https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83962
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2018-01-23
CC| |abel at gcc dot gnu.org,
| |amonakov at gcc dot gnu.org,
| |jakub at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Bisecting, but the ICE is very old.
Anyway, what I see is we have EDGE_FALLTHRU edge as the single edge out of a bb
not ending with any jump, then in reorder_basic_blocks ->
set_edge_can_fallthru_flag EDGE_CAN_FALLTHRU is ored with it. Then
cfg_layout_finalize -> fixup_reorder_chain -> force_nonfallthru_and_redirect
clears the EDGE_FALLTHRU bit in and adds an unconditional jump to the target
bb.
But then selective scheduling (sched2 pass) creates bogus:
(code_label 51 29 50 8 7 (nil) [0 uses])
(note 50 51 98 8 [bb 8] NOTE_INSN_BASIC_BLOCK)
(insn 98 50 82 8 (set (reg:SI 3 3 [orig:123 iftmp.0_4 ] [123])
(reg:SI 10 10 [orig:140 iftmp.0_4 ] [140])) 487 {*movsi_internal1}
(nil))
;; succ: 10 [always] (CAN_FALLTHRU)
;; lr out 1 [1] 3 [3] 7 [7] 8 [8] 9 [9] 10 [10] 73 [5]
(barrier 82 98 32)
;; basic block 10, loop depth 0, maybe hot
;; prev block 8, next block 1, flags: (REACHABLE, RTL, MODIFIED)
;; pred: 5 (CAN_FALLTHRU)
;; 7 (CAN_FALLTHRU)
;; 8 [always] (CAN_FALLTHRU)
;; 4 (CAN_FALLTHRU)
;; 6 (CAN_FALLTHRU)
;; bb 10 artificial_defs: { }
;; bb 10 artificial_uses: { u-1(1){ }}
;; lr in 1 [1] 3 [3] 7 [7] 8 [8] 9 [9] 10 [10] 73 [5]
;; lr use 1 [1] 3 [3]
;; lr def 3 [3] 4 [4] 76 [ca]
(code_label 32 82 33 10 3 (nil) [4 uses])
out of that, if the edge falls through into the next bb, it should be
EDGE_FALLTHRU, and the barrier should be left out, or there should be an
unconditional jump other passes could remove as unneeded later on.