Re: building LLVM threads gets killed

When I discovered that you were building debug versions
of devel/llvm60 that also meant that my Pine64+ 2GB
specifics do not make a useful comparison (do not
show how small an environment can be for a build).

Last I tried such a debug build was on a powerpc64
with 16 MiBytes RAM and it took having a swap space
of 10 MiBytes (9 insufficient) but it was devel/llvm40
and the machine was not in use for anything else
significant and was using UFS, unlike your more
involved context (ZFS, large applications running).

For your usage, it is not clear to me that 32 GiBytes
is all that much RAM, nor that 32GiByte RAM + 2 GiByte
swap is all that big of a total. To me it seemed odd
that your swap was so small for what you described as
running in your environment vs. the amount of RAM.

If it turns out handy, the stress test (scaled to 32
GiBytes) should show how your context handles low free
RAM over extended periods. (Without the complications
or delays of a build being involved.) (The Mark Johnston
patches expose more than just any "was killed" notice.)

My expectation is that the stress test would over time
show OOM kills even if you expanded the swap space so
that it clearly had notable space not in use at the
time.

It may be that setting vm.pageout_oom_seq large enough
would help your builds complete by tolerating low
free RAM for a longer time.

Something I only just learned was that by default lld
is multi-threaded. I've been told that this mode of
operation uses more RAM but I've no knowledge of my
own about the tradeoffs. They might be bigger for
debug builds than for non-debug builds. Doing a quick
experiment showed lld using 5 threads or so unless
I used -Wl,--no-threads . (I was using the cc/c++
command interface.)

===
Mark Millard
marklmi at yahoo.com
( dsl-only.net went
away in early 2018-Mar)