How to design a replacement for C++
apenwarr.ca
apenwarr.ca
Especially his comment about garbage collection: if you're not going to have garbage collection, frankly, what's the fucking point? You might as well continue to use C/C++ then.
He is demonstrably wrong that you cannot reliably combine garbage collected an non-garbage collected memory with a mark sweep collector! My C++ to Io binding library builds C++ objects in script and stores script objects in C++, and Io uses a mark-sweep collector and I am using it right freaking now this very minute in a game I'm programming without trouble. I regularly profile it in valgrind; there are no memory corruptions. The solution? Give the script-created C++ objects the same semantics as C++ code does. C++ temporaries are created in script with a "temp" keyword and get passed by value. Long lived C++ objects are created with a "new" keyword (in script) which stores a pointer to the C++ object. You have to manually call "delete" (again, in script) on the script-object-containing-the-C++-pointer to delete the objects. Here's the beauty of the logic: as long as you are holding the pointer, the script object will not be collected. If you do allow the script object to be collected, by definition that means it is unreachable, so by definition you FORGOT to call delete on it, ergo the system could then throw an exception if an undeleted C++ pointer is ever collected. So rather than a problem, it actually helps you catch undeleted memory.
For C++ objects that hold script objects, I simply force any C++ object that wants to hold a pointer to a script object to implement my IMarkable interface, and the script binding is smart enough to know that if a C++ implements IMarkable, it gets included in the mark-sweep so it can mark off its owned script objects.
Written by Mozilla employees, and very much like Go, except that shared state is immutable by default, no global GC, and no null pointers. It does have fine struct layout control, RAII always, and type-parametric user code (purely structural too!).
It does have a real module system for grownups, which even if it doesn't preclude your fetish for #include, it will at least make you feel bad about it.
Maybe you like [lambdas], maybe you don't, maybe you think
they're God's gift to programming and any language
without them is an infidel. But adding them would be
harmless, anyway.
Otherwise no issues, but it does feel like the author is advocating C++0x minus heavy template metaprogramming - which is a perfectly fine language in my book. Why not just use that and get on with life?
Templates might have been harmless if there was a way to do basic obvious stuff (callback functions, strings) without using them.
interface 'a collection {
void sort(lt:'a -> 'a -> bool);
}
foos.sort(fn x y -> x.bar() > y.bar());
bool sort_by_bar_gt(x:Foo, y:Foo) {
return x.bar() > y.bar();
}
foos.sort(sort_by_bar_gt);
I would like to stress: I think lambdas are a requirement for any language, I just take issue with the argument presented to convince lambda haters.
Working with C/C++ for embedded systems often means you have to have total control of the memory layout of your objects. You sometimes want to define classes which have data members that directly map on to something in memory, and you have to know exactly how the class will be layed out in memory, and more importantly, that it won't take up extra memory without your knowledge.
Specifically, his point #5, "Automatic vtable generation.", might get in the way. C++ gets away with vtables because they aren't included by default, only included when you define a virtual method. The same needs to be true of any replacement.
2. Hahahahahaha
3. Aside from the fact that his wording is backwards from his meaning, I agree that it should be possible to write code that doesn't ever garbage collect.
4. Ok
5. Does this guy know the function of a linker?
6. Again, he asserts something you are to avoid, when really he should be emphasizing what you need to support: code without runtime checks.
7. Wrong. This is one of the few opportunities you have to beat C at its own game. C hardcodes assumptions about the stack that prevent certain low level optimizations.
His #8 is then a summary of the entire article, and really is the only valid point he's made. C is close to a sweet spot for wrapping low level code in somewhat higher level clothing. You can't beat it at the same level of simplicity (not if you count the cost of any potential switch anyway). The way to beat C is to naturally support the high level / low level hybrid that he's doing with python+C in a single language.
Would you please expand on this? I can't think of any assumptions C makes about the stack that C++ wouldn't make; neither can I think of low-level optimizations that said assumptions prohibit. I'm curious about what you have in mind.
This statement is simply wrong, and saying "End of discussion" is useless, if not plain stupid.
It's like saying: 1+1 = 3, End of discussion.
Some of the "functions" in there are macros.
You can import C headers without your language being a superset of C.
Are you sure about this?
#define getc(stream) ...
