Jonathan Wakely <jwakely@redhat.com> writes:

> On Sat, 29 Apr 2023 at 11:24, Arsen Arsenović via Libstdc++ <
> libstdc++@gcc.gnu.org> wrote:
>
>> libstdc++-v3/ChangeLog:
>>
>>         * libsupc++/typeinfo: Switch to bits/version.h for
>>         __cpp_lib_constexpr_typeinfo.
>>
>>
> Does this change have an impact on compilation speed?
> With this change we'll be re-including bits/version.h multiple times in
> most compilations, and unlike other headers the preprocessor can't optimize
> away the second and subsequent times its' included, because the header
> isn't idempotent.
> It will only affect the preprocessing phase, which is a fraction of the
> time taken by template instantiation and middle end optimizations, but I'd
> like to know it's not *too* expensive before committing to this approach.
>

Whoops, I forgot to provide these measurements.

  ~/gcc/build$ time for x in $(seq 1 100); do _pfx/bin/g++ -x c++ - \
               -S -o /dev/null <<< '#include <bits/stdc++.h>'; done
  
  real	3m4.104s
  user	2m53.394s
  sys	0m10.132s

  ~/gcc/pristine-bld$ time for x in $(seq 1 100); do _pfx/bin/g++ \
               -x c++ - -S -o /dev/null \
               <<< '#include <bits/stdc++.h>'; done
  
  real	2m56.430s
  user	2m46.898s
  sys	0m8.967s

The top is a build of gcc with (a rebase of) this patchset, and the
bottom is the merge base.  The configuration for both was:

  ../configure --enable-languages=c++ \
               --disable-libsanitizers \
               --prefix="$(pwd)/_pfx"

This should imply that the difference is minor compared to compilation
time.

Of course, for preprocessing, the difference is far more significant
(same testcase, but with -E): 0m17.248s vs 0m28.382s.

This seems okay to me.
-- 
Arsen Arsenović