LLVMValueRef
ac_build_imsb(struct ac_llvm_context *ctx,
	      LLVMValueRef arg,
	      LLVMTypeRef dst_type)
{
	const char *intr_name = (HAVE_LLVM < 0x0400) ? "llvm.AMDGPU.flbit.i32" :
						       "llvm.amdgcn.sffbh.i32";
	LLVMValueRef msb = ac_build_intrinsic(ctx, intr_name,
					      dst_type, &arg, 1,
					      AC_FUNC_ATTR_READNONE);

	/* The HW returns the last bit index from MSB, but NIR/TGSI wants
	 * the index from LSB. Invert it by doing "31 - msb". */
	msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
			   msb, "");

	LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
	LLVMValueRef cond = LLVMBuildOr(ctx->builder,
					LLVMBuildICmp(ctx->builder, LLVMIntEQ,
						      arg, LLVMConstInt(ctx->i32, 0, 0), ""),
					LLVMBuildICmp(ctx->builder, LLVMIntEQ,
						      arg, all_ones, ""), "");

	return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
}

static LLVMValueRef make_cmp_value(compile_t* c, bool sign,
  LLVMValueRef l_value, LLVMValueRef r_value, LLVMRealPredicate cmp_f,
  LLVMIntPredicate cmp_si, LLVMIntPredicate cmp_ui)
{
  if((l_value == NULL) || (r_value == NULL))
    return NULL;

  if(LLVMIsConstant(l_value) && LLVMIsConstant(r_value))
  {
    if(is_fp(l_value))
      return LLVMConstFCmp(cmp_f, l_value, r_value);

    if(sign)
      return LLVMConstICmp(cmp_si, l_value, r_value);

    return LLVMConstICmp(cmp_ui, l_value, r_value);
  }

  if(is_fp(l_value))
    return LLVMBuildFCmp(c->builder, cmp_f, l_value, r_value, "");

  if(sign)
    return LLVMBuildICmp(c->builder, cmp_si, l_value, r_value, "");

  return LLVMBuildICmp(c->builder, cmp_ui, l_value, r_value, "");
}

/* Find the last bit opposite of the sign bit. */
static void emit_imsb(const struct lp_build_tgsi_action *action,
		      struct lp_build_tgsi_context *bld_base,
		      struct lp_build_emit_data *emit_data)
{
	struct gallivm_state *gallivm = bld_base->base.gallivm;
	LLVMBuilderRef builder = gallivm->builder;
	LLVMValueRef arg = emit_data->args[0];

	LLVMValueRef msb =
		lp_build_intrinsic(builder, "llvm.AMDGPU.flbit.i32",
				emit_data->dst_type, &arg, 1,
				LLVMReadNoneAttribute);

	/* The HW returns the last bit index from MSB, but TGSI wants
	 * the index from LSB. Invert it by doing "31 - msb". */
	msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
			   msb, "");

	/* If arg == 0 || arg == -1 (0xffffffff), return -1. */
	LLVMValueRef all_ones = lp_build_const_int32(gallivm, -1);

	LLVMValueRef cond =
		LLVMBuildOr(builder,
			    LLVMBuildICmp(builder, LLVMIntEQ, arg,
					  bld_base->uint_bld.zero, ""),
			    LLVMBuildICmp(builder, LLVMIntEQ, arg,
					  all_ones, ""), "");

	emit_data->output[emit_data->chan] =
		LLVMBuildSelect(builder, cond, all_ones, msb, "");
}

static void emit_ssg(const struct lp_build_tgsi_action *action,
		     struct lp_build_tgsi_context *bld_base,
		     struct lp_build_emit_data *emit_data)
{
	LLVMBuilderRef builder = bld_base->base.gallivm->builder;

	LLVMValueRef cmp, val;

	if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_I64SSG) {
		cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int64_bld.zero, "");
		val = LLVMBuildSelect(builder, cmp, bld_base->int64_bld.one, emit_data->args[0], "");
		cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int64_bld.zero, "");
		val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int64_bld.elem_type, -1, true), "");
	} else if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_ISSG) {
		cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int_bld.zero, "");
		val = LLVMBuildSelect(builder, cmp, bld_base->int_bld.one, emit_data->args[0], "");
		cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
		val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
	} else { // float SSG
		cmp = LLVMBuildFCmp(builder, LLVMRealOGT, emit_data->args[0], bld_base->base.zero, "");
		val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
		cmp = LLVMBuildFCmp(builder, LLVMRealOGE, val, bld_base->base.zero, "");
		val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
	}

