Re: [PATCH 11/18] seqcount: Introduce raw_write_seqcount_barrier()

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Jun 17, 2015 at 11:02:14AM -0700, Paul E. McKenney wrote:
> On Wed, Jun 17, 2015 at 07:11:40PM +0200, Peter Zijlstra wrote:
> > On Wed, Jun 17, 2015 at 09:37:31AM -0700, Paul E. McKenney wrote:
> > > The point of std::atomic<> (and of the equivalent C11 syntax) is to
> > > force the compiler to suppress optimizations that are unsafe for shared
> > > variables.  We get more or less the same effect with volatile, protests
> > > from compiler people notwithstanding.
> > > 
> > > I often tell the compiler guys that they have to expect make -some-
> > > concessions for being 30 years late to the concurrency party, but
> > > it nevertheless makes sense to future-proof our code where it is
> > > reasonable to do so.
> > 
> > Right, so in that regards I would request the compiler option (and or
> > #pragma) that disables all the out-of-thin-air nonsense.
> 
> OK.  What is the form of the #pragma?  If it focuses on a specific
> access, we are likely to get a lot of pushback.

I didn't have anything specific in mind; other than

#pragma no_speculative_stores_ever

Which would forbid all these retarded 'optimizations' for the entire
translation unit.

> > Because while they hide behind their undefined behaviour, the fact is
> > that all of their machines for the past 30 odd years have been relying
> > on this 'undefined' behaviour to work. This being the machines they've
> > been typing their useless specs on :-)
> 
> Maybe I can scare them into doing all their work on UP systems.  ;-)
> 
> Interestingly enough, LLVM is taking a slightly different approach.
> Rather than invoke undefined behavior, they say that data races result
> in random bits being loaded.  Not that it makes much difference to the
> health and well-being of the software, mind you...

I'm not sure I follow that argument.

> > I doubt there's a single OS kernel (that supports SMP configurations)
> > that does not rely on a whole host of 'undefined' behaviour.
> 
> An alternative approach would be a compiler switch (or similar) that
> changed the default atomic access from SC to relaxed.  Then shared
> variables could be marked atomic, and normal C code could be used to
> access them, but without the compiler emitting memory barriers all over
> the place (yes, even on x86).

See, I don;'t think that is a realistic approach. Who is going to audit
our ~16 million lines of code to mark all shared variables? Or all the
other existing code bases that rely on this behaviour?