Iirc mingw links its libc statically, that's probably why. Fwiw I still remember that an empty VB6 project with one form compiles to 20kb (16 if you select "optimize for size"). It still needs its VM, which is a dll that was considerably larger. I think about 1mb, but it's been a while.
We shipped our VB6 software on a 1.44 MB floppy disk. Actually, we had to add some random stuff to the installer to make our software fit on two floppies instead of one, so that customers would think our application was more complex. Our CEO wanted to make sure that the end user had to eject disk 1 and insert disk 2 during the installation. Sometimes you want to go big. So size matters.
That's amazing! It reminds me of when a client once asked to make the red more red. I asked him to look at it again the next day and he said it was much better. I didn't have to do anything.
So, this uses the DLL version of the Microsoft CRT. Don’t know if that counts as cheating :) Why is a CRT needed here anyway, beyond very basic stuff like startup code and memcpy?.. Is MiniCRT[1] an option?
It possibly isn't, and you can typically do without by telling VC++ to build without linking to it - though the compiler has a bad habit of generating calls to memcpy, so you may have to provide your own implementation of that (and hope the compiler doesn't spot what you're doing and helpfully replace it with a call to memcpy). But aside from that, it's easy enough. You don't need the CRT startup code; you can use WriteConsole to print to stdout/stderr; you can use LocalAlloc/LocalFree to allocate memory; there's a few string routines in the Win32 DLLs that are a bit like the C equivalents, if nothing in the ordinary Win32 API does what you want.
I used to be into this stuff and created a number of useful <10 KB EXEs this way. I'm not sure it's really worth it though! I started out writing code for computers where you'd have like 16 KB free RAM, and this wasn't much, but for most purposes you didn't actually have to sweat every single last byte. So on the same basis, now that even my laptop (which is over 10 years old) has 16 GB RAM, I am not massively inclined to worry about anything less than 1 MB.
The downside to me knowing how small programs can be if I don't link the CRT is that I can't help but feel like what I've written is terribly bloated if I don't make an effort to avoid the CRT.
I blame people who constantly comment about how bloated software is now for making overly conscious of program size even when its down at a point where it doesn't matter :')
While MSDN is a bit impractical to browse (there's simply so much stuff in there) it's usually the best place to go to for documentation when doing Windows dev.
The magic numbers (hashes and UUIDs) sort of make sense because there's a slightly adversarial interaction between developers and Microsoft. If you just had a compatibleWindowsMin and compatibleWindowsMax field with version numbers, people would just go ahead and put "9999" in the max field, and then OS upgrades break applications again. By using UUIDs instead, an application developer can't intentionally or unintentionally declare compatibility with unreleased Windows versions.
As well as information about side by side assemblies (ie: library versions), a Windows manifest file has various settings and a declaration of compatible versions of Windows that affect Windows' handling of the app, such as whether it can handle paths over MAX_PATH, if it is hi-dpi aware, or if it knows about themed controls.
If an EXE doesn't have a manifest file, Windows assumes that it's ancient, so it falls back to conservative defaults like ye olde USER controls to try and avoid breaking it.
You can do it without a manifest too, it involves calling "CreateActCtxA" and "ActivateActCtx" with the appropriate Activation Context object. Manifest file is much easier.
Looks like you are linking to static libraries. You should link to DLL not to static libraries - this is will cut down on the application size dramatically.
That seems backwards. If you need to ship the DLLs with the program anyway (they are not part of the operating system, after all), they will contain their full functionality, each one potentially with its own C runtime, etc. If you statically compile everything into a single EXE, there will be only a single C runtime, all unused functions can be trivially removed, etc.
DLLs only reduce size if their code is meant to be shared between different programs.
Yes they are. Exercising the native Windows API is the entire point of this project, and the only artifact it builds is an executable.
edit: See the thread; I had the wrong end here. I haven't worked with Win32 or C in so long I'd forgotten what balls of fishhooks and hair they both tend to be.
The CRT is not part of the operating system, unless you count the UCRT on Windows 10 onwards (yes there is also a MSVCRT copy in the Windows folder that Microsoft strongly discourages you from using, so let's ignore that for now). So unless you link against the system-provided UCRT, you will have to either ship a dynamic or a static copy of the C runtime, and linking it statically will be smaller because that will only contain the few string and time functions used by the program, instead of the whole CRT.
The idea that something which ships with the OS, as a compatibility shim for legacy but still intentionally supported applications, can be called "not part of the OS," is religious. Strongly discouraged or otherwise, I believe it to be in use here.
Not sure. I haven't really done anything serious with MinGW, or Cygwin or even Windows really, in at least a decade now. But there's not really a lot going on in the build here, so I would imagine with your much more recent experience you'd be better able to interpret what's there.
edit: Wait, are you even discussing the old (okay, ancient) Win32 API? I'm confused, but as I said, it's been a very long time since I attended the space.
The important thing in the context of the original question is that mingw uses its own C runtime (MSVCRT is old and incomplete - you would not be able to use it as a drop-in replacement for a modern CRT). You can link the mingw CRT statically or dynamically, and OP suggested that linking it dynamically would reduce the size. But that is only true for the main executable. You would still have to ship the mingw CRT alongside, which in return would increase the distribution size of the program again.
I really didn't want to get into the specifics, because the situation is already complex enough without MSVCRT, but: Yes, it is. If you wanted to do a size-optimzied version of this particular program, it would be a good enough choice.
Regarding your edit: the CRT is responsible for headers such as time.h and string.h, which you can find being used in todo.h. Their implementations are not provided by the operating system but by the C runtime typically supplied by the compiler (mingw in this particular case). Other headers such as windows.h belong to the WinAPI, and result in functions being imported from OS libraries such as user32.dll. They are obviously not linked statically into the program.
Say what you like about the modern JS ecosystem, the practical requirement of a single exhaustive declaration of all external dependencies and their sources is not to be sneered at.
i get you, but this also introduces a whole other set of issues, like "dependency hell", where if you update one package you need to re-jig everything else. Things were just moving much slower back "in the days"
Dep version clashes were a problem until iirc some time about mid-late last decade, say ~2017.
These days it's perfectly solved at some cost in storage and memory, which in entire fairness are lately about as cheap as any other resource. There is also lately a movement in the space significantly away from "taking dependencies," a phrase I quote to highlight its pejorative connotation: presumptively undesirable, however necessary.
If a large language model is satisfactorily efficient for 2025, then to scruple over the disk space and RAM used in the current model of exhaustive dependency resolution employed by modern JS/TS package managers, seems impossible to regard as other than ideologically motivated.
In any case, I took some care earlier to avoid talking beyond the package.json dependency declaration model, in hopes of forestalling precisely this bikeshed. If you want to talk further about things that don't interest me in this context, please find someone else with whom to do so.
(I don't wish to seem rude in this. Only that I have had that discussion too many more times than I can count, and I have long since stopped expecting it to vary noticeably in a given iteration.)
If you mean maintaining separate versions of the same dependency, that doesn't "perfectly solve" the problem because you still can't take an object returned from one library (or something that depends on it and exposes that dependency) and pass it to another library without breakage, which can often be silent and subtle.
In Windows, the APIs provided by the OS are IMO better than these C standard headers.
WinAPI date/time is more precise, the kernel keeps int64 number of 100 nanosecond ticks. WinAPI supports arbitrary timezones, has functions to convert between timezones, has functions to deal with daylight saving.
About strings, WinAPI can convert strings between encodings, case-insensitive comparisons (including a version to use arbitrary locale), Unicode normalization, etc.
MinGW does not provide its own CRT. Historically, it used msvcrt.dll, which is the old and outdated CRT for VC++ 6.0 that was included in Windows a long time ago. These days it can also use uCRT, and that is indeed the new default:
Some people value being able to support older OSes without shipping the entire UCRT as an optional component. Granted, this will become less and less relevant in the future...
At this point, the only Windows version which doesn't ship uCRT that's not out of support is Windows Server 2012 R2 VL, and even that is only supported through the paid Extended Security Updates program which ends next year.
Not really, because this is Windows we are talking about.
A traditonal Windows application would be using the Windows APIs, and not the C standard library, e.g. FillMemory() instead of memset(), thus there is no DLLs to ship with the application.
This code is not from the Petzold books. It literally includes and uses string.h, stdlib.h and time.h. It is not using WinAPI equivalents of C functions.
No, linking to a static version of the CRT is a good thing, it cuts out the unused code. If you dynamic link to MSVCRxx/VCRUNTIME, you force the user to download that exact DLL from Microsoft. Dynamic linking to MSVCRT doesn't have that problem, but it's very hard to do in Visual Studio.
The only time you really can't static link to the CRT is LGPL compliance?
It looks looks mingw actually has several options to base its own CRT on (including MSVCRT, UCRT and various VS runtimes). Still, in the context of OP's original post, we are talking about the mingw DLLs that may or may not redirect most of their duties to Microsoft DLLs, depending on how mingw is configured.
Windows has been shipping an up-to-date, modern CRT in the box for a decade now (Win10+), and MinGW will even dynamically link to it by default. Even on out-of-support OSes like Vista and Win7, users who have all the security updates installed will have it. So you have to unwind all the way back to WinXP for a version of Windows that doesn't have uCRT out of the box.
