Hi! When create_preheader is called with CP_FALLTHRU_PREHEADERS and the loop header is cold, but the bb dominating the loop is hot (or vice versa) and there are just 2 incoming edges into the header, we use split_edge, which unfortunately make the new bb use the partition of the source rather than destination, so this newly created preheader is never fallthru without jump, because the edge is EDGE_CROSSING.
Fixed by forcing to use make_forwarder_block in that case, that uses the partition of the bb on which it is called (i.e. loop header). Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2018-04-06 Jakub Jelinek <ja...@redhat.com> PR rtl-optimization/84872 * cfgloopmanip.c (create_preheader): Use make_forwarder_block even if nentry == 1 when CP_FALLTHRU_PREHEADERS and single_entry is EDGE_CROSSING edge. * gcc.dg/graphite/pr84872.c: New test. --- gcc/cfgloopmanip.c.jj 2018-01-03 10:19:56.065534103 +0100 +++ gcc/cfgloopmanip.c 2018-04-06 12:57:53.141922946 +0200 @@ -1494,7 +1494,9 @@ create_preheader (struct loop *loop, int mfb_kj_edge = loop_latch_edge (loop); latch_edge_was_fallthru = (mfb_kj_edge->flags & EDGE_FALLTHRU) != 0; - if (nentry == 1) + if (nentry == 1 + && ((flags & CP_FALLTHRU_PREHEADERS) == 0 + || (single_entry->flags & EDGE_CROSSING) == 0)) dummy = split_edge (single_entry); else { --- gcc/testsuite/gcc.dg/graphite/pr84872.c.jj 2018-04-06 13:06:07.017014715 +0200 +++ gcc/testsuite/gcc.dg/graphite/pr84872.c 2018-04-06 13:05:46.358011993 +0200 @@ -0,0 +1,19 @@ +/* PR rtl-optimization/84872 */ +/* { dg-do compile { target pthread } } */ +/* { dg-options "-O1 -floop-parallelize-all -freorder-blocks-and-partition -fschedule-insns2 -fselective-scheduling2 -fsel-sched-pipelining -fno-tree-dce" } */ + +void +foo (int x) +{ + int a[2]; + int b, c = 0; + + for (b = 0; b < 2; ++b) + a[b] = 0; + for (b = 0; b < 2; ++b) + a[b] = 0; + + while (c < 1) + while (x < 1) + ++x; +} Jakub