/**
 * Swizzle a vector consisting of an array of XYZW structs.
 *
 * This fills a vector of dst_len length with the swizzled channels from src.
 *
 * e.g. with swizzles = { 2, 1, 0 } and swizzle_count = 6 results in
 *      RGBA RGBA = BGR BGR BG
 *
 * @param swizzles        the swizzle array
 * @param num_swizzles    the number of elements in swizzles
 * @param dst_len         the length of the result
 */
LLVMValueRef
lp_build_swizzle_aos_n(struct gallivm_state* gallivm,
                       LLVMValueRef src,
                       const unsigned char* swizzles,
                       unsigned num_swizzles,
                       unsigned dst_len)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMValueRef shuffles[LP_MAX_VECTOR_WIDTH];
   unsigned i;

   assert(dst_len < LP_MAX_VECTOR_WIDTH);

   for (i = 0; i < dst_len; ++i) {
      int swizzle = swizzles[i % num_swizzles];

      if (swizzle == LP_BLD_SWIZZLE_DONTCARE) {
         shuffles[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
      } else {
         shuffles[i] = lp_build_const_int32(gallivm, swizzle);
      }
   }

   return LLVMBuildShuffleVector(builder, src, LLVMGetUndef(LLVMTypeOf(src)), LLVMConstVector(shuffles, dst_len), "");
}

int main (void) {
    LLVMModuleRef module = LLVMModuleCreateWithName("kal");
    LLVMBuilderRef builder = LLVMCreateBuilder();
//	LLVMInitializeNativeTarget();

	LLVMTypeRef funcType = LLVMFunctionType(LLVMVoidType(), NULL, 0, 0);
	LLVMValueRef func = LLVMAddFunction(module, "main", funcType);
	LLVMSetLinkage(func, LLVMExternalLinkage);
	LLVMBasicBlockRef block = LLVMAppendBasicBlock(func, "entry");
	LLVMPositionBuilderAtEnd(builder, block);

	LLVMValueRef cond = LLVMBuildICmp(builder, LLVMIntNE, LLVMConstInt(LLVMInt32Type(), 2, 0), LLVMConstInt(LLVMInt32Type(), 1, 0), "ifcond");

	LLVMValueRef owning_block = LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder));	//TODO: WRONG??
	//LLVMValueRef owning_block = LLVMBasicBlockAsValue(LLVMGetPreviousBasicBlock(LLVMGetInsertBlock(builder)));
	// 2. Generate new blocks for cases.
	LLVMBasicBlockRef then_ref = LLVMAppendBasicBlock(owning_block, "then");
	LLVMBasicBlockRef else_ref = LLVMAppendBasicBlock(owning_block, "else");
	LLVMBasicBlockRef merge_ref = LLVMAppendBasicBlock(owning_block, "ifmerge");

	// 3. Branch conditionally on then or else.
	LLVMBuildCondBr(builder, cond, then_ref, else_ref);

	// 4. Build then branch prologue.
	LLVMPositionBuilderAtEnd(builder, then_ref);

	LLVMValueRef hi1 = LLVMBuildXor(builder, LLVMGetUndef(LLVMInt32Type()), LLVMGetUndef(LLVMInt32Type()), "subtmp");

	// 5. Connect then branch to merge block.
	LLVMBuildBr(builder, merge_ref);

	then_ref = LLVMGetInsertBlock(builder);

	// 6. Build else branch prologue.
	LLVMPositionBuilderAtEnd(builder, else_ref);

	LLVMValueRef hi2 = LLVMBuildXor(builder, LLVMGetUndef(LLVMInt32Type()), LLVMGetUndef(LLVMInt32Type()), "subtmp2");

	// 7. Connect else branch to merge block.
	LLVMBuildBr(builder, merge_ref);

	else_ref = LLVMGetInsertBlock(builder);
	// 8. Position ourselves after the merge block.
	LLVMPositionBuilderAtEnd(builder, merge_ref);
	// 9. Build the phi node.
//	LLVMValueRef phi = LLVMBuildPhi(builder, LLVMDoubleType(), "phi");
	// 10. Add incoming edges.
//	LLVMAddIncoming(phi, &hi1, &then_ref, 1);
//	LLVMAddIncoming(phi, &hi2, &else_ref, 1);