	emit_data->output[emit_data->chan] = val;
}

/**
 * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
 * two (adjacent) mipmap level indexes.  Later, we'll sample from those
 * two mipmap levels and interpolate between them.
 */
void
lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
                           unsigned unit,
                           LLVMValueRef lod_ipart,
                           LLVMValueRef *lod_fpart_inout,
                           LLVMValueRef *level0_out,
                           LLVMValueRef *level1_out)
{
   LLVMBuilderRef builder = bld->gallivm->builder;
   struct lp_build_context *int_bld = &bld->int_bld;
   struct lp_build_context *float_bld = &bld->float_bld;
   LLVMValueRef last_level;
   LLVMValueRef clamp_min;
   LLVMValueRef clamp_max;

   *level0_out = lod_ipart;
   *level1_out = lp_build_add(int_bld, lod_ipart, int_bld->one);

   last_level = bld->dynamic_state->last_level(bld->dynamic_state,
                                               bld->gallivm, unit);

   /*
    * Clamp both lod_ipart and lod_ipart + 1 to [0, last_level], with the
    * minimum number of comparisons, and zeroing lod_fpart in the extreme
    * ends in the process.
    */

   /* lod_ipart < 0 */
   clamp_min = LLVMBuildICmp(builder, LLVMIntSLT,
                             lod_ipart, int_bld->zero,
                             "clamp_lod_to_zero");

   *level0_out = LLVMBuildSelect(builder, clamp_min,
                                 int_bld->zero, *level0_out, "");

   *level1_out = LLVMBuildSelect(builder, clamp_min,
                                 int_bld->zero, *level1_out, "");

   *lod_fpart_inout = LLVMBuildSelect(builder, clamp_min,
                                      float_bld->zero, *lod_fpart_inout, "");

   /* lod_ipart >= last_level */
   clamp_max = LLVMBuildICmp(builder, LLVMIntSGE,
                             lod_ipart, last_level,
                             "clamp_lod_to_last");

   *level0_out = LLVMBuildSelect(builder, clamp_max,
                                 last_level, *level0_out, "");

   *level1_out = LLVMBuildSelect(builder, clamp_max,
                                 last_level, *level1_out, "");

   *lod_fpart_inout = LLVMBuildSelect(builder, clamp_max,
                                      float_bld->zero, *lod_fpart_inout, "");

   lp_build_name(*level0_out, "sampler%u_miplevel0", unit);
   lp_build_name(*level1_out, "sampler%u_miplevel1", unit);
   lp_build_name(*lod_fpart_inout, "sampler%u_mipweight", unit);
}

static LLVMValueRef
translateStructBinOp(NodeKind Op, LLVMValueRef ValueE1, LLVMValueRef ValueE2) {
  switch (Op) {
    case EqOp:   return LLVMBuildICmp(Builder, LLVMIntEQ, ValueE1, ValueE2, "");
    case DiffOp: return LLVMBuildICmp(Builder, LLVMIntNE, ValueE1, ValueE2, "");
    default:     return NULL;
  }
}

LLVMValueRef gendesc_istrait(compile_t* c, LLVMValueRef desc, ast_t* type)
{
  // Get the trait identifier.
  reach_type_t* t = reach_type(c->reach, type);
  assert(t != NULL);
  LLVMValueRef trait_id = LLVMConstInt(c->i32, t->type_id, false);

  // Read the count and the trait list from the descriptor.
  LLVMValueRef count = desc_field(c, desc, DESC_TRAIT_COUNT);
  LLVMValueRef list = desc_field(c, desc, DESC_TRAITS);

  LLVMBasicBlockRef entry_block = LLVMGetInsertBlock(c->builder);
  LLVMBasicBlockRef cond_block = codegen_block(c, "cond");
  LLVMBasicBlockRef body_block = codegen_block(c, "body");
  LLVMBasicBlockRef post_block = codegen_block(c, "post");
  LLVMBuildBr(c->builder, cond_block);

  // While the index is less than the count, check an ID.
  LLVMPositionBuilderAtEnd(c->builder, cond_block);
  LLVMValueRef phi = LLVMBuildPhi(c->builder, c->i32, "");
  LLVMValueRef zero = LLVMConstInt(c->i32, 0, false);
  LLVMAddIncoming(phi, &zero, &entry_block, 1);

  LLVMValueRef test = LLVMBuildICmp(c->builder, LLVMIntULT, phi, count, "");
  LLVMBuildCondBr(c->builder, test, body_block, post_block);

