Hacker News Reader

Comment by buserror | Hacker News Reader

Case ranges are in gcc for a million years. And also nested functions. I used nested functions a lot but, it was never ported to clang which makes it a bit of a problem for portability, also, nowadays the linker complains about having to make the stack executable, which is another problem; so I had to stop using them.

I also don't understand why the "C standard" hasn't evolved further -- there was a discussion lately about the committee being out of touch. Some of these extensions makes life so much easier. Instead we get stupid stuff nobody need, or have been in a header for 200 years (ie, <pthread.h>, whoohoo, I'm delighted this standard header is now a... standard!)

For nested/sub function, it makes so much sense it's not even funny. Having micro-callbacks right there before the call to qsort() (or any other) is SO much better than having to farm out a context, yet-another-static-function etc...

View on HN

> (ie, <pthread.h>, whoohoo, I'm delighted this standard header is now a... standard!

<pthreads.h> is POSIX, it has never been - and still don't is! - part of the C standard. In particular, MSVC only supports it through third party libraries.

However, C11 added cross-platform thread support with <threads.h>: https://en.cppreference.com/w/c/thread. I wouldn't call that "stupid stuff nobody needs".

> and still don't is!

"isn't," not "don't is"

You're right, though

d'oh

> I'm delighted this standard header is now a... standard!

The POSIX standard isn't the C standard though (for better or worse).

Especially on Windows with MSVC "obvious" things like this are traditionally a royal PITA (which also means that a lot of C code coming from the UNIX world simply doesn't build on MSVC even though MSVC is now a decent standard-C compiler again).

But yeah, the C standard is just the minimal common subset of what C compilers offer. The good stuff is all in Clang and GCC extensions.

> But yeah, the C standard is just the minimal common subset of what C compilers offer.

Why is this? Most other standards — for languages or otherwise — seem to run ahead of implementation, as meetings of major players agreeing on what they'll all implement next, with the standard constraining the implementations on how to do that.

But the C standard is seemingly instead just an external, descriptivist report on "what you can get away with writing in C while keeping it completely portable."

Don't know, most likely reason is probably that the C standard was always an afterthought and 'reactive' in the C world (e.g. the language already existed for nearly two decades before the first standard was released).

TBH I think the general idea of "let's only standardise what has already been implemented in real world extensions and proven its worth" isn't too bad, especially when looking at the mess that the C++ standard committee made of C++ in the last decade.

But even with this pragmatic approach I imagine it's a highly political balancing act to get multiple powerful compiler vendors to actually agree on the same thing.

See for instance the C23 auto proposal which contains a special exception for Clang's behaviour which clashes with GCC's behaviour:

https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3007.htm

Problem is, that approach forces the committee to roll in stuff from the C++ spec -- stuff that would have been shot in flame if you had hinted about that sort of things as a proposal C extension..

For example the [[keywords]] that had was rolled in because it was already out there -- imagine proposing that out of the blue and see a people self-propel into orbit...

On other hand the Case A ... B: syntax that has been out there for 30+ years, well there's a proposal for it but it won't be THAT because well it hasn't been invented here really, we'll make up a different syntax for it instead...

Well there is a reason for that. The A ... B syntax requires that you use spaces between A, the three dots and B, else it will be lexed as a single number. C is supposed to ignore whitespace, it wouldn't make sense to add a feature that requires whitespace.

It doesn't requires a space, you could modify the lexer to recognize '..' as a prefix and solve that problem, it's just a choice they made to refuse that it could be implemented by just having one character forward looking lexer.

It is not because it is implemented like that in gcc (requiring a space, possibly so they didnt have to modify the lexer) that it ought to be specced like that.

All C lexers are greedy, i.e. they extend the current token as far as they can. So "0..9" is split as a token "0." (numeric literal) followed by token "." (operator) followed by token "9" (numeric literal).

To do what you suggest, the lexers would need a major redesign, not just a small change.

C still supports EBCDIC, and other potential non-ASCII character sets one might invent in the future.

