diff --git a/gen/dvalue.cpp b/gen/dvalue.cpp
index cf0a4bdd4e..e0057654da 100644
--- a/gen/dvalue.cpp
+++ b/gen/dvalue.cpp
@@ -16,6 +16,8 @@ DVarValue::DVarValue(Type* t, VarDeclaration* vd, LLValue* llvmValue)
 : DValue(t), var(vd), val(llvmValue)
 {
     assert(isaPointer(llvmValue));
+    assert(!isSpecialRefVar(vd) ||
+        isaPointer(isaPointer(llvmValue)->getElementType()));
 }
 
 DVarValue::DVarValue(Type* t, LLValue* llvmValue)
@@ -27,6 +29,8 @@ DVarValue::DVarValue(Type* t, LLValue* llvmValue)
 LLValue* DVarValue::getLVal()
 {
     assert(val);
+    if (var && isSpecialRefVar(var))
+        return DtoLoad(val);
     return val;
 }
 
@@ -34,9 +38,14 @@ LLValue* DVarValue::getRVal()
 {
     assert(val);
     Type* bt = type->toBasetype();
+
+    LLValue* tmp = val;
+    if (var && isSpecialRefVar(var))
+        tmp = DtoLoad(tmp);
+
     if (DtoIsPassedByRef(bt))
-        return val;
-    return DtoLoad(val);
+        return tmp;
+    return DtoLoad(tmp);
 }
 
 /////////////////////////////////////////////////////////////////////////////////////////////////
diff --git a/gen/functions.cpp b/gen/functions.cpp
index 287464dc3f..5e540c2174 100644
--- a/gen/functions.cpp
+++ b/gen/functions.cpp
@@ -815,8 +815,7 @@ void DtoDefineFunction(FuncDeclaration* fd)
     {
         DtoNestedInit(fd->vresult);
     } else if (fd->vresult) {
-        fd->vresult->ir.irLocal = new IrLocal(fd->vresult);
-        fd->vresult->ir.irLocal->value = DtoAlloca(fd->vresult->type, fd->vresult->toChars());
+        DtoVarDeclaration(fd->vresult);
     }
 
     // copy _argptr and _arguments to a memory location
diff --git a/gen/llvmhelpers.cpp b/gen/llvmhelpers.cpp
index 15283c7778..de90a17f6f 100644
--- a/gen/llvmhelpers.cpp
+++ b/gen/llvmhelpers.cpp
@@ -1013,6 +1013,110 @@ void DtoConstInitGlobal(VarDeclaration* vd)
 /*////////////////////////////////////////////////////////////////////////////////////////
 //      DECLARATION EXP HELPER
 ////////////////////////////////////////////////////////////////////////////////////////*/
+
+// TODO: Merge with DtoRawVarDeclaration!
+void DtoVarDeclaration(VarDeclaration* vd)
+{
+    assert(!vd->isDataseg() && "Statics/globals are handled in DtoDeclarationExp.");
+    assert(!vd->aliassym && "Aliases are handled in DtoDeclarationExp.");
+
+    Logger::println("vdtype = %s", vd->type->toChars());
+
+#if DMDV2
+    if (vd->nestedrefs.dim)
+#else
+    if (vd->nestedref)
+#endif
+    {
+        Logger::println("has nestedref set (referenced by nested function/delegate)");
+        assert(vd->ir.irLocal);
+        DtoNestedInit(vd);
+    }
+    else if(vd->ir.irLocal)
+    {
+        // Nothing to do if it has already been allocated.
+    }
+#if DMDV2
+    /* Named Return Value Optimization (NRVO):
+        T f(){
+            T ret;        // &ret == hidden pointer
+            ret = ...
+            return ret;    // NRVO.
+        }
+    */
+    else if (gIR->func()->retArg && gIR->func()->decl->nrvo_can && gIR->func()->decl->nrvo_var == vd) {
+        assert(!isSpecialRefVar(vd) && "Can this happen?");
+        vd->ir.irLocal = new IrLocal(vd);
+        vd->ir.irLocal->value = gIR->func()->retArg;
+    }
+#endif
+    // normal stack variable, allocate storage on the stack if it has not already been done
+    else {
+        vd->ir.irLocal = new IrLocal(vd);
+
+#if DMDV2
+        /* NRVO again:
+            T t = f();    // t's memory address is taken hidden pointer
+        */
+        ExpInitializer *ei = 0;
+        if (vd->type->toBasetype()->ty == Tstruct && vd->init &&
+            !!(ei = vd->init->isExpInitializer()))
+        {
+            if (ei->exp->op == TOKconstruct) {
+                AssignExp *ae = static_cast<AssignExp*>(ei->exp);
+                if (ae->e2->op == TOKcall) {
+                    CallExp *ce = static_cast<CallExp *>(ae->e2);
+                    TypeFunction *tf = static_cast<TypeFunction *>(ce->e1->type->toBasetype());
+                    if (tf->ty == Tfunction && tf->fty.arg_sret) {
+                        LLValue* const val = ce->toElem(gIR)->getLVal();
+                        if (isSpecialRefVar(vd))
+                        {
+                            vd->ir.irLocal->value = DtoAlloca(
+                                vd->type->pointerTo(), vd->toChars());
+                            DtoStore(val, vd->ir.irLocal->value);
+                        }
+                        else
+                        {
+                            vd->ir.irLocal->value = val;
+                        }
+                        goto Lexit;
+                    }
+                }
+            }
+        }
+#endif
+
+        Type* type = isSpecialRefVar(vd) ? vd->type->pointerTo() : vd->type;
+        LLType* lltype = DtoType(type);
+
+        llvm::Value* allocainst;
+        if(gTargetData->getTypeSizeInBits(lltype) == 0)
+            allocainst = llvm::ConstantPointerNull::get(getPtrToType(lltype));
+        else
+            allocainst = DtoAlloca(type, vd->toChars());
+
+        vd->ir.irLocal->value = allocainst;
+
+        DtoDwarfLocalVariable(allocainst, vd);
+    }
+
+    if (Logger::enabled())
+        Logger::cout() << "llvm value for decl: " << *vd->ir.irLocal->value << '\n';
+
+    DtoInitializer(vd->ir.irLocal->value, vd->init); // TODO: Remove altogether?
+
+#if DMDV2
+Lexit:
+    /* Mark the point of construction of a variable that needs to be destructed.
+     */
+    if (vd->edtor && !vd->noscope)
+    {
+        // Put vd on list of things needing destruction
+        gIR->varsInScope().push_back(vd);
+    }
+#endif
+}
+
 DValue* DtoDeclarationExp(Dsymbol* declaration)
 {
     Logger::print("DtoDeclarationExp: %s\n", declaration->toChars());
@@ -1036,123 +1140,8 @@ DValue* DtoDeclarationExp(Dsymbol* declaration)
         }
         else
         {
-            if (global.params.llvmAnnotate)
-                DtoAnnotation(declaration->toChars());
-
-            Logger::println("vdtype = %s", vd->type->toChars());
-
-            // ref vardecls are generated when DMD lowers foreach to a for statement,
