------- Comment #37 from steven at gcc dot gnu dot org 2006-01-13 23:26 -------
I think I know what the problem is. At the point where we error in cfgrtl, we
have the following basic block:
Breakpoint 3, rtl_verify_flow_info_1 () at cfgrtl.c:2051
2051 error ("wrong amount of branch edges after conditional jump
%i", bb->index);
(gdb) l
2046 err = 1;
2047 }
2048 if (n_branch != 1 && any_condjump_p (BB_END (bb))
2049 && JUMP_LABEL (BB_END (bb)) != BB_HEAD (fallthru->dest))
2050 {
2051 error ("wrong amount of branch edges after conditional jump
%i", bb->index);
2052 err = 1;
2053 }
2054 if (n_call && !CALL_P (BB_END (bb)))
2055 {
(gdb) p debug_bb(bb)
;; basic block 7, loop depth 0, count 0
;; prev block 1, next block 9
;; pred: 1 [100.0%] (fallthru)
;; succ: 9 [100.0%] (fallthru) 8 (true)
;; Registers live at start: 1 [1] 31 [31] 67 [ap] 113 [sfp] 123 127 128 129
139
(note 89 68 90 7 [bb 7] NOTE_INSN_BASIC_BLOCK)
(insn 90 89 91 7 (set (reg:CC 140)
(compare:CC (reg:SI 139)
(const_int 1 [0x1]))) -1 (nil)
(nil))
(jump_insn 91 90 94 7 (set (pc)
(if_then_else (gt (reg:CC 140)
(const_int 0 [0x0]))
(label_ref:SI 93)
(pc))) 472 {*rs6000.md:12875} (nil)
(nil))
;; Registers live at end: 1 [1] 31 [31] 67 [ap] 113 [sfp] 123 127 128 129 139
Note the "(true)" flag on the edge to block 8. That is the EDGE_TRUE_VALUE
flag, which is not supposed to appear on trees. We fail to compute n_branch
properly because of this:
if ((e->flags & ~(EDGE_DFS_BACK
| EDGE_CAN_FALLTHRU
| EDGE_IRREDUCIBLE_LOOP
| EDGE_LOOP_EXIT
| EDGE_CROSSING)) == 0)
n_branch++;
Aha, but we have (e->flags & ~(EDGE_DFS_BACK
| EDGE_CAN_FALLTHRU
| EDGE_IRREDUCIBLE_LOOP
| EDGE_LOOP_EXIT
| EDGE_CROSSING)) == EDGE_TRUE_VALUE)
so n_branch stays at 0. Then we go on checking with the wrong n_branch, and we
get to the point where we fail.
My best bet: loop versioning still has bits that assume it works on trees only.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24626