("Free-standing implementations" like kernels have a lot of freedom to muck with the standard library. But that's not really relevant.)
#define getc fgetc_or_similar
#include <stdio.h>
#undef getc
int main(void) {
getc(stdin);
return 0;
}
Opt-out garbage collection, type inference, operator overloading, full-blown OO but you can use covers if you just want to have methods on primitive types, some type inference, saner syntax, you can include C headers all you want (but it has a real, good module system), one-time declarations in any order, objects are by-reference....
Besides, you can replace the SDK pretty much as you want, you're not tied to any 'threads' implementation, the proof being: there's actually an OS written in ooc! http://github.com/tsion/oos
(Of course I'm barely scratching ooc's features right now)
For the second question, the article's answer is: kernels, drivers, highly performance-sensitive code like game engines, virtual machines, some kinds of networking code, and so on. And for me in particular, it also includes new plugins to existing C-based legacy systems, including Microsoft Office.
I think this is the first step to discussing "a replacement". I fear that every C++ programmer would give a different answer to each question.
The ideal replacement in the article can be summarised as most/all language features of C, plus some chosen subset of the features of C++.
Although with most of them implemented differently to how C++ has implemented them. Presumably that better implementation is to be done without compromising it's C-like qualities. Which sounds to me like the same trap that C++ itself fell in.
Google Go is fast because it stole ideas from Wirth and followers (Turbo Pascal, Modula) who actually took the care to provide sane module linkage for decades. The proper path to the evolution of C was to keep includes and classic linking but to allow "module linking" too. All that was being done for decades, only not in C and C++.
I personally prefer C + C libraries + the non-invasive suggestions which are mentioned which are already found in C++, Obj-C, etc.
I don't know any other programming environment that can claim that without adding tons of complication (the C/python interface layer is the simplest I've ever seen, except maybe for c/tcl).
Were it possible to have Python code be in some non-root node, it'd be great (e.g., for parsing text, or doing something algorithmic). As it is, when everything has to be fed through the interpreter, the design becomes top-heavy from a performance standpoint because the slow stuff is on top. And any time you want to optimize something but then use some Python, you're stuck (it becomes part of the root subtree again).
At least that was my experience. But Python does encourage interesting alternatives for performance (with generators, etc.). And if you're already mostly doing Python anyway (I'd say the threshold has to be 70%), then it seems to make sense (v. the opportunity cost of switching contexts between languages -- if you are spending more than 30% in C, I find it difficult).
But it's a really interesting issue. C, C++, and Obj-C all take a much longer time to develop up front. But if you refactor into libraries, maybe not in the long-run...
Additionally, all IO code in the standard library and most C-extensions that I've seen release the GIL around blocking IO calls. This allows multi-threading to be used for the standard blocking IO use cases. Further, the computationally intensive routines in numpy and scipy that I’m familiar with also release the GIL to facilitate multithreading at the C-level
The GIL only becomes an issue if you want multiple threads to be executing Python bytecode concurrently. I don’t believe this is a common problem as one rarely runs computationally intensive code at the python level. Such computationally intensive work is generally performed through C extensions or through external C functions (as accessed through ctypes). As long as such C extensions properly release the GIL, multithreading at the C-level is unaffected.
If what you're doing in C is numerical or computational in a functional way that doesn't need a hierarchy of types -- then keeping all your types in Python and calling into C for IO, computation, etc., makes a lot of sense.
Where I've run into problems is if the C modules themselves get sophisticated and start having scene graphs or hierarchies of shapes, etc. Because then it makes sense to mirror the hierarchy in Python -- but then you get real performance issues.
So I guess if you can ensure that your C modules have a clear functional interface (or at least one whose side effects are clearly defined) and doesn't involve anything but a shallow type hierarchy, then Python + C is all good for that. I.e., the C modules have to be very well-defined or have just about no 'code smell' (at least near their interface with Python). Arguably you can always do this if you make an effort to refactor your C modules. But yeah, because Python must always be at the root node, one is sort of constrained to serve the 'top', so to speak. And that may encourage a sort of top-down design which is less amenable to bottom-up programming (as mentioned in On Lisp), etc.
