I attached the patch instead of pushing to Git, as I'm not satisfied yet with it. It's works stable here (passes `validate` & `nofib`), but I'm not sure about the `unsafeCoerce#`.
In my first version of this patch, I used a 3-tuple `(# Int#, ByteArray#, Word# #)` as return value, but that wasted 1 return register
#8647: Reduce allocations in `integer-gmp` --------------------------------------------+------------------------------ Reporter: hvr | Owner: Type: task | Status: patch Priority: normal | Milestone: 7.8.1 Component: libraries (other) | Version: 7.6.3 Resolution: | Keywords: integer- Operating System: Unknown/Multiple | gmp Type of failure: Runtime performance bug | Architecture: x86_64 Test Case: | (amd64) Blocking: | Difficulty: Unknown | Blocked By: | Related Tickets: #8638 --------------------------------------------+------------------------------ Comment (by hvr): Replying to [comment:6 hvr]: per `Integer` result, and still had to return a NULL pointer for `ByteArray#` for the case it wasn't used. After a conversation with Duncan, I'm now convinced that the `(# Int#, ByteArray# #)` + `unsafeCoerce#` hack is really bad, as the GC could try to follow the `ByteArray#` pointer even though it was really just a `Word#` in disguise. Otoh, passing a 0-pointer for the 3-tuple variant has the same danger. So now I'm left wondering if I can somehow statically allocate a dummy `ByteArray#` I can return in Cmm when there's no need to allocate a proper `ByteArray#` in order to avoid confusing the GC. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/8647#comment:7 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler