Sorry for the mess up, but I accidentally did a reply instead of reply-all. This includes my response to Richard Smith, his reply to me, and my reply to his reply.<br><br><div class="gmail_quote">On Mon, Apr 4, 2011 at 4:39 PM, Richard Smith <span dir="ltr"><<a href="mailto:richard@metafoo.co.uk">richard@metafoo.co.uk</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;"><div class="im">On Mon, April 4, 2011 22:57, Richard Trieu wrote:<br>
> On Mon, Apr 4, 2011 at 1:19 PM, Richard Smith <<a href="mailto:richard@metafoo.co.uk">richard@metafoo.co.uk</a>><br>
> wrote:<br>
>> On Wed, 30 Mar 2011, Richard Trieu <<a href="mailto:rtrieu@google.com">rtrieu@google.com</a>> wrote:<br>
>>> Detect when the string "<::" is found in code after a cast or<br>
>>> template name and is interpreted as "[:" because of the digraph "<:".<br>
>>> When found,<br>
>>> give an error with a fix-it to add whitespace between the "<" and<br>
>>> "::".<br>
>>> Also, repair the token stream and recover parsing afterwards.<br>
>><br>
>> I don't think that this is the right fix. In C++0x, '<::' (not followed<br>
>> by ':' or '>') is lexed as '<' followed by '::' (see [lex.pptoken]p3<br>
>> bullet 2, or<br>
>> <a href="http://www.open-std.org/jtc1/sc22/wg21/docs/cwg_defects.html#1104" target="_blank">http://www.open-std.org/jtc1/sc22/wg21/docs/cwg_defects.html#1104</a>).<br>
>> Implementing that rule (for both C++0x and earlier) seems to me like a<br>
>> better fit, even though it breaks legal-but-pathological C++98 programs<br>
>> like this:<br>
>><br>
>> int a[42]; #define A(n) a<::##:n:><br>
>> int b = A(b);<br>
><br>
</div><div class="im">> I don't think it's a good idea to break valid code. Also, your link<br>
> shows that fix in the lexer is not yet in the standard.<br>
<br>
</div>Have a look at [lex.pptoken]p3 in N3242, the latest public draft of the<br>
C++0x standard. It's there already :)<br></blockquote><div>Ah, okay. I see it now. I thought that you had linked to a discussion of possible features.</div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div class="im"><br>
> Even if it were, it<br>
> should only be applied to C++0x with the parser handling this error in<br>
> other versions.<br>
<br>
</div>I don't have a particularly strong opinion on this. Personally, I'm much<br>
more interested in recovering nicely for code which accidentally contains<br>
'<::' and means '< ::' than I am in preserving the semantics of<br>
theoretical code where '<::' is supposed to be a digraph, but I'm not sure<br>
what's the best fit for clang's goals (my preference would be that this is<br>
an ExtWarn with a fixit outside of C++0x).<br>
<br>
If you still want to implement it this way, I have some comments on the<br>
patch itself:<br>
<br>
It would be nice to also cover the comparison case: '::a <::b'.</blockquote><div>Maybe later. This patch was limited to prevent unforeseen impacts to other code. </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<br>
Index: lib/Parse/ParseExprCXX.cpp<br>
===================================================================<br>
--- lib/Parse/ParseExprCXX.cpp (revision 128572)<br>
+++ lib/Parse/ParseExprCXX.cpp (working copy)<br>
@@ -20,6 +20,45 @@<br>
<br>
using namespace clang;<br>
<br>
+static int SelectDigraphErrorMessage(tok::TokenKind Kind) {<br>
+ switch (Kind) {<br>
+ case tok::kw_template: return 0;<br>
+ case tok::kw_const_cast: return 1;<br>
+ case tok::kw_dynamic_cast: return 2;<br>
+ case tok::kw_reinterpret_cast: return 3;<br>
+ case tok::kw_static_cast: return 4;<br>
+ default:<br>
