[PATCH] Defer codegen of dllexport/attribute(used) inline method definitions

Hans Wennborg hans at chromium.org
Thu Jun 5 18:48:54 PDT 2014


================
Comment at: lib/CodeGen/ModuleBuilder.cpp:92
@@ -91,3 @@
-      if (!D->isDependentContext() &&
-          (D->hasAttr<UsedAttr>() || D->hasAttr<ConstructorAttr>() ||
-           D->hasAttr<DLLExportAttr>())) {
----------------
Richard Smith wrote:
> An alternative fix would be to check `D->isUsed()` instead of checking for `UsedAttr` here. That'd avoid the locality degradation and memory allocation here.
> 
> Or just remove this check and unconditionally call `EmitTopLevelDecl`? That would better match what we do for out-of-line method definitions, and would avoid duplicating this between here and `ASTContext::DeclMustBeEmitted`.
Checking D->isUsed() certainly sounds simpler :-) I'll try that out.

We can't unconditionally call EmitTopLevelDecl, because it will call DeclMustBeEmitted, which might calculate the gva linkage, and then we fail this test from SemaCXX/undefined-internal.cpp :(

      namespace test7 {
        typedef struct {
          void bar();
          void foo() { bar(); }
        } A;
      }

I suspect that test doesn't work well with the "used" attribute either.

================
Comment at: test/CodeGenCXX/attr-used.cpp:21
@@ +20,3 @@
+  // We must delay emission of bar() until foo() has had its body parsed,
+  // othersie foo() would not be emitted.
+  void __attribute__((used)) bar() { foo(); }
----------------
Richard Smith wrote:
> Reid Kleckner wrote:
> > "otherwise"
> Typo "othersie"
Done.

http://reviews.llvm.org/D4038






More information about the cfe-commits mailing list