Muchos fixos. gears kinda works. and cases don't crash.

glorious

9 files changed, 1149 insertions, 778 deletions

diff --git a/src/mesa/pipe/draw/draw_vertex_shader_llvm.c b/src/mesa/pipe/draw/draw_vertex_shader_llvm.c
index 2ff35ace24..d658544f30 100644
--- a/src/mesa/pipe/draw/draw_vertex_shader_llvm.c
+++ b/src/mesa/pipe/draw/draw_vertex_shader_llvm.c
@@ -141,7 +141,8 @@ void draw_vertex_shader_queue_flush_llvm(struct draw_context *draw)
    /* batch execute the shaders on all the vertices */
    ga_llvm_prog_exec(prog, inputs, outputs, consts,
                      draw->vs.queue_nr,
-                     draw->vertex_shader->state->num_inputs);
+                     draw->vertex_shader->state->num_inputs,
+                     draw->vertex_info.num_attribs);
 
    /* FIXME: finish conversion */
    /* dests = outputs */
diff --git a/src/mesa/pipe/llvm/instructions.cpp b/src/mesa/pipe/llvm/instructions.cpp
index 2dd1d1861c..3105c49bcc 100644
--- a/src/mesa/pipe/llvm/instructions.cpp
+++ b/src/mesa/pipe/llvm/instructions.cpp
@@ -3,13 +3,17 @@
 #include <llvm/CallingConv.h>
 #include <llvm/Constants.h>
 #include <llvm/DerivedTypes.h>
+#include <llvm/Function.h>
 #include <llvm/InstrTypes.h>
 #include <llvm/Instructions.h>
 
 using namespace llvm;
 
-Instructions::Instructions(llvm::Module *mod, llvm::BasicBlock *block)
-   :  m_mod(mod), m_block(block), m_idx(0)
+
+Function* makeLitFunction(Module *mod);
+
+Instructions::Instructions(llvm::Module *mod, llvm::Function *func, llvm::BasicBlock *block)
+   :  m_mod(mod), m_func(func), m_block(block), m_idx(0)
 {
    m_floatVecType = VectorType::get(Type::FloatTy, 4);
 
@@ -18,6 +22,8 @@ Instructions::Instructions(llvm::Module *mod, llvm::BasicBlock *block)
    m_llvmPow   = 0;
    m_llvmFloor = 0;
    m_llvmFlog  = 0;
+   m_llvmLit  = 0;
+   m_fmtPtr = 0;
 }
 
 llvm::Value * Instructions::add(llvm::Value *in1, llvm::Value *in2)
@@ -128,7 +134,7 @@ llvm::Value * Instructions::vectorFromVals(llvm::Value *x, llvm::Value *y,
                                m_block);
    if (w)
       res = new InsertElementInst(res, w, unsigned(3),
-                                  name("vecxyw"),
+                                  name("vecxyzw"),
                                   m_block);
    return res;
 }
@@ -159,9 +165,36 @@ llvm::Value *Instructions::callFAbs(llvm::Value *val)
    return call;
 }
 
-llvm::Value * Instructions::lit(llvm::Value *in1)
+llvm::Value * Instructions::lit(llvm::Value *in)
 {
-   return in1;
+#if 1
+   printVector(in);
+   return in;
+
+   ExtractElementInst *x = new ExtractElementInst(in, unsigned(0),
+                                                  name("x"),
+                                                  m_block);
+   
+   ExtractElementInst *y = new ExtractElementInst(in, unsigned(1),
+                                                  name("y"),
+                                                  m_block);
+   
+   ExtractElementInst *w = new ExtractElementInst(in, unsigned(3),
+                                                  name("w"),
+                                                  m_block);
+   return vectorFromVals(ConstantFP::get(Type::FloatTy, APFloat(1.f)),
+                         ConstantFP::get(Type::FloatTy, APFloat(0.f)),
+                         ConstantFP::get(Type::FloatTy, APFloat(0.f)),
+                         ConstantFP::get(Type::FloatTy, APFloat(1.f)));
+#else
+   if (!m_llvmLit) {
+      m_llvmLit = makeLitFunction(m_mod);
+   }
+   CallInst *call = new CallInst(m_llvmLit, in, name("litres"), m_block);
+   //call->setCallingConv(CallingConv::C);
+   call->setTailCall(false);
+   return call;
+#endif
 }
 
 llvm::Value * Instructions::sub(llvm::Value *in1, llvm::Value *in2)
@@ -515,3 +548,357 @@ llvm::Value * Instructions::max(llvm::Value *in1, llvm::Value *in2)
    return vectorFromVals(selx, sely, selz, selw);
 }
 
+
+/*
+  Generated from:
+extern float exp2f(float x);
+extern float log2f(float x);
+
+float4 lit(float4 tmp)
+{
+    tmp.w = (tmp.w < -128.0) ? -128.0f : ((tmp.w > 128.0f) ? 128.f : tmp.w);
+    float4 result;
+    result.x = 1.0;
+    result.y = tmp.x;
+    result.z = (tmp.x > 0) ? exp2f(tmp.w * log2f(tmp.y)) : 0.0;
+    result.w = 1.0;
+    return result;
+}
+with:
+clang --emit-llvm lit.c |llvm-as|opt -std-compile-opts|llvm2cpp -gen-function -for=lit
+
+*/
+Function* makeLitFunction(Module *mod) {
+
+// Type Definitions
+ArrayType* ArrayTy_0 = ArrayType::get(IntegerType::get(8), 27);
+
+PointerType* PointerTy_1 = PointerType::get(ArrayTy_0);
+
+ArrayType* ArrayTy_2 = ArrayType::get(IntegerType::get(8), 28);
+
+PointerType* PointerTy_3 = PointerType::get(ArrayTy_2);
+
+ArrayType* ArrayTy_4 = ArrayType::get(IntegerType::get(8), 8);
+
+PointerType* PointerTy_5 = PointerType::get(ArrayTy_4);
+
+ArrayType* ArrayTy_6 = ArrayType::get(IntegerType::get(8), 33);
+
+PointerType* PointerTy_7 = PointerType::get(ArrayTy_6);
+
+std::vector<const Type*>FuncTy_8_args;
+FuncTy_8_args.push_back(Type::FloatTy);
+FuncTy_8_args.push_back(Type::FloatTy);
+ParamAttrsList *FuncTy_8_PAL = 0;
+FunctionType* FuncTy_8 = FunctionType::get(
+  /*Result=*/Type::FloatTy,
+  /*Params=*/FuncTy_8_args,
+  /*isVarArg=*/false,
+  /*ParamAttrs=*/FuncTy_8_PAL);
+
+std::vector<const Type*>FuncTy_10_args;
+ParamAttrsList *FuncTy_10_PAL = 0;
+FunctionType* FuncTy_10 = FunctionType::get(
+  /*Result=*/IntegerType::get(32),
+  /*Params=*/FuncTy_10_args,
+  /*isVarArg=*/true,
+  /*ParamAttrs=*/FuncTy_10_PAL);
+
+PointerType* PointerTy_9 = PointerType::get(FuncTy_10);
+
+PointerType* PointerTy_11 = PointerType::get(IntegerType::get(8));
+
+PointerType* PointerTy_12 = PointerType::get(FuncTy_8);
+
+VectorType* VectorTy_13 = VectorType::get(Type::FloatTy, 4);
+
+std::vector<const Type*>FuncTy_14_args;
+FuncTy_14_args.push_back(VectorTy_13);
+ParamAttrsList *FuncTy_14_PAL = 0;
+FunctionType* FuncTy_14 = FunctionType::get(
+  /*Result=*/VectorTy_13,
+  /*Params=*/FuncTy_14_args,
+  /*isVarArg=*/false,
+  /*ParamAttrs=*/FuncTy_14_PAL);
+
+
+// Function Declarations
+
+Function* func_approx = new Function(
+  /*Type=*/FuncTy_8,
+  /*Linkage=*/GlobalValue::ExternalLinkage,
+  /*Name=*/"approx", mod); 
+func_approx->setCallingConv(CallingConv::C);
+
+Function* func_printf = mod->getFunction("printf");
+
+Function* func_powf = new Function(
+  /*Type=*/FuncTy_8,
+  /*Linkage=*/GlobalValue::ExternalLinkage,
+  /*Name=*/"powf", mod); // (external, no body)
+func_powf->setCallingConv(CallingConv::C);
+
+Function* func_lit = new Function(
+  /*Type=*/FuncTy_14,
+  /*Linkage=*/GlobalValue::ExternalLinkage,
+  /*Name=*/"lit", mod); 
+func_lit->setCallingConv(CallingConv::C);
+
+// Global Variable Declarations
+
+
+GlobalVariable* gvar_array__str = new GlobalVariable(
+/*Type=*/ArrayTy_0,
+/*isConstant=*/true,
+/*Linkage=*/GlobalValue::InternalLinkage,
+/*Initializer=*/0, // has initializer, specified below
+/*Name=*/".str",
+mod);
+
+GlobalVariable* gvar_array__str1 = new GlobalVariable(
+/*Type=*/ArrayTy_2,
+/*isConstant=*/true,
+/*Linkage=*/GlobalValue::InternalLinkage,
+/*Initializer=*/0, // has initializer, specified below
+/*Name=*/".str1",
+mod);
+
+GlobalVariable* gvar_array__str2 = new GlobalVariable(
+/*Type=*/ArrayTy_4,
+/*isConstant=*/true,
+/*Linkage=*/GlobalValue::InternalLinkage,
+/*Initializer=*/0, // has initializer, specified below
+/*Name=*/".str2",
+mod);
+
+GlobalVariable* gvar_array__str3 = new GlobalVariable(
+/*Type=*/ArrayTy_6,
+/*isConstant=*/true,
+/*Linkage=*/GlobalValue::InternalLinkage,
+/*Initializer=*/0, // has initializer, specified below
+/*Name=*/".str3",
+mod);
+
+// Constant Definitions
+Constant* const_array_15 = ConstantArray::get("After test with '%f' '%f'\x0A", true);
+Constant* const_array_16 = ConstantArray::get("Calling pow with '%f' '%f'\x0A", true);
+Constant* const_array_17 = ConstantArray::get("IN LIT\x0A", true);
+Constant* const_array_18 = ConstantArray::get("About to approx  with '%f' '%f'\x0A", true);
+ConstantFP* const_float_19 = ConstantFP::get(Type::FloatTy, APFloat(-1.280000e+02f));
+ConstantFP* const_float_20 = ConstantFP::get(Type::FloatTy, APFloat(1.280000e+02f));
+std::vector<Constant*> const_ptr_21_indices;
+Constant* const_int32_22 = Constant::getNullValue(IntegerType::get(32));
+const_ptr_21_indices.push_back(const_int32_22);
+const_ptr_21_indices.push_back(const_int32_22);
+Constant* const_ptr_21 = ConstantExpr::getGetElementPtr(gvar_array__str, &const_ptr_21_indices[0], const_ptr_21_indices.size() );
+Constant* const_float_23 = Constant::getNullValue(Type::FloatTy);
+std::vector<Constant*> const_ptr_24_indices;
+const_ptr_24_indices.push_back(const_int32_22);
+const_ptr_24_indices.push_back(const_int32_22);
+Constant* const_ptr_24 = ConstantExpr::getGetElementPtr(gvar_array__str1, &const_ptr_24_indices[0], const_ptr_24_indices.size() );
+std::vector<Constant*> const_ptr_25_indices;
+const_ptr_25_indices.push_back(const_int32_22);
+const_ptr_25_indices.push_back(const_int32_22);
+Constant* const_ptr_25 = ConstantExpr::getGetElementPtr(gvar_array__str2, &const_ptr_25_indices[0], const_ptr_25_indices.size() );
+std::vector<Constant*> const_packed_26_elems;
+ConstantFP* const_float_27 = ConstantFP::get(Type::FloatTy, APFloat(1.000000e+00f));
+const_packed_26_elems.push_back(const_float_27);
+UndefValue* const_float_28 = UndefValue::get(Type::FloatTy);
+const_packed_26_elems.push_back(const_float_28);
+const_packed_26_elems.push_back(const_float_28);
+const_packed_26_elems.push_back(const_float_27);
+Constant* const_packed_26 = ConstantVector::get(VectorTy_13, const_packed_26_elems);
+ConstantInt* const_int32_29 = ConstantInt::get(APInt(32,  "1", 10));
+ConstantInt* const_int32_30 = ConstantInt::get(APInt(32,  "3", 10));
+std::vector<Constant*> const_ptr_31_indices;
+const_ptr_31_indices.push_back(const_int32_22);
+const_ptr_31_indices.push_back(const_int32_22);
+Constant* const_ptr_31 = ConstantExpr::getGetElementPtr(gvar_array__str3, &const_ptr_31_indices[0], const_ptr_31_indices.size() );
+ConstantInt* const_int32_32 = ConstantInt::get(APInt(32,  "2", 10));
+std::vector<Constant*> const_packed_33_elems;
+const_packed_33_elems.push_back(const_float_27);
+const_packed_33_elems.push_back(const_float_23);
+const_packed_33_elems.push_back(const_float_23);
+const_packed_33_elems.push_back(const_float_27);
+Constant* const_packed_33 = ConstantVector::get(VectorTy_13, const_packed_33_elems);
+
+// Global Variable Definitions
+gvar_array__str->setInitializer(const_array_15);
+gvar_array__str1->setInitializer(const_array_16);
+gvar_array__str2->setInitializer(const_array_17);
+gvar_array__str3->setInitializer(const_array_18);
+
+// Function Definitions
+
+// Function: approx (func_approx)
+{
+  Function::arg_iterator args = func_approx->arg_begin();
+  Value* float_a = args++;
+  float_a->setName("a");
+  Value* float_b = args++;
+  float_b->setName("b");
+  
+  BasicBlock* label_entry = new BasicBlock("entry",func_approx,0);
+  
+  // Block entry (label_entry)
+  FCmpInst* int1_cmp = new FCmpInst(FCmpInst::FCMP_OLT, float_b, const_float_19, "cmp", label_entry);
+  SelectInst* float_b_addr_0 = new SelectInst(int1_cmp, const_float_19, float_b, "b.addr.0", label_entry);
+  FCmpInst* int1_cmp3 = new FCmpInst(FCmpInst::FCMP_OGT, float_b_addr_0, const_float_20, "cmp3", label_entry);
+  SelectInst* float_b_addr_1 = new SelectInst(int1_cmp3, const_float_20, float_b_addr_0, "b.addr.1", label_entry);
+  CastInst* double_conv = new FPExtInst(float_a, Type::DoubleTy, "conv", label_entry);
+  CastInst* double_conv8 = new FPExtInst(float_b_addr_1, Type::DoubleTy, "conv8", label_entry);
+  std::vector<Value*> int32_call_params;
+  int32_call_params.push_back(const_ptr_21);
+  int32_call_params.push_back(double_conv);
+  int32_call_params.push_back(double_conv8);
+  CallInst* int32_call = new CallInst(func_printf, int32_call_params.begin(), int32_call_params.end(), "call", label_entry);
+  int32_call->setCallingConv(CallingConv::C);
+  int32_call->setTailCall(true);
+  FCmpInst* int1_cmp11 = new FCmpInst(FCmpInst::FCMP_OLT, float_a, const_float_23, "cmp11", label_entry);
+  SelectInst* float_a_addr_0 = new SelectInst(int1_cmp11, const_float_23, float_a, "a.addr.0", label_entry);
+  CastInst* double_conv16 = new FPExtInst(float_a_addr_0, Type::DoubleTy, "conv16", label_entry);
+  std::vector<Value*> int32_call19_params;
+  int32_call19_params.push_back(const_ptr_24);
+  int32_call19_params.push_back(double_conv16);
+  int32_call19_params.push_back(double_conv8);
+  CallInst* int32_call19 = new CallInst(func_printf, int32_call19_params.begin(), int32_call19_params.end(), "call19", label_entry);
+  int32_call19->setCallingConv(CallingConv::C);
+  int32_call19->setTailCall(true);
+  std::vector<Value*> float_call22_params;
+  float_call22_params.push_back(float_a_addr_0);
+  float_call22_params.push_back(float_b_addr_1);
+  CallInst* float_call22 = new CallInst(func_powf, float_call22_params.begin(), float_call22_params.end(), "call22", label_entry);
+  float_call22->setCallingConv(CallingConv::C);
+  float_call22->setTailCall(true);
+  new ReturnInst(float_call22, label_entry);
+  
+}
+
+// Function: lit (func_lit)
+{
+  Function::arg_iterator args = func_lit->arg_begin();
+  Value* packed_tmp = args++;
+  packed_tmp->setName("tmp");
+  
+  BasicBlock* label_entry_35 = new BasicBlock("entry",func_lit,0);
+  BasicBlock* label_ifthen = new BasicBlock("ifthen",func_lit,0);
+  BasicBlock* label_UnifiedReturnBlock = new BasicBlock("UnifiedReturnBlock",func_lit,0);
+  
+  // Block entry (label_entry_35)
+  CallInst* int32_call_36 = new CallInst(func_printf, const_ptr_25, "call", label_entry_35);
+  int32_call_36->setCallingConv(CallingConv::C);
+  int32_call_36->setTailCall(true);
+  ExtractElementInst* float_tmp7 = new ExtractElementInst(packed_tmp, const_int32_22, "tmp7", label_entry_35);
+  FCmpInst* int1_cmp_37 = new FCmpInst(FCmpInst::FCMP_OGT, float_tmp7, const_float_23, "cmp", label_entry_35);
+  new BranchInst(label_ifthen, label_UnifiedReturnBlock, int1_cmp_37, label_entry_35);
+  
+  // Block ifthen (label_ifthen)
+  InsertElementInst* packed_tmp12 = new InsertElementInst(const_packed_26, float_tmp7, const_int32_29, "tmp12", label_ifthen);
+  ExtractElementInst* float_tmp14 = new ExtractElementInst(packed_tmp, const_int32_29, "tmp14", label_ifthen);
+  CastInst* double_conv15 = new FPExtInst(float_tmp14, Type::DoubleTy, "conv15", label_ifthen);
+  ExtractElementInst* float_tmp17 = new ExtractElementInst(packed_tmp, const_int32_30, "tmp17", label_ifthen);
+  CastInst* double_conv18 = new FPExtInst(float_tmp17, Type::DoubleTy, "conv18", label_ifthen);
+  std::vector<Value*> int32_call19_39_params;
+  int32_call19_39_params.push_back(const_ptr_31);
+  int32_call19_39_params.push_back(double_conv15);
+  int32_call19_39_params.push_back(double_conv18);
+  CallInst* int32_call19_39 = new CallInst(func_printf, int32_call19_39_params.begin(), int32_call19_39_params.end(), "call19", label_ifthen);
+  int32_call19_39->setCallingConv(CallingConv::C);
+  int32_call19_39->setTailCall(true);
+  FCmpInst* int1_cmp_i = new FCmpInst(FCmpInst::FCMP_OLT, float_tmp17, const_float_19, "cmp.i", label_ifthen);
+  SelectInst* float_b_addr_0_i = new SelectInst(int1_cmp_i, const_float_19, float_tmp17, "b.addr.0.i", label_ifthen);
+  FCmpInst* int1_cmp3_i = new FCmpInst(FCmpInst::FCMP_OGT, float_b_addr_0_i, const_float_20, "cmp3.i", label_ifthen);
+  SelectInst* float_b_addr_1_i = new SelectInst(int1_cmp3_i, const_float_20, float_b_addr_0_i, "b.addr.1.i", label_ifthen);
+  CastInst* double_conv8_i = new FPExtInst(float_b_addr_1_i, Type::DoubleTy, "conv8.i", label_ifthen);
+  std::vector<Value*> int32_call_i_params;
+  int32_call_i_params.push_back(const_ptr_21);
+  int32_call_i_params.push_back(double_conv15);
+  int32_call_i_params.push_back(double_conv8_i);
+  CallInst* int32_call_i = new CallInst(func_printf, int32_call_i_params.begin(), int32_call_i_params.end(), "call.i", label_ifthen);
+  int32_call_i->setCallingConv(CallingConv::C);
+  int32_call_i->setTailCall(true);
+  FCmpInst* int1_cmp11_i = new FCmpInst(FCmpInst::FCMP_OLT, float_tmp14, const_float_23, "cmp11.i", label_ifthen);
+  SelectInst* float_a_addr_0_i = new SelectInst(int1_cmp11_i, const_float_23, float_tmp14, "a.addr.0.i", label_ifthen);
+  CastInst* double_conv16_i = new FPExtInst(float_a_addr_0_i, Type::DoubleTy, "conv16.i", label_ifthen);
+  std::vector<Value*> int32_call19_i_params;
+  int32_call19_i_params.push_back(const_ptr_24);
+  int32_call19_i_params.push_back(double_conv16_i);
+  int32_call19_i_params.push_back(double_conv8_i);
+  CallInst* int32_call19_i = new CallInst(func_printf, int32_call19_i_params.begin(), int32_call19_i_params.end(), "call19.i", label_ifthen);
+  int32_call19_i->setCallingConv(CallingConv::C);
+  int32_call19_i->setTailCall(true);
+  std::vector<Value*> float_call22_i_params;
+  float_call22_i_params.push_back(float_a_addr_0_i);
+  float_call22_i_params.push_back(float_b_addr_1_i);
+  CallInst* float_call22_i = new CallInst(func_powf, float_call22_i_params.begin(), float_call22_i_params.end(), "call22.i", label_ifthen);
+  float_call22_i->setCallingConv(CallingConv::C);
+  float_call22_i->setTailCall(true);
+  InsertElementInst* packed_tmp26 = new InsertElementInst(packed_tmp12, float_call22_i, const_int32_32, "tmp26", label_ifthen);
+  new ReturnInst(packed_tmp26, label_ifthen);
+  
+  // Block UnifiedReturnBlock (label_UnifiedReturnBlock)
+  new ReturnInst(const_packed_33, label_UnifiedReturnBlock);
+  
+}
+
+return func_lit;
+
+}
+
+void Instructions::printVector(llvm::Value *val)
+{
+   static const char *frmt = "Vector is [%f, %f, %f, %f]\x0A";
+
+   if (!m_fmtPtr) {
+      Constant *format = ConstantArray::get(frmt, true);
+      ArrayType *arrayTy = ArrayType::get(IntegerType::get(8), strlen(frmt) + 1);
+      GlobalVariable* globalFormat = new GlobalVariable(
+         /*Type=*/arrayTy,
+         /*isConstant=*/true,
+         /*Linkage=*/GlobalValue::InternalLinkage,
+         /*Initializer=*/0, // has initializer, specified below
+         /*Name=*/name(".str"),
+         m_mod);
+      globalFormat->setInitializer(format);
+
+      Constant* const_int0 = Constant::getNullValue(IntegerType::get(32));
+      std::vector<Constant*> const_ptr_21_indices;
+      const_ptr_21_indices.push_back(const_int0);
+      const_ptr_21_indices.push_back(const_int0);
+      m_fmtPtr = ConstantExpr::getGetElementPtr(globalFormat,
+                                                &const_ptr_21_indices[0], const_ptr_21_indices.size());
+   }
+
+   Function *func_printf = m_mod->getFunction("printf");
+   ExtractElementInst *x = new ExtractElementInst(val, unsigned(0),
+                                                  name("extractx"),
+                                                  m_block);
+   ExtractElementInst *y = new ExtractElementInst(val, unsigned(1),
+                                                  name("extracty"),
+                                                  m_block);
+   ExtractElementInst *z = new ExtractElementInst(val, unsigned(2),
+                                                  name("extractz"),
+                                                  m_block);
+   ExtractElementInst *w = new ExtractElementInst(val, unsigned(3),
+                                                  name("extractw"),
+                                                  m_block);
+   CastInst *dx = new FPExtInst(x, Type::DoubleTy, name("dx"), m_block);
+   CastInst *dy = new FPExtInst(y, Type::DoubleTy, name("dy"), m_block);
+   CastInst *dz = new FPExtInst(z, Type::DoubleTy, name("dz"), m_block);
+   CastInst *dw = new FPExtInst(w, Type::DoubleTy, name("dw"), m_block);
+   std::vector<Value*> params;
+   params.push_back(m_fmtPtr);
+   params.push_back(dx);
+   params.push_back(dy);
+   params.push_back(dz);
+   params.push_back(dw);
+   CallInst* call = new CallInst(func_printf, params.begin(), params.end(), "printf", m_block);
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(true);
+}
+
+llvm::Value * Instructions::swizzleWrite(llvm::Value *in, int mask)
+{
+    
+}
diff --git a/src/mesa/pipe/llvm/instructions.h b/src/mesa/pipe/llvm/instructions.h
index dd1e76728a..bb71cc8520 100644
--- a/src/mesa/pipe/llvm/instructions.h
+++ b/src/mesa/pipe/llvm/instructions.h
@@ -7,12 +7,13 @@
 
