Re: [PATCH][BZ #19490] Add unwind descriptors for x86_64 _mcount and __fentry__

[Paul Pluzhnikov <ppluzhnikov at google dot com>]

On Mon, Feb 22, 2016 at 1:14 AM, Mike Frysinger <vapier@gentoo.org> wrote:
> On 23 Jan 2016 14:52, Paul Pluzhnikov wrote:
...
>> +     cfi_rel_offset (rdi, 32)
>> +     cfi_rel_offset (r8, 40)
>> +     cfi_rel_offset (r9, 48)
>
> don't we usually interleave the insns & cfi calls so that it's harder
> for them to get out of sync ?

I don't believe we are very consistent here. E.g. in sysdeps/i386/submul_1.S:

ENTRY (__mpn_submul_1)
        pushl   %edi
        cfi_adjust_cfa_offset (4)
        pushl   %esi
        cfi_adjust_cfa_offset (4)
        pushl   %ebp
        cfi_adjust_cfa_offset (4)
        pushl   %ebx
        cfi_adjust_cfa_offset (4)
        cfi_rel_offset (edi, 12)
        cfi_rel_offset (esi, 8)
        cfi_rel_offset (ebp, 4)
        cfi_rel_offset (ebx, 0)

But x86_64 does seem to interleave, so I've updated the patch to do the same.

>>       .type C_SYMBOL_NAME(__fentry__), @function
>>       .align ALIGNARG(4)
>>  C_LABEL(__fentry__)
>> -     /* Allocate space for 7 registers.  */
>> +     cfi_startproc
>> +     /* Allocate space for 7 registers (+8 for proper stack alignment).  */
>>       subq    $64,%rsp
>
> mmm, 56 is used above w/_mcount and is 8 byte aligned.  are you saying
> we need 16 byte alignment and thus _mcount should be fixed ?

No: the difference between _mcount and __fentry__ is that the former
is called after function prologue. That is, on entry into _mcount
(%RSP & 0xF) == 8 (as is usual for x86_64), but on entry into
__fentry__ (%RSP & 0xF) == 0 (which is special).

Note that I am only changing comment here, not the actual alignment.

Thanks,

2016-02-26  Paul Pluzhnikov  <ppluzhnikov@google.com>

        [BZ #19490]
        * sysdeps/x86_64/_mcount.S (_mcount): Add unwind descriptor.
        (__fentry__): Likewise

-- 
Paul Pluzhnikov

Attachment: glibc-bz19490-20160226.txt
Description: Text document