1. Can embed and call C code with no overhead, uses C representation internally
2. Embedded C can use CPP, otherwise the language has macros
3. GC is optional
4. No "system" thread
5. Optional standard library
6. Static typing, linear types, types as propositions, programs as proof.
7. ML-style exceptions, not insane
8. Examples from K&R C translated into ATS http://www.ats-lang.org/EXAMPLE/KernighanRitchie/
I'd like to see an example, if anyone has one. But for me, the template system is the best thing about C++ - they are much closer in functionality to Lisp macros than preprocessor macros are!!
http://www.tiobe.com/index.php/content/paperinfo/tpci/index....
Outstanding.
Even in my case, I was happily coding in C# in the last 3 years, and in January - bam! - I am on another project written in, well, C++.
Thank you. I am not looking forward to the day when people begin to advocate Java or something like that for these applications, and I am glad to see this explicitly acknowledged.
Can't do operator overloading though because they want to compile down to C without mangling names
7. One-time declaration/definition of functions.
Aren't these in conflict? A C-style preprocessor is crippled without the header-file model. You'd do better to replace it with some other kind of metaprogramming.
I think you might be right though, I'd loose the cpp preprocessor but add built-in hygienic macros and constants to make up for the loss.
Of course, the above model will most likely increase (re-)compilation times.
- I don't think .NET-style generics can be implemented in a statically typed language that doesn't have a common base class (e.g. everything derives from class Object). This is an oversimplification to say this but at some level generics are implemented as syntatic sugar for typecasts between Object and the type parameter type. If you want generics instead of templates, a lot of C#'s design gets pulled in along for the ride. Nothing wrong with that (C# is a fine language), just as long as you understand it will happen and why. But also be aware that it means you will never be able to make a generic array class that performs as fast as a C++ STL vector does - which violates your own starting criteria, that it has to be as fast as C++ to succeed.
- Most of the time I wish pass-by-reference were implicit, but pass by value has some special consequences in C++ that make it unlike many other languages. I was able to write a class that does fast bit-twiddling on itself, with its internal representation being an integer, and the entire thing was no more expensive in memory or speed than C functions bit twiddling on an int. So I was able to get abstraction and raw speed with no compromises. That's a rare use case, but I think it is illustrative of why pass by value is in C++ - you don't need it most of the time, but when you need it, you really need it.
As for not being able to reseat a reference, I found that frustating too until I started learning LLVM and about SSA (static single assignment). The common answer to why you can't reseat a C++ reference is just that it has something to do with compiler optimizations. That common answer is too weak; it has EVERYTHING to do with how the compiler optimizes the code. Basically, if you're asking why can't I reseat a reference, you don't understand what a reference is. But it's not your fault: what you really don't understand is that C++ uses the word reference to refer to a thing that is not exactly like references as used in other languages. I used to think a reference was syntactic sugar for a pointer; that's not quite true. A C++ reference literally IS the object; it's just converted to a pointer during the process of a function call. The reason you can't reseat the reference is that if you could the static analysis of the program becomes immeasurably harder and later in the code the compiler cannot be absolutely sure that it knows what the reference really refers too.
In a way, not being able to reseat a C++ reference is a lot like not being able to assign to a constant defined by a macro or not being able to assign a new address to a function at runtime. It's part of a class of things that happen at a stage before runtime, and to change the system so that it would be possible would require introducing a level of indirection.
CLR (.NET) generics compile separate code for each value type parameter, so you can have eg. vectors of structs, with the same performance as in C++. (for reference types, it uses the same code with an additional hidden "generic context" parameter, but that's an implementation detail). CLR has pass-by-value too, and although older versions of the MS.NET JIT did not inline such functions, the current version should have performance similar to C++.
Unfortunately, few programmers know how to get the maximum performance out of the CLR.
About not being to able to reassign references in C++ - the main reason is probably because that would require a separate operator, since assigning to a reference calls operator= on the class.
C++ has subtle interactions in many areas of the language, especially operator overloading, allowing you to override even assignment. If C++ can not do something apparently simple, there is probably a good reason for that, only obvious to those who know the spec inside out.
Fortran for example has this - pointers in fortran behave like value types (like C++ references) when performing operations on them, however there is a special => operator to change the actual pointer. I never used operator overloading in fortran but I don't think it would be much of an issue, you would be able to overload the operator for anything except a pointer type.
