I pointed out two shortcomings of the current semantics already. Even with the proposed symbol offsets, the prefix data must start with a target-dependent frontend-constructed instruction, which is subsequently not used. I.e. it requires carefully constructed garbage bytes. To me,

#4213: LLVM: Add support for TNTC to LLVM compiler suite -------------------------------------+------------------------------------ Reporter: dterei | Owner: dterei Type: feature request | Status: new Priority: low | Milestone: 7.6.2 Component: Compiler (LLVM) | Version: 6.13 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: Unknown/Multiple Type of failure: None/Unknown | Difficulty: Unknown Test Case: | Blocked By: Blocking: | Related Tickets: -------------------------------------+------------------------------------ Comment (by bgamari): Replying to [comment:30 altaic]: that sounds kind of... not good. I could be mistaken but I don't believe this should be necessary with symbol offsets as the inliner won't see the info table at all (since the symbol points to the entrypoint of the function as it should). -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/4213#comment:31 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler