My C23 Conventions

This is one of a three-part mini series about how I’m using modern C as of late 2026.

Before C23, I was taken in by How I Program C for years and years, and would even feel guilty for choosing C99 instead of -ansi -pedantic. But holy crap man I’m done being a doomsday prepper! These days I maintain a large-enough project with enough users where I could stand less foot-shooting. (I’ve also got bills to pay.)

So here’s three modern C conventions which started to rise naturally from adding -std=c23 to my zig build:

1. Unused? Maybe. (Yes)#

Our first convention is centered around [[nodiscard]]. If I draw an app window or panel I require the caller to interrogate the rect of the thing that was just rendered:

[[nodiscard]] rect_t draw_screen(/* ... */) {
    rect_t screen_r = { };
    // Draw stuff
    // ...
    return screen_r;
}

It’s a genuine business logic bug if the caller doesn’t capture the rectangle for later processing. Therefore [[nodiscard]] has been such a godsend for fixing bugs… Except there’s a choice few instances where I will draw something with no real use for the resulting rect. I’d rather ignore the return value:

enum screen_opt popup_options = /* ... */ ;
draw_screen(popup_options, /* ... */); // ERROR: damn you [[nodiscard]]

Clang will error out here. I could maybe capture the var and never use it but…

rect_t rect = draw_screen(/* ... */); // ERROR: unused variable (I have `-Werror` enabled)

Pre-C23 one might solve this through an UNUSED() macro:

#ifndef UNUSED
#define UNUSED(x) (void)(x)
#endif

rect_t rect = draw_screen(/* ... */);
UNUSED(rect); // Silences the compiler, but now I'm sad.

But now with C23 we can delete the macro in favor of the native [[maybe_unused]]:

[[maybe_unused]] rect_t rect = draw_screen(...); // OK

Phew I feel better. I’ve also noticed this is happening all over the place: lots of different routines returning different types that are marked with [[nodiscard]]. So for all the unused cases why not just go for:

[[maybe_unused]] auto _ = foo(); // A new pattern emerges

Bonus: For any macro fans out there a fellow indie dev (Zardium) offered an additional improvement. From our chat server:

Zardium: #define _ [[maybe_unused]] auto CONCAT_MACRO(zz,COUNTER)

Abner: ahh, unique varname. nice one

Zardium: you could take the auto out if you don’t want it to look so weird. then it acts exactly like the _ in other languages, where it can be redefined

Abner: Yep totally. Although here I was celebrating my removal of a macro… XD

2. More Perfect Unions#

I’ve been cheated and lied to: designated initializers DO NOT clear my struct to zero and I can’t even trust {0} anymore. (Talk about a betrayal.) The problem is I’ll often have data types like this contrived timestamp:

typedef struct {
    Type type;
    char buf[MAX];
    union {
        time_t time;
        float dt;
        date_t date;
    };
} stamp_t;

And if I partially initialize one:

stamp_t stamp = { .type = STAMP_TYPE_BLA_BLA };

While it does mean we successfully zeroed out buf (thanks to empty initialization), those union members are just a different beast. It also doesn’t matter if we go with stamp_t stamp = {0} instead. My dear reader: I’ve made the costly mistake of assuming date would clear to zero!

As I Spongebob-panicked in public, Lobsters member accelbread offered a sage revelation:

accelbread: It used to be that GCC would fully zero the union anyways, and clang would only zero the time_t, though around a year ago they swapped behaviors iirc. Either way, standard only requires the time_t be zeroed and not whole union. GCC has a flag to always zero whole union. Also C23 initialization always inits whole union

Wait WHAT? clang and gcc SWAPPED what now? I won’t investigate further. Everything about this is dangerous af unless we hug C23 tightly and never let go:

stamp_t stamp = { }; // Fully clears to zero (C23)
stamp.type = STAMP_TYPE_BLA_BLA;

Accelbread went on to say two more things:

1. “Note the C23 zero init whole thing is nothing in the brackets, not 0 in brackets; { 0 } does same thing as before C23”
2. “Also if you control union layout, putting largest member first is safest”

Yessir.

3. Enums In C Were Magic (And I Don’t Like It)#

If I want to customize some dialog box I’ll use bit flags:

enum {
    OPT_AUTOSCROLL = (1 << 0),
    OPT_NOCURSOR   = (1 << 1),
    OPT_CENTERED   = (1 << 2),
    OPT_NOCLOSE    = (1 << 3),
    // ...
};

[[nodiscard]] bool draw_warning(const char* msg, int opt);

Even if we disagree with this approach lots of relevant libraries or projects do it:

_ = draw_warning("you done messed up A-A-RON", OPT_CENTERED | OPT_NOCLOSE);

Of course, erroneous bit shifting will land anyone into trouble and yield comical numbers. Thankfully C enums are a plain int underneath so the compiler will warn us if we defined a bad constant inside the enum that won’t fit.

:)

I’m kidding. Sigh. Well, I’m growing fearful of sage revelations but this one comes from Modern C by Jens Gustedt:

Jens: If enumeration constants do not fit into signed int, if possible, the enumeration type is adjusted such that it can store all enumeration constants.

Noooooo!

Jens: The property that an an enumeration type is adjusted such that it fits all its constants could have surprising effects

sobs

Jens: So, it is preferable to specify the underlying integer type explicitly

THERE IS HOPE?

enum options : int {
    OPT_AUTOSCROLL = (1 << 0),
    OPT_NOCURSOR   = (1 << 1),
    OPT_CENTERED   = (1 << 2),
    OPT_NOCLOSE    = (1 << 3),
    // ...
};

It compiles! We shoot. We score. We stick with C23.

Epilogue#

In part two of our mini series I’ll discuss three more conventions that I’ve fleshed out thanks to modern C: starting with safer macro usage (forgive me Zig folks.)

P.S. If you want to shape this discussion maybe join the Handmade Cities server!

-Abner