Re: [patch] PR56572 flatten unnecessary nested transactions after inlining

[Aldy Hernandez <aldyh at redhat dot com>]

On 01/06/14 13:40, Richard Henderson wrote:
> On 12/19/2013 11:06 AM, Richard Biener wrote:
> > Aldy Hernandez <aldyh@redhat.com> wrote:
> > > I'd still like to catch the common cases, like I do with this patch.
> > >
> > > Perhaps we move this code to the .tmmark pass and handle the
> > > uninstrumented case.  rth?
>
> tmmark is way way later than you'd want.  I believe that ipa_tm is the right
> place.  That's where we generate clones.  The clones know a-priori that they're
> called within a transaction and thus all internal transations may be
> eliminated.  And thus any inlining that would happen after ipa_tm would
> properly inline the clone, and thus no more need be done.

But ipa_tm still runs before the latter inliner:

...
u.c.041i.tmipa
u.c.043i.whole-program
u.c.044i.profile_estimate
u.c.045i.devirt
u.c.046i.cp
u.c.048i.inline

So even though my proposed patch works in the supplied testcase (because 
the early inliner ran before .tmipa and inlined the nested transaction), 
if the latter inliner added a nested transaction we're in the same boat. 
 I just saw this happen with LTO, as Richi pointed out.

What do you suggest?