There's absolutely no reason to statically link CRT in a Win32 app today. Especially not if your goal is to minimize .exe size.
A lot of it is due to the platform and executable format. Things can be much more lightweight when there's no information for stack traces, no dynamic linking infrastructure, no exception handling tables (necessary even in C in case exceptions traverse a c function,) etc.
A processor? Luuuxury! In my time we worked twenty-six hours a day, did all the calculations with pen and paper and would be thrilled to use an abacus!
Ah, back in my early existence, we didn't have time and all these superficial dimensions. Ontological creation out of nothing was all one needed, but it looks like it's all lost art now.
This reminds me of the days when all of a sudden win32 programming in assembly became hip enough, probably as a response to the increasing size of shareware downloads ('twas the dark time of MFC).
Combined with early Palm Pilot 68k programming, those were the last hurrahs of non-retrocomputing asm I can remember.
I'm spending some time this evening debugging a failure I have with an emulator I've written - it emulates a system running a Z80 processor with 64k of RAM.
Sometimes I too take a step back and look at the way things have changed. But then again we've made a lot of progress for the size-changes I guess.
I’m surprised it’s that big to be honest. I was expecting it to be smaller or half the size to be taken by some app icon. I remember writing this kind of stuff back in the days and it was smaller.
I tried to reproduce this binary to see what the 278 KB was being taken up by. The first obstacle that I ran into was that the build.bat file doesn't work if you have git configured to use core.autocrlf=false. Changing that to core.autocrlf=true and recloning was sufficient to get me building.
I'm using x86_64-15.1.0-release-win32-seh-msvcrt-rt_v12-rev0.7z from https://github.com/niXman/mingw-builds-binaries/releases/tag... as the toolchain. This produces a 102 KB .exe file. Right off the bat we are doing much better than the claimed 278 KB. Maybe the author is using a different toolchain or different settings? Exact steps to reproduce would be welcome.
We can improve this by passing some switches to GCC.
I also used mingw and yet I arrived at different results. Maybe it was a different version, or a different distro of MinGW, or a 32-bit vs. 64-bit issue, or I'm linking against a different CRT. Without details from OP, we can't really tell.
The v0.1 release from yesterday, at the time of posting, was 278 KB. The latest release, v0.3 from 9 hours ago, adds -Os -s and UPX to compress down to 27 KB.
I'd suggest `-Oz` instead, as it optimises for size above all else at the cost of performance, unlike `-Os` which is less aggressive (but likely produces similar code anyway). `-Oz` is somewhat new if I remember correctly, so it depends on the GCC version.
What's modern about it? Also you could have used C++ instead to remove some potential issues, and those global variables...
Use std::string and std::array or std::list, some anonymous namespaces, remove all the malloc, etc. Your code would be half the size and still compile to the same assembly language without the bugs.
It’s not modern but there’s some value to programming close to the metal so you understand what’s really going on. The problem domain is simple and easy to keep in your head. Good learning exercise. I wonder if there’s an assembly language version of something like this.
> Use std::string and std::array or std::list /.../
While nothing is modern about this approach, if we're going the WINAPI route, there's very little to be gained by using std::string instead of the LPWSTR that WINAPI offers (and plays nicely with). I would definitely avoid plain C strings (char[]) but rather use the wide version (which is what LPWSTR is under the hood). But for std::array or std::list, I don't see how the codebase would vastly benefit from them.
To note that even the latest editions of the famous Petzold book, before the C# edition, switched to using the common C subset from C++ and compiling with Visual C++ in C++ mode, instead of plain old C.
Already by Windows 95 timeframe, only C diehards would be writing Windows applications still in C instead of VB, Delphi and C++, as the more mainstream languages, there were obviously other ones to chose from as well.
As others have said, doing so in pure C and linking dynamically, you can easily remain under 20 KiB, at least on Linux, but Windows should be even simpler since it ships with much more out of the box as part of the OS.
In any event, I salute the effort! You can try the linking options I mentioned at the end of my article, it should help getting the size down.
Well, my somewhat extended TUI (ncurses) TODO program is 15K. Linux. Not statically linked though. I did not get around to build ncurses yet with musl.
Hello friends, I made this app just to try it out and have some fun, haha, but the comments are right, something like this could have been done more sensibly with C++ or other languages, ahaha.
This is exactly how I've learned to create my first Windows programs about 30 years ago, except that I'd use a C++ compiler.
I am not sure why but I believe writing C style code with a C++ compiler was how the windows API was documented to be used. I think Microsoft just went with the idea that C++ was an improved superset of C so should be used even for C-style code.
> I think Microsoft just went with the idea that C++ was an improved superset of C so should be used even for C-style code.
