It seems that gcc-4.1 is getting getting more agressive using the powerpc fused-multiply-add (fmadd) which impacts the code gen for the bessel functions in ./dbl-64/e_j1.c, ./dbl-64/e_j1.c, ./flt-32/e_j1f.c, ./flt-32/e_j1f.c. The changes the rounding a bit and results in a 1-2 bit (ulp) difference in the result from the code generated by gcc-3.4. This only seemd to impacts the y1() yn() functions.
Created attachment 677 [details] patch for sysdeps/powerpc/fpu/libm-test-ulps This patch expands the results range by 1-2 ulps so that test-float and test-double pass when glibc is built with either gcc-3.4 or gcc-4.1
Subject: Bug 1384 CVSROOT: /cvs/glibc Module name: libc Changes by: roland@sources.redhat.com 2005-09-28 20:05:13 Modified files: sysdeps/powerpc/fpu: libm-test-ulps Log message: 2005-09-26 Steven Munroe <sjmunroe@us.ibm.com> [BZ #1384] * sysdeps/powerpc/fpu/libm-test-ulps: Adjust expected y1, y1f, yn, ynf results. Patches: http://sources.redhat.com/cgi-bin/cvsweb.cgi/libc/sysdeps/powerpc/fpu/libm-test-ulps.diff?cvsroot=glibc&r1=1.13&r2=1.14
Apparently checked in.
Subject: Bug 1384 CVSROOT: /cvs/glibc Module name: libc Branch: glibc-2_3-branch Changes by: roland@sources.redhat.com 2005-10-17 04:33:03 Modified files: sysdeps/powerpc/fpu: libm-test-ulps Log message: 2005-09-26 Steven Munroe <sjmunroe@us.ibm.com> [BZ #1384] * sysdeps/powerpc/fpu/libm-test-ulps: Adjust expected y1, y1f, yn, ynf results. Patches: http://sources.redhat.com/cgi-bin/cvsweb.cgi/libc/sysdeps/powerpc/fpu/libm-test-ulps.diff?cvsroot=glibc&only_with_tag=glibc-2_3-branch&r1=1.11.2.1&r2=1.11.2.2