[PATCH] D23130: [Clang-tidy] Add a check for definitions in the global namespace.

[Aaron Ballman via cfe-commits]

aaron.ballman added a reviewer: aaron.ballman.
aaron.ballman added inline comments.


================
Comment at: clang-tidy/google/GlobalNamesCheck.cpp:62
+            Result.SourceManager->getExpansionLoc(D->getLocStart()))) {
+      // unless that file is a header.
+      if (!utils::isSpellingLocInHeaderFile(
----------------
s/unless/Unless


================
Comment at: clang-tidy/google/GlobalNamesCheck.cpp:84
+  // Ignore decls with internal linkage.
+  if (!D->isExternallyVisible())
+    return;
----------------
I wonder if this check would have better performance by making this a local AST matcher instead? UsingDecl and UsingDirectiveDecl do not have linkage, so it may require adjusting the matcher somewhat.


================
Comment at: clang-tidy/google/GlobalNamesCheck.cpp:96
+    // main() should be in the global namespace.
+    if (FDecl->isMain())
+      return;
----------------
malcolm.parsons wrote:
> Should `isMSVCRTEntryPoint()` be checked too?
Yes, it should.


================
Comment at: clang-tidy/google/GoogleTidyModule.cpp:68
+    CheckFactories.registerCheck<readability::GlobalNamesCheck>(
+        "google-global-names");
     CheckFactories.registerCheck<clang::tidy::readability::FunctionSizeCheck>(
----------------
Given that this was shipped under the old name, I think we need to figure out our policy for how to handle this. It also comes up in D26511, so I would like us to be consistent with what we do.


================
Comment at: test/clang-tidy/google-global-names.cpp:13-14
+// CHECK-MESSAGES: :[[@LINE-1]]:5: warning: 'i' declared in the global namespace
+extern int ii = 0;
+// CHECK-MESSAGES: :[[@LINE-1]]:12: warning: 'ii' declared in the global namespace
+
----------------
bkramer wrote:
> aaron.ballman wrote:
> > bkramer wrote:
> > > aaron.ballman wrote:
> > > > aaron.ballman wrote:
> > > > > bkramer wrote:
> > > > > > aaron.ballman wrote:
> > > > > > > This strikes me as being intentional enough to warrant not diagnosing because of the `extern` keyword.
> > > > > > The only case I see where this pattern is valuable is interfacing with C code. Not sure yet if we want to allow that or enforce extern "C" instead. Ideas?
> > > > > > 
> > > > > > an extern global in the global namespace still feels like something we should warn on :|
> > > > > Yet externs in the global namespace do happen for valid reasons (such as not breaking ABIs by putting the extern definition into a namespace or changing the language linkage) -- I'm trying to think of ways we can allow the user to silence this diagnostic in those cases. I feel like in cases where the user writes "extern", they're explicitly specifying their intent and that doesn't seem like a case to warn them about, in some regards. It would give us two ways to silence the diagnostic (well, three, but two are morally close enough):
> > > > > 
> > > > > 1) Put it into a namespace
> > > > > 2) Slap `extern` on it if it is global for C++ compatibility (such as ABIs)
> > > > > 3) Slap `extern "C"` on it if it global for C compatibility
> > > > > 
> > > > > I suppose we could require `extern "C++"` instead of `extern`, but I don't think that's a particularly common use of the language linkage specifier?
> > > > I still think that a user explicitly writing 'extern' is expecting external linkage and all that goes along with it.
> > > I disagree. If this is a special variable to be accessed via dlopen it should be extern "C". If not it should be in a namespace.
> > I'm thinking more that it's not a special variable to be accessed via dlopen, but instead is a special implementation detail that doesn't need to be exposed via a header file. For instance, see ClangTidyMain.cpp. Such uses have no way to silence this check, which may be fine since it's in the Google module. I merely wanted to point it out because it does happen for valid reasons, and having a way to silence a diagnostic aside from NOLINT is nice.
> That's an example of something that should be in a namespace. Currently it's polluting the global namespace which will break anyone who happens to use the same name for a symbol anywhere else. I think that flagging those cases is useful.
Okay, that's fair. I was thinking that since it was in the global namespace, it was there for a reason. After looking a bit more closely, I think you're correct, this would be a true-positive. You've convinced me and I retract my concerns. :-)


https://reviews.llvm.org/D23130