[clangd] Use in-memory preambles in clangd.

[ilya-biryukov]

After looking at the numbers, I think we should make this configurable. I'll update this change accordingly. Not sure what should be the default, though. I'm currently erring on the side of "in-memory by default".

Current implementation eats almost twice as much memory, which quickly adds up (tried comparing with earlier builds on ClangdUnit.cpp).
Even without proper benchmarks, it's clear that if we run out of memory (and we easily could if run on a machine with, say, 8GB of RAM) on-disk preambles are an easy way out of trouble.

[sammccall]

Heh, I was just going to ask :-)

As a very first step, can we make this configurable but off-by-default? That will address the pressing diskless-servers need.

In a typical workstation scenario, disk is plentiful (if slow) and RAM is scarce - very likely we're running browser + editor (VSCode is another browser really) + clangd + build system, and clangd is likely the least important of these...
So if we're going to increase mem usage by a significant amount I think we want to measure it carefully first.

[ilya-biryukov]

Made in-memory preambles optional (on-disk by default).

[ilya-biryukov]

In D39843#920568, @sammccall wrote:

As a very first step, can we make this configurable but off-by-default? That will address the pressing diskless-servers need.

I think on-disk by default makes sense now until we do the measurements and have a the data to prove the on-memory is really better.

[sammccall]

As discussed, we should probably bundle together some related options into structs to make plumbing easier. Not a blocker here though.

If these options are pure-data, I think it's OK if some layers end up with access to other layers' options.

clangd/ClangdUnit.cpp

238

ref is still here :-)

clangd/tool/ClangdMain.cpp

48 ↗

(On Diff #122845)

Couple of questions about the flag syntax:

• I wonder about the granularity - if we had another temp storage (indexes?), would we want another flag or something shared like -in-memory-storage
• Having this as a bool seems confusing compared to -pch-storage={memory/disk}. (We could also overload such a flag to allow setting the location, but nobody's asked for that yet)

[ilya-biryukov]

• Removed /*ref*/.
• Changed the command-line flag: it's -pch-storage now instead of -in-memory-pchs.
• Actually pass the StoreInMemory flag to PrecompiledPreamble::Build.
• Fixed a new usage of ClangdServer after rebase (In ClangdTests).

clangd/ClangdUnit.cpp

238

Thx. Missed that. Removed.

clangd/tool/ClangdMain.cpp

48 ↗

(On Diff #122845)

• Indexes are fundamentally different because we actually want to persist them across multiple sessions. But I see your point. I'd go with another flag for greater flexibility. We do want clangd work great by-default, so those flags are mostly for people who know what they're doing and can figure out which flags they want to change and why. Maybe I'm not right, though, and people would want those kinds of switches.
• Totally like the -pch-storage={memory/disk} more, updated the change accordingly.

clangd/ClangdUnit.cpp

prepareCompilerInstance(std::unique_ptr<clang::CompilerInvocation> CI,
   assert(VFS && "VFS is null");
   assert(!CI->getPreprocessorOpts().RetainRemappedFileBuffers &&
          "Setting RetainRemappedFileBuffers to true will cause a memory leak "
          "of ContentsBuffer");
 
   // NOTE: we use Buffer.get() when adding remapped files, so we have to make
   // sure it will be released if no error is emitted.
   if (Preamble) {
-    Preamble->AddImplicitPreamble(*CI, Buffer.get());
+    Preamble->AddImplicitPreamble(*CI, /*ref*/ VFS, Buffer.get());
   } else {
     CI->getPreprocessorOpts().addRemappedFile(
         CI->getFrontendOpts().Inputs[0].getFile(), Buffer.get());
   }
 
   auto Clang = llvm::make_unique<CompilerInstance>(PCHs);
   Clang->setInvocation(std::move(CI));
   Clang->createDiagnostics(&DiagsClient, false);
auto DoRebuildPreamble = [&]() -> std::shared_ptr<const PreambleData> {
       std::vector<DiagWithFixIts> PreambleDiags;
       StoreDiagsConsumer PreambleDiagnosticsConsumer(/*ref*/ PreambleDiags);
       IntrusiveRefCntPtr<DiagnosticsEngine> PreambleDiagsEngine =
           CompilerInstance::createDiagnostics(
               &CI->getDiagnosticOpts(), &PreambleDiagnosticsConsumer, false);
       CppFilePreambleCallbacks SerializedDeclsCollector;
       auto BuiltPreamble = PrecompiledPreamble::Build(
           *CI, ContentsBuffer.get(), Bounds, *PreambleDiagsEngine, VFS, PCHs,
-          SerializedDeclsCollector);
+          /*StoreInMemory=*/true, SerializedDeclsCollector);
 
       if (BuiltPreamble) {
         return std::make_shared<PreambleData>(
             std::move(*BuiltPreamble),
             SerializedDeclsCollector.takeTopLevelDeclIDs(),
             std::move(PreambleDiags));
       } else {
         return nullptr;