From fd17ad236a234533a1b2de58f35575c4034fa6bc Mon Sep 17 00:00:00 2001
From: Evghenii <egaburov@dds.nl>
Date: Tue, 12 Nov 2013 14:05:12 +0100
Subject: [PATCH] export functions are now also generated... next add proper
 CDP calls..

---
 ctx.cpp    | 12 +++++++++---
 func.cpp   | 32 +++++++++++++++++++++-----------
 module.cpp |  1 +
 3 files changed, 31 insertions(+), 14 deletions(-)

diff --git a/ctx.cpp b/ctx.cpp
index 45130b66..8355bbc6 100644
--- a/ctx.cpp
+++ b/ctx.cpp
@@ -3595,7 +3595,7 @@ FunctionEmitContext::LaunchInst(llvm::Value *callee,
     }
     else /* isPTX ==  true */
     {
-      assert(0);  /* must only be called in export */
+      //assert(0);  /* must only be called in export */
     //  assert(g->target->getISA() != Target::NVPTX64);
 
       if (callee == NULL) {
@@ -3606,6 +3606,7 @@ FunctionEmitContext::LaunchInst(llvm::Value *callee,
       launchedTasks = true;
 
       AssertPos(currentPos, llvm::isa<llvm::Function>(callee));
+#if 0
       llvm::Type *argType =
         (llvm::dyn_cast<llvm::Function>(callee))->arg_begin()->getType();
       AssertPos(currentPos, llvm::PointerType::classof(argType));
@@ -3614,6 +3615,7 @@ FunctionEmitContext::LaunchInst(llvm::Value *callee,
       AssertPos(currentPos, llvm::StructType::classof(pt->getElementType()));
       llvm::StructType *argStructType =
         static_cast<llvm::StructType *>(pt->getElementType());
+#endif
 
       llvm::Function *falloc = m->module->getFunction("CUDAAlloc");
       AssertPos(currentPos, falloc != NULL);
@@ -3648,9 +3650,13 @@ FunctionEmitContext::LaunchInst(llvm::Value *callee,
        * this, and other parts are reverse enginered via
        * cpp(clang++ -S -emit-llvm)->IR and then IR(llc -march=cpp)->cpp 
        */
-      for (unsigned int i = 0; i < argVals.size(); ++i) 
+      llvm::Function *func = llvm::dyn_cast<llvm::Function>(callee);
+      llvm::Function::arg_iterator argIter =func->arg_begin();
+      llvm::Function::arg_iterator argIterEnd =func->arg_end();
+      int i = 0;
+      for (; argIter != argIterEnd; argIter++, i++)
       {
-        llvm::Type*           type = argStructType->getElementType(i);
+        llvm::Type*           type = argIter->getType(); //argStructType->getElementType(i);
         llvm::Value* ptr_arg1_addr = AllocaInst(type, "argptr");
         StoreInst(argVals[i], ptr_arg1_addr);
 
diff --git a/func.cpp b/func.cpp
index 6417df7a..e2e8ba84 100644
--- a/func.cpp
+++ b/func.cpp
@@ -354,15 +354,16 @@ Function::emitCode(FunctionEmitContext *ctx, llvm::Function *function,
             Assert(++argIter == function->arg_end());
           }
 
-#if 0
-          llvm::NamedMDNode* annotations =
-            m->module->getOrInsertNamedMetadata("nvvm.annotations");
-          llvm::SmallVector<llvm::Value*, 3> av;
-          av.push_back(function);
-          av.push_back(llvm::MDString::get(*g->ctx, "kernel"));
-          av.push_back(llvm::ConstantInt::get(llvm::IntegerType::get(*g->ctx,32), 1));
-          annotations->addOperand(llvm::MDNode::get(*g->ctx, av)); 
-#endif
+          if (g->target->isPTX() && g->target->getISA() == Target::NVPTX64)
+          {
+            llvm::NamedMDNode* annotations =
+              m->module->getOrInsertNamedMetadata("nvvm.annotations");
+            llvm::SmallVector<llvm::Value*, 3> av;
+            av.push_back(function);
+            av.push_back(llvm::MDString::get(*g->ctx, "kernel"));
+            av.push_back(llvm::ConstantInt::get(llvm::IntegerType::get(*g->ctx,32), 1));
+            annotations->addOperand(llvm::MDNode::get(*g->ctx, av)); 
+          }
         }
     }
     else {
@@ -538,7 +539,6 @@ Function::GenerateIR() {
      /* export function with NVPTX64 target should be emitted host architecture */
 #if 0
     const FunctionType *func_type= CastType<FunctionType>(sym->type);
-    /* exported functions are not supported yet */
     if (g->target->getISA() == Target::NVPTX64 && func_type->isExported)
       return;
 #endif
@@ -567,7 +567,7 @@ Function::GenerateIR() {
         // the application can call it
         const FunctionType *type = CastType<FunctionType>(sym->type);
         Assert(type != NULL);
-        if (type->isExported) { // && g->target->getISA() != Target::NVPTX64) {
+        if (type->isExported) { // && g->target->getISA() != Target::VPTX64) {
             if (!type->isTask) {
                 llvm::FunctionType *ftype = type->LLVMFunctionType(g->ctx, true);
                 llvm::GlobalValue::LinkageTypes linkage = llvm::GlobalValue::ExternalLinkage;
@@ -602,6 +602,16 @@ Function::GenerateIR() {
                             FATAL("Function verificication failed");
                         }
                     }
+                    if (g->target->isPTX() && g->target->getISA() == Target::NVPTX64)
+                    {
+                      llvm::NamedMDNode* annotations =
+                        m->module->getOrInsertNamedMetadata("nvvm.annotations");
+                      llvm::SmallVector<llvm::Value*, 3> av;
+                      av.push_back(appFunction);
+                      av.push_back(llvm::MDString::get(*g->ctx, "kernel"));
+                      av.push_back(llvm::ConstantInt::get(llvm::IntegerType::get(*g->ctx,32), 1));
+                      annotations->addOperand(llvm::MDNode::get(*g->ctx, av)); 
+                    }
                 }
             }
         }
diff --git a/module.cpp b/module.cpp
index def0c6d7..743e3adf 100644
--- a/module.cpp
+++ b/module.cpp
@@ -796,6 +796,7 @@ Module::AddFunctionDeclaration(const std::string &name,
 #else // LLVM 3.1 and 3.3+
         function->addFnAttr(llvm::Attribute::AlwaysInline);
 #endif
+     /* evghenii: on PTX target this must not be used, cause crash, dunno why */
     if (functionType->isTask && g->target->getISA() != Target::NVPTX64)
         // This also applies transitively to members I think?
 #if defined(LLVM_3_1)