How does ld and gold deal with C++ duplicate templates?

[Ian Lance Taylor]

On Wed, Sep 16, 2020 at 12:38 PM sotrdg sotrdg <euloanty@live.com> wrote:
>
> But gold linker is said to be removed in the future since no one is willing to maintain it any more. BTW, I am using MinGW on Windows where gold does not seem to be available.

Seems a pity, but OK.  It's true that I don't have time to maintain
it.  And it's true that it doesn't work on Windows.

> Does ld provide the same options?

Not to my knowledge.

> And why these optimizations are not applied by default?

Because they measurably increase link times.

Ian


> From: Ian Lance Taylor
> Sent: Wednesday, September 16, 2020 15:23
> To: sotrdg sotrdg
> Cc: binutils@sourceware.org
> Subject: Re: How does ld and gold deal with C++ duplicate templates?
>
>
>
> On Wed, Sep 16, 2020 at 2:26 AM sotrdg sotrdg via Binutils
> <binutils@sourceware.org> wrote:
> >
> > Like the bug I reported a month ago.
> >
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96577
> >
> >
> >
> > https://godbolt.org/z/haG4E7
> >
> > https://godbolt.org/z/M8e7Kb
> >
> >
> >
> > First:
> >
> >
> >
> > #include<vector>
> >
> > #include<array>
> >
> > #include<algorithm>
> >
> >
> >
> > template<std::contiguous_iterator Iter>
> >
> > void f(Iter a,Iter b)
> >
> > {
> >
> >     std::sort(a,b);
> >
> > }
> >
> >
> >
> > void g(std::vector<int>::iterator b,std::vector<int>::iterator e)
> >
> > {
> >
> >     f(b,e);
> >
> > }
> >
> >
> >
> > void h(int* b,int* e)
> >
> > {
> >
> >     f(b,e);
> >
> > }
> >
> >
> >
> > Second:
> >
> >
> >
> >
> >
> > #include<vector>
> >
> > #include<array>
> >
> > #include<algorithm>
> >
> >
> >
> > template<std::contiguous_iterator Iter>
> >
> > void f(Iter a,Iter b)
> >
> > {
> >
> >     std::sort(a,b);
> >
> > }
> >
> >
> >
> > void g(std::vector<int>::iterator b,std::vector<int>::iterator e)
> >
> > {
> >
> >     f(std::to_address(b),std::to_address(e));
> >
> > }
> >
> >
> >
> > void h(int* b,int* e)
> >
> > {
> >
> >     f(b,e);
> >
> > }
> >
> >
> >
> >
> >
> > The two sort function should generate exactly the same binary. However the first which is none to_address version generates twice as much as code as the second one.
> >
> > Will the GNU ld remove the duplication definitions?
>
>
> For gold, see the --icf option.
>
> Ian
>
>