[LLVMdev] Reviving the new LLVM concurrency model

[Eli Friedman]

On Mon, Aug 22, 2011 at 11:17 AM, Jianzhou Zhao <jianzhou at seas.upenn.edu> wrote:
> On Mon, Aug 22, 2011 at 1:02 PM, Eli Friedman <eli.friedman at gmail.com> wrote:
>> On Mon, Aug 22, 2011 at 9:55 AM, Jianzhou Zhao <jianzhou at seas.upenn.edu> wrote:
>>> In the definition of 'monotonic' ordering,
>>> ... "If an address is written monotonically by one thread, and other
>>> threads monotonically read that address repeatedly, the other threads
>>> must eventually see the write"...
>>
>> It's supposed to mean that if you have a something like looks like a
>> spinloop with monotonic reads, it shouldn't spin forever if the value
>> changes.  I'll take another look at rewording that.
>>
>>> Does this mean if a thread does multi-writes monotonically, monotonic
>>> reads from other threads should see all of them? But intuitively, it
>>> seems to be fine for a read to ``miss'' some of the writes as long as
>>> the writes seen are monotonic in the sense that later reads should see
>>> the same write of earlier reads, or any write monotonically after the
>>> writes seen.
>>>
>>> In the case there is only one monotonic write, what does 'eventually'
>>> mean? Can we know a write must be seen when some condition holds, for
>>> example, a number of instructions executed, the thread that did the
>>> write executes a fence, ...?
>>>
>>> C++ memory model does not have ``unordered'', and "monotonic", but
>>> have "modification ordering" (is it same to the relaxed atomic
>>> variables the LLVM IR mentions?). If I am compiling C++ to LLVM, can
>>> all modification atomic be compiled to monotonic? And when should we
>>> use "unordered"?
>>
>> http://llvm.org/docs/Atomics.html is an attempt to make things much
>> more straightforward than the stuff in LangRef.
>
> This is cool.
>
> At the end of the "optimization outside atomic" section there are
> discussions about "returning undef". Is it the following correct?
> * a store/store data race in LLVM leads to undefined behaviors,

What exactly is a store-store "race"?  That sounds wrong.

> * a store/load data race does not result in undefined behavior, but
> the load returns undef
> * if two memory accesses are of data races, then at least one of them
> is NonAtomic.

The model isn't really defined in terms of races, but these two sound
roughly correct.

> My question is suppose a load L and a store S have a data race, and L
> runs earlier than S in an execution, L is well-ordered with earlier
> writes by happens-before, then at the point when L runs, but S has not
> run yet, should the L also return undef or what ever write it can see
> w/o races so far?
>
> Although non-synchronized writes from other threads may propagate to
> another thread in different orders, but the writes that a read from a
> different thread can see should have already executed before the read
> (in a global time). So in the above case, it seems fine to allow the
> load to return a 'defined' value. Is there any case that makes 'undef'
> possible?

If there is a load and a store to the same address with no
happens-before relationship, the load returns undef.  "L runs earlier
than S in an execution" doesn't make sense; if the load and store
aren't atomic, the compiler is allowed to, for example, rematerialize
the load, so that it happens both before and after the store.


-Eli