#7794: GHCi "Prelude.undefined" exceptions on ARM; ByteCodeItbls.mkJumpToAddr unimplemented -------------------------+-------------------------------------------------- Reporter: cjwatson | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: GHCi | Version: 7.6.2 Keywords: | Os: Linux Architecture: arm | Failure: GHCi crash Difficulty: Unknown | Testcase: Blockedby: | Blocking: Related: | -------------------------+-------------------------------------------------- Changes (by igloo): * difficulty: => Unknown Old description:

A number of non-trivial operations fail in ghci on ARM (specifically armhf on both Debian experimental and Ubuntu raring). I initially noticed this as conduit's doctests failing, but reduced this to:

$ ghci GHCi, version 7.6.2: http://www.haskell.org/ghc/ :? for help Loading package ghc-prim ... linking ... done. Loading package integer-gmp ... linking ... done. Loading package base ... linking ... done. Prelude> data Type a = Nothing *** Exception: Prelude.undefined

This is a singularly unhelpful exception, coming as it does without context, but I eventually tracked it down to ByteCodeItbls.mkJumpToAddr having no specific implementation for ARM and thus falling back to the default of "undefined". I've tried adding this:

+#elif arm_TARGET_ARCH +type ItblCode = Word32 +mkJumpToAddr a + = [ 0xe51ff004 -- ldr pc, [pc, #-4] # pc reads as <current insn>+8 + , fromIntegral (ptrToInt a) ]

This definitely changes the nature of the problem, so I think I'm on the right track, and I'm reasonably confident that that's the correct branch implementation; but when I actually try it in practice I get crashes (SIGSEGV/SIGILL/SIGBUS) with trashed stack traces, so I'm clearly not done yet. Friends of mine suggest that I may need to __clear_cache or equivalent, which sounds plausible, so my current line of attack is figuring out how to glue that in; but if any actual GHC hackers can spot something obviously missing in the meantime then that would be great.

New description: A number of non-trivial operations fail in ghci on ARM (specifically armhf on both Debian experimental and Ubuntu raring). I initially noticed this as conduit's doctests failing, but reduced this to: {{{ $ ghci GHCi, version 7.6.2: http://www.haskell.org/ghc/ :? for help Loading package ghc-prim ... linking ... done. Loading package integer-gmp ... linking ... done. Loading package base ... linking ... done. Prelude> data Type a = Nothing *** Exception: Prelude.undefined }}} This is a singularly unhelpful exception, coming as it does without context, but I eventually tracked it down to ByteCodeItbls.mkJumpToAddr having no specific implementation for ARM and thus falling back to the default of "undefined". I've tried adding this: {{{ +#elif arm_TARGET_ARCH +type ItblCode = Word32 +mkJumpToAddr a + = [ 0xe51ff004 -- ldr pc, [pc, #-4] # pc reads as <current insn>+8 + , fromIntegral (ptrToInt a) ] }}} This definitely changes the nature of the problem, so I think I'm on the right track, and I'm reasonably confident that that's the correct branch implementation; but when I actually try it in practice I get crashes (SIGSEGV/SIGILL/SIGBUS) with trashed stack traces, so I'm clearly not done yet. Friends of mine suggest that I may need to `__clear_cache` or equivalent, which sounds plausible, so my current line of attack is figuring out how to glue that in; but if any actual GHC hackers can spot something obviously missing in the meantime then that would be great. -- -- Ticket URL: http://hackage.haskell.org/trac/ghc/ticket/7794#comment:2 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler