Re: [eigen] Dropping C++14 compatibility discussion

On Wed, 8 Apr 2020 at 17:20, Martin Beeger <martin.beeger@xxxxxxxxx> wrote:

Hallo!

This is continuation of the discussion from last year about
compatibility, but I wanted to bring it up again, because I have some
new data from my experience.

Our company uses Eigen extensively and has a larger ecosystem built on
it, as do many of the voices on this list. We are in the embedded world,
so we are often hindered to adapt quickly to new tools, very much like
the HPC community has been.
But we managed to move to C++14 with our codebase in 2019. When doing
that, we somewhat monitored performance and compile time during the
adaption.

When we starting compiling our C++98 codebase as is with C+14,
performance already went up slightly. C++14 does silently move or elide
copies, and modern compiler have better optimizers, and the quality of
implementation of STL types got better. So if you are at all bound by
performance of STL types, I would really recommend using a new compiler
in C++14 mode even with your old code.

This was the obvious part. Another part was less obvious. We profiled
compile time, which didn't change much after the switch and started to
look into stuff which was expensive to compile. And then selectively in
places, where the compile time was bad we simplified the code using
C++14 features to improve compile time. This allowed us with very
localized changes to cut our compile time by almost in half, while, at
the same time, making the code in question often a lot simpler and new
features (e.g. for performance improvements) easier to implement.

This is also the experience which was observed with other template heavy
codebases like boost::mpl in comparison to hana and other stuff. The
gains in simplicity and compile time, especially from constexpr and
lambda features are not minor. They can often cut your code in half and
more than double the compile time.

The problem is, I am at a point where its hard to do much more to
improve the compile time of my codebase significantly, because most of
the compile time is brought in by 2 libraries: boost.test and Eigen. We
will most likely at some point abandon boost.test due to this, like many
others have already (which slowed the development and improvement of
boost.test further, while the alternatives got better and getting into
this downwards spiral). I would be vary happy if I am not forced to
abandon Eigen after the great 10 years we had with this library. In
order to avoid this I pulled a lot of tricks like explicit template
instatiations, tricks to reduce includes, even pimpl-like encapsulation
at performance cost to isolate from the problem, but that gets you only
so far.

I may well be that my use case is special, but I strongly assume that
new users which today want to adopt Eigen and have to look into its
internals (as you inevitable need to do at some points), will see how
its written and quickly run for alternatives. This amount of macros,
boilerplate and similar stuff will be an argument against this great
library some day and this day may already have come.

The important part is here: Eigen compile time and internal _expression_
template engine code readability it was great by 2010s standards, it was
ok by 2015 standards, it is borderline by 2020 standards, and unless
something changes, it will be unacceptable by 2025 standards, unless the
Eigen library moves along.

As C++ users, we do care about backwards compatibility greatly and that
is even good for me, but we should not go the C way and care about too
ancient compilers. The C++ comittee doesn't (that why int is required to
be 2s complement in C++20), so Eigen library maintainers IMHO should
follow.

What are the chances to get Eigen 3.4 out of the door with C++98 support
and drop it on the devel branch afterwards and jump to C++14?
What is the Eigen promises about how old yours compilers may be?
Can we explicitly agree on a statement like: We vow to support up to 3
year old compilers (or 5 years)?

If we could agree on clear and conservative rules like we will go 3
years back or 5 years back and state these on the Eigen front page,
Eigen user may look at our codebase and be much more willing to accept
older standards code, knowing that it will improve over time and that
the user gets some useful guarantees about the future in return.
There is agrument to be made that if you use a 5+ years old compiler,
you really do not care about performance, something Eigen uses generally
care about. So you are not in a targeted user group of Eigen. Little HPC
clusters do not at all offer any way to install a more recent compilers
that 5 years (this has changed a lot from a decade ago) and even in the
embedded world, vendors tend to drop support or upgrade for platforms
with more than 5 years old toolchains too (this has also very much
changed from 10 years ago).

Eigen users should be able to get a clear answer on the question when we
drop C++98 (of if). That belongs on the front page IMHO.

Kind regards,
Martin