[OpenRISC] [PATCH v4 04/13] or1k: startup and dynamic linking code

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 04/01/2022 00:08, Stafford Horne wrote:
> On Tue, Jan 04, 2022 at 10:28:15AM +0900, Stafford Horne wrote:
>> On Mon, Jan 03, 2022 at 03:17:15PM -0300, Adhemerval Zanella wrote:
>>>
>>>
>>> On 29/12/2021 01:42, Stafford Horne via Libc-alpha wrote:
>>>> +/* Return nonzero iff ELF header is compatible with the running host.  */
>>>> +static inline int __attribute__ ((unused))
>>>> +elf_machine_matches_host (const Elf32_Ehdr *ehdr)
>>>> +{
>>>> +  return ehdr->e_machine == EM_OPENRISC;
>>>> +}
>>>> +
>>>> +static inline Elf32_Addr *
>>>> +or1k_get_got (void)
>>>> +{
>>>> +  Elf32_Addr *got;
>>>> +
>>>> +  asm ("l.jal    0x8\n"
>>>> +       " l.movhi %0, gotpchi(_GLOBAL_OFFSET_TABLE_-4)\n"
>>>> +       "l.ori    %0, %0, gotpclo(_GLOBAL_OFFSET_TABLE_+0)\n"
>>>> +       "l.add    %0, %0, r9\n"
>>>> +       : "=r" (got) : : "r9");
>>>
>>>
>>> I think this requires a "memory" clobber.
>>
>> Why would it?  This does not do any memory loads or stores.
>>
>> Let me test with "memory" clobber and see what the difference is.
> 
> If seem that with the "memory" clobber the compiler can no longer optimize away
> calling or1k_get_got the second time in elf_machine_load_address.
> 
> I will leave it out unless there is a good reason for it.

Fair enough, I had the impression it does access memory but it seems it
just my lack of or1k assembly knowledge. Sorry for the noise.