[pandaman]

I have memories of people trying Java for the same purpose. I believe, the C# will ultimately meet the same fate. Here is why (I believe that):

Switching from C/C++ to a GC language is nothing like switching from Assembler to C/C++. We did not write Assembler because we liked it. We wrote it because we targeted 8/16-bit archs, which did not fit C/C++ memory model quite well. Because of this Assembler had huge speed and size advantage. As soon as flat memory archs became viable almost everybody dropped Assembler without regret. We lost a bit of performance (which could have been easily regained by rewriting inner loops in Assembler) in exchange of reusable, portable code, type system and terse program texts. These allowed to write much more complex games by engaging bigger teams who could sensibly collaborate over a much bigger project in C/C++.

What is the transition to GC from C++? You sacrifice a lot of performance for freedom from managing memory. I don't know about other people but having shipped few AAA games my only concern was fitting shit into available memory (which GC does not help, to say the least). You allocate memory on load. If you load multiple levels you nuke the previous level's memory. If you stream you do the same but your "levels" are now "segments" in a pool. Of course, if you want to go full-GC and allocate your structures byte by byte, uint32_t by uint32_t - be my guest, not sure how you plan to compete but, hey, load times (same as frame rate) are not that important as they say on the internet :) Nevertheless, as soon as 32-bit targets become completely obsolete people will find out that you can pre-allocate everything off-line in 64-bit so all you will have to do at run-time is map/unmap pages (which GC won't help either). In conclusion: one dropped performance on the floor (due to cache issues, SIMD issues, off-line vs JIT quality issues etc) and picked up the solution to the problem one should not have had in the first place.

[nonsince]

C# is, among GC languages, particularly bloated with regards to memory. There is huge amounts of runtime type information (including stringified versions of everything) and you have close to no control over memory management without hooking into ugly APIs. I wish that Swift or Rust had been usable early enough to get swept up in the hype train so that we could avoid this crap. Also, it would mean I didn't have to use OOP to write games.

[golergka]

C# specifies the language, not the runtime. Apart from .NET and Mono, you can compile C# to cpp (what Unity does under the hood, btw).

> Also, it would mean I didn't have to use OOP to write games.

There are many instances where other programming paradigms are more usable - but in my experience, gameplay logic, specifically, maps very well onto classic OOP concepts.

[nonsince]

If gameplay mapped well onto OOP then the big game companies would use it, and not data-driven programming (specifically, entity-component systems). I think that there are a couple things that require OOP-style dynamic dispatch, but they don't need to be pervasive, even a function pointer stored as a struct member is more than enough to provide all the benefits of OOP without having to take on all the drawbacks.

[jameshart]

You should take a look at what is going on in the latest dotnet core CLR and C# compiler around 'ref-like' types and Span<>s. Basically a way to let you write 'safe' C# code that interacts with memory that is allocated in different ways (for example with controlled allocation outside of the GC).

Since C# is a multiparadigm language with strong functional as well as OO fundamentals, I suspect it will make it possible for libraries and frameworks to be built which enable allocation free code in C# gameloops. It's not just game engines which have these requirements - the same functionality is also critical for high performance servers, and machine learning applications, so this is a language feature with a lot of demand.

[golergka]

You're forgetting that this transition takes place in game engines: so the critical path is still mostly C++, while C# is responsible for game logic and most of it is run on callbacks, not in the update loop.

Same goes for memory: textures and other assets are managed through the engine; C# only contains game logic data and handles to the assets.

[animal531]

I was once working on a solo indie game using C# with OpenGL bindings. If you create it along the normal program lines of object instantiate/use/destroy (and repeat), then you're soon going to run into serious issues.

At the time I had many issues with the GC, Large Object Heap and fragmentation (although this had still been in a much older .NET version).

For a game project, both for speed and memory, you want as much as possible to be pre-allocated (and all at once). During run-time you want as few objects as possible being instantiated, that way you save cycles on both the objects, and eventually on the GC. Of course it's not always that easy.

After that it depends on your targets, you'll never fit a Unity C# game as tightly into any target as a native app, but it's a trade-off that's worth making in many cases.

[pjmlp]

Java only failed because Sun didn't care enough for pushing it into game developers after their initial presentations. Actually they didn't care for any other scenario that wasn't JEE, in terms of putting resources into it.

Switching from Assembly into C, Pascal, Modula-2, QuickBasic was mostly an amateur thing until SDKs started to be based on C.

