After r308322, clang hangs on compiling lib/IR/Function.cpp from llvm 3.5.2

[DimitryAndric]

Bugzilla Link	33900
Resolution	FIXED
Resolved on	Aug 02, 2017 03:27
Version	trunk
OS	All
Blocks	#33196
Attachments	Preprocessed Function.cpp and reproduction script
CC	@d0k,@emaste,@zmodem,@RKSimon,@rotateright,@ZviRackover

Extended Description

I recently imported llvm/clang trunk r308421 into FreeBSD 12-CURRENT, and afterwards, I got a report that building the port for llvm 3.5.2 was hanging: https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=219139#c13 (it gets automatically killed after 7200 seconds, as a "runaway build").

It turned out it was hanging on compiling lib/IR/Function.cpp, and bisecting shows this started at https://reviews.llvm.org/rL308322. Since this commit seems to only change two defaults (MaxLoadsPerMemcmp and MaxLoadsPerMemcmpOptSize), I suspect the hang exposes another, deeper lying issue.

I've attached a tarball with the fully preprocessed Function.cpp, and a shell script for reproduction. In the mean time, I'm trying to reduce the test case, but it is tricky to do with a hang. :) Normally, the .ii file should compile within ~20 seconds, depending on the speed of the machine.

[d0k]

r308891 should avoid runaway builds, I'm not sure if it's fast enough though.

[zmodem]

r308891 should avoid runaway builds, I'm not sure if it's fast enough though.

Merged to 5.0 in r309131.

Dimitry: can you confirm whether this fixes the issue?

[DimitryAndric]

r308891 should avoid runaway builds, I'm not sure if it's fast enough though.

Merged to 5.0 in r309131.

Dimitry: can you confirm whether this fixes the issue?

Yes, I should have reported that here earlier, sorry. After r308891, the test case finished compilation in ~46s (on a i7-3770 CPU @ 3.40GHz).

With its follow-up https://reviews.llvm.org/D35830 (which has been committed in r308986), the time is reduced even further, to ~38s. I think that one has also been nominated for merging?

[zmodem]

r308891 should avoid runaway builds, I'm not sure if it's fast enough though.

Merged to 5.0 in r309131.

Dimitry: can you confirm whether this fixes the issue?

Yes, I should have reported that here earlier, sorry. After r308891, the
test case finished compilation in ~46s (on a i7-3770 CPU @ 3.40GHz).

With its follow-up https://reviews.llvm.org/D35830 (which has been committed
in r308986), the time is reduced even further, to ~38s. I think that one
has also been nominated for merging?

Yes, r308986 was merged in r309127.

[ZviRackover]

mentioned in issue llvm/llvm-bugzilla-archive#34032