	LLVMDumpModule(module);
	LLVMDisposeBuilder(builder);
	LLVMDisposeModule(module);

	return 0;
}

LLVMValueRef
lp_build_broadcast(struct gallivm_state *gallivm,
                   LLVMTypeRef vec_type,
                   LLVMValueRef scalar)
{
   LLVMValueRef res;

   if (LLVMGetTypeKind(vec_type) != LLVMVectorTypeKind) {
      /* scalar */
      assert(vec_type == LLVMTypeOf(scalar));
      res = scalar;
   } else {
      LLVMBuilderRef builder = gallivm->builder;
      const unsigned length = LLVMGetVectorSize(vec_type);
      LLVMValueRef undef = LLVMGetUndef(vec_type);
      /* The shuffle vector is always made of int32 elements */
      LLVMTypeRef i32_type = LLVMInt32TypeInContext(gallivm->context);
      LLVMTypeRef i32_vec_type = LLVMVectorType(i32_type, length);

      assert(LLVMGetElementType(vec_type) == LLVMTypeOf(scalar));

      res = LLVMBuildInsertElement(builder, undef, scalar, LLVMConstNull(i32_type), "");
      res = LLVMBuildShuffleVector(builder, res, undef, LLVMConstNull(i32_vec_type), "");
   }

   return res;
}

static LLVMValueRef build_cube_intrinsic(struct gallivm_state *gallivm,
					 LLVMValueRef in[3])
{
	if (HAVE_LLVM >= 0x0309) {
		LLVMTypeRef f32 = LLVMTypeOf(in[0]);
		LLVMValueRef out[4];

		out[0] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubetc",
					    f32, in, 3, LLVMReadNoneAttribute);
		out[1] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubesc",
					    f32, in, 3, LLVMReadNoneAttribute);
		out[2] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubema",
					    f32, in, 3, LLVMReadNoneAttribute);
		out[3] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubeid",
					    f32, in, 3, LLVMReadNoneAttribute);

		return lp_build_gather_values(gallivm, out, 4);
	} else {
		LLVMValueRef c[4] = {
			in[0],
			in[1],
			in[2],
			LLVMGetUndef(LLVMTypeOf(in[0]))
		};
		LLVMValueRef vec = lp_build_gather_values(gallivm, c, 4);

		return lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.cube",
					  LLVMTypeOf(vec), &vec, 1,
					  LLVMReadNoneAttribute);
	}
}

/**
 * Pack first element of aos values,
 * pad out to destination size.
 * i.e. x1 _ _ _ x2 _ _ _ will become x1 x2 _ _
 */
LLVMValueRef
lp_build_pack_aos_scalars(struct gallivm_state *gallivm,
                          struct lp_type src_type,
                          struct lp_type dst_type,
                          const LLVMValueRef src)
{
   LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
   LLVMValueRef undef = LLVMGetUndef(i32t);
   LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
   unsigned num_src = src_type.length / 4;
   unsigned num_dst = dst_type.length;
   unsigned i;

   assert(num_src <= num_dst);

   for (i = 0; i < num_src; i++) {
      shuffles[i] = LLVMConstInt(i32t, i * 4, 0);
   }
   for (i = num_src; i < num_dst; i++) {
      shuffles[i] = undef;
   }

   if (num_dst == 1) {
      return LLVMBuildExtractElement(gallivm->builder, src, shuffles[0], "");
   }
   else {
      return LLVMBuildShuffleVector(gallivm->builder, src, src,
                                    LLVMConstVector(shuffles, num_dst), "");
   }
}

static LLVMValueRef
emit_array_fetch(
	struct lp_build_tgsi_context *bld_base,
	unsigned File, enum tgsi_opcode_type type,
	struct tgsi_declaration_range range,
	unsigned swizzle)
{
	struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
	struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
	LLVMBuilderRef builder = bld_base->base.gallivm->builder;

	unsigned i, size = range.Last - range.First + 1;
	LLVMTypeRef vec = LLVMVectorType(tgsi2llvmtype(bld_base, type), size);
	LLVMValueRef result = LLVMGetUndef(vec);

	struct tgsi_full_src_register tmp_reg = {};
	tmp_reg.Register.File = File;

	for (i = 0; i < size; ++i) {
		tmp_reg.Register.Index = i + range.First;
		LLVMValueRef temp = emit_fetch(bld_base, &tmp_reg, type, swizzle);
		result = LLVMBuildInsertElement(builder, result, temp,
			lp_build_const_int32(gallivm, i), "");
	}
	return result;
}

void
lp_build_context_init(struct lp_build_context *bld,
                      struct gallivm_state *gallivm,
                      struct lp_type type)
{
   bld->gallivm = gallivm;
   bld->type = type;

   bld->int_elem_type = lp_build_int_elem_type(gallivm, type);
   if (type.floating)
      bld->elem_type = lp_build_elem_type(gallivm, type);
   else
      bld->elem_type = bld->int_elem_type;

   if (type.length == 1) {
      bld->int_vec_type = bld->int_elem_type;
      bld->vec_type = bld->elem_type;
   }
   else {
      bld->int_vec_type = LLVMVectorType(bld->int_elem_type, type.length);
      bld->vec_type = LLVMVectorType(bld->elem_type, type.length);
   }

   bld->undef = LLVMGetUndef(bld->vec_type);
   bld->zero = LLVMConstNull(bld->vec_type);
   bld->one = lp_build_one(gallivm, type);
}

/**
 * Gather elements from scatter positions in memory into a single vector.
 * Use for fetching texels from a texture.
 * For SSE, typical values are length=4, src_width=32, dst_width=32.
 *
 * @param length length of the offsets
 * @param src_width src element width in bits
 * @param dst_width result element width in bits (src will be expanded to fit)
 * @param base_ptr base pointer, should be a i8 pointer type.
 * @param offsets vector with offsets
 */
LLVMValueRef
lp_build_gather(struct gallivm_state *gallivm,
                unsigned length,
                unsigned src_width,
                unsigned dst_width,
                LLVMValueRef base_ptr,
                LLVMValueRef offsets)
{
   LLVMValueRef res;

   if (length == 1) {
      /* Scalar */
      return lp_build_gather_elem(gallivm, length,
                                  src_width, dst_width,
                                  base_ptr, offsets, 0);
   } else {
      /* Vector */

      LLVMTypeRef dst_elem_type = LLVMIntTypeInContext(gallivm->context, dst_width);
      LLVMTypeRef dst_vec_type = LLVMVectorType(dst_elem_type, length);
      unsigned i;

      res = LLVMGetUndef(dst_vec_type);
      for (i = 0; i < length; ++i) {
         LLVMValueRef index = lp_build_const_int32(gallivm, i);
         LLVMValueRef elem;
         elem = lp_build_gather_elem(gallivm, length,
                                     src_width, dst_width,
                                     base_ptr, offsets, i);
         res = LLVMBuildInsertElement(gallivm->builder, res, elem, index, "");
      }
   }

   return res;
}

LLVMValueRef
ac_build_gather_values_extended(struct ac_llvm_context *ctx,
				LLVMValueRef *values,
				unsigned value_count,
				unsigned value_stride,
				bool load)
{
	LLVMBuilderRef builder = ctx->builder;
	LLVMValueRef vec = NULL;
	unsigned i;

	if (value_count == 1) {
		if (load)
			return LLVMBuildLoad(builder, values[0], "");
		return values[0];
	} else if (!value_count)
		unreachable("value_count is 0");

	for (i = 0; i < value_count; i++) {
		LLVMValueRef value = values[i * value_stride];
		if (load)
			value = LLVMBuildLoad(builder, value, "");

		if (!i)
			vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
		LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
		vec = LLVMBuildInsertElement(builder, vec, value, index, "");
	}
	return vec;
}

