Multiple issues with ClangSharp

[kaby76]

I'm looking to replace my SWIG generated and ManagedCuda wrappers for CUDA with a ClangSharp-generated wrapper (SWIG rules are just to arcane). ClangSharpPInvokeGenerator looks promising, but there are several issues:

1. ClangSharp seems to be in two locations: https://github.com/Microsoft/ClangSharp and as a sub project in LLVM https://github.com/Microsoft/LLVMSharp/tree/master/tools/ClangSharpPInvokeGenerator
   Which is it? I assume https://github.com/Microsoft/ClangSharp is the correct source as it was more recently updated.

2. In cuda.h, there are many typedef declarations, for example:

typedef enum CUaddress_mode_enum { CU_TR_ADDRESS_MODE_WRAP = 0, /**< Wrapping address mode */ CU_TR_ADDRESS_MODE_CLAMP = 1, /**< Clamp to edge address mode */ CU_TR_ADDRESS_MODE_MIRROR = 2, /**< Mirror address mode */ CU_TR_ADDRESS_MODE_BORDER = 3 /**< Border address mode */ } CUaddress_mode;

The generator handles the enum, but not the typedef. A "using ... = ..." might be appropriate.

3. The generator does not handle C unions, nor anonymous types well. In cuda.h, there is the following decl:

union { struct { CUarray hArray; /**< CUDA array */ } array; struct { CUmipmappedArray hMipmappedArray; /**< CUDA mipmapped array */ } mipmap; struct { CUdeviceptr devPtr; /**< Device pointer */ CUarray_format format; /**< Array format */ unsigned int numChannels; /**< Channels per array element */ size_t sizeInBytes; /**< Size in bytes */ } linear; struct { CUdeviceptr devPtr; /**< Device pointer */ CUarray_format format; /**< Array format */ unsigned int numChannels; /**< Channels per array element */ size_t width; /**< Width of the array in elements */ size_t height; /**< Height of the array in elements */ size_t pitchInBytes; /**< Pitch between two rows in bytes */ } pitch2D; struct { int reserved[32]; } reserved; } res;

The generator bypasses unions completely. For anonymous structs (e.g., contained in the above union), the generator assigns the name "" as the anonymous type name, but it should make it unique, eg "" + int_id++, because after creating the type, it tests to see if the type was declared previously, and thinks that it was already declared.

I'm planning on a patch for ClangSharp to fix these problems in the next couple days as I'd like to get the tool to work with CUDA asap so I can get back to work on my C#/CIL compiler for GPUs.

Ken Domino

[kaby76]

https://github.com/Microsoft/ClangSharp/blob/master/ClangSharpPInvokeGenerator/Program.cs#L118

The size of "arr" is hardwired to 3 for the call to clang.parseTranslationUnit2; it should be arr.Length. "-I" is an option (not parsed as a required parameter). So without -I parameter, the generator crashes. There are probably similar issues with other parameters.

For the following decls,

#pragma once

typedef int IntType;
extern "C" {
__declspec(dllexport) void foo(IntType v);
}

the ClangSharp generator outputs

namespace ZZZ
{
using System;
using System.Runtime.InteropServices;

public partial struct IntType
{
    public IntType(int value)
    {
        this.Value = value;
    }

    public int Value;
}

public static partial class XXX
{
    private const string libraryPath = "ValueTypes";

    [DllImport(libraryPath, EntryPoint = "foo", CallingConvention = CallingConvention.Cdecl)]
    public static extern void foo(IntType @v);

}

}

Although this code works, it would be nice if it handles base types like int, long, etc., directly instead of creating a struct value type.

[kaby76]

I've decided to write an entirely new P/Invoke generator based on Clang AST matchers and code templates. Therefore, I am closing this issue.