Only clear setting when the force flag is set (as suggested offline by @jingham).

LGTM.

Can you provide some more information about why this is needed?

The patch looks fine but since I don't know much about opencl I'll leave the LGTM to someone that actually knows this code.

The code and test look correct, so this LGTM but I'll leave it open for now in case someone else wants to have a look too.

Can you add a test case please?

LGTM, Thx!

In D52783#1253806, @whisperity wrote:

Are you sure the diff is okay? I see only a move on VirtualFileSystem.cpp but I really think there were namespace clang { lines in it too.

• Fix a few formatting issues.

LGTM

Thanks for working on this Stefan, I'm really excited about this feature!

In D52772#1252656, @jingham wrote:

Would it be possible for the exporter to notice empty settings and write "settings clear" instead?

In D52783#1253573, @grimar wrote:

Seems there are a lot of changes caused by clang-formatting. I suggest to not mix the cosmetic and functionality changes and either revert them (remove from this patch) or commit as NFC separately at first.

Split off settings set change into separate review: D52772

Thanks a lot for the feedback! I clearly overlooked some stuff when inspiring myself on the other CommandObject code.

• Rebase

• Make the different tests standalone.

• Rename file

In D52603#1247775, @RKSimon wrote:

Use .ll file extension not .test

• Add assertions per Greg's suggestion.

My first observations:

Thanks!

In D52543#1246877, @clayborg wrote:

It is the SymbolVendor's job to do the locking. And in many cases it already is. I stopped with inlined comments after a few comments as it would apply to this entire patch.

Thanks for the information, Greg!

In D48393#1139327, @labath wrote:

It's even more complicated than that, in case you really have reference cycles, you can have multiple threads starting parsing from different points in that cycle, and getting deadlocked waiting for the DIE_IS_BEING_PARSED results from each other.

The only sane algorithm I can come up right now is to make the list of parsed dies local to each thread/parsing entity (e.g. via a "visited" list), and only update the global map once the parsing has completed (successfully or not). This can potentially duplicate some effort where one thread parses a type only to find out that it has already been parsed, but hopefully that is not going to be the common case. The alternative is some complicated resource cycle detection scheme.

Thanks a lot for the review feedback everyone!

LGTM, Thanks!

• Add check for equal tags to cover case where both are subprograms, both are formal params.
• Ran verifier over clang, no errors.
• Ran verifier over GCC generated binary I had laying around, also no errors.

I'm having some trouble with the test case. Based on the initialization code I assume I'm not supposed to destroy the SBDebugger singleton shared by the LLDB test suite. If I do it anyway the test crashes with the an exception:

ping

• Use the new dump method that passes the correct DI dump options.
• Update test to expect verbose output.

Make sure addUnit keeps the DWARFUnitVector sorted.

Thanks Vedant, this looks very useful!

Generally this looks good, but I'd like for the other to have a look first (at this and the other patch) before accepting.

Hey Arnaud, let me know if you want me to commit this for you.

Rebase

Use separate vectors for compile and type units.