memory | STL compatible C++ memory allocator | Monitoring library

You can now update Android Studio normally.

Thank you all for helping to bring this problem to Google's attention.

NaN is a numerical or computational concept: something that is "not a number". Eg 0/0 is NAN, because the result of this computation is undefined (but note that 1/0 is Inf, or infinity, and similarly -1/0 is -Inf).

The way that R handles these concepts internally isn't something that you should ever be concerned about.

GCC 12 (currently trunk) warns about it:

As to why it's not an error, it's because C++ is not required to do bounds checking on arrays. You could use a vector and use the at function, which throws exceptions if you go outside the bounds, but arrays do not.

When you attempt to access memory you don’t have access to, such as accessing an array out of bounds or dereferencing an invalid pointer. The signal SIGSEGV is standardized but different OS might implement it differently. "Segmentation fault" is mainly a term used in *nix systems, Windows calls it "access violation".

Why is the process in undefined behavior state after that point?

Because one or several of the variables in the program didn’t behave as expected. Let’s say you have some array that is supposed to store a number of values, but you didn’t allocate enough room for all them. So only those you allocated room for get written correctly, and the rest written out of bounds of the array can hold any values. How exactly is the OS to know how critical those out of bounds values are for your application to function? It knows nothing of their purpose.

Furthermore, writing outside allowed memory can often corrupt other unrelated variables, which is obviously dangerous and can cause any random behavior. Such bugs are often hard to track down. Stack overflows for example are such segmentation faults prone to overwrite adjacent variables, unless the error was caught by protection mechanisms.

If we look at the behavior of "bare metal" microcontroller systems without any OS and no virtual memory features, just raw physical memory - they will just silently do exactly as told - for example, overwriting unrelated variables and keep on going. Which in turn could cause disastrous behavior in case the application is mission-critical.

Why is it not recoverable?

Because the OS doesn’t know what your program is supposed to be doing.

Though in the "bare metal" scenario above, the system might be smart enough to place itself in a safe mode and keep going. Critical applications such as automotive and med-tech aren’t allowed to just stop or reset, as that in itself might be dangerous. They will rather try to "limp home" with limited functionality.

Why does this solution avoid that unrecoverable state? Does it even?

That solution is just ignoring the error and keeps on going. It doesn’t fix the problem that caused it. It’s a very dirty patch and setjmp/longjmp in general are very dangerous functions that should be avoided for any purpose.

We have to realize that a segmentation fault is a symptom of a bug, not the cause.

This version seems pretty robust, with the actual call to SendAsync wrapped in an adapter class.

I accept this doesn't necessarily answer the question directly, but in the absence of being able to replicate your problem exactly, offers an alternative way of structuring the code that may eliminate the problem.

The TLDR is that loads which miss all levels of the TLB (and so require a page walk) and which are separated by address unknown stores can't execute in parallel, i.e., the loads are serialized and the memory level parallelism (MLP) factor is capped at 1. Effectively, the stores fence the loads, much as lfence would.

The slow version of your insert function results in this scenario, while the other two don't (the store address is known). For large region sizes the memory access pattern dominates, and the performance is almost directly related to the MLP: the fast versions can overlap load misses and get an MLP of about 3, resulting in a 3x speedup (and the narrower reproduction case we discuss below can show more than a 10x difference on Skylake).

The underlying reason seems to be that the Skylake processor tries to maintain page-table coherence, which is not required by the specification but can work around bugs in software.

For those who are interested, we'll dig into the details of what's going on.

I could reproduce the problem immediately on my Skylake i7-6700HQ machine, and by stripping out extraneous parts we can reduce the original hash insert benchmark to this simple loop, which exhibits the same issue: