------- Comment #3 from rguenth at gcc dot gnu dot org 2007-11-10 14:31 ------- tree-ssa-sink is the culprit (we re-compute aliasing after PRE, right before sink, so an error there might be the bug as well):
Sinking newdim_31 = newdim_1 + -1 from bb 15 to bb 12 Sinking # SFT.2356_213 = VDEF <SFT.2356_209> D.45672_138->idx = D.45677_152 from bb 12 to bb 14 Sinking D.45677_152 = prephitmp.2402_141 from bb 12 to bb 14 Sinking # SFT.2356_209 = VDEF <SFT.2356_197> D.45676_147->val = 1.0e+0 from bb 12 to bb 14 Sinking # SFT.2355_200 = VDEF <SFT.2355_186> D.45672_138->val = 1.0e+0 from bb 10 to bb 14 Sinking # SFT.2354_136 = VDEF <SFT.2354_139> D.37036.D.34678.m_elem = D.45669_134 from bb 10 to bb 14 Sinking # SFT.2355_186 = VDEF <SFT.2355_161> D.37036.themem.val = 0.0 from bb 10 to bb 14 virtual void soplex::SoPlex::reDim() (this) at least ptr->val with only a single SFT looks suspicious. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34048
