On one particular project from 1995 where the hardware was very cost optimised, the C program compiled to 1800 bytes which meant we could save nearly a dollar by buying micro-controllers with 2KB flash rather than 4KB flash. We manufactured 20,000 units with this cheaper chip. 2 years down the line we needed a simple code change to increase the UART baud rate to the host, a change that should have resulted in the same sized binary, but instead increased it to 2300 bytes due to a newer C compiler. We ended up tweaking the assembly file and running an assembler, then praying there would be no more changes!
I have always over specified the micro-controllers a little from that point, and kept a copy of the original dev environment, luckily all my projects are now EOL as I am retired.
Correct, I have thousands of tank temperature controllers still out there, still working fine where the End Of Life was 3 years ago. EOL just means support for spares and software updates cannot be guaranteed past that point, and is mainly tied to the EOL of the specific micro-controller used.
One of my best rescue jobs involved doing this in 1999, yes that 1999. The client had shuttered their development department years before but was expecting to continue happily supporting and selling their simple enough alarm system products indefinitely. Testing revealed that come 2000 the alarms would just fire continually. Whoops. Fortunately there was one dev PC they'd decided to keep and not touch. Found the offending .c code and the corresponding offending machine code after some disassembly. A little bit of creative assembly language was required to squeeze an extra check in but really no big deal and the day was saved. I remember the client manager being ridiculously happy and grateful.
Whilst I disapprove of your use of the word "just", which I am strongly of the opinion should be banned in engineering circles...
I have done something similar, albeit in a different context, to fix the behaviour of a poorly performing SQL query embedded in a binary for which the source code was not easily available (as in: it turned out that the version in source control wasn't the version running in production and it would have been quite a lot of work to reverse engineer the production version and retrofit its changes back to the source - and, yes, this is as bad as you think it is).
When I initially suggesting monkey patching the binary there was all manner of screaming and objections from my colleagues but they were eventually forced to concede that it was the pragmatic and sensible thing to do.
Visual C++ 6 was the first C(++) compiler I used. I'm fairly certain it had auto completion (Intellisense).
Casey Muratori would point out the debugger ran faster on hardware from the era than modern versions run on today's hardware, though I don't have a link to the side–by–side video comparison.
Edit: Casey Muratori showing off the speed of visual studio 6 on a Pentium something after ranting about it: Jump to 36:08 in https://youtu.be/GC-0tCy4P1U — earlier section of the video is how it is today (or when the video was made)
The VS debugger got an order of magnitude slower in the transition from VS6 to Visual Studio .NET. It's been sped up a bit but is still nowhere near as fast as the VS6 debugger at responding to step commands, debug output, or conditional breakpoints. In VS.NET you could be waiting as long as a full second on a contemporary dev machine for the debugger to finish stepping forward one line.
Funny thing is that at the time, I was lamenting how much slower VC6 was than VC4. Macro playback, for instance, got much slower in VC6. It's all relative.
absolutely! When launching my editor today, it phones home, it checks location, it does everything that an editor SHOULDN'T DO. Not to mention the extensions...
Software today is a horrible bloated mess on top of horrible bloated messes.
It's really ironic that this appeared on the front page when it did, because I've spent the last couple days replacing the ZQuake sound system with FMOD and Atmoky TrueSpatial for HRTF and such. This was my first time ever working on a code base from 1996-2000. And in pure C no less. C feels so foreign to me since I'm so used to writing in C++ and Zig and such. But it was still really fun!
And I mean it doesn't seem super impressive, but it's something. Lol
you could have chosen a far worse codebase than Quake from the 90s. Quake was pretty clean in comparison. Sensible use of macros for doing things. A type system that made sense.
I mean... It's ugly from a modern standpoint. Zero encapsulation to speak of. Global state everywhere. But doing the modifications I did were pretty easy all things considered. So yeah, the code is clean, if "ugly" when viewed from a modern standpoint. It was definitely fun to do either way! Idk what other changes I'll make, we'll see. Especially since I don't know the architecture very well yet
It definitely was an amazing codebase for the time. You didn’t need to get hung up on architecture because it is very singular… it’s just a level, you, and the entities that were created when the level loaded.
There’s no pre-caching, no virtual textures, no shaders (there are materials for later quake 3), it’s just pure load -> set -> loop. The “client” renders, the “server” has the state.
