If the workspace folder is a symlink and the client doesn't follow it.
Treat /tmp/symlink/ as canonical and convert every /tmp/real/ path to
/tmp/symlink/.
Combined Daan De Meyer's #227 with other simplification
* USE_SHARED_LLVM is deleted in favor of LLVM_LINK_LLVM_DYLIB
* LLVM_ENABLE_RTTI is deleted as it is provided by LLVMConfig.cmake
* Only direct Clang/LLVM dependencies are required in target_link_libraries
* Restrict -DCLANG_RESOURCE_DIRECTORY= to src/utils.cc
cache.hierarchicalPath: store cache files as $directory/a/b/c.cc.blob to
work around NAME_MAX limitation.
cache.retainInMemory: after this number of loads, keep a copy of file
index in memory. If set to 1, it avoids cache corruption if the index
file is changed after the initial load, which may happen if several
language clients open the same project and share the same cache
directory.
Also rename cacheDirectory cacheFormat to cache.{directory,format}
Also makes textDocument/definition in comments takes scope resolution
into account, and use detailed names (for Func, without parameters) for workspace/symbol
* Look for .cquery in any directory above the source file in the hierarchy.
Currently cquery only reads compiler arguments (.cquery) from project
root. Under some circumstances (e.g. remote compiling), generating a
compilation database with correct path in it is non-trivial, and
allowing per directory compile arguments usually helps.
* unused var buf
The heuristic may not be perfect, but it is probably good enough.
clang_codeCompleteAt seems to get slower as more and more CXUnsavedFile
instances are uploaded.
This function has a huge stack frame (> 8k, thanks to tinydir), and is
called recursively. I have seen this function causing some stack
overflows. So convert it to BFS to avoid that.