 namespace llvm {
    class VectorType;
+   class Function;
 }
 
 class Instructions
 {
 public:
-   Instructions(llvm::Module *mod, llvm::BasicBlock *block);
+   Instructions(llvm::Module *mod, llvm::Function *func, llvm::BasicBlock *block);
 
    llvm::Value *add(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *dp3(llvm::Value *in1, llvm::Value *in2);
@@ -33,6 +34,9 @@ public:
    llvm::Value *rcp(llvm::Value *in);
    llvm::Value *rsq(llvm::Value *in);
    llvm::Value *sub(llvm::Value *in1, llvm::Value *in2);
+   llvm::Value *swizzleWrite(llvm::Value *in, int mask);
+
+   void printVector(llvm::Value *val);
 private:
    const char *name(const char *prefix);
 
@@ -46,6 +50,7 @@ private:
                                llvm::Value *z, llvm::Value *w=0);
 private:
    llvm::Module *m_mod;
+   llvm::Function *m_func;
    char        m_name[32];
    llvm::BasicBlock *m_block;
    int               m_idx;
@@ -57,6 +62,9 @@ private:
    llvm::Function   *m_llvmPow;
    llvm::Function   *m_llvmFloor;
    llvm::Function   *m_llvmFlog;
+   llvm::Function   *m_llvmLit;
+
+   llvm::Constant   *m_fmtPtr;
 };
 
 #endif
diff --git a/src/mesa/pipe/llvm/llvm_base_shader.cpp b/src/mesa/pipe/llvm/llvm_base_shader.cpp
index 58dd57f6b9..f1854b5c37 100644
--- a/src/mesa/pipe/llvm/llvm_base_shader.cpp
+++ b/src/mesa/pipe/llvm/llvm_base_shader.cpp
@@ -6,59 +6,47 @@ Module* createBaseShader() {
   Module* mod = new Module("Shader");
   
   // Type Definitions
-  ArrayType* ArrayTy_0 = ArrayType::get(IntegerType::get(8), 20);
-
-  ArrayType* ArrayTy_2 = ArrayType::get(IntegerType::get(8), 31);
+  ArrayType* ArrayTy_0 = ArrayType::get(IntegerType::get(8), 24);
   
-  std::vector<const Type*>FuncTy_4_args;
-  VectorType* VectorTy_7 = VectorType::get(Type::FloatTy, 4);
+  PointerType* PointerTy_1 = PointerType::get(ArrayTy_0);
   
-  ArrayType* ArrayTy_6 = ArrayType::get(VectorTy_7, 16);
+  ArrayType* ArrayTy_2 = ArrayType::get(IntegerType::get(8), 20);
   
-  PointerType* PointerTy_5 = PointerType::get(ArrayTy_6);
+  PointerType* PointerTy_3 = PointerType::get(ArrayTy_2);
   
-  FuncTy_4_args.push_back(PointerTy_5);
-  ArrayType* ArrayTy_10 = ArrayType::get(Type::FloatTy, 4);
+  ArrayType* ArrayTy_4 = ArrayType::get(IntegerType::get(8), 68);
   
-  ArrayType* ArrayTy_9 = ArrayType::get(ArrayTy_10, 16);
+  PointerType* PointerTy_5 = PointerType::get(ArrayTy_4);
   
-  PointerType* PointerTy_8 = PointerType::get(ArrayTy_9);
+  ArrayType* ArrayTy_6 = ArrayType::get(IntegerType::get(8), 33);
   
-  FuncTy_4_args.push_back(PointerTy_8);
-  FuncTy_4_args.push_back(IntegerType::get(32));
-  FuncTy_4_args.push_back(IntegerType::get(32));
-  ParamAttrsList *FuncTy_4_PAL = 0;
-  FunctionType* FuncTy_4 = FunctionType::get(
-    /*Result=*/Type::VoidTy,
-    /*Params=*/FuncTy_4_args,
-    /*isVarArg=*/false,
-    /*ParamAttrs=*/FuncTy_4_PAL);
+  PointerType* PointerTy_7 = PointerType::get(ArrayTy_6);
   
-  PointerType* PointerTy_12 = PointerType::get(VectorTy_7);
+  std::vector<const Type*>FuncTy_8_args;
+  VectorType* VectorTy_11 = VectorType::get(Type::FloatTy, 4);
   
-  std::vector<const Type*>FuncTy_13_args;
-  FuncTy_13_args.push_back(PointerTy_12);
-  PointerType* PointerTy_14 = PointerType::get(ArrayTy_10);
+  ArrayType* ArrayTy_10 = ArrayType::get(VectorTy_11, 16);
   
-  FuncTy_13_args.push_back(PointerTy_14);
-  FuncTy_13_args.push_back(IntegerType::get(32));
-  ParamAttrsList *FuncTy_13_PAL = 0;
-  FunctionType* FuncTy_13 = FunctionType::get(
-    /*Result=*/Type::VoidTy,
-    /*Params=*/FuncTy_13_args,
-    /*isVarArg=*/false,
-    /*ParamAttrs=*/FuncTy_13_PAL);
-  
-  std::vector<const Type*>FuncTy_15_args;
-  FuncTy_15_args.push_back(PointerTy_14);
-  FuncTy_15_args.push_back(PointerTy_12);
-  FuncTy_15_args.push_back(IntegerType::get(32));
-  ParamAttrsList *FuncTy_15_PAL = 0;
-  FunctionType* FuncTy_15 = FunctionType::get(
+  PointerType* PointerTy_9 = PointerType::get(ArrayTy_10);
+  
+  FuncTy_8_args.push_back(PointerTy_9);
+  ArrayType* ArrayTy_14 = ArrayType::get(Type::FloatTy, 4);
+  
+  ArrayType* ArrayTy_13 = ArrayType::get(ArrayTy_14, 16);
+  
+  PointerType* PointerTy_12 = PointerType::get(ArrayTy_13);
+  
+  FuncTy_8_args.push_back(PointerTy_12);
+  FuncTy_8_args.push_back(IntegerType::get(32));
+  FuncTy_8_args.push_back(IntegerType::get(32));
+  ParamAttrsList *FuncTy_8_PAL = 0;
+  FunctionType* FuncTy_8 = FunctionType::get(
     /*Result=*/Type::VoidTy,
-    /*Params=*/FuncTy_15_args,
+    /*Params=*/FuncTy_8_args,
     /*isVarArg=*/false,
-    /*ParamAttrs=*/FuncTy_15_PAL);
+    /*ParamAttrs=*/FuncTy_8_PAL);
+  
+  PointerType* PointerTy_15 = PointerType::get(Type::FloatTy);
   
   std::vector<const Type*>FuncTy_17_args;
   ParamAttrsList *FuncTy_17_PAL = 0;
@@ -68,25 +56,44 @@ Module* createBaseShader() {
     /*isVarArg=*/true,
     /*ParamAttrs=*/FuncTy_17_PAL);
   
-  std::vector<const Type*>FuncTy_19_args;
-  FuncTy_19_args.push_back(PointerTy_8);
-  FuncTy_19_args.push_back(PointerTy_8);
-  FuncTy_19_args.push_back(PointerTy_14);
-  FuncTy_19_args.push_back(IntegerType::get(32));
-  FuncTy_19_args.push_back(IntegerType::get(32));
-  ParamAttrsList *FuncTy_19_PAL = 0;
-  FunctionType* FuncTy_19 = FunctionType::get(
+  PointerType* PointerTy_16 = PointerType::get(FuncTy_17);
+  
+  PointerType* PointerTy_18 = PointerType::get(IntegerType::get(8));
+  
+  PointerType* PointerTy_19 = PointerType::get(VectorTy_11);
+  
+  std::vector<const Type*>FuncTy_20_args;
+  FuncTy_20_args.push_back(PointerTy_19);
+  PointerType* PointerTy_21 = PointerType::get(ArrayTy_14);
+  
+  FuncTy_20_args.push_back(PointerTy_21);
+  FuncTy_20_args.push_back(IntegerType::get(32));
+  ParamAttrsList *FuncTy_20_PAL = 0;
+  FunctionType* FuncTy_20 = FunctionType::get(
     /*Result=*/Type::VoidTy,
-    /*Params=*/FuncTy_19_args,
+    /*Params=*/FuncTy_20_args,
     /*isVarArg=*/false,
-    /*ParamAttrs=*/FuncTy_19_PAL);
-  
-  ArrayType* ArrayTy_21 = ArrayType::get(ArrayTy_6, 2048);
+    /*ParamAttrs=*/FuncTy_20_PAL);
+  
+  std::vector<const Type*>FuncTy_22_args;
+  FuncTy_22_args.push_back(PointerTy_21);
+  FuncTy_22_args.push_back(PointerTy_19);
+  FuncTy_22_args.push_back(IntegerType::get(32));
+  ParamAttrsList *FuncTy_22_PAL = 0;
+  FunctionType* FuncTy_22 = FunctionType::get(
+    /*Result=*/Type::VoidTy,
+    /*Params=*/FuncTy_22_args,
+    /*isVarArg=*/false,
+    /*ParamAttrs=*/FuncTy_22_PAL);
   
   std::vector<const Type*>FuncTy_23_args;
   FuncTy_23_args.push_back(PointerTy_12);
   FuncTy_23_args.push_back(PointerTy_12);
-  FuncTy_23_args.push_back(PointerTy_12);
+  FuncTy_23_args.push_back(PointerTy_21);
+  FuncTy_23_args.push_back(IntegerType::get(32));
+  FuncTy_23_args.push_back(IntegerType::get(32));
+  FuncTy_23_args.push_back(IntegerType::get(32));
+  FuncTy_23_args.push_back(IntegerType::get(32));
   ParamAttrsList *FuncTy_23_PAL = 0;
   FunctionType* FuncTy_23 = FunctionType::get(
     /*Result=*/Type::VoidTy,
@@ -94,41 +101,62 @@ Module* createBaseShader() {
     /*isVarArg=*/false,
     /*ParamAttrs=*/FuncTy_23_PAL);
   
+  ArrayType* ArrayTy_25 = ArrayType::get(ArrayTy_10, 2048);
+  
+  PointerType* PointerTy_24 = PointerType::get(ArrayTy_25);
+  
+  ArrayType* ArrayTy_27 = ArrayType::get(VectorTy_11, 32);
+  
+  PointerType* PointerTy_26 = PointerType::get(ArrayTy_27);
+  
+  std::vector<const Type*>FuncTy_29_args;
+  FuncTy_29_args.push_back(PointerTy_19);
+  FuncTy_29_args.push_back(PointerTy_19);
+  FuncTy_29_args.push_back(PointerTy_19);
+  ParamAttrsList *FuncTy_29_PAL = 0;
+  FunctionType* FuncTy_29 = FunctionType::get(
+    /*Result=*/Type::VoidTy,
+    /*Params=*/FuncTy_29_args,
+    /*isVarArg=*/false,
+    /*ParamAttrs=*/FuncTy_29_PAL);
+  
+  PointerType* PointerTy_28 = PointerType::get(FuncTy_29);
+  
   
   // Function Declarations
   
   Function* func_from_array = new Function(
-    /*Type=*/FuncTy_4,
+    /*Type=*/FuncTy_8,
     /*Linkage=*/GlobalValue::ExternalLinkage,
     /*Name=*/"from_array", mod); 
   func_from_array->setCallingConv(CallingConv::C);
   
+  Function* func_printf = new Function(
+    /*Type=*/FuncTy_17,
+    /*Linkage=*/GlobalValue::ExternalLinkage,
+    /*Name=*/"printf", mod); // (external, no body)
+  func_printf->setCallingConv(CallingConv::C);
+  
   Function* func_from_consts = new Function(
-    /*Type=*/FuncTy_13,
+    /*Type=*/FuncTy_20,
     /*Linkage=*/GlobalValue::ExternalLinkage,
     /*Name=*/"from_consts", mod); 
   func_from_consts->setCallingConv(CallingConv::C);
   
   Function* func_to_array = new Function(
-    /*Type=*/FuncTy_15,
+    /*Type=*/FuncTy_22,
     /*Linkage=*/GlobalValue::ExternalLinkage,
     /*Name=*/"to_array", mod); 
   func_to_array->setCallingConv(CallingConv::C);
   
-  Function* func_printf = new Function(
-    /*Type=*/FuncTy_17,
-    /*Linkage=*/GlobalValue::ExternalLinkage,
-    /*Name=*/"printf", mod); // (external, no body)
-  func_printf->setCallingConv(CallingConv::C);
-  
   Function* func_run_vertex_shader = new Function(
-    /*Type=*/FuncTy_19,
+    /*Type=*/FuncTy_23,
     /*Linkage=*/GlobalValue::ExternalLinkage,
     /*Name=*/"run_vertex_shader", mod); 
   func_run_vertex_shader->setCallingConv(CallingConv::C);
   
   Function* func_execute_shader = new Function(
-    /*Type=*/FuncTy_23,
+    /*Type=*/FuncTy_29,
     /*Linkage=*/GlobalValue::ExternalLinkage,
     /*Name=*/"execute_shader", mod); // (external, no body)
   func_execute_shader->setCallingConv(CallingConv::C);
@@ -145,34 +173,89 @@ Module* createBaseShader() {
   mod);
   
   GlobalVariable* gvar_array__str1 = new GlobalVariable(
-  /*Type=*/ArrayTy_2,
+  /*Type=*/ArrayTy_0,
   /*isConstant=*/true,
   /*Linkage=*/GlobalValue::InternalLinkage,
   /*Initializer=*/0, // has initializer, specified below
   /*Name=*/".str1",
   mod);
   
+  GlobalVariable* gvar_array__str2 = new GlobalVariable(
+  /*Type=*/ArrayTy_2,
+  /*isConstant=*/true,
+  /*Linkage=*/GlobalValue::InternalLinkage,
+  /*Initializer=*/0, // has initializer, specified below
+  /*Name=*/".str2",
+  mod);
+  
+  GlobalVariable* gvar_array__str3 = new GlobalVariable(
+  /*Type=*/ArrayTy_4,
+  /*isConstant=*/true,
+  /*Linkage=*/GlobalValue::InternalLinkage,
+  /*Initializer=*/0, // has initializer, specified below
+  /*Name=*/".str3",
+  mod);
+  
+  GlobalVariable* gvar_array__str4 = new GlobalVariable(
+  /*Type=*/ArrayTy_6,
+  /*isConstant=*/true,
+  /*Linkage=*/GlobalValue::InternalLinkage,
+  /*Initializer=*/0, // has initializer, specified below
+  /*Name=*/".str4",
+  mod);
+  
+  GlobalVariable* gvar_array__str5 = new GlobalVariable(
+  /*Type=*/ArrayTy_0,
+  /*isConstant=*/true,
+  /*Linkage=*/GlobalValue::InternalLinkage,
+  /*Initializer=*/0, // has initializer, specified below
+  /*Name=*/".str5",
+  mod);
+  
   // Constant Definitions
-  Constant* const_array_24 = ConstantArray::get("DEST = %f %f %f %f\x0A", true);
-  Constant* const_array_25 = ConstantArray::get("XXXXXXXXXXX run_vertex_shader\x0A", true);
-  Constant* const_int32_26 = Constant::getNullValue(IntegerType::get(32));
-  UndefValue* const_packed_27 = UndefValue::get(VectorTy_7);
-  ConstantInt* const_int32_28 = ConstantInt::get(APInt(32,  "1", 10));
-  ConstantInt* const_int32_29 = ConstantInt::get(APInt(32,  "2", 10));
-  ConstantInt* const_int32_30 = ConstantInt::get(APInt(32,  "3", 10));
-  std::vector<Constant*> const_ptr_31_indices;
-  const_ptr_31_indices.push_back(const_int32_26);
-  const_ptr_31_indices.push_back(const_int32_26);
-  Constant* const_ptr_31 = ConstantExpr::getGetElementPtr(gvar_array__str, &const_ptr_31_indices[0], const_ptr_31_indices.size() );
-  std::vector<Constant*> const_ptr_32_indices;
-  const_ptr_32_indices.push_back(const_int32_26);
-  const_ptr_32_indices.push_back(const_int32_26);
-  Constant* const_ptr_32 = ConstantExpr::getGetElementPtr(gvar_array__str1, &const_ptr_32_indices[0], const_ptr_32_indices.size() );
-  ConstantInt* const_int32_33 = ConstantInt::get(APInt(32,  "4", 10));
+  Constant* const_array_30 = ConstantArray::get("FRA(%d %d) %f %f %f %f\x0A", true);
+  Constant* const_array_31 = ConstantArray::get("VCONST(%d) %f %f %f %f\x0A", true);
+  Constant* const_array_32 = ConstantArray::get("DEST = %f %f %f %f\x0A", true);
+  Constant* const_array_33 = ConstantArray::get("XXX LLVM run_vertex_shader vertices = %d, inputs = %d, consts = %d\x0A", true);
+  Constant* const_array_34 = ConstantArray::get("               after conversion\x0A", true);
+  Constant* const_array_35 = ConstantArray::get("after executing shader\x0A", true);
+  Constant* const_int32_36 = Constant::getNullValue(IntegerType::get(32));
+  UndefValue* const_packed_37 = UndefValue::get(VectorTy_11);
+  ConstantInt* const_int32_38 = ConstantInt::get(APInt(32,  "1", 10));
+  ConstantInt* const_int32_39 = ConstantInt::get(APInt(32,  "2", 10));
+  ConstantInt* const_int32_40 = ConstantInt::get(APInt(32,  "3", 10));
+  std::vector<Constant*> const_ptr_41_indices;
+  const_ptr_41_indices.push_back(const_int32_36);
+  const_ptr_41_indices.push_back(const_int32_36);
+  Constant* const_ptr_41 = ConstantExpr::getGetElementPtr(gvar_array__str, &const_ptr_41_indices[0], const_ptr_41_indices.size() );
+  std::vector<Constant*> const_ptr_42_indices;
+  const_ptr_42_indices.push_back(const_int32_36);
+  const_ptr_42_indices.push_back(const_int32_36);
+  Constant* const_ptr_42 = ConstantExpr::getGetElementPtr(gvar_array__str1, &const_ptr_42_indices[0], const_ptr_42_indices.size() );
+  std::vector<Constant*> const_ptr_43_indices;
+  const_ptr_43_indices.push_back(const_int32_36);
+  const_ptr_43_indices.push_back(const_int32_36);
+  Constant* const_ptr_43 = ConstantExpr::getGetElementPtr(gvar_array__str2, &const_ptr_43_indices[0], const_ptr_43_indices.size() );
+  std::vector<Constant*> const_ptr_44_indices;
+  const_ptr_44_indices.push_back(const_int32_36);
+  const_ptr_44_indices.push_back(const_int32_36);
+  Constant* const_ptr_44 = ConstantExpr::getGetElementPtr(gvar_array__str3, &const_ptr_44_indices[0], const_ptr_44_indices.size() );
+  std::vector<Constant*> const_ptr_45_indices;
+  const_ptr_45_indices.push_back(const_int32_36);
+  const_ptr_45_indices.push_back(const_int32_36);
+  Constant* const_ptr_45 = ConstantExpr::getGetElementPtr(gvar_array__str4, &const_ptr_45_indices[0], const_ptr_45_indices.size() );
+  std::vector<Constant*> const_ptr_46_indices;
+  const_ptr_46_indices.push_back(const_int32_36);
+  const_ptr_46_indices.push_back(const_int32_36);
+  Constant* const_ptr_46 = ConstantExpr::getGetElementPtr(gvar_array__str5, &const_ptr_46_indices[0], const_ptr_46_indices.size() );
   
