refcount: Use performBlockingMajorGC if available

[wenkokke]

No description provided.

[wenkokke]

@dcoutts This uses performBlockingMajorGC when available and uses your trick otherwise. I've provided a bit of explanation, and removed the first assert, since asserting twice seems superfluous. It'd give you fail-fast behaviour, but in the common case you hope to not fail.

[wenkokke]

@dcoutts This still requires yield in order to pass the test suite, even when using performBlockingMajorGC.

I'll look into what it actually does. In the meantime, this is WIP.

[wenkokke]

From cursory reading of GHC source, performBlockingMajorGC appears mostly to be the correct alternative to use for compatibility with non-moving concurrent GC. So we should probably keep your hack and combine it with preferring performBlockingMajorGC if it’s available?

[dcoutts]

I think this whole thing could be a lot simpler.

How about:

import qualified System.Mem

and then

-- Later GHC versions have 'performBlockingMajorGC' which does a full GC even
-- when using the concurrent incremental collector. This is what we want to ensure
-- all forgotten refs are found.
#if MIN_VERSION_base(4,20,0)
performBlockingMajorGC = System.Mem.performBlockingMajorGC
#else
performBlockingMajorGC = System.Mem.performMajorGC
#endif

and otherwise in checkForgottenRefs, keep it the same as originally but replace performMajorGC with performBlockingMajorGC. And the comment you added is good too.

But there's nothing magic about performBlockingMajorGC that means we don't need yield or that means we should do it once vs twice. The performBlockingMajorGC is just the same as performMajorGC in older GHC versions except that it provides the same guarantee for the incremental GC as the non-incremental: that it scans the whole heap.

[dcoutts]

LGTM now. Please squash and merge.

[dcoutts]

Unclear what's up with stylish-haskell.

[wenkokke]

Unclear what's up with stylish-haskell.

stylish-haskell really messed up owing to the fact that I didn't run it.