Space leak in QuickCheck test

[jacereda]

The error:

shake-test: Stack space overflow: current size 33568 bytes.

Reducing the number of QuickCheck tests to 500 makes it pass.

[jacereda]

Forgot to say that I'm building with ghc 8 in case it matters.

[ndmitchell]

Does stubbing out the property still fail? Eg could it be a Mac specific QuickCheck bug? To stub out properly you would need the same type and some kind of deep seq to force the input.

[jacereda]

Do you mean something like this?

Success{} <- quickCheckWithResult stdArgs{maxSuccess=1000} $ \(Pattern p) (Path x) -> p `deepseq` x `deepseq` (property True)

That passes with -K1K

[ndmitchell]

Can you reinstate the property but print p and x before so we can see which value it fails on.

[jacereda]

With print (p,x), this is the last pair printed:

("b*\\/**ab//b\\\\\\b\\a\\b*\\b/baab\\b/*a///\\b\\\\/b/*/aaaa*b/b\\b\\\\a\\*bb/*a//\\///\\/**bb\\/aab//\\\\\\**","\\\\/baa\\/ab\\a/\\ab\\a\\/\\aaa///\\ab///\\a\\bb/b\\a/\\/\\/\\///a\\abbb/\\/ab\\//baab/\\\\//a\\a\\a\\//b\\/b\\\\\\/\\/")

[jacereda]

I have reduced the problem to:

Success{} <- quickCheckWithResult stdArgs{maxSuccess=1000} $ \(Pattern p) (Path x) -> label "foo" True

[jacereda]

Changing label "foo" True to property True terminates properly. I guess it's normal since quickcheck is collecting those labels. Isn't reducing maxSuccess to 500 an option here? Would be good to know why you don't get this same error though...

[ndmitchell]

If you change it to \(p :: Double) (x :: Double) do you still get the overflow? What version of QuickCheck are you using?

[ndmitchell]

Reducing maxSuccess to 500 is certainly an option, but I'd like to understand what's causing it first. It seems something has a space leak, and from where you are getting to, that something is looking increasingly like QuickCheck, and thus would be good to fix upstream.

[jacereda]

QuickCheck version is 2.8.2, but I found the reason.
I was building with stack test --profile. As soon as I changed my build command to stack test the problem disappeared.
Is there some way to detect profiling in cabal so a larger stack can be used there?

[ndmitchell]

Thanks, I'll try and reproduce later. If profile is the culprit it's probably a space leak in QC that gets fixed by optimisation. Should still be quite easy to track down and fix.

[ndmitchell]

OK, I can reproduce this in QuickCheck alone, so nothing to do with Shake. Will track it down there.

[ndmitchell]

Space leak diagnosed and patched, pull request at nick8325/quickcheck#93.

[ndmitchell]

My thoughts are if this gets patched quickly, do nothing to Shake, if not, remove the label call until a release is made.