On Saturday, 17 June 2017 at 21:49:29 UTC, Walter Bright wrote:
It's not quite the same. I spend 99.99% of my time programming
working with code, not makefiles.
Martin and Sebastian can correct me if I'm wrong, but unless I
am, Martin, Sebastian, and myself spend a scary amount of time
wrestling with makefiles. We have had bad production issues due
to makefiles, such as missing dependencies causing inconsistent
or broken builds (notably observable when something works only
without -j), typos or undefined variables resulting in bugs being
silently ignored (hey, I filed a PR to fix one of those not half
an hour ago - dmd#6916), CI checks being accidentally completely
switched off (which happened more than once!), the win32.mak /
win64.mak / posix.mak mess, super-ugly hacks due to limitations
of Make that slow down the build unconditionally and/or make
everything much more fragile and complicated... and I could go on.
Which is not to say that I think we need to get off makefiles
ASAP. Migrating to anything else is going to undoubtedly incur a
massive migration cost and for a long time, a big maintenance
cost (especially if it involves dogfooding). I agree with many of
your arguments as well.
I think you're not encountering much of the problems with
makefiles because you're mainly dealing with DMD, whose build
process is relatively simple - and even there, the build process
is mostly maintained these days by other people. In comparison,
the build process for dlang.org has been getting much more
complicated with time (partially because of some technical debt
and fragmentation, but also simply because we want to do more
things like nice runnable examples, various pre-processing /
post-processing / validation, building the spec from the compiler
source, etc.
If anyone wants to SERIOUSLY propose a plan to migrate from
makefiles, I'd be interested to look at it. However, its benefits
obviously must outweigh the maintenance costs of current
makefiles, as well as not raising the contribution barrier too
much.
Some requirements would be:
- Obvious syntax - making a change such as adding a new target or
dependency should be obvious just from looking at the existing
code.
- Performance - shouldn't have considerable overhead; but also,
it shouldn't take 10 seconds to "rebuild" the build tool itself
after every modification.
- Cross-platform support - this may seem obvious, but it should
support at least all platforms DMD supports. One of the
suggestions for replacement in a similar recent thread
(http://forum.dlang.org/post/ohkkqj$21lg$1...@digitalmars.com),
buildsome, shows no indication of any Windows support at all, not
even plans to work on it.
- Minimal dependencies - ideally it should work with as few as
possible external dependencies that one would typically need to
install to build D. Windows is very often overlooked when
evaluating this requirement, for example although Python is
ubiquitous on POSIX systems, it's much less so on Windows.
That doesn't leave many candidates. reggae might be the closest
to satisfying all of the above, though I haven't looked too
closely at it. For one thing, I don't really understand the need
for build system generators - seems like an extra step and
imposing an implementation detail on the user.
