Re: static_branch/jump_label vs branch merging

From: Peter Zijlstra <peterz@infradead.org>
To: Nick Desaulniers <ndesaulniers@google.com>
Cc: Florian Weimer <fweimer@redhat.com>,
	Ard Biesheuvel <ardb@kernel.org>,
	linux-toolchains@vger.kernel.org,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	Josh Poimboeuf <jpoimboe@redhat.com>,
	Jason Baron <jbaron@akamai.com>,
	"Steven Rostedt (VMware)" <rostedt@goodmis.org>,
	Segher Boessenkool <segher@kernel.crashing.org>,
	dmalcolm@redhat.com
Subject: Re: static_branch/jump_label vs branch merging
Date: Fri, 9 Apr 2021 22:11:01 +0200	[thread overview]
Message-ID: <YHC01St3wfXOlRUb@hirez.programming.kicks-ass.net> (raw)
In-Reply-To: <CAKwvOdnBXyR6gPgQjaXbHF8Ozx9Kk=OKgPv8_P7=jvvsWRVHEg@mail.gmail.com>

On Fri, Apr 09, 2021 at 12:33:29PM -0700, Nick Desaulniers wrote:
> On Fri, Apr 9, 2021 at 4:18 AM Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > On Fri, Apr 09, 2021 at 12:55:18PM +0200, Florian Weimer wrote:
> > > * Ard Biesheuvel:
> > >
> > > > Wouldn't that require the compiler to interpret the contents of the
> > > > asm() block?
> > >
> > > Yes and no.  It would require proper toolchain support, so in this case
> > > a new ELF relocation type, with compiler, assembler, and linker support
> > > to generate those relocations and process them.  As far as I understand
> > > it, the kernel doesn't do things this way.
> >
> > I don't think that all is needed. All we need is for the compiler to
> > recognise that:
> >
> >         if (cond) {
> >                 stmt-A;
> >         }
> >         if (cond) {
> >                 stmt-B;
> >         }
> >
> > both cond are equivalent and hence can merge the blocks like:
> >
> >         if (cond) {
> >                 stmt-A;
> >                 stmt-B;
> >         }
> >
> > But because @cond is some super opaque asm crap, the compiler throws up
> > it's imaginry hands and says it cannot possibly tell and leaves them as
> > is.
> 
> Right, because if `cond` has side effects (such as is implied by asm
> statements that are volatile qualified), suddenly those side effects
> would only occur once whereas previously they occurred twice.
> 
> Since asm goto is implicitly volatile qualified, it sounds like
> removing the implicit volatile qualifier from asm goto might help?
> Then if there were side effects but you forgot to inform the compiler
> that there were via an explicit volatile qualifier, and it performed
> the suggested merge, oh well.

So, as mentioned elsewhere in this thread, it would be nice if either
the pure or const function attribute could over-ride/constrain that
volatile side effect.

I'm fine with things going side-ways if we get it wrong, that's more or
less the game we're playing anyway ;-)