-O results in incorrect behavior

[RichardG]

HUnit 1.2.2.1 ships with a set of unit tests. Incorrect behavior occurs when these tests are compiled into a program using Cabal and GHC 6.12.1 (Haskell Platform 2010.1).

Correct behavior occurs when these tests are:

• compiled using ghc --make using GHC 6.12.1.
• run using GHCi 6.12.1.
• compiled into a program using Cabal and GHC 6.10.4 (Haskell Platform 2009.2).
• compiled using ghc --make using GHC 6.10.4.
• run using GHCi 6.10.4.

This behavior appears to involve interactions between the Testable and Assertable classes and instances in Test/HUnit/Base.hs. This affects the correctness of the programs.

[RichardG]

Script of running various tests with GHC 6.10.4

[RichardG]

Script of running various tests with GHC 6.12.1

[RichardG]

Simplified test case

[RichardG]

Script of simplified test case

[RichardG]

I've attached a somewhat simplified test case (it's HUnit with a bunch of stuff ripped out). The tests/SimpleTest.hs file has 3 tests in it which compare the expected and actual results of a failed assert. These are:

expectProblem Failure 0 "test" (TestCase $ assertFailure "test"),
expectProblem Failure 0 "test" (test $ assertFailure "test"),
expectProblem Failure 0 "test" ((TestCase . (>>= assert)) $ (assertFailure "test"))

The first test gives the expected result. The second and third tests display the same erroneous behavior that is described in the original report (i.e., it behaves incorrectly only when compiled with GHC 6.12.1 and Cabal).

(The third test in the list is the second test where test has been replaced with, what I assume are, the appropriate functions.)

[RichardG]

Further experimentation has shown the behavior differs when the -O flag is used. I.e.,

ghc --make -i.:tests -cpp -DBASE4 tests/HUnitTest98.lhs

has the correct behavior but

ghc --make -i.:tests -cpp -DBASE4 -O tests/HUnitTest98.lhs

does not.

The severity of this issue is compounded by the fact that Cabal uses the -O flag by default.

[RichardG]

This was reproduced with the 6.12.2 snapshot (10-Apr-2010 15:57) as compiled with GHC 6.10.4 (Haskell Platform 2009.2).

[RichardG]

Ticket #3983 may be related.

[simonpj]

I suspect this is related to #3959, which I'm working on.

Simon

[igloo]

I can reproduce this on amd64/Linux with the 6.12 branch, but not the HEAD.

[simonmar]

This won't be fixed for 6.12.3, but we almost have a fix ready for HEAD.

[simonpj]

I believe this is now fixed by the fix to #3959. Could you see if it works now, please?

Thanks

Simon

[RichardG]

Tested with HUnit 1.2.2.3 and GHC 7.0.0.20100924 on Mac OS X 10.6.4 Server. Issue was not reproduced.

[simonpj]

OK so it looks as if it's fixed. Good. I'm clsoing te ticket.

Simon