[clang][cas] Use module cache key as context hash

This improves canonicalization, reducing the number of cache misses when
using compilation caching. For leaf modules this makes no difference
since the extra compilation would be cached anyway, but when a module is
imported the context hash is generally part of the path of the pcm,
which then effects every downstream importer. This was previously
causing spurious cache misses when changing options like
-fmessage-length that are otherwise canonicalized away.

rdar://105844077

Author: benlangmuir

clang/lib/Tooling/DependencyScanning/ModuleDepCollector.cpp

@@ -98,7 +98,8 @@ ModuleDepCollector::makeInvocationForModuleBuildWithoutOutputs(
   // differ between identical modules discovered from different translation
   // units.
   CI.getFrontendOpts().Inputs.clear();
-  CI.getFrontendOpts().OutputFile.clear();
+  // CAS: OutputFile cannot be empty when computing a cache key.
+  CI.getFrontendOpts().OutputFile = "-";
   // FIXME: a build system may want to provide a new path.
   CI.getFrontendOpts().IndexUnitOutputPath.clear();
 
@@ -118,7 +119,9 @@ ModuleDepCollector::makeInvocationForModuleBuildWithoutOutputs(
     CI.getDiagnosticOpts().DiagnosticSerializationFile = "-";
   if (!CI.getDependencyOutputOpts().OutputFile.empty())
     CI.getDependencyOutputOpts().OutputFile = "-";
-  CI.getDependencyOutputOpts().Targets.clear();
+  // CAS: -MT must be preserved for cache key.
+  if (!CI.getDependencyOutputOpts().Targets.empty())
+    CI.getDependencyOutputOpts().Targets = {"-"};
 
   CI.getFrontendOpts().ProgramAction = frontend::GenerateModule;
   CI.getFrontendOpts().ARCMTAction = FrontendOptions::ARCMT_None;
@@ -289,6 +292,22 @@ static std::string getModuleContextHash(const ModuleDeps &MD,
       HashBuilder;
   SmallString<32> Scratch;
 
+  auto FormatHash = [&](llvm::BLAKE3Result<16> Hash) {
+    std::array<uint64_t, 2> Words;
+    static_assert(sizeof(Hash) == sizeof(Words), "Hash must match Words");
+    std::memcpy(Words.data(), Hash.data(), sizeof(Hash));
+    return toString(llvm::APInt(sizeof(Words) * 8, Words), 36,
+                    /*Signed=*/false);
+  };
+
+  if (MD.ModuleCacheKey) {
+    // Cache keys have better canonicalization, so use them as the context hash.
+    // This reduces the number of modules needed when compatible configurations
+    // are used (e.g. change in -fmessage-length).
+    HashBuilder.add(*MD.ModuleCacheKey);
+    return FormatHash(HashBuilder.final());
+  }
+
   // Hash the compiler version and serialization version to ensure the module
   // will be readable.
   HashBuilder.add(getClangFullRepositoryVersion());
@@ -324,13 +343,30 @@ static std::string getModuleContextHash(const ModuleDeps &MD,
   }
 
   HashBuilder.add(EagerLoadModules);
+  return FormatHash(HashBuilder.final());
+}
 
-  llvm::BLAKE3Result<16> Hash = HashBuilder.final();
-  std::array<uint64_t, 2> Words;
-  static_assert(sizeof(Hash) == sizeof(Words), "Hash must match Words");
-  std::memcpy(Words.data(), Hash.data(), sizeof(Hash));
-  return toString(llvm::APInt(sizeof(Words) * 8, Words), 36, /*Signed=*/false);
+#ifndef NDEBUG
+static void checkCompileCacheKeyMatch(cas::ObjectStore &CAS,
+                                      StringRef OldKeyStr, cas::CASID NewKey) {
+  if (NewKey.toString() == OldKeyStr)
+    return;
+
+  // Mismatched keys, report error.
+  auto OldKey = CAS.parseID(OldKeyStr);
+  if (!OldKey)
+    llvm::report_fatal_error(OldKey.takeError());
+  SmallString<256> Err;
+  llvm::raw_svector_ostream OS(Err);
+  OS << "Compile cache key for module changed; previously:";
+  if (auto E = printCompileJobCacheKey(CAS, *OldKey, OS))
+    OS << std::move(E);
+  OS << "\nkey is now:";
+  if (auto E = printCompileJobCacheKey(CAS, NewKey, OS))
+    OS << std::move(E);
+  llvm::report_fatal_error(OS.str());
 }
+#endif
 
 void ModuleDepCollector::associateWithContextHash(const CompilerInvocation &CI,
                                                   ModuleDeps &Deps) {
@@ -552,17 +588,27 @@ ModuleDepCollectorPP::handleTopLevelModule(const Module *M) {
   if (auto E = MDC.Controller.finalizeModuleInvocation(CI, MD))
     Diags.Report(diag::err_cas_depscan_failed) << std::move(E);
 
+  // Compute the cache key, if needed. Requires dependencies.
+  if (MDC.ScanInstance.getFrontendOpts().CacheCompileJob) {
+    auto &CAS = MDC.ScanInstance.getOrCreateObjectStore();
+    if (auto Key = createCompileJobCacheKey(CAS, Diags, CI))
+      MD.ModuleCacheKey = Key->toString();
+  }
+
   MDC.associateWithContextHash(CI, MD);
 
   // Finish the compiler invocation. Requires dependencies and the context hash.
   MDC.addOutputPaths(CI, MD);
 
-  // Compute the cache key, if needed. Requires dependencies and outputs.
-  if (MDC.ScanInstance.getFrontendOpts().CacheCompileJob) {
+#ifndef NDEBUG
+  // Verify the key has not changed with final arguments.
+  if (MD.ModuleCacheKey) {
     auto &CAS = MDC.ScanInstance.getOrCreateObjectStore();
-    if (auto Key = createCompileJobCacheKey(CAS, Diags, CI))
-      MD.ModuleCacheKey = Key->toString();
+    auto Key = createCompileJobCacheKey(CAS, Diags, CI);
+    assert(Key);
+    checkCompileCacheKeyMatch(CAS, *MD.ModuleCacheKey, *Key);
   }
+#endif
 
   MD.BuildArguments = CI.getCC1CommandLine();
 

clang/test/ClangScanDeps/modules-cas-context-hash.c

@@ -63,14 +63,38 @@
 //--- cdb1.json.template
 [{
   "directory": "DIR",
-  "command": "clang -Xclang -fcas-plugin-path -Xclang /1 -Xclang -fcas-plugin-option -Xclang a=x -fsyntax-only DIR/tu.c -fmodules -fimplicit-module-maps -fmodules-cache-path=DIR/cache1",
+  "arguments": [
+    "clang",
+    "-fsyntax-only",
+    "DIR/tu.c",
+    "-fmodules",
+    "-fimplicit-module-maps",
+
+    "-Xclang", "-fcas-plugin-path", "-Xclang", "/1",
+    "-Xclang", "-fcas-plugin-option", "-Xclang", "a=x",
+    "-fmodules-cache-path=DIR/cache1",
+    "-fmessage-length=1",
+    "-fcolor-diagnostics",
+  ],
   "file": "DIR/tu.c"
 }]
 
 //--- cdb2.json.template
 [{
   "directory": "DIR",
-  "command": "clang -Xclang -fcas-plugin-path -Xclang /2 -Xclang -fcas-plugin-option -Xclang b=y -fsyntax-only DIR/tu.c -fmodules -fimplicit-module-maps -fmodules-cache-path=DIR/cache2",
+  "arguments": [
+    "clang",
+    "-fsyntax-only",
+    "DIR/tu.c",
+    "-fmodules",
+    "-fimplicit-module-maps",
+
+    "-Xclang", "-fcas-plugin-path", "-Xclang", "/2",
+    "-Xclang", "-fcas-plugin-option", "-Xclang", "b=y",
+    "-fmodules-cache-path=DIR/cache2",
+    "-fmessage-length=2",
+    "-fno-color-diagnostics",
+  ],
   "file": "DIR/tu.c"
 }]