It promises e.g. that 0..9 is contiguous, but e.g. "case 'A' ... 'Z'" would behave differently.

That has clearly prevented dozens of other languages to support that feature... because no parser would be able to tell you that range isn't valid at compile time?

In EBCDIC, ‘Z’ > ‘A’, so the range 'A' ... 'Z' is valid, but may not do what the programmer intended.

Most other languages never supported EBCDIC or dropped it long ago.

(Also: 'A' ... 'Z' may not do what the programmer intended in most other languages, too. Even ignoring Unicode, quite a few encodings have accented letters that are outside of that range)

One problem is that the important compilers are today mostly controlled by C++ people, and smaller C compilers and C users are underrepresented. Every change in C that is different to what C++ has already is an uphill battle.

> the C23 auto proposal which contains a special exception for Clang's behaviour which clashes with GCC's behaviour

Can you explain what that is? There’s supreme awkwardness around allowing `auto int i;` at block scope to work as before (with `auto` being ignored), but I’ve stared at that proposal a fair bit and haven’t noticed such an exception there.

The details are buried in the proposal and make for an interesting read, but the gist is that Clang's C frontend has a "C++ style" auto keyword where a pointer can be declared both with "auto" and "auto*" (and this was already baked into the existing __auto_type extension before C23 came along).

For instance this is valid Clang C23 code (and entirely C23 standard-compliant):

    int i = 0;
    auto p0 = &i;
    auto* p1 = &i;

...both p0 and p1 have the same type of int*.

In GCC this code doesn't compile because GCC already had a simpler "C style" __auto_type extension which doesn't allow the form "auto*", but this behaviour is also entirely valid C23 ¯\_(ツ)_/¯.

(personally I prefer GCC's version as it makes more sense for C)

This will cause some head-scratching when code that builds on Clang doesn't build on GCC, especially since Clang usually tries to emulate GCC behaviour. But OTH I think/hope that C's auto will only be used when absolutely necessary in "type-generic" macros, and no "almost always auto" nonsense which was fashionable in the C++ world for a while.

Im in the wg14 and this is not true. The later versions of the C standard are not dependable and diverges quite a bit from implementations. There are lots of things in the C standard text that you should not depend on. There are in fact extensions that are more commonly implemented than standard features, and some standard features that hav no known implementations.

I hope to work on a guide for developers who want a “dependable-C” where they can guarantee it will compile and run correctly in as many implementations as possible.

I agree. Nested functions are very useful often lead to substantially better code. With GCC 14 you can also create heap trampolines, so an executable stack is not required anymore when you take the address of a nested function (if you don't take the address, it isn't required anyway).

“ Nested functions can even goto back into their parent function, allowing for nonlocal exits to break out of nested functions like Smalltalk blocks, allowing control flow-like functions to be built using them.”

Wild!

Makes one wonder how something could have evolved if it took C and made it more “functional” and/or smalltalk-ish, without going the dynamic way of ObjectiveC.

I don’t exactly love C++ but I find it very useful. Maybe I could have loved that other never conceived C offspring.

Note that this functionality is already in C — in the form of the setjmp/longjmp functions, which indeed allow you to have a parent function that contains "global" error-handling, and then a hierarchy of child functions designed to only be executed from that parent, that can "jump out" to the parent's error handler. (But where "parent" and "child" here are just a conceptual relationship the programmer is keeping track of, not anything the compiler knows about.)

The difference is that High C gives you lexically-nested function definitions, thus making labels into things like local variables that have scope + shadowing. And so it becomes sensible — necessary, really, to have a coherent semantics — to extend the (runtime, dynamic) longjmp, into a compile-time-targeted, lexical, syntax-sugared version that jumps to in-scope labels (incl. ones defined outside the current lexically-nested function definition.)

If you didn't, then you'd have to decide on what a `goto` to an in-scope but out-of-function label would do instead — compiler error, maybe? — and whatever you choose, it would probably be just as hard to implement as just making it work. (Most of the effort is in extending the CFG to allow some analysis pass to notice that this is happening, after all.)

Reply on news.ycombinator.com