Awhile back, I wrote some C# code that interacts with C/C++ code by reading struct layouts directly. This is unsafe because C structs don't have guaranteed layouts unless you use nonstandard pragmas. So I had the C code bypass regular struct {} and lay out data manually by pointer arithmetic, making it consistent and safe. This worked great, until I discovered that booleans are a different size in C and C++. I fixed this too, but only by doing horrible things I'm only 95% certain actually worked
0 Boosts
0 Likes
@beatrix *thinks*
Actually… I think protobuf would be a good fit for our situation. I'm not sure why we didn't go with protobuf. I think maybe it's because I thought our needs were simple and protobuf was complicated, but our needs have grown more complex since this project was started.
I just asked the "client", who says that they've actually surveyed C# implementations of Protobuf and as of last survey they all seem to be very hard to use compared to how easy serialization normally is in C#.
@mcc “C structs don’t have guaranteed layouts”
So how does repr(C) work, then? Does it only work with a specific set of compilers? Is it based on platform conventions?
@bytex64 I may be about to say something horribly wrong, but the thing about repr(C) is that it guarantees you are using the same layout as the platform's standard C layout, but DOES NOT convey any information about what this platform is. This means repr(C) can talk to repr(C) but neither one can totally reliably talk to C# which is using pure pointer manipulation.
P/Invoke *can* talk to repr(C)… unless you fall into the hole with the disaster around boolean sizes.
@dotstdy @bytex64 Okay so on the one hand you are entirely right, but on the other hand: Microsoft, creator of this problem, is both the platform defining body who decides the ABI the C compilers are downstream from, AND the author of the currently most common compiler on the platform, AND the body who both defined and implemented the entire C# language
@dotstdy It's not that simple. P/invoke by default maps C# bool to win32 Bool32 *in arguments to functions*, meaning if you want to write C++ code containing an extern "C" function with a "bool" argument, you need to [MarshalAs(UnmanagedType.I1)] the definition on the C# side.
Which isn't that complicated, but it is *very nonobvious*, and we didn't realize it was true until one *year* after the first bug introduced to this codebase caused by not realizing it.
@ireneista @mcc @bytex64 IME you only run into ambiguities when you break alignment. If you keep all fields aligned, packing is a non-issue. IMO that's far better than languages like java that actively prevent you from knowing struct layouts.
@hyc @ireneista @bytex64 …which… alignment
@ireneista @bytex64 How do you read a struct without knowing its layout?
@mcc @bytex64 you guess. you start by guessing the fields occur in the order they're declared, which turns out to almost always be true.
if you dig into the details of what Rust is doing, it turns out each Rust type has a specific alignment that repr(C) uses, and those are just unchangeable from Rust's perspective. there are additional declarations that appear to change it but they explicitly cannot make the kinds of changes that map to actual variation in C compilers. we ran into this recently with our Vulkan thing.
I'm thinking about this because I'm having to add to that code right now and it's miserable. I'm not sure why I'm telling you this. I guess I want to stress there are reasons I'm now on team "Nothing but Rust. Nothing but Rust forever. No more C/C++. Rust. What? Okay, Go or Zig are probably fine. Just no C"
@mcc one day someone will invent a C compiler that lets you write C and native Rust together and they will call it Crust
And then... no one will be laughing anymore
@brandonscript this would make me sad
@mcc to be fair if you want layout to be predictable in rust you also need to use C rules...
@mcc The problem is your code won't work on a PDP-10 or a DEC Alpha.
@mcc isn’t this what protobuf and frens are for..?