Re: libjava configure question

[Andrew Haley <aph at redhat dot com>]

Andrew Haley writes:
 > Bryce McKinlay writes:
 >  > Andrew Haley wrote:
 >  > 
 >  > >Tom Tromey writes:
 >  > > > >>>>> "Andrew" == Andrew Haley <aph@redhat.com> writes:
 >  > > > 
 >  > > > Andrew> One other thing: I don't think we can do this in
 >  > > > Andrew> general and meet even the more relaxed requirements of
 >  > > > Andrew> java.lang.Math, because the x86 has a restricted range
 >  > > > Andrew> of allowable arguments to its trig functions.
 >  > > > 
 >  > > > I thought the Math restrictions were designed to explicitly allow use
 >  > > > of the FPU on x86.
 >  > >
 >  > >Yes, but not without some wrapper code.  The crucial issue here is
 >  > >full range reduction, and the floating-point instructions don't do
 >  > >that by themselves.
 >  > 
 >  > I'm sure the math library functions do, though - we'd still likely
 >  > get a performance improvement from using those directly, even
 >  > without inlining of FP instructions - since they'll be optimized
 >  > for each architecture.
 > 
 > Yes, that's probably true.  We'd have to make sure that calls to
 > java::lang::Math::sin() etc. were replaced by sin() and we'd also have
 > either to remove our own definitions of the functions or make them
 > weak.
 > 
 > All quite possible, but perhaps not top of the list.

I just tried, and I discovered that even without any optimizations
gcc/glibc fails dismally to meet even the relaxed java.lang.Math
accuracy requirement for trig functions.  We should keep our own math
library, because it's better.

Andrew.