[cfe-dev] Windows Lit problems again. What should we do?

[Francois Pichet]

> I might find some time to investigate this today. Is it deterministic?
> Is any particular test most likely to trigger it?

The tests that trigger the failure are:
\tools\clang\test\CodeGenObjC\bitfield-access.m
\tools\clang\test\CodeGenObjC\image-info.m
\tools\clang\test\CodeGenObjC\metadata_symbols.m
\tools\clang\test\CodeGenObjC\ns-constant-strings.m

Some observation:
- It is not deterministic. Sometimes all the tests pass, most of the
time at least 1 of them will fail.
- The tests will always pass if run alone.
- The test will always pass if -j=1
- Doing a sleep long enough before renamePathOnDisk will fix the problem.
- changing "< %t %s" with "-input-file=%t %s" will fix the problems.

My theory is that lit, python or windows is holding a handle to '%t'
during the < redirect operation for too long which cause the second
run of clang to fail.