Re: [PATCH 01/05] Math optimizations for POWER8

[Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>]

On 29-03-2014 11:12, Andreas Schwab wrote:
> Adhemerval Zanella <azanella@linux.vnet.ibm.com> writes:
>
>> POWER8 follows ISA 2.07 and it adds the mfvsrd/mtvsrd instruction class to
>> transfer data between GRP register to/from FPR register without the need of
>> memory access. These instructions shows a better latency than the sequence
>> std,...,ld which may create load-hit-stores and stalls in some CPUs. 
>>
>> These patches add a slight more optimized implementations of
>> isnan/isinf/finite/llround/llrint using these instruction to gains some
>> cycles by avoiding the possible load-hit-store of FP<->INT conversion.
>>
>> Tested on powerpc64 and powerpc64le.
> Failure: float isinf (inf) == 1
> Failure: float isinf (-inf) == -1
> Failure: float isnan (qNaN)
> Failure: float isnan (sNaN)
> Failure: float isnan (-qNaN)
> Failure: float isnan (-sNaN)
> Failure: float isnan (NAN)
> Failure: float isnan (-NAN)
> Failure: float isnan (nan (""))
> Failure: float isnan (-nan (""))
> Failure: float isinf (HUGE_VALx) == +1
> Failure: float isinf (-HUGE_VALx) == -1
> Failure: double isinf (inf) == 1
> Failure: double isinf (-inf) == -1
> Failure: double isnan (qNaN)
> Failure: double isnan (sNaN)
> Failure: double isnan (-qNaN)
> Failure: double isnan (-sNaN)
> Failure: double isnan (NAN)
> Failure: double isnan (-NAN)
> Failure: double isnan (nan (""))
> Failure: double isnan (-nan (""))
> Failure: double isinf (HUGE_VALx) == +1
> Failure: double isinf (-HUGE_VALx) == -1
> Failure:  float x = (float) (double) qNaN, isnan
> Failure:  float x = (float) (double) sNaN, isnan
> Failure:  float x = (float) (double) Inf
> Failure:  double x = (double) (float) qNaN, isnan
> Failure:  double x = (double) (float) sNaN, isnan
> Failure:  double x = (double) (float) Inf
> Failure: long double isnan (NAN)
> Failure: long double isnan (-NAN)
> Failure:  float x = (float) (long double) sNaN, isnan
> Failure:  float x = (float) (long double) Inf
> Failure:  double x = (double) (long double) qNaN, isnan
> Failure:  double x = (double) (long double) sNaN, isnan
> Failure:  double x = (double) (long double) Inf
>
> Failure: float isnan (qNaN)
> Failure: float isnan (-qNaN)
> Failure: float isnan (sNaN)
> Failure: float isnan (-sNaN)
> Failure: float isfinite (qNaN)
> Failure: float isfinite (-qNaN)
> Failure: float isfinite (sNaN)
> Failure: float isfinite (-sNaN)
> Failure: double isnan (qNaN)
> Failure: double isnan (-qNaN)
> Failure: double isnan (sNaN)
> Failure: double isnan (-sNaN)
> Failure: double isfinite (qNaN)
> Failure: double isfinite (-qNaN)
> Failure: double isfinite (sNaN)
> Failure: double isfinite (-sNaN)
>
> Andreas.
>
Fix by 757d9dd5c3efa56fac75965abc014faaae7b7895. It was due a quick fix I did to make it
work by old binutils I wrongly assumed it would work effortlessly on LE environment. Thanks
for catching it.