Same with C++, we were the outliers until console vendors started pushing C++ into devs. See Mike Acton's talks about about how he enjoyed having been forced to move from C into C++.

The thing with GC is not the freedom to manage memory, you still need to take care about how memory is managed. What one gets is type safety and an uniform way to manage memory across libraries.

You know exactly who is responsible for releasing it, there are no double deallocations, releasing invalid addresses, or having to deal with library specific smart pointers (CPtr, QSmartPtr, unique_ptr, my_in-house_ptr, ...).

> due to cache issues, SIMD issues, off-line vs JIT quality issues etc

Having a GC doesn't preclude lack of support for cache handling or SIMD.

Many GC based languages have such support and actually C# 7 with Microsoft's toolchain (not Unity's one) does offer such features, even if they can still be improved.

As for JIT quality, it is no different from any other compiler, otherwise as I referred, why not discuss C performance using MS-DOS compiler's benchmarks.

[flohofwoe]

One problem I have with memory-managed languages for writing high-performance-code is that you need to give up the simplicity of not having to care about memory management. You need to know exactly how the garbage collector in this specific language (and current version of the language) works, and how to build your code around that. The resulting 'mental burden' is the same or even bigger as in C or C++, and the resulting code often looks worse then C++ code which does the same. An extreme example of this is asm.js, much higher performance than typical JS since it only uses numbers, not JS objects (and thus basically removes the need for garbage collection), but so much different from manually written JS that it is basically a different language.

As for smart pointers in C++: they are only useful in a code architecture that essentially emulates C# or Java (all objects as single entities on the heap) using C or C++ in such a "memory-managed style" really doesn't make much sense since it will be slower then a GC. The point about manual memory management is to reduce dynamic allocations as much as possible and have all your data either on the stack or in long-lived, big memory chunks. If you need to allocate and deallocate all the time, or track the ownership and lifetime for every little memory chunk, you're not using C/C++ to its advantage.

PS: C# for tools is great though (not so much because of the language, but because of the standard framework, which is much nicer than the C++ stdlib).

[edit: typos]

[pjmlp]

Just like you need to know how malloc()/new/STL allocators for your specific C or C++ compiler are implemented.

And most implementations actually suck at multi-threaded code, to the point there are companies selling better implementations of them.

> The point about manual memory management is to reduce dynamic allocations as much as possible and have all your data either on the stack or in long-lived, big memory chunks.

Also possible in GC enabled languages, grated not available in all of them.

As former C++ dev, I can guarantee you that unless you write 100% of the code yourself, there will be leaks, double frees, delete calls instead of delete[] ones, releasing unallocated memory and totally lack of control of third party leaks.

[flohofwoe]

The point I'm trying to make is that you loose the advantage of a garbage collected language once performance becomes important, so there's no point in using such a language for high-performance code in the first place.

Integrating third party dependencies into C++ is its real achilles heel, I agree. I only accept dependencies that come with source and either don't allocate dynamic memory at all, or allow full control over allocation by providing custom hooks. I also learned over time that typical C code is usually much easier to integrate than typical C++ code (which very often is an over-engineered template mess).

The way to deal effectively with memory management bugs is to allocate as little as possible. There are good analyzer and runtime tools (for instance the new runtime memory debuggers in Xcode and Visual Studio, the static analyzers, the clang address sanitizer in Xcode etc), these provide a much better view on what's going on under the hood than most managed-language-tools. And those tools make it quite trivial to catch most memory-related bugs. If you're doing hundreds-of-thousands of allocations it becomes much harder though, because than the problems will be lost in all the noise.

It is a valid argument though to use higher-level languages in high-level gameplay code, but this should only do scripting-style stuff, glueing subsystems and game objects together. Up in this area I don't want to care about ownership and lifetimes. It's important to find the balance though where the high-level code should better be moved into a low-level, central system.

[pandaman]

>Same with C++, we were the outliers until console vendors started pushing C++ into devs. See Mike Acton's talks about about how he enjoyed having been forced to move from C into C++.

The first (major) console C++ SDK library is GNM/GNMX for PS4. Are you saying C++ games has been outliers till 201x? I guess this depends on how you define outliers then. Other than the guy casting his game writing on youtube (Code Hero?) I don't know of any significant titles in C past 2000. I have not seen all the games but I shipped a few myself and have friends working on many more and nobody that I know had been writing pure C. Think of all vtbl-> you'd have to write if you are targeting Windows/Xbox just for the pleasure to name your source file .c instead of .cpp!

