@Bjorn, Thanks for reviewing and accepting the patch.

Updated with test from Bjorn Pettersson which is much more accurate and clearer. Thanks!

Yeah, you're right. Only one loop has to be checked in this case. I'll update the test as per your suggestion. Thank you!

I have updated the test to not run the optimizer. The test I had added previously for checking if the unroller is respecting the pragma is useful I think. Not sure where that can be added though.
I guess it's independent of this patch anyway. If the patch and test is okay, will update bugzilla as well.

Update the tests.

I encountered the issue while working with the unroller and found that it was not following the pragma info, and traced it back to the issue with metadata.
As far as I understood, for for-loops and while-loops, we add the metadata only to the loop back-edge. So it would make sense to keep them consistent.
I'm not an expert in clang, and do not know how we can detect such problems.

Added to Bugzilla,
https://bugs.llvm.org/show_bug.cgi?id=38011

Do I need to add specific reviewers?

I had based it on the other tests in clang/test/CodeGen.
Do we not need the -o to output to standard output?
Or did you mean something else?

Add tests and the patch.

Add tests.

Phab is the correct way to submit patches.
But having a bugreport in bugzilla is good too.
But the test will be needed regardless of the patch submission method.
And yes, please do always upload all patches with full context (-U99999).

Add full context

It's a patch for a bug in clang.
I have requested for a Bugzilla account, however thought of putting up the
patch in the meantime.
Do I need to mark it '[Private]'?

Adding Greg for review.

The temp directory was chosen since it's portable. There's no syslog.h in Windows and also no directory like /var/log I think.

I think the default being the current directory "." would not be a good idea, if lldb-mi does not have the permission to write there. Unlike lldb, this is more important for lldb-mi, especially while debugging from a GUI, since the user wouldn't know what's happening and why lldb-mi failed.

Thanks for fixing this. I haven't had a chance to test this patch with our cross-platform issues yet, have to look at some other issues first.
Just a couple of doubts I had.

You're right. That should be replaced to looping through registers one by one and marking them valid.
I'll modify it.

Yeah, updating the PathMappingList does seem more appropriate. We specifically had issues with interoperability handling paths between different operating systems. If your patch is ready, please put it up and we can test it out.