GCC's -ffast-math behavior

[Andrew Haley]

On 02/09/2012 10:07 AM, xunxun wrote:
> 于 2012/2/9 17:47, Andrew Haley 写道:
>> On 02/09/2012 09:35 AM, xunxun wrote:
>>> 于 2012/2/9 17:21, Andrew Haley 写道:
>>>> On 02/09/2012 08:33 AM, xunxun wrote:
>>>>>           When using -ffast-math, gcc don't generate the math function
>>>>> symbol: U _exp
>>>> No, it doesn't.  Instead gcc uses the F2XM1 instruction.  Why would
>>>> you want to call a library when gcc has an instruction to do the
>>>> job?
>>>>
>>> Because other math lib works faster than gcc itself (even with
>>> fastmath), and I want to use fastmath to make other caculation faster, too.
>> Hmm, I think that'll be difficult.  We tend to assume that when a
>> processor has built-in instructions to do something, that's the
>> fastest way to do it.  It's usually true, and I am wondering what
>> tricks Intel uses.  Granted, the floating-point transcendental
>> instructions aren't super-fast, and perhaps Intel doesn't optimize
>> them any more.
>>
> Thank you for the explanation.
> 
> I think I can separate all the math functions from other code, put the 
> math functions within one lib, and don't use fastmath to build the lib. :)

Okay.  Can you tell us how much faster than the builtins the Intel lib
actually, is, and how you measured that?

Andrew.