Old code no longer working on any DMD compilers

Jonathan M Davis newsgroup.d at jmdavisprog.com
Fri Sep 6 07:03:21 UTC 2019

On Thursday, September 5, 2019 11:59:30 PM MDT Jamie via Digitalmars-d-learn 
> On Friday, 6 September 2019 at 00:41:12 UTC, Jonathan M Davis
> wrote:
> > On Thursday, September 5, 2019 6:24:07 PM MDT Jamie via
> >
> > Digitalmars-d-learn wrote:
> >> /home/jamie/.dvm/compilers/dmd-2.077.0/linux/bin/../../src/phobos/std/m
> >> ath .d(3702): Error: fmodl cannot be interpreted at compile time,
> >> because it
> >> has no available source code called from here: called from
> >> here:
> >> fmod(cast(real)to(n),
> >> cast(real)to(2))
> >>
> >> Thanks
> >
> > Based on that implementation, fmod would work at compile time
> > if you compile on Windows and use Microsoft's runtime instead
> > of Digital Mars runtime (so, if it was compiled as 64-bit or it
> > used whatever the flag is to force the COFF format instead of
> > OMF for 32-bit). It won't work on any other platform at compile
> > time, because core.stdc.math.fmodl is a C function, and C
> > functions cannot be called at compile time.
> >
> > By any chance were you compiling this on Windows previously
> > rather than Linux? If so, that's likely why the code worked
> > before and doesn't now. If you were always compiling on Linux,
> > then I don't know why you were able to compile your code
> > before. A quick glance at the repo history shows that fmod has
> > had that same implementation since 2012, so there should be no
> > way that it was callable on any Linux system even 6 years ago,
> > let alone 6 months ago.
> >
> > - Jonathan M Davis
> Thanks for the reply. I haven't used Windows in the past two
> years so that isn't the issue in my case. After your comments I
> looked deeper, and I think my issue is related to compile time
> values, and me misunderstanding default arguments for structs.
> import std.stdio;
> import std.conv: to;
> import std.math: fmod;
> void main()
> {
>      /// without default arguments
>      S s = S(64, 64);            // case 1: works
>      // static S s = S(64, 64);     // case 2: doesn't work
>      /// with default arguments
>      // static S s = S();           // case 3: works
> }
> struct S
> {
>      this(size_t N, size_t M) {
>      // this(size_t N=64, size_t M=64) {
>          f(N, M);
>      }
> }
> void f(size_t N, size_t M)
> {
>      int kj(int n)
>      {
>          return to!int(fmod(n, 2));
>      }
>      int k = kj(2);
> }
> If I'm understanding it correctly now, in case 1 the struct
> constructor isn't called at compile time, so the underlying C
> function (fmod) is not called. In case 2 because the struct is
> called in a static region, the constructor is called at compile
> time and fmod is called so it breaks. In case 3, with default
> struct arguments, I thought that the constructor I have defined
> was being called, however the default constructor was being
> called (this()) so fmod wasn't being called.
> The reason why my old code worked was because it used the default
> arguments and I wasn't actually calling the constructor I
> defined. When I removed the default arguments in the constructor
> and tried case 2 it obviously didn't work.
> Am I understanding correctly? Thanks

Structs in D do not have default constructors. They have init values, where
the value is what their members are directly initialized to. e.g.

struct S
    int i = 5;
    bool b;

void main()
    assert(S.init.i == 5);
    assert(S.init.b == false);
    S s;
    assert(s.i == 5);
    assert(s.b == false);

The init value must be known at compile time, so if you directly initialize
any members with anything, that expression must be evaluated at compile time
(meaning that you can't do something like use a C function to directly
initialize a member variable).

If you declare a constructor for a struct with no parameters, it will
_never_ be called. Similarly, a constructor with all default arguments will
never be called unless it's explicitly called with at least one argument.
And in fact, at least with the version of the compiler I have locally, the
compiler gives an error if I try to declare a struct constructor with either
no parameters or where all of its parameters have default arguments.

Under normal circumstances, in D, all variables are default initialized if
they're not explicitly initialized. So, when you have something like

void main()
    int i;
    static int j;

both i and j are default initialized to the init value of int (which is 0).
However, local variables are different from other variables in that if you
explicitly initialize them, that's done at runtime. So, if you have

void main()
    int i = foo();
    static int j = bar();

then foo will be called at runtime to initialize i, whereas because j is
static, its initial value must be known at compile time, and so bar is
called at compile time just like happens with member variables. Similarly,
enums must have their value known at compile time, so the code to determine
their value is run at compile time (though since an enum has no address, its
value essentially gets copy-pasted everywhere the enum is used).

If a struct's constructor is called to initialize a variable, then it's
called whenever is appropriate based on the variable being initialized. e.g.

void main()
    MyStruct a = MyStruct(42);
    static MyStruct b = MyStruct(42);

for a, the constructor is called at runtime to initialize the variable,
whereas for b, it's called at compile time. Similarly, if a member variable
is a struct, and it's directly initialized with a constructor call, then
that constructor call is done at compile time, whereas if it were done in
the constructor of the enclosing type, then it would be done whenever that
constructor was called. So, whether a constructor is called at compile time
or runtime depends entirely on the context in which it's used - but because
structs don't have default constructors, you'll never get a constructor
being called implicitly.

Unlike D structs, D classes _can_ have constructors with no parameters, but
that's because they're reference types, and their init value is null (and
even then, the constructor isn't a default constructor, because it still has
te be called explicitly - a class variable with no constructor call is
null). The reason that structs don't have default constructors is because D
requires that the init value of a type has to be known at compile time, and
having a piece of code run when a struct isn't explicitly initialized is
incompatible with that.

- Jonathan M Davis

More information about the Digitalmars-d-learn mailing list