And as a consequence, for a long time their official C compiler was stuck on C89, while other platforms already had full C99 support and beyond. I believe their support for newer C standards has gotten better since then, but AFAIK they still don't have full C99 support.
C++ actually makes a lot of sense specifically for Win32 API because RAII takes care of releasing all the numerous handles at the right time in the right manner. Also, things like string operations are a pain in pure C (indeed, this app uses stuff like strcpy which is a recipe for buffer overruns etc).
I remember doing that for some custom installer I wrote. It felt like a good idea for 5mins until it got flagged by a bunch of anti virus software… had to sign the installer in the end and spent a lot of time reporting false positives.
Checks out: blurry fonts in scaled dpi, no Tab support, can't Ctrl-A select text in text fields and do all the other stuff that pre-modern frameworks offered you, errors on adding a row, ...
This code dynamically checks for and calls one of the following: user32:SetProcessDpiAwarenessContext, shcore:SetProcessDpiAwareness, then user32:SetProcessDPIAware. If the Windows version is extremely old and doesn't implement any of those (Windows XP or earlier), it won't call anything.
Was going to point out would have to be an embedded program that can be booted into without need of an OS; but then program would still rely on various external api/abi interfaces aka bios, drivers & protocols (usb, scsi, pci, ansi, etc). Given today's hardware, think there's not really a realizstic thing as a realistic reall world 'pure C' program -- way to much knowledge / effort even if exclude NDA's. VM environment for 'hardware' as software still has abi interface, so not 'pure' C at source code level.
The colons there don't represent C++. That's just a way of referring to a windows API function that exists in a specific DLL (in this case "user32"). Because the functions used here do not exist in older versions of Windows, the linked code dynamically loads user32.dll and tries to get the address of those functions so they can be called. That's why you need to know which Windows DLL they exist in.
It's a little more complicated if you are to be using themes, GDI and common controls.
Some time ago I put together this example: https://github.com/tringi/win32-dpi
The high DPI support in Windows went through quite an evolution since XP, but mostly to fix what app programmers messed up. You can have nice and crisp XP at 250% dpi if you do things right, e.g.: https://x.com/TheBobPony/status/1733196004881482191/photo/1
By "actual code" I meant the assembly that the application logic compiles down to, not the entire executable. But as far as the entire package goes, compiling it using clang with some flags I can get down to 19.5k without any effort. If I wanted to waste time on this, ripping out the CRT entirely and getting it to 16k would probably take less than an hour.
That's more than 10x bigger than I expected, given that all it seems to do is manipulate a list view. Something like this should be doable in under 10KB.
Instead of laboriously calling CreateWindow() for every control, traditionally we would lay out a dialog resource in a .rc file (Visual Studio still has the dialog editor to do it visually) and then use CreateDialog() instead of CreateWindow(). This will create all the controls for you. Add an application manifest and you can get modern UI styling and high-DPI support.
You also get automatic tabbing between controls, and a few other keyboard shortcuts this way. Note that resizing them still needs to be done manually if you want that, but that's usually easy and not more than a few hundred bytes of code.
However, this approach is easily translatable to a language that has decent FFI, and requires nothing else: no resource compiler and linker to make a resource DLL.
Resource files and their binary format are not a good API.
If you have those CreateWindow calls in a decently high level language, you can probably meta-program some resource-like DSL that fits right in the language.
You don't need a "resource DLL"; the compiled .rc file gets linked directly into the binary, and any Win32 C toolchain is capable of doing that, including MinGW.
As API goes, I don't see what's wrong with it (anymore so than Win32 in general). And you do get quite a lot for free, as GP mentioned. Hi-DPI, for example - .rc files use "dialog units" to measure all widgets, which, unlike raw pixel values you pass to CreateWindow, are DPI-independent.
I don't understand this scenario. If you're writing an .rc file, then by definition you are building your own app (or library, although having one that displays dialogs is rather unusual).
Are you talking about localizable text, by chance?
What if the app isn't in a language that links to an .exe, where we can include the .res file compiled from .rc?
The language may have an .exe run-time, but we don't get to link that.
In that case if we want to use a .rc file anyway, it seems we would want to create a resource DLL out of it and dynamically load it.
(I'm aware that one reason for resource DLL's is for localizable strings. That use makes sense even of programs which can statically link the GUI layout resources.)
Ah, I see what you mean now. But the context is a Win32 app that is very deliberately written in C, so my take was from that perspective - i.e. there's no good reason to avoid .rc files and instead handwrite CreateWindow, handle DPI etc in this case.
However, even in the case that you describe, you don't actually have to use a resource DLL. CreateDialogIndirect can be used to create a dialog from an array of DLGITEMTEMPLATE structs, each describing a single control. I'm not entirely sure, but I suspect that compiled dialogs are basically just a dump of that array (but if not you could always just create it manually).
