Re: ifcvt/crossjump patch: Fix PR 42496, 21803

[Eric Botcazou <ebotcazou at adacore dot com>]

> The move_across_if changes (motivated by failures) use
> control_flow_insn_p which I believe takes care of the issue for that
> part.  Do you feel anything else is necessary in the context of this patch?

No, probably not.

> > The ATTRIBUTE_UNUSED on 'mode' in flow_find_cross_jump is already bogus,
> > so the new one in flow_find_head_matching_sequence is as well.
>
> Eliminated the new one.

Please remove the old one as well.

> > There is no comment in flow_find_head_matching_sequence about the purpose
> > of the EH edges checking code.
>
> It's basically a sanity check; I'm not even sure it can trigger.  I
> looked at outgoing_edges_match to find things I may need to test for;
> this was one the pieces I added.  I've added a small comment now; let me
> know if you feel something else is necessary.

OK, but add one more sentence explaining what the tests are testing.

> Retested this version with arm-linux-gnueabi
> (qemu-system-armv7{arch=armv7-a/thumb,thumb,}).  No new failures (one
> set of tests now times out on a different multilib, compared to a clean
> baseline).  Ok?

OK modulo the nits above and

+/* When comparing two insns in flow_find_cross_jump or
+   flow_find_head_matching_sequence, esure the notes match.  */
+static void
+merge_notes (rtx i1, rtx i2)

"When comparing insns I1 and I2 in..., ensure the notes match" + missing blank 
line after the comment.

-- 
Eric Botcazou