-            // and this is a hack to support them for this case only
-            if(vd->isRef())
-            {
-                if (!vd->ir.irLocal)
-                    vd->ir.irLocal = new IrLocal(vd);
-
-                ExpInitializer* ex = vd->init->isExpInitializer();
-                assert(ex && "ref vars must have expression initializer");
-                assert(ex->exp);
-                AssignExp* as = ex->exp->isAssignExp();
-                assert(as && "ref vars must be initialized by an assign exp");
-                DValue *val = as->e2->toElem(gIR);
-                if (val->isLVal())
-                {
-                    vd->ir.irLocal->value = val->getLVal();
-                }
-                else
-                {
-                    LLValue *newVal = DtoAlloca(val->type);
-                    DtoStore(val->getRVal(), newVal);
-                    vd->ir.irLocal->value = newVal;
-                }
-            }
-
-            // referenced by nested delegate?
-        #if DMDV2
-            if (vd->nestedrefs.dim) {
-        #else
-            if (vd->nestedref) {
-        #endif
-                Logger::println("has nestedref set");
-                assert(vd->ir.irLocal);
-                DtoNestedInit(vd);
-            // is it already allocated?
-            } else if(vd->ir.irLocal) {
-                // nothing to do...
-            }
-#if DMDV2
-            /* Named Return Value Optimization (NRVO):
-                T f(){
-                    T ret;        // &ret == hidden pointer
-                    ret = ...
-                    return ret;    // NRVO.
-                }
-            */
-            else if (gIR->func()->retArg && gIR->func()->decl->nrvo_can && gIR->func()->decl->nrvo_var == vd) {
-                vd->ir.irLocal = new IrLocal(vd);
-                vd->ir.irLocal->value = gIR->func()->retArg;
-            }
-#endif
-            // normal stack variable, allocate storage on the stack if it has not already been done
-            else if(!vd->isRef()) {
-                vd->ir.irLocal = new IrLocal(vd);
-
-#if DMDV2
-                /* NRVO again:
-                    T t = f();    // t's memory address is taken hidden pointer
-                */
-                ExpInitializer *ei = 0;
-                if (vd->type->toBasetype()->ty == Tstruct && vd->init &&
-                    !!(ei = vd->init->isExpInitializer()))
-                {
-                    if (ei->exp->op == TOKconstruct) {
-                        AssignExp *ae = static_cast<AssignExp*>(ei->exp);
-                        if (ae->e2->op == TOKcall) {
-                            CallExp *ce = static_cast<CallExp *>(ae->e2);
-                            TypeFunction *tf = static_cast<TypeFunction *>(ce->e1->type->toBasetype());
-                            if (tf->ty == Tfunction && tf->fty.arg_sret) {
-                                vd->ir.irLocal->value = ce->toElem(gIR)->getLVal();
-                                goto Lexit;
-                            }
-                        }
-                    }
-                }
-#endif
-
-                LLType* lltype = DtoType(vd->type);
-
-                llvm::Value* allocainst;
-                if(gTargetData->getTypeSizeInBits(lltype) == 0)
-                    allocainst = llvm::ConstantPointerNull::get(getPtrToType(lltype));
-                else
-                    allocainst = DtoAlloca(vd->type, vd->toChars());
-
-                //allocainst->setAlignment(vd->type->alignsize()); // TODO
-                vd->ir.irLocal->value = allocainst;
-
-                DtoDwarfLocalVariable(allocainst, vd);
-            }
-            else
-            {
-                assert(vd->ir.irLocal->value);
-            }
-
-            if (Logger::enabled())
-                Logger::cout() << "llvm value for decl: " << *vd->ir.irLocal->value << '\n';
-            if (!vd->isRef())
-                DtoInitializer(vd->ir.irLocal->value, vd->init); // TODO: Remove altogether?
-
-#if DMDV2
-        Lexit:
-            /* Mark the point of construction of a variable that needs to be destructed.
-             */
-            if (vd->edtor && !vd->noscope)
-            {
-                // Put vd on list of things needing destruction
-                gIR->varsInScope().push_back(vd);
-            }
-#endif
+            DtoVarDeclaration(vd);
         }
-
         return new DVarValue(vd->type, vd, vd->ir.getIrValue());
     }
     // struct declaration
@@ -1921,6 +1910,13 @@ void callPostblit(Loc &loc, Expression *exp, LLValue *val)
 
 //////////////////////////////////////////////////////////////////////////////////////////
 
+bool isSpecialRefVar(VarDeclaration* vd)
+{
+    return (vd->storage_class & STCref) && (vd->storage_class & STCforeach);
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////
+
 void printLabelName(std::ostream& target, const char* func_mangle, const char* label_name)
 {
     target << gTargetMachine->getMCAsmInfo()->getPrivateGlobalPrefix() <<
diff --git a/gen/llvmhelpers.h b/gen/llvmhelpers.h
index ab80da921b..1fd85a0735 100644
--- a/gen/llvmhelpers.h
+++ b/gen/llvmhelpers.h
@@ -98,6 +98,7 @@ void DtoResolveDsymbol(Dsymbol* dsym);
 void DtoConstInitGlobal(VarDeclaration* vd);
 
 // declaration inside a declarationexp
+void DtoVarDeclaration(VarDeclaration* var);
 DValue* DtoDeclarationExp(Dsymbol* declaration);
 LLValue* DtoRawVarDeclaration(VarDeclaration* var, LLValue* addr = 0);
 
@@ -158,6 +159,14 @@ LLValue* makeLValue(Loc& loc, DValue* value);
 void callPostblit(Loc &loc, Expression *exp, LLValue *val);
 #endif
 
+/// Returns whether the given variable is a DMD-internal "ref variable".
+///
+/// D doesn't have reference variables (the ref keyword is only usable in
+/// function signatures and foreach headers), but the DMD frontend internally
+/// creates them in cases like lowering a ref foreach to a for loop or the
+/// implicit __result variable for ref-return functions with out contracts.
+bool isSpecialRefVar(VarDeclaration* vd);
+
 ////////////////////////////////////////////
 // gen/tocall.cpp stuff below
 ////////////////////////////////////////////
diff --git a/gen/nested.cpp b/gen/nested.cpp
index 1e4213abd8..7688d6dab7 100644
--- a/gen/nested.cpp
+++ b/gen/nested.cpp
@@ -211,7 +211,7 @@ DValue* DtoNestedVariable(Loc loc, Type* astype, VarDeclaration* vd, bool byref)
         val = DtoGEPi(val, 0, vd->ir.irLocal->nestedIndex, vd->toChars());
         Logger::cout() << "Addr: " << *val << '\n';
         Logger::cout() << "of type: " << *val->getType() << '\n';
-        if (vd->ir.irLocal->byref || byref) {
+        if (byref || (vd->isParameter() && vd->ir.irParam->arg->byref)) {
             val = DtoAlignedLoad(val);
             //dwarfOpDeref(dwarfAddr);
             Logger::cout() << "Was byref, now: " << *val << '\n';
@@ -251,27 +251,7 @@ void DtoNestedInit(VarDeclaration* vd)
         DtoAlignedStore(val, gep);
     }
     else if (nestedCtx == NCHybrid) {
-        assert(vd->ir.irLocal->value && "Nested variable without storage?");
-
-        if (!vd->isParameter() && (vd->isRef() || vd->isOut())) {
-            unsigned vardepth = vd->ir.irLocal->nestedDepth;
-
-            LLValue* val = NULL;
-            // Retrieve frame pointer
-            if (vardepth == irfunc->depth) {
-                val = nestedVar;
-            } else {
-                FuncDeclaration *parentfunc = getParentFunc(vd, true);
-                assert(parentfunc && "No parent function for nested variable?");
-
-                val = DtoGEPi(nestedVar, 0, vardepth);
-                val = DtoAlignedLoad(val, (std::string(".frame.") + parentfunc->toChars()).c_str());
-            }
-            val = DtoGEPi(val, 0, vd->ir.irLocal->nestedIndex, vd->toChars());
-            storeVariable(vd, val);
-        } else {
-            // Already initialized in DtoCreateNestedContext
-        }
+        // Already initialized in DtoCreateNestedContext.
     }
     else {
         assert(0 && "Not implemented yet");
@@ -492,18 +472,13 @@ static void DtoCreateNestedContextType(FuncDeclaration* fd) {
                             type = type->getContainedType(0);
                         else
                             type = DtoType(vd->type);
-                        vd->ir.irParam->byref = false;
                     } else {
-                        vd->ir.irParam->byref = true;
                     }
                     types.push_back(type);
-                } else if (vd->isRef() || vd->isOut()) {
-                    // Foreach variables can also be by reference, for instance.
+                } else if (isSpecialRefVar(vd)) {
                     types.push_back(DtoType(vd->type->pointerTo()));
-                    vd->ir.irLocal->byref = true;
                 } else {
                     types.push_back(DtoType(vd->type));
-                    vd->ir.irLocal->byref = false;
                 }
                 if (Logger::enabled()) {
                     Logger::println("Nested var: %s", vd->toChars());
@@ -692,36 +667,26 @@ void DtoCreateNestedContext(FuncDeclaration* fd) {
                 if (vd->isParameter()) {
                     Logger::println("nested param: %s", vd->toChars());
                     LOG_SCOPE
-                    LLValue* value = vd->ir.irLocal->value;
-                    if (llvm::isa<llvm::AllocaInst>(llvm::GetUnderlyingObject(value))) {
+                    IrParameter* parm = vd->ir.irParam;
+
+                    if (parm->arg->byref)
+                    {
+                        storeVariable(vd, gep);
+                    }
+                    else
+                    {
                         Logger::println("Copying to nested frame");
                         // The parameter value is an alloca'd stack slot.
                         // Copy to the nesting frame and leave the alloca for
                         // the optimizers to clean up.
-                        assert(!vd->ir.irLocal->byref);
-                        DtoStore(DtoLoad(value), gep);
-                        gep->takeName(value);
-                        vd->ir.irLocal->value = gep;
-                    } else {
-                        Logger::println("Adding pointer to nested frame");
-                        // The parameter value is something else, such as a
-                        // passed-in pointer (for 'ref' or 'out' parameters) or
-                        // a pointer arg with byval attribute.
-                        // Store the address into the frame.
-                        assert(vd->ir.irLocal->byref);
-                        storeVariable(vd, gep);
+                        DtoStore(DtoLoad(parm->value), gep);
+                        gep->takeName(parm->value);
+                        parm->value = gep;
                     }
-                } else if (vd->isRef() || vd->isOut()) {
-                    // This slot is initialized in DtoNestedInit, to handle things like byref foreach variables
-                    // which move around in memory.
-                    assert(vd->ir.irLocal->byref);
                 } else {
                     Logger::println("nested var:   %s", vd->toChars());
-                    if (vd->ir.irLocal->value)
-                        Logger::cout() << "Pre-existing value: " << *vd->ir.irLocal->value << '\n';
                     assert(!vd->ir.irLocal->value);
                     vd->ir.irLocal->value = gep;
-                    assert(!vd->ir.irLocal->byref);
                 }
 