That doesn't matter; you can still use Visual Studio to create the .rc file. This technique still works great for MinGW-based projects. The important thing is that Visual Studio has a .rc dialog editor.
There are several .rc editors for Windows outside of VS that incorporate visual dialog editing. Pelles C would be one example that still gets regular updates.
Thanks for the tip about Pelles C. I just fired it up and it's a little clunky (so is Visual Studio) but I was able to load and edit a .rc dialog resource. Visual Studio's old dialog editor is pretty crashy these days so it's great to have an alternative.
Great respect! I've tried many times, without final result. I'll try to use this for learning purposes!
Btw. I like how Inno Setup used some very old Delphi 2 compiler to create exe so small it would fit without breaking the zip compliancy. I read it somewhere 10+ years ago, so not sure if this is still the case, but still. And the initial dialog was done in pure winapi.h (of course it was winapi.pas which made everything more difficult for me to learn from)
I’m pretty sure an app as mediocre as this would take up less code in React, or even plain JavaScript. The UI is a single table and a few inputs and buttons, and its main way of communicating with the outside world is message boxes - trivial to do in a web browser.
Interesting! I was just looking at raylib for this kind of thing: super light weight, cross platform, reliable, ideally C-based method to get GUI.
Raylib and raygui is truly incredible from my point of view. I succeeded in getting the macOS and Windows builds going on a bunch of cute little novel (not stock standard in the repo) examples in a matter of hours with AI help. I'm inspired by all I can do with this.
For ages I felt "cut off" from the world of Desktop GUI because it was so verbose, and had high friction - need a bunch of tooling, set up, and so on. And then everything was fragile. I like to work quickly, with feedback, and PoCs and results. I think in raylib I have found a method that really achieves this. For instance, check out this tiny little "text_input.c"
#define RAYGUI_IMPLEMENTATION
#include <raylib.h>
#include "deps/raygui.h"
#define WINDOW_WIDTH 800
#define WINDOW_HEIGHT 600
#define MAX_INPUT_CHARS 32
int main(void) {
InitWindow(WINDOW_WIDTH, WINDOW_HEIGHT, "Text Input Demo");
SetTargetFPS(60);
// Load a larger font for better text appearance
Font font = LoadFontEx("resources/fonts/arial.ttf", 32, 0, 0);
if (font.texture.id == 0) {
font = GetFontDefault(); // Fallback to default font if loading fails
}
// Set the font for raygui controls
GuiSetFont(font);
// Customize raygui styles (using BGR order for hex values)
GuiSetStyle(BUTTON, TEXT_ALIGNMENT, TEXT_ALIGN_CENTER);
GuiSetStyle(BUTTON, BASE_COLOR_NORMAL, 0x50AF4CFF); // Green (B=80, G=175, R=76, A=255) ‚Üí R=76, G=175, B=80
GuiSetStyle(BUTTON, TEXT_COLOR_NORMAL, 0xFFFFFFFF); // White text
GuiSetStyle(BUTTON, BASE_COLOR_PRESSED, 0x6ABB66FF); // Lighter green
GuiSetStyle(BUTTON, TEXT_COLOR_PRESSED, 0xFFFFFFFF);
GuiSetStyle(BUTTON, BASE_COLOR_FOCUSED, 0x84C781FF); // Hover color
GuiSetStyle(BUTTON, TEXT_COLOR_FOCUSED, 0xFFFFFFFF);
GuiSetStyle(BUTTON, BORDER_WIDTH, 2);
GuiSetStyle(BUTTON, BORDER_COLOR_NORMAL, 0x327D2EFF); // Dark green border
// Adjust font size for raygui controls (optional, since font is already 32pt)
GuiSetStyle(DEFAULT, TEXT_SIZE, 20); // Slightly smaller for button and text box to fit better
GuiSetStyle(DEFAULT, TEXT_SPACING, 1);
char inputText[MAX_INPUT_CHARS + 1] = "\0"; // Buffer for text input
bool textBoxEditMode = false; // Tracks if the text box is being edited
bool messageSubmitted = false; // Tracks if a message has been submitted
float effectTimer = 0.0f; // Timer for the flash effect
const float effectDuration = 0.5f; // Flash duration in seconds
while (!WindowShouldClose()) {
// Update effect timer
if (effectTimer > 0) {
effectTimer -= GetFrameTime();
}
BeginDrawing();
// Set background color based on effect
if (effectTimer > 0) {
ClearBackground((Color){ 255, 255, 150, 255 }); // Yellow flash (RGB order)
} else {
ClearBackground(RAYWHITE);
}
// Center the text box and button
int textBoxWidth = 200;
int textBoxHeight = 40;
int buttonWidth = 120;
int buttonHeight = 40;
int textBoxX = (WINDOW_WIDTH - textBoxWidth) / 2;
int textBoxY = (WINDOW_HEIGHT - textBoxHeight) / 2 - 40;
int buttonX = (WINDOW_WIDTH - buttonWidth) / 2;
int buttonY = textBoxY + textBoxHeight + 10;
// Draw the text box
if (GuiTextBox((Rectangle){ (float)textBoxX, (float)textBoxY, textBoxWidth, textBoxHeight }, inputText, MAX_INPUT_CHARS, textBoxEditMode)) {
textBoxEditMode = !textB...
