Re: How to define a symbol with absolute address for AArch64?

[Richard Earnshaw <Richard.Earnshaw@foss.arm.com>]

On 12/09/2023 15:23, Sebastian Huber wrote:
> On 12.09.23 15:46, Xi Ruoyao wrote:
>> On Tue, 2023-09-12 at 14:58 +0200, Sebastian Huber wrote:
>>> On 12.09.23 13:21, Richard Earnshaw wrote:
>>>>
>>>> On 12/09/2023 11:02, Sebastian Huber wrote:
>>>>> Hello,
>>>>>
>>>>> I would like to define a global symbol with an absolute address in an
>>>>> assembly/C source file for the AArch64 target. This works for all
>>>>> other architectures I tried so far, but not for AArch64:
>>>>>
>>>>> extern char abs_symbol[];
>>>>> extern char abs_symbol_2[];
>>>>>
>>>>> __asm__(
>>>>>     "\t.globl abs_symbol\n"
>>>>>     "\t.set abs_symbol, 0x123\n"
>>>>> );
>>>>>
>>>>> unsigned long f_abs_symbol(void)
>>>>> {
>>>>>     return (unsigned long)abs_symbol;
>>>>> }
>>>>>
>>>>> unsigned long f_abs_symbol_2(void)
>>>>> {
>>>>>     return (unsigned long)abs_symbol_2;
>>>>> }
>>>>>
>>>>> unsigned long _start(void)
>>>>> {
>>>>>     return f_abs_symbol() + f_abs_symbol_2();
>>>>> }
>>>>>
>>>>> aarch64-rtems6-gcc abs.c -Wl,--gc-sections 
>>>>> -Wl,--defsym=abs_symbol_2=291
>>>> Have you tried -mcmodel=large?  With that I get:
>>> I get the same result with -mcmodel=large. With -mcmodel=tiny I get:
>>>
>>> aarch64-rtems6-gcc abs.c -Wl,--gc-sections -Wl,--defsym=abs_symbol_2=291
>>> -mcmodel=tiny
>>> /tmp/ccKUnvyq.o: in function `f_abs_symbol_2':
>>> abs.c:(.text+0x8): relocation truncated to fit: R_AARCH64_ADR_PREL_LO21
>>> against symbol `abs_symbol_2' defined in*ABS*  section in a.out
>>> collect2: error: ld returned 1 exit status
>>>
>>> Is this a tool bug?
>> No, it's how code models are defined.  GCC documentation says clearly:
>>
>>         -mcmodel=tiny
>>             Generate code for  the  tiny  code  model.   The  program  
>> and  its
>>             statically  defined  symbols  must  be  within  1MB  of 
>> each other.
>>             Programs can be statically or dynamically linked.
>>
>> Here the text is located at 0x400000 but abs_symbol_2 is at 0x123, thus
>> violating the definition of -mcmodel=tiny.
> 
> Yes, this makes sense.
> 
>>
>>> Is there some way to make this working with -mcmodel=small?
>> No because -mcmodel=small only assumes the program and the symbols are
>> within*a*  4GB range - for example it's allowed to be [47GB, 51GB).
>> This is different from the default code model of RISC-V (-
>> mcmodel=medlow) where the symbols must be in [-2GB, 2GB).
>>
>> If you really think GCC should support this you can ask GCC for adding a
>> new code model.  Anyway this is not a linker issue because the adrp-add
>> pairs are generated by GCC.
> 
> Thanks for the explanation.
> 
> I guess what reliably works across code models is using the address of 
> an existing symbol and then add a constant which fits into the code 
> model. To retrieve the constant, we just have to subtract the address of 
> the other symbol.
> 

There are various ways you can code this:

1) An alternative when using the large code model is to just write

extern char abs_symbol[];

char * const abs_sym_ptr = abs_symbol;

unsigned long f_abs_symbol(void)
{
   return (unsigned long)abs_sym_ptr;
}

Which assembles to:
0000000000400000 <f_abs_symbol>:
   400000:       90000000        adrp    x0, 400000 <f_abs_symbol>
   400004:       91004000        add     x0, x0, #0x10
   400008:       f9400000        ldr     x0, [x0]
   40000c:       d65f03c0        ret
   400010:       00000123        .word   0x00000123
   400014:       00000000        .word   0x00000000

2) If you really want to use other code models, you can try things like:

extern char abs_symbol[];

char * const volatile  abs_sym_ptr = abs_symbol;

unsigned long f_abs_symbol(void)
{
   return (unsigned long)abs_sym_ptr;
}

The "const volatile" forces the compiler not to try to inline the 
pointer value dereference, so you end up with:

0000000000400000 <f_abs_symbol>:
   400000:       90000000        adrp    x0, 400000 <f_abs_symbol>
   400004:       f9400c00        ldr     x0, [x0, #24]
   400008:       d65f03c0        ret

...

Disassembly of section .rodata:

0000000000400018 <abs_sym_ptr>:
   400018:       00000123 00000000                       #.......


In other respects, though, the pointer is const, so attempts to modify 
it via normal C code will be inhibited by the compiler.

This should work with both the small and tiny models as well as the 
large model.

R.