Re: [Qemu-devel] [RFC] target-ppc/fpu_helper.c: Use C99 code to speed up floating point unit

[Programmingkid <programmingkidx@gmail.com>]

On Dec 3, 2016, at 3:44 AM, Paolo Bonzini wrote:

> 
> 
> On 03/12/2016 06:59, Programmingkid wrote:
>> The floating point code used in fpu_helper.c can be sped up by using
>> the IEEE 754 support added to the C99 standard. To test this code out
>> simply set and unset the I_NEED_SPEED macro. The program to test out
>> each version of the helper_fmadd() function is below the patch. It
>> needs to be ran in the guest. The emulator to use is qemu-system-ppc.
>> I used a Mac OS X guest, but the test program would compile on a
>> Linux guest.
>> 
>> This patch does make the fused multiply-add instruction fmadd work
>> faster and still give a correct result.
>> 
>> This documentation might be of help to those who want to learn more
>> about C99's IEEE 754 support: 
>> http://grouper.ieee.org/groups/754/meeting-materials/2001-07-18-c99.pdf
> 
> You're undoing what was done in 2007:
> 
> commit 76a66253e5e48f1744f689041c1c21cedcaff630
> Author: j_mayer <j_mayer@c046a42c-6fe2-441c-8c8c-71466251a162>
> Date:   Wed Mar 7 08:32:30 2007 +0000
> 
>    Great PowerPC emulation code resynchronisation and improvments:
> ...
>    - Micro-operation fixes:
> ...
>      * use softfloat routines for all floating-point operations
> 
> Paolo

Yes it would be. The commit message never stated why he wanted to switch
to floating point softfloat routines. These are my guesses:

- Different versions of gcc might have produced different results?
- Different hosts produced different results and instead wanted consistency?
- Author did not know about the C99 numerics support?

The C99 standard will keep results accurate on both different versions of 
gcc and different host architecture. It will also give us speed over
the softfloat routines.