Summary:
  The commit fixes incorrect code generation of
  integer-gmp package on ia64 due to C prototypes mismatch.
  Before the patch prototypes for "foreign import prim" were:
      StgWord poizh[];
  After the patch they became:
      StgFunPtr poizh();

Long story:

Consider the following simple example:

    {-# LANGUAGE MagicHash, GHCForeignImportPrim, UnliftedFFITypes #-}
    module M where
    import GHC.Prim -- Int#
    foreign import prim "poizh" poi# :: Int# -> Int#

Before the patch unregisterised build generated the
following 'poizh' reference:
    EI_(poizh); /* StgWord poizh[]; */
    FN_(M_poizh_entry) {
    // ...
    JMP_((W_)&poizh);
    }

After the patch it looks this way:
    EF_(poizh); /* StgFunPtr poizh(); */
    FN_(M_poizh_entry) {
    // ...
    JMP_((W_)&poizh);
    }

On ia64 it leads to different relocation types being generated:
  incorrect one:
    addl r14 = @ltoffx(poizh#)
    ld8.mov r14 = [r14], poizh# ; r14 = address-of 'poizh#'
  correct one:
    addl r14 = @ltoff(@fptr(poizh#)), gp ; r14 = address-of-thunk 'poizh#'
    ld8 r14 = [r14]

'@fptr(poizh#)' basically instructs assembler to creates
another obect consisting of real address to 'poizh' instructions
and module address. That '@fptr' object is used as a function "address"
This object is different for every module referencing 'poizh' symbol.

All indirect function calls expect '@fptr' object. That way
call site reads real destination address and set destination
module address in 'gp' register from '@fptr'.

Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>