int PASCAL WinMain(HINSTANCE hInstance, HINSTANCE hPrevInstance, LPSTR lpszCmdLine, int nCmdShow);
If you know, you know.
I'm to understand that entire divisions of Microsoft itself no longer know how to code anywhere near this level, which is why many of their flagship applications (looking at you, Microsoft Teams (work or school)) are Electron monstrosities you can watch draw themselves like Windows 1.0 apps -- on modern multicore hardware.
I'm apparently so ancient I am just looking at that line of code thinking "isn't that the standard entrypoint for a Windows app?", lol; is there something else to know? ;P Oh! I guess, staring at it, isn't that wrong?... if you use LPCTSTR, one would expect that you are trying to be agnostic to UNICODE, but then you also have to use _tWinMain, no? (And, after considering that, I went and double-checked, and the C is also incorrect in this context.)
Funnily enough, the "ANSI" versions of Win32 functions are no longer considered a legacy thing, now that the OS offers UTF-8 as the supported encoding, and apps can explicitly request it (regardless of non-Unicode codepage setting) via manifest:
It's probably too late, but this qualifies for "Show HN" if you update the title to have the prefix "Show HN: ".[0]
I think the size of the source is actually more impressive than the size of the binary. I'm impressed that you can implement the whole thing in what looks like about 1 KLOC in just four .c files.
Fun project! I think the smallest I ever shrunk a win32 application was on the order of 2-4kb by writing in ASM. It was a great illustration of why 10x binary size is actually a great trade-off in terms of productivity.
249 comments
[ 2.9 ms ] story [ 77.9 ms ] threadhttps://forum.exetools.com/showthread.php?p=127834
[1] http://www.malsmith.net/minicrt/
I used to be into this stuff and created a number of useful <10 KB EXEs this way. I'm not sure it's really worth it though! I started out writing code for computers where you'd have like 16 KB free RAM, and this wasn't much, but for most purposes you didn't actually have to sweat every single last byte. So on the same basis, now that even my laptop (which is over 10 years old) has 16 GB RAM, I am not massively inclined to worry about anything less than 1 MB.
I blame people who constantly comment about how bloated software is now for making overly conscious of program size even when its down at a point where it doesn't matter :')
An example from one of my projects: https://pastebin.com/Jvjn5C6S
You need to reference it from your resource source like so: https://pastebin.com/8FUi4tMz
And then compile that into an object file with windres: x86_64-w64-mingw32-windres rsrc/metadata/windows.rc -o winbuild/windowsrc.o
And link it with your project like you would any other object file.
While MSDN is a bit impractical to browse (there's simply so much stuff in there) it's usually the best place to go to for documentation when doing Windows dev.
https://learn.microsoft.com/en-us/windows/win32/sbscs/applic...
If an EXE doesn't have a manifest file, Windows assumes that it's ancient, so it falls back to conservative defaults like ye olde USER controls to try and avoid breaking it.
DLLs only reduce size if their code is meant to be shared between different programs.
Yes they are. Exercising the native Windows API is the entire point of this project, and the only artifact it builds is an executable.
edit: See the thread; I had the wrong end here. I haven't worked with Win32 or C in so long I'd forgotten what balls of fishhooks and hair they both tend to be.
Not sure. I haven't really done anything serious with MinGW, or Cygwin or even Windows really, in at least a decade now. But there's not really a lot going on in the build here, so I would imagine with your much more recent experience you'd be better able to interpret what's there.
edit: Wait, are you even discussing the old (okay, ancient) Win32 API? I'm confused, but as I said, it's been a very long time since I attended the space.
I hope that clears my point up.
Say what you like about the modern JS ecosystem, the practical requirement of a single exhaustive declaration of all external dependencies and their sources is not to be sneered at.
These days it's perfectly solved at some cost in storage and memory, which in entire fairness are lately about as cheap as any other resource. There is also lately a movement in the space significantly away from "taking dependencies," a phrase I quote to highlight its pejorative connotation: presumptively undesirable, however necessary.