   // Global Variable Definitions
-  gvar_array__str->setInitializer(const_array_24);
-  gvar_array__str1->setInitializer(const_array_25);
+  gvar_array__str->setInitializer(const_array_30);
+  gvar_array__str1->setInitializer(const_array_31);
+  gvar_array__str2->setInitializer(const_array_32);
+  gvar_array__str3->setInitializer(const_array_33);
+  gvar_array__str4->setInitializer(const_array_34);
+  gvar_array__str5->setInitializer(const_array_35);
   
   // Function Definitions
   
@@ -190,158 +273,187 @@ Module* createBaseShader() {
     
     BasicBlock* label_entry = new BasicBlock("entry",func_from_array,0);
     BasicBlock* label_forbody6 = new BasicBlock("forbody6",func_from_array,0);
-    BasicBlock* label_forinc57 = new BasicBlock("forinc57",func_from_array,0);
-    BasicBlock* label_afterfor60 = new BasicBlock("afterfor60",func_from_array,0);
+    BasicBlock* label_forinc70 = new BasicBlock("forinc70",func_from_array,0);
+    BasicBlock* label_afterfor73 = new BasicBlock("afterfor73",func_from_array,0);
     
     // Block entry (label_entry)
-    ICmpInst* int1_cmp = new ICmpInst(ICmpInst::ICMP_SGT, int32_count, const_int32_26, "cmp", label_entry);
-    ICmpInst* int1_cmp5 = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_attribs, const_int32_26, "cmp5", label_entry);
+    ICmpInst* int1_cmp = new ICmpInst(ICmpInst::ICMP_SGT, int32_count, const_int32_36, "cmp", label_entry);
+    ICmpInst* int1_cmp5 = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_attribs, const_int32_36, "cmp5", label_entry);
     BinaryOperator* int1_bothcond = BinaryOperator::create(Instruction::And, int1_cmp, int1_cmp5, "bothcond", label_entry);
-    new BranchInst(label_forbody6, label_afterfor60, int1_bothcond, label_entry);
+    new BranchInst(label_forbody6, label_afterfor73, int1_bothcond, label_entry);
     
     // Block forbody6 (label_forbody6)
-    Argument* fwdref_35 = new Argument(IntegerType::get(32));
-    Argument* fwdref_36 = new Argument(IntegerType::get(32));
-    PHINode* int32_i_013_0 = new PHINode(IntegerType::get(32), "i.013.0", label_forbody6);
-    int32_i_013_0->reserveOperandSpace(3);
-    int32_i_013_0->addIncoming(const_int32_26, label_entry);
-    int32_i_013_0->addIncoming(fwdref_35, label_forinc57);
-    int32_i_013_0->addIncoming(fwdref_36, label_forbody6);
-    
-    Argument* fwdref_37 = new Argument(IntegerType::get(32));
+    Argument* fwdref_48 = new Argument(IntegerType::get(32));
+    Argument* fwdref_49 = new Argument(IntegerType::get(32));
+    PHINode* int32_i_014_0 = new PHINode(IntegerType::get(32), "i.014.0", label_forbody6);
+    int32_i_014_0->reserveOperandSpace(3);
+    int32_i_014_0->addIncoming(const_int32_36, label_entry);
+    int32_i_014_0->addIncoming(fwdref_48, label_forinc70);
+    int32_i_014_0->addIncoming(fwdref_49, label_forbody6);
+    
+    Argument* fwdref_50 = new Argument(IntegerType::get(32));
     PHINode* int32_j_03_0 = new PHINode(IntegerType::get(32), "j.03.0", label_forbody6);
     int32_j_03_0->reserveOperandSpace(3);
-    int32_j_03_0->addIncoming(fwdref_37, label_forbody6);
-    int32_j_03_0->addIncoming(const_int32_26, label_forinc57);
-    int32_j_03_0->addIncoming(const_int32_26, label_entry);
+    int32_j_03_0->addIncoming(fwdref_50, label_forbody6);
+    int32_j_03_0->addIncoming(const_int32_36, label_forinc70);
+    int32_j_03_0->addIncoming(const_int32_36, label_entry);
     
-    Argument* fwdref_38 = new Argument(VectorTy_7);
-    PHINode* packed_vec_01_0 = new PHINode(VectorTy_7, "vec.01.0", label_forbody6);
+    Argument* fwdref_51 = new Argument(VectorTy_11);
+    PHINode* packed_vec_01_0 = new PHINode(VectorTy_11, "vec.01.0", label_forbody6);
     packed_vec_01_0->reserveOperandSpace(3);
-    packed_vec_01_0->addIncoming(fwdref_38, label_forbody6);
-    packed_vec_01_0->addIncoming(const_packed_27, label_entry);
-    packed_vec_01_0->addIncoming(fwdref_38, label_forinc57);
+    packed_vec_01_0->addIncoming(fwdref_51, label_forbody6);
+    packed_vec_01_0->addIncoming(const_packed_37, label_entry);
+    packed_vec_01_0->addIncoming(fwdref_51, label_forinc70);
     
     std::vector<Value*> ptr_arraydecay11_indices;
-    ptr_arraydecay11_indices.push_back(int32_i_013_0);
+    ptr_arraydecay11_indices.push_back(int32_i_014_0);
     ptr_arraydecay11_indices.push_back(int32_j_03_0);
-    ptr_arraydecay11_indices.push_back(const_int32_26);
+    ptr_arraydecay11_indices.push_back(const_int32_36);
     Instruction* ptr_arraydecay11 = new GetElementPtrInst(ptr_ainputs, ptr_arraydecay11_indices.begin(), ptr_arraydecay11_indices.end(), "arraydecay11", label_forbody6);
     LoadInst* float_tmp13 = new LoadInst(ptr_arraydecay11, "tmp13", false, label_forbody6);
-    InsertElementInst* packed_tmp15 = new InsertElementInst(packed_vec_01_0, float_tmp13, const_int32_26, "tmp15", label_forbody6);
+    InsertElementInst* packed_tmp15 = new InsertElementInst(packed_vec_01_0, float_tmp13, const_int32_36, "tmp15", label_forbody6);
     std::vector<Value*> ptr_arrayidx23_indices;
-    ptr_arrayidx23_indices.push_back(int32_i_013_0);
+    ptr_arrayidx23_indices.push_back(int32_i_014_0);
     ptr_arrayidx23_indices.push_back(int32_j_03_0);
-    ptr_arrayidx23_indices.push_back(const_int32_28);
+    ptr_arrayidx23_indices.push_back(const_int32_38);
     Instruction* ptr_arrayidx23 = new GetElementPtrInst(ptr_ainputs, ptr_arrayidx23_indices.begin(), ptr_arrayidx23_indices.end(), "arrayidx23", label_forbody6);
     LoadInst* float_tmp24 = new LoadInst(ptr_arrayidx23, "tmp24", false, label_forbody6);
-    InsertElementInst* packed_tmp26 = new InsertElementInst(packed_tmp15, float_tmp24, const_int32_28, "tmp26", label_forbody6);
+    InsertElementInst* packed_tmp26 = new InsertElementInst(packed_tmp15, float_tmp24, const_int32_38, "tmp26", label_forbody6);
     std::vector<Value*> ptr_arrayidx34_indices;
-    ptr_arrayidx34_indices.push_back(int32_i_013_0);
+    ptr_arrayidx34_indices.push_back(int32_i_014_0);
     ptr_arrayidx34_indices.push_back(int32_j_03_0);
-    ptr_arrayidx34_indices.push_back(const_int32_29);
+    ptr_arrayidx34_indices.push_back(const_int32_39);
     Instruction* ptr_arrayidx34 = new GetElementPtrInst(ptr_ainputs, ptr_arrayidx34_indices.begin(), ptr_arrayidx34_indices.end(), "arrayidx34", label_forbody6);
     LoadInst* float_tmp35 = new LoadInst(ptr_arrayidx34, "tmp35", false, label_forbody6);
-    InsertElementInst* packed_tmp37 = new InsertElementInst(packed_tmp26, float_tmp35, const_int32_29, "tmp37", label_forbody6);
+    InsertElementInst* packed_tmp37 = new InsertElementInst(packed_tmp26, float_tmp35, const_int32_39, "tmp37", label_forbody6);
     std::vector<Value*> ptr_arrayidx45_indices;
-    ptr_arrayidx45_indices.push_back(int32_i_013_0);
+    ptr_arrayidx45_indices.push_back(int32_i_014_0);
     ptr_arrayidx45_indices.push_back(int32_j_03_0);
-    ptr_arrayidx45_indices.push_back(const_int32_30);
+    ptr_arrayidx45_indices.push_back(const_int32_40);
     Instruction* ptr_arrayidx45 = new GetElementPtrInst(ptr_ainputs, ptr_arrayidx45_indices.begin(), ptr_arrayidx45_indices.end(), "arrayidx45", label_forbody6);
     LoadInst* float_tmp46 = new LoadInst(ptr_arrayidx45, "tmp46", false, label_forbody6);
-    InsertElementInst* packed_tmp48 = new InsertElementInst(packed_tmp37, float_tmp46, const_int32_30, "tmp48", label_forbody6);
-    std::vector<Value*> ptr_arrayidx54_indices;
-    ptr_arrayidx54_indices.push_back(int32_i_013_0);
-    ptr_arrayidx54_indices.push_back(int32_j_03_0);
-    Instruction* ptr_arrayidx54 = new GetElementPtrInst(ptr_res, ptr_arrayidx54_indices.begin(), ptr_arrayidx54_indices.end(), "arrayidx54", label_forbody6);
-    new StoreInst(packed_tmp48, ptr_arrayidx54, false, label_forbody6);
-    BinaryOperator* int32_inc = BinaryOperator::create(Instruction::Add, int32_j_03_0, const_int32_28, "inc", label_forbody6);
-    ICmpInst* int1_cmp511 = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc, int32_num_attribs, "cmp511", label_forbody6);
-    new BranchInst(label_forbody6, label_forinc57, int1_cmp511, label_forbody6);
-    
-    // Block forinc57 (label_forinc57)
-    BinaryOperator* int32_inc59 = BinaryOperator::create(Instruction::Add, int32_i_013_0, const_int32_28, "inc59", label_forinc57);
-    ICmpInst* int1_cmp21 = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc59, int32_count, "cmp21", label_forinc57);
-    new BranchInst(label_forbody6, label_afterfor60, int1_cmp21, label_forinc57);
-    
-    // Block afterfor60 (label_afterfor60)
-    new ReturnInst(label_afterfor60);
+    InsertElementInst* packed_tmp48 = new InsertElementInst(packed_tmp37, float_tmp46, const_int32_40, "tmp48", label_forbody6);
+    CastInst* double_conv = new FPExtInst(float_tmp13, Type::DoubleTy, "conv", label_forbody6);
+    CastInst* double_conv55 = new FPExtInst(float_tmp24, Type::DoubleTy, "conv55", label_forbody6);
+    CastInst* double_conv58 = new FPExtInst(float_tmp35, Type::DoubleTy, "conv58", label_forbody6);
+    CastInst* double_conv61 = new FPExtInst(float_tmp46, Type::DoubleTy, "conv61", label_forbody6);
+    std::vector<Value*> int32_call_params;
+    int32_call_params.push_back(const_ptr_41);
+    int32_call_params.push_back(int32_i_014_0);
+    int32_call_params.push_back(int32_j_03_0);
+    int32_call_params.push_back(double_conv);
+    int32_call_params.push_back(double_conv55);
+    int32_call_params.push_back(double_conv58);
+    int32_call_params.push_back(double_conv61);
+    CallInst* int32_call = new CallInst(func_printf, int32_call_params.begin(), int32_call_params.end(), "call", label_forbody6);
+    int32_call->setCallingConv(CallingConv::C);
+    int32_call->setTailCall(true);
+    std::vector<Value*> ptr_arrayidx67_indices;
+    ptr_arrayidx67_indices.push_back(int32_i_014_0);
+    ptr_arrayidx67_indices.push_back(int32_j_03_0);
+    Instruction* ptr_arrayidx67 = new GetElementPtrInst(ptr_res, ptr_arrayidx67_indices.begin(), ptr_arrayidx67_indices.end(), "arrayidx67", label_forbody6);
+    StoreInst* void_52 = new StoreInst(packed_tmp48, ptr_arrayidx67, false, label_forbody6);
+    BinaryOperator* int32_inc = BinaryOperator::create(Instruction::Add, int32_j_03_0, const_int32_38, "inc", label_forbody6);
+    ICmpInst* int1_cmp512 = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc, int32_num_attribs, "cmp512", label_forbody6);
+    new BranchInst(label_forbody6, label_forinc70, int1_cmp512, label_forbody6);
+    
+    // Block forinc70 (label_forinc70)
+    BinaryOperator* int32_inc72 = BinaryOperator::create(Instruction::Add, int32_i_014_0, const_int32_38, "inc72", label_forinc70);
+    ICmpInst* int1_cmp23 = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc72, int32_count, "cmp23", label_forinc70);
+    new BranchInst(label_forbody6, label_afterfor73, int1_cmp23, label_forinc70);
+    
+    // Block afterfor73 (label_afterfor73)
+    new ReturnInst(label_afterfor73);
     
     // Resolve Forward References
-    fwdref_36->replaceAllUsesWith(int32_i_013_0); delete fwdref_36;
-    fwdref_38->replaceAllUsesWith(packed_tmp48); delete fwdref_38;
-    fwdref_37->replaceAllUsesWith(int32_inc); delete fwdref_37;
-    fwdref_35->replaceAllUsesWith(int32_inc59); delete fwdref_35;
+    fwdref_49->replaceAllUsesWith(int32_i_014_0); delete fwdref_49;
+    fwdref_51->replaceAllUsesWith(packed_tmp48); delete fwdref_51;
+    fwdref_50->replaceAllUsesWith(int32_inc); delete fwdref_50;
+    fwdref_48->replaceAllUsesWith(int32_inc72); delete fwdref_48;
     
   }
   
   // Function: from_consts (func_from_consts)
   {
     Function::arg_iterator args = func_from_consts->arg_begin();
-    Value* ptr_res_43 = args++;
-    ptr_res_43->setName("res");
-    Value* ptr_ainputs_44 = args++;
-    ptr_ainputs_44->setName("ainputs");
-    Value* int32_count_45 = args++;
-    int32_count_45->setName("count");
-    
-    BasicBlock* label_entry_46 = new BasicBlock("entry",func_from_consts,0);
+    Value* ptr_res_56 = args++;
+    ptr_res_56->setName("res");
+    Value* ptr_ainputs_57 = args++;
+    ptr_ainputs_57->setName("ainputs");
+    Value* int32_count_58 = args++;
+    int32_count_58->setName("count");
+    
+    BasicBlock* label_entry_59 = new BasicBlock("entry",func_from_consts,0);
     BasicBlock* label_forbody = new BasicBlock("forbody",func_from_consts,0);
     BasicBlock* label_afterfor = new BasicBlock("afterfor",func_from_consts,0);
     
-    // Block entry (label_entry_46)
-    ICmpInst* int1_cmp_47 = new ICmpInst(ICmpInst::ICMP_SGT, int32_count_45, const_int32_26, "cmp", label_entry_46);
-    new BranchInst(label_forbody, label_afterfor, int1_cmp_47, label_entry_46);
+    // Block entry (label_entry_59)
+    ICmpInst* int1_cmp_60 = new ICmpInst(ICmpInst::ICMP_SGT, int32_count_58, const_int32_36, "cmp", label_entry_59);
+    new BranchInst(label_forbody, label_afterfor, int1_cmp_60, label_entry_59);
     
     // Block forbody (label_forbody)
-    Argument* fwdref_49 = new Argument(IntegerType::get(32));
+    Argument* fwdref_62 = new Argument(IntegerType::get(32));
     PHINode* int32_i_02_0 = new PHINode(IntegerType::get(32), "i.02.0", label_forbody);
     int32_i_02_0->reserveOperandSpace(2);
-    int32_i_02_0->addIncoming(const_int32_26, label_entry_46);
-    int32_i_02_0->addIncoming(fwdref_49, label_forbody);
+    int32_i_02_0->addIncoming(const_int32_36, label_entry_59);
+    int32_i_02_0->addIncoming(fwdref_62, label_forbody);
     
-    Argument* fwdref_51 = new Argument(VectorTy_7);
-    PHINode* packed_vec_01_0_50 = new PHINode(VectorTy_7, "vec.01.0", label_forbody);
-    packed_vec_01_0_50->reserveOperandSpace(2);
-    packed_vec_01_0_50->addIncoming(const_packed_27, label_entry_46);
-    packed_vec_01_0_50->addIncoming(fwdref_51, label_forbody);
+    Argument* fwdref_64 = new Argument(VectorTy_11);
+    PHINode* packed_vec_01_0_63 = new PHINode(VectorTy_11, "vec.01.0", label_forbody);
+    packed_vec_01_0_63->reserveOperandSpace(2);
+    packed_vec_01_0_63->addIncoming(const_packed_37, label_entry_59);
+    packed_vec_01_0_63->addIncoming(fwdref_64, label_forbody);
     
     std::vector<Value*> ptr_arraydecay_indices;
     ptr_arraydecay_indices.push_back(int32_i_02_0);
-    ptr_arraydecay_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay = new GetElementPtrInst(ptr_ainputs_44, ptr_arraydecay_indices.begin(), ptr_arraydecay_indices.end(), "arraydecay", label_forbody);
+    ptr_arraydecay_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay = new GetElementPtrInst(ptr_ainputs_57, ptr_arraydecay_indices.begin(), ptr_arraydecay_indices.end(), "arraydecay", label_forbody);
     LoadInst* float_tmp5 = new LoadInst(ptr_arraydecay, "tmp5", false, label_forbody);
-    InsertElementInst* packed_tmp7 = new InsertElementInst(packed_vec_01_0_50, float_tmp5, const_int32_26, "tmp7", label_forbody);
+    InsertElementInst* packed_tmp7 = new InsertElementInst(packed_vec_01_0_63, float_tmp5, const_int32_36, "tmp7", label_forbody);
     std::vector<Value*> ptr_arrayidx12_indices;
     ptr_arrayidx12_indices.push_back(int32_i_02_0);
-    ptr_arrayidx12_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx12 = new GetElementPtrInst(ptr_ainputs_44, ptr_arrayidx12_indices.begin(), ptr_arrayidx12_indices.end(), "arrayidx12", label_forbody);
-    LoadInst* float_tmp13_52 = new LoadInst(ptr_arrayidx12, "tmp13", false, label_forbody);
-    InsertElementInst* packed_tmp15_53 = new InsertElementInst(packed_tmp7, float_tmp13_52, const_int32_28, "tmp15", label_forbody);
+    ptr_arrayidx12_indices.push_back(const_int32_38);
+    Instruction* ptr_arrayidx12 = new GetElementPtrInst(ptr_ainputs_57, ptr_arrayidx12_indices.begin(), ptr_arrayidx12_indices.end(), "arrayidx12", label_forbody);
+    LoadInst* float_tmp13_65 = new LoadInst(ptr_arrayidx12, "tmp13", false, label_forbody);
+    InsertElementInst* packed_tmp15_66 = new InsertElementInst(packed_tmp7, float_tmp13_65, const_int32_38, "tmp15", label_forbody);
     std::vector<Value*> ptr_arrayidx20_indices;
     ptr_arrayidx20_indices.push_back(int32_i_02_0);
-    ptr_arrayidx20_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx20 = new GetElementPtrInst(ptr_ainputs_44, ptr_arrayidx20_indices.begin(), ptr_arrayidx20_indices.end(), "arrayidx20", label_forbody);
+    ptr_arrayidx20_indices.push_back(const_int32_39);
+    Instruction* ptr_arrayidx20 = new GetElementPtrInst(ptr_ainputs_57, ptr_arrayidx20_indices.begin(), ptr_arrayidx20_indices.end(), "arrayidx20", label_forbody);
     LoadInst* float_tmp21 = new LoadInst(ptr_arrayidx20, "tmp21", false, label_forbody);
-    InsertElementInst* packed_tmp23 = new InsertElementInst(packed_tmp15_53, float_tmp21, const_int32_29, "tmp23", label_forbody);
+    InsertElementInst* packed_tmp23 = new InsertElementInst(packed_tmp15_66, float_tmp21, const_int32_39, "tmp23", label_forbody);
     std::vector<Value*> ptr_arrayidx28_indices;
     ptr_arrayidx28_indices.push_back(int32_i_02_0);
-    ptr_arrayidx28_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx28 = new GetElementPtrInst(ptr_ainputs_44, ptr_arrayidx28_indices.begin(), ptr_arrayidx28_indices.end(), "arrayidx28", label_forbody);
+    ptr_arrayidx28_indices.push_back(const_int32_40);
+    Instruction* ptr_arrayidx28 = new GetElementPtrInst(ptr_ainputs_57, ptr_arrayidx28_indices.begin(), ptr_arrayidx28_indices.end(), "arrayidx28", label_forbody);
     LoadInst* float_tmp29 = new LoadInst(ptr_arrayidx28, "tmp29", false, label_forbody);
-    InsertElementInst* packed_tmp31 = new InsertElementInst(packed_tmp23, float_tmp29, const_int32_30, "tmp31", label_forbody);
-    GetElementPtrInst* ptr_arrayidx34_54 = new GetElementPtrInst(ptr_res_43, int32_i_02_0, "arrayidx34", label_forbody);
-    new StoreInst(packed_tmp31, ptr_arrayidx34_54, false, label_forbody);
-    BinaryOperator* int32_indvar_next = BinaryOperator::create(Instruction::Add, int32_i_02_0, const_int32_28, "indvar.next", label_forbody);
-    ICmpInst* int1_exitcond = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next, int32_count_45, "exitcond", label_forbody);
+    InsertElementInst* packed_tmp31 = new InsertElementInst(packed_tmp23, float_tmp29, const_int32_40, "tmp31", label_forbody);
+    CastInst* double_conv_67 = new FPExtInst(float_tmp5, Type::DoubleTy, "conv", label_forbody);
+    CastInst* double_conv37 = new FPExtInst(float_tmp13_65, Type::DoubleTy, "conv37", label_forbody);
+    CastInst* double_conv40 = new FPExtInst(float_tmp21, Type::DoubleTy, "conv40", label_forbody);
+    CastInst* double_conv43 = new FPExtInst(float_tmp29, Type::DoubleTy, "conv43", label_forbody);
+    std::vector<Value*> int32_call_68_params;
+    int32_call_68_params.push_back(const_ptr_42);
+    int32_call_68_params.push_back(int32_i_02_0);
+    int32_call_68_params.push_back(double_conv_67);
+    int32_call_68_params.push_back(double_conv37);
+    int32_call_68_params.push_back(double_conv40);
+    int32_call_68_params.push_back(double_conv43);
+    CallInst* int32_call_68 = new CallInst(func_printf, int32_call_68_params.begin(), int32_call_68_params.end(), "call", label_forbody);
+    int32_call_68->setCallingConv(CallingConv::C);
+    int32_call_68->setTailCall(true);
+    GetElementPtrInst* ptr_arrayidx46 = new GetElementPtrInst(ptr_res_56, int32_i_02_0, "arrayidx46", label_forbody);
+    StoreInst* void_69 = new StoreInst(packed_tmp31, ptr_arrayidx46, false, label_forbody);
+    BinaryOperator* int32_indvar_next = BinaryOperator::create(Instruction::Add, int32_i_02_0, const_int32_38, "indvar.next", label_forbody);
+    ICmpInst* int1_exitcond = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next, int32_count_58, "exitcond", label_forbody);
     new BranchInst(label_afterfor, label_forbody, int1_exitcond, label_forbody);
     
     // Block afterfor (label_afterfor)
     new ReturnInst(label_afterfor);
     
     // Resolve Forward References
-    fwdref_51->replaceAllUsesWith(packed_tmp31); delete fwdref_51;
-    fwdref_49->replaceAllUsesWith(int32_indvar_next); delete fwdref_49;
+    fwdref_64->replaceAllUsesWith(packed_tmp31); delete fwdref_64;
+    fwdref_62->replaceAllUsesWith(int32_indvar_next); delete fwdref_62;
     
   }
   
@@ -352,457 +464,418 @@ Module* createBaseShader() {
     ptr_dests->setName("dests");
     Value* ptr_in = args++;
     ptr_in->setName("in");
-    Value* int32_num_attribs_58 = args++;
-    int32_num_attribs_58->setName("num_attribs");
+    Value* int32_num_attribs_72 = args++;
+    int32_num_attribs_72->setName("num_attribs");
     
-    BasicBlock* label_entry_59 = new BasicBlock("entry",func_to_array,0);
-    BasicBlock* label_forbody_60 = new BasicBlock("forbody",func_to_array,0);
-    BasicBlock* label_afterfor_61 = new BasicBlock("afterfor",func_to_array,0);
+    BasicBlock* label_entry_73 = new BasicBlock("entry",func_to_array,0);
+    BasicBlock* label_forbody_74 = new BasicBlock("forbody",func_to_array,0);
+    BasicBlock* label_afterfor_75 = new BasicBlock("afterfor",func_to_array,0);
     
-    // Block entry (label_entry_59)
-    ICmpInst* int1_cmp_62 = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_attribs_58, const_int32_26, "cmp", label_entry_59);
-    new BranchInst(label_forbody_60, label_afterfor_61, int1_cmp_62, label_entry_59);
+    // Block entry (label_entry_73)
+    ICmpInst* int1_cmp_76 = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_attribs_72, const_int32_36, "cmp", label_entry_73);
+    new BranchInst(label_forbody_74, label_afterfor_75, int1_cmp_76, label_entry_73);
     
-    // Block forbody (label_forbody_60)
-    Argument* fwdref_64 = new Argument(IntegerType::get(32));
-    PHINode* int32_i_01_0 = new PHINode(IntegerType::get(32), "i.01.0", label_forbody_60);
+    // Block forbody (label_forbody_74)
+    Argument* fwdref_78 = new Argument(IntegerType::get(32));
+    PHINode* int32_i_01_0 = new PHINode(IntegerType::get(32), "i.01.0", label_forbody_74);
     int32_i_01_0->reserveOperandSpace(2);
-    int32_i_01_0->addIncoming(const_int32_26, label_entry_59);
-    int32_i_01_0->addIncoming(fwdref_64, label_forbody_60);
-    
-    std::vector<Value*> ptr_arraydecay_65_indices;
-    ptr_arraydecay_65_indices.push_back(int32_i_01_0);
-    ptr_arraydecay_65_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay_65 = new GetElementPtrInst(ptr_dests, ptr_arraydecay_65_indices.begin(), ptr_arraydecay_65_indices.end(), "arraydecay", label_forbody_60);
-    GetElementPtrInst* ptr_arrayidx6 = new GetElementPtrInst(ptr_in, int32_i_01_0, "arrayidx6", label_forbody_60);
-    LoadInst* packed_tmp7_66 = new LoadInst(ptr_arrayidx6, "tmp7", false, label_forbody_60);
-    ExtractElementInst* float_tmp9 = new ExtractElementInst(packed_tmp7_66, const_int32_26, "tmp9", label_forbody_60);
-    CastInst* double_conv = new FPExtInst(float_tmp9, Type::DoubleTy, "conv", label_forbody_60);
-    ExtractElementInst* float_tmp11 = new ExtractElementInst(packed_tmp7_66, const_int32_28, "tmp11", label_forbody_60);
-    CastInst* double_conv12 = new FPExtInst(float_tmp11, Type::DoubleTy, "conv12", label_forbody_60);
-    ExtractElementInst* float_tmp14 = new ExtractElementInst(packed_tmp7_66, const_int32_29, "tmp14", label_forbody_60);
-    CastInst* double_conv15 = new FPExtInst(float_tmp14, Type::DoubleTy, "conv15", label_forbody_60);
-    ExtractElementInst* float_tmp17 = new ExtractElementInst(packed_tmp7_66, const_int32_30, "tmp17", label_forbody_60);
-    CastInst* double_conv18 = new FPExtInst(float_tmp17, Type::DoubleTy, "conv18", label_forbody_60);
-    std::vector<Value*> int32_call_params;
-    int32_call_params.push_back(const_ptr_31);
-    int32_call_params.push_back(double_conv);
-    int32_call_params.push_back(double_conv12);
-    int32_call_params.push_back(double_conv15);
-    int32_call_params.push_back(double_conv18);
-    CallInst* int32_call = new CallInst(func_printf, int32_call_params.begin(), int32_call_params.end(), "call", label_forbody_60);
-    int32_call->setCallingConv(CallingConv::C);
-    int32_call->setTailCall(true);
-    new StoreInst(float_tmp9, ptr_arraydecay_65, false, label_forbody_60);
+    int32_i_01_0->addIncoming(const_int32_36, label_entry_73);
+    int32_i_01_0->addIncoming(fwdref_78, label_forbody_74);
+    
+    std::vector<Value*> ptr_arraydecay_79_indices;
+    ptr_arraydecay_79_indices.push_back(int32_i_01_0);
+    ptr_arraydecay_79_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay_79 = new GetElementPtrInst(ptr_dests, ptr_arraydecay_79_indices.begin(), ptr_arraydecay_79_indices.end(), "arraydecay", label_forbody_74);
+    GetElementPtrInst* ptr_arrayidx6 = new GetElementPtrInst(ptr_in, int32_i_01_0, "arrayidx6", label_forbody_74);
+    LoadInst* packed_tmp7_80 = new LoadInst(ptr_arrayidx6, "tmp7", false, label_forbody_74);
+    ExtractElementInst* float_tmp9 = new ExtractElementInst(packed_tmp7_80, const_int32_36, "tmp9", label_forbody_74);
+    CastInst* double_conv_81 = new FPExtInst(float_tmp9, Type::DoubleTy, "conv", label_forbody_74);
+    ExtractElementInst* float_tmp11 = new ExtractElementInst(packed_tmp7_80, const_int32_38, "tmp11", label_forbody_74);
+    CastInst* double_conv12 = new FPExtInst(float_tmp11, Type::DoubleTy, "conv12", label_forbody_74);
+    ExtractElementInst* float_tmp14 = new ExtractElementInst(packed_tmp7_80, const_int32_39, "tmp14", label_forbody_74);
+    CastInst* double_conv15 = new FPExtInst(float_tmp14, Type::DoubleTy, "conv15", label_forbody_74);
+    ExtractElementInst* float_tmp17 = new ExtractElementInst(packed_tmp7_80, const_int32_40, "tmp17", label_forbody_74);
+    CastInst* double_conv18 = new FPExtInst(float_tmp17, Type::DoubleTy, "conv18", label_forbody_74);
+    std::vector<Value*> int32_call_82_params;
+    int32_call_82_params.push_back(const_ptr_43);
+    int32_call_82_params.push_back(double_conv_81);
+    int32_call_82_params.push_back(double_conv12);
+    int32_call_82_params.push_back(double_conv15);
+    int32_call_82_params.push_back(double_conv18);
+    CallInst* int32_call_82 = new CallInst(func_printf, int32_call_82_params.begin(), int32_call_82_params.end(), "call", label_forbody_74);
+    int32_call_82->setCallingConv(CallingConv::C);
+    int32_call_82->setTailCall(true);
+    StoreInst* void_83 = new StoreInst(float_tmp9, ptr_arraydecay_79, false, label_forbody_74);
     std::vector<Value*> ptr_arrayidx24_indices;
     ptr_arrayidx24_indices.push_back(int32_i_01_0);
-    ptr_arrayidx24_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx24 = new GetElementPtrInst(ptr_dests, ptr_arrayidx24_indices.begin(), ptr_arrayidx24_indices.end(), "arrayidx24", label_forbody_60);
-    new StoreInst(float_tmp11, ptr_arrayidx24, false, label_forbody_60);
-    std::vector<Value*> ptr_arrayidx28_69_indices;
-    ptr_arrayidx28_69_indices.push_back(int32_i_01_0);
-    ptr_arrayidx28_69_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx28_69 = new GetElementPtrInst(ptr_dests, ptr_arrayidx28_69_indices.begin(), ptr_arrayidx28_69_indices.end(), "arrayidx28", label_forbody_60);
-    new StoreInst(float_tmp14, ptr_arrayidx28_69, false, label_forbody_60);
+    ptr_arrayidx24_indices.push_back(const_int32_38);
+    Instruction* ptr_arrayidx24 = new GetElementPtrInst(ptr_dests, ptr_arrayidx24_indices.begin(), ptr_arrayidx24_indices.end(), "arrayidx24", label_forbody_74);
+    StoreInst* void_84 = new StoreInst(float_tmp11, ptr_arrayidx24, false, label_forbody_74);
+    std::vector<Value*> ptr_arrayidx28_85_indices;
+    ptr_arrayidx28_85_indices.push_back(int32_i_01_0);
+    ptr_arrayidx28_85_indices.push_back(const_int32_39);
+    Instruction* ptr_arrayidx28_85 = new GetElementPtrInst(ptr_dests, ptr_arrayidx28_85_indices.begin(), ptr_arrayidx28_85_indices.end(), "arrayidx28", label_forbody_74);
+    StoreInst* void_86 = new StoreInst(float_tmp14, ptr_arrayidx28_85, false, label_forbody_74);
     std::vector<Value*> ptr_arrayidx32_indices;
     ptr_arrayidx32_indices.push_back(int32_i_01_0);
-    ptr_arrayidx32_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx32 = new GetElementPtrInst(ptr_dests, ptr_arrayidx32_indices.begin(), ptr_arrayidx32_indices.end(), "arrayidx32", label_forbody_60);
-    new StoreInst(float_tmp17, ptr_arrayidx32, false, label_forbody_60);
-    BinaryOperator* int32_indvar_next_72 = BinaryOperator::create(Instruction::Add, int32_i_01_0, const_int32_28, "indvar.next", label_forbody_60);
-    ICmpInst* int1_exitcond_73 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next_72, int32_num_attribs_58, "exitcond", label_forbody_60);
-    new BranchInst(label_afterfor_61, label_forbody_60, int1_exitcond_73, label_forbody_60);
+    ptr_arrayidx32_indices.push_back(const_int32_40);
+    Instruction* ptr_arrayidx32 = new GetElementPtrInst(ptr_dests, ptr_arrayidx32_indices.begin(), ptr_arrayidx32_indices.end(), "arrayidx32", label_forbody_74);
+    StoreInst* void_87 = new StoreInst(float_tmp17, ptr_arrayidx32, false, label_forbody_74);
+    BinaryOperator* int32_indvar_next_88 = BinaryOperator::create(Instruction::Add, int32_i_01_0, const_int32_38, "indvar.next", label_forbody_74);
+    ICmpInst* int1_exitcond_89 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next_88, int32_num_attribs_72, "exitcond", label_forbody_74);
+    new BranchInst(label_afterfor_75, label_forbody_74, int1_exitcond_89, label_forbody_74);
     
-    // Block afterfor (label_afterfor_61)
-    new ReturnInst(label_afterfor_61);
+    // Block afterfor (label_afterfor_75)
+    new ReturnInst(label_afterfor_75);
     
     // Resolve Forward References
-    fwdref_64->replaceAllUsesWith(int32_indvar_next_72); delete fwdref_64;
+    fwdref_78->replaceAllUsesWith(int32_indvar_next_88); delete fwdref_78;
     
   }
   
   // Function: run_vertex_shader (func_run_vertex_shader)
   {
     Function::arg_iterator args = func_run_vertex_shader->arg_begin();
-    Value* ptr_ainputs_76 = args++;
-    ptr_ainputs_76->setName("ainputs");
-    Value* ptr_dests_77 = args++;
-    ptr_dests_77->setName("dests");
+    Value* ptr_ainputs_92 = args++;
+    ptr_ainputs_92->setName("ainputs");
+    Value* ptr_dests_93 = args++;
+    ptr_dests_93->setName("dests");
     Value* ptr_aconsts = args++;
     ptr_aconsts->setName("aconsts");
-    Value* int32_count_78 = args++;
-    int32_count_78->setName("count");
-    Value* int32_num_attribs_79 = args++;
-    int32_num_attribs_79->setName("num_attribs");
-    
-    BasicBlock* label_entry_80 = new BasicBlock("entry",func_run_vertex_shader,0);
+    Value* int32_num_vertices = args++;
+    int32_num_vertices->setName("num_vertices");
+    Value* int32_num_inputs = args++;
+    int32_num_inputs->setName("num_inputs");
+    Value* int32_num_attribs_94 = args++;
+    int32_num_attribs_94->setName("num_attribs");
+    Value* int32_num_consts = args++;
+    int32_num_consts->setName("num_consts");
+    
+    BasicBlock* label_entry_95 = new BasicBlock("entry",func_run_vertex_shader,0);
     BasicBlock* label_forbody6_i = new BasicBlock("forbody6.i",func_run_vertex_shader,0);
-    BasicBlock* label_forinc57_i = new BasicBlock("forinc57.i",func_run_vertex_shader,0);
-    BasicBlock* label_forbody_i11_preheader = new BasicBlock("forbody.i11.preheader",func_run_vertex_shader,0);
+    BasicBlock* label_forinc70_i = new BasicBlock("forinc70.i",func_run_vertex_shader,0);
+    BasicBlock* label_from_array_exit = new BasicBlock("from_array.exit",func_run_vertex_shader,0);
+    BasicBlock* label_forbody_i15 = new BasicBlock("forbody.i15",func_run_vertex_shader,0);
+    BasicBlock* label_from_consts_exit = new BasicBlock("from_consts.exit",func_run_vertex_shader,0);
     BasicBlock* label_forbody_preheader = new BasicBlock("forbody.preheader",func_run_vertex_shader,0);
     BasicBlock* label_forbody_us = new BasicBlock("forbody.us",func_run_vertex_shader,0);
     BasicBlock* label_to_array_exit_us = new BasicBlock("to_array.exit.us",func_run_vertex_shader,0);
     BasicBlock* label_forbody_i_us = new BasicBlock("forbody.i.us",func_run_vertex_shader,0);
-    BasicBlock* label_forbody_81 = new BasicBlock("forbody",func_run_vertex_shader,0);
-    BasicBlock* label_afterfor_82 = new BasicBlock("afterfor",func_run_vertex_shader,0);
-    
-    // Block entry (label_entry_80)
-    AllocaInst* ptr_inputs = new AllocaInst(ArrayTy_21, "inputs", label_entry_80);
-    AllocaInst* ptr_consts = new AllocaInst(ArrayTy_6, "consts", label_entry_80);
-    AllocaInst* ptr_results = new AllocaInst(ArrayTy_21, "results", label_entry_80);
-    CallInst* int32_call_83 = new CallInst(func_printf, const_ptr_32, "call", label_entry_80);
-    int32_call_83->setCallingConv(CallingConv::C);
-    int32_call_83->setTailCall(false);
-    ICmpInst* int1_cmp_i = new ICmpInst(ICmpInst::ICMP_SGT, int32_count_78, const_int32_26, "cmp.i", label_entry_80);
-    ICmpInst* int1_cmp5_i = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_attribs_79, const_int32_26, "cmp5.i", label_entry_80);
-    BinaryOperator* int1_bothcond_i = BinaryOperator::create(Instruction::And, int1_cmp5_i, int1_cmp_i, "bothcond.i", label_entry_80);
-    new BranchInst(label_forbody6_i, label_forbody_i11_preheader, int1_bothcond_i, label_entry_80);
+    BasicBlock* label_forbody_96 = new BasicBlock("forbody",func_run_vertex_shader,0);
+    BasicBlock* label_afterfor_97 = new BasicBlock("afterfor",func_run_vertex_shader,0);
+    
+    // Block entry (label_entry_95)
+    AllocaInst* ptr_inputs = new AllocaInst(ArrayTy_25, "inputs", label_entry_95);
+    AllocaInst* ptr_consts = new AllocaInst(ArrayTy_27, "consts", label_entry_95);
+    AllocaInst* ptr_results = new AllocaInst(ArrayTy_25, "results", label_entry_95);
+    std::vector<Value*> int32_call_98_params;
+    int32_call_98_params.push_back(const_ptr_44);
+    int32_call_98_params.push_back(int32_num_vertices);
+    int32_call_98_params.push_back(int32_num_inputs);
+    int32_call_98_params.push_back(int32_num_consts);
+    CallInst* int32_call_98 = new CallInst(func_printf, int32_call_98_params.begin(), int32_call_98_params.end(), "call", label_entry_95);
+    int32_call_98->setCallingConv(CallingConv::C);
+    int32_call_98->setTailCall(false);
+    ICmpInst* int1_cmp_i = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_vertices, const_int32_36, "cmp.i", label_entry_95);
+    ICmpInst* int1_cmp5_i = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_inputs, const_int32_36, "cmp5.i", label_entry_95);
+    BinaryOperator* int1_bothcond_i = BinaryOperator::create(Instruction::And, int1_cmp5_i, int1_cmp_i, "bothcond.i", label_entry_95);
+    new BranchInst(label_forbody6_i, label_from_array_exit, int1_bothcond_i, label_entry_95);
     
     // Block forbody6.i (label_forbody6_i)
-    Argument* fwdref_85 = new Argument(IntegerType::get(32));
-    Argument* fwdref_86 = new Argument(IntegerType::get(32));
-    PHINode* int32_i_013_0_i_ph = new PHINode(IntegerType::get(32), "i.013.0.i.ph", label_forbody6_i);
-    int32_i_013_0_i_ph->reserveOperandSpace(3);
-    int32_i_013_0_i_ph->addIncoming(const_int32_26, label_entry_80);
-    int32_i_013_0_i_ph->addIncoming(fwdref_85, label_forinc57_i);
-    int32_i_013_0_i_ph->addIncoming(fwdref_86, label_forbody6_i);
-    
-    Argument* fwdref_87 = new Argument(IntegerType::get(32));
+    Argument* fwdref_100 = new Argument(IntegerType::get(32));
+    Argument* fwdref_101 = new Argument(IntegerType::get(32));
+    PHINode* int32_i_014_0_i_ph = new PHINode(IntegerType::get(32), "i.014.0.i.ph", label_forbody6_i);
+    int32_i_014_0_i_ph->reserveOperandSpace(3);
+    int32_i_014_0_i_ph->addIncoming(const_int32_36, label_entry_95);
+    int32_i_014_0_i_ph->addIncoming(fwdref_100, label_forinc70_i);
+    int32_i_014_0_i_ph->addIncoming(fwdref_101, label_forbody6_i);
+    
+    Argument* fwdref_102 = new Argument(IntegerType::get(32));
     PHINode* int32_j_03_0_i = new PHINode(IntegerType::get(32), "j.03.0.i", label_forbody6_i);
     int32_j_03_0_i->reserveOperandSpace(3);
-    int32_j_03_0_i->addIncoming(fwdref_87, label_forbody6_i);
-    int32_j_03_0_i->addIncoming(const_int32_26, label_forinc57_i);
-    int32_j_03_0_i->addIncoming(const_int32_26, label_entry_80);
+    int32_j_03_0_i->addIncoming(fwdref_102, label_forbody6_i);
+    int32_j_03_0_i->addIncoming(const_int32_36, label_forinc70_i);
+    int32_j_03_0_i->addIncoming(const_int32_36, label_entry_95);
     
