On sparc, which uses ldbl-128, the tests using cos() on pi/2 give inaccurate results. For example: Test "cos (pi/2) == 0x1.cd129024e088a67cc74020bbea64p-115": ildouble: 267676 ldouble: 267676 ... Test "sincos (pi/2, &sin_res, &cos_res) puts 0x1.cd129024e088a67cc74020bbea64p-115 in cos_res": ildouble: 267676 ldouble: 267676 ... Function "cos": ildouble: 267676 ldouble: 267676 ... Function "sincos": ildouble: 267676 ldouble: 267676
I'll be disabling this test for ldbl-128 until we fix this.
Fixed for 2.18 by: commit d0213cd0b629cae95ed9607fcf26793a3a0d686c Author: Joseph Myers <joseph@codesourcery.com> Date: Thu May 9 21:28:54 2013 +0000 Fix ldbl-128 cos range reduction near pi/2 (bug 15429).