------- Comment #2 from lloyd at randombit dot net 2007-06-30 22:11 ------- The behavior still exists in the 4.3 20070622 snapshot. It does not occur using -march=core2 (the actual CPU in question). The bad value results when using -ftree-vectorize and -march or -mtune =nocona. -O, -O2, or -O3 is also required; a binary compiled with -Os produces the correct value.
-- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31966