Re: [patch] fix verify_rtl_sharing handling of SEQUENCEs

[Steven Bosscher <stevenb dot gcc at gmail dot com>]

On Tue, Apr 16, 2013 at 6:39 PM, Steven Bosscher wrote:
>> My new delay branch scheduler uses TODO_verify_rtl_sharing but it
>> turns out that verify_rtl_sharing doesn't handle SEQUENCEs correctly:
>> Clearing the used-flags is done correctly, but verifying insns in the
>> SEQUENCE fails. The problem is that every insn in the SEQUENCE is
>> marked used via PATTERN(SEQUENCE) and also via PREV_INSN/NEXT_INSN of
>> the insns in the SEQUENCE.
>>
>> Fixed with the attached patch. Bootstrapepd&tested on
>> sparc64-unknown-linux-gnu, and cross-built&tested on mips-elf, both
>> with TODO_verify_rtl_sharing added to the passes in reorg.c.
>> Will commit as obvious.
>
> Andreas Krebbel's patch
> (http://gcc.gnu.org/ml/gcc-patches/2013-04/msg00714.html) makes the
> problem go away for me, so I'm not going to commit this patch after
> all.

Hmm, no it hasn't.

What happens is this:

reset_all_used_flags resets the "used" flags via mark_used_flags,
which doesn't mark or unmark insns:

    case DEBUG_INSN:
    case INSN:
    case JUMP_INSN:
    case CALL_INSN:
    case NOTE:
    case LABEL_REF:
    case BARRIER:
      /* The chain of insns is not being copied.  */
      return;

But verify_rtx_sharing sets the "used" flag on insns if they are
reached via a SEQUENCE. So the first verify_rtl_sharing call with
SEQUENCEs in the insn chain passes, but a second call will fail
because the "used" flags on insns in the SEQUENCE isn't cleared.

So, updated patch attached, will commit after testing on sparc64-linux.

Ciao!
Steven

Attachment: fix_verify_rtl_sharing.diff.txt
Description: Text document