[flang] [llvm] [flang-rt] Runtime implementation of extended intrinsic function SECNDS() (PR #152021)

[Peter Klausler via llvm-commits]

klausler wrote:

> > I think that you can replace that loop with just a single compare & swap if zero. Only the first thread to ever execute it would succeed, and all later ones would fail and return the first thread's value.
> 
> The other threads still need to wait for the initialization to happen, so I think the loop is unavoidable. (In early implementation, I didn't have a loop and threads came in quicker than the time value could be computed.) Also, I decided to put even `compare_exchange_strong()` inside the loop, because I read somewhere that even this function can still fail under some conditions (for example, see section in https://dev.to/pauljlucas/advanced-thread-safety-in-c-3ap5).

The atomic compare & swap if zero would be the initialization, I mean.  The flow would be like:
```
  auto UTCmidnight{atomicRead(atomicVar}};
  if (UTCmidnight == 0) { // not initialized
    // not initialized yet
    UTCmidnight = computation;
    CHECK(UTCmidnight > 0);
    auto value{atomicCompareAndSwap(atomicVar, 0, UTCmidnight)};  // only stores a value once ever
    if (oldValue != 0) {
      // lost the race, my value wasn't stored
      UTCmidnight = oldValue;
    }
  }
  // all threads have the same UTCmidnight value from this point


https://github.com/llvm/llvm-project/pull/152021