[PATCH v1 4/4] LoongArch: Add testsuits for TLSDESC in gas and ld.

[mengqinggang]

在 2023/12/2 下午3:44, Xi Ruoyao 写道:
> On Sat, 2023-12-02 at 15:38 +0800, Xi Ruoyao wrote:
>> On Fri, 2023-12-01 at 17:04 +0800, Lulu Cai wrote:
>>> +	pcalau12i       $a0,%desc_pc_hi20(var)
>>> +	ld.d    $a1,$a0,%desc_ld_pc_lo12(var)
>>> +	addi.d  $a0,$a0,%desc_add_pc_lo12(var)
>>> +	jirl    $ra,$a1,%desc_call(var)
>> Should we consider using a different link register for desc_call?  Say
>>
>> jirl $t0,$a1,%desc_call(var)
>>
>> And when the TLS descriptor function returns, use "jr $t0" instead of
>> "ret".  The advantage would be avoiding saving $ra just because TLS
>> descriptor is used.
> Hmm, maybe better:
>
> jirl $a1, $a1, %desc_call(var)
>
> Just reuse a1 as it's clobbered anyway.  Then __tlsdesc_static becomes
> something like
>
> __tlsdesc_static:
>      ld.d    $a0, $a0, 0
>      jr      $a1


According to my colleague's suggestion, the jirl instruction must use 
the $ra register (jirl $zero, $ra, 0)
in order to perform hardware instruction prefetching.
We need to do some performance testing to decide which register to use.


>