Thoroughly Understanding C++ ABI (2024) | Hacker News Reader

Thoroughly Understanding C++ ABI (2024)

ykiko.me

83 pointsrramadass88 comments

C++ ABIs are why Win32 settled on things like COM to do cross-module passing of objects. Because nobody could agree on a standard, the COM standard enforced a very specific calling convention, and a very specific vtable layout for COM objects.

Not only, there is also XPC and IO/DriverKit on Apple, Android IPC, D-BUS, FIDL on Fucshia,....

Also C ABI also does not exist, people keep mistaking the ABI of their favourite C compiler with the OS ABI, which only overlap if the OS was written in C to start with.

For example on mainframes and micros, naturally not written in C, it is either a bytecode based ABI like TIMI on IBM i, or language environments like on z/OS, ClearPath MCP, OS 2200 and so forth.

And as a reminder, from a famous WG14 and WG21 member, and former Rust contributor,

"To Save C, We Must Save ABI"

https://thephd.dev/to-save-c-we-must-save-abi-fixing-c-funct...

Uh, GNU does define an OS ABI...

Documentation....

You're gonna have to be more verbose, sorry.

It is kind of obvious, GNU OS ABI documentation.

What about it?

Your assertion.

I'm gonna need you to write full sentences because I still have exactly 0 clue what point you're making

Nah, you are on a trolling vibe.

All you need to do is clarify the first post on the thread with documentation.

Sometimes I do miss these Usenet like rounds.

I genuinely did not understand that you were asking for documentation.

I believe GNU follows the SysV ABI on amd64: https://refspecs.linuxfoundation.org/elf/x86_64-abi-0.95.pdf

Which GNU OS was the point, there isn't any GNU OS, unless we talk about Hurd I guess.

That ABI is known for UNIX systems yes, although it isn't uniformly implemented, and again comes back to the whole OS written in C, like UNIX naturally is given it is the reason C exists in first place.

Which as mentioned, is not followed by every UNIX variant anyway.

Now GNU OS, still don't know what that is, Linux kernel with glibc?

Then again, using other libc doesn't require the same ABI.

System V ABI is pretty standardized.

UNIX, the reason why C exists.

Yeah, I'm not interested in decoding what you meant by your aphorism. Either say what you mean or don't.

Great, so lets save us both time.

Given OOP, the only ABI for an object is a vtable and that is left unspecified (for good reason) by the C++ standard.

COM said, "objects are cool and modules are cool so we need an ABI for objects" and specified a vtable. It isn't the only programming model which continued from the same observation: Python did too with its absolutely abysmal leaky PyObject ABI

SOM predates COM, and is even cooler, as it supported class inheritance, metaclasses and Smalltalk as target language.

On OS/2, Smalltalk enjoyed a role similar to what would be .NET and VB on Windows.

Sure you can use delegation with aggregation, and type libraries (nowadays .NET metadata), but still isn't as ergonomic.

COM designers were inspired by C++ but their goals were different. They were designing a language-neutral binary standard (aka ABI) and hence took C++'s approach and tweaked it for their specification.

Here is an earlier comment of mine with some details - https://news.ycombinator.com/item?id=49142987

And COM enforces lifecycle rules, transparent remoting, uniform activation (constructors basically), security primitives, marshalling, transparent async calls (yes, you can call any method asynchronously), and a ton of other things. It's actually very good and it's a shame most don't know about it and most of the rest sneer at it.

I agree!

(Though you cannot call just any method asynchronously; the interface's IDL needs specific annotations and the underlying object needs to implement ICallFactory.)

Just add async_uuid to your IDL locally and rerun MIDL. The server doesn't know or care that it's being called asynchronously. It's a pure client thing. If you're calling an IFoo in a different apartment, you can do the call asynchronously by changing your client configuration without the IFoo implementation having to be aware that anything is going on.

You can also use IRpcChannelBuffer3 to make the call asynchronously no matter what the IDL says. You'll have to write the proxy plumbing yourself, but it'll work, and you don't even need to rerun MIDL.

Well said.

COM was one of the best realizations of "Component-based Software Engineering" and Brad Cox's "Software-IC" model. It was a binary standard and so components could be written in any language and yet be assured of perfect interoperability provided you followed all the rules/conventions. There was a bunch of boiler-plate for the framework itself but once you understood it, everything was smooth sailing in your language of choice.

