http://gcc.gnu.org/bugzilla/show_bug.cgi?id=61058
--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Jeffrey A. Law from comment #3)
> What do you need me to confirm? I can confirm that you're not supposed to
> have BARRIERS in the middle of a block.
Confirmation of invalid barrier location, so this PR can be qualified as
rtl-optimization problem.
>
> THe RTL in question:
>
> .file "j.c"
> .text
> .globl foo
> .type foo, @function
> foo:
> pushq %rbp
> (note 1 0 3 NOTE_INSN_DELETED)
>
> (note 3 1 9 2 [bb 2] NOTE_INSN_BASIC_BLOCK)
>
> (insn/f 9 3 10 2 (set (mem:DI (pre_dec:DI (reg/f:DI 7 sp)) [0 S8 A8])
> (reg/f:DI 6 bp)) j.c:2 65 {*pushdi2_rex64}
> (nil))
>
> (insn/f 10 9 5 2 (set (reg/f:DI 6 bp)
> (reg/f:DI 7 sp)) j.c:2 89 {*movdi_internal}
> (nil))
>
> (barrier 5 10 11)
>
> (note 11 5 2 2 NOTE_INSN_PROLOGUE_END)
>
> (note 2 11 8 2 NOTE_INSN_FUNCTION_BEG)
>
> (note 8 2 0 NOTE_INSN_DELETED)
>
>
> We're in distance_agu_use_in_bb. We're passing it insn 10 as INSN and the
> BARRIER as START. We try to look at BLOCK_FOR_INSN (start), after that,
> we're well into undefined territory.
>
> But this really looks like x86 backend breakage. Refer to the above RTL and
> look at this call site:
>
> 17976 if (insn != BB_END (bb))
> 17977 distance = distance_agu_use_in_bb (regno0, insn, distance,
> 17978 NEXT_INSN (insn),
> 17979 &found, &redefined);
>
> Clearly if NEXT_INSN (insn) is a BARRIER, then nothing good can happen.
The accessors don't agree:
(gdb) p debug_rtx (insn)
(insn/f 10 9 5 2 (set (reg/f:DI 6 bp)
(reg/f:DI 7 sp)) pr61058.c:2 89 {*movdi_internal}
(nil))
$3 = void
(gdb) p debug_rtx (bb->il.x.head_)
(note 3 1 9 2 [bb 2] NOTE_INSN_BASIC_BLOCK)
$4 = void
(gdb) p debug_rtx (bb->il.x.rtl->end_)
(note 2 11 8 2 NOTE_INSN_FUNCTION_BEG)
$5 = void
According to BB_END, which points to (note 2), barrier is inside block.
