dstep not converting `const T x[]` to `const (T)* x`

[dataPulverizer]

dstep not converting `const T x[]` to `const (T)* x`

[ciechowoj]

Could you give an example that reproduces this bug? E.g. a code snippet with the problematic types?

[dataPulverizer]

An example is converting a function like this:

int gsl_dft_complex_forward (const double data[], const size_t stride, const size_t n, double result[]);

The output you get is this:

int gsl_dft_complex_forward (<unimplemented> data, const size_t stride, const size_t n, <unimplemented> result);

[mihails-strasuns]

Works for me on master. Very likely that I have accientally fixed it as a side effect of a35b6ea

[ciechowoj]

Works for me on master. Very likely that I have accientally fixed it as a side effect of a35b6ea

On master it converts:

int func (const double data[], double result[]);

to

int func (const double[] data, double[] result);

@Dicebot, @dataPulverizer
Is it correct? Shouldn't it convert the C-arrays to pointers. I suppose D-arrays are slightly different, as there is a way to get the length of the array...

[jacob-carlborg]

@ciechowoj, yes I think it needs to translate the C arrays to pointers. A D array is twice as large as a pointer (the extra length as you mentioned).

[mihails-strasuns]

Yep: https://dlang.org/spec/interfaceToC.html#passing_d_array

[ciechowoj]

@jacob-carlborg

I'm trying to fix this, don't you know what the following comment from dstep/translatorType.d means?

        // extern static arrays (which are normally present in bindings)
        // have same ABI as extern dynamic arrays, size is only checked
        // against declaration in header. As it is not possible in D
        // to define static array with ABI of dynamic one, only way is to
        // abandon the size information
        return elementType ~ "[]";

[mihails-strasuns]

It is an old comment from me and I think it is wrong - I didn't know at the time of writing it that ref T[42] will use same ABI as T[42] in C. Of course it likely should have used * instead of [] anyway as mentioned by @jacob-carlborg

[mihails-strasuns]

(but original use case I had was for global extern variables, not for function parameters - please check if ABI is different there)

[ciechowoj]

Hmm, I do not know at the moment how to check it other way than experimentally.
Compiled with gcc -c lib.c && dmd test.d lib.o && ./test

// lib.c

const int tab[] = {
    42,
    42,
    24
};

Segmentation fault.

//  test.d

extern (C) extern __gshared const (int)[] tab;

int main()
{
    import std.stdio;

    writeln(tab[0]);

    return 0;
}

Segmentation fault.

//  test.d

extern (C) extern __gshared const (int)* tab;

int main()
{
    import std.stdio;

    writeln(tab[0]);

    return 0;
}

42

//  test.d

extern (C) extern __gshared const (int)[3] tab;

int main()
{
    import std.stdio;

    writeln(tab[0]);

    return 0;
}

It seems like the only way to translate this is to compute the size of an array somehow during translation...

[ciechowoj]

Hmm, worse if the array is declared with extern already in the header file...

extern const int tab[];

[ciechowoj]

On forum.dlang they suggested to declare an one element array instead.

https://forum.dlang.org/post/[email protected]

[jacob-carlborg]

What if you make it a pointer?

[dataPulverizer]

Sorry for my silence on this I think the conversion const T[] to const (T)* x is the right thing to do. The Case of the global variable is an edge case that should probably be put in the readme file to make users aware of this issue. Anyway global variables like that are seriously NASTY!

[ciechowoj]

What if you make it a pointer?

Segmentation fault.

//  test.d

extern (C) extern __gshared const (int)* tab;

int main()
{
    import std.stdio;

    writeln(tab[0]);

    return 0;
}

[mihails-strasuns]

Hm, how about extern (C) extern __gshared const int[0] tab? With some luck it may behave thesame as variadic length struct trick.

[ciechowoj]

Hm, how about extern (C) extern __gshared const int[0] tab? With some luck it may behave the same as variadic length struct trick.

dmd will not allow you to access directly, statically checking bounds, warning that the array is too short. On the forum I got an advice to use length 1, but there is the same issue with element access. The workaround is to use .ptr for indexing, e.g. tab.ptr[42].

[ciechowoj]

I suppose we could fix it for the time being by declaring an array with 0 or 1 length.

[dataPulverizer]

Is this something that needs fixing in D itself? Is the inability to consistently allow the replacement of const T x[] with const (T)* x not a bug?

[jacob-carlborg]

@dataPulverizer sounds like a bug to me.

[ciechowoj]

Is this something that needs fixing in D itself? Is the inability to consistently allow the replacement of const T x[] with const (T)* x not a bug?

@dataPulverizer sounds like a bug to me.

Just for curiosity, do you have an idea how it could be correctly solved/handled in D?

[dataPulverizer]

@ciechowoj nothing at all comes to mind

[jacob-carlborg]

Not sure if it's related. On OS X, DMD outputs the extern variable to __textcoal_nt section while compiling the C code, Clang outputs the extern variable to the __text section.