One of the things i always advise people is not to focus only on the current way of doing things but to study older well-known libraries/frameworks/architectures/kernels/etc. to really understand "Software Engineering" from many perspectives. That is where insight comes from and real understanding happens.

For people interested in understanding COM, see the classic Essential COM by Don Box.

COM is heavily used in Windows today, no need for past tense.

Not directly by app developers. Everybody uses only WinRT/.NET based managed languages/runtimes/libraries/frameworks etc.

Who uses WinRT besides Microsoft employees, or APIs that are only available in WinRT like MIDI 2.0?

Regardless, WinRT is COM.

It only adds IInspectable as additional interface alongside IUnknown, and .NET metadata files instead of type libraries.

I said "WinRT/.NET based managed languages/runtimes/libraries/frameworks" implying that it is the higher-level constructs that are used by app developers.

If you were to interview any "standard" windows programmer today i can almost guarantee they know nothing about COM even though WinRT itself is an evolution of COM and .NET also provides COM interop. Only the older senior programmers who have used it directly know of it.

Ignoring C++ devs on purpose?

Also no one uses WinUI, if that is what you are implying with WinRT, only Microsoft employees on Windows team forced to deal with that clusterfuck framework after Project Reunion pivoted into WinAppSDK.

What do you think Qt, VCL, FireMonkey, ImGui, wxWindgets, JUCE make use of?

What are you talking about? None of it is relevant to the point i am making.

Also you do not need COM to program in C/C++ directly on Windows API since they are all C interfaces anyway. COM is just an architecture you choose to use or not depending upon your needs.

Ever since managed languages/virtual machines became "standard" programming platforms most Windows programmers only write to these. Everything is at such a high-level now that only curious programmers delve deeper into the rabbit-hole.

You definitely do, unless you are stuck in Windows XP view of computing world.

It is like using ReactOS instead.

All Windows APIs since Vista are delivered via COM.

How much Windows development are you doing actually?

Not that much apparently.

Also someone has to surface Windows APIs to managed languages, they don't appear by magic.

> All Windows APIs since Vista are delivered via COM.

Nope, This right here tells me you do not have much actual programming experience in Win32/Win64 apis nor of Windows Internals. The C-style windows apis are from kernel32/gdi32/user32/etc. user-mode dlls which call into intermediate ntdll/win32u dlls which then calls into kernel mode ntoskernel.exe/win32k.sys. With modern Windows there are another layer of abstractions with "Windows API Sets" (https://learn.microsoft.com/en-us/windows/win32/apiindex/win...) which decouple those user api from their actual implementation dlls.

COM is at user-mode and so interfaces to the first set of dlls only. Since the windows api breadth is vast not all of them are exposed via COM. WinRT uses/enhances classic COM but also calls Win32 api as needed. So Win32 api and WinRT api coexist with the latter providing the "modern" way to api access - https://en.wikipedia.org/wiki/Windows_Runtime But because WinRT is oriented towards secure sandboxed apps many low-level Win32 apis dealing with memory management, thread/process manipulations, system hooking etc. are limited/removed entirely from WinRT api.

For your edification see this detailed older article; Turning to the past to power Windows’ future: An in-depth look at WinRT - https://arstechnica.com/features/2012/10/windows-8-and-winrt...

Your claims must be backed up with references before asking others about their experience. I only see you name dropping and quoting historical data in all your comments which is not relevant to discussing/understanding anything.

Every new API since around NT5.2/6.0 era had to get explicit waiver to be provided as anything other than COM interface. Some additionally have weird hybridization going on IIRC that lets you have some kind of faster path but still has COM exposed (IIRC that's only DirectX?)

That is for new APIs, not extensions of existing APIs.

IIRC, most non-COM methods you'll see for those are wrappers around COM calls. Back when WinRT was promoted as the next big thing, once you got outside of the small circle of WinRT propaganda there was explicit acknowledgement that all of the APIs are actually COM based enabling continued development of non-.NET compilers.

Of course underneath it all, especially once you get to userland-kernespace interactions you go back to combinations of IOCTL, memory mapped IO, some basic object calls and undocumented fun of OpenVMS-derived IPC, but a considerable chunk of that is not exposed or documented for non-blessed programmers.

