[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Lightning] Lightning extensions
From: |
Paulo César Pereira de Andrade |
Subject: |
Re: [Lightning] Lightning extensions |
Date: |
Thu, 23 Sep 2010 05:05:02 -0300 |
Em 21 de setembro de 2010 05:02, Paolo Bonzini <address@hidden> escreveu:
> On 09/12/2010 05:03 PM, Paulo César Pereira de Andrade wrote:
>>
>> It should be very easy, like reviewing 10K lines of assembly code;-)
>> Well, ok, a very tough job. Do you mind if I import the lightning test
>> tool I added to my project to my lightning tree in github? It should not
>> be installed, but building it should make testing very easier, as well as
>> writing small tests to check some construct.
>
> Go ahead.
Just did it :-) Initially, I added it to a new contrib/check directory, and
do not build it by default. For now it is there more for the commodity
of having it easily available.
I also just merged the work branch into the master branch in
http://github.com/pcpa/lightning
while it should still require a lot of work to review, I don't think I will
backup from the changes there at least for my own use, so, merged
to master :-) But I may remake patches for easier integration if
required.
Besides the pseudo assembler testing tool, I also added the new
jit_{sin,cos,tan}r_{f,d} for i386 and x86_64. They are undocumented,
and basically just enable the code that was previously under an "#if 0".
It uses the x87, and the inline functions sse_from_x87 and x87_from_sse
if using sse. It should match the documented procedure in the Intel
manual for values too large, but for those, does not match the software
implementation in x86_64 glibc, neither mpfr results. For the very uncommon
values I tested, it matches the values returned by glibc's libm 32 bit. It
should be small enough to be worth not making a function call, and even
better by ensuring register values are saved, but, nothing prevents one from
creating a jit function just to call it, as it does not destroy the contents of
any register.
> Paolo
Thanks,
Paulo