ranges in phobos increase exe size

[WalterBright]

not much to be done about that.

[MartinNowak]

:-(

not much to be done about that.

Well you don't use datetime at all, so we need to find out how to keep that dead code out of the binary.

[WalterBright]

It may have been an accumulation of things that this change just put past the trigger.

[dnadlinger]

--gc-sections seems to be quite an effective way to work around the codegen issues. I can't easily check with the 2.068 Phobos on LDC right now, but on 2.067 the Linux x86_64 executable for that test is only 30% the size of DMD's.

[MartinNowak]

We had little to no effect of gc-sections with dmd, mostly because it already uses multilibs.
I think the root problem is that the ModuleInfo hard-links imported modules and all classes within a module. The classes vtable the drags all the member functions and lots of other stuff.
I spent some time to make weak external references work, but there were concerns to use that b/c it might break Object.factory.
#4638

[dnadlinger]

Well, all I'm saying is that LDC's executable is much smaller, and --gc-sections certainly contributes to that.

[WalterBright]

Object.factory in hindsight was a mistake. It should only work with classes marked with export. It's probably worthwhile to fix that.

[MartinNowak]

Might make an interesting story for the next release, though export still needs a lot of fixes. Also there are more important topics than binary size.

[mihails-strasuns-sociomantic]

on 2.067 the Linux x86_64 executable for that test is only 30% the size of DMD's

Still about the same ratio for 2.068 (with both linked to shared phobos library)

-rwxr-xr-x 1 mihails mihails 139K Aug 20 18:34 hello-dmd
-rwxr-xr-x 1 mihails mihails  47K Aug 20 18:33 hello-ldc

(for a simple app that uses some template utilities)

void main ()
{
    import std.stdio, std.algorithm;
    writeln([ 1, 2, 3 ].map!(a => 2*a));
}