Right; the links i provided above explain it in more detail.

WinRT is made up of component dlls and therefore exposes functional interface bundles. But the implementation of these components mostly took the path WinRT -> Win32 api -> ntdll (aka "Windows Native api" which is undocumented). But since MS wanted to push WinRT they allowed some paths to be WinRT -> ntdll. AFAIK this is what the mandate/waiver was for.

So you have the current situation that some Win32 apis have no analogues in WinRT and vice-versa. The DirectX path uses "Nano-COM" (a stripped down version of COM) which calls into user-mode runtime dlls which then takes another path to the kernel.

Lets not even get into how .NET projections/interfaces/wrappers do their job ;-)

The key point to know is that COM is only a packaging architecture/framework/middleware for structuring at a user-mode binary level.

It is all quite elaborate and complicated and so making a one-line statement like "everything is COM" is sheer cluelessness.

Which shows how many devs never learnt about Windows development past Windows XP, given that COM is nowadays the main API delivery mechanism, as they redid Longhorn ideas in C++.

> and a very specific vtable layout for COM objects

Fun fact: Microsoft developed COM based on how Zortech C++ virtual functions worked. (It predated Microsoft C++ by years.)

You mean... using a vtable? Are you claiming to have invented the concept of vtables?

Yes, the vtable. Microsoft at the time equipped their COM team with Zortech C++ and so they used the Zortech vtable. No, I did not invent the concept of vtables.

AFAIK, the first design for implementation of vtables was by Bjarne Stroustrup in his paper Multiple Inheritance for C++ in 1986/1987.

Pdf at https://www.google.com/goto?url=CAESawHuR6pN4SDa5tLkqrSuoaot... and reissue at https://www.researchgate.net/publication/2396782_Multiple_In...

I read this in one of the documentation books (still have it somewhere) which came with SCO Unix (don't recall the date). This book contained articles on C++ only and was an introduction to the language and compiler (since the language was very new). I also remember that it had an article on Template mechanism.

Next was Jan Gray from Microsoft's excellent 1994 article in MSDN titled C++: Under the Hood which explained class layouts/vtables/etc. It is still a excellent read today. Pdf at https://www.google.com/goto?url=CAESewHuR6pNy2XLp9fzgwckhWpg...

Finally of course you have Inside the C++ Object Model by Stanley Lippman explaining everything in detail.

More like the specific design of a vtable

And Delphi came along and used the same layout, so an interfaced Delphi object has beautiful COM interop.

Which is also used by C++ Builder.

In a tragic way, given the prevalence of COM in Windows, especially since Vista, someone at Microsoft should offer a few Delphi and C++ Builder licenses to the teams responsible for doing COM tooling.

Because MFC/OLE, ATL, WTL, WRL, C++/WinRT all have their sharp edges and could be so much better, if someone actually cared about productivity and framework ergonomics.

Does the C++ standard guarantee binary ABI? Isn't that something that compiler and library contributors handle separately? So I think the arguments you see online are more accurate—vendors are the ones maintaining it.

In other words, the ABI we rely on today isn't really part of the C+ standard—it's more like the Itnaium C++ ABI or the MSVC C++ ABI.

In the end, I think the ABI stays stable because of community conventions established by compiler vendors.

It doesn't, and FWIW the article says that. A meta note: ABI stability is a holy war inside the C++ community. Those against it argue that it's holding back real language evolution. Herb Sutter is even working on his own C++ offshoot that shows what could be.

Afaik Herb Sutter largely stopped working on Cppfront a few months ago, and was never quite supposed to be a real thing, just a personal playground (or so Sutter claims at this point, anyway).

Afaik Carbon is at this point the only attempt at a successor language that's still going?

Probably, but note that Google says they are they main customer anyway, and it is an experiment above anything else.

Anyone that can reach out to Rust, Go, Java, C#, whatever, should do that preferably.

There are also some efforts to tame existing C++ with profiles, and replacing UB with erroneous behaviour, and that's about it.

No, ABI is not part of the C++ standard, although there are parts of C++ that do go "this is an ABI thing" (e.g., [[no_unique_address]]).

