Re: [gold] 2.27 generates GOT where 2.26 did not

[Andrew Pinski <pinskia@gmail.com>]

On Thu, Aug 11, 2016 at 5:31 PM, Jethro Beekman <binutils@jbeekman.nl> wrote:
> I have these two files:
>
> // asm_var.S
> .section .rodata
> .global asm_var
> .protected asm_var
> .size asm_var, 8
> asm_var:
>         .org .+8
>
> // get_addr.c
> extern unsigned long asm_var;
>
> void* get_addr() {
>         return &asm_var;
> }
>
> That I compile with:
> gcc -O3 -fuse-ld=gold -fPIC -shared -nostdlib -Wl,-Bsymbolic asm_var.S get_addr.c
>
> With gold 2.26.1 (and most likely older versions as well), this results in:
>
> 00000000000002f0 <get_addr>:
>  2f0:   48 8d 05 01 00 00 00    lea    0x1(%rip),%rax    # 2f8 <asm_var>
>  2f7:   c3                      retq
>
> Yet with 2.27 this results in:
>
> 0000000000000340 <get_addr>:
>  340:   48 8b 05 99 1c 00 00    mov    0x1c99(%rip),%rax # 1fe0 <_DYNAMIC+0x110>
>  347:   c3                      retq
>
> What's at 1fe0?
>
> 000000001fe0  000500000006 R_X86_64_GLOB_DAT 0000000000000348 asm_var + 0
>
> Is this a regression in 2.27? I really don't want to use the GOT. Is there a
> flag I can pass to get the old behavior?

IIRC this is correct behavior for x86_64 due to copy relocs.

Thanks,
Andrew

>
> Jethro Beekman