  // The phi node is the index. Get ID and compare it.
  LLVMPositionBuilderAtEnd(c->builder, body_block);

  LLVMValueRef gep[2];
  gep[0] = LLVMConstInt(c->i32, 0, false);
  gep[1] = phi;

  LLVMValueRef id_ptr = LLVMBuildInBoundsGEP(c->builder, list, gep, 2, "");
  LLVMValueRef id = LLVMBuildLoad(c->builder, id_ptr, "");
  LLVMValueRef test_id = LLVMBuildICmp(c->builder, LLVMIntEQ, id, trait_id,
    "");

  // Add one to the phi node.
  LLVMValueRef one = LLVMConstInt(c->i32, 1, false);
  LLVMValueRef inc = LLVMBuildAdd(c->builder, phi, one, "");
  LLVMAddIncoming(phi, &inc, &body_block, 1);

  // Either to the post block or back to the condition.
  LLVMBuildCondBr(c->builder, test_id, post_block, cond_block);

  LLVMPositionBuilderAtEnd(c->builder, post_block);
  LLVMValueRef result = LLVMBuildPhi(c->builder, c->i1, "");
  LLVMAddIncoming(result, &test, &cond_block, 1);
  LLVMAddIncoming(result, &test_id, &body_block, 1);

  return result;
}

bool JITImpl::
emitJumpToNextFragment(InstructionOpcode opc, const Operands &operands,
                       JITCoreInfo &coreInfo, uint32_t nextPc,
                       JITFunctionInfo *caller)
{
  std::set<uint32_t> successors;
  if (!getSuccessors(opc, operands, nextPc, successors))
    return false;
  unsigned numSuccessors = successors.size();
  if (numSuccessors == 0)
    return false;
  std::set<uint32_t>::iterator it = successors.begin();
  ++it;
  if (it != successors.end()) {
    LLVMValueRef args[] = {
      threadParam
    };
    LLVMValueRef nextPc = emitCallToBeInlined(functions.jitGetPc, args, 1);
    for (;it != successors.end(); ++it) {
      LLVMValueRef cmp =
        LLVMBuildICmp(builder, LLVMIntEQ, nextPc,
                      LLVMConstInt(LLVMTypeOf(nextPc), *it, false), "");
      LLVMBasicBlockRef trueBB = appendBBToCurrentFunction(builder, "");
      LLVMBasicBlockRef afterBB = appendBBToCurrentFunction(builder, "");
      LLVMBuildCondBr(builder, cmp, trueBB, afterBB);
      LLVMPositionBuilderAtEnd(builder, trueBB);
      emitJumpToNextFragment(coreInfo, *it, caller);
      LLVMPositionBuilderAtEnd(builder, afterBB);
    }
  }
  emitJumpToNextFragment(coreInfo, *successors.begin(), caller);
  return true;
}

LLVMValueRef gen_not(compile_t* c, ast_t* ast)
{
  LLVMValueRef value = gen_expr(c, ast);

  if(value == NULL)
    return NULL;

  ast_t* type = ast_type(ast);

  if(is_bool(type))
  {
    if(LLVMIsAConstantInt(value))
    {
      if(is_always_true(value))
        return LLVMConstInt(c->ibool, 0, false);

      return LLVMConstInt(c->ibool, 1, false);
    }

    LLVMValueRef test = LLVMBuildICmp(c->builder, LLVMIntEQ, value,
      LLVMConstInt(c->ibool, 0, false), "");
    return LLVMBuildZExt(c->builder, test, c->ibool, "");
  }

  if(LLVMIsAConstantInt(value))
    return LLVMConstNot(value);

  return LLVMBuildNot(c->builder, value, "");
}

static void emit_icmp(const struct lp_build_tgsi_action *action,
		      struct lp_build_tgsi_context *bld_base,
		      struct lp_build_emit_data *emit_data)
{
	unsigned pred;
	LLVMBuilderRef builder = bld_base->base.gallivm->builder;
	LLVMContextRef context = bld_base->base.gallivm->context;

	switch (emit_data->inst->Instruction.Opcode) {
	case TGSI_OPCODE_USEQ:
	case TGSI_OPCODE_U64SEQ: pred = LLVMIntEQ; break;
	case TGSI_OPCODE_USNE:
	case TGSI_OPCODE_U64SNE: pred = LLVMIntNE; break;
	case TGSI_OPCODE_USGE:
	case TGSI_OPCODE_U64SGE: pred = LLVMIntUGE; break;
	case TGSI_OPCODE_USLT:
	case TGSI_OPCODE_U64SLT: pred = LLVMIntULT; break;
	case TGSI_OPCODE_ISGE:
	case TGSI_OPCODE_I64SGE: pred = LLVMIntSGE; break;
	case TGSI_OPCODE_ISLT:
	case TGSI_OPCODE_I64SLT: pred = LLVMIntSLT; break;
	default:
		assert(!"unknown instruction");
		pred = 0;
		break;
	}