This was my (hopeful) first thought on seeing this; his recent posts have been Quake-related. I do hope this is a harbinger of another installment. His others have been excellent.
> The first batches of Quake executables, quake.exe and vquake.exe were programmed on HP 712-60 running NeXT and cross-compiled with DJGPP running on a DEC Alpha server 2100A.
Is that accurate? I thought DJGPP only ran on and for PC compatible x86. ID had Alpha for things like running qbps and light and vis (these took for--ever to run, so the alpha SMP was really useful), but for building the actual DOS binaries, surely this was DJGPP on x86 PC?
Was DJGPP able to run on Alpha for cross compilation? I'm skeptical, but I could be wrong.
I thought the same thing. There wouldn't be a huge advantage to cross-compiling in this instance since the target platform can happily run the compiler?
Running your builds on a much larger, higher performance server — using a real, decent, stable multi-user OS with proper networking — is a huge advantage.
Yes, but the gains may be lost in the logistics of shipping the build binary back to the PC for actual execution.
An incremental build of C (not C++) code is pretty fast, and was pretty fast back then too.
In q1source.zip this article links to is only 198k lines spread across 384 files. The largest file is 3391 lines. Though the linked q1source.zip is QW and WinQuake, so not exactly the DJGPP build. (quote the README: "The original dos version of Quake should also be buildable from these
sources, but we didn't bother trying").
It's just not that big a codebase, even by 1990s standards. It was written by just a small team of amazing coders.
I mean correct me if you have actual data to prove me wrong, but my memory at the time is that build times were really not a problem. C is just really fast to build. Even back in, was it 1997, when the source code was found laying around on an ftp server or something: https://www.wired.com/1997/01/hackers-hack-crack-steal-quake...
"Shipping" wouldn't be a problem, they could just run it from a network drive. Their PCs were networked, they needed to test deathmatches after all ;)
And the compilation speed difference wouldn't be small. The HP workstations they were using were "entry level" systems with (at max spec) a 100MHz CPU. Their Alpha server had four CPUs running at probably 275MHz. I know which system I would choose for compiles.
> I mean correct me if you have actual data to prove me wrong, but my memory at the time is that build times were really not a problem.
I never had cause to build quake, but my Linux kernel builds took something like 3-4 hours on an i486. It was a bit better on the dual socket pentium I had at work, but it was still painfully slow.
I specifically remember setting up gcc cross toolchains to build Linux binaries on our big iron ultrasparc machines because the performance difference was so huge — more CPUs, much faster disks, and lots more RAM.
That gap disappeared pretty quickly as we headed into the 2000s, but in 1997 it was still very large.
> (Visual Studio 6) I never used it but it must have felt like a dream at the time.
I used it in the mid-90's and yes, it was eye opening. On the other hand, I was an Emacs user in uni, and by studying a bit the history of Emacs (especially Lucid Emacs) I came to understand that the concepts in Visual Studio were nothing new.
On the third hand, I hated customizing Emacs, which did not have "batteries included" for things like "jump to definition", not to mention a package manager. So the only times in the late-90s I got all the power of modern IDEs was when I was doing something that needed Windows and Visual Studio.
I've only played Quake I (and a modified version of it at that which had accessibility features). I did purchase quake II and III from Steam a few years ago, but it's much harder to play them because they have no accessibility to speak of (and I'm not entirely certain where to begin to try to replicate what was done with my version of Quake I). Quake in general has always been an insanely fun game for me, and I started playing it in like 2010. I still love playing it even now because it's got something to it that most other games I have just lack. Don't ask me to explain what it is because I can't really put it into words but...
there was another article where someone bootstrapped the very first version of gcc that had the i386 backend added to it, and it turns out there was a bug in the codegen. I'll try to find it...
EDIT: Found in, infact there was a HN discussion about an article referencing the original article:
Build environment archaeology like this matters more than people realize. Modern CI assumes containers solve reproducibility, but compiler version differences, libc variants, and even CPU instruction sets can silently change binary output. The detail about needing to reinstall Windows NT just to add a second CPU shows how tightly coupled OS and hardware were — there was no abstraction layer pretending otherwise. Exact toolchain reproduction isn't nostalgia; it's the only way to validate that a specific binary came from specific source.
