Fix a clang bug in lambda capture of 'this'

[twoh]

Summary

(This is a fix for Bug 27797 https://llvm.org/bugs/show_bug.cgi?id=27797).

Currently when RebuildLambdaScopeInfo() function in lib/Sema/SemaDecl.cpp observes lambda capturing 'this', it calls Sema::getCurrentThisType() to get the type of 'this'. However, clang (revision 269789) crashes with an assertion failure inside Sema::getCurrentThisType() if template instantiation creates nested lambdas. Inside, Sema::getCurrentThisType(), there is an assertion saying that getCurLambda() never returns nullptr, but nest lambdas created by template instantiation makes getCurLambda() returns nullptr.

Actually, even without the assertion failure, calling Sema::getCurrentThisType() from RebuildLambdaScopeInfo() seems wrong. When there are nested lambdas, what is required from Sema::getCurrentThisType() is a type of 'this' for nesting lambda, while what is supposed to be returned from Sema::getCurrentThisType() is a type of 'this' for nested lambda.

This patch addresses this issue and makes RebuildLambdaScopeInfo() compute the correct 'this' type.

[dblaikie]

The patch'll need a test case (in clang/tests) - I've not looked at the
change/have much opinion there, just suggesting adding a test at minimum so
when someone does come to review it it's more complete/closer to/ready to
commit.

[faisalv]

Thanks for the work here - but this bug is a duplicate of this:
https://llvm.org/bugs/show_bug.cgi?id=27507, which unearthed some
other subtle bugs with the cv capture of *this by copy.
I've submitted a patch for review here: http://reviews.llvm.org/D19783
that addresses both issues - about 2-3 weeks ago - given the
nastiness of this bug, the fact that it was introduced by one of my
patches and that there is now a second report of the bug - and that I
feel reasonably good about the fix - I'll commit this patch before the
weekend, unless Richard raises an issue.

Faisal Vali

[twoh]

Thank you for your comments. @faisalv, it is great that you already submitted a patch. Let me see if your patch resolves the issue I have. Thanks!

[faisalv]

That feedback would be greatly appreciated - thanks Taewook!
Faisal Vali

[faisalv]

Should be fixed by: http://reviews.llvm.org/D19783.

Thanks again for your feedback and all the time you put into this =)

lib/Sema/SemaDecl.cpp

if (C.capturesVariable()) {
       const bool ByRef = C.getCaptureKind() == LCK_ByRef;
       LSI->addCapture(VD, /*IsBlock*/false, ByRef,
           /*RefersToEnclosingVariableOrCapture*/true, C.getLocation(),
           /*EllipsisLoc*/C.isPackExpansion()
                          ? C.getEllipsisLoc() : SourceLocation(),
           CaptureType, /*Expr*/ nullptr);
 
     } else if (C.capturesThis()) {
+      QualType ThisTy = CallOperator->getThisType(S.Context);
+      QualType BaseTy = ThisTy->getPointeeType();
+      if (C.getCaptureKind() == LCK_StarThis &&
+          CallOperator->isConst() &&
+          !BaseTy.isConstQualified()) {
+        BaseTy.addConst();
+        ThisTy = S.Context.getPointerType(BaseTy);
+      }
       LSI->addThisCapture(/*Nested*/ false, C.getLocation(),
-                              S.getCurrentThisType(), /*Expr*/ nullptr,
+                              ThisTy, /*Expr*/ nullptr,
                               C.getCaptureKind() == LCK_StarThis);
     } else {
       LSI->addVLATypeCapture(C.getLocation(), I->getType());
     }
     ++I;
   }
 }