ghc-lib-parser causes segfaults when used together with the `ghc` library on ghc-9.10

[AndreasPK]

The FastString table built into GHC changed recently.

Since this table is shared by all versions of the GHC source code via an RTS symbol this causes segfaults when ghc-lib-parser-9.12 and ghc-9.10 are used together since they use different memory layouts for this table.

The fix for this is either to change GHC.Data.FastString such that the sharedCAF mechanism is not used. However that means FastStrings created by ghc-lib-parser and ghc are never allowed to be mixed.

The other alternative is to use CPP to adjust the memory layout of the faststring table depending on the GHC version used.

This is the change in question: https://gitlab.haskell.org/ghc/ghc/-/commit/f2cc1107790d42fee1a11d5b16bc282d31ea6f78

---

The simplest option of course would be to mark ghc-lib-parser-9.12 unbuildable on ghc-9.10, but perhaps also a bit of a shame.

[AndreasPK]

Related HLS bug where this was discovered: haskell/haskell-language-server#4674

[shayne-fletcher]

The other alternative is to use CPP to adjust the memory layout of the faststring table depending on the GHC version used.

what would that cpp look like and where would it go? FastString.hs? we can have ghc-lib generate it. care to offer a patch?

[vaibhavsagar]

I don't know where to put this as a PR, but I'm guessing a patch would look something like this:

diff --git a/compiler/GHC/Data/FastString.hs b/compiler/GHC/Data/FastString.hs
index 017a8d3926a..c72a24cf16b 100644
--- a/compiler/GHC/Data/FastString.hs
+++ b/compiler/GHC/Data/FastString.hs
@@ -310,6 +310,7 @@ data FastStringTable = FastStringTable
   {-# UNPACK #-} !FastMutInt
   -- ^ The unique ID counter shared with all buckets
   --
+#if MIN_VERSION_ghc(9,12,0)
   -- We unpack the 'FastMutInt' counter as it is always consumed strictly.
   {-# NOUNPACK #-} !FastMutInt
   -- ^ Number of computed z-encodings for all buckets.
@@ -318,6 +319,9 @@ data FastStringTable = FastStringTable
   -- in 'mkFastStringWith' and needs to be boxed any way.
   -- If this is unpacked, then we box this single 'FastMutInt' once for each
   -- allocated FastString.
+#else
+  {-# UNPACK #-} !FastMutInt -- number of computed z-encodings for all buckets
+#endif
   (Array# (IORef FastStringTableSegment)) -- ^  concurrent segments

 data FastStringTableSegment = FastStringTableSegment

[shayne-fletcher]

cc @dylant-da

[dylant-da]

Seen this just now - I'll take a look tomorrow to see what I can understand of the issue and the proposed fix.

[AndreasPK]

What @vaibhavsagar posted seems right to me.

[hasufell]

What is the status of fixing this?

[dylant-da]

Apologies, this fell off my radar - taking a look at incorporating this change later today.

[dylant-da]

Gonna need a bit longer to make the time for this.