Hacker News Reader

Comment by EdSchouten | Hacker News Reader

> What is much less prevalent is a demand from average C++ users for memory safety features; they’re much more concerned about compilation speed. When most C++ developers haven’t adopted tools like Coverity and C++ core guidelines checkers, it is hard to claim that memory safety features substantially improve their lives at least from their point of view.

I don’t really agree with this. There’s also a group of developers that want more safety, but don’t feel like setting up (or in the case of Coverity purchasing) additional tooling. Some people just want a decent out of the box experience.

View on HN

I'll make a similar, but fundamentally different, claim to what you're quoting: There isn't much demand for half assed memory safety features. And C++ memory safety features are - nearly by definition - half assed. They're opt-in (meaning all the third-party and system code you link against doesn't have coverage), slow (meaning you can't use it in production), and tend to catch the trivial bugs rather than the hard-to-find ones that keep you up at night.

Reminds me of "type-safety" features in duck type languages.

That sentence stood out to me as well. I don't think C++ developers is necessarily the best people to ask when discussing whether memory safety is important and urgent to work on or not. A lot of developers are relatively shielded from the consequences of bugs and security vulnerabilities, either by their programs not being exposed to the wild in any major sense or by bureaucracy being in between them and any consequences.

At some point, the deficiencies of a language becomes a concern for its end-users rather than its developers.

> the deficiencies of a language becomes a concern for its end-users rather than its developers.

I think that this level of professional malpractice is something that can't be solved by language choice.

But the level of damage can be reduced by language choice

Maybe, but surely the better answer is for devs to do better.

That actually seems unrealistic to me. Of course, people will always become better at doing the things they do, and we can always try harder, but many of these issues take time to consider and find during code reviews, and a few just slip by. That's why tools like Coverity exist, but you have to spend time and money to set those up, meaning it's only done when absolutely necessary (or one dedicated person is really pushing for it). Choosing a different language is basically free in the beginning, and it will impact which sort of bugs will be caught by default and which won't. C++ is also massively hurt in this regard by not having a package manager. JS is a very risky language by default, but a few tools just with their default settings will already help massively. I guess Nix could be considered the missing package manager for C and C++, but it's still niche and definitely not a "default" like pip, npm or cargo.

Yeah that stood out to me too. I guess the problem is that only a small percentage of developers both understand the need for memory safety and are vocal about it. That is very different from the number of users who would benefit from better memory safety.

When someone encounters a crash, their actionable item isn’t to go ask the C++ committee for better safety. That’s an unassailable wall for most developers, assuming they even understand what’s happening and that there are better options out there.

C++ Developers aren't complaining for two reasons:

1. Most of those who care about memory safety have already left the building.

2. Those who are still in the building and care don't speak up because their colleagues don't value it, and many of their colleagues will view them as having lower competence.

Yes

It is not about what we want. It is what we need

Not real.y "we". I have lost any interest in C++ now. I loved it, in the day. Time has moved on

I think it's less about not wanting memory safety and more that compilation speeds are so time-wastingly abysmal that it's a few orders of magnitude more important to address.

The whole deal with interpreted languages was to avoid sitting there for a few minutes every time you make a god damn one letter change, with the unfortunate but usually acceptable trade-off of some execution speed.

And the additional trade-off that some bugs are only noticed at runtime when that particular line is executed while it could have been noticed by the compiler of a strongly typed language. Pytype helps but at this point you have a static analyzer that potentially runs as slow as a compiler without the additional performance benefit.

Idk as long as you can develop at speed I don't see why a static analyser that's a few times slower than compiling couldn't run on the latest commit overnight? More as a sonarqube type thing I suppose.

There’s no good reason for type checking to be super slow. I’m no fan of Go, but the language compiles insanely fast while being fully statically typed.

As I understand it, C++’s slow compilation comes from the fact that it usually parses all of your header files n times instead of once. This isn’t a problem with static typing. It’s a problem with C++, and to a lesser extent C.

> As I understand it, C++’s slow compilation comes from the fact that it usually parses all of your header files n times instead of once.

That's one of the things that can slow compilation down but it's definitely not the only one. It helps that precompiled headers (and maybe modules?) can go a long way towards reducing and possibly eliminating these costs as well.

I think some (most?) of the larger remaining costs revolve around template instantiation, especially it impacts link times as well due to the fact that the linker needs to do extra work to eliminate redundant instantiations.

> due to the fact that the linker needs to do extra work to eliminate redundant instantiations.

Yeah, I see this as another consequence of C++'s poor compilation model:

- Compilation is slow because a template class in your header file gets compiled N times (maybe with precompiled headers). The compiler produces N object files filled with redundant code.

- Then the linker is slow because it needs to parse all those object files, and filter out all the redundant code that you just wasted time generating.

Its a bad design.

I'm not sure I'd call the design "bad". At the very least it's a product of the design constraints, and I'm not sure there's an obviously better implementation without sacrificing something else. I think separate compilation and monomorphization are the biggest contributors, but I wouldn't be surprised if there was something I was forgetting.

Somewhat related, there was some work in Rust about sharing monomorphized generics across crates, but it appears it was not a universal win at the time[0]. I'm not sure if anything has changed since that point, unfortunately, or if something similar could be applied to C++ somehow.

[0]: https://github.com/rust-lang/rust/issues/47317#issuecomment-...