> The detail about needing to reinstall Windows NT just to add a second CPU shows how tightly coupled OS and hardware were — there was no abstraction layer pretending otherwise.
In this case there was: the reason you need to reinstall to go from uniprocessor to SMP was because NT shipped with two HALs (Hardware Abstarction Layer): one supporting just a single processor, and one supporting more than one.
The SMP one had all the code for things like CPU synchronization and interrupt routing, while the UP one did not.
If they'd packed everything into one HAL, single-processor systems would have to take the performance hit of all the synchronization code even though it wasn't necessary. Memory usage would be higher too. I expect that you probably could run the SMP HAL on a UP system (unless Microsoft put extra code in to make it not let you), but you wouldn't really want to do that, as it would be slower and require more RAM.
So it wasn't that those abstraction layers didn't exist back then. It was that abstraction layers can be expensive. This is still true today, of course, but we have the cycles and memory to spare, more or less, which was very much not the case then.
> If they'd packed everything into one HAL, single-processor systems would have to take the performance hit of all the synchronization code even though it wasn't necessary. Memory usage would be higher too.
Linux also used to be like this, but these days has unified MP/UP kernels; on single-CPU systems (or if you give nosmp), the extra code is patched away at boot time. It wouldn't have been an unheard of technique at the time.
I actually would love this to be built in to a language/compiler. A lot of times when I’m building a single-threaded program but I’m using libraries written by other people. These libraries don’t know whether they are being incorporated into programs with single thread or not. So they either take the performance penalty of assuming multi-threaded (the approach by std::shared_ptr) or they give callers choice by making two implementations (Rust Arc and Rc). But the latter doesn’t actually work because this needs to be a global setting, not just a decision made at a local call site. It won’t work if such a library is a transitive dependency.
Glibc has a bunch of tests throughout the codebase where it checks if there have been any threads started besides the main one. I don’t really know how effective they are from a performance perspective. (In principle, turning fgetc into getc_unlocked, for instance, could be quite beneficial.) Microsoft used to have a single-threaded C runtime, but it was done away with some time ago, I’m guessing because they started putting things into the platform that would start and manage random threads outside the programmer’s control.
They could have shipped both HALs. Or made it easy to switch which one was in use without reinstalling.
CDs were around and hard drives weren’t that small at the time. (Or maybe the really early SMP versions predated widespread availability of CD-ROMs, but I remember dealing with this nonsense and reinstalling from an MSDN CD set.)
With NT4, I'm pretty sure both HALs were on the CD-ROM (unless you had an exotic system with a custom HAL, which came with its own install media). Keep in mind your use case is approximately nobody, you either had a SMP system or you didn't.
It was really not that rare to want to move a disk from one system to another. Except that there was an obnoxiously high chance that Windows would refuse to boot.
there is something to be said about old windows installation CDs being essentially modern-day equivalents of immutable docker layers - i don't think one could say that about modern windows, but then i'm not super clued in into ms stuff.
I have always over specified the micro-controllers a little from that point, and kept a copy of the original dev environment, luckily all my projects are now EOL as I am retired.
I doubt that everything you ever worked on is end-of-life. Some of it is still out there...
I have done something similar, albeit in a different context, to fix the behaviour of a poorly performing SQL query embedded in a binary for which the source code was not easily available (as in: it turned out that the version in source control wasn't the version running in production and it would have been quite a lot of work to reverse engineer the production version and retrofit its changes back to the source - and, yes, this is as bad as you think it is).
When I initially suggesting monkey patching the binary there was all manner of screaming and objections from my colleagues but they were eventually forced to concede that it was the pragmatic and sensible thing to do.
Casey Muratori would point out the debugger ran faster on hardware from the era than modern versions run on today's hardware, though I don't have a link to the side–by–side video comparison.
Edit: Casey Muratori showing off the speed of visual studio 6 on a Pentium something after ranting about it: Jump to 36:08 in https://youtu.be/GC-0tCy4P1U — earlier section of the video is how it is today (or when the video was made)
Funny thing is that at the time, I was lamenting how much slower VC6 was than VC4. Macro playback, for instance, got much slower in VC6. It's all relative.
Software today is a horrible bloated mess on top of horrible bloated messes.
And I mean it doesn't seem super impressive, but it's something. Lol
Descent on the other hand...
