> > --Boundary-00=_dFRYDl9RKCduOdb > Content-Type: text/plain; > charset="us-ascii" > Content-Transfer-Encoding: 7bit > Content-Disposition: inline > > Hi, > > We have run into an excess precision problem on IA-64, similar to the > infamous > problem on IA-32. A C testcase is attached. > > [EMAIL PROTECTED]:~/EA26-004> gcc -v > Using built-in specs. > Target: ia64-sgi-linux-gnu > Configured with: /home01/botcazou/cvs/gcc/configure ia64-sgi-linux-gnu > --prefix=/nile.build/botcazou/fsf/install_ia64 --disable-nls > --disable-libmudflap --disable-checking --enable-languages=c,c++ > Thread model: posix > gcc version 4.1.0 20051024 (experimental) > [EMAIL PROTECTED]:~/EA26-004> gcc -o t t.c > [EMAIL PROTECTED]:~/EA26-004> ./t > [EMAIL PROTECTED]:~/EA26-004> gcc -o t t.c -O > [EMAIL PROTECTED]:~/EA26-004> ./t > Aborted > > The bottom line is that: > > fmpy.s f8 = f8, f15 > fma.s f8 = f11, f12, f8 > > doesn't always give 0.0 in f8 when f8 = f15 = f11 = -f12, because the > computations are done in "infinite precision" and only rounded at the end. > > > What should we do about that? Conditionalize the combined FP operations on > -ffast-math or something along these lines?
This seems like any other target which has a fused multiply and add instruction like PPC. Maybe a target option to turn on and off the fma instruction like there is for PPC. -- Pinski
