Hi!
As mentioned in the PR, on alpha (but I don't see a reason why it can't
occur on most other targets except for i?86/x86_64, cr16, m32c, moxie and
msp430) if a function has non-NULL forced_labels (contains computed goto),
we can get ICEs due to incorrect EDGE_CROSSING flags on some edges.
The problem is that if not HAS_LONG_UNCOND_BRANCH,
fix_crossing_unconditional_branches changes normal unconditional jumps into
indirect jumps that jump to a single label, because normal jump might not
be able to jump to very far other section. But, turning the jump into
indirect one means computed_jump_p is true on those, and such jumps are
supposed to have edges to all forced_labels, but
fix_crossing_unconditional_branches doesn't add those and just if say
post-reload splitter attempts to change something, we can end up calling
make_edges and that might not give the right flags to the edges.
As the indirect jump jumps to a single label only, I don't see why we should
require edges to all forced_labels, we know the single label to which it can
jump. So the following patch fixes that by setting JUMP_LABEL on the
indirect jump, which results e.g. into computed_jump_p no longer returning
true for it.
Bootstrapped/regtested on x86_64-linux and i686-linux (where it isn't
called, sure) and Uros has bootstrapped/regtested this on Alpha. Ok for
trunk?
2014-01-17 Jakub Jelinek <ja...@redhat.com>
PR rtl-optimization/57763
* bb-reorder.c (fix_crossing_unconditional_branches): Set JUMP_LABEL
on the new indirect jump_insn.
--- gcc/bb-reorder.c.jj 2014-01-03 11:41:01.000000000 +0100
+++ gcc/bb-reorder.c 2014-01-16 12:49:27.695377777 +0100
@@ -2183,6 +2183,8 @@ fix_crossing_unconditional_branches (voi
emit_insn_before (indirect_jump_sequence, last_insn);
delete_insn (last_insn);
+ JUMP_LABEL (jump_insn) = label;
+
/* Make BB_END for cur_bb be the jump instruction (NOT the
barrier instruction at the end of the sequence...). */
Jakub
