It's a regression in glibc, what do you want us to do?

Created attachment 65787 [details]
sincos

Would it be possible to use following assembler code?

Not really, we don't do assembler.

(In reply to comment #3)
> Not really, we don't do assembler.

That's not completely true, Albert: SplashMath.h uses already assembler code (under some circumstances). So it's at least thinkable to do something like 
#if __GNUC__ && __i386__
// use __asm__
#else
// use c-code
#endif
But I would agree if You mean that we prefer to wait for a glibc correction.

As said, try to get the glibc people to fix their stuff, i'm sure we're not the only ones suffering from this performance degradation.