PowerPC : Undefined reference to `__sync_fetch_and_xor_8'

[erikd]

Compiling on PowerPC with git head (0c9c2d899e63b810), compile terminates with:

/home/ghc-upstream/libraries/ghc-prim/dist-install/build/
    libHSghcpr_FgrV6cgh2JHBlbcx1OSlwt-ghc7.9.20141214.so: undefined
    reference to `__sync_fetch_and_xor_8'
/home/ghc-upstream/libraries/ghc-prim/dist-install/build/
    libHSghcpr_FgrV6cgh2JHBlbcx1OSlwt-ghc7.9.20141214.so: undefined
    reference to `__sync_fetch_and_and_8'
/home/ghc-upstream/libraries/ghc-prim/dist-install/build/
    libHSghcpr_FgrV6cgh2JHBlbcx1OSlwt-ghc7.9.20141214.so: undefined
    reference to `__sync_fetch_and_nand_8'
/home/ghc-upstream/libraries/ghc-prim/dist-install/build/
    libHSghcpr_FgrV6cgh2JHBlbcx1OSlwt-ghc7.9.20141214.so: undefined
    reference to `__sync_val_compare_and_swap_8'
/home/ghc-upstream/libraries/ghc-prim/dist-install/build/
    libHSghcpr_FgrV6cgh2JHBlbcx1OSlwt-ghc7.9.20141214.so: undefined
    reference to `__sync_fetch_and_sub_8'
/home/ghc-upstream/libraries/ghc-prim/dist-install/build/
    libHSghcpr_FgrV6cgh2JHBlbcx1OSlwt-ghc7.9.20141214.so: undefined
    reference to `__sync_fetch_and_add_8'
/home/ghc-upstream/libraries/ghc-prim/dist-install/build/
    libHSghcpr_FgrV6cgh2JHBlbcx1OSlwt-ghc7.9.20141214.so: undefined
    reference to `__sync_fetch_and_or_8'

Looks like some new primops that need to be implemented for powerpc.

[erikd]

There is a quick and *very* dirty hack to get GHC compiling here ( ​https://phabricator.haskell.org/D560#15871 ) but it doesn't actually fix the problem, just prevent the compile error.

[tibbe]

I believe we fixed this on x86 by requiring i686. Can't find the ticket right now.

Another alternative is to actually implement it in the PowerPC but modifying the native code generator. We could lock at the instruction sequence GCC output for these and also take a hint from the x86 implementation I wrote.

[tibbe]

The x86 issue was fixed in #9346.

[erikd]

It seems that these __sync_fetch_* function were used around gcc 4.4 but have been replaced in gcc 4.8 with a different set of functions __atomic_*.

Working on a patch to sort this out.

[erikd]

For some reason GCC doesn't seem to have 64 bit versions of any of the atomic functions.

[erikd]

Bah, turns out it was a lack of the 64 bit versions of these that was the problem all along.

[erikd]

It seems the 64 bit primpos are *only* needed on 32 bit architectures and then only for the native code generators. For x86 via the native code generator we needed #9346 as mentioned by @tibbe, but the arm backend did need anything because that omly compilers via LLVM.

That leaves PowerPC as the remaining 32 bit architecture with an NCG. However, if I hack the 64 bit primops like (for example) this:

StgWord64
hs_atomic_add64(volatile StgWord64 *x, StgWord64 val)
{
#if WORD_SIZE_IN_BITS == 32
        printf ("%s : %s not implemented.\n", __FILE__, __func__) ;
        exit (1);
#else
  return __sync_fetch_and_add(x, val);
#endif
}

then building the compiler runs to completion and futhermore, the one test for these atomic primops, testsuite/tests/concurrent/should_run/AtomicPrimops.hs passes. It turns out this test operates on Int which on 32 bit platforms is a 32 bit value.

This makes me wonder if its worth the trouble of implementing these if they aren't used anywhere.

[erikd]

Replying to erikd:

It seems the 64 bit primpos are *only* needed on 32 bit architectures and then only for the native code generators.

Correction, the 64 bit primops are *not* needed on 32 bit architectures because there is currently no way to call them from Haskell code.

[Erik de Castro Lopo <erikd@…>]

In 19440ae2bb256f75934949ae57934caee3831a80/ghc:

ghc-prim : Hide 64 bit primops when the word size is 32 bits (fixes #9886).

Summary:
These primops were failing to compile on PowerPC (32 bit). There is also
currently no way to call into these primops from Haskell code. Currently,
the *only* way to call any of these C hs_atomic_* functions is via the
fetch*IntArray primops which are only defined for Int values and Int is
always the native word size.

When these functions can be called (and tested) from Haskell code, then
it will be worth while implementing them.

Test Plan:
    Compile and run on x86, x86_64, powerpc and arm:
    testsuite/tests/concurrent/should_run/AtomicPrimops.hs

Reviewers: tibbe, austin

Subscribers: thomie

Differential Revision: https://phabricator.haskell.org/D702

GHC Trac Issues: #9886

[thoughtpolice]

Merged (via bd785d101766a1207b8d27a45b6ae0bf83cd678a), thanks Erik.