Don't free() the handle after dlclose

[garrison]

Of the possible code paths, there was only one that freed the handle after calling dlclose. This change removes that free() statement, fixing an issue that was recently brought to light by JuliaLang/julia#10808.

[tkelman]

Should this be conditional on VERSION?

[garrison]

Sure, this introduces a (very small) memory leak for old VERSIONs, but there were already code paths in BinDeps that leaked the same memory. So I don't think it's worth adding a conditional on VERSION.

[garrison]

What would people think of doing a BinDeps release now that this change is merged? Right now the released version of BinDeps doesn't jive with julia master (see e.g. @jiahao's encounter with this bug).

[tkelman]

Are we positive this is the right fix, before or after JuliaLang/julia#10832 or equivalent gets merged? I'm very very hesitant to change any behavior of BinDeps on release Julia, since it's critical infrastructure, either directly or indirectly, for such a large portion of packages. If BinDeps was compensating for a memory leak in base and we're not going to backport the change to base (and if we did, we'd have to bump REQUIRE here to reflect that minimum version of Julia), I'd rather leave that in place.

[garrison]

There were two symptoms in BinDeps that resulted from JuliaLang/julia#10808 being merged. The first resulted from JuliaLang/julia@11a0f43, the fix for which is in JuliaLang/julia#10832 (still waiting to be merged, though I am sure this is the correct fix).

The other resulted from JuliaLang/julia@fcf4a47, which is fixed in this pull request. Yes, BinDeps was compensating for a memory leak in base, though not fully -- there are in fact two other calls to Libc.dlclose in dependencies.jl which were not accompanied by a call to Libc.free, each of which (before and after this pull request) leaks two pointers of memory on Julia 0.3.

[mlubin]

Is this memory leak just applicable during Pkg.build? It seems pretty harmless, I'd rather have travis working again for NLopt/Cbc on Julia nightly.

[tkelman]

I agree it's not good to leave this broken on nightly for too long. If the memory leak on 0.3 is small and only occurs during Pkg.build then I guess it's not that bad, but "there were already leaks elsewhere" doesn't seem like a very good reason to knowingly introduce new leaks.

[ihnorton]

I vote to add a conditional for 0.3 only (don't worry about 0.4-dev sub-versions), and then tag BinDeps.

[garrison]

I concur with @ihnorton.

[tkelman]

That works for me.

[ihnorton]

Done.

[tkelman]

didn't help for long, tuple changes caused more breakage