-    Argument* fwdref_88 = new Argument(VectorTy_7);
-    PHINode* packed_vec_01_0_i = new PHINode(VectorTy_7, "vec.01.0.i", label_forbody6_i);
+    Argument* fwdref_103 = new Argument(VectorTy_11);
+    PHINode* packed_vec_01_0_i = new PHINode(VectorTy_11, "vec.01.0.i", label_forbody6_i);
     packed_vec_01_0_i->reserveOperandSpace(3);
-    packed_vec_01_0_i->addIncoming(fwdref_88, label_forbody6_i);
-    packed_vec_01_0_i->addIncoming(const_packed_27, label_entry_80);
-    packed_vec_01_0_i->addIncoming(fwdref_88, label_forinc57_i);
+    packed_vec_01_0_i->addIncoming(fwdref_103, label_forbody6_i);
+    packed_vec_01_0_i->addIncoming(const_packed_37, label_entry_95);
+    packed_vec_01_0_i->addIncoming(fwdref_103, label_forinc70_i);
     
     std::vector<Value*> ptr_arraydecay11_i_indices;
-    ptr_arraydecay11_i_indices.push_back(int32_i_013_0_i_ph);
+    ptr_arraydecay11_i_indices.push_back(int32_i_014_0_i_ph);
     ptr_arraydecay11_i_indices.push_back(int32_j_03_0_i);
-    ptr_arraydecay11_i_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay11_i = new GetElementPtrInst(ptr_ainputs_76, ptr_arraydecay11_i_indices.begin(), ptr_arraydecay11_i_indices.end(), "arraydecay11.i", label_forbody6_i);
+    ptr_arraydecay11_i_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay11_i = new GetElementPtrInst(ptr_ainputs_92, ptr_arraydecay11_i_indices.begin(), ptr_arraydecay11_i_indices.end(), "arraydecay11.i", label_forbody6_i);
     LoadInst* float_tmp13_i = new LoadInst(ptr_arraydecay11_i, "tmp13.i", false, label_forbody6_i);
-    InsertElementInst* packed_tmp15_i = new InsertElementInst(packed_vec_01_0_i, float_tmp13_i, const_int32_26, "tmp15.i", label_forbody6_i);
+    InsertElementInst* packed_tmp15_i = new InsertElementInst(packed_vec_01_0_i, float_tmp13_i, const_int32_36, "tmp15.i", label_forbody6_i);
     std::vector<Value*> ptr_arrayidx23_i_indices;
-    ptr_arrayidx23_i_indices.push_back(int32_i_013_0_i_ph);
+    ptr_arrayidx23_i_indices.push_back(int32_i_014_0_i_ph);
     ptr_arrayidx23_i_indices.push_back(int32_j_03_0_i);
-    ptr_arrayidx23_i_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx23_i = new GetElementPtrInst(ptr_ainputs_76, ptr_arrayidx23_i_indices.begin(), ptr_arrayidx23_i_indices.end(), "arrayidx23.i", label_forbody6_i);
+    ptr_arrayidx23_i_indices.push_back(const_int32_38);
+    Instruction* ptr_arrayidx23_i = new GetElementPtrInst(ptr_ainputs_92, ptr_arrayidx23_i_indices.begin(), ptr_arrayidx23_i_indices.end(), "arrayidx23.i", label_forbody6_i);
     LoadInst* float_tmp24_i = new LoadInst(ptr_arrayidx23_i, "tmp24.i", false, label_forbody6_i);
-    InsertElementInst* packed_tmp26_i = new InsertElementInst(packed_tmp15_i, float_tmp24_i, const_int32_28, "tmp26.i", label_forbody6_i);
+    InsertElementInst* packed_tmp26_i = new InsertElementInst(packed_tmp15_i, float_tmp24_i, const_int32_38, "tmp26.i", label_forbody6_i);
     std::vector<Value*> ptr_arrayidx34_i_indices;
-    ptr_arrayidx34_i_indices.push_back(int32_i_013_0_i_ph);
+    ptr_arrayidx34_i_indices.push_back(int32_i_014_0_i_ph);
     ptr_arrayidx34_i_indices.push_back(int32_j_03_0_i);
-    ptr_arrayidx34_i_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx34_i = new GetElementPtrInst(ptr_ainputs_76, ptr_arrayidx34_i_indices.begin(), ptr_arrayidx34_i_indices.end(), "arrayidx34.i", label_forbody6_i);
+    ptr_arrayidx34_i_indices.push_back(const_int32_39);
+    Instruction* ptr_arrayidx34_i = new GetElementPtrInst(ptr_ainputs_92, ptr_arrayidx34_i_indices.begin(), ptr_arrayidx34_i_indices.end(), "arrayidx34.i", label_forbody6_i);
     LoadInst* float_tmp35_i = new LoadInst(ptr_arrayidx34_i, "tmp35.i", false, label_forbody6_i);
-    InsertElementInst* packed_tmp37_i = new InsertElementInst(packed_tmp26_i, float_tmp35_i, const_int32_29, "tmp37.i", label_forbody6_i);
+    InsertElementInst* packed_tmp37_i = new InsertElementInst(packed_tmp26_i, float_tmp35_i, const_int32_39, "tmp37.i", label_forbody6_i);
     std::vector<Value*> ptr_arrayidx45_i_indices;
-    ptr_arrayidx45_i_indices.push_back(int32_i_013_0_i_ph);
+    ptr_arrayidx45_i_indices.push_back(int32_i_014_0_i_ph);
     ptr_arrayidx45_i_indices.push_back(int32_j_03_0_i);
-    ptr_arrayidx45_i_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx45_i = new GetElementPtrInst(ptr_ainputs_76, ptr_arrayidx45_i_indices.begin(), ptr_arrayidx45_i_indices.end(), "arrayidx45.i", label_forbody6_i);
+    ptr_arrayidx45_i_indices.push_back(const_int32_40);
+    Instruction* ptr_arrayidx45_i = new GetElementPtrInst(ptr_ainputs_92, ptr_arrayidx45_i_indices.begin(), ptr_arrayidx45_i_indices.end(), "arrayidx45.i", label_forbody6_i);
     LoadInst* float_tmp46_i = new LoadInst(ptr_arrayidx45_i, "tmp46.i", false, label_forbody6_i);
-    InsertElementInst* packed_tmp48_i = new InsertElementInst(packed_tmp37_i, float_tmp46_i, const_int32_30, "tmp48.i", label_forbody6_i);
-    std::vector<Value*> ptr_arrayidx54_i_indices;
-    ptr_arrayidx54_i_indices.push_back(const_int32_26);
-    ptr_arrayidx54_i_indices.push_back(int32_i_013_0_i_ph);
-    ptr_arrayidx54_i_indices.push_back(int32_j_03_0_i);
-    Instruction* ptr_arrayidx54_i = new GetElementPtrInst(ptr_inputs, ptr_arrayidx54_i_indices.begin(), ptr_arrayidx54_i_indices.end(), "arrayidx54.i", label_forbody6_i);
-    new StoreInst(packed_tmp48_i, ptr_arrayidx54_i, false, label_forbody6_i);
-    BinaryOperator* int32_inc_i = BinaryOperator::create(Instruction::Add, int32_j_03_0_i, const_int32_28, "inc.i", label_forbody6_i);
-    ICmpInst* int1_cmp511_i = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc_i, int32_num_attribs_79, "cmp511.i", label_forbody6_i);
-    new BranchInst(label_forbody6_i, label_forinc57_i, int1_cmp511_i, label_forbody6_i);
-    
-    // Block forinc57.i (label_forinc57_i)
-    BinaryOperator* int32_inc59_i = BinaryOperator::create(Instruction::Add, int32_i_013_0_i_ph, const_int32_28, "inc59.i", label_forinc57_i);
-    ICmpInst* int1_cmp21_i = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc59_i, int32_count_78, "cmp21.i", label_forinc57_i);
-    new BranchInst(label_forbody6_i, label_forbody_i11_preheader, int1_cmp21_i, label_forinc57_i);
-    
-    // Block forbody.i11.preheader (label_forbody_i11_preheader)
-    std::vector<Value*> ptr_arraydecay_i3_indices;
-    ptr_arraydecay_i3_indices.push_back(const_int32_26);
-    ptr_arraydecay_i3_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay_i3 = new GetElementPtrInst(ptr_aconsts, ptr_arraydecay_i3_indices.begin(), ptr_arraydecay_i3_indices.end(), "arraydecay.i3", label_forbody_i11_preheader);
-    LoadInst* float_tmp5_i = new LoadInst(ptr_arraydecay_i3, "tmp5.i", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp7_i4 = new InsertElementInst(const_packed_27, float_tmp5_i, const_int32_26, "tmp7.i4", label_forbody_i11_preheader);
+    InsertElementInst* packed_tmp48_i = new InsertElementInst(packed_tmp37_i, float_tmp46_i, const_int32_40, "tmp48.i", label_forbody6_i);
+    CastInst* double_conv_i = new FPExtInst(float_tmp13_i, Type::DoubleTy, "conv.i", label_forbody6_i);
+    CastInst* double_conv55_i = new FPExtInst(float_tmp24_i, Type::DoubleTy, "conv55.i", label_forbody6_i);
+    CastInst* double_conv58_i = new FPExtInst(float_tmp35_i, Type::DoubleTy, "conv58.i", label_forbody6_i);
+    CastInst* double_conv61_i = new FPExtInst(float_tmp46_i, Type::DoubleTy, "conv61.i", label_forbody6_i);
+    std::vector<Value*> int32_call_i_params;
+    int32_call_i_params.push_back(const_ptr_41);
+    int32_call_i_params.push_back(int32_i_014_0_i_ph);
+    int32_call_i_params.push_back(int32_j_03_0_i);
+    int32_call_i_params.push_back(double_conv_i);
+    int32_call_i_params.push_back(double_conv55_i);
+    int32_call_i_params.push_back(double_conv58_i);
+    int32_call_i_params.push_back(double_conv61_i);
+    CallInst* int32_call_i = new CallInst(func_printf, int32_call_i_params.begin(), int32_call_i_params.end(), "call.i", label_forbody6_i);
+    int32_call_i->setCallingConv(CallingConv::C);
+    int32_call_i->setTailCall(false);
+    std::vector<Value*> ptr_arrayidx67_i_indices;
+    ptr_arrayidx67_i_indices.push_back(const_int32_36);
+    ptr_arrayidx67_i_indices.push_back(int32_i_014_0_i_ph);
+    ptr_arrayidx67_i_indices.push_back(int32_j_03_0_i);
+    Instruction* ptr_arrayidx67_i = new GetElementPtrInst(ptr_inputs, ptr_arrayidx67_i_indices.begin(), ptr_arrayidx67_i_indices.end(), "arrayidx67.i", label_forbody6_i);
+    StoreInst* void_104 = new StoreInst(packed_tmp48_i, ptr_arrayidx67_i, false, label_forbody6_i);
+    BinaryOperator* int32_inc_i = BinaryOperator::create(Instruction::Add, int32_j_03_0_i, const_int32_38, "inc.i", label_forbody6_i);
+    ICmpInst* int1_cmp512_i = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc_i, int32_num_inputs, "cmp512.i", label_forbody6_i);
+    new BranchInst(label_forbody6_i, label_forinc70_i, int1_cmp512_i, label_forbody6_i);
+    
+    // Block forinc70.i (label_forinc70_i)
+    BinaryOperator* int32_inc72_i = BinaryOperator::create(Instruction::Add, int32_i_014_0_i_ph, const_int32_38, "inc72.i", label_forinc70_i);
+    ICmpInst* int1_cmp23_i = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc72_i, int32_num_vertices, "cmp23.i", label_forinc70_i);
+    new BranchInst(label_forbody6_i, label_from_array_exit, int1_cmp23_i, label_forinc70_i);
+    
+    // Block from_array.exit (label_from_array_exit)
+    ICmpInst* int1_cmp_i4 = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_consts, const_int32_36, "cmp.i4", label_from_array_exit);
+    new BranchInst(label_forbody_i15, label_from_consts_exit, int1_cmp_i4, label_from_array_exit);
+    
+    // Block forbody.i15 (label_forbody_i15)
+    Argument* fwdref_108 = new Argument(IntegerType::get(32));
+    PHINode* int32_i_02_0_i = new PHINode(IntegerType::get(32), "i.02.0.i", label_forbody_i15);
+    int32_i_02_0_i->reserveOperandSpace(2);
+    int32_i_02_0_i->addIncoming(const_int32_36, label_from_array_exit);
+    int32_i_02_0_i->addIncoming(fwdref_108, label_forbody_i15);
+    
+    Argument* fwdref_109 = new Argument(VectorTy_11);
+    PHINode* packed_vec_01_0_i5 = new PHINode(VectorTy_11, "vec.01.0.i5", label_forbody_i15);
+    packed_vec_01_0_i5->reserveOperandSpace(2);
+    packed_vec_01_0_i5->addIncoming(const_packed_37, label_from_array_exit);
+    packed_vec_01_0_i5->addIncoming(fwdref_109, label_forbody_i15);
+    
+    std::vector<Value*> ptr_arraydecay_i6_indices;
+    ptr_arraydecay_i6_indices.push_back(int32_i_02_0_i);
+    ptr_arraydecay_i6_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay_i6 = new GetElementPtrInst(ptr_aconsts, ptr_arraydecay_i6_indices.begin(), ptr_arraydecay_i6_indices.end(), "arraydecay.i6", label_forbody_i15);
+    LoadInst* float_tmp5_i = new LoadInst(ptr_arraydecay_i6, "tmp5.i", false, label_forbody_i15);
+    InsertElementInst* packed_tmp7_i7 = new InsertElementInst(packed_vec_01_0_i5, float_tmp5_i, const_int32_36, "tmp7.i7", label_forbody_i15);
     std::vector<Value*> ptr_arrayidx12_i_indices;
-    ptr_arrayidx12_i_indices.push_back(const_int32_26);
-    ptr_arrayidx12_i_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx12_i = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx12_i_indices.begin(), ptr_arrayidx12_i_indices.end(), "arrayidx12.i", label_forbody_i11_preheader);
-    LoadInst* float_tmp13_i5 = new LoadInst(ptr_arrayidx12_i, "tmp13.i5", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp15_i6 = new InsertElementInst(packed_tmp7_i4, float_tmp13_i5, const_int32_28, "tmp15.i6", label_forbody_i11_preheader);
+    ptr_arrayidx12_i_indices.push_back(int32_i_02_0_i);
+    ptr_arrayidx12_i_indices.push_back(const_int32_38);
+    Instruction* ptr_arrayidx12_i = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx12_i_indices.begin(), ptr_arrayidx12_i_indices.end(), "arrayidx12.i", label_forbody_i15);
+    LoadInst* float_tmp13_i8 = new LoadInst(ptr_arrayidx12_i, "tmp13.i8", false, label_forbody_i15);
+    InsertElementInst* packed_tmp15_i9 = new InsertElementInst(packed_tmp7_i7, float_tmp13_i8, const_int32_38, "tmp15.i9", label_forbody_i15);
     std::vector<Value*> ptr_arrayidx20_i_indices;
-    ptr_arrayidx20_i_indices.push_back(const_int32_26);
-    ptr_arrayidx20_i_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx20_i = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx20_i_indices.begin(), ptr_arrayidx20_i_indices.end(), "arrayidx20.i", label_forbody_i11_preheader);
-    LoadInst* float_tmp21_i = new LoadInst(ptr_arrayidx20_i, "tmp21.i", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp23_i = new InsertElementInst(packed_tmp15_i6, float_tmp21_i, const_int32_29, "tmp23.i", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx28_i7_indices;
-    ptr_arrayidx28_i7_indices.push_back(const_int32_26);
-    ptr_arrayidx28_i7_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx28_i7 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx28_i7_indices.begin(), ptr_arrayidx28_i7_indices.end(), "arrayidx28.i7", label_forbody_i11_preheader);
-    LoadInst* float_tmp29_i = new LoadInst(ptr_arrayidx28_i7, "tmp29.i", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp31_i = new InsertElementInst(packed_tmp23_i, float_tmp29_i, const_int32_30, "tmp31.i", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx34_i8_indices;
-    ptr_arrayidx34_i8_indices.push_back(const_int32_26);
-    ptr_arrayidx34_i8_indices.push_back(const_int32_26);
-    Instruction* ptr_arrayidx34_i8 = new GetElementPtrInst(ptr_consts, ptr_arrayidx34_i8_indices.begin(), ptr_arrayidx34_i8_indices.end(), "arrayidx34.i8", label_forbody_i11_preheader);
-    new StoreInst(packed_tmp31_i, ptr_arrayidx34_i8, false, label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arraydecay_i3_1_indices;
-    ptr_arraydecay_i3_1_indices.push_back(const_int32_28);
-    ptr_arraydecay_i3_1_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay_i3_1 = new GetElementPtrInst(ptr_aconsts, ptr_arraydecay_i3_1_indices.begin(), ptr_arraydecay_i3_1_indices.end(), "arraydecay.i3.1", label_forbody_i11_preheader);
-    LoadInst* float_tmp5_i_1 = new LoadInst(ptr_arraydecay_i3_1, "tmp5.i.1", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp7_i4_1 = new InsertElementInst(packed_tmp31_i, float_tmp5_i_1, const_int32_26, "tmp7.i4.1", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx12_i_1_indices;
-    ptr_arrayidx12_i_1_indices.push_back(const_int32_28);
-    ptr_arrayidx12_i_1_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx12_i_1 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx12_i_1_indices.begin(), ptr_arrayidx12_i_1_indices.end(), "arrayidx12.i.1", label_forbody_i11_preheader);
-    LoadInst* float_tmp13_i5_1 = new LoadInst(ptr_arrayidx12_i_1, "tmp13.i5.1", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp15_i6_1 = new InsertElementInst(packed_tmp7_i4_1, float_tmp13_i5_1, const_int32_28, "tmp15.i6.1", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx20_i_1_indices;
-    ptr_arrayidx20_i_1_indices.push_back(const_int32_28);
-    ptr_arrayidx20_i_1_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx20_i_1 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx20_i_1_indices.begin(), ptr_arrayidx20_i_1_indices.end(), "arrayidx20.i.1", label_forbody_i11_preheader);
-    LoadInst* float_tmp21_i_1 = new LoadInst(ptr_arrayidx20_i_1, "tmp21.i.1", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp23_i_1 = new InsertElementInst(packed_tmp15_i6_1, float_tmp21_i_1, const_int32_29, "tmp23.i.1", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx28_i7_1_indices;
-    ptr_arrayidx28_i7_1_indices.push_back(const_int32_28);
-    ptr_arrayidx28_i7_1_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx28_i7_1 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx28_i7_1_indices.begin(), ptr_arrayidx28_i7_1_indices.end(), "arrayidx28.i7.1", label_forbody_i11_preheader);
-    LoadInst* float_tmp29_i_1 = new LoadInst(ptr_arrayidx28_i7_1, "tmp29.i.1", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp31_i_1 = new InsertElementInst(packed_tmp23_i_1, float_tmp29_i_1, const_int32_30, "tmp31.i.1", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx34_i8_1_indices;
-    ptr_arrayidx34_i8_1_indices.push_back(const_int32_26);
-    ptr_arrayidx34_i8_1_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx34_i8_1 = new GetElementPtrInst(ptr_consts, ptr_arrayidx34_i8_1_indices.begin(), ptr_arrayidx34_i8_1_indices.end(), "arrayidx34.i8.1", label_forbody_i11_preheader);
-    new StoreInst(packed_tmp31_i_1, ptr_arrayidx34_i8_1, false, label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arraydecay_i3_2_indices;
-    ptr_arraydecay_i3_2_indices.push_back(const_int32_29);
-    ptr_arraydecay_i3_2_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay_i3_2 = new GetElementPtrInst(ptr_aconsts, ptr_arraydecay_i3_2_indices.begin(), ptr_arraydecay_i3_2_indices.end(), "arraydecay.i3.2", label_forbody_i11_preheader);
-    LoadInst* float_tmp5_i_2 = new LoadInst(ptr_arraydecay_i3_2, "tmp5.i.2", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp7_i4_2 = new InsertElementInst(packed_tmp31_i_1, float_tmp5_i_2, const_int32_26, "tmp7.i4.2", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx12_i_2_indices;
-    ptr_arrayidx12_i_2_indices.push_back(const_int32_29);
-    ptr_arrayidx12_i_2_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx12_i_2 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx12_i_2_indices.begin(), ptr_arrayidx12_i_2_indices.end(), "arrayidx12.i.2", label_forbody_i11_preheader);
-    LoadInst* float_tmp13_i5_2 = new LoadInst(ptr_arrayidx12_i_2, "tmp13.i5.2", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp15_i6_2 = new InsertElementInst(packed_tmp7_i4_2, float_tmp13_i5_2, const_int32_28, "tmp15.i6.2", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx20_i_2_indices;
-    ptr_arrayidx20_i_2_indices.push_back(const_int32_29);
-    ptr_arrayidx20_i_2_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx20_i_2 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx20_i_2_indices.begin(), ptr_arrayidx20_i_2_indices.end(), "arrayidx20.i.2", label_forbody_i11_preheader);
-    LoadInst* float_tmp21_i_2 = new LoadInst(ptr_arrayidx20_i_2, "tmp21.i.2", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp23_i_2 = new InsertElementInst(packed_tmp15_i6_2, float_tmp21_i_2, const_int32_29, "tmp23.i.2", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx28_i7_2_indices;
-    ptr_arrayidx28_i7_2_indices.push_back(const_int32_29);
-    ptr_arrayidx28_i7_2_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx28_i7_2 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx28_i7_2_indices.begin(), ptr_arrayidx28_i7_2_indices.end(), "arrayidx28.i7.2", label_forbody_i11_preheader);
-    LoadInst* float_tmp29_i_2 = new LoadInst(ptr_arrayidx28_i7_2, "tmp29.i.2", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp31_i_2 = new InsertElementInst(packed_tmp23_i_2, float_tmp29_i_2, const_int32_30, "tmp31.i.2", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx34_i8_2_indices;
-    ptr_arrayidx34_i8_2_indices.push_back(const_int32_26);
-    ptr_arrayidx34_i8_2_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx34_i8_2 = new GetElementPtrInst(ptr_consts, ptr_arrayidx34_i8_2_indices.begin(), ptr_arrayidx34_i8_2_indices.end(), "arrayidx34.i8.2", label_forbody_i11_preheader);
-    new StoreInst(packed_tmp31_i_2, ptr_arrayidx34_i8_2, false, label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arraydecay_i3_3_indices;
-    ptr_arraydecay_i3_3_indices.push_back(const_int32_30);
-    ptr_arraydecay_i3_3_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay_i3_3 = new GetElementPtrInst(ptr_aconsts, ptr_arraydecay_i3_3_indices.begin(), ptr_arraydecay_i3_3_indices.end(), "arraydecay.i3.3", label_forbody_i11_preheader);
-    LoadInst* float_tmp5_i_3 = new LoadInst(ptr_arraydecay_i3_3, "tmp5.i.3", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp7_i4_3 = new InsertElementInst(packed_tmp31_i_2, float_tmp5_i_3, const_int32_26, "tmp7.i4.3", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx12_i_3_indices;
-    ptr_arrayidx12_i_3_indices.push_back(const_int32_30);
-    ptr_arrayidx12_i_3_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx12_i_3 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx12_i_3_indices.begin(), ptr_arrayidx12_i_3_indices.end(), "arrayidx12.i.3", label_forbody_i11_preheader);
-    LoadInst* float_tmp13_i5_3 = new LoadInst(ptr_arrayidx12_i_3, "tmp13.i5.3", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp15_i6_3 = new InsertElementInst(packed_tmp7_i4_3, float_tmp13_i5_3, const_int32_28, "tmp15.i6.3", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx20_i_3_indices;
-    ptr_arrayidx20_i_3_indices.push_back(const_int32_30);
-    ptr_arrayidx20_i_3_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx20_i_3 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx20_i_3_indices.begin(), ptr_arrayidx20_i_3_indices.end(), "arrayidx20.i.3", label_forbody_i11_preheader);
-    LoadInst* float_tmp21_i_3 = new LoadInst(ptr_arrayidx20_i_3, "tmp21.i.3", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp23_i_3 = new InsertElementInst(packed_tmp15_i6_3, float_tmp21_i_3, const_int32_29, "tmp23.i.3", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx28_i7_3_indices;
-    ptr_arrayidx28_i7_3_indices.push_back(const_int32_30);
-    ptr_arrayidx28_i7_3_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx28_i7_3 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx28_i7_3_indices.begin(), ptr_arrayidx28_i7_3_indices.end(), "arrayidx28.i7.3", label_forbody_i11_preheader);
-    LoadInst* float_tmp29_i_3 = new LoadInst(ptr_arrayidx28_i7_3, "tmp29.i.3", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp31_i_3 = new InsertElementInst(packed_tmp23_i_3, float_tmp29_i_3, const_int32_30, "tmp31.i.3", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx34_i8_3_indices;
-    ptr_arrayidx34_i8_3_indices.push_back(const_int32_26);
-    ptr_arrayidx34_i8_3_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx34_i8_3 = new GetElementPtrInst(ptr_consts, ptr_arrayidx34_i8_3_indices.begin(), ptr_arrayidx34_i8_3_indices.end(), "arrayidx34.i8.3", label_forbody_i11_preheader);
-    new StoreInst(packed_tmp31_i_3, ptr_arrayidx34_i8_3, false, label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arraydecay_i3_4_indices;
-    ptr_arraydecay_i3_4_indices.push_back(const_int32_33);
-    ptr_arraydecay_i3_4_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay_i3_4 = new GetElementPtrInst(ptr_aconsts, ptr_arraydecay_i3_4_indices.begin(), ptr_arraydecay_i3_4_indices.end(), "arraydecay.i3.4", label_forbody_i11_preheader);
-    LoadInst* float_tmp5_i_4 = new LoadInst(ptr_arraydecay_i3_4, "tmp5.i.4", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp7_i4_4 = new InsertElementInst(packed_tmp31_i_3, float_tmp5_i_4, const_int32_26, "tmp7.i4.4", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx12_i_4_indices;
-    ptr_arrayidx12_i_4_indices.push_back(const_int32_33);
-    ptr_arrayidx12_i_4_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx12_i_4 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx12_i_4_indices.begin(), ptr_arrayidx12_i_4_indices.end(), "arrayidx12.i.4", label_forbody_i11_preheader);
-    LoadInst* float_tmp13_i5_4 = new LoadInst(ptr_arrayidx12_i_4, "tmp13.i5.4", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp15_i6_4 = new InsertElementInst(packed_tmp7_i4_4, float_tmp13_i5_4, const_int32_28, "tmp15.i6.4", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx20_i_4_indices;
-    ptr_arrayidx20_i_4_indices.push_back(const_int32_33);
-    ptr_arrayidx20_i_4_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx20_i_4 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx20_i_4_indices.begin(), ptr_arrayidx20_i_4_indices.end(), "arrayidx20.i.4", label_forbody_i11_preheader);
-    LoadInst* float_tmp21_i_4 = new LoadInst(ptr_arrayidx20_i_4, "tmp21.i.4", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp23_i_4 = new InsertElementInst(packed_tmp15_i6_4, float_tmp21_i_4, const_int32_29, "tmp23.i.4", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx28_i7_4_indices;
-    ptr_arrayidx28_i7_4_indices.push_back(const_int32_33);
-    ptr_arrayidx28_i7_4_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx28_i7_4 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx28_i7_4_indices.begin(), ptr_arrayidx28_i7_4_indices.end(), "arrayidx28.i7.4", label_forbody_i11_preheader);
-    LoadInst* float_tmp29_i_4 = new LoadInst(ptr_arrayidx28_i7_4, "tmp29.i.4", false, label_forbody_i11_preheader);
-    InsertElementInst* packed_tmp31_i_4 = new InsertElementInst(packed_tmp23_i_4, float_tmp29_i_4, const_int32_30, "tmp31.i.4", label_forbody_i11_preheader);
-    std::vector<Value*> ptr_arrayidx34_i8_4_indices;
-    ptr_arrayidx34_i8_4_indices.push_back(const_int32_26);
-    ptr_arrayidx34_i8_4_indices.push_back(const_int32_33);
-    Instruction* ptr_arrayidx34_i8_4 = new GetElementPtrInst(ptr_consts, ptr_arrayidx34_i8_4_indices.begin(), ptr_arrayidx34_i8_4_indices.end(), "arrayidx34.i8.4", label_forbody_i11_preheader);
-    new StoreInst(packed_tmp31_i_4, ptr_arrayidx34_i8_4, false, label_forbody_i11_preheader);
-    new BranchInst(label_forbody_preheader, label_afterfor_82, int1_cmp_i, label_forbody_i11_preheader);
+    ptr_arrayidx20_i_indices.push_back(int32_i_02_0_i);
+    ptr_arrayidx20_i_indices.push_back(const_int32_39);
+    Instruction* ptr_arrayidx20_i = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx20_i_indices.begin(), ptr_arrayidx20_i_indices.end(), "arrayidx20.i", label_forbody_i15);
+    LoadInst* float_tmp21_i = new LoadInst(ptr_arrayidx20_i, "tmp21.i", false, label_forbody_i15);
+    InsertElementInst* packed_tmp23_i = new InsertElementInst(packed_tmp15_i9, float_tmp21_i, const_int32_39, "tmp23.i", label_forbody_i15);
+    std::vector<Value*> ptr_arrayidx28_i10_indices;
+    ptr_arrayidx28_i10_indices.push_back(int32_i_02_0_i);
+    ptr_arrayidx28_i10_indices.push_back(const_int32_40);
+    Instruction* ptr_arrayidx28_i10 = new GetElementPtrInst(ptr_aconsts, ptr_arrayidx28_i10_indices.begin(), ptr_arrayidx28_i10_indices.end(), "arrayidx28.i10", label_forbody_i15);
+    LoadInst* float_tmp29_i = new LoadInst(ptr_arrayidx28_i10, "tmp29.i", false, label_forbody_i15);
+    InsertElementInst* packed_tmp31_i = new InsertElementInst(packed_tmp23_i, float_tmp29_i, const_int32_40, "tmp31.i", label_forbody_i15);
+    CastInst* double_conv_i11 = new FPExtInst(float_tmp5_i, Type::DoubleTy, "conv.i11", label_forbody_i15);
+    CastInst* double_conv37_i = new FPExtInst(float_tmp13_i8, Type::DoubleTy, "conv37.i", label_forbody_i15);
+    CastInst* double_conv40_i = new FPExtInst(float_tmp21_i, Type::DoubleTy, "conv40.i", label_forbody_i15);
+    CastInst* double_conv43_i = new FPExtInst(float_tmp29_i, Type::DoubleTy, "conv43.i", label_forbody_i15);
+    std::vector<Value*> int32_call_i12_params;
+    int32_call_i12_params.push_back(const_ptr_42);
+    int32_call_i12_params.push_back(int32_i_02_0_i);
+    int32_call_i12_params.push_back(double_conv_i11);
+    int32_call_i12_params.push_back(double_conv37_i);
+    int32_call_i12_params.push_back(double_conv40_i);
+    int32_call_i12_params.push_back(double_conv43_i);
+    CallInst* int32_call_i12 = new CallInst(func_printf, int32_call_i12_params.begin(), int32_call_i12_params.end(), "call.i12", label_forbody_i15);
+    int32_call_i12->setCallingConv(CallingConv::C);
+    int32_call_i12->setTailCall(false);
+    std::vector<Value*> ptr_arrayidx46_i_indices;
+    ptr_arrayidx46_i_indices.push_back(const_int32_36);
+    ptr_arrayidx46_i_indices.push_back(int32_i_02_0_i);
+    Instruction* ptr_arrayidx46_i = new GetElementPtrInst(ptr_consts, ptr_arrayidx46_i_indices.begin(), ptr_arrayidx46_i_indices.end(), "arrayidx46.i", label_forbody_i15);
+    StoreInst* void_110 = new StoreInst(packed_tmp31_i, ptr_arrayidx46_i, false, label_forbody_i15);
+    BinaryOperator* int32_indvar_next25 = BinaryOperator::create(Instruction::Add, int32_i_02_0_i, const_int32_38, "indvar.next25", label_forbody_i15);
+    ICmpInst* int1_exitcond26 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next25, int32_num_consts, "exitcond26", label_forbody_i15);
+    new BranchInst(label_from_consts_exit, label_forbody_i15, int1_exitcond26, label_forbody_i15);
+    
+    // Block from_consts.exit (label_from_consts_exit)
+    CallInst* int32_call9 = new CallInst(func_printf, const_ptr_45, "call9", label_from_consts_exit);
+    int32_call9->setCallingConv(CallingConv::C);
+    int32_call9->setTailCall(false);
+    new BranchInst(label_forbody_preheader, label_afterfor_97, int1_cmp_i, label_from_consts_exit);
     
     // Block forbody.preheader (label_forbody_preheader)
-    new BranchInst(label_forbody_us, label_forbody_81, int1_cmp5_i, label_forbody_preheader);
+    std::vector<Value*> ptr_arraydecay21_indices;
+    ptr_arraydecay21_indices.push_back(const_int32_36);
+    ptr_arraydecay21_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay21 = new GetElementPtrInst(ptr_consts, ptr_arraydecay21_indices.begin(), ptr_arraydecay21_indices.end(), "arraydecay21", label_forbody_preheader);
+    ICmpInst* int1_cmp_i1 = new ICmpInst(ICmpInst::ICMP_SGT, int32_num_attribs_94, const_int32_36, "cmp.i1", label_forbody_preheader);
+    new BranchInst(label_forbody_us, label_forbody_96, int1_cmp_i1, label_forbody_preheader);
     
     // Block forbody.us (label_forbody_us)
-    Argument* fwdref_99 = new Argument(IntegerType::get(32));
-    PHINode* int32_i_012_0_us = new PHINode(IntegerType::get(32), "i.012.0.us", label_forbody_us);
-    int32_i_012_0_us->reserveOperandSpace(2);
-    int32_i_012_0_us->addIncoming(const_int32_26, label_forbody_preheader);
-    int32_i_012_0_us->addIncoming(fwdref_99, label_to_array_exit_us);
-    
-    std::vector<Value*> ptr_arraydecay9_us_indices;
-    ptr_arraydecay9_us_indices.push_back(const_int32_26);
-    ptr_arraydecay9_us_indices.push_back(int32_i_012_0_us);
-    ptr_arraydecay9_us_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay9_us = new GetElementPtrInst(ptr_inputs, ptr_arraydecay9_us_indices.begin(), ptr_arraydecay9_us_indices.end(), "arraydecay9.us", label_forbody_us);
-    std::vector<Value*> ptr_arraydecay13_us_indices;
-    ptr_arraydecay13_us_indices.push_back(const_int32_26);
-    ptr_arraydecay13_us_indices.push_back(int32_i_012_0_us);
-    ptr_arraydecay13_us_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay13_us = new GetElementPtrInst(ptr_results, ptr_arraydecay13_us_indices.begin(), ptr_arraydecay13_us_indices.end(), "arraydecay13.us", label_forbody_us);
-    std::vector<Value*> void_100_params;
-    void_100_params.push_back(ptr_arraydecay13_us);
-    void_100_params.push_back(ptr_arraydecay9_us);
-    void_100_params.push_back(ptr_arrayidx34_i8);
-    CallInst* void_100 = new CallInst(func_execute_shader, void_100_params.begin(), void_100_params.end(), "", label_forbody_us);
-    void_100->setCallingConv(CallingConv::C);
-    void_100->setTailCall(false);
+    Argument* fwdref_114 = new Argument(IntegerType::get(32));
+    PHINode* int32_i_016_0_us = new PHINode(IntegerType::get(32), "i.016.0.us", label_forbody_us);
+    int32_i_016_0_us->reserveOperandSpace(2);
+    int32_i_016_0_us->addIncoming(const_int32_36, label_forbody_preheader);
+    int32_i_016_0_us->addIncoming(fwdref_114, label_to_array_exit_us);
+    
+    std::vector<Value*> ptr_arraydecay14_us_indices;
+    ptr_arraydecay14_us_indices.push_back(const_int32_36);
+    ptr_arraydecay14_us_indices.push_back(int32_i_016_0_us);
+    ptr_arraydecay14_us_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay14_us = new GetElementPtrInst(ptr_inputs, ptr_arraydecay14_us_indices.begin(), ptr_arraydecay14_us_indices.end(), "arraydecay14.us", label_forbody_us);
+    std::vector<Value*> ptr_arraydecay18_us_indices;
+    ptr_arraydecay18_us_indices.push_back(const_int32_36);
+    ptr_arraydecay18_us_indices.push_back(int32_i_016_0_us);
+    ptr_arraydecay18_us_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay18_us = new GetElementPtrInst(ptr_results, ptr_arraydecay18_us_indices.begin(), ptr_arraydecay18_us_indices.end(), "arraydecay18.us", label_forbody_us);
+    std::vector<Value*> void_115_params;
+    void_115_params.push_back(ptr_arraydecay18_us);
+    void_115_params.push_back(ptr_arraydecay14_us);
+    void_115_params.push_back(ptr_arraydecay21);
+    CallInst* void_115 = new CallInst(func_execute_shader, void_115_params.begin(), void_115_params.end(), "", label_forbody_us);
+    void_115->setCallingConv(CallingConv::C);
+    void_115->setTailCall(false);
+    CallInst* int32_call22_us = new CallInst(func_printf, const_ptr_46, "call22.us", label_forbody_us);
+    int32_call22_us->setCallingConv(CallingConv::C);
+    int32_call22_us->setTailCall(false);
     new BranchInst(label_forbody_i_us, label_forbody_us);
     
     // Block to_array.exit.us (label_to_array_exit_us)
-    BinaryOperator* int32_inc_us = BinaryOperator::create(Instruction::Add, int32_i_012_0_us, const_int32_28, "inc.us", label_to_array_exit_us);
-    ICmpInst* int1_cmp18_us = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc_us, int32_count_78, "cmp18.us", label_to_array_exit_us);
-    new BranchInst(label_forbody_us, label_afterfor_82, int1_cmp18_us, label_to_array_exit_us);
+    BinaryOperator* int32_inc_us = BinaryOperator::create(Instruction::Add, int32_i_016_0_us, const_int32_38, "inc.us", label_to_array_exit_us);
+    ICmpInst* int1_cmp22_us = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc_us, int32_num_vertices, "cmp22.us", label_to_array_exit_us);
+    new BranchInst(label_forbody_us, label_afterfor_97, int1_cmp22_us, label_to_array_exit_us);
     
     // Block forbody.i.us (label_forbody_i_us)
-    Argument* fwdref_103 = new Argument(IntegerType::get(32));
+    Argument* fwdref_118 = new Argument(IntegerType::get(32));
     PHINode* int32_i_01_0_i_us = new PHINode(IntegerType::get(32), "i.01.0.i.us", label_forbody_i_us);
     int32_i_01_0_i_us->reserveOperandSpace(2);
-    int32_i_01_0_i_us->addIncoming(const_int32_26, label_forbody_us);
-    int32_i_01_0_i_us->addIncoming(fwdref_103, label_forbody_i_us);
+    int32_i_01_0_i_us->addIncoming(const_int32_36, label_forbody_us);
+    int32_i_01_0_i_us->addIncoming(fwdref_118, label_forbody_i_us);
     
