Re: [Qemu-devel] [Consult] tilegx: About floating point instructions

[Chen Gang]

On 8/16/15 09:41, Chen Gang wrote:
> On 8/16/15 02:16, Chen Gang wrote:
>>
>> On 8/15/15 23:47, Richard Henderson wrote:
>>> On Aug 15, 2015 2:56 AM, Chen Gang <address@hidden>
>>>> Oh, we are unlucky, after continue gcc testsuite, add/sub floating point 
>>>> insns also can be mixed together! The related C code, -save-temps, and 
>>>> objdump files are in attachments (is it gcc's issue? I guess not). 
>>>>
>>>> So, I guess, we have to 'crack' all floating point insns, precisely, or 
>>>> we can not pass gcc testsuite. 
>>>>
>>>
>>> If you go back to my first message to you on the subject, you'll find that 
>>> my suggestion was to not split the operation at all, using move for pack1.  
>>> Which would nicely handle any such interleaving.
>>>
>>
>> OK, thanks, but for float(uns)sisf2 and float(uns)sidf2, we can not only
>> simply move.  :-(
>>
>> But what you said is really quite valuable to me!! we can treat the flag
>> as a caller saved context, then can let the caller can use callee freely
>> (in fact, I guess, the real hardware treats it as caller context, too).
>>
>>  - we have to define the flag format based on the existing format in the
>>    related docs and tilegx.md (reserve 0-20 and 25-31 bits).
>>
>>  - We can only use 21-24 for mark addsub, mul, or typecast result. If
>>    21-24 bits are all zero, it means typecast result. For fsingle: 32-63
>>    bits is the input integer; for fdouble: srca is the input integer.
>>
>>  - For addsub and mul result, we use 32-63 bits for an index of resource
>>    handler (like 'fd' returned by open). fsingle_addsub2, fsingle_mul1,
>>    fdouble_mul_flags, fdouble_addsub allocate resource, and pack1 free.
>>
>> But if caller "make mistakes", our implementation can not avoid related
>> resource leak (but the real hardware can, it also lets caller save all
>> related resources; when it needs them, it can let caller pass them to).
>>
> 
> If we assume that the optimization for the floating point insns can not
> cross the basic blocks (I guess so), we can reset all related resources
> when start a basic block.
> 

Oh, sorry, even qemu itself, my split a basic block into 2 basic blocks,
when the basic block is too big. And we also have to assume a same value
may call fdouble_pack1 individually with multiple times.

So for the resource management, we can do like this:

 - For fsingle, it can be saved in 32-63 bits of caller context (it is
   float32 which is 32-bit).

 - For fdouble, we can allocate a 'bit' buffer for it (e.g. 8KB), when
   the saved values count overflow 1K, let it roundup to 0 again -- of
   cause, the old 1Kth value should be already useless.

I guess, in this way, we can emulate the tilgex floating points insns!!
:-)


Thanks.
-- 
Chen Gang

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