Loss of qualifiers in function parameters in Clang template instantiations about calypso

[Syniurge]

Hi Kelly,

I started working on the cppaggregate branch, only showcase.d works for the time being but it should catch up quickly since this simplifies things a great deal (the cost being that we'll get different semantics for D classes and for C++ classes, but no more ambiguities between whether a class is value or reference based on C++ compilation options).

I'll have a look at list.d once I bring back vector, string, bitset, deque and map.

[wilsonk]

Ok, sounds good Elie.

[Syniurge]

The bug seems to come from DMD: the deduced type for _Predicate throws away the const qualifier in « bool function(const(int) value) », so the previously matching remove_if instance gets discarded too by TypeFunction::callMatch().

Removing const in single_digit() makes list.d compile and run. I'll track where does const disappear tomorrow.

[Syniurge]

Okay, it's not DMD but this line in Sema while sending the (correct) template argument to instantiate remove_if<...>:

    ArgType = Context.getCanonicalType(ArgType);

Prior to this call:

 (gdb) call ArgType.dump()
 PointerType 0xc0d86f0 'PointerType FunctionProtoType BuiltinType _Bool (*)(const BuiltinType int)'
 `-FunctionProtoType 0xc0d8680 'FunctionProtoType BuiltinType _Bool (const BuiltinType int)' cdecl
   |-BuiltinType 0xa33e520 'BuiltinType _Bool'
   `-QualType 0xa33e5a1 'const BuiltinType int' const
     `-BuiltinType 0xa33e5a0 'BuiltinType int'

Then:

 (gdb) call ArgType.dump()
 PointerType 0xc0d86c0 'PointerType FunctionProtoType BuiltinType _Bool (*)(BuiltinType int)'
 `-FunctionProtoType 0xc0d8640 'FunctionProtoType BuiltinType _Bool (BuiltinType int)' cdecl
   |-BuiltinType 0xa33e520 'BuiltinType _Bool'
   `-BuiltinType 0xa33e5a0 'BuiltinType int'

Clang bug or expected behavior? If it's a bug it's not causing too much harm in Clang which is probably why it's still here, and if it's expected behavior I can't understand why and accustoming the Calypso side to this strange behavior of getCanonicalType won't be easy. In any case I think we should stick with the workaround for now.

[Syniurge]

It's still in latest Clang 3.7 and it looks like stripping the parameters from qualifiers in canonical function type is sought after..:

http://clang.llvm.org/doxygen/classclang_1_1ASTContext.html#aebe1c0862e34693bddd4683b51558e35 (ASTContext::getCanonicalParamType)

[wilsonk]

Looks like just using the workaround is the way to go. I tried compiling the cppaggregate branch to test this and I ran into an error in mtype.c...I suppose you must have this fixed on your end??

diff --git a/dmd2/mtype.c b/dmd2/mtype.c
index 19fcc2b..d87862b 100644
--- a/dmd2/mtype.c
+++ b/dmd2/mtype.c
@@ -9797,7 +9797,7 @@ AggregateDeclaration *getAggregateSym(Type *t)

TypeClass *isClassValue(Type *t)
{

• if (t->ty != Tclass) return false;
• if (t->ty != Tclass) return NULL;
  TypeClass* tc = (TypeClass*)t;
  return tc->byRef() ? NULL : tc;
  }

Anyways, it looks like things are back to about the same, for libstdc++ tests, as they were a couple weeks ago. I'll see if I can get around to testing some other libs. I can see a few errors in my qt5demo extended version, so some rewriting will be needed.

Is there anything else left to do on this branch or is it going to merge with master right away?

[Syniurge]

Good catch, and no it's just that my compiler is a cool dude :)

On to the Qt5 demo, and then it'll be on par with master so ready to merge.

[Syniurge]

This was definitely fixed with the switch from DMD's to Clang's template argument deduction.