ABI is more of OS-level thing. Most systems these days follow the SysV ABI, which is largely defined by the hardware manufacturers via the processor-specific supplements (the x86-64 one is here: https://gitlab.com/x86-psABIs/x86-64-ABI). These largely delegate C++-specific conventions to the Itanium C++ ABI (which they likely directly reference), although the ARM ecosystem uses a somewhat different layout for the exception handling tables. Microsoft uses a different ABI for both the underlying C ABI and for the C++ compatibility layer built on top of the C ABI.

One of the issues that crops up is that vendors end up needing to add extensions to the ABI for various reasons, and these extensions tend to end up being incompatible, since they're added before they've had a time to be standardized. 16-bit floats is a particular historical bugbear, as is the C23 _BitInt stuff.

Thank you for the kind comment. I'll make a note of it and look into it.

One more addition.

There is often a distinction made between "language ABI" and "library ABI" in documentation which is highly confusing.

The answer is: library API + compiler ABI = library ABI

GNU libstdc++ ABI Policy and Guidelines - https://gcc.gnu.org/onlinedocs/gcc-9.2.0/libstdc++/manual/ma...

No, neither does the C standard.

With exception of Swift, D, and bytecode based languages, the ABI is left to the vendors.

For those that think ISO/IEC 9899:2024 PDF has anything related to ABI, the actual ABI used by C compilers, is the OS ABI, if the OS was written in C to start with, and naturally this overlaps quite nicely with UNIX like OSes, and Windows.

It isn't like that on other platforms that decided to either use other systems languages, or expose their OS APIs in a different way, e.g. mainframes, micros, Android, ChromeOS, WebOS,...

Isn't this one reason why C is the de facto FFI? OSes happen to use it in their OS ABI and so everything is forced to speak it

OS written the C, which is what many overlook.

If you are on e.g. z/OS you would be using ILE, Integrated Language Environment, on ChromeOS JS/WASM, on Android either DEX or JNI,...

Android is Linux and it's written in C. Bionic is still a posix libc.

Android uses the Linux kernel as implementation detail, it could be changed with Zircon if Google cared, bionic is partially POSIX, only the required for ISO C, ISO C++, ART and NDK native APIs is exposed.

80% of Android APIs are exposed via Java, and even native ones have to use JNI.

Additionally a lot of the APIs if you go underneath, including native code components, are exposed through few variants of Binder (BeOS-derived IPC for the uninitiated) including userspace drivers.

Binder APIs very much do not follow C ABI other than the part where you deal with wrapper around ioctl()

This is entirely unrelated to the point made.

Binder is just C++ with a thin IPC compiler, more or less similar to any IPC idl. Could be sunrpc, but Google wanted a c++ wrapper.

But in any case this is completely irrelevant, because there is no such thing as a "userspace driver" (unless you're speaking about hurd or plan9). Drivers are by definition kernel-space code (which on Linux is either built into the kernel, or is a loadable kernel module). Those binder wrappers are just a convenience layer, underneath it's all still Linux, /dev/video0, v4l2, and so on.

Linux actually has pretty long history of userspace drivers, and multiple kernel drivers have as their main if not only logic a way to link hardware interface with userspace app, or just providing transport interface. And android heavily uses such interfaces from day 0, which is why you have separate binder instance on Android for drivers.

As for Binder being C++, no it isn't. Historically it has connection with C++ due to how it was used, but like other IPC it's not bound to C++, and arguably given how it is structured compared to some other options there's even less C++ specific stuff in it.

Also, IIRC the use of Binder predates acquisition of Danger (?) by Google

It's just demagoguery. Windows could switch to Linux kernel if Microsoft cared enough to invest into Wine. But this will not happen any time soon.

All smartphone drivers would have to be rewritten for a different kernel.

One reason. The other is C is easy. Any language can implement it and have confidence it can work. (there are other options that are just as easy, but C won). C++ is much harder because of things like function overloading and exceptions - none of it is difficult, but there is a lot more to do and thus more opportunities for bugs.

I would go far as to say that modern C++ is an Eldritch horror of a programming experience

Unfortunately no one is going to rewrite LLVM, GCC, CUDA, DirectX, Metal, SYSCL, V8,... into something else.

Same applies to C and UNIX/POSIX, OpenGL/Vulkan/OpenCL/NVN/LibGNM(X),...

