[Bug target/65496] New: ICE: in maybe_record_trace_start, at dwarf2cfi.c:2318 with -O3 -fsched2-use-superblocks -mavx512dq --param=max-pending-list-length=0

[Bug target/65496] New: ICE: in maybe_record_trace_start, at dwarf2cfi.c:2318 with -O3 -fsched2-use-superblocks -mavx512dq --param=max-pending-list-length=0

[zsojka at seznam dot cz]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65496

            Bug ID: 65496
           Summary: ICE: in maybe_record_trace_start, at dwarf2cfi.c:2318
                    with -O3 -fsched2-use-superblocks -mavx512dq
                    --param=max-pending-list-length=0
           Product: gcc
           Version: 5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zsojka at seznam dot cz

Created attachment 35075
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35075&action=edit
reduced testcase

Compiler output:
$ gcc -O3 -fsched2-use-superblocks -mavx512dq --param=max-pending-list-length=0
testcase.c         
testcase.c: In function 'foo':
testcase.c:6:1: internal compiler error: in maybe_record_trace_start, at
dwarf2cfi.c:2318
 }
 ^
0x8843d0 maybe_record_trace_start
        /mnt/svn/gcc-trunk/gcc/dwarf2cfi.c:2318
0x885348 scan_trace
        /mnt/svn/gcc-trunk/gcc/dwarf2cfi.c:2496
0x885c9a create_cfi_notes
        /mnt/svn/gcc-trunk/gcc/dwarf2cfi.c:2650
0x885c9a execute_dwarf2_frame
        /mnt/svn/gcc-trunk/gcc/dwarf2cfi.c:3006
0x885c9a execute
        /mnt/svn/gcc-trunk/gcc/dwarf2cfi.c:3486
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <http://gcc.gnu.org/bugs.html> for instructions.

$ gcc -v
Using built-in specs.
COLLECT_GCC=/mnt/svn/gcc-trunk/binary-latest/bin/gcc
COLLECT_LTO_WRAPPER=/mnt/svn/gcc-trunk/binary-221530-lto-fortran-checking-yes-rtl-df/libexec/gcc/x86_64-unknown-linux-gnu/5.0.0/lto-wrapper
Target: x86_64-unknown-linux-gnu
Configured with: /mnt/svn/gcc-trunk//configure --enable-checking=yes,rtl,df
--enable-languages=c,c++,lto,fortran
--prefix=/mnt/svn/gcc-trunk/binary-221530-lto-fortran-checking-yes-rtl-df/
--without-cloog --without-ppl
Thread model: posix
gcc version 5.0.0 20150320 (experimental) (GCC) 

Tested revisions:
r221530 - ICE

[Bug target/65496] ICE: in maybe_record_trace_start, at dwarf2cfi.c:2318 with -O3 -fsched2-use-superblocks -mavx512dq --param=max-pending-list-length=0

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65496

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-03-20
                 CC|                            |jakub at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Started with r217331.  Not really a regression, as GCC 4.9 doesn't have
-mavx512dq support.

[Bug target/65496] ICE: in maybe_record_trace_start, at dwarf2cfi.c:2318 with -O3 -fsched2-use-superblocks -mavx512dq --param=max-pending-list-length=0

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65496

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rth at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
And the problem is that with these weirdo scheduling options the scheduler
schedules the
(insn/f:TI 620 60 37 2 (set (reg:DI 39 r10)
        (plus:DI (reg/f:DI 7 sp)
            (const_int 8 [0x8]))) pr65496.c:3 214 {*leadi}
     (nil))
instruction before the shrink-wrapping conditional jump:
(jump_insn:TI 38 45 43 2 (set (pc)
        (if_then_else (eq (reg:CCZ 17 flags)
                (const_int 0 [0]))
            (label_ref:DI 645)
            (pc))) pr65496.c:4 613 {*jcc_1}
     (expr_list:REG_DEAD (reg:CCZ 17 flags)
        (int_list:REG_BR_PROB 900 (nil)))
 -> 645)
At the end of the function we have:
(insn/f:TI 641 640 645 40 (set (reg/f:DI 7 sp)
        (plus:DI (reg:DI 39 r10)
            (const_int -8 [0xfffffffffffffff8]))) pr65496.c:6 214 {*leadi}
     (expr_list:REG_DEAD (reg:DI 39 r10)
        (expr_list:REG_CFA_DEF_CFA (plus:DI (reg/f:DI 7 sp)
                (const_int 8 [0x8]))
            (nil))))
(code_label 645 641 644 41 106 "" [1 uses])
(note 644 645 663 41 [bb 41] NOTE_INSN_BASIC_BLOCK)
(jump_insn 663 644 643 41 (parallel [
            (simple_return)
            (unspec [
                    (const_int 0 [0])
                ] UNSPEC_REP)
        ]) pr65496.c:6 682 {simple_return_internal_long}
     (nil))
(barrier 643 663 355)

And the problem is that because the r10 = rsp + 8 frame related instruction is
moved before the shrink-wrapping jump we assume CFA is in r10 + 0, but before
fallthru into the return we have a CFA restore note to use r8 + 8 again as CFA
(not really needed in this case, but desirable for the case when the r10 = rsp
+ 8 instruction isn't moved.

Richard, any thoughts what to do about this?  Avoid scheduling frame related
instructions across conditional jumps?  Something else?

[Bug target/65496] ICE: in maybe_record_trace_start, at dwarf2cfi.c:2318 with -O3 -fsched2-use-superblocks -mavx512dq --param=max-pending-list-length=0

[rth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65496

--- Comment #3 from Richard Henderson <rth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #2)
> Richard, any thoughts what to do about this?  Avoid scheduling frame related
> instructions across conditional jumps?  Something else?

Yes, for the short term that will have to be a requirement in order to
keep the unwind info happy.

Something for the next stage1 is

(1) Avoiding the use of bare /f, and thus also REG_FRAME_RELATED_EXPR,
using instead always REG_CFA_*.  I suspect that this insn 620 shouldn't
actually be frame related at all, but is a part of a larger dwarf
expression we're intending to construct.

(2) Once we have an unambiguous note for everything, we can allow these
insns to be scheduled across a conditional jump if we strip and collect
those notes and place them after the conditional jump, probably on a
fake insn like

  (insn/f (use (const_int 0))
     (expr-list:REG_CFA_...
       (expr-list:REG_CFA_...
         (expr-list:REG_CFA_...))))