On Fri, 14 Oct 2022, Ivan Perez wrote:
On 2022-10-12 14:00, Scott Talbert wrote:
On Tue, 11 Oct 2022, Ivan Perez wrote:
On 2022-10-11 19:34, Scott Talbert wrote:
On Tue, 11 Oct 2022, Ivan Perez wrote:
On 2022-10-11 14:00, Scott Talbert wrote:
The biggest issue currently is
with what4 - it isn't building on all architectures yet, which is
preventing copilot from being built on all architectures.
You can check the package tracker here (and see testing migrations
section): https://tracker.debian.org/pkg/haskell-copilot
Ok. I'm not very familiar with the tracker's interface.
Is what4 itself not building, or is it not building because a dependency
of what4 is not building?
There may be dependencies we or they can drop (we can't drop all of
what4).
I expect future releases to have fewer dependencies, not more. I've been
trying to prune that a lot.
It's what4 itself. It is building, but tests are not working on
anything but amd64. I'll try to sort that out soon.
See: https://buildd.debian.org/status/package.php?p=haskell-what4
Ok. I've made the what4 team aware. They are on standby. If there's an
actual bug being detected by the tests, they can work on trying to fix it.
I ended up disabling the template_tests and expr-builder-smtlib2 tests
on everything but amd64. It looks like the what4 team has been
working on the tests (in git), so perhaps we can revisit after the
next release.
Now, I am down to just a failure on armhf where the build is timing
out. That one file ([57 of 79] Compiling What4.Expr.Builder (
src/What4/Expr/Builder.hs, dist-ghc/build/What4/Expr/Builder.o,
dist-ghc/build/What4/Expr/Builder.dyn_o )) takes a REALLY long time to
compile!
I asked about this. Without access to that architecture, they can't figure
out why that file takes so long to compile on armhf but not e.g., armel.
It does take a long time to compile on armel as well - but not sure why it
seems slightly faster on armel. In any event, we tried building what4 on
armhf and it actually succeeded this time. So, we're past that issue now.
copilot-theorem and copilot-language built, and -libraries is building
now. :)
It's a large file (over 4K locs), and uses a lot of extensions and imports.
Nothing else stands out as suspicious to me.
Could it be just swapping? GHC leaves some byproducts of earlier files
compiled in memory. The first time I compiled what4, GHC's memory consumption
was up to 35% on my Macbook. The second time, it was below 30%. It swapped in
the first case.
I'm able to partition Builder into several modules. Consumption goes down to
21% on my mac. I don't know if that would help. Is there a simple way to try
it on an amd64 machine to make sure it's worth the effort of making this
change official in what4?
That particular file takes a long time to compile even on amd64, so it
seems likely whatever improvements made to amd64 performance would be
applicable to other architectures too. I do have access to all of the
Debian architectures so if you want me to try something on an arm box, let
me know.
Scott