     std::vector<Value*> ptr_arraydecay_i_us_indices;
-    ptr_arraydecay_i_us_indices.push_back(int32_i_012_0_us);
+    ptr_arraydecay_i_us_indices.push_back(int32_i_016_0_us);
     ptr_arraydecay_i_us_indices.push_back(int32_i_01_0_i_us);
-    ptr_arraydecay_i_us_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay_i_us = new GetElementPtrInst(ptr_dests_77, ptr_arraydecay_i_us_indices.begin(), ptr_arraydecay_i_us_indices.end(), "arraydecay.i.us", label_forbody_i_us);
+    ptr_arraydecay_i_us_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay_i_us = new GetElementPtrInst(ptr_dests_93, ptr_arraydecay_i_us_indices.begin(), ptr_arraydecay_i_us_indices.end(), "arraydecay.i.us", label_forbody_i_us);
     std::vector<Value*> ptr_arrayidx6_i_us_indices;
-    ptr_arrayidx6_i_us_indices.push_back(const_int32_26);
-    ptr_arrayidx6_i_us_indices.push_back(int32_i_012_0_us);
+    ptr_arrayidx6_i_us_indices.push_back(const_int32_36);
+    ptr_arrayidx6_i_us_indices.push_back(int32_i_016_0_us);
     ptr_arrayidx6_i_us_indices.push_back(int32_i_01_0_i_us);
     Instruction* ptr_arrayidx6_i_us = new GetElementPtrInst(ptr_results, ptr_arrayidx6_i_us_indices.begin(), ptr_arrayidx6_i_us_indices.end(), "arrayidx6.i.us", label_forbody_i_us);
     LoadInst* packed_tmp7_i_us = new LoadInst(ptr_arrayidx6_i_us, "tmp7.i.us", false, label_forbody_i_us);
-    ExtractElementInst* float_tmp9_i_us = new ExtractElementInst(packed_tmp7_i_us, const_int32_26, "tmp9.i.us", label_forbody_i_us);
-    CastInst* double_conv_i_us = new FPExtInst(float_tmp9_i_us, Type::DoubleTy, "conv.i.us", label_forbody_i_us);
-    ExtractElementInst* float_tmp11_i_us = new ExtractElementInst(packed_tmp7_i_us, const_int32_28, "tmp11.i.us", label_forbody_i_us);
+    ExtractElementInst* float_tmp9_i_us = new ExtractElementInst(packed_tmp7_i_us, const_int32_36, "tmp9.i.us", label_forbody_i_us);
+    CastInst* double_conv_i2_us = new FPExtInst(float_tmp9_i_us, Type::DoubleTy, "conv.i2.us", label_forbody_i_us);
+    ExtractElementInst* float_tmp11_i_us = new ExtractElementInst(packed_tmp7_i_us, const_int32_38, "tmp11.i.us", label_forbody_i_us);
     CastInst* double_conv12_i_us = new FPExtInst(float_tmp11_i_us, Type::DoubleTy, "conv12.i.us", label_forbody_i_us);
-    ExtractElementInst* float_tmp14_i_us = new ExtractElementInst(packed_tmp7_i_us, const_int32_29, "tmp14.i.us", label_forbody_i_us);
+    ExtractElementInst* float_tmp14_i_us = new ExtractElementInst(packed_tmp7_i_us, const_int32_39, "tmp14.i.us", label_forbody_i_us);
     CastInst* double_conv15_i_us = new FPExtInst(float_tmp14_i_us, Type::DoubleTy, "conv15.i.us", label_forbody_i_us);
-    ExtractElementInst* float_tmp17_i_us = new ExtractElementInst(packed_tmp7_i_us, const_int32_30, "tmp17.i.us", label_forbody_i_us);
+    ExtractElementInst* float_tmp17_i_us = new ExtractElementInst(packed_tmp7_i_us, const_int32_40, "tmp17.i.us", label_forbody_i_us);
     CastInst* double_conv18_i_us = new FPExtInst(float_tmp17_i_us, Type::DoubleTy, "conv18.i.us", label_forbody_i_us);
-    std::vector<Value*> int32_call_i_us_params;
-    int32_call_i_us_params.push_back(const_ptr_31);
-    int32_call_i_us_params.push_back(double_conv_i_us);
-    int32_call_i_us_params.push_back(double_conv12_i_us);
-    int32_call_i_us_params.push_back(double_conv15_i_us);
-    int32_call_i_us_params.push_back(double_conv18_i_us);
-    CallInst* int32_call_i_us = new CallInst(func_printf, int32_call_i_us_params.begin(), int32_call_i_us_params.end(), "call.i.us", label_forbody_i_us);
-    int32_call_i_us->setCallingConv(CallingConv::C);
-    int32_call_i_us->setTailCall(false);
-    new StoreInst(float_tmp9_i_us, ptr_arraydecay_i_us, false, label_forbody_i_us);
+    std::vector<Value*> int32_call_i3_us_params;
+    int32_call_i3_us_params.push_back(const_ptr_43);
+    int32_call_i3_us_params.push_back(double_conv_i2_us);
+    int32_call_i3_us_params.push_back(double_conv12_i_us);
+    int32_call_i3_us_params.push_back(double_conv15_i_us);
+    int32_call_i3_us_params.push_back(double_conv18_i_us);
+    CallInst* int32_call_i3_us = new CallInst(func_printf, int32_call_i3_us_params.begin(), int32_call_i3_us_params.end(), "call.i3.us", label_forbody_i_us);
+    int32_call_i3_us->setCallingConv(CallingConv::C);
+    int32_call_i3_us->setTailCall(false);
+    StoreInst* void_119 = new StoreInst(float_tmp9_i_us, ptr_arraydecay_i_us, false, label_forbody_i_us);
     std::vector<Value*> ptr_arrayidx24_i_us_indices;
-    ptr_arrayidx24_i_us_indices.push_back(int32_i_012_0_us);
+    ptr_arrayidx24_i_us_indices.push_back(int32_i_016_0_us);
     ptr_arrayidx24_i_us_indices.push_back(int32_i_01_0_i_us);
-    ptr_arrayidx24_i_us_indices.push_back(const_int32_28);
-    Instruction* ptr_arrayidx24_i_us = new GetElementPtrInst(ptr_dests_77, ptr_arrayidx24_i_us_indices.begin(), ptr_arrayidx24_i_us_indices.end(), "arrayidx24.i.us", label_forbody_i_us);
-    new StoreInst(float_tmp11_i_us, ptr_arrayidx24_i_us, false, label_forbody_i_us);
+    ptr_arrayidx24_i_us_indices.push_back(const_int32_38);
+    Instruction* ptr_arrayidx24_i_us = new GetElementPtrInst(ptr_dests_93, ptr_arrayidx24_i_us_indices.begin(), ptr_arrayidx24_i_us_indices.end(), "arrayidx24.i.us", label_forbody_i_us);
+    StoreInst* void_120 = new StoreInst(float_tmp11_i_us, ptr_arrayidx24_i_us, false, label_forbody_i_us);
     std::vector<Value*> ptr_arrayidx28_i_us_indices;
-    ptr_arrayidx28_i_us_indices.push_back(int32_i_012_0_us);
+    ptr_arrayidx28_i_us_indices.push_back(int32_i_016_0_us);
     ptr_arrayidx28_i_us_indices.push_back(int32_i_01_0_i_us);
-    ptr_arrayidx28_i_us_indices.push_back(const_int32_29);
-    Instruction* ptr_arrayidx28_i_us = new GetElementPtrInst(ptr_dests_77, ptr_arrayidx28_i_us_indices.begin(), ptr_arrayidx28_i_us_indices.end(), "arrayidx28.i.us", label_forbody_i_us);
-    new StoreInst(float_tmp14_i_us, ptr_arrayidx28_i_us, false, label_forbody_i_us);
+    ptr_arrayidx28_i_us_indices.push_back(const_int32_39);
+    Instruction* ptr_arrayidx28_i_us = new GetElementPtrInst(ptr_dests_93, ptr_arrayidx28_i_us_indices.begin(), ptr_arrayidx28_i_us_indices.end(), "arrayidx28.i.us", label_forbody_i_us);
+    StoreInst* void_121 = new StoreInst(float_tmp14_i_us, ptr_arrayidx28_i_us, false, label_forbody_i_us);
     std::vector<Value*> ptr_arrayidx32_i_us_indices;
-    ptr_arrayidx32_i_us_indices.push_back(int32_i_012_0_us);
+    ptr_arrayidx32_i_us_indices.push_back(int32_i_016_0_us);
     ptr_arrayidx32_i_us_indices.push_back(int32_i_01_0_i_us);
-    ptr_arrayidx32_i_us_indices.push_back(const_int32_30);
-    Instruction* ptr_arrayidx32_i_us = new GetElementPtrInst(ptr_dests_77, ptr_arrayidx32_i_us_indices.begin(), ptr_arrayidx32_i_us_indices.end(), "arrayidx32.i.us", label_forbody_i_us);
-    new StoreInst(float_tmp17_i_us, ptr_arrayidx32_i_us, false, label_forbody_i_us);
-    BinaryOperator* int32_indvar_next_108 = BinaryOperator::create(Instruction::Add, int32_i_01_0_i_us, const_int32_28, "indvar.next", label_forbody_i_us);
-    ICmpInst* int1_exitcond_109 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next_108, int32_num_attribs_79, "exitcond", label_forbody_i_us);
-    new BranchInst(label_to_array_exit_us, label_forbody_i_us, int1_exitcond_109, label_forbody_i_us);
-    
-    // Block forbody (label_forbody_81)
-    Argument* fwdref_111 = new Argument(IntegerType::get(32));
-    PHINode* int32_i_012_0 = new PHINode(IntegerType::get(32), "i.012.0", label_forbody_81);
-    int32_i_012_0->reserveOperandSpace(2);
-    int32_i_012_0->addIncoming(const_int32_26, label_forbody_preheader);
-    int32_i_012_0->addIncoming(fwdref_111, label_forbody_81);
-    
-    std::vector<Value*> ptr_arraydecay9_indices;
-    ptr_arraydecay9_indices.push_back(const_int32_26);
-    ptr_arraydecay9_indices.push_back(int32_i_012_0);
-    ptr_arraydecay9_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay9 = new GetElementPtrInst(ptr_inputs, ptr_arraydecay9_indices.begin(), ptr_arraydecay9_indices.end(), "arraydecay9", label_forbody_81);
-    std::vector<Value*> ptr_arraydecay13_indices;
-    ptr_arraydecay13_indices.push_back(const_int32_26);
-    ptr_arraydecay13_indices.push_back(int32_i_012_0);
-    ptr_arraydecay13_indices.push_back(const_int32_26);
-    Instruction* ptr_arraydecay13 = new GetElementPtrInst(ptr_results, ptr_arraydecay13_indices.begin(), ptr_arraydecay13_indices.end(), "arraydecay13", label_forbody_81);
-    std::vector<Value*> void_112_params;
-    void_112_params.push_back(ptr_arraydecay13);
-    void_112_params.push_back(ptr_arraydecay9);
-    void_112_params.push_back(ptr_arrayidx34_i8);
-    CallInst* void_112 = new CallInst(func_execute_shader, void_112_params.begin(), void_112_params.end(), "", label_forbody_81);
-    void_112->setCallingConv(CallingConv::C);
-    void_112->setTailCall(false);
-    BinaryOperator* int32_inc_113 = BinaryOperator::create(Instruction::Add, int32_i_012_0, const_int32_28, "inc", label_forbody_81);
-    ICmpInst* int1_cmp18 = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc_113, int32_count_78, "cmp18", label_forbody_81);
-    new BranchInst(label_forbody_81, label_afterfor_82, int1_cmp18, label_forbody_81);
-    
-    // Block afterfor (label_afterfor_82)
-    new ReturnInst(label_afterfor_82);
+    ptr_arrayidx32_i_us_indices.push_back(const_int32_40);
+    Instruction* ptr_arrayidx32_i_us = new GetElementPtrInst(ptr_dests_93, ptr_arrayidx32_i_us_indices.begin(), ptr_arrayidx32_i_us_indices.end(), "arrayidx32.i.us", label_forbody_i_us);
+    StoreInst* void_122 = new StoreInst(float_tmp17_i_us, ptr_arrayidx32_i_us, false, label_forbody_i_us);
+    BinaryOperator* int32_indvar_next_123 = BinaryOperator::create(Instruction::Add, int32_i_01_0_i_us, const_int32_38, "indvar.next", label_forbody_i_us);
+    ICmpInst* int1_exitcond_124 = new ICmpInst(ICmpInst::ICMP_EQ, int32_indvar_next_123, int32_num_attribs_94, "exitcond", label_forbody_i_us);
+    new BranchInst(label_to_array_exit_us, label_forbody_i_us, int1_exitcond_124, label_forbody_i_us);
+    
+    // Block forbody (label_forbody_96)
+    Argument* fwdref_126 = new Argument(IntegerType::get(32));
+    PHINode* int32_i_016_0 = new PHINode(IntegerType::get(32), "i.016.0", label_forbody_96);
+    int32_i_016_0->reserveOperandSpace(2);
+    int32_i_016_0->addIncoming(const_int32_36, label_forbody_preheader);
+    int32_i_016_0->addIncoming(fwdref_126, label_forbody_96);
+    
+    std::vector<Value*> ptr_arraydecay14_indices;
+    ptr_arraydecay14_indices.push_back(const_int32_36);
+    ptr_arraydecay14_indices.push_back(int32_i_016_0);
+    ptr_arraydecay14_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay14 = new GetElementPtrInst(ptr_inputs, ptr_arraydecay14_indices.begin(), ptr_arraydecay14_indices.end(), "arraydecay14", label_forbody_96);
+    std::vector<Value*> ptr_arraydecay18_indices;
+    ptr_arraydecay18_indices.push_back(const_int32_36);
+    ptr_arraydecay18_indices.push_back(int32_i_016_0);
+    ptr_arraydecay18_indices.push_back(const_int32_36);
+    Instruction* ptr_arraydecay18 = new GetElementPtrInst(ptr_results, ptr_arraydecay18_indices.begin(), ptr_arraydecay18_indices.end(), "arraydecay18", label_forbody_96);
+    std::vector<Value*> void_127_params;
+    void_127_params.push_back(ptr_arraydecay18);
+    void_127_params.push_back(ptr_arraydecay14);
+    void_127_params.push_back(ptr_arraydecay21);
+    CallInst* void_127 = new CallInst(func_execute_shader, void_127_params.begin(), void_127_params.end(), "", label_forbody_96);
+    void_127->setCallingConv(CallingConv::C);
+    void_127->setTailCall(false);
+    CallInst* int32_call22 = new CallInst(func_printf, const_ptr_46, "call22", label_forbody_96);
+    int32_call22->setCallingConv(CallingConv::C);
+    int32_call22->setTailCall(false);
+    BinaryOperator* int32_inc_128 = BinaryOperator::create(Instruction::Add, int32_i_016_0, const_int32_38, "inc", label_forbody_96);
+    ICmpInst* int1_cmp22 = new ICmpInst(ICmpInst::ICMP_SLT, int32_inc_128, int32_num_vertices, "cmp22", label_forbody_96);
+    new BranchInst(label_forbody_96, label_afterfor_97, int1_cmp22, label_forbody_96);
+    
+    // Block afterfor (label_afterfor_97)
+    new ReturnInst(label_afterfor_97);
     
     // Resolve Forward References
-    fwdref_86->replaceAllUsesWith(int32_i_013_0_i_ph); delete fwdref_86;
-    fwdref_88->replaceAllUsesWith(packed_tmp48_i); delete fwdref_88;
-    fwdref_87->replaceAllUsesWith(int32_inc_i); delete fwdref_87;
-    fwdref_85->replaceAllUsesWith(int32_inc59_i); delete fwdref_85;
-    fwdref_99->replaceAllUsesWith(int32_inc_us); delete fwdref_99;
-    fwdref_103->replaceAllUsesWith(int32_indvar_next_108); delete fwdref_103;
-    fwdref_111->replaceAllUsesWith(int32_inc_113); delete fwdref_111;
+    fwdref_101->replaceAllUsesWith(int32_i_014_0_i_ph); delete fwdref_101;
+    fwdref_103->replaceAllUsesWith(packed_tmp48_i); delete fwdref_103;
+    fwdref_102->replaceAllUsesWith(int32_inc_i); delete fwdref_102;
+    fwdref_100->replaceAllUsesWith(int32_inc72_i); delete fwdref_100;
+    fwdref_109->replaceAllUsesWith(packed_tmp31_i); delete fwdref_109;
+    fwdref_108->replaceAllUsesWith(int32_indvar_next25); delete fwdref_108;
+    fwdref_114->replaceAllUsesWith(int32_inc_us); delete fwdref_114;
+    fwdref_118->replaceAllUsesWith(int32_indvar_next_123); delete fwdref_118;
+    fwdref_126->replaceAllUsesWith(int32_inc_128); delete fwdref_126;
     
   }
   
diff --git a/src/mesa/pipe/llvm/llvm_builtins.c b/src/mesa/pipe/llvm/llvm_builtins.c
index e135b4781d..82ae59446e 100644
--- a/src/mesa/pipe/llvm/llvm_builtins.c
+++ b/src/mesa/pipe/llvm/llvm_builtins.c
@@ -102,14 +102,14 @@ void from_array(float4 (*res)[16], float (*ainputs)[16][4],
          vec.y = ainputs[i][j][1];
          vec.z = ainputs[i][j][2];
          vec.w = ainputs[i][j][3];
-         //printf("FAR(%d %d) %f %f %f %f\n", i, j, vec.x, vec.y, vec.z, vec.w);
+         printf("FRA(%d %d) %f %f %f %f\n", i, j, vec.x, vec.y, vec.z, vec.w);
          res[i][j] = vec;
       }
    }
 }
 
 void from_consts(float4 *res, float (*ainputs)[4],
-                int count)
+                 int count)
 {
    for (int i = 0; i < count; ++i) {
       float4 vec;
@@ -117,7 +117,7 @@ void from_consts(float4 *res, float (*ainputs)[4],
       vec.y = ainputs[i][1];
       vec.z = ainputs[i][2];
       vec.w = ainputs[i][3];
-      //printf("VCONST %f %f %f %f\n", vec.x, vec.y, vec.z, vec.w);
+      printf("VCONST(%d) %f %f %f %f\n", i, vec.x, vec.y, vec.z, vec.w);
       res[i] = vec;
    }
 }
@@ -141,20 +141,25 @@ extern void execute_shader(float4 *dests, float4 *inputs,
 void run_vertex_shader(float (*ainputs)[16][4],
                        float (*dests)[16][4],
                        float (*aconsts)[4],
-                       int count,
-                       int num_attribs)
+                       int num_vertices,
+                       int num_inputs,
+                       int num_attribs,
+                       int num_consts)
 {
    float4  inputs[16*32*4][16];
-   float4  consts[16];
+   float4  consts[32];
    float4  results[16*32*4][16];
 
-   printf("XXXXXXXXXXX run_vertex_shader\n");
-   from_array(inputs, ainputs, count, num_attribs);
-   from_consts(consts, aconsts, 5);
-   for (int i = 0; i < count; ++i) {
+   printf("XXX LLVM run_vertex_shader vertices = %d, inputs = %d, consts = %d\n",
+          num_vertices, num_inputs, num_consts);
+   from_array(inputs, ainputs, num_vertices, num_inputs);
+   from_consts(consts, aconsts, num_consts);
+   printf("               after conversion\n");
+   for (int i = 0; i < num_vertices; ++i) {
       float4 *in  = inputs[i];
       float4 *res = results[i];
       execute_shader(res, in, consts);
+      printf("after executing shader\n");
       to_array(dests[i], res, num_attribs);
    }
 }
diff --git a/src/mesa/pipe/llvm/llvmtgsi.cpp b/src/mesa/pipe/llvm/llvmtgsi.cpp
index 9b4a5f024e..2a86a4e442 100644
--- a/src/mesa/pipe/llvm/llvmtgsi.cpp
+++ b/src/mesa/pipe/llvm/llvmtgsi.cpp
@@ -125,6 +125,10 @@ translate_instruction(llvm::Module *module,
                       struct tgsi_full_instruction *fi)
 {
    llvm::Value *inputs[4];
+   inputs[0] = 0;
+   inputs[1] = 0;
+   inputs[2] = 0;
+   inputs[3] = 0;
    printf("translate instr START\n");
    for (int i = 0; i < inst->Instruction.NumSrcRegs; ++i) {
       struct tgsi_full_src_register *src = &inst->FullSrcRegisters[i];
@@ -168,7 +172,10 @@ translate_instruction(llvm::Module *module,
       inputs[i] = val;
    }
 
+   if (inputs[0])
+      instr->printVector(inputs[0]);
    llvm::Value *out = 0;
+   printf("Opcode is %d\n", inst->Instruction.Opcode);
    switch (inst->Instruction.Opcode) {
    case TGSI_OPCODE_ARL:
       break;
@@ -512,32 +519,19 @@ translate_instruction(llvm::Module *module,
 
       if (dst->DstRegister.File == TGSI_FILE_OUTPUT) {
          printf("--- storing to %d %p\n", dst->DstRegister.Index, out);
-         storage->store(dst->DstRegister.Index, out);
+         storage->store(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
       } else if (dst->DstRegister.File == TGSI_FILE_TEMPORARY) {
-         storage->setTempElement(dst->DstRegister.Index, out);
+         storage->setTempElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
       } else {
          fprintf(stderr, "ERROR: unsupported LLVM destination!");
       }
-
-#if 0
-      if (dst->DstRegister.WriteMask != TGSI_WRITEMASK_XYZW) {
-         if (dst->DstRegister.WriteMask & TGSI_WRITEMASK_X) {
-         }
-         if (dst->DstRegister.WriteMask & TGSI_WRITEMASK_Y) {
-         }
-         if (dst->DstRegister.WriteMask & TGSI_WRITEMASK_Z) {
-         }
-         if (dst->DstRegister.WriteMask & TGSI_WRITEMASK_W) {
-         }
-      }
-#endif
    }
    printf("translate instr END\n");
 }
 
 
 static llvm::Module *
-tgsi_to_llvm(const struct tgsi_token *tokens)
+tgsi_to_llvm(struct ga_llvm_prog *prog, const struct tgsi_token *tokens)
 {
    llvm::Module *mod = createBaseShader();
    struct tgsi_parse_context parse;
@@ -572,7 +566,7 @@ tgsi_to_llvm(const struct tgsi_token *tokens)
    fi = tgsi_default_full_instruction();
    fd = tgsi_default_full_declaration();
    Storage storage(label_entry, ptr_OUT, ptr_IN, ptr_CONST);
-   Instructions instr(mod, label_entry);
+   Instructions instr(mod, shader, label_entry);
    while(!tgsi_parse_end_of_tokens(&parse)) {
       tgsi_parse_token(&parse);
 
@@ -601,178 +595,14 @@ tgsi_to_llvm(const struct tgsi_token *tokens)
       }
    }
 