LLVMValueRef
lp_build_intrinsic_map(struct gallivm_state *gallivm,
                       const char *name,
                       LLVMTypeRef ret_type,
                       LLVMValueRef *args,
                       unsigned num_args)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMTypeRef ret_elem_type = LLVMGetElementType(ret_type);
   unsigned n = LLVMGetVectorSize(ret_type);
   unsigned i, j;
   LLVMValueRef res;

   assert(num_args <= LP_MAX_FUNC_ARGS);

   res = LLVMGetUndef(ret_type);
   for(i = 0; i < n; ++i) {
      LLVMValueRef index = lp_build_const_int32(gallivm, i);
      LLVMValueRef arg_elems[LP_MAX_FUNC_ARGS];
      LLVMValueRef res_elem;
      for(j = 0; j < num_args; ++j)
         arg_elems[j] = LLVMBuildExtractElement(builder, args[j], index, "");
      res_elem = lp_build_intrinsic(builder, name, ret_elem_type, arg_elems, num_args, 0);
      res = LLVMBuildInsertElement(builder, res, res_elem, index, "");
   }

   return res;
}

/**
 * Expands src vector from src.length to dst_length
 */
LLVMValueRef
lp_build_pad_vector(struct gallivm_state *gallivm,
                       LLVMValueRef src,
                       struct lp_type src_type,
                       unsigned dst_length)
{
   LLVMValueRef undef = LLVMGetUndef(lp_build_vec_type(gallivm, src_type));
   LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
   unsigned i;

   assert(dst_length <= Elements(elems));
   assert(dst_length > src_type.length);

   if (src_type.length == dst_length)
      return src;

   /* If its a single scalar type, no need to reinvent the wheel */
   if (src_type.length == 1) {
      return lp_build_broadcast(gallivm, LLVMVectorType(lp_build_elem_type(gallivm, src_type), dst_length), src);
   }

   /* All elements from src vector */
   for (i = 0; i < src_type.length; ++i)
      elems[i] = lp_build_const_int32(gallivm, i);

   /* Undef fill remaining space */
   for (i = src_type.length; i < dst_length; ++i)
      elems[i] = lp_build_const_int32(gallivm, src_type.length);

   /* Combine the two vectors */
   return LLVMBuildShuffleVector(gallivm->builder, src, undef, LLVMConstVector(elems, dst_length), "");
}

/**
 * Combined extract and broadcast (mere shuffle in most cases)
 */
LLVMValueRef
lp_build_extract_broadcast(struct gallivm_state *gallivm,
                           struct lp_type src_type,
                           struct lp_type dst_type,
                           LLVMValueRef vector,
                           LLVMValueRef index)
{
   LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
   LLVMValueRef res;

   assert(src_type.floating == dst_type.floating);
   assert(src_type.width    == dst_type.width);

   assert(lp_check_value(src_type, vector));
   assert(LLVMTypeOf(index) == i32t);

   if (src_type.length == 1) {
      if (dst_type.length == 1) {
         /*
          * Trivial scalar -> scalar.
          */

         res = vector;
      }
      else {
         /*
          * Broadcast scalar -> vector.
          */

         res = lp_build_broadcast(gallivm,
                                  lp_build_vec_type(gallivm, dst_type),
                                  vector);
      }
   }
   else {
      if (dst_type.length > 1) {
         /*
          * shuffle - result can be of different length.
          */

         LLVMValueRef shuffle;
         shuffle = lp_build_broadcast(gallivm,
                                      LLVMVectorType(i32t, dst_type.length),
                                      index);
         res = LLVMBuildShuffleVector(gallivm->builder, vector,
                                      LLVMGetUndef(lp_build_vec_type(gallivm, src_type)),
                                      shuffle, "");
      }
      else {
         /*
          * Trivial extract scalar from vector.
          */
          res = LLVMBuildExtractElement(gallivm->builder, vector, index, "");
      }
   }

   return res;
}

LLVMValueRef gen_tuple(compile_t* c, ast_t* ast)
{
  ast_t* child = ast_child(ast);

  if(ast_sibling(child) == NULL)
    return gen_expr(c, child);

  deferred_reification_t* reify = c->frame->reify;

  ast_t* type = deferred_reify(reify, ast_type(ast), c->opt);

