[clang] Control analysis-based diagnostics with #pragma (PR #136323)

[Artem Dergachev via cfe-commits]

================
@@ -202,6 +202,43 @@ class SemaPPCallbacks : public PPCallbacks {
       break;
     }
   }
+  void PragmaDiagnostic(SourceLocation Loc, StringRef Namespace,
+                        diag::Severity Mapping, StringRef Str) override {
+    // If one of the analysis-based diagnostics was enabled while processing
+    // a function, we want to note it in the analysis-based warnings so they
+    // can be run at the end of the function body even if the analysis warnings
+    // are disabled at that point.
+    SmallVector<diag::kind, 256> GroupDiags;
+    diag::Flavor Flavor =
+        Str[1] == 'W' ? diag::Flavor::WarningOrError : diag::Flavor::Remark;
+    StringRef Group = Str.substr(2);
+
+    if (S->PP.getDiagnostics().getDiagnosticIDs()->getDiagnosticsInGroup(
+            Flavor, Group, GroupDiags))
+      return;
+
+    for (diag::kind K : GroupDiags) {
+      // Note: the cases in this switch should be kept in sync with the
+      // diagnostics in AnalysisBasedWarnings::getPolicyInEffectAt().
+      AnalysisBasedWarnings::Policy &Override =
+          S->AnalysisWarnings.getPolicyOverrides();
+      switch (K) {
+      default: break;
----------------
haoNoQ wrote:

So every new analysis-based warning needs to be added to this list right? How weird are the consequences of forgetting it?

It might be a good idea to make a unified list of analysis-based warnings. (Or like, make it a flag in `Diagnostic...Kinds.td`.) Then it may be possible to build a runtime assertion to confirm that as long as `AnalysisBasedWarnings::IssueWarnings()` is on the stack, all warnings emitted by clang are on that list.

(Probably out of scope for this patch. Sounds like a lot of work. There could be easier ways to check this. Could also be too restrictive in practice.)

https://github.com/llvm/llvm-project/pull/136323