>What one gets is type safety

Thanks, we've got this in C++ already.

>You know exactly who is responsible for releasing it, there are no double deallocations, releasing invalid addresses, or having to deal with library specific smart pointers (CPtr, QSmartPtr, unique_ptr, my_in-house_ptr, ...).

Never had been a problem for me. The only game I touched that had smart pointers had them because it used NetImmerse and it got cancelled anyways.

>Having a GC doesn't preclude lack of support for cache handling or SIMD.

I don't know what is cache handling, to be honest, but what I meant is that, since GC moves stuff around memory, it's likely to create cache hazards. And since it probably uses a single pool, you cannot change cache policies on per-entity basis.

>As for JIT quality, it is no different from any other compiler, otherwise as I referred, why not discuss C performance using MS-DOS compiler's benchmarks.

Really? You believe JIT will be on-par with a profile assisted, bruteforcing compiler, which can spend a week optimizing a single function?

[pjmlp]

> The first (major) console C++ SDK library is GNM/GNMX for PS4.

The major C++ SDK was the DirectX SDK for the Sega Saturn.

> Are you saying C++ games has been outliers till 201x?

Yes, the code was basically C compiled with C++ compiler.

> >What one gets is type safety

> Thanks, we've got this in C++ already.

Not when what most write is actually C compiled with C++ compiler.

> I don't know what is cache handling, to be honest, but what I meant is that, since GC moves stuff around memory

Is the ability to write cache friendly code.

If you don't want the GC to move memory around then don't allocate it via the GC.

Many GC enabled languages also allow for global statics and stack allocation. Even C# has some support here, even if it isn't comparable to what Modula-3 or D allow for.

Also there is also the possibility to just allocate it off GC heap.

> Really? You believe JIT will be on-par with a profile assisted, bruteforcing compiler, which can spend a week optimizing a single function?

No, but JITs can also make use of PGO just like AOT compilers. IBM J9 JIT and .NET RyuJIT have such support.

Also many GC enabled languages, including C# do have AOT compilers to native code code as well, it is not as JIT is the only viable approach.

[pandaman]

>The major C++ SDK was the DirectX SDK for the Sega Saturn.

a) there was no DirectX SDK for Sega Saturn. b) DirectX SDK (including one for Sega Dreamcast) is not C++. It has C++ bindings but is usable from C.

>Yes, the code was basically C compiled with C++ compiler.

You mean if I have taken that code and compiled with C compiler it worked? You realize even the DirectX C++ wrapper is not already C, right? You know classes, overloads, namespaces are not C?

>Not when what most write is actually C compiled with C++ compiler.

Could you explain exactly how this works? The C++ compiler uses ML to recognize that the code is not exactly following Alexandrescu's book and turns off typechecks? I seriously just don't understand what you mean here. I usually take "C compiled with C++ compiler", "C with classes" etc as "not enough GoF patterns for my taste" but you are making some other claim here it seems.

>If you don't want the GC to move memory around then don't allocate it via the GC.

So, why do you want GC in the first place? For types, which somehow disappeared from C++?

>No, but JITs can also make use of PGO just like AOT compilers.

How exactly does it work? The code stops executing for a week, the profiler gets run to under user credentials and then JIT finally decides?

[pjmlp]

> there was no DirectX SDK for Sega Saturn.

Broken memories, I didn't bother to search for it (Saturn vs Dreamcast).

> You mean if I have taken that code and compiled with C compiler it worked?

Of course it would, that was one of the design goals of C++.

C90 is mostly a C++98 subset, except for stronger type conversion rules (no implicit void* conversions), precedence order for operator ?: and typedef/struct namespaces.

> You realize even the DirectX C++ wrapper is not already C, right? You know classes, overloads, namespaces are not C?

Yes, but COM is also callable from C by design. Also I have seen many codebases that have restricted C++ code to calling DX APIs, with everything else being compilable by a C compiler as well.

> Could you explain exactly how this works? ...

1 - Rename .c translation units to .cpp, .cxx, .C

2 - Invoke C++ compiler on them

3 - Fix compiler errors related to semantic differences in C subset of C++

4 - Forbid use of any C++ specific feature beyond those required to use the OS SDK.

> So, why do you want GC in the first place? For types, which somehow disappeared from C++?

Productivity.

> How exactly does it work? The code stops executing for a week, the profiler gets run to under user credentials and then JIT finally decides?

