[Bug rtl-optimization/57268] [4.9 Regression] c nested loops hang compiler in sched-deps.c

["dtemirbulatov at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57268

Dinar Temirbulatov <dtemirbulatov at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dtemirbulatov at gmail dot com

--- Comment #3 from Dinar Temirbulatov <dtemirbulatov at gmail dot com> ---
(In reply to Jakub Jelinek from comment #2)
> This is weird, I've tried to bisect it, r197678 still compiled it without
> hanging, r197681 ICEd somewhere in lra-constraints.c, but if I rebuild
> lra-constraints.o with -O0, it instead ICEs on frame reg uses during
> final_scan_insns (i.e. LRA hasn't replaced the frame pointer with hard frame
> pointer or stack pointer), r197696 still ICEs, r197700 hangs.
> BTW, doing -da before it hangs, I'm seeing right in *.split4 dump REG_EQUIV
> notes like:
> (insn 20109 70 20110 19 (set (reg/f:DI 38 r9 [19927])
>         (plus:DI (reg/f:DI 7 sp)
>             (const_int 16 [0x10]))) 254 {*leadi}
>      (expr_list:REG_EQUIV (plus:DI (reg/f:DI 20 frame)
>             (const_int -8 [0xfffffffffffffff8]))
>         (nil)))
> wonder if it isn't a bug that RA hasn't replaced the eliminable register in
> the note with sp (or hfp).

Problem was triggered by
Author: amker <amker@138bc75d-0d04-0410-961f-82ee72b054a4>
Date:   Thu Apr 11 03:55:14 2013 +0000

         PR target/56124
         * ira-costs.c (scan_one_insn): Check whether the source rtx of
         loading has side effect.

And the problem in my view is in the scheduling 2 pass, it creates too long
read dependents list without flushing.