Thus analysers and language subsetting tools, enforcing language style guides and safer coding practices.

Oh come on; that is BS and you know it.

It is complex/complicated/baroque but for the power, flexibility, real-world-usage that it provides it is worth the effort. The mistake that people make is to try and learn it all at the same time which overwhelms them.

Use a good book like Discovering Modern C++: An Intensive Course for Scientists, Engineers, and Programmers by Peter Gottschling and you should have no problem at all.

For context: I have over a decade of professional experience with c++ in large scale software projects and it is by far the language over which I have the most mastery.

And yet you make such statements ;-)

I often see on HN these sorts of comments whenever C++ is brought up and it really needs to stop. It is wrong, adds no value to the discussion and does a grave injustice to the programming community.

We should be encouraging programmers to learn C/C++ since just knowing those two languages enables one to program MCU/embedded/desktop/server machines across all layers of software from apps/scientific software/OS/system software/bare-metal in the real-world. One can imagine the job opportunities for a programmer with such a skillset as market needs change. With both Hardware and Software evolving so rapidly nowadays, C/C++ are the one constant interface language in the industry.

I would encourage programmers to learn c++ because it's important, but I'll never lie about what it's like to actually use it

You don't have to lie or play-up/play-down the language; just say it as it is with no painting of dragons and fear-mongering.

It is just another language and not rocket science.

Easy and still expressive, so you can map many things to C. And while the language standard does not define an ABI it takes a lot of care not to require any ABI break.

> With exception of Swift, D, and bytecode based languages, the ABI is left to the vendors.

I don’t see exceptions there. Apple defines the ABI of Apple’s Swift’s implementation, Walter Bright (or his team) defines that of D, Python defines its ABI, etc.

If you were to write a Swift/D/etc compiler, you’re free to define your own ABI. Disadvantage is that you would give up linking with code compiled by the other compiler (workarounds such as pragmas, C++’s extern "C", are possible)

You're missing who defines the language standard.

You're free to do whatever you feel like on your implementation, including not being compliant with the official language standard.

The C++ standard doesn't say anything about binary ABI (as you say generally either Itanium or MSVC, though there are other options they are rare). However the people who write that standard are very sensitive to the vendors and users of C++ who depend on a stable binary ABI. Thus they take extreme care to ensure that no change to the standard breaks binary ABI.

The C++ standard did force gcc to break the ABI of std::string (copy on write was banned - for good reason). They then watched the gcc community work through 10 years of pain to make the transition. They are also well aware that python 3 broke compatibility with Python 2 - and again it resulted in 10 years of pain for the python community to deal with that. With this history there are a lot of experts strongly against any breaking change. It might happen anyway, but only with strong justification and likely an attempt at a migration of some form (what? there are a lot of examples of migration plans that don't work that they are aware of)

Again, it is not because of convention, it is because of painful experience from those who break it.

> They then watched the gcc community work through 10 years of pain to make the transition.

During the same time the MSVC compiler broke ABI multiple times and the world didn't end.

The fact that breaking ABI was such a shitshow for libstc++ is more related to the way C++/ABI/SOs is/are handled in Linux...

In the time since that VC++ has switched to a frozen ABI model that’s much closer to what Linux distros have. Each version of VC++ being ABI-incompatible and requiring each application to be built with a single version and ship its c++ runtime had some significant pros and cons.

Microsoft never said it outright, but I’ve always assumed the major con that made them switch was that people didn’t buy new Visual Studio versions because upgrading VS required also upgrading all of your binary blob closed-source dependencies.

The reason the world didn't end is because people refused to upgrade Visual Studio, in particular Microsoft's bigger customers. Microsoft surveyed their users for why people weren't making the switch to VS 2013 despite all the "nice" and "new" features introduced and the predominant reason was due to all the ABI breakages.

For VS 2015 they adopted a policy of having a stable ABI and have preserved ABI compatibility for over 10 years now [1].

[1] https://devblogs.microsoft.com/cppblog/binary-compatibility-...

The part where Windows applications built with C++ generally do not share C++ ABI DLLs with each other and the operating system except for certain well isolated ones is major part of it.

GCC changing C++ ABI usually meant horrific time for every C++ application on Linux unless you use something like Nix so you don't chance loading incompatible binaries.

