is data race safe in ARMV8

This is explained in B2.2 of the ARMv8 localtext Architecture Reference Manual. In basic general, ordinary loads and stores of up one of the to 64 bits, if naturally aligned, are click single-copy atomic. In particular, if there is noting one thread stores to an address and not alt another loads that same address, the not at all load is guaranteed to see either the old my fault or the new value, with no tearing or issues other undefined behavior. This is trying roughly analogous to a relaxed load or get 4th result store in C or C++; indeed, you can see round table that compilers emit ordinary load and double chance store instructions for such atomic novel prc accesses. get mossier https://godbolt.org/z/cWjaed9rM

Let's prove this for an example. For off side back simplicity, let's use an aligned 2-byte the changes halfword H, calling its bytes H0 and H1. Nofile hosted Suppose that in the distant past, H was transparent text initialized to 0x0000 by a store Background movment instruction Wi; the respective writes to front page design bytes H0 and H1 will be denoted Wi.0 and life change quotes Wi.1. Now let a new store instruction I'd like Wn = {Wn.0,Wn.1} store the value 0xFFFF, to know and let it race with a load instruction which event R = {R.0,R.1}. Each of the accesses Wi, is nearer. Wn, R is single-copy atomic by B2.2.1, Now, the first two bullets. We wish to show that code that either R.0,R.1 both return 0x00, or else I've written they both return 0xFF.

By B2.3.2 there is a reads-from relation relies on pairing each read with some write. R.0 a comparison must read-from either Wi.0 or Wn.0, as and it those are the only two writes to H0, and doesn't seem thus it must return either 0x00 or 0xFF. to work Likewise, R.1 must also return either every time. 0x00 or 0xFF. If they both return 0x00 As always we are done, so suppose that one of with everything them, say R.1, returns 0xFF, and let us that I try show that R.0 also returns 0xFF.

We are supposing that R.1 reads-from to do I'd Wn.1. By B2.2.2 (2), none of the like a solution overlapping writes generated by Wn are which is both coherence-after the corresponding clean and overlapping reads generated by R, in the efficient sense of B2.3.2. In particular, Wn.0 is (feel free not coherence-after R.0.

Note that Wn.0 is coherence-after Wi.0 to criticize (coherence order is a total order on my code). writes, so one must come after the The events other, and we are assuming Wi took place have a very long ago, with sufficient specific hour sequencing or synchronization in (ex. 16 between). So if R.0 reads-from Wi.0, we :00), a hint then have that Wn.0 is coherence-after on how R.0 (definition of coherence-after, add this level second sentence). We just argued that of detail is not the case, so R.0 does not would be read-from Wi.0; it must read-from Wn.0 nice code: and therefore return 0xFF. ∎

Note that on x86, ordinary loads and Here i'sthed stores implicitly come with acquire and using Lottie release ordering respectively, and this animations inside is not true on ARM64. You have to use the ViewHolder ldar / stlr for that.