                 if (global.params.symdebug) {
diff --git a/gen/toir.cpp b/gen/toir.cpp
index 6064b17df9..c348dd1061 100644
--- a/gen/toir.cpp
+++ b/gen/toir.cpp
@@ -583,6 +583,30 @@ DValue* AssignExp::toElem(IRState* p)
         return newlen;
     }
 
+    // Can't just override ConstructExp::toElem because not all TOKconstruct
+    // operations are actually instances of ConstructExp... Long live the DMD
+    // coding style!
+    if (op == TOKconstruct)
+    {
+        if (e1->op == TOKvar)
+        {
+            VarExp* ve = (VarExp*)e1;
+            if (ve->var->storage_class & STCref)
+            {
+                // Note that the variable value is accessed directly (instead
+                // of via getLValue(), which would perform a load from the
+                // uninitialized location), and that rhs is stored as an l-value!
+
+                IrLocal* const local = ve->var->ir.irLocal;
+                assert(local && "ref var must be local and already initialized");
+
+                DValue* rhs = e2->toElem(p);
+                DtoStore(rhs->getLVal(), local->value);
+                return rhs;
+            }
+        }
+    }
+
     Logger::println("performing normal assignment");
 
     DValue* l = e1->toElem(p);
diff --git a/ir/irvar.cpp b/ir/irvar.cpp
index f41f6acb22..e338f12359 100644
--- a/ir/irvar.cpp
+++ b/ir/irvar.cpp
@@ -31,7 +31,6 @@ IrGlobal::IrGlobal(VarDeclaration* v): IrVar(v)
 IrLocal::IrLocal(VarDeclaration* v) : IrVar(v)
 {
     nestedIndex = -1;
-    byref = false;
 }
 
 //////////////////////////////////////////////////////////////////////////////
diff --git a/ir/irvar.h b/ir/irvar.h
index 3d66cc39cd..680c112be0 100644
--- a/ir/irvar.h
+++ b/ir/irvar.h
@@ -28,8 +28,8 @@ struct IrLocal : IrVar
 {
     IrLocal(VarDeclaration* v);
 
-    bool byref;         // Not used for -nested-ctx=array
-    int nestedDepth;    // ditto
+    // Used for hybrid nested context creation.
+    int nestedDepth;
     int nestedIndex;
 };