> I'm not sure I'd call the design "bad". At the very least it's a product of the design constraints, and I'm not sure there's an obviously better implementation without sacrificing something else.

It was a product of the design constraints in the 70s when memory was expensive, and compilers couldn't store a whole program in memory during compilation.

The problem C++ has now is that the preprocessor operates on the raw text of a header file (which is a relic from C). This means the same header file can generate totally different source code each time its included in your program. C++ can't change that behaviour without breaking backwards compatibility. So headers get parsed over and over again "just in case" - wasting time producing excess code that just gets stripped back out again by the linker.

The way C++ works doesn't make any sense now that memory is so much cheaper. Go, C#, java, rust, zig - basically every compiled language younger than C++ compiles faster than C++ because these languages don't contain C++'s design mistake.

Rust doesn't share monomorphized generics across crates, but at least each crate is compiled as a single compilation unit.

This was already the case with languages like Modula-2 and Object Pascal in the 1980's, C++ works that way because it was designed to be a drop-in in UNIX/C without additional requirements.

> As I understand it, C++’s slow compilation comes from the fact that it usually parses all of your header files n times instead of once.

Sort of. The primary issues are:

1) The C/C++ grammar is garbage.

Note that every single modern language has grammatical constructs so that you can figure out what is "type" and what is "name" without parsing the universe. "typedef" makes that damn near impossible in C without parsing the universe, and C++ takes that to a whole new level of special.

2) C++ monomorphization

You basically compile up your tempate for the universe of every type that works, and then you optimize down to the one you actually use. This means that you can wind up with M*N*O*P versions of a function of which you use only 1. That's a lot of extra work that simply gets thrown away.

The monomorphization seems to be the biggest compile time problem. It's why Rust struggles with compile times while something like Zig blazes through things--both of those have modern grammars that don't suck.

1. No, the grammar is not the issue per se. As you say, C has the same problem, and C code invariably compiles dozens of times faster than C++, and both Zig and Rust have modern grammars, but Zig compiles about as quickly as C and Rust is only somewhat faster than C++ (depending on features used).

2. This is incorrect. What's happening is that each template instantiation for a new set of template arguments requires reprocessing the template to check types and generate code, and also that is done per translation unit instead of per project. Each distinct template instantiation increases the compilation time a bit, much more than it takes to parse the use itself. That's why it's easy to have a small C++ source that takes several seconds to compile.

> 1. No, the grammar is not the issue per se. As you say, C has the same problem, and C code invariably compiles dozens of times faster than C++, and both Zig and Rust have modern grammars, but Zig compiles about as quickly as C and Rust is only somewhat faster than C++ (depending on features used).

Sorry, the C++ grammar is terrible. There are lots of things where C++ can't figure out whether something is a class or name or template or constructor call until looking way far down the chain.

However, you are the first person I think I have ever heard claim that Rust is faster than C++. Rust is notoriously slow to compile.

Zig generally compiles much faster than most C projects I've used. However, that is difficult to lay at the hands of C as a lot of those are the build system being obtuse.

>Sorry, the C++ grammar is terrible.

When did I say otherwise? What I said was that it's not the main cause of C++'s long compilation times. The grammar causes other problems, such as making it more difficult to write parsers for IDEs, and creating things like the most vexing parse.

>However, you are the first person I think I have ever heard claim that Rust is faster than C++. Rust is notoriously slow to compile.

It's kind of a mixed bag. Given two projects of similar complexity, one in C++ and one in Rust, the one written in Rust will take longer to compile if organized as a single crate, because right now there's no way to parallelize compilation within a single crate. However, compiling the C++ version will definitely be the larger computational task, and would take longer if done in a single thread. Both contain Turing-complete meta-languages, so both can make compiling a fixed length source take an arbitrarily long time. Rust's type system is more complex, but I think C++'s templates win out on computational load. You're running a little dynamically typed script inside the compiler every time you instantiate a template.

(one of the reasons Go compiles fast is its compiler is really bare bones, comparatively speaking it does very little in the optimization area vs what you would see out of .NET and JVM implementations, not even mentioning GCC or LLVM)

It isn't just that, but there are a number of ways that templates can cause superlinear type checking behavior.

I know I'm weird, but as long as I can compile individual units and link them to other already-compiled units then compilation speed is something I don't care about much at all.

That's fine and dandy as long as you don't use C++ with templates and other stuff that needs 90% of the code to be in the headers.

Any recompile triggers a lot of recompilation.

And no, I can't change the code the other hundred devs are writing in my company.

In C++ there is no individual units. In C it's somewhat plausible with some discipline.

> In C++ there is no individual units

There's a still a notion of the compilation unit, which is the cpp file on which you invoke the compiler. Due to templates being prevalent in C++, included files contain the implementation as well; so any changes you make there lead to re-compiling a lot more compilation units than, for example, in C.

That is kind of restating my point. There might be these translation units, but they're tightly coupled to all the other translation units, making them not very individual IMO.

yes, that's why I used the disclaimer "as long as I can compile individual units and link them".

The problem with C++ templates is accurate, and is one of the several reasons why I avoid using templates in my own C++ code. I don't have that freedom in the code I write for my employer, though.

> In C++ there is no individual units.

Yes, there are. Those unit boundaries are often blurred by other C++ features, but they do exist.

Reply on news.ycombinator.com