If a large language model is satisfactorily efficient for 2025, then to scruple over the disk space and RAM used in the current model of exhaustive dependency resolution employed by modern JS/TS package managers, seems impossible to regard as other than ideologically motivated.
In any case, I took some care earlier to avoid talking beyond the package.json dependency declaration model, in hopes of forestalling precisely this bikeshed. If you want to talk further about things that don't interest me in this context, please find someone else with whom to do so.
(I don't wish to seem rude in this. Only that I have had that discussion too many more times than I can count, and I have long since stopped expecting it to vary noticeably in a given iteration.)
If you mean maintaining separate versions of the same dependency, that doesn't "perfectly solve" the problem because you still can't take an object returned from one library (or something that depends on it and exposes that dependency) and pass it to another library without breakage, which can often be silent and subtle.
In Windows, the APIs provided by the OS are IMO better than these C standard headers.
WinAPI date/time is more precise, the kernel keeps int64 number of 100 nanosecond ticks. WinAPI supports arbitrary timezones, has functions to convert between timezones, has functions to deal with daylight saving.
About strings, WinAPI can convert strings between encodings, case-insensitive comparisons (including a version to use arbitrary locale), Unicode normalization, etc.
https://learn.microsoft.com/en-us/cpp/porting/upgrade-your-c...
MinGW does not provide its own CRT. Historically, it used msvcrt.dll, which is the old and outdated CRT for VC++ 6.0 that was included in Windows a long time ago. These days it can also use uCRT, and that is indeed the new default:
https://github.com/mingw-w64/mingw-w64/blob/master/mingw-w64...
edit: Saw that you just now edited your comment, glad we're on the same page now.
A traditonal Windows application would be using the Windows APIs, and not the C standard library, e.g. FillMemory() instead of memset(), thus there is no DLLs to ship with the application.
As can be seen on Petzold books examples.
The only time you really can't static link to the CRT is LGPL compliance?
There's absolutely no reason to statically link CRT in a Win32 app today. Especially not if your goal is to minimize .exe size.
15kB quickrun.exe :) C, pure Win32 API.. No hacks to shrink binary, Mingw32 compiler.
Its GUI app to quickly launch any application via alias.
You get that for free on Windows.
no exception handling tables (necessary even in C in case exceptions traverse a c function,
Not necessary if you're using pure C. SEH is rarely necessary either.
Combined with early Palm Pilot 68k programming, those were the last hurrahs of non-retrocomputing asm I can remember.
Sometimes I too take a step back and look at the way things have changed. But then again we've made a lot of progress for the size-changes I guess.
Is it due to MinGw maybe?
I'm using x86_64-15.1.0-release-win32-seh-msvcrt-rt_v12-rev0.7z from https://github.com/niXman/mingw-builds-binaries/releases/tag... as the toolchain. This produces a 102 KB .exe file. Right off the bat we are doing much better than the claimed 278 KB. Maybe the author is using a different toolchain or different settings? Exact steps to reproduce would be welcome.
We can improve this by passing some switches to GCC.
If all you are interested in is a small .exe size, there is plenty of room for improvement here.I wonder if they are compiling with debugging symbols? I don't know how much this would change things in vanilla C but that would be my first guess
Link-Time Optimisation with "-flto" might also help, depending on how the libraries were built.
Random example: https://news.ycombinator.com/item?id=39547329
What's modern about it? Also you could have used C++ instead to remove some potential issues, and those global variables...
Use std::string and std::array or std::list, some anonymous namespaces, remove all the malloc, etc. Your code would be half the size and still compile to the same assembly language without the bugs.
While nothing is modern about this approach, if we're going the WINAPI route, there's very little to be gained by using std::string instead of the LPWSTR that WINAPI offers (and plays nicely with). I would definitely avoid plain C strings (char[]) but rather use the wide version (which is what LPWSTR is under the hood). But for std::array or std::list, I don't see how the codebase would vastly benefit from them.
https://github.com/microsoft/wil
What about them? In a 500 loc app there is no practical difference and there's only ~20 of them with clear purpose.
> Use std::string <...> or std::list <...> remove all the malloc, etc
> still compile to the same assembly language without the bugs.
I see you have no clue what those things actually compile down to.
Already by Windows 95 timeframe, only C diehards would be writing Windows applications still in C instead of VB, Delphi and C++, as the more mainstream languages, there were obviously other ones to chose from as well.
Very nostalgic OP, warms my heart 10/10
As others have said, doing so in pure C and linking dynamically, you can easily remain under 20 KiB, at least on Linux, but Windows should be even simpler since it ships with much more out of the box as part of the OS.
In any event, I salute the effort! You can try the linking options I mentioned at the end of my article, it should help getting the size down.
I am not sure why but I believe writing C style code with a C++ compiler was how the windows API was documented to be used. I think Microsoft just went with the idea that C++ was an improved superset of C so should be used even for C-style code.
