Re: It's time to kill statistical profiling

On Sun, 20 Jun 2021 at 17:08, Steve Kargl
<[email protected]> wrote:
>
> Are there plans to replace the _p.a with something
> that allows profiling code?  If not, are you upstreaming
> patches to GCC to disable -pg?  AFAIK, 'gcc10 -pg ... -lm'
> causes gcc to look for at least libc_p.a and libm_p.a.

I believe this is how GCC operates indeed, but I'm not aware of the
implementation or specific details. Clang behaves as you describe --
here is all behaviour change based on OPT_pg, from
lib/Driver/ToolChains/FreeBSD.cpp:

    if (!Args.hasArg(options::OPT_shared)) {
      if (Args.hasArg(options::OPT_pg))
        crt1 = "gcrt1.o";
      else if (IsPIE)
        crt1 = "Scrt1.o";
      else
        crt1 = "crt1.o";
    }
...
      if (Args.hasArg(options::OPT_pg))
        CmdArgs.push_back("-lm_p");
      else
        CmdArgs.push_back("-lm");

and similar for -lgcc_p, -lgcc_eh_p, -lpthread_p, -lc_p, -lc++_p, -lstdc++_p.

This support was initially introduced upstream in:

commit 66f2276aee67738d116d26494d8a78fc6528586b
Author: Roman Divacky <[email protected]>
Date:   Thu Feb 10 16:59:40 2011 +0000

    Adjust the object files to be linked in when mcount profiling
    is specified in the FreeBSD linker driver.

    llvm-svn: 125285

and was implemented for OpenBSD later that year, but no other OS does
this. I'm not sure exactly what happens elsewhere - my guess is that
the runtime support exists in libc but it is built without -pg so
coverage will not include libc itself.