Linker issues around integer-gmp on MacOS

[harpocrates]

I'm not really sure if this is a Cabal issue or a GHC one, but since old-build sometimes succeeds where new-build segfaults, I'm going to assume it might be Cabal. I ran across this when trying to get the Haddock test suite to run with GHC 8.4.1, you can see what I tried there: haskell/haddock#767.

I've only seen this happen on MacOS - everything worked just fine in my Ubuntu VM.

---

	binary GHC-8.4	manually-built GHC-8.4
new-style commands	crash	won't compile
old-style commands	crash	works

For each cell in this table I started by building Haddock (either the ghc-8.4 or the official 1.19.0.1 release branch - same result).

• new-style:

   cabal new-build -w ghc-8.4
  

• old-style:

   cabal sandbox init
   cabal sandbox add-source haddock-library
   cabal sandbox add-source haddock-api
   cabal install -j4 --dependencies-only -w ghc-8.4
   cabal configure -w ghc-8.4
   cabal build -j4
  

Then, I created a test file with some TH in it:

-- Main.hs
{-# LANGUAGE TemplateHaskell #-}
main = pure ()
pure []

Finally, I ran the compiled haddock with Main.hs as an argument:

• new-style:

   ./dist-newstyle/build/x86_64-osx/ghc-8.4.1/haddock-2.19.0.1/c/haddock/build/haddock/haddock Main.hs
  

• old-style:

   ./dist/build/haddock/haddock Main.hs
  

While I was able to get this to work with a manually-built GHC-8.4 and with the old-style commands, everything else crashed when I called when I tried to run the haddock binary with an error of the form:

lookupSymbol failed in relocateSection (RELOC_GOT)
/Users/atheriault/.stack/programs/x86_64-osx/ghc-8.4.1/lib/ghc-8.4.1/integer-gmp-1.0.1.0/HSinteger-gmp-1.0.1.0.o: unknown symbol `___gmp_rands'
haddock: unable to load package `integer-gmp-1.0.1.0'

[hvr]

/cc @bgamari does this ring any bells?

[angerman]

I've seem something like with with in-tree-gmp.

@harpocrates what's the exact error with your manually built 8.4 and the new-style haddock?
The same just with integer gmp at a different location?

Also, could you share both binaries somewhere?

lookupSymbol failed in relocateSection (RELOC_GOT)

This, I think, comes from ghcs internal linker.

[harpocrates]

Woops! I goofed. I don't get an error with the new-style commands + manually built 8.4 - I get a compile error (sorry - I miscopied the summary from the Haddock thread):

Building library for haddock-api-2.19.0.1..
<command line>: cannot satisfy -package-id ghc-paths-0.1.0.9-f1132a3630dfd069ac6c748c11485c2bc7a77709c7f89a58
360749883bab0d93:
    ghc-paths-0.1.0.9-f1132a3630dfd069ac6c748c11485c2bc7a77709c7f89a58360749883bab0d93 is unusable due to sha
dowed dependencies:
      base-4.11.0.0
    (use -v for more information)
cabal: Failed to build haddock-api-2.19.0.1 (which is required by exe:haddock
from haddock-2.19.0.1).

(Worth noting I'm building the Haddock 2.19.0.1 release and GHC 8.4.1 release tags, so dependencies really should be working out)

[angerman]

@harpocrates that looks more like your dist/dist-newstyle folders are somehow messed up. You might want to nuke them and try again. I'm not sure cabal will clean them out properly if you use the same compiler version from two different places.

[harpocrates]

@angerman I just did a fresh clone and new-build with only my manually built 8.4. I'm seeing exactly the same compile error as before.

(I've also been regularly nuking my dist/dist-newstyle folders between different experiments.)

[angerman]

Hmm... I’ll try to reproduce this later.

[harpocrates]

Worth mentioning this isn't a blocker for me anymore - I'm just falling back onto my Ubuntu VM whenever I hit this. That said, I'm willing to put time into stepping through any sort of debugging ideas people may have.

[LeventErkok]

I'm still being hit by this, as it manifests itself on doctest: sol/doctest#199

Essentially, with GHC 8.4.1, doctest on Mac OSX simply doesn't run. Replicating from that ticket, it fails with:

$ doctest Data/SBV/Examples/Polynomials/Polynomials.hs
doctest:
lookupSymbol failed in relocateSection (RELOC_GOT)
/usr/local/lib/ghc-8.4.1/integer-gmp-1.0.1.0/HSinteger-gmp-1.0.1.0.o: unknown symbol `___gmp_rands'
doctest: doctest: unable to load package `integer-gmp-1.0.1.0'

Any help would be most appreciated!

[simonmichael]

I see the same thing with hledger-lib doctests on mac with GHC 8.4.1, but FWIW I am using stack and the GHC it installed, not cabal old or new-build.

[LeventErkok]

anyone knows if GHC 8.4.2 will fix this problem?

[gbaz]

@bgamari ?

[harpocrates]

I opened a GHC trac ticket for a very similar (maybe the same?) bug here and I think that is going to be fixed in 8.4.2.

[LeventErkok]

I'm afraid the problem is still there with ghc 8.4.2:

doctest:
lookupSymbol failed in relocateSection (RELOC_GOT)
/usr/local/lib/ghc-8.4.2/integer-gmp-1.0.2.0/HSinteger-gmp-1.0.2.0.o: unknown symbol `___gmp_rands'
doctest: doctest: unable to load package `integer-gmp-1.0.2.0'

Can anyone else confirm?

[harpocrates]

I can confirm. 😢

[LeventErkok]

@harpocrates @bgamari Since the GHC bug you referenced is already closed as fixed, should we open a new one? I’m happy to do so, but you have more context. Let me know.

[harpocrates]

I've reported this as a GHC API bug: https://ghc.haskell.org/trac/ghc/ticket/15105.

[23Skidoo]

If this is a GHC bug, can this ticket be closed?

[23Skidoo]

Closing in favour of https://ghc.haskell.org/trac/ghc/ticket/15105.