CPLGenerateTempFilename volatile and/or CPLAtomicInc

[Kurt Schwehr]

In r15574, you put a volatile modifier on nTempFileCounter. This looks to be protect from multithreaded race cases between callers of CPLGenerateTempFilename where more than one thread might get the same tempfile number. Would it be better to call CPLAtomicInc and use the result of that? If so, should I drop the volatile?

e.g.

    static volatile int nTempFileCounter = 0;
    CPLString osFilename;
    osFilename.Printf( "%s%u_%d", pszStem,
                       (unsigned int)(CPLGetPID() & 0xFFFFFFFFU), nTempFileCounter++ );

Would become this.

    static int nTempFileCounter = 0;
    CPLString osFilename;
    osFilename.Printf( "%s%u_%d", pszStem,
                       static_cast<unsigned int>( CPLGetPID() & 0xFFFFFFFFU),
                       CPLAtomicInc(nTempFileCounter) );

A second question on this code:

Why the mask on the CPLGetPID()?

[Even Rouault]

+1 for CPLAtomicInc (but use &nTempFileCounter as argument)

I think the mask was when using with some sanitzation runtime because CPLGetPID() could return a value larger than a unsigned int. But A better fix would be to replace %u with CPL_FRMT_GIB and dropping the cast

[Kurt Schwehr]

Bugger... forgot to include this ticket in the commit message of r31513 on trunk.

[Kurt Schwehr]

Submitted r31516 to make a TODO that all this can be done in C++11 with std::atomic.

[Kurt Schwehr]

For reference, see Herb Sutter's response about volatile:

​http://stackoverflow.com/questions/2479067/why-is-the-volatile-qualifier-used-through-out-stdatomic

[Even Rouault]

I assume this can be closed now ?

[Kurt Schwehr]

Closing. Resolved with r31513.