[PATCH] D10599: [OPENMP 4.0] Initial support for '#pragma omp declare simd' directive.

[Richard Smith via cfe-commits]

rsmith added inline comments.

================
Comment at: include/clang/Basic/Attr.td:2098-2102
@@ +2097,7 @@
+  let Documentation = [OMPDeclareSimdDocs];
+  let AdditionalMembers = [{
+  void printPrettyPragma(raw_ostream &OS, const PrintingPolicy &Policy) const {
+    OS << "\n";
+  }
+  }];
+}
----------------
Can we instead handle this in the generated code? Presumably we'd get this wrong in the same way for all attributes with pragma spelling?

================
Comment at: lib/Parse/ParseOpenMP.cpp:133
@@ +132,3 @@
+          << getOpenMPDirectiveName(OMPD_declare_simd);
+      if (!SkipUntil(tok::annot_pragma_openmp_end, StopBeforeMatch))
+        return DeclGroupPtrTy();
----------------
Hmm, on reflection it would be better to use:

  while (Tok.isNot(tok::annot_pragma_openmp_end))
    ConsumeAnyToken();

We know that there is an `annot_pragma_openmp_end` token coming, but `SkipUntil` might stop early if there's (say) a stray right-paren in the token stream.

================
Comment at: lib/Parse/ParseOpenMP.cpp:144-145
@@ +143,4 @@
+    } else {
+      // Here we expect to see some function declaration.
+      while (Tok.isNot(tok::r_brace) && !isEofOrEom() && !Ptr) {
+        if (AS == AS_none) {
----------------
I don't think your example is reasonable to accept. Right now (prior to your patch), we have two fundamentally different kinds of pragmas:

1) Lexer-level pragmas. These can appear anywhere in the token stream, and take effect immediately. These are handled entirely by the lexer.

2) Declaration-like pragmas. These can only appear where a declaration is permitted, and act as if they declare ... something. We transform these into tokens and handle them in the parser.

You're adding a third kind of pragma, an attribute-like pragma, that can only appear before declarations. As such, if your pragma appears before a pragma of kind 2 (such as `#pragma GCC visibility`), it should act as an attribute applying to *that* declaration-like entity.

So I think the only thing you should allow between your pragma and the declaration to which it applies is more attribute-like pragmas.


http://reviews.llvm.org/D10599