[PATCH] libstdc++: use new built-in trait __is_reference

[Patrick Palka]

On Mon, Mar 20, 2023 at 5:56 AM Ken Matsui <kmatsui@cs.washington.edu> wrote:
>
> > Does it actually make compilation faster though?
> >
> > Has it been measured?
>
> In my understanding, what I have implemented so far is so simple that
> it does not affect the speed. These traits are what Partick kindly
> recommended to get started. As explained on the GSoC page, some traits
> might involve expensive instantiation of multiple class templates. So
> IMHO, implementing built-in traits for those traits can make
> compilation cheaper.
>
> I have not measured it, but Patrick might have done?

Although the native implementation of using two partial
specializations is very efficient, I'd suspect using a built-in would
be even better because it'd avoid the work of having to figure out
which of the two partial specializations to instantiate.

I contrived the attached benchmark to measure this, which instantiates
~160k specializations of is_reference_v (it was simpler to write a
benchmark for the variable template version). On my machine gcc trunk
(release build) as well as Clang 15 use between 10-15% less
time/memory with -DUSE_BUILTIN than without for this benchmark, so
this suggests the built-in improves compile time/memory usage for this
trait by at least 10-15%.  For more complicated traits the improvement
should be even better.

>
> On Mon, Mar 20, 2023 at 2:14 AM Jonathan Wakely <jwakely.gcc@gmail.com> wrote:
> >
> > On Mon, 20 Mar 2023 at 08:08, Xi Ruoyao via Libstdc++
> > <libstdc++@gcc.gnu.org> wrote:
> > >
> > > On Mon, 2023-03-20 at 01:03 -0700, Ken Matsui wrote:
> > > > Oops, I assumed those were my email... Thank you for your heads up and
> > > > your comments!
> > > >
> > > > > Bad ChangeLog format.  You should have a tab (not 4 or 8 spaces, nor
> > > > > nothing) to indent the ChangeLog content.
> > > >
> > > > Do you mean like the following?
> > > >
> > > > ```
> > > > libstdc++-v3/ChangeLog:
> > > >
> > > > [TAB]* include/std/type_traits (is_reference): Use __is_reference
> > > > built-in
> > > > trait.
> > > > ```
> > >
> > > Yep.
> > >
> > > > > Is there any benefit to use a builtin, instead of the existing
> > > > > implementation?  I can see no but maybe I'm stupid.
> > > >
> > > > My patches are based on the GSoC project "C++: Implement compiler
> > > > built-in traits for the standard library traits". These built-in
> > > > traits basically make the compilation faster.
> > > >
> > > > https://gcc.gnu.org/wiki/SummerOfCode
> > >
> > > Ok, to me making compilation faster is a valid reason.
> >
> > Does it actually make compilation faster though?
> >
> > Has it been measured?
> >
> > > > > The patch fails to apply.  It seems because your mail client inserted an
> > > > > additional newline before "b/".  Try to use git-send-email or configure
> > > > > the mail client properly.
> > > >
> > > > Let me try to use git-send-email instead. I stupidly don't understand
> > > > how to use them, so I was making my patches manually...
> > >
> > > Or adjust the mail client correctly.  You can send the patch to yourself
> > > first and see if it's not "mangled" by the mail client when you debug
> > > such an issue...
> > >
> > > But when you finally end up sending 10 patches in a series you'll find
> > > git send-email much easier :).
> >
> > Figuring out how to generate proper patches is an important part of
> > contributing to GCC, so part of any GSoC project.
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: is_reference_benchmark.C
Type: text/x-c++src
Size: 89639 bytes
Desc: not available
URL: <https://gcc.gnu.org/pipermail/libstdc++/attachments/20230320/604cb4af/attachment-0001.bin>