Comment by tiberius_p
2 days ago
I'm not up to date with the latest developments in C++ but would't it be straightforward to do something like "#pragma pointer_safety strong" which would force the compiler to only accept the use of smart pointers or something along those lines. Was anything like this proposed so far?
You might be interested in this talk[0] by a WebKit engineer on how they're implementing similar approaches using libTooling and their own smart pointer types.
For example, their tooling prevents code like this:
from compiling, forcing you to explicitly create an owning local reference, like so:
unless it's a trivial inlined function, like a simple getter.
[0]https://www.youtube.com/watch?v=RLw13wLM5Ko
I was going to link to this.
My interpretation of Geoff's presentation is that some version of profiles might work, at least in the sense of making it possible to write C++ code that is substantially safer than what we have today.
Geoff's stuff is mostly about heuristics. For his purpose that makes sense. If Apple are spending say $1Bn on security problems and Geoff spends $1M cutting such problems by 90% that's money well spent. The current direction of profiles is big on heuristics. Easy quick wins to maybe get C++ under the "Radioactively unsafe" radar even if it can't pass for safe.
The most hopeful thing I saw in Geoff's talk was cultural. It sounds like Geoff's team wanted to get to safer code. Things went faster than expected, people landed "me too" unsolicited patches, that sort of thing. Of course this is self-reported, but assuming Geoff wasn't showing us a very flattering portrait of a grim reality, which I can't see any incentive for, this team sounds like while they'd get value from Rust they're delivering many of the same security benefits in C++ anyway.
Bureaucrats don't like culture because it's hard to measure. "Make sure you hire programmers with a good culture" is hard to chart. You're probably going to end up running some awful quiz your team hates "Answer D, A, B, B, E, B to get 100%". Whereas "Use Rust not C++" is measurable, team A has 93% of code in Rust, but team B scored 94.5% so that's more Rust, they win.
3 replies →
I don't think that really accomplishes anything. If you interpret it broadly enough to meaningfully improve safety you have to ban so much stuff that no codebase will ever turn it on. It's a pretty straightforward locally-verifiable property as well, so people who really want it don't need a pragma to enforce it.
The problem with this would probably be that you usually have to use some libraries with C APIs and regular pointers.
You could compile your program with address sanitizer then it at least crashes in a defined way at runtime when memory corruption would happen. TCC (tiny C compiler initially written by fabrice bellard) also has such a feature I think.
This of course makes it significantly slower.
> pragma pointer_safety strong" which would force the compiler to only accept the use of smart pointers
You’d possibly just be trading one problem for another though - ask anyone who’s had to debug a shared ownership issue.
That's what the "Profiles" feature is. The problem is that any nontrivial real world program in a non-GC language needs non-owning reference types to perform well, and you can't express the rules for safe use of non-owning references without augmenting the language. People have tried. You need something more sophisticated than using smart pointers for everything. In the limit, smart pointers for everything is just called "Python".
What infuriates me about the C++ safety situation is that C++ is by and large a better, more expressive language than Rust is, particularly with respect to compile time type level metaprogramming. And I am being walked hands handcuffed behind my back, alongside everyone else, into the Rust world with its comparatively anemic proc macro shit because the C++ committee can't be bothered to care about memory safety.
Because of the C++ standards committee's misfeasance, I'm going to have to live in a world where I don't get to use some of my favorite programming techniques.
> You need something more sophisticated than using smart pointers for everything. In the limit, smart pointers for everything is just called "Python".
I don't see how that follows at all. What makes Python Python (and slow!) is dynamic dispatch everywhere down to the most primitive things. Refcounted smart pointers are a very minor thing in the big picture, which is why we've seen Python implementations without them (Jython, IronPython). Performance-wise, yes, refcounting certainly isn't cheap, but you just do that and keep everything else C++-like, the overall performance profile of such a language is still much closer to C++ than to something like Python.
You can also have refcounting + something like `ref` types in modern C# (which are essentially restricted-lifetime zero-overhead pointers with inferred or very simplistic lifetimes):
https://learn.microsoft.com/en-us/dotnet/csharp/language-ref...
It doesn't cover all the cases that a full-fledged borrow checked with explicit lifetime annotations can, but it does cover quite a few; perhaps enough to adopt the position that refcounting is "good enough" for the rest.
Python doesn't have lvalues in the way that C++ and Rust do. You can't refcount everything and still pass lvalues to subobjects. If lvalues to subobjects are important, you need borrow checking.
> Jython, IronPython
Both of which have modern, concurrent, parallel, and generational garbage collectors.
> In the limit, smart pointers for everything is just called "Python".
To be more precise, it's old Python. Recent versions of Python use a gc.
> And I am being walked hands handcuffed behind my back, alongside everyone else, into the Rust world with its comparatively anemic proc macro shit because the C++ committee can't be bothered to care about memory safety.
Out of curiosity (as someone working on static analysis), what properties would you like your compiler to check?
To be even more precise:
> Reference counting is the primary mechanism of garbage collection, however, it doesn’t work when the references have cycles between them and for handling such cases it employs the GC.
I've been thinking for a while now about using dependant typing to enforce good numerics in numerics kernels. Wouldn't it be nice if we could propagate value bounds and make catastrophic cancellation a type error?
Have you worked much with SAL and MIDL from Microsoft? Using SAL (an aesthetically hideous but conceptually beautiful macro based gradual typing system for C and C++) overlay guarantees about not only reference safety but also sign comparison restriction, maximum buffer sizes, and so on.
9 replies →