Re: [RFC v2 0/6] x86: dynamic indirect branch promotion

From: Josh Poimboeuf
Date: Thu Jan 03 2019 - 15:31:15 EST


On Thu, Jan 03, 2019 at 06:30:08PM +0000, Nadav Amit wrote:
> > On Jan 3, 2019, at 10:10 AM, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> >
> > On Mon, Dec 31, 2018 at 07:53:06PM +0000, Nadav Amit wrote:
> >>> On Dec 31, 2018, at 11:51 AM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> >>>
> >>> On Sun, Dec 30, 2018 at 11:20 PM Nadav Amit <namit@xxxxxxxxxx> wrote:
> >>>> This is a revised version of optpolines (formerly named retpolines) for
> >>>> dynamic indirect branch promotion in order to reduce retpoline overheads
> >>>> [1].
> >>>
> >>> Some of your changelogs still call them "relpolines".
> >>>
> >>> I have a crazy suggestion: maybe don't give them a cute name at all?
> >>> Where it's actually necessary to name them (like in a config option),
> >>> use a description like CONFIG_DYNAMIC_DEVIRTUALIZATION or
> >>> CONFIG_PATCH_INDIRECT_CALLS or something.
> >
> > Cute or not, naming is important.
> >
> > If you want a description instead of a name, it will be a challenge to
> > describe it in 2-3 words.
> >
> > I have no idea what "dynamic devirtualization" means.
> >
> > "Patch indirect calls" doesn't fully describe it either (and could be
> > easily confused with static calls and some other approaches).
> >
> >> Iâm totally fine with that (donât turn me into a "marketingâ guy). It was
> >> just a way to refer to the mechanism. I need more feedback about the more
> >> fundamental issues to go on.
> >
> > Naming isn't marketing. It's a real issue: it affects both usability
> > and code readability.
>
> Well, allow me to be on the fence not this one.
>
> I look for the path of least resistance. I think it would be easiest if I
> first manage to make Joshâs static calls to be less intrusive. For that, I
> try to add in the GCC plugin an attribute to annotate the function pointers
> whose calls should be promoted. However, I donât manage to get the
> declaration from the call instruction's rtx. If anyone has a pointer on how
> it can be done, thatâs would be very helpful.

Ok. FYI, I'll be posting v3 in the next few days or so.

--
Josh