Sorry, please ignore my previous comment. I was looking at the wrong place.

Here are a couple of examples I found running the regression tests:

Call CheckPlaceholderExpr instead of pushing an unevaluated evaluation context.

The patch looks largely fine to me.

Just to clarify, this patch doesn't change clang's handling of C++ unions that have non-static data members with non-trivial special member functions.

ping

LGTM

LGTM

I've reverted the changes I made to test/SemaCUDA/call-host-fn-from-device.cu since r342749 fixed the overload resolution bug.

In D54539#1298949, @rjmccall wrote:

This should not be changing the actual @-encoding value. If we want to use the @-encoding in the symbol name, we can change it when naming the symbol, provided that \1 is not legal in the @encoding.

http://wg21.link/p0968r0#2227 says:

LGTM

Address review comments. Fix bugs in the handling of __block variables that have reference types.

Point the diagnostic at either the relevant init list element or at the close brace.

Move the call that checks the element's destructor into InitListChecker::CheckStructUnionTypes.

The GNUstep documentation I found replaces '@' with '\1'. Would that fix the problem?

ping

In D47757#1211276, @tra wrote:

I've confirmed that the patch does not break anything in our CUDA code, so it's good to go as far as CUDA is concerned.

Thanks. Yes, it's a typo.

The code you showed does compile with or without -fcuda-is-device after applying my patch.

In D47757#1204561, @tra wrote:

It's a regression. There's a decent chance it breaks someone and this patch, if committed by itself, will end up being rolled back.

@tra and @rsmith: Can we move forward and fix the incorrect cuda diagnostics in a separate patch?

Mark the block descriptor global variable as unnamed_addr.

Try harder to shorten the names of block descriptor global variables.

A few points I forgot to mention:

Replace the two flags, EHOnly and DisposeCannotThrow, passed to pushCaptureCleanup with a single flag ForCopyHelper.

Call MarkFunctionReferenced to mark the function __builtin_operator_new or __builtin_operator_delete will call as referenced.

Define functions inline in the header file.

ping

Address review comments.

Address review comments.

Modify getBlockVarCopyInits and setBlockVarCopyInits to get and set both the copy expression and the boolean flag that indicates whether the expression can throw or not.

Change the value type of BlockVarCopyInits to PointerIntPair<Expr *, 1, bool>. The boolean flag indicates whether the copy expression can throw. Serialize/deserialize the copy expression and the boolean flag and add a regression test to test/PCH.

Since BlockVarCopyInits is a map with key VarDecl *, I think we want to add a flag to VarDecl (NonParmVarDeclBits) that indicates whether the copy expression can throw or not. Or is there a reason to store the bit in BlockDecl::Capture instead?

I thought it was okay to skip the work done by ResolveExceptionSpec in IRGen as long as the exception specifications that are needed have already been resolved in Sema. But calling Sema::canThrow in Sema::CheckCompleteVariableDeclaration and storing the result in BlockDecl::Capture is clearly the better solution since it doesn't introduce the complexity introduced in the updated patch.

Remove a stale comment and add an assertion to check the destructor's exception specification has been resolved.