Java Optimizations and the JMM(playingwithpointers.com) |
Java Optimizations and the JMM(playingwithpointers.com) |
This is a wrong transformation:
void writer(Tuple tuple) {
tuple.nonVolatileF = 5;
tuple.volatileF = 1;
}
void reader(Tuple tuple) {
int regNV = 0;
int cache = tuple.nonVolatileF;
int regV = tuple.volatileF;
if (regV == 1) { regNV = cache; }
// computation that uses both of the above
}
> But you said that the JMM guarantees that all reads in C_i - C_(i-1) > will see writes in C_(i-1); this means that C_3 which reads > `tuple.nonVolatileF` will see the write to that variable in C_2, no?
Talking about such things in English is ambiguous. The formal statement is
"For any read r ∈ C_i −C_(i−1), we have W_i(r) ∈ C_(i−1) and W(r) ∈ C_(i−1)".
A clearer way to state the clause in English is "writes seen by reads in C_i - C_(i-1) belong to C_(i-1);". I can justify an execution as long as reads (ultimately) seen any write in the commit previous to the one it is in, subject to happens before consistency. To prevent causality loops, the JMM allows a read to see a write that happened before it in the commit that introduces it. Then, to allow certain interesting optimizations, the JMM allows us the bait-and-switch the write a read saw -- "Each read r in C_i − C_(i−1) must see writes in C_(i−1) in both E_i and E, but may see a different write in E_i from the one it sees in E.".
The non-volatile read is _allowed_ to see the non-volatile write (i.e. such an execution exists, this is the question I ask in the exercise), but it doesn't have to. The transform in question is illegal because the transformed program allows behavior that the untransformed program didn't allow -- the transform breaks semantics.
> Since you're constructing an execution, why didn't you interleave > the execution to make this trivially true?
I don't understand this, make /what/ trivially true? In any case, the transformed program has a data race, and so observationally it doesn't need to have a sequentially consistent execution. Specifically, it is allowed to show behavior that cannot be described by any interleaving of the instructions streams of the individual threads.
> Also, can't you wrap the writer in an atomic block during > transformation?
What purpose would that serve?
And this is why debugging this kind of issue in practice is fun: You have very little or non-obvious guarantees here.
If you think about constructing adversary proofs or decided that your program is a little gnome bound to make you insane, this can be rephrased as: In this example, the nonvolatile read can choose any value of any write as long as it is in a happened before relationship to the nv-read. It might pick every value, every second value, or just that one troublesome valu and never change again.
one note: >This blog post tries to show how some re-orderings (that may be done in as part of optimizing a Java program for performance) that intuitively seem illegal are, in fact, illegal, by deriving that illegality from directly the JMM spec.
Should the first 'illegal' read legal?
Well it depends. :) I did mean "illegal", though there certainly are many illegal transforms that look legal and vice versa.