diff --git a/ctx.cpp b/ctx.cpp
index 43d5771b..df1129b8 100644
--- a/ctx.cpp
+++ b/ctx.cpp
@@ -3617,8 +3617,12 @@ FunctionEmitContext::LaunchInst(llvm::Value *callee,
 
       AssertPos(currentPos, llvm::isa<llvm::Function>(callee));
       std::vector<llvm::Type*> argTypes;
-      for (unsigned int i = 0; i < argVals.size(); i++)
-        argTypes.push_back(argVals[i]->getType());
+
+      llvm::Function *F = llvm::dyn_cast<llvm::Function>(callee);
+      const unsigned int nArgs = F->arg_size();
+      llvm::Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
+      for (; I != E; ++I) 
+        argTypes.push_back(I->getType());
       llvm::Type *st = llvm::StructType::get(*g->ctx, argTypes);
       llvm::StructType *argStructType = static_cast<llvm::StructType *>(st);
       llvm::Value *structSize = g->target->SizeOf(argStructType, bblock);
@@ -3661,6 +3665,13 @@ FunctionEmitContext::LaunchInst(llvm::Value *callee,
         // don't need to do masked store here, I think
         StoreInst(argVals[i], ptr);
       }
+      if (nArgs == argVals.size() + 1) {
+        // copy in the mask
+        llvm::Value *mask = GetFullMask();
+        llvm::Value *ptr = AddElementOffset(argmem, argVals.size(), NULL,
+            "funarg_mask");
+        StoreInst(mask, ptr);
+      }
       BranchInst(if_false);
 
       /**********************/
diff --git a/decl.cpp b/decl.cpp
index 7b321dde..0366dd73 100644
--- a/decl.cpp
+++ b/decl.cpp
@@ -531,12 +531,6 @@ Declarator::InitFromType(const Type *baseType, DeclSpecs *ds) {
         returnType = returnType->ResolveUnboundVariability(Variability::Varying);
 
         bool isTask =     ds && ((ds->typeQualifiers & TYPEQUAL_TASK) != 0);
-
-#if 1   /* evghenii: without this, PTX fails on some examples, like deferred, with #if 0 */
-        if (isTask && g->target->getISA() == Target::NVPTX)
-          ds->typeQualifiers |= TYPEQUAL_UNMASKED;
-#endif
-
         bool isExternC =  ds && (ds->storageClass == SC_EXTERN_C);
         bool isExported = ds && ((ds->typeQualifiers & TYPEQUAL_EXPORT) != 0);
         bool isUnmasked = ds && ((ds->typeQualifiers & TYPEQUAL_UNMASKED) != 0);
diff --git a/func.cpp b/func.cpp
index bff29e32..3923e780 100644
--- a/func.cpp
+++ b/func.cpp
@@ -359,7 +359,7 @@ Function::emitCode(FunctionEmitContext *ctx, llvm::Function *function,
         // entire thing inside code that tests to see if the mask is all
         // on, all off, or mixed.  If this is a simple function, then this
         // isn't worth the code bloat / overhead.
-        bool checkMask = (type->isTask == true && g->target->getISA() != Target::NVPTX) ||
+        bool checkMask = (type->isTask == true) ||
             (
 #if defined(LLVM_3_1)
               (function->hasFnAttr(llvm::Attribute::AlwaysInline) == false)
diff --git a/module.cpp b/module.cpp
index 38853419..4be06df7 100644
--- a/module.cpp
+++ b/module.cpp
@@ -795,6 +795,7 @@ Module::AddFunctionDeclaration(const std::string &name,
 #else // LLVM 3.1 and 3.3+
         function->addFnAttr(llvm::Attribute::AlwaysInline);
 #endif
+    /* evghenii: fails function verification when "if" executed in nvptx target */
     if (functionType->isTask && g->target->getISA() != Target::NVPTX)
         // This also applies transitively to members I think?
 #if defined(LLVM_3_1)