	LLVMValueRef v = LLVMBuildICmp(builder, pred,
			emit_data->args[0], emit_data->args[1],"");

	v = LLVMBuildSExtOrBitCast(builder, v,
			LLVMInt32TypeInContext(context), "");

	emit_data->output[emit_data->chan] = v;
}

static void emit_minmax_int(const struct lp_build_tgsi_action *action,
			    struct lp_build_tgsi_context *bld_base,
			    struct lp_build_emit_data *emit_data)
{
	LLVMBuilderRef builder = bld_base->base.gallivm->builder;
	LLVMIntPredicate op;

	switch (emit_data->info->opcode) {
	default:
		assert(0);
	case TGSI_OPCODE_IMAX:
	case TGSI_OPCODE_I64MAX:
		op = LLVMIntSGT;
		break;
	case TGSI_OPCODE_IMIN:
	case TGSI_OPCODE_I64MIN:
		op = LLVMIntSLT;
		break;
	case TGSI_OPCODE_UMAX:
	case TGSI_OPCODE_U64MAX:
		op = LLVMIntUGT;
		break;
	case TGSI_OPCODE_UMIN:
	case TGSI_OPCODE_U64MIN:
		op = LLVMIntULT;
		break;
	}

	emit_data->output[emit_data->chan] =
		LLVMBuildSelect(builder,
				LLVMBuildICmp(builder, op, emit_data->args[0],
					      emit_data->args[1], ""),
				emit_data->args[0],
				emit_data->args[1], "");
}

/**
 * Return (scalar-cast)val ? true : false;
 */
LLVMValueRef
lp_build_any_true_range(struct lp_build_context *bld,
                        unsigned real_length,
                        LLVMValueRef val)
{
   LLVMBuilderRef builder = bld->gallivm->builder;
   LLVMTypeRef scalar_type;
   LLVMTypeRef true_type;

   assert(real_length <= bld->type.length);

   true_type = LLVMIntTypeInContext(bld->gallivm->context,
                                    bld->type.width * real_length);
   scalar_type = LLVMIntTypeInContext(bld->gallivm->context,
                                      bld->type.width * bld->type.length);
   val = LLVMBuildBitCast(builder, val, scalar_type, "");
   /*
    * We're using always native types so we can use intrinsics.
    * However, if we don't do per-element calculations, we must ensure
    * the excess elements aren't used since they may contain garbage.
    */
   if (real_length < bld->type.length) {
      val = LLVMBuildTrunc(builder, val, true_type, "");
   }
   return LLVMBuildICmp(builder, LLVMIntNE,
                        val, LLVMConstNull(true_type), "");
}

/* this is ffs in C */
static void emit_lsb(const struct lp_build_tgsi_action *action,
		     struct lp_build_tgsi_context *bld_base,
		     struct lp_build_emit_data *emit_data)
{
	struct gallivm_state *gallivm = bld_base->base.gallivm;
	LLVMBuilderRef builder = gallivm->builder;
	LLVMValueRef args[2] = {
		emit_data->args[0],

		/* The value of 1 means that ffs(x=0) = undef, so LLVM won't
		 * add special code to check for x=0. The reason is that
		 * the LLVM behavior for x=0 is different from what we
		 * need here. However, LLVM also assumes that ffs(x) is
		 * in [0, 31], but GLSL expects that ffs(0) = -1, so
		 * a conditional assignment to handle 0 is still required.
		 */
		LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
	};

	LLVMValueRef lsb =
		lp_build_intrinsic(gallivm->builder, "llvm.cttz.i32",
				emit_data->dst_type, args, ARRAY_SIZE(args),
				LLVMReadNoneAttribute);

	/* TODO: We need an intrinsic to skip this conditional. */
	/* Check for zero: */
	emit_data->output[emit_data->chan] =
		LLVMBuildSelect(builder,
				LLVMBuildICmp(builder, LLVMIntEQ, args[0],
					      bld_base->uint_bld.zero, ""),
				lp_build_const_int32(gallivm, -1), lsb, "");
}

LLVMValueRef gen_if(struct node *ast)
{
	LLVMValueRef condition, func;
	LLVMBasicBlockRef then_block, else_block, end;

	condition = codegen(ast->one);
	condition = LLVMBuildICmp(builder, LLVMIntNE, condition, CONST(0), "");
	func = LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder));

	then_block = LLVMAppendBasicBlock(func, "");
	else_block = LLVMAppendBasicBlock(func, "");
	end = LLVMAppendBasicBlock(func, "");
	LLVMBuildCondBr(builder, condition, then_block, else_block);

	LLVMPositionBuilderAtEnd(builder, then_block);
	codegen(ast->two);
	LLVMBuildBr(builder, end);

	LLVMPositionBuilderAtEnd(builder, else_block);
	if (ast->three)
		codegen(ast->three);
	LLVMBuildBr(builder, end);

	LLVMPositionBuilderAtEnd(builder, end);

	return NULL;
}

static LLVMValueRef
CreateFibFunction(LLVMModuleRef M, LLVMContextRef Context)
{
	LLVMBuilderRef B = LLVMCreateBuilderInContext(Context);

	// Create the fib function and insert it into module M. This function is said
	// to return an int and take an int parameter.
	LLVMTypeRef  ParamTypes[] = {LLVMInt32TypeInContext(Context)};
	LLVMTypeRef  ReturnType   = LLVMInt32TypeInContext(Context);
	LLVMTypeRef  FunctionTy   = LLVMFunctionType(ReturnType, ParamTypes, 1, 0);
	LLVMValueRef FibF         = LLVMAddFunction(M, "fib", FunctionTy);

	// Add a basic block to the function.
	LLVMBasicBlockRef BB = LLVMAppendBasicBlockInContext(Context, FibF, "EntryBlock");

	// Get pointers to the constants.
	LLVMValueRef One = LLVMConstInt(LLVMInt32TypeInContext(Context), 1, 0);
	LLVMValueRef Two = LLVMConstInt(LLVMInt32TypeInContext(Context), 2, 0);

	// Get pointer to the integer argument of the add1 function...
	LLVMValueRef ArgX = LLVMGetFirstParam(FibF); // Get the arg.
	LLVMSetValueName(ArgX, "AnArg");             // Give it a nice symbolic name for fun.

	// Create the true_block.
	LLVMBasicBlockRef RetBB = LLVMAppendBasicBlockInContext(Context, FibF, "return");

	// Create an exit block.
	LLVMBasicBlockRef RecurseBB = LLVMAppendBasicBlockInContext(Context, FibF, "recurse");

	// Create the "if (arg <= 2) goto exitbb"
	LLVMPositionBuilderAtEnd(B, BB);
	LLVMValueRef CondInst = LLVMBuildICmp(B, LLVMIntSLE, ArgX, Two, "cond");
	LLVMBuildCondBr(B, CondInst, RetBB, RecurseBB);

	// Create: ret int 1
	LLVMPositionBuilderAtEnd(B, RetBB);
	LLVMBuildRet(B, One);

	// create fib(x-1)
	LLVMPositionBuilderAtEnd(B, RecurseBB);
	LLVMValueRef Sub       = LLVMBuildSub(B, ArgX, One, "arg");
	LLVMValueRef CallFibX1 = LLVMBuildCall(B, FibF, &Sub, 1, "fibx1");
	LLVMSetTailCall(CallFibX1, 1);

	// create fib(x-2)
	LLVMPositionBuilderAtEnd(B, RecurseBB);
	Sub                    = LLVMBuildSub(B, ArgX, Two, "arg");
	LLVMValueRef CallFibX2 = LLVMBuildCall(B, FibF, &Sub, 1, "fibx2");
	LLVMSetTailCall(CallFibX2, 1);

	// fib(x-1)+fib(x-2)
	LLVMPositionBuilderAtEnd(B, RecurseBB);
	LLVMValueRef Sum = LLVMBuildAdd(B, CallFibX1, CallFibX2, "addresult");

	// Create the return instruction and add it to the basic block
	LLVMPositionBuilderAtEnd(B, RecurseBB);
	LLVMBuildRet(B, Sum);

	return FibF;
}

void
lp_build_loop_end_cond(struct lp_build_loop_state *state,
                       LLVMValueRef end,
                       LLVMValueRef step,
                       LLVMIntPredicate llvm_cond)
{
   LLVMBuilderRef builder = state->gallivm->builder;
   LLVMValueRef next;
   LLVMValueRef cