[clang] [llvm] [clang] Support multiple in-process cc1 jobs in a ToolSession (PR #222531)

[Alexandre Ganea via cfe-commits]

================
@@ -283,15 +278,23 @@ int cc1_main(ArrayRef<const char *> Argv, const char *Argv0, void *MainAddr) {
   // Create the actual diagnostics engine.
   Clang->createDiagnostics();
 
+  DiagsBuffer->FlushDiagnostics(Clang->getDiagnostics());
+  if (!Success)
+    return 1;
+
+  // Initialize process-global state only after the early-return paths above.
+  // Everything below reaches the corresponding cleanup before returning.
+  if (!Clang->getFrontendOpts().TimeTracePath.empty()) {
+    llvm::timeTraceProfilerInitialize(
+        Clang->getFrontendOpts().TimeTraceGranularity, Argv0,
+        Clang->getFrontendOpts().TimeTraceVerbose);
+  }
+
   // Set an error handler, so that any LLVM backend diagnostics go through our
   // error handler.
   llvm::install_fatal_error_handler(LLVMErrorHandler,
----------------
aganea wrote:

Please use a `CrashRecoveryContextCleanupRegistrar` along with the installation of the fatal error handler, otherwise `llvm::remove_fatal_error_handler` below would never be called.

Additionally, I wonder if we shouln't stop further commands and consider the process as "contamined" after the first crash, in the same way LLD does with `canRunAgain`? Regardless of that choice, I would still do the fatal error cleanup anyway.

https://github.com/llvm/llvm-project/pull/222531