This is not true. Java generics work this way (google "type erasure"), but C# generics are implemented with specialized code, so that e.g. List<int>.Add has its own IL implementation distinct from the one for List<Bitmap>.Add. This lets the JIT do type-specific optimizations, like not having to box the ints in a List<int>.
The specialization code also tries to share code in cases where the type parameters don't matter (e.g. so that you don't have a million identical implementations of List.Count hanging around).
2. Fuck the C preprocesser. An architecture based on #include files is a joke; you will never make something better than C++ if you stick with that compilation model, you'll only make another C++ (corollary to the phrase, those who don't understand Unix are doomed to re-implement it poorly). Here's an idea: sandbox the god damn shit preprocessor and parse its output to provide input to the FFI (foreign function interface) JIT compiler. Under no circumstances allow macros defined in an #include to spill into the language proper. This is not negotiable.
3. No, mark-sweep garbage collection can be made to coexist with C++-style memory management.
4. If you aren't going to include built-in threading or some better form of concurrency, why bother? Unless you have a time machine and intend to go back to 1980 to introduce your language there.
5. I don't think it's that clear cut (about requiring (or not) a standard library) but I won't argue this point.
6. I think you're right in the sense that if you do include dynamic typing it is going to lead to creating something completely different than the stated goal. I'd say it would be a better language, but it would not be the correct language for replacing a statically typed one.
7. I think exceptions are an important language advance, but I can't dispute that they can cause a mess of problems when it comes to cross-language (or cross-thread) boundaries.
8. If you look at C, it is not too verbose for coding the types of concepts which it supports (function calls with a small number of control-flow statements). But it does not allow you to define new concepts. Of course C is terse for calling functions: those are C's atoms, it's fundamental language element! So IMO comparing to C code based on how easy it is to call a function is the wrong thing to think about. True improvements would come at the macro scale from better organization of groups of functions (e.g. into classes) or at a semantic level, by not needing to specify a function explicitly at all (e.g. operator overloading).
I think exceptions are an important language advance, but
I can't dispute that they can cause a mess of problems
when it comes to cross-language (or cross-thread)
boundaries.
doSomething();
int result = doSomething();
if (result != OK) return result;
This is simply doesn't work. And few people programming C/C++ appear do it correctly; I'd imagine that many security flaws come from improperly proceeding code that should have checked a return code.
Haskell does a pretty good job, I think. With Maybe or Either, you have to check it to extract your value, but if you are doing a lot of things that will result in/use the results of Maybe/Either, you can just operate in a Monad that makes them the default.
Similarly, languages with multiple return values make you explicitly ignore the return value.
Exceptions have their benefits, but I have a personal preference for return values because they make it explicit in the interface. Also, I find that it makes more sense to handle various success and failure scenarios in the same bit of code instead of writing a best-case flow and assuming that exceptions will be thrown and errors will be dealt with somehow.
Exceptions are great if you can't continue due to programmer error, but most cases where people use exceptions are just expected execution paths that (in my view) make no sense to handle via non-local control flow.
One problem with exceptions is that they can only be one deep. Ie. if you are executing the code invoked by handling an exception, what should happen if another exception gets thrown. Java says forget about the old exception (ie. lose information), C++ says crash. Maybe there is something in between?
With or without exceptions, being error-safe is still really hard. I find you have to think "transactionally". For every operation you perform you have to think about what you need to clean up if it fails, so you focus on specifying what you want to do, not how you want to do it.
4. If you are going to need threads by default (e.g. for your GC thread) you are breaking lots of UNIX, notably fork(), and require far too much infrastructure for some kernel/embedded programming. Yes, a good concurrency story is needed, but your program should be able to run without threads.
GC depends on being able to find all valid pointers. C++ allows pointer games, like the xor trick for doubly-linked lists, which hide pointers. (It uses a single pointer field in each element to represent both the forward and backward pointer.)
GC can't collect things that are algorithmically inaccessible. There are algorithms that are efficient because they don't kill pointers to objects that are no longer in use.
I like GC, but there are costs.
closure int foo(int x) {return x + y;}
int(*foo_p)(int) = malloc(sizeof(foo));
memcpy(foo_p,&foo,sizeof(foo));
return foo_p;