PGO data generated by the JIT compiler gets updated after each application execution and is used as input for optimization selection just like in an AOT compiler by a multi-stage compiler.

Feel free to read Android 7 ART source code to learn how about a possible implementation.

[pandaman]

>Of course it would, that was one of the design goals of C++. >C90 is mostly a C++98 subset, except for stronger type conversion rules (no implicit void* conversions), precedence order for operator ?: and typedef/struct namespaces.

I think either I don't understand something you are trying to say or you are confused. C being a subset of C++ (one of design goals) does not imply C++ is a subset of C. C++ code in general cannot be compiled with C compiler without rewriting.

>Yes, but COM is also callable from C by design. Also I have seen many codebases that have restricted C++ code to calling DX APIs, with everything else being compilable by a C compiler as well.

You have blah->Foo(bar); in C++. In C it won't compile by any design. That code has to be rewritten as blah->vtbl->Foo(blah,bar). It is not C code as you claimed. It won't compile with C compiler. It's C++.

>> Could you explain exactly how this works? ... >1 - Rename .c translation units to .cpp, .cxx, .C

I asked how type checking disappears in this process, not how you compile C with C++....

>PGO data generated by the JIT compiler gets updated after each application execution and is used as input for optimization selection just like in an AOT compiler by a multi-stage compiler.

How can a JIT compiler obtain the PGO data in the first place? Is it running under profiler all the time? You realize that you've just refuted your claim about performance not being affected, right?

[douche]

> Not when what most write is actually C compiled with C++ compiler.

Is this moving the goal posts a little bit? C++ now doesn't necessarily look anything like C++ written 20 years ago. Way back, people were writing C+, or C-with-classes, but that's what C++ was, in that era.

[pjmlp]

Not at all, just because C++11, C++14 and C++17 offer lots of improvements over C++98, doesn't mean people use them.

I can assure you that at enterprise level, every time I have to integrate C++ code with our Java or .NET stacks, it looks exactly like C++ written 20 years ago, even if it was actually written last week.

Just because C++ has been vastly improved, doesn't mean everyone using it are adopting the new features, some people rather stay in Python 2 forever.

[wiz21c]

> Really? You believe JIT will be on-par with a profile assisted, bruteforcing compiler, which can spend a week optimizing a single function?

I was in game development looong ago (C++ / assembler, lots of 3D software rendering). What kind of compiler is that ? Could you give some pointers ?

[pandaman]

All modern compilers do this (MSVC, clang, I believe gcc too). Search for "PGO" (Profile Guided Optimization). Also, not just for C++ - PS3 shader compiler did brute force instruction scheduling optimization (using shaderperf instead of profiler).

[wiz21c]

damn time to refresh my knowledge :-) That's what happens when you do CRUD too long :-)

Last time I optimized code the hard way was using VTune to channel the right operation in the right pipeline.

[maccard]

In UE4 everything is garbage collected. Lots of high performance, AAA games using it.

[caseymarquis]

Maybe Nim would be a good GC included C++ replacement then? Its GC can reliably be constrained to a maximum run time of 2ms, and it has strong macro support.

[pandaman]

IMO, 2ms is too much for strong macro support. Even if GC had 0 run-time cost I'd still rather use C++ since 99.9% of my code is creating memory structures of particular layout (I write graphics). To write the same code in a GC language I'd have to map memory to a file and then use the strong macro system to emulate native struct type from C++ :)

[nonsince]

This is Nim vs C#, not Nim vs C++. If you're writing hot loops and graphics code a GC is utterly anathema to your goals. I think given 5/10 years Rust will be the language of choice here, but the ecosystem is nowhere near good enough yet.

[pandaman]

As of now I don't see any problems with C++. I doubt Rust is going to offer something so much better that people will abandon the tools they already have for these benefits. Though, who knows? I never touched Rust.

[pjmlp]

It just needs a OS company to pick it up and push it down dev throats, like it happened before with C and C++.

If you want the goodies on Windows you need C++, if you want the goodies on iOS and OS X you needed Objective-C and now Swift as well, the goodies on Android require Java, the goodies on consoles moved from C to C++.

Same here, if tomorrow one of those companies decides their SDK should be Language X, anyone that wants a piece of the pie will learn Language X.

[TuringTest]

This could come from a IoT company that does things well. A company providing unsupervised devices would benefit from a OS with a strong approach to avoiding security bugs.