When the C++ standards committee make decisions crippling the language to preserve ABI (see std::regex), it's defacto part of the standard.

As always, read wikipedia and follow links as needed - https://en.wikipedia.org/wiki/Application_binary_interface

One nice thing to remember is for projects you control, you can dictate the ABI. There isn't a law enforcement bureau that will arrest you for passing parameters in whatever register suits you. You just have to imagine life differently than they did in the 70s when they dreamed up dynamic libraries for reasons that are largely no longer valid.

This technique is often used for security-sensitive projects involving binary protection, obfuscation etc.

Program Obfuscation via ABI Debiasing (pdf) - https://www.google.com/goto?url=CAESYgHuR6pNsi8_4x_I2WBy7lmf...

PS: James Coplien in his excellent Advanced C++ Programming Styles and Idioms shows many techniques one of which is actually replacing vtable entries at runtime to mimic features from more dynamical runtime languages (eg. Smalltalk).

GNU libstdc++ "Dual ABI" issue - https://gcc.gnu.org/onlinedocs/libstdc++/manual/using_dual_a...

I understand that Move Semantics led to better implementations and that this is why things changed. I know that in this case, problems arose when transitioning from C++98 to what's commonly called Modern C++ (C++11). However, I'm not very familiar with what the specific issues were with COW. If you happen to know of any documents that describe this problem in more detail, I'd appreciate it if you could let me know. I'm sorry for always asking questions. I've looked up a few documents, but since you clearly know much more about this than I do, I'm asking to study it myself as well. If it's too much trouble, feel free not to reply.

The changes to std::string were made because "the [pre-C++11] definition of basic_string allows only very limited concurrent access to strings. Such limited concurrency will inhibit performance in multi-threaded applications." The proposed fix was to make "all iterator and element access operations safely concurrently executable." [0]. This involved limiting the circumstances in which iterators could be invalidated, which among other things meant that non-const operator[] could no longer invalidate iterators, which meant copy-on-write was no longer a viable implementation [1].

[0]: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n26...

[1]: https://stackoverflow.com/q/12199710

Thank you for the kind comment. !

Asking questions is always a good thing so don't worry about it. It makes us think and check our own understanding too, so we also learn something :-)

A nice succinct post by Pavel Khaipov; Copy-on-Write in Modern C++: Why the Standard Dropped It - and Why Qt Still Uses It - https://www.linkedin.com/posts/khaipov_copy-on-write-in-mode...

A detailed post by Andrii Nikishaiev; Why Copy-On-Write (COW) removed in C++? - https://www.linkedin.com/pulse/why-copy-on-write-cow-removed...

The std document on Concurrency Modifications to Basic String - https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n25...

Finally, C++11 tightened the specification for std::string viz. contiguous memory, some methods must have O(1) access time, when references/iterators/pointers can/cannot be invalidated etc. which doomed COW.

I'm sorry for always bothering you, senior. I've learned quite a lot from the books you recommended. I've especially been studying low-level topics lately, and the books you suggested have been very helpful. Have a great day.

I am glad that you found my comments useful. I use HN as my "blog" and generally include references to books/papers/articles as a pointer to further information on a topic. You will find links to lots of books if you search my comments for key phrases (eg. embedded, book etc.).

One thing i try to do is point people to lesser-known but very good and advanced books. I generally find that most HN book suggestions refer only to popular books most of which are just so-so while many great books remain unknown which is a shame. For example how many people know of Operating Systems in Depth by Thomas Doeppner? It has lots of helpful illustrations, compares concepts in both Linux/Windows implementations and is an all-round great book. I really understood the nuances of signal handling between kernel/user modes when i read it here.

C++ famously has no ABI, it is the implementations of C++ that have ABIs. The reason the situation is so bad is that it's considered explicitly out of scope of the language (it targets an abstract machine).

Neither of the ISO C and ISO C++ standards define an ABI.

Right, C doesn't specify an ABI either but (many) OSes do and they provide their ABIs in C so C became the defacto ABI of a given platform.

Yes and no. The Linux kernel barely interacts with the C calling convention ABI (vDSO), and you can interact with the OS without going through a C library. I guess the layout of some structs are exposed through some binary protocols, but most of them are not C function calls.

Reply on news.ycombinator.com