x86 (32-bit only) has its own implementations of cexp, cexpf and cexpl which use fsincos with inaccurate argument reduction and so suffer from the same problems sincos did when it used fsincos (e.g., cexp (0x1p1023 * I) has no relation to the correct value). These should probably be removed (with tests for arguments with large imaginary part added similar to those for sincos) so the generic C version is used for x86.
Fixed by: commit 1a4ac776ebc9bb07287f59f63e473db531318dff Author: Joseph Myers <joseph@codesourcery.com> Date: Wed Mar 21 15:28:05 2012 +0000 Remove inaccurate x86 cexp implementations (bug 13883).