On Sat, May 26, 2012 at 5:00 AM, Eric Botcazou <ebotca...@adacore.com> wrote: > The problem is that the new call to cleanup_cfg in gimple_expand_cfg has > short-circuited the machinery that emits nops to carry goto locus at -O0. > The machinery works in CFGLAYOUT mode, but here we're still in CFGRTL. > > The attached patch makes it so that forwarder blocks are not deleted by > try_optimize_cfg if CLEANUP_NO_INSN_DEL and partially extends the above > machinery to the CFGRTL mode. > > Bootstrapped/regtested on x86_64-suse-linux, applied on the mainline. > > > 2012-05-26 Eric Botcazou <ebotca...@adacore.com> > > * cfgcleanup.c (try_optimize_cfg): Do not delete forwarder blocks > if CLEANUP_NO_INSN_DEL. > * cfgrtl.c (unique_locus_on_edge_between_p): New function extracted > from cfg_layout_merge_blocks. > (emit_nop_for_unique_locus_between): New function. > (rtl_merge_blocks): Invoke emit_nop_for_unique_locus_between. > (cfg_layout_merge_blocks): Likewise. > >
This caused: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54456 -- H.J.