It definitely was an amazing codebase for the time. You didn’t need to get hung up on architecture because it is very singular… it’s just a level, you, and the entities that were created when the level loaded.
There’s no pre-caching, no virtual textures, no shaders (there are materials for later quake 3), it’s just pure load -> set -> loop. The “client” renders, the “server” has the state.
Is that accurate? I thought DJGPP only ran on and for PC compatible x86. ID had Alpha for things like running qbps and light and vis (these took for--ever to run, so the alpha SMP was really useful), but for building the actual DOS binaries, surely this was DJGPP on x86 PC?
Was DJGPP able to run on Alpha for cross compilation? I'm skeptical, but I could be wrong.
Edit: Actually it looks like you could. But did they? https://www.delorie.com/djgpp/v2faq/faq22_9.html
An incremental build of C (not C++) code is pretty fast, and was pretty fast back then too.
In q1source.zip this article links to is only 198k lines spread across 384 files. The largest file is 3391 lines. Though the linked q1source.zip is QW and WinQuake, so not exactly the DJGPP build. (quote the README: "The original dos version of Quake should also be buildable from these sources, but we didn't bother trying").
It's just not that big a codebase, even by 1990s standards. It was written by just a small team of amazing coders.
I mean correct me if you have actual data to prove me wrong, but my memory at the time is that build times were really not a problem. C is just really fast to build. Even back in, was it 1997, when the source code was found laying around on an ftp server or something: https://www.wired.com/1997/01/hackers-hack-crack-steal-quake...
And the compilation speed difference wouldn't be small. The HP workstations they were using were "entry level" systems with (at max spec) a 100MHz CPU. Their Alpha server had four CPUs running at probably 275MHz. I know which system I would choose for compiles.
I never had cause to build quake, but my Linux kernel builds took something like 3-4 hours on an i486. It was a bit better on the dual socket pentium I had at work, but it was still painfully slow.
I specifically remember setting up gcc cross toolchains to build Linux binaries on our big iron ultrasparc machines because the performance difference was so huge — more CPUs, much faster disks, and lots more RAM.
That gap disappeared pretty quickly as we headed into the 2000s, but in 1997 it was still very large.
I used it in the mid-90's and yes, it was eye opening. On the other hand, I was an Emacs user in uni, and by studying a bit the history of Emacs (especially Lucid Emacs) I came to understand that the concepts in Visual Studio were nothing new.
On the third hand, I hated customizing Emacs, which did not have "batteries included" for things like "jump to definition", not to mention a package manager. So the only times in the late-90s I got all the power of modern IDEs was when I was doing something that needed Windows and Visual Studio.
Gonna warm that up when the kids get a bit older and we start doing LAN parties.
That and Quake World Team Fortress.
I bet there are still servers out there, at that.
https://q2online.net/action
https://store.steampowered.com/app/1978800/AQtion/
there was another article where someone bootstrapped the very first version of gcc that had the i386 backend added to it, and it turns out there was a bug in the codegen. I'll try to find it...
EDIT: Found in, infact there was a HN discussion about an article referencing the original article:
https://miyuki.github.io/2017/10/04/gcc-archaeology-1.html
https://news.ycombinator.com/item?id=39901290
In this case there was: the reason you need to reinstall to go from uniprocessor to SMP was because NT shipped with two HALs (Hardware Abstarction Layer): one supporting just a single processor, and one supporting more than one.
The SMP one had all the code for things like CPU synchronization and interrupt routing, while the UP one did not.
If they'd packed everything into one HAL, single-processor systems would have to take the performance hit of all the synchronization code even though it wasn't necessary. Memory usage would be higher too. I expect that you probably could run the SMP HAL on a UP system (unless Microsoft put extra code in to make it not let you), but you wouldn't really want to do that, as it would be slower and require more RAM.
So it wasn't that those abstraction layers didn't exist back then. It was that abstraction layers can be expensive. This is still true today, of course, but we have the cycles and memory to spare, more or less, which was very much not the case then.
Linux also used to be like this, but these days has unified MP/UP kernels; on single-CPU systems (or if you give nosmp), the extra code is patched away at boot time. It wouldn't have been an unheard of technique at the time.
CDs were around and hard drives weren’t that small at the time. (Or maybe the really early SMP versions predated widespread availability of CD-ROMs, but I remember dealing with this nonsense and reinstalling from an MSDN CD set.)