http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57763
--- Comment #18 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Created attachment 31852 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=31852&action=edit gcc49-pr57763.patch Untested fix. So, from what I can understand, the really computed jumps are just fine, the problem is that on some targets (well, most of them except for a few exceptions that define HAS_LONG_UNCOND_BRANCH to non-zero) fix_crossing_unconditional_branches turns the unconditional normal crossing jumps into indirect jumps. Those then satisfy computed_jump_p predicate, and if current function has forced_labels, then make_edge later on wants to see edges from those jumps to all the forced_labels. But, while those crossing jumps are indirect jumps, they actually can't jump to arbitrary labels, they will only jump to the single label they load a few insns before the indirect jump into register, so there is no point in having edges from them to all the forced_labels, the insns are indirect just because normal uncondjumps aren't guaranteed to be able to jump to arbitrary code distances. Attached is an attempt to fix this just by setting JUMP_LABEL on the indirect jump, then computed_jump_p predicate will be false and the various places in RTL optimizers will be told where it jumps to. Not 100% how adding the JUMP_LABEL plays with e.g. LABEL_NUSES, should it increment it or not, or do only LABEL_REFs count for that? In any case, the patch fixes the bug in the cross compiler, can't test it on Alpha though, perhaps could test say on ppc64 and s390x which also don't define HAS_LONG_UNCOND_BRANCH to non-zero. But, our testsuite coverage for -freorder-blocks-and-partition is very limited anyway.
