convert pointers in function arguments to addrspace(3). still there is poroblem with shared memory. need to figure out which one ..

ctx.cpp

@@ -1890,14 +1890,21 @@ static llvm::Value* lConvertToGenericPtr(FunctionEmitContext *ctx, llvm::Value *
       addressSpace == 3 ? "__cvt_loc2gen" : "__cvt_const2gen");
   std::vector<llvm::Value *> __cvt2gen_args;
   __cvt2gen_args.push_back(value);
+#if 0
   value = ctx->CallInst(__cvt2gen, NULL, __cvt2gen_args, "cvt2gen_call");
+#else
+  value = llvm::CallInst::Create(__cvt2gen, __cvt2gen_args, "cvt2gen_call", ctx->GetCurrentBasicBlock());
+#endif
 
   /* convert i64* to elTy* */
   llvm::PointerType *typeElPtr = llvm::PointerType::get(typeEl, 0);
   value  = ctx->BitCastInst(value, typeElPtr, "cvtLoc2Gen_i642ptr");
 
-  /* add warp offset to the pointer */
-  return lAddWarpOffset(ctx, value);
+  /* add warp offset to the pointer for local memory */
+  if (addressSpace == 3)
+    value = lAddWarpOffset(ctx, value);
+
+  return value;
 }
 
 llvm::Value *
@@ -3394,13 +3401,19 @@ lCalleeArgCount(llvm::Value *callee, const FunctionType *funcType) {
 
 llvm::Value *
 FunctionEmitContext::CallInst(llvm::Value *func, const FunctionType *funcType,
-                              const std::vector<llvm::Value *> &args,
+                              const std::vector<llvm::Value *> &args_in,
                               const char *name) {
     if (func == NULL) {
         AssertPos(currentPos, m->errorCount > 0);
         return NULL;
     }
 
+    std::vector<llvm::Value *> args = args_in;
+    /* NVPTX:
+     * Convert all pointers to addrspace(0) 
+     */
+    for (unsigned int i = 0; i < args.size(); i++)
+      args[i] = lConvertToGenericPtr(this, args[i], currentPos);
     std::vector<llvm::Value *> argVals = args;
     // Most of the time, the mask is passed as the last argument.  this
     // isn't the case for things like intrinsics, builtins, and extern "C"

examples_ptx/deferred/kernels.ispc

@@ -506,9 +506,10 @@ RenderTile(uniform int num_groups_x, uniform int num_groups_y,
     uniform float cameraProj_32 = inputHeader.cameraProj[3][2];
 
     // Light intersection: figure out which lights illuminate this tile.
-#ifdef __NVPTX__
+#if 1 //def __NVPTX__
     uniform int * uniform tileLightIndices = uniform new uniform int [MAX_LIGHTS];
-#else
+#define MALLOC
+#else /* shared memory doesn't full work... why? */
     uniform int tileLightIndices[MAX_LIGHTS];  // Light list for the tile
 #endif
     uniform int numTileLights = 
@@ -532,7 +533,7 @@ RenderTile(uniform int num_groups_x, uniform int num_groups_y,
               cameraProj_00, cameraProj_11, cameraProj_22, cameraProj_32,
               tileLightIndices, numTileLights, visualizeLightCount, 
               framebuffer_r, framebuffer_g, framebuffer_b);
-#ifdef __NVPTX__
+#ifdef MALLOC
     delete tileLightIndices;
 #endif
 }

stmt.cpp

@@ -278,7 +278,7 @@ DeclStmt::EmitCode(FunctionEmitContext *ctx) const {
               int addressSpace;
               if (sym->type->IsConstType())
               {
-#if 0          /* current, addressSpace = 4 generates a compilation fails as it can't be passed as a function arg:S */
+#if 1          /* current, addressSpace = 4 generates a compilation fails as it can't be passed as a function arg:S */
                 addressSpace = 4; /* constant */
 #else
                 addressSpace = 0; /* use global for now */
@@ -333,7 +333,7 @@ DeclStmt::EmitCode(FunctionEmitContext *ctx) const {
               sym->storagePtr =
                 new llvm::GlobalVariable(*m->module, llvmTypeUn,
                     sym->type->IsConstType(),
-                    llvm::GlobalValue::PrivateLinkage, 
+                    llvm::GlobalValue::InternalLinkage, 
                     cinit,
                     llvm::Twine("local_") +
                     llvm::Twine(sym->pos.first_line) +