[Gcl-devel] Re: mips64 assembler

[Camm Maguire]

Greetings!

David Daney <address@hidden> writes:

> On 10/15/2010 10:11 AM, Camm Maguire wrote:
>> Greetings!  What is wrong with this stub attempting a jump to contents
>> of register $t0?
>>
>> (gdb) p/x *(ul *)address@hidden
>> $3 = {0x3c080077, /*lui t0,0x77*/
>>        0x2508a170, /*addui t0,t0,0xa170*/
>>        0x8d08a288, /*lw t0,-23928(t0) */
>>        0x1000008   /* jr t0*/ }
>
> What instruction follows this JR?  It will always be executed in the
> delay slot of the jump.  If it causes an exception, like BREAK or an
> illegal instruction would, $pc will be left pointing at the JR so that
> it can retry the whole thing once the mess is cleaned up.
>
> David Daney
>

Thanks so much -- guessed this from an earlier post you sent shortly
before receiving this.  Am now uploading gcl native relocation for
mips using a local got table for each loaded .o file.  Looks solid on
gabrielli. 

The idea is to write a little stub in the got table for entries
referring to e.g. _setjmp.  In the stub I load the address from the
global table into t9 and jump.  Presupposes running with LD_BIND_NOW,
as a pointer to the .MIPS.stub will not work as the gp register has
been reset.  Wish I knew how to do this from within the C program
instead of having to rely on shell script wrappers.

I still overwrite GOT16/LO16 with HI16/LO16 on 32bit, as I am too lazy
at the moment to understand what the meaning of addends in the code:

    lw t0,1(gp)   R_MIPS_GOT16 /*add one page to the symbol address?*/
    addiu t0,t0,xxxx R_MIPS_LO16 /*presumably this needs to push the carry*/
                                 /*bit above like with HI*/

I'm guessing that any GOT16 needs a got entry for each symbol times
the highest addend used, on one loads addr,addr+4096,....

This snippet summarizes the reloc logic:

#define GOT_RELOC(r) 
(ELF_R_TYPE(r->r_info)==R_MIPS_CALL16||ELF_R_TYPE(r->r_info)==R_MIPS_GOT16)

    case R_MIPS_JALR:
      break;
    case R_MIPS_32:
      add_val(where,~0L,s);
      break;
    case R_MIPS_GOT16:
    case R_MIPS_CALL16:
      if (!sym1[ELF_R_SYM(r->r_info)].st_shndx) { /*Better way to
                                                    separate GOT16
                                                    from GOT16/LO16? */
        gote=got;
        got+=sym1[ELF_R_SYM(r->r_info)].st_size-1;
        *got=s;
        store_val(where,MASK(16),(got-gote)*sizeof(*got));
        if (s>=stub1 && s<stube)
          massert(!write_stub(s,got));
        break;
      }
      massert(ELF_R_TYPE(r->r_info)==R_MIPS_GOT16);
      store_val(where,0xffe00000,0x3c000000); 
      r->r_info=ELF_R_INFO(ELF_R_SYM(r->r_info),R_MIPS_HI16);
    case R_MIPS_HI16:
      if (!s) s=gpd=(ul)got-(ul)where;
      if (!hr) hr=r;
      if (a) add_vals(where,MASK(16),(s>>16)+a);
      break;
    case R_MIPS_LO16:
      if (!s) s=gpd;
      a=*where&MASK(16);
      if (a&0x8000) a|=0xffff0000; 
      a+=s&MASK(16);
      a+=(a&0x8000)<<1; 
      store_val(where,MASK(16),a);
      a=0x10000|(a>>16);
      for (hr=hr ? hr : r;--r>=hr && ELF_R_TYPE(r->r_info)==R_MIPS_HI16;)
        relocate(sym1,r,a,start,got,gote);
      hr=NULL;gpd=0;
      break;


Take care,

>>
>> (gdb) c
>> Program received signal SIGTRAP, Trace/breakpoint trap.
>> 0x01094148 in ?? () /*this is the jump instruction*/
>> (gdb) i reg t0
>> t0: 0x2ae13bf0      /*this is the correct address*/
>>
>> Any help most appreciated!
>>
>> Take care,
>>
>>
>> David Daney<address@hidden>  writes:
>>
>>> On 09/22/2010 04:05 PM, David Daney wrote:
>>> [...]
>>>> [11] .init PROGBITS 10000458 000458 000078 00 AX 0 0 8
>>>> [12] .plt PROGBITS 100004e0 0004e0 000030 00 AX 0 0 32
>>>> [13] .text PROGBITS 10000510 000510 0002c0 00 AX 0 0 16
>>>> [14] .MIPS.stubs PROGBITS 100007d0 0007d0 000020 00 AX 0 0 4
>>>> [15] .fini PROGBITS 100007f0 0007f0 000048 00 AX 0 0 8
>>>> [16] .rodata PROGBITS 10000838 000838 000020 00 A 0 0 8
>>>> [17] .eh_frame PROGBITS 10000858 000858 000004 00 A 0 0 4
>>>> [18] .ctors PROGBITS 1001085c 00085c 000008 00 WA 0 0 4
>>> [...]
>>>>
>>>>
>>>> Look there, the PLT is section 17.
>>>>
>>>>
>>>
>>> Well really it is section 12.  But what is a difference of 5 among friends?
>>>
>>> David Daney
>>>
>>>
>>>
>>>
>>>
>>
>
>
>
>
>

-- 
Camm Maguire                                        address@hidden
==========================================================================
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah