quit without saving configuration file changes, for example), the compatibility assistant appears. Because Windows thinks the program is an installer, and it hasn't written any files to disk, it assumes that something must have gone wrong with installation, and it might be a compatibility problem with a program designed for an older version of Windows. The assistant is supposed to help you resolve the problems you've encountered.
To work around this requires an addition to the manifest file to state that Vista (and Windows 7) are supported OSes; therefore, if no files are written, it's no problem. Here's the new version of the manifest:
<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
<assembly xmlns="urn:schemas-microsoft-com:asm.v1" manifestVersion="1.0">
<assemblyIdentity version="0.0.0.0" processorArchitecture="X86"
name="foo.exe" type="win32"/>
<trustInfo xmlns="urn:schemas-microsoft-com:asm.v3">
<security>
<requestedPrivileges>
<requestedExecutionLevel level="asInvoker" uiAccess="false" />
</requestedPrivileges>
</security>
</trustInfo>
<!-- Stop the Program Compatibility Assistant appearing: -->
<compatibility xmlns="urn:schemas-microsoft-com:compatibility.v1">
<application>
<supportedOS Id="{35138b9a-5d96-4fbd-8e2d-a2440225f93a}"/> <!-- 7 -->
<supportedOS Id="{e2011457-1546-43c5-a5fe-008deee3d3f0}"/> <!-- Vista -->
</application>
</compatibility>
</assembly>References
Creative defacement
Something funny I saw attached to a sign on the car park down the road from my flat:
