When linking a call to a file in the same module that requires a long branch, gold inserts a jump to 8 bytes into the stub.

This came up with a ghc update in Ubuntu, but it's possible to recreate by hand.

Here are my test files:

(hirsute-ppc64el)root@juju-b11c42-ubuntu-26:/build/haskell-network-byte-order-RQo4iL/haskell-network-byte-order-0.1.5/gold-test-case# cat main.c
void large(void);
void local(void);

int
main(int argc, char** argv)
{
        local();
        return 0;
}

(hirsute-ppc64el)root@juju-b11c42-ubuntu-26:/build/haskell-network-byte-order-RQo4iL/haskell-network-byte-order-0.1.5/gold-test-case# cat large.S
        .machine power8
        .abiversion 2
        .section        ".text"
        .align 2
large:
#include "nops16M.h"
(hirsute-ppc64el)root@juju-b11c42-ubuntu-26:/build/haskell-network-byte-order-RQo4iL/haskell-network-byte-order-0.1.5/gold-test-case# cat target.c
#include <stdio.h>
void local(void);
int a_global;

void local(void)
{
        printf("a_global %d\n", a_global);
}

I made nops16M.h like this:

for x in `seq 1024`; do echo nop >> nops1k.h; done
for x in `seq 1024`; do echo '#include "nops1k.h"' >> nops1M.h; done
for x in `seq 16`; do echo '#include "nops1M.h"' >> nops16M.h; done


Then:

# gcc -c -o main.o main.c
# gcc -c -o large.o large.S
# gcc -c -o target.o target.c
# gcc -fuse-ld=gold -o test-gold main.c large.o target.o
# ./test-gold
Segmentation fault (core dumped)

And from objdump -D test-gold:

000000000000073c <main>:
     73c:       02 04 4c 3c     addis   r2,r12,1026
     740:       c4 78 42 38     addi    r2,r2,30916
     744:       a6 02 08 7c     mflr    r0
     748:       10 00 01 f8     std     r0,16(r1)
     74c:       f8 ff e1 fb     std     r31,-8(r1)
     750:       c1 ff 21 f8     stdu    r1,-64(r1)
     754:       78 0b 3f 7c     mr      r31,r1
     758:       78 1b 69 7c     mr      r9,r3
     75c:       20 00 9f f8     std     r4,32(r31)
     760:       e6 01 09 7c     mtfprwz f0,r9
     764:       2c 00 3f 39     addi    r9,r31,44
     768:       ae 4f 00 7c     stfiwx  f0,0,r9
     76c:       9d 00 00 48     bl      808 <00000000.long_branch.4000810+0x8>
     770:       00 00 00 60     nop
     774:       00 00 20 39     li      r9,0
     778:       78 4b 23 7d     mr      r3,r9
     77c:       40 00 3f 38     addi    r1,r31,64
     780:       10 00 01 e8     ld      r0,16(r1)
     784:       a6 03 08 7c     mtlr    r0
     788:       f8 ff e1 eb     ld      r31,-8(r1)
     78c:       20 00 80 4e     blr
     790:       00 00 00 00     .long 0x0
     794:       00 00 00 01     .long 0x1000000
     798:       80 01 00 01     .long 0x1000180
     79c:       00 00 00 00     .long 0x0

and

0000000000000800 <00000000.long_branch.4000810>:
     800:       70 80 82 e9     ld      r12,-32656(r2)
     804:       a6 03 89 7d     mtctr   r12
     808:       20 04 80 4e     bctr
     80c:       00 00 00 00     .long 0x0

This doesn't happen with -fuse-ld=bfd.

There are some recent commits that look like they might address this (specifically fa40fbe484954c560ab1c0ff4bc1b2eeb1511344 which says in part

            (Target_powerpc::Branch_info::make_stub): Don't add local
            entry offset to long branch dest passed to
            add_long_branch_entry.  Do pass st_other bits.

) but I built gold from git and this still happens.