https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95654
--- Comment #4 from Tobias Burnus <burnus at gcc dot gnu.org> --- (In reply to Tobias Burnus from comment #3) > Created attachment 49222 [details] > Slightly reduced example, compile with gfortran -fopenmp -O1 -ftracer On the host side, a single BB gets inserted – but more interesting is the device side: For the nvptx lto1, all dumped trees are identical until profile_estimate. The result (the non-64 value) slightly varies, but one can disable disable app three passes until tree-tracer to get still a fail: gfortran -fopenmp -ftracer -O1 test.f90 -foffload='-fdisable-tree-dom3 -fdisable-tree-profile_estimate -fdisable-tree-fixup_cfg3 -fdisable-tree-strlen1 -fdisable-tree-copyprop5 -fdisable-tree-wrestr^Ct -fdisable-tree-dse3 -fdump-tree-all -fdisable-tree-cddce3 -fdisable-tree-forwprop4 -fdisable-tree-phiopt4 -fdisable-tree-fab1 -fdisable-tree-dce7 -fdisable-tree-crited1 -fdisable-tree-uncprop1 -fdisable-tree-local-pure-const2 -fdisable-tree-nrv -fdisable-tree-isel -fdisable-tree-optimized' Adding -foffload=-fdisable-tree-tracer unsurprisingly makes all tree dumps identical and yields the expected 64.