Good point! Thanks.

Abandoned in favor of: https://reviews.llvm.org/D67475

I realized there was an issue with adding the -o /dev/null to arguemnts in case when there was a -MD flag as Driver later adds -MT with the last -o value in case there's no -MT or -MQ. That means that for headers the target would always be /dev/null.
@dexonsmith pointed out that we could just conditionally add -MT in the adjustment. Also, we can use build directory instead of CDB index.

In D67197#1664565, @tschuett wrote:

Thanks for doing that! But it looks odd to attach the dependencies to the output file.

clang -MD -c foo.c -o foo.o

will attach the headers/dependencies to foo.c. foo.c includes the headers. It does not attach the headers to foo.o
It will work for make-like build systems, but it looks odd.

• prefix dependency targets with CDB index

• rebased
• updated tests

In D67197#1658553, @arphaman wrote:

Instead of basing the new -MT name off the output, can you base it on the name of the input filename instead to simplify this logic?

• renamed functions to suggest their limitations
• added one more test

Rebased on TOT and removed noise from the patch as I've already landed couple refactoring patches.

Thanks for polishing this!

Add test.

In D66859#1650481, @gribozavr wrote:

In D66859#1649868, @jkorous wrote:

Hmm, actually if we replace Buffer in the interface with a llvm::function_ref<void(llvm::raw_ostream &)> callback it would cover couple other existing uses of createUniqueFile.

That would make a lot of sense if you're going to do the cleanup and refactor those callers to use new API. Good idea!

However, I think just from the API simplicity perspective, it would make sense to have two overloaded APIs if calling this function with a precomputed buffer is common.

There are actually couple other fatal-error-handling strategies. But all the other I've seen raise a signal (mostly SIGABRT; SIGILL in case of LLVM_BUILTIN_TRAP). For the purpose of having a way how to ask libclang to crash in a way that might lead to core dump this seems fine. Uninstalling the handler doesn't lead to libclang just calling exit in case of any crash though. Maybe I should rename the functions? Open to suggestions...

It turned out that recovery context matters as it messes with signals.

Added a proper test.

How to minimize a patch to nothing:

Hmm, actually if we replace Buffer in the interface with a llvm::function_ref<void(llvm::raw_ostream &)> callback it would cover couple other existing uses of createUniqueFile.

Addressed comments.

Good points as always. Thanks!

Otherwise LGTM. Thanks for refactoring this!

• Split off change to IndexDataConsumer.
• Rebase.

I'll start splitting off some self-standing parts that won't be affected by hypothetical changes at the FrontendAction/ASTConsumer level.