  // If we contain '_', we have no usable value.
  if(contains_dontcare(type))
  {
    ast_free_unattached(type);
    return GEN_NOTNEEDED;
  }

  reach_type_t* t = reach_type(c->reach, type);
  compile_type_t* c_t = (compile_type_t*)t->c_type;
  int count = LLVMCountStructElementTypes(c_t->primitive);
  size_t buf_size = count * sizeof(LLVMTypeRef);
  LLVMTypeRef* elements = (LLVMTypeRef*)ponyint_pool_alloc_size(buf_size);
  LLVMGetStructElementTypes(c_t->primitive, elements);

  LLVMValueRef tuple = LLVMGetUndef(c_t->primitive);
  int i = 0;

  while(child != NULL)
  {
    LLVMValueRef value = gen_expr(c, child);

    if(value == NULL)
    {
      ponyint_pool_free_size(buf_size, elements);
      return NULL;
    }

    // We'll have an undefined element if one of our source elements is a
    // variable declaration. This is ok, since the tuple value will never be
    // used.
    if(value == GEN_NOVALUE || value == GEN_NOTNEEDED)
    {
      ponyint_pool_free_size(buf_size, elements);
      return value;
    }

    ast_t* child_type = deferred_reify(reify, ast_type(child), c->opt);
    value = gen_assign_cast(c, elements[i], value, child_type);
    ast_free_unattached(child_type);
    tuple = LLVMBuildInsertValue(c->builder, tuple, value, i++, "");
    child = ast_sibling(child);
  }

  ponyint_pool_free_size(buf_size, elements);
  return tuple;
}

static LLVMValueRef
emit_fetch(
	struct lp_build_tgsi_context *bld_base,
	const struct tgsi_full_src_register *reg,
	enum tgsi_opcode_type type,
	unsigned swizzle)
{
	struct radeon_llvm_context * ctx = radeon_llvm_context(bld_base);
	struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
	LLVMBuilderRef builder = bld_base->base.gallivm->builder;
	LLVMValueRef result, ptr;

	if (swizzle == ~0) {
		LLVMValueRef values[TGSI_NUM_CHANNELS];
		unsigned chan;
		for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
			values[chan] = emit_fetch(bld_base, reg, type, chan);
		}
		return lp_build_gather_values(bld_base->base.gallivm, values,
					      TGSI_NUM_CHANNELS);
	}

	if (reg->Register.Indirect) {
		struct tgsi_declaration_range range = get_array_range(bld_base,
			reg->Register.File, &reg->Indirect);
		return LLVMBuildExtractElement(builder,
			emit_array_fetch(bld_base, reg->Register.File, type, range, swizzle),
			emit_array_index(bld, &reg->Indirect, reg->Register.Index - range.First),
			"");
	}

	switch(reg->Register.File) {
	case TGSI_FILE_IMMEDIATE: {
		LLVMTypeRef ctype = tgsi2llvmtype(bld_base, type);
		return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
	}

	case TGSI_FILE_INPUT:
		result = ctx->inputs[radeon_llvm_reg_index_soa(reg->Register.Index, swizzle)];
		break;

	case TGSI_FILE_TEMPORARY:
		ptr = lp_get_temp_ptr_soa(bld, reg->Register.Index, swizzle);
		result = LLVMBuildLoad(builder, ptr, "");
		break;

	case TGSI_FILE_OUTPUT:
		ptr = lp_get_output_ptr(bld, reg->Register.Index, swizzle);
		result = LLVMBuildLoad(builder, ptr, "");
		break;

