Modified file to provide mapping provided by kernel object to get arguments from

parent instead of getInValAt() function

llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp

@@ -54,9 +54,10 @@ public:
 // calls
 class Kernel {
 public:
-  Kernel(Function* _KF, std::vector<unsigned>* _inArgMap = NULL,
-         unsigned _gridDim = 0, std::vector<Value*>
-         _globalWGSize = std::vector<Value*>(), unsigned _blockDim = 0,
+  Kernel(Function* _KF, std::vector<unsigned> _inArgMap =
+         std::vector<unsigned>(), unsigned _gridDim = 0, std::vector<Value*>
+         _globalWGSize = std::vector<Value*>(),
+         unsigned _blockDim = 0,
          std::vector<Value*> _localWGSize = std::vector<Value*>())
     : KernelFunction(_KF), inArgMap(_inArgMap),
       gridDim(_gridDim), globalWGSize(_globalWGSize),
@@ -69,11 +70,15 @@ public:
   }
 
   Function* KernelFunction;
-  std::vector<unsigned>* inArgMap;
+  std::vector<unsigned> inArgMap;
   unsigned gridDim;
   unsigned blockDim;
   std::vector<Value*> globalWGSize;
   std::vector<Value*> localWGSize;
+
+  std::vector<unsigned> getInArgMap() {
+    return inArgMap;
+  }
 };
 
 // Helper function declarations
@@ -469,7 +474,11 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
   std::vector<OutputPtr> OutputPointers;
   for(unsigned i=0; i<CF->getFunctionType()->getNumParams(); i++) {
 
-    Value* inputVal = getInValueAt(C, i, F_X86, RI);
+    // The kernel object gives us the mapping of arguments from kernel launch
+    // node function (F_X86) to kernel (kernel->KF)
+    Value* inputVal = getArgumentAt(F_X86, kernel->getInArgMap()[i]);
+    DEBUG(errs() << "\tArgument "<< i<< " = "  << *inputVal << "\n");
+
     // input value has been obtained.
     // Check if input is a scalar value or a pointer operand
     // For scalar values such as int, float, etc. the size is simply the size of
@@ -484,10 +493,10 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
 
       Argument* A = getArgumentAt(CF, i);
       if(isOutput == True) {
-        errs() << *A << " is an OUTPUT argument\n";
+        DEBUG(errs() << *A << " is an OUTPUT argument\n");
       }
       if(isInput == True) {
-        errs() << *A << " is an INPUT argument\n";
+        DEBUG(errs() << *A << " is an INPUT argument\n");
       }
 
 
@@ -495,7 +504,7 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
                              Type::getInt8PtrTy(M.getContext()),
                              inputVal->getName()+".i8ptr",
                              RI);
-      Value* inputSize = getInValueAt(C, i+1, F_X86, RI);
+      Value* inputSize = getArgumentAt(F_X86, kernel->getInArgMap()[i+1]);
       assert(inputSize->getType() == Type::getInt64Ty(M.getContext())
              && "Pointer type input must always be followed by size (integer type)");
       Value* setInputArgs[] = {GraphID,
@@ -551,9 +560,9 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
                              };
 
     d_Output = CallInst::Create(llvm_visc_ptx_argument_ptr,
-                                          ArrayRef<Value*>(setOutputArgs, 6),
-                                          "d_output."+CF->getName(),
-                                          RI);
+                                ArrayRef<Value*>(setOutputArgs, 6),
+                                "d_output."+CF->getName(),
+                                RI);
   }
 
   // Enqueue kernel
@@ -601,9 +610,10 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
   // Read all the pointer arguments which had side effects i.e., had out
   // attribute
   for(auto output: OutputPointers) {
-    errs() << "Read: " << *output.d_ptr << "\n";
-    errs() << "\t To: " << *output.h_ptr << "\n";
-    errs() << "\t #bytes: " << *output.bytes << "\n";
+    DEBUG(errs() << "Read: " << *output.d_ptr << "\n");
+    DEBUG(errs() << "\tTo: " << *output.h_ptr << "\n");
+    DEBUG(errs() << "\t#bytes: " << *output.bytes << "\n");
+
     Value* GetOutputArgs[] = {GraphID, output.h_ptr, output.d_ptr, output.bytes};
     CallInst* CI = CallInst::Create(llvm_visc_ptx_getOutput,
                                     ArrayRef<Value*>(GetOutputArgs, 4),
@@ -688,14 +698,13 @@ void CodeGenTraversal::codeGen(DFInternalNode* N) {
     DEBUG(errs() << "Found intermediate node. Getting size parameters.\n");
     // Keep track of the arguments order.
 
-    std::vector<unsigned>* map1 = N->getInArgMap();
-    std::vector<unsigned>* map2 = kernel->getInArgMap();
+    std::vector<unsigned> map1 = N->getInArgMap();
+    std::vector<unsigned> map2 = kernel->getInArgMap();
 
     // The limit is the size of map2, because this is the number of kernel arguments
     for (unsigned i = 0; i < map2.size(); i++) {
       map2[i] = map1[map2[i]];
     }
-    free(map1);
 
   }