[cfe-commits] PATCH #1 for "[sema++] clang should give a fixit for */& mismatch"

[Douglas Gregor]

On Jul 14, 2011, at 7:26 PM, Anna Zaks wrote:

> This patch is a first step towards providing FixItHints in case a function call is missing one/several * or & operators (See http://llvm.org/PR5941).
> 
> With this fix, the error message will display the function call. It will be followed by the list of notes - one for each possible candidate function declaration. If a function call can be fixed to match the particular declaration, a FixItHint will be provided along with the note. 

Very cool. Comments below.

> I also bumped up the number of MaxFixItHints to 6 (we need 2 FixIts per incorrect argument).
> 
> The remaining steps are:
> 1) Use FixIts when determining the order in which the notes should be displayed.
> 2) Find out if the same FixIts can be used with errors other then "error: no matching function for call to foo".


+def note_ovl_candidate_bad_conv_fixit : Note<"candidate "
+    "%select{function|function|constructor|"
+    "function |function |constructor |"
+    "constructor (the implicit default constructor)|"
+    "constructor (the implicit copy constructor)|"
+    "constructor (the implicit move constructor)|"
+    "function (the implicit copy assignment operator)|"
+    "function (the implicit move assignment operator)|"
+    "constructor (inherited)}0%1"
+    " not viable: no known conversion from %2 to %3 for "
+    "%select{%ordinal5 argument|object argument}4; "
+    "did you miss * or & operator%s6?">;

The diagnostic should suggest either * or &, depending on what would actually fix the code, rather than requiring the user to think about it further. This could be done with some kind of select, e.g.,

	%select{|;dereference the argument with *|;take the address of the argument with &}7

which has the advantage of being extensible (e.g., in case we want to also suggest things like () to turn a function expression into a function call) and could even be tacked on to note_ovl_candidate_bad_conv rather than creating a new diagnostic.

Index: lib/Sema/SemaOverload.cpp
===================================================================
--- lib/Sema/SemaOverload.cpp	(revision 135234)
+++ lib/Sema/SemaOverload.cpp	(working copy)
@@ -6730,6 +6730,54 @@
 
 namespace {
 
+/// Update the Hints with FixIts to correct the given conversion. Return true
+/// on success.
+static bool TryToFixBadConversion(const Sema &S,
+                                  const ImplicitConversionSequence &Conv,
+                                  llvm::SmallVector<FixItHint, 4> &Hints) {
+  assert(Conv.isBad());
+  const Expr *Arg = Conv.Bad.FromExpr;
+  if (!Arg)
+    return false;
+
+  // The conversion is from argument type to parameter type.
+  const CanQualType FromQTy = S.Context.getCanonicalType(Conv.Bad
+                                                         .getFromType());
+  const CanQualType ToQTy = S.Context.getCanonicalType(Conv.Bad.getToType());
+
+  const SourceLocation Begin = Arg->getSourceRange().getBegin();
+  const SourceLocation End = S.PP.getLocForEndOfToken(Arg->getSourceRange()
+                                                      .getEnd());
+
+  // Check if the argument needs to be dereferenced
+  // (type * -> type) or (type * -> type &).
+  if (const CanQual<PointerType> FromPtrTy = FromQTy->getAs<PointerType>()) {
+    const CanQualType ToTy = ToQTy.getNonReferenceType();
+    const CanQualType FromTy = FromPtrTy->getPointeeType();
+    if (ToTy.getUnqualifiedType() == FromTy.getUnqualifiedType() &&
+        ToTy.isAtLeastAsQualifiedAs(FromTy)) {

It's probably worth loosening the type checking here slightly, to support (for example) derived-to-base conversions and Objective-C object pointer conversions. I suspect you'll want to factor that check out into a separate routine, since a similar check is needed for &.

+      Hints.push_back(FixItHint::CreateInsertion(Begin, "*("));
+      Hints.push_back(FixItHint::CreateInsertion(End, ")"));
+      return true;
+    }
+  }

The parentheses are necessary only in a few cases… can we perhaps whitelist common kinds of expressions (DeclRefExprs, ParenExprs, etc.) where the parentheses aren't needed, and then just emit the simple "*" in those cases? We want to try fairly hard to introduce the simplest, correct Fix-It we can.

+
+  // Check if the pointer to the argument needs to be passed
+  // (type -> type *) or (type & -> type *).
+  if (const CanQual<PointerType> ToPtrTy = ToQTy->getAs<PointerType>()) {
+    const CanQualType FromTy = FromQTy.getNonReferenceType();
+    const CanQualType ToTy = ToPtrTy->getPointeeType();
+    if (FromTy.getUnqualifiedType() == ToTy.getUnqualifiedType() &&
+        ToTy.isAtLeastAsQualifiedAs(FromTy)) {
+      Hints.push_back(FixItHint::CreateInsertion(Begin, "&("));
+      Hints.push_back(FixItHint::CreateInsertion(End, ")"));
+      return true;
+    }
+  }
+

For this case, we'll also need to check whether the expression is an lvalue (Expr::isLValue()), so that we don't produce a bogus fix-it for something like:

	void f(int*);
	void g() { f(17); } // can't apply the & operator here

Your patch is looking quite good; thanks for working on this!

	- Doug