-#if 0
-   // Type Definitions
-   ArrayType* ArrayTy_0 = ArrayType::get(IntegerType::get(8), 19);
-
-   PointerType* PointerTy_1 = PointerType::get(ArrayTy_0);
-
-   VectorType* VectorTy_4 = VectorType::get(Type::FloatTy, 4);
-
-   PointerType* PointerTy_3 = PointerType::get(VectorTy_4);
-
-
-   VectorType* VectorTy_5 = VectorType::get(IntegerType::get(32), 4);
-
-
-   PointerType* PointerTy_8 = PointerType::get(IntegerType::get(8));
-
-
-   // Global Variable Declarations
-
-
-   GlobalVariable* gvar_array__str = new GlobalVariable(
-      /*Type=*/ArrayTy_0,
-      /*isConstant=*/true,
-      /*Linkage=*/GlobalValue::InternalLinkage,
-      /*Initializer=*/0, // has initializer, specified below
-      /*Name=*/".str",
-      mod);
-
-   GlobalVariable* gvar_array__str1 = new GlobalVariable(
-      /*Type=*/ArrayTy_0,
-      /*isConstant=*/true,
-      /*Linkage=*/GlobalValue::InternalLinkage,
-      /*Initializer=*/0, // has initializer, specified below
-      /*Name=*/".str1",
-      mod);
-
-   GlobalVariable* gvar_array__str2 = new GlobalVariable(
-      /*Type=*/ArrayTy_0,
-      /*isConstant=*/true,
-      /*Linkage=*/GlobalValue::InternalLinkage,
-      /*Initializer=*/0, // has initializer, specified below
-      /*Name=*/".str2",
-      mod);
-
-   // Constant Definitions
-   Constant* const_array_9 = ConstantArray::get("const %f %f %f %f\x0A", true);
-   Constant* const_array_10 = ConstantArray::get("resul %f %f %f %f\x0A", true);
-   Constant* const_array_11 = ConstantArray::get("outpu %f %f %f %f\x0A", true);
-   UndefValue* const_packed_12 = UndefValue::get(VectorTy_4);
-   Constant* const_packed_13 = Constant::getNullValue(VectorTy_5);
-   std::vector<Constant*> const_packed_14_elems;
-   ConstantInt* const_int32_15 = ConstantInt::get(APInt(32,  "1", 10));
-   const_packed_14_elems.push_back(const_int32_15);
-   const_packed_14_elems.push_back(const_int32_15);
-   const_packed_14_elems.push_back(const_int32_15);
-   const_packed_14_elems.push_back(const_int32_15);
-   Constant* const_packed_14 = ConstantVector::get(VectorTy_5, const_packed_14_elems);
-   std::vector<Constant*> const_packed_16_elems;
-   ConstantInt* const_int32_17 = ConstantInt::get(APInt(32,  "2", 10));
-   const_packed_16_elems.push_back(const_int32_17);
-   const_packed_16_elems.push_back(const_int32_17);
-   const_packed_16_elems.push_back(const_int32_17);
-   const_packed_16_elems.push_back(const_int32_17);
-   Constant* const_packed_16 = ConstantVector::get(VectorTy_5, const_packed_16_elems);
-   std::vector<Constant*> const_packed_18_elems;
-   ConstantInt* const_int32_19 = ConstantInt::get(APInt(32,  "3", 10));
-   const_packed_18_elems.push_back(const_int32_19);
-   const_packed_18_elems.push_back(const_int32_19);
-   const_packed_18_elems.push_back(const_int32_19);
-   const_packed_18_elems.push_back(const_int32_19);
-   Constant* const_packed_18 = ConstantVector::get(VectorTy_5, const_packed_18_elems);
-   std::vector<Constant*> const_ptr_20_indices;
-   Constant* const_int32_21 = Constant::getNullValue(IntegerType::get(32));
-   const_ptr_20_indices.push_back(const_int32_21);
-   const_ptr_20_indices.push_back(const_int32_21);
-   Constant* const_ptr_20 = ConstantExpr::getGetElementPtr(gvar_array__str, &const_ptr_20_indices[0], const_ptr_20_indices.size() );
-   UndefValue* const_double_22 = UndefValue::get(Type::DoubleTy);
-   std::vector<Constant*> const_ptr_23_indices;
-   const_ptr_23_indices.push_back(const_int32_21);
-   const_ptr_23_indices.push_back(const_int32_21);
-   Constant* const_ptr_23 = ConstantExpr::getGetElementPtr(gvar_array__str1, &const_ptr_23_indices[0], const_ptr_23_indices.size() );
-   std::vector<Constant*> const_ptr_24_indices;
-   const_ptr_24_indices.push_back(const_int32_21);
-   const_ptr_24_indices.push_back(const_int32_21);
-   Constant* const_ptr_24 = ConstantExpr::getGetElementPtr(gvar_array__str2, &const_ptr_24_indices[0], const_ptr_24_indices.size() );
-
-   // Global Variable Definitions
-   gvar_array__str->setInitializer(const_array_9);
-   gvar_array__str1->setInitializer(const_array_10);
-   gvar_array__str2->setInitializer(const_array_11);
-
-   // Function Definitions
-
-   // Function: execute_shader (func_execute_shader)
-   {
-      // Block entry (label_entry)
-      LoadInst* packed_tmp1 = new LoadInst(ptr_IN, "tmp1", false, label_entry);
-      ShuffleVectorInst* packed_tmp3 = new ShuffleVectorInst(packed_tmp1, const_packed_12, const_packed_13, "tmp3", label_entry);
-      LoadInst* packed_tmp6 = new LoadInst(ptr_CONST, "tmp6", false, label_entry);
-      BinaryOperator* packed_mul = BinaryOperator::create(Instruction::Mul, packed_tmp3, packed_tmp6, "mul", label_entry);
-      ShuffleVectorInst* packed_tmp8 = new ShuffleVectorInst(packed_tmp1, const_packed_12, const_packed_14, "tmp8", label_entry);
-      GetElementPtrInst* ptr_arrayidx10 = new GetElementPtrInst(ptr_CONST, const_int32_15, "arrayidx10", label_entry);
-      LoadInst* packed_tmp11 = new LoadInst(ptr_arrayidx10, "tmp11", false, label_entry);
-      BinaryOperator* packed_mul12 = BinaryOperator::create(Instruction::Mul, packed_tmp8, packed_tmp11, "mul12", label_entry);
-      BinaryOperator* packed_add = BinaryOperator::create(Instruction::Add, packed_mul12, packed_mul, "add", label_entry);
-      ShuffleVectorInst* packed_tmp15 = new ShuffleVectorInst(packed_tmp1, const_packed_12, const_packed_16, "tmp15", label_entry);
-      GetElementPtrInst* ptr_arrayidx17 = new GetElementPtrInst(ptr_CONST, const_int32_17, "arrayidx17", label_entry);
-      LoadInst* packed_tmp18 = new LoadInst(ptr_arrayidx17, "tmp18", false, label_entry);
-      BinaryOperator* packed_mul19 = BinaryOperator::create(Instruction::Mul, packed_tmp15, packed_tmp18, "mul19", label_entry);
-      BinaryOperator* packed_add21 = BinaryOperator::create(Instruction::Add, packed_mul19, packed_add, "add21", label_entry);
-      ShuffleVectorInst* packed_tmp25 = new ShuffleVectorInst(packed_tmp1, const_packed_12, const_packed_18, "tmp25", label_entry);
-      GetElementPtrInst* ptr_arrayidx27 = new GetElementPtrInst(ptr_CONST, const_int32_19, "arrayidx27", label_entry);
-      LoadInst* packed_tmp28 = new LoadInst(ptr_arrayidx27, "tmp28", false, label_entry);
-      BinaryOperator* packed_mul29 = BinaryOperator::create(Instruction::Mul, packed_tmp25, packed_tmp28, "mul29", label_entry);
-      BinaryOperator* packed_add31 = BinaryOperator::create(Instruction::Add, packed_mul29, packed_add21, "add31", label_entry);
-      StoreInst* void_25 = new StoreInst(packed_add31, ptr_OUT, false, label_entry);
-      GetElementPtrInst* ptr_arrayidx33 = new GetElementPtrInst(ptr_OUT, const_int32_15, "arrayidx33", label_entry);
-      GetElementPtrInst* ptr_arrayidx35 = new GetElementPtrInst(ptr_IN, const_int32_15, "arrayidx35", label_entry);
-      LoadInst* packed_tmp36 = new LoadInst(ptr_arrayidx35, "tmp36", false, label_entry);
-      StoreInst* void_26 = new StoreInst(packed_tmp36, ptr_arrayidx33, false, label_entry);
-      std::vector<Value*> int32_call_params;
-      int32_call_params.push_back(const_ptr_20);
-      int32_call_params.push_back(const_double_22);
-      int32_call_params.push_back(const_double_22);
-      int32_call_params.push_back(const_double_22);
-      int32_call_params.push_back(const_double_22);
-      //CallInst* int32_call = new CallInst(func_printf, int32_call_params.begin(), int32_call_params.end(), "call", label_entry);
-      //int32_call->setCallingConv(CallingConv::C);
-      //int32_call->setTailCall(true);
-      ExtractElementInst* float_tmp52 = new ExtractElementInst(packed_tmp1, const_int32_21, "tmp52", label_entry);
-      CastInst* double_conv53 = new FPExtInst(float_tmp52, Type::DoubleTy, "conv53", label_entry);
-      ExtractElementInst* float_tmp55 = new ExtractElementInst(packed_tmp1, const_int32_15, "tmp55", label_entry);
-      CastInst* double_conv56 = new FPExtInst(float_tmp55, Type::DoubleTy, "conv56", label_entry);
-      ExtractElementInst* float_tmp58 = new ExtractElementInst(packed_tmp1, const_int32_17, "tmp58", label_entry);
-      CastInst* double_conv59 = new FPExtInst(float_tmp58, Type::DoubleTy, "conv59", label_entry);
-      ExtractElementInst* float_tmp61 = new ExtractElementInst(packed_tmp1, const_int32_19, "tmp61", label_entry);
-      CastInst* double_conv62 = new FPExtInst(float_tmp61, Type::DoubleTy, "conv62", label_entry);
-      std::vector<Value*> int32_call63_params;
-      int32_call63_params.push_back(const_ptr_23);
-      int32_call63_params.push_back(double_conv53);
-      int32_call63_params.push_back(double_conv56);
-      int32_call63_params.push_back(double_conv59);
-      int32_call63_params.push_back(double_conv62);
-      //CallInst* int32_call63 = new CallInst(func_printf, int32_call63_params.begin(), int32_call63_params.end(), "call63", label_entry);
-      //int32_call63->setCallingConv(CallingConv::C);
-      //int32_call63->setTailCall(true);
-      ExtractElementInst* float_tmp65 = new ExtractElementInst(packed_add31, const_int32_21, "tmp65", label_entry);
-      CastInst* double_conv66 = new FPExtInst(float_tmp65, Type::DoubleTy, "conv66", label_entry);
-      ExtractElementInst* float_tmp68 = new ExtractElementInst(packed_add31, const_int32_15, "tmp68", label_entry);
-      CastInst* double_conv69 = new FPExtInst(float_tmp68, Type::DoubleTy, "conv69", label_entry);
-      ExtractElementInst* float_tmp71 = new ExtractElementInst(packed_add31, const_int32_17, "tmp71", label_entry);
-      CastInst* double_conv72 = new FPExtInst(float_tmp71, Type::DoubleTy, "conv72", label_entry);
-      ExtractElementInst* float_tmp74 = new ExtractElementInst(packed_add31, const_int32_19, "tmp74", label_entry);
-      CastInst* double_conv75 = new FPExtInst(float_tmp74, Type::DoubleTy, "conv75", label_entry);
-      std::vector<Value*> int32_call76_params;
-      int32_call76_params.push_back(const_ptr_24);
-      int32_call76_params.push_back(double_conv66);
-      int32_call76_params.push_back(double_conv69);
-      int32_call76_params.push_back(double_conv72);
-      int32_call76_params.push_back(double_conv75);
-      //CallInst* int32_call76 = new CallInst(func_printf, int32_call76_params.begin(), int32_call76_params.end(), "call76", label_entry);
-      //int32_call76->setCallingConv(CallingConv::C);
-      //int32_call76->setTailCall(true);
-   }
-#endif
-
    new ReturnInst(label_entry);
 
    //TXT("\ntgsi-dump end -------------------\n");
 
    tgsi_parse_free(&parse);
 
+   prog->num_consts = storage.numConsts();
+
    std::cout<<"XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"<<std::endl;
    std::cout<<*mod<<std::endl;
    std::cout<<"YYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYY"<<std::endl;
@@ -786,10 +616,9 @@ ga_llvm_from_tgsi(struct pipe_context *pipe, const struct tgsi_token *tokens)
    struct ga_llvm_prog *ga_llvm =
       (struct ga_llvm_prog *)malloc(sizeof(struct ga_llvm_prog));
    fprintf(stderr, "DUMPX \n");
-   //tgsi_dump(tokens, TGSI_DUMP_VERBOSE);
    tgsi_dump(tokens, 0);
    fprintf(stderr, "DUMPEND \n");
-   llvm::Module *mod = tgsi_to_llvm(tokens);
+   llvm::Module *mod = tgsi_to_llvm(ga_llvm, tokens);
 
    /* Run optimization passes over it */
    PassManager passes;
@@ -831,25 +660,29 @@ void ga_llvm_prog_delete(struct ga_llvm_prog *prog)
 }
 
 typedef void (*vertex_shader_runner)(float (*ainputs)[PIPE_MAX_SHADER_INPUTS][4],
-                                  float (*dests)[PIPE_MAX_SHADER_INPUTS][4],
-                                  float (*aconsts)[4],
-                                  int count,
-                                  int num_attribs);
+                                     float (*dests)[PIPE_MAX_SHADER_INPUTS][4],
+                                     float (*aconsts)[4],
+                                     int num_vertices,
+                                     int num_inputs,
+                                     int num_attribs,
+                                     int num_consts);
 
 int ga_llvm_prog_exec(struct ga_llvm_prog *prog,
                       float (*inputs)[PIPE_MAX_SHADER_INPUTS][4],
                       float (*dests)[PIPE_MAX_SHADER_INPUTS][4],
                       float (*consts)[4],
-                      int count,
+                      int num_vertices,
+                      int num_inputs,
                       int num_attribs)
 {
    std::cout << "---- START LLVM Execution "<<std::endl;
    vertex_shader_runner runner = reinterpret_cast<vertex_shader_runner>(prog->function);
-   runner(inputs, dests, consts, count, num_attribs);
+   runner(inputs, dests, consts, num_vertices, num_inputs,
+          num_attribs, prog->num_consts);
 
    std::cout << "---- END LLVM Execution "<<std::endl;
 
-   for (int i = 0; i < count; ++i) {
+   for (int i = 0; i < num_vertices; ++i) {
       for (int j = 0; j < num_attribs; ++j) {
          printf("OUT(%d, %d) [%f, %f, %f, %f]\n", i, j,
                 dests[i][j][0], dests[i][j][1],
diff --git a/src/mesa/pipe/llvm/llvmtgsi.h b/src/mesa/pipe/llvm/llvmtgsi.h
index f957a69a72..2b30f7ca70 100644
--- a/src/mesa/pipe/llvm/llvmtgsi.h
+++ b/src/mesa/pipe/llvm/llvmtgsi.h
@@ -16,6 +16,7 @@ struct ga_llvm_prog {
    void *module;
    void *engine;
    void *function;
+   int   num_consts;
 };
 struct ga_llvm_prog *
 ga_llvm_from_tgsi(struct pipe_context *pipe, const struct tgsi_token *tokens);
@@ -26,7 +27,8 @@ int ga_llvm_prog_exec(struct ga_llvm_prog *prog,
                       float (*inputs)[PIPE_MAX_SHADER_INPUTS][4],
                       float (*dests)[PIPE_MAX_SHADER_INPUTS][4],
                       float (*consts)[4],
-                      int count,
+                      int num_vertices,
+                      int num_inputs,
                       int num_attribs);
 
 #if defined __cplusplus
diff --git a/src/mesa/pipe/llvm/storage.cpp b/src/mesa/pipe/llvm/storage.cpp
index c6e86ea4b4..dc7c1f3928 100644
--- a/src/mesa/pipe/llvm/storage.cpp
+++ b/src/mesa/pipe/llvm/storage.cpp
@@ -1,5 +1,6 @@
 #include "storage.h"
 
+#include "pipe/tgsi/exec/tgsi_token.h"
 #include <llvm/BasicBlock.h>
 #include <llvm/Module.h>
 #include <llvm/Value.h>
@@ -16,7 +17,8 @@ Storage::Storage(llvm::BasicBlock *block, llvm::Value *out,
                                          llvm::Value *in, llvm::Value *consts)
    : m_block(block), m_OUT(out),
      m_IN(in), m_CONST(consts),
-     m_temps(32)
+     m_temps(32), m_dstCache(32),
+     m_idx(0)
 {
    m_floatVecType = VectorType::get(Type::FloatTy, 4);
    m_intVecType   = VectorType::get(IntegerType::get(32), 4);
@@ -25,6 +27,7 @@ Storage::Storage(llvm::BasicBlock *block, llvm::Value *out,
    m_undefIntVec   = UndefValue::get(m_intVecType);
 
    m_shuffleId = 0;
+   m_numConsts = 0;
 }
 
 //can only build vectors with all members in the [0, 9] range
@@ -69,15 +72,11 @@ llvm::Value *Storage::inputElement(int idx)
    if (m_inputs.find(idx) != m_inputs.end()) {
       return m_inputs[idx];
    }
-   char ptrName[13];
-   char name[9];
-   snprintf(ptrName, 13, "input_ptr%d", idx);
-   snprintf(name, 9, "input%d", idx);
    GetElementPtrInst *getElem = new GetElementPtrInst(m_IN,
                                                       constantInt(idx),
-                                                      ptrName,
+                                                      name("input_ptr"),
                                                       m_block);
-   LoadInst *load = new LoadInst(getElem, name,
+   LoadInst *load = new LoadInst(getElem, name("input"),
                                  false, m_block);
    m_inputs[idx] = load;
    return load;
@@ -85,6 +84,7 @@ llvm::Value *Storage::inputElement(int idx)
 
 llvm::Value *Storage::constElement(int idx)
 {
+   m_numConsts = ((idx + 1) > m_numConsts) ? (idx + 1) : m_numConsts;
    if (m_consts.find(idx) != m_consts.end()) {
       return m_consts[idx];
    }
@@ -123,18 +123,70 @@ llvm::Value *Storage::tempElement(int idx) const
    return ret;
 }
 
-void Storage::setTempElement(int idx, llvm::Value *val)
+void Storage::setTempElement(int idx, llvm::Value *val, int mask)
 {
+   if (mask != TGSI_WRITEMASK_XYZW) {
+      llvm::Value *templ = m_temps[idx];
+      val = maskWrite(val, mask, templ);
+   }
    m_temps[idx] = val;
 }
 
-void Storage::store(int dstIdx, llvm::Value *val)
+void Storage::store(int dstIdx, llvm::Value *val, int mask)
 {
-   char ptrName[13];
-   snprintf(ptrName, 13, "out_ptr%d", dstIdx);
+   if (mask != TGSI_WRITEMASK_XYZW) {
+      llvm::Value *templ = m_dstCache[dstIdx];
+      val = maskWrite(val, mask, templ);
+   }
+
    GetElementPtrInst *getElem = new GetElementPtrInst(m_OUT,
                                                       constantInt(dstIdx),
-                                                      ptrName,
+                                                      name("out_ptr"),
                                                       m_block);
-   new StoreInst(val, getElem, false, m_block);
+   StoreInst *st = new StoreInst(val, getElem, false, m_block);
+   //m_dstCache[dstIdx] = st;
+}
+
+llvm::Value *Storage::maskWrite(llvm::Value *src, int mask, llvm::Value *templ)
+{
+   llvm::Value *dst = templ;
+   if (!dst)
+      dst = Constant::getNullValue(m_floatVecType);
+   if ((mask & TGSI_WRITEMASK_X)) {
+      llvm::Value *x = new ExtractElementInst(src, unsigned(0),
+                                              name("x"), m_block);
+      dst = new InsertElementInst(dst, x, unsigned(0),
+                                  name("dstx"), m_block);
+   }
+   if ((mask & TGSI_WRITEMASK_Y)) {
+      llvm::Value *y = new ExtractElementInst(src, unsigned(1),
+                                              name("y"), m_block);
+      dst = new InsertElementInst(dst, y, unsigned(1),
+                                  name("dsty"), m_block);
+   }
+   if ((mask & TGSI_WRITEMASK_Z)) {
+      llvm::Value *z = new ExtractElementInst(src, unsigned(2),
+                                              name("z"), m_block);
+      dst = new InsertElementInst(dst, z, unsigned(2),
+                                  name("dstz"), m_block);
+   }
+   if ((mask & TGSI_WRITEMASK_W)) {
+      llvm::Value *w = new ExtractElementInst(src, unsigned(3),
+                                              name("w"), m_block);
+      dst = new InsertElementInst(dst, w, unsigned(3),
+                                  name("dstw"), m_block);
+   }
+   return dst;
+}
+
+const char * Storage::name(const char *prefix)
+{
+   ++m_idx;
+   snprintf(m_name, 32, "%s%d", prefix, m_idx);
+   return m_name;
+}
+
+int Storage::numConsts() const
+{
+   return m_numConsts;
 }
diff --git a/src/mesa/pipe/llvm/storage.h b/src/mesa/pipe/llvm/storage.h
index b69c8d614e..f39d308fdd 100644
--- a/src/mesa/pipe/llvm/storage.h
+++ b/src/mesa/pipe/llvm/storage.h
@@ -27,12 +27,17 @@ public:
    llvm::Value *constElement(int idx);
 
    llvm::Value *tempElement(int idx) const;
-   void setTempElement(int idx, llvm::Value *val);
+   void setTempElement(int idx, llvm::Value *val, int mask);
 
    llvm::Value *shuffleVector(llvm::Value *vec, int shuffle);
 
 
-   void store(int dstIdx, llvm::Value *val);
+   void store(int dstIdx, llvm::Value *val, int mask);
+
+   int numConsts() const;
+private:
+   llvm::Value *maskWrite(llvm::Value *src, int mask, llvm::Value *templ);
+   const char *name(const char *prefix);
 private:
    llvm::BasicBlock *m_block;
    llvm::Value *m_OUT;
@@ -42,6 +47,7 @@ private:
    std::map<int, llvm::ConstantInt*> m_constInts;
    std::map<int, llvm::Constant*>    m_intVecs;
    std::vector<llvm::Value*>         m_temps;
+   std::vector<llvm::Value*>         m_dstCache;
    LoadMap                           m_inputs;
    LoadMap                           m_consts;
 
@@ -52,6 +58,10 @@ private:
    llvm::Value      *m_undefIntVec;
 
    int         m_shuffleId;
+   char        m_name[32];
+   int         m_idx;
+
+   int         m_numConsts;
 };
 
 #endif