ghc 7.6.3 and 7.4.2 hang on -O2

[tsuraan]

I haven't had any luck finding a minimal test case for this, but the following replicates the issue on my two Linux machines running ghc 7.6.3:

git clone ​https://github.com/tsuraan/optimal-blocks
cd optimal-blocks
git checkout bceee69f9f03d2e559f96cdf58412ada09c96a8c
cabal sandbox init
cabal install --only-dependencies
cabal configure
cabal build

Cabal build will successfully build Algorithm.OptimalBlocks.SipHash and Algorithm.OptimalBlocks.BuzzHash, and it will attempt to build Algorithm.OptimalBlocks, but it will hang forever, using 100% of a single core. I have no idea how to debug this, but in src/Algorithm/OptimalBlocks.hs, if you replace the "in Blocks (reverse rlist) end" with "in Blocks [] bs", it will compile. Also, if that same file starts with the line "{-# OPTIONS_GHC -O1 #-}" it will compile.

This also happens on my Mac running OSX 10.8.5 and ghc 7.4.2 without llvm; I have to tweak the .cabal file a little to drop the lower bound on base and to get rid of the "-fllvm" flag, but if I do that I see the exact same behaviour.

[gelisam]

I can reproduce the issue with 7.6.3 (removing the -fllvm flag), but not with 7.8.20140130 (also removing the upper bound on base). Must have been fixed already.

[thoughtpolice]

This looks similar to #7944 and #5550 - can you see where the compiler halts during compilation using -v3 or something?

In any case, if it's fixed by 7.8, all the better.

[gelisam]

You are right, it hangs at the same spot as #7944:

...
*** SpecConstr:
Result size of SpecConstr

[thoughtpolice]

Right, then we know what's wrong. Unfortunately, there isn't really a good 'fix' that's possible to backport to fix SpecConstr for 7.6 users - in the mean time, you'll have to get by with -O1 or (-O2 -fno-spec-constr, I suppose). Alternatively, you can also try tuning the constructor specialisation threshold as noted in #7944.

There won't be a 7.6.4 release, and as this is fixed in HEAD, I'm marking this as a duplicate.

[simonpj]

I believe this is finally fixed; see #8852. Can you try now, with HEAD?

Simon

[thoughtpolice]

I believe this is fixed in HEAD after a quick attempt.