Re: [Qemu-devel] [PATCH] target-tilegx: Execute _start and reach to __libc_start_main successfully

[Chen Gang S]

On 02/27/2015 12:31 AM, Richard Henderson wrote:
> On 02/25/2015 03:44 PM, Chen Gang S wrote:
>> OK, thanks. After check ISA document again, for me, we have to still use
>> "y0, y1, y2", e.g. de5e598034ac3000 { fnop ; jalr r12 ; st r10, r11 }
>>
>>  If y0 -> y1 -> y2:
>>
>>  - if jalr succeeds, it will write pc to sp stack, but sp is not changed
>>    (just like lr, pc, they are buffered to tcg temporary variables).
>>
>>  - if st fails, as the result, we can still say the whole bundle is not
>>    execute (it has already written pc to sp stack, but sp isn't changed,
>>    so it is still OK).
>>
>>  If y0 -> y2 -> y1:
>>
>>  - if st succeeds, it will write data to the useful memory.
>>
>>  - if jalr fails (e.g. sp stack is full, which may cause memory access
>>    issue), we can not restore the bundle.
> 
> You need to re-check the ISA document.  JALR does not write to the "real" 
> stack
> at all, and cannot raise any kind of exception.
> 
> Section 2.1.2.3 clearly defines pushReturnStack as part of the branch
> prediction mechanism on the cpu.  It can be completely ignored for QEMU.
> 

OK, thanks. What you said above sounds reasonable to me.


Thanks.
-- 
Open, share, and attitude like air, water, and life which God blessed.