And as a consequence, for a long time their official C compiler was stuck on C89, while other platforms already had full C99 support and beyond. I believe their support for newer C standards has gotten better since then, but AFAIK they still don't have full C99 support.
Other language doesn't fundamentally change anything if you want to use win32 API, if anything it would make things more confusing.
People often fall prey to C++isms, and they would have made the whole thing an even more confusing mess (to people not familiar with win32 API).
This is a very cute thing to do and some familiarity with win32 APIs is a nice basic competency thing, regardless of what other people think.
WTL (https://en.wikipedia.org/wiki/Windows_Template_Library) is the oldschool way to do low-level Win32 coding in C++.
Checks out: blurry fonts in scaled dpi, no Tab support, can't Ctrl-A select text in text fields and do all the other stuff that pre-modern frameworks offered you, errors on adding a row, ...
> modern
In what way?
(A lot of what you mention is missing is trivial to add, especially tabbing between controls.)
This code dynamically checks for and calls one of the following: user32:SetProcessDpiAwarenessContext, shcore:SetProcessDpiAwareness, then user32:SetProcessDPIAware. If the Windows version is extremely old and doesn't implement any of those (Windows XP or earlier), it won't call anything.
The high DPI support in Windows went through quite an evolution since XP, but mostly to fix what app programmers messed up. You can have nice and crisp XP at 250% dpi if you do things right, e.g.: https://x.com/TheBobPony/status/1733196004881482191/photo/1
Can you share how you can compile to only 10kb?
Resource files and their binary format are not a good API.
If you have those CreateWindow calls in a decently high level language, you can probably meta-program some resource-like DSL that fits right in the language.
As API goes, I don't see what's wrong with it (anymore so than Win32 in general). And you do get quite a lot for free, as GP mentioned. Hi-DPI, for example - .rc files use "dialog units" to measure all widgets, which, unlike raw pixel values you pass to CreateWindow, are DPI-independent.
A program like this nicely translates to that situation.
Are you talking about localizable text, by chance?
The language may have an .exe run-time, but we don't get to link that.
In that case if we want to use a .rc file anyway, it seems we would want to create a resource DLL out of it and dynamically load it.
(I'm aware that one reason for resource DLL's is for localizable strings. That use makes sense even of programs which can statically link the GUI layout resources.)
However, even in the case that you describe, you don't actually have to use a resource DLL. CreateDialogIndirect can be used to create a dialog from an array of DLGITEMTEMPLATE structs, each describing a single control. I'm not entirely sure, but I suspect that compiled dialogs are basically just a dump of that array (but if not you could always just create it manually).
Sure there was K&R C, which each OS outside UNIX cherry picked what would be available.
Additionally outside UNIX clones, the tradition among vendors has been that the C compiler is responsible for the standard library, not the platform.
Thus the C library was provided by Borland, Watcom, Symantec, Microsoft, Green Hills, Zortech,....
Note it was the same on Mac OS, until MPW came to be.
As it was in IBM and Unisys, micros and mainframes.
VMS before OpenVMS.
And so on.
Since Windows 10, you have the Universal C Runtime as well.
They are using gcc.
Btw. I like how Inno Setup used some very old Delphi 2 compiler to create exe so small it would fit without breaking the zip compliancy. I read it somewhere 10+ years ago, so not sure if this is still the case, but still. And the initial dialog was done in pure winapi.h (of course it was winapi.pas which made everything more difficult for me to learn from)
Raylib and raygui is truly incredible from my point of view. I succeeded in getting the macOS and Windows builds going on a bunch of cute little novel (not stock standard in the repo) examples in a matter of hours with AI help. I'm inspired by all I can do with this.
For ages I felt "cut off" from the world of Desktop GUI because it was so verbose, and had high friction - need a bunch of tooling, set up, and so on. And then everything was fragile. I like to work quickly, with feedback, and PoCs and results. I think in raylib I have found a method that really achieves this. For instance, check out this tiny little "text_input.c"
I'm to understand that entire divisions of Microsoft itself no longer know how to code anywhere near this level, which is why many of their flagship applications (looking at you, Microsoft Teams (work or school)) are Electron monstrosities you can watch draw themselves like Windows 1.0 apps -- on modern multicore hardware.
EDIT: more correct function sig
https://learn.microsoft.com/en-us/windows/apps/design/global...
It's probably too late, but this qualifies for "Show HN" if you update the title to have the prefix "Show HN: ".[0]
I think the size of the source is actually more impressive than the size of the binary. I'm impressed that you can implement the whole thing in what looks like about 1 KLOC in just four .c files.
[0] https://news.ycombinator.com/showhn.html