Work around a race condition in lldbtest.py for Windows.

[amccarth]

Summary

Allows about 10 passing tests to succeed rather than fail due to log file juggling.

[sas]

Can you just explain why this race happens?
Otherwise, looks good.

[zturner]

Windows is well-known (I wouldn't say documented necessarily, but it seems
everyone has run into this problem before) where there is a very short
window after a process exits where some of the handles that were used by
the file (e.g. open files, the image file itself, etc) are still locked by
the operating system. It doesn't happen every time, or even on every
machine (I've never been able to reproduce Adrian's problem on my Windows
10 machine, for example)

Maybe it's due to the search indexer, maybe it's due to antivirus, maybe
it's due to a bug in the kernel. LLVM just had a similar problem just this
week and fixed it with a similar workaround (
http://llvm.org/viewvc/llvm-project?rev=246708&view=rev)

[amccarth]

I've put more detail in the comment.

[sas]

Wow, this is funny.

Thanks for the explanation @zturner and thanks for adding the comment @amccarth.

lldb/trunk/test/lldbtest.py

def dumpSessionInfo(self):
                         # guaranteed to be possible on Windows, but we need this to work anyway, so just remove the
                         # destination first if it already exists.
                         os.remove(dst)
 
                     os.rename(src, dst)
         else:
             # success!  (and we don't want log files) delete log files
             for log_file in log_files_for_this_test:
+                try:
+                    os.unlink(log_file)
+                except:
+                    # We've seen consistent unlink failures on Windows, perhaps because the
+                    # just-created log file is being scanned by anti-virus.  Empirically, this
+                    # sleep-and-retry approach allows tests to succeed much more reliably.
+                    # Attempts to figure out exactly what process was still holding a file handle
+                    # have failed because running instrumentation like Process Monitor seems to
+                    # slow things down enough that the problem becomes much less consistent.
+                    time.sleep(0.5)
                     os.unlink(log_file)
 
     # ====================================================
     # Config. methods supported through a plugin interface
     # (enables reading of the current test configuration)
     # ====================================================
 
     def getArchitecture(self):
         """Returns the architecture in effect the test suite is running with."""