------- Comment #2 from jakub at gcc dot gnu dot org 2008-07-02 09:22 -------
This is caused by the implicit gotos at the end of basic blocks.
See http://gcc.gnu.org/ml/gcc-patches/2007-04/msg01840.html
During cfgexpand this can be IMHO fixed by setting curr_location before
emitting the jump, not after it (otherwise e->goto_locus is IMHO useless,
as most likely the following statements will change curr_location before
anything is emitted):
--- gcc/cfgexpand.c.jj 2008-06-06 09:17:07.000000000 +0200
+++ gcc/cfgexpand.c 2008-07-02 10:43:54.000000000 +0200
@@ -1610,9 +1610,9 @@ expand_gimple_basic_block (basic_block b
if (e && e->dest != bb->next_bb)
{
- emit_jump (label_rtx_for_bb (e->dest));
if (e->goto_locus)
set_curr_insn_source_location (e->goto_locus);
+ emit_jump (label_rtx_for_bb (e->dest));
e->flags &= ~EDGE_FALLTHRU;
}
Not sure if expand_gimple_cond_expr should stay as is or be modified as well.
This fixes this problem for the few early RTL passes, but then we go into
cfglayout mode and the JUMP is removed again.
force_nonfallthru_and_redirect then when going out of cfglayout mode recreates
the JUMP again, but completely ignores e->goto_locus:
if (target == EXIT_BLOCK_PTR)
{
#ifdef HAVE_return
emit_jump_insn_after_noloc (gen_return (), BB_END (jump_block));
#else
gcc_unreachable ();
#endif
}
else
{
rtx label = block_label (target);
emit_jump_insn_after_noloc (gen_jump (label), BB_END (jump_block));
JUMP_LABEL (BB_END (jump_block)) = label;
LABEL_NUSES (label)++;
}
One problem is that we have just locus for the edges, not BLOCK, so if
force_nonfallthru_and_redirect were to call emit_jump_insn_after_setloc
instead,
the question is what BLOCK should be used for it (e.g. pick the block from insn
before the jump (if any)? What to do if there is none in the basic block?).
--
jakub at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rakdver at gcc dot gnu dot
| |org, hubicka at gcc dot gnu
| |dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36690