+ assert(0 && "Unknown type for digraph error message.");<br>
+ return -1;<br>
+ }<br>
+}<br>
+<br>
+// Suggest fixit for "<::" after a cast.<br>
+static void FixDigraph(Parser &P, Preprocessor &PP, Token &DigraphToken,<br>
+ Token &ColonToken, tok::TokenKind Kind, bool<br>
AtDigraph) {<br>
+ // Pull '<:' and ':' off token stream.<br>
+ if (!AtDigraph)<br>
+ PP.Lex(DigraphToken);<br>
+ PP.Lex(ColonToken);<br>
+<br>
+ SourceRange Range;<br>
+ Range.setBegin(DigraphToken.getLocation());<br>
+ Range.setEnd(ColonToken.getLocation());<br>
+ P.Diag(DigraphToken.getLocation(), diag::err_missing_whitespace_digraph)<br>
+ << SelectDigraphErrorMessage(Kind)<br>
+ << FixItHint::CreateReplacement(Range, "< ::");<br>
+<br>
+ // Reuse tokens to preserve source location and macro instantiation<br>
+ // information.<br>
+ ColonToken.setKind(tok::coloncolon);<br>
<br>
The source location for this token will be off by one character. Later<br>
diagnostics could be confusing. </blockquote><div>The source location would be off by 1, pointing to the second colon instead of the first. Would that cause confusion? </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<br>
+ DigraphToken.setKind(tok::less);<br>
+<br>
+ // Push new tokens back to token stream.<br>
+ PP.EnterToken(ColonToken);<br>
+ if (!AtDigraph)<br>
+ PP.EnterToken(DigraphToken);<br>
+}<br>
+<br>
/// \brief Parse global scope or nested-name-specifier if present.<br>
///<br>
/// Parses a C++ global scope specifier ('::') or nested-name-specifier<br>
(which<br>
@@ -287,6 +326,28 @@<br>
continue;<br>
}<br>
<br>
+ // Check for '<::' which should be '< ::' instead of '[:' when following<br>
+ // a template name.<br>
+ if (Next.is(tok::l_square) && Next.getLength() == 2) {<br>
+ Token SecondToken = GetLookAheadToken(2);<br>
+ if (SecondToken.is(tok::colon)) {<br>
<br>
This recovery codepath should only be entered if the '[' and ':' are<br>
adjacent in the source file.<br></blockquote><div>I'll have a look into it. I'm guessing whitespace would throw off the parser?</div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<br>
+ TemplateTy Template;<br>
+ UnqualifiedId TemplateName;<br>
+ TemplateName.setIdentifier(&II, Tok.getLocation());<br>
+ bool MemberOfUnknownSpecialization;<br>
+ if (Actions.isTemplateName(getCurScope(), SS,<br>
+ /*hasTemplateKeyword=*/false,<br>
+ TemplateName,<br>
+ ObjectType,<br>
+ EnteringContext,<br>
+ Template,<br>
+ MemberOfUnknownSpecialization)) {<br>
+ FixDigraph(*this, PP, Next, SecondToken, tok::kw_template,<br>
+ /*AtDigraph*/false);<br>
+ }<br>
+ }<br>
+ }<br>
+<br>
// nested-name-specifier:<br>
// type-name '<'<br>
if (Next.is(tok::less)) {<br>
@@ -453,6 +514,13 @@<br>
SourceLocation OpLoc = ConsumeToken();<br>
SourceLocation LAngleBracketLoc = Tok.getLocation();<br>
<br>
+ // Check for "<::" which is parsed as "[:". If found, fix token stream,<br>
+ // diagnose error, suggest fix, and recover parsing.<br>
+ Token Next = NextToken();<br>
+ if (Tok.is(tok::l_square) && Tok.getLength() == 2 &&<br>
Next.is(tok::colon)) {<br>
<br>
Likewise here.<br>
<br>
+ FixDigraph(*this, PP, Tok, Next, Kind, /*AtDigraph*/true);<br>
+ }<br>
+<br>
if (ExpectAndConsume(tok::less, diag::err_expected_less_after, CastName))<br>
return ExprError();<br>
<br>
<br>
Incidentally, did you intentionally remove cfe-commits from CC?<br></blockquote><div><br></div><div>No, I just clicked Reply instead of Reply-All. They're like right next to each other.</div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<font color="#888888"><br>
Richard<br>
<br>
</font></blockquote></div>Other Richard