Re: [Qemu-devel] [PATCH v2 4/7] softfloat: fallback to __int128 maths for s390x and others

[Alex Bennée]

Richard Henderson <address@hidden> writes:

> On 1/17/19 7:23 AM, Alex Bennée wrote:
>> Apparently some versions of clang can't handle inline assembly with
>> __int128 parameters, especially on s390. Instead of hand-coding the
>> s390 divide provide a generic fallback for anything that provides
>> __int128 capable maths.
>>
>> Signed-off-by: Alex Bennée <address@hidden>
>> Cc: Thomas Huth <address@hidden>
>> ---
>>  include/fpu/softfloat-macros.h | 10 ++++------
>>  1 file changed, 4 insertions(+), 6 deletions(-)
>>
>> diff --git a/include/fpu/softfloat-macros.h b/include/fpu/softfloat-macros.h
>> index b1d772e6d4..1a43609eef 100644
>> --- a/include/fpu/softfloat-macros.h
>> +++ b/include/fpu/softfloat-macros.h
>> @@ -641,12 +641,6 @@ static inline uint64_t udiv_qrnnd(uint64_t *r, uint64_t 
>> n1,
>>      uint64_t q;
>>      asm("divq %4" : "=a"(q), "=d"(*r) : "0"(n0), "1"(n1), "rm"(d));
>>      return q;
>> -#elif defined(__s390x__)
>> -    /* Need to use a TImode type to get an even register pair for DLGR.  */
>> -    unsigned __int128 n = (unsigned __int128)n1 << 64 | n0;
>> -    asm("dlgr %0, %1" : "+r"(n) : "r"(d));
>> -    *r = n >> 64;
>> -    return n;
>>  #elif defined(_ARCH_PPC64) && defined(_ARCH_PWR7)
>>      /* From Power ISA 2.06, programming note for divdeu.  */
>>      uint64_t q1, q2, Q, r1, r2, R;
>> @@ -663,6 +657,10 @@ static inline uint64_t udiv_qrnnd(uint64_t *r, uint64_t 
>> n1,
>>      }
>>      *r = R;
>>      return Q;
>> +#elif defined(CONFIG_INT128)
>> +    unsigned __int128 n = (unsigned __int128)n1 << 64 | n0;
>> +    *r = n % d;
>> +    return n / d;
>>  #else
>
> I thought that we'd shown that, at least at present, no compiler is taking
> advantage of hardware insns for this, and is promoting this to a full 128-bit
> divide.  And further that the version using 64-bit arithmetic was competitive
> with the hardware insn.

Yeah it seems so. While Thomas' numbers weren't convincing the
CONFIG_INT128 fallback did trigger on my SynQuacer an knocked off about
2 MFlops of it's admittedly slow performance. Amusingly of course it's
faster under translation because of the hardware fall back:

07:44:44 address@hidden:~/l/q/t/fp] (8973c1e5…) + ./fp-bench -o div -p double
13.28 MFlops
07:44:49 address@hidden:~/l/q/t/fp] (8973c1e5…) + ./fp-bench -o div -p double 
-t host
498.20 MFlops
07:44:53 address@hidden:~/l/q/t/fp] (8973c1e5…) + 
../../aarch64-linux-user/qemu-aarch64  ./fp-bench -o div -p double -t host
52.71 MFlops

I'll drop this and use Thomas' #elif defined(__s390x__) &&
!defined(__clang__) version in the pull-request.

--
Alex Bennée