	default:
		return LLVMGetUndef(tgsi2llvmtype(bld_base, type));
	}

	return bitcast(bld_base, type, result);
}

/**
 * Expands src vector from src.length to dst_length
 */
LLVMValueRef
lp_build_pad_vector(struct gallivm_state *gallivm,
                    LLVMValueRef src,
                    unsigned dst_length)
{
   LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
   LLVMValueRef undef;
   LLVMTypeRef type;
   unsigned i, src_length;

   type = LLVMTypeOf(src);

   if (LLVMGetTypeKind(type) != LLVMVectorTypeKind) {
      /* Can't use ShuffleVector on non-vector type */
      undef = LLVMGetUndef(LLVMVectorType(type, dst_length));
      return LLVMBuildInsertElement(gallivm->builder, undef, src, lp_build_const_int32(gallivm, 0), "");
   }

   undef      = LLVMGetUndef(type);
   src_length = LLVMGetVectorSize(type);

   assert(dst_length <= Elements(elems));
   assert(dst_length >= src_length);

   if (src_length == dst_length)
      return src;

   /* All elements from src vector */
   for (i = 0; i < src_length; ++i)
      elems[i] = lp_build_const_int32(gallivm, i);

   /* Undef fill remaining space */
   for (i = src_length; i < dst_length; ++i)
      elems[i] = lp_build_const_int32(gallivm, src_length);

   /* Combine the two vectors */
   return LLVMBuildShuffleVector(gallivm->builder, src, undef, LLVMConstVector(elems, dst_length), "");
}

/**
 * @param n  is the number of pixels processed
 * @param packed  is a <n x i32> vector with the packed YUYV blocks
 * @param i  is a <n x i32> vector with the x pixel coordinate (0 or 1)
 * @return  a <4*n x i8> vector with the pixel RGBA values in AoS
 */
LLVMValueRef
lp_build_fetch_subsampled_rgba_aos(struct gallivm_state *gallivm,
                                   const struct util_format_description *format_desc,
                                   unsigned n,
                                   LLVMValueRef base_ptr,
                                   LLVMValueRef offset,
                                   LLVMValueRef i,
                                   LLVMValueRef j)
{
   LLVMValueRef packed;
   LLVMValueRef rgba;
   struct lp_type fetch_type;

   assert(format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED);
   assert(format_desc->block.bits == 32);
   assert(format_desc->block.width == 2);
   assert(format_desc->block.height == 1);

   fetch_type = lp_type_uint(32);
   packed = lp_build_gather(gallivm, n, 32, fetch_type, TRUE, base_ptr, offset, FALSE);

   (void)j;

   switch (format_desc->format) {
   case PIPE_FORMAT_UYVY:
      rgba = uyvy_to_rgba_aos(gallivm, n, packed, i);
      break;
   case PIPE_FORMAT_YUYV:
      rgba = yuyv_to_rgba_aos(gallivm, n, packed, i);
      break;
   case PIPE_FORMAT_R8G8_B8G8_UNORM:
      rgba = rgbg_to_rgba_aos(gallivm, n, packed, i);
      break;
   case PIPE_FORMAT_G8R8_G8B8_UNORM:
      rgba = grgb_to_rgba_aos(gallivm, n, packed, i);
      break;
   case PIPE_FORMAT_G8R8_B8R8_UNORM:
      rgba = grbr_to_rgba_aos(gallivm, n, packed, i);
      break;
   case PIPE_FORMAT_R8G8_R8B8_UNORM:
      rgba = rgrb_to_rgba_aos(gallivm, n, packed, i);
      break;
   default:
      assert(0);
      rgba =  LLVMGetUndef(LLVMVectorType(LLVMInt8TypeInContext(gallivm->context), 4*n));
      break;
   }

   return rgba;
}

LLVMValueRef
lp_build_gather_values(struct gallivm_state * gallivm,
                       LLVMValueRef * values,
                       unsigned value_count)
{
   LLVMTypeRef vec_type = LLVMVectorType(LLVMTypeOf(values[0]), value_count);
   LLVMBuilderRef builder = gallivm->builder;
   LLVMValueRef vec = LLVMGetUndef(vec_type);
   unsigned i;

   for (i = 0; i < value_count; i++) {
      LLVMValueRef index = lp_build_const_int32(gallivm, i);
      vec = LLVMBuildInsertElement(builder, vec, values[i], index, "");
   }
   return vec;
}

void JITImpl::reclaimUnreachableFunctions(JITCoreInfo &coreInfo)
{
  std::vector<uint32_t> &unreachableFunctions = coreInfo.unreachableFunctions;
  std::map<uint32_t,JITFunctionInfo*> &functionMap = coreInfo.functionMap;
  for (std::vector<uint32_t>::iterator it = unreachableFunctions.begin(),
       e = unreachableFunctions.end(); it != e; ++it) {
    std::map<uint32_t,JITFunctionInfo*>::iterator entry = functionMap.find(*it);
    if (entry == functionMap.end())
      continue;
    LLVMValueRef value = entry->second->value;
    LLVMFreeMachineCodeForFunction(executionEngine, value);
    LLVMReplaceAllUsesWith(value, LLVMGetUndef(LLVMTypeOf(value)));
    LLVMDeleteFunction(value);
    delete entry->second;
    functionMap.erase(entry);
  }
  unreachableFunctions.clear();
}

/**
 * @param n  is the number of pixels processed
 * @param packed  is a <n x i32> vector with the packed YUYV blocks
 * @param i  is a <n x i32> vector with the x pixel coordinate (0 or 1)
 * @return  a <4*n x i8> vector with the pixel RGBA values in AoS
 */
LLVMValueRef
lp_build_fetch_subsampled_rgba_aos(LLVMBuilderRef builder,
                                   const struct util_format_description *format_desc,
                                   unsigned n,
                                   LLVMValueRef base_ptr,
                                   LLVMValueRef offset,
                                   LLVMValueRef i,
                                   LLVMValueRef j)
{
   LLVMValueRef packed;
   LLVMValueRef rgba;

   assert(format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED);
   assert(format_desc->block.bits == 32);
   assert(format_desc->block.width == 2);
   assert(format_desc->block.height == 1);

   packed = lp_build_gather(builder, n, 32, 32, base_ptr, offset);

   (void)j;

   switch (format_desc->format) {
   case PIPE_FORMAT_UYVY:
      rgba = uyvy_to_rgba_aos(builder, n, packed, i);
      break;
   case PIPE_FORMAT_YUYV:
      rgba = yuyv_to_rgba_aos(builder, n, packed, i);
      break;
   case PIPE_FORMAT_R8G8_B8G8_UNORM:
      rgba = rgbg_to_rgba_aos(builder, n, packed, i);
      break;
   case PIPE_FORMAT_G8R8_G8B8_UNORM:
      rgba = grgb_to_rgba_aos(builder, n, packed, i);
      break;
   default:
      assert(0);
      rgba =  LLVMGetUndef(LLVMVectorType(LLVMInt8Type(), 4*n));
      break;
   }

   return rgba;
}

static LLVMValueRef emit_swizzle(
	struct lp_build_tgsi_context * bld_base,
        LLVMValueRef value,
	unsigned swizzle_x,
	unsigned swizzle_y,
	unsigned swizzle_z,
	unsigned swizzle_w)
{
	LLVMValueRef swizzles[4];
	LLVMTypeRef i32t =
		LLVMInt32TypeInContext(bld_base->base.gallivm->context);

	swizzles[0] = LLVMConstInt(i32t, swizzle_x, 0);
	swizzles[1] = LLVMConstInt(i32t, swizzle_y, 0);
	swizzles[2] = LLVMConstInt(i32t, swizzle_z, 0);
	swizzles[3] = LLVMConstInt(i32t, swizzle_w, 0);

	return LLVMBuildShuffleVector(bld_base->base.gallivm->builder,
		value,
		LLVMGetUndef(LLVMTypeOf(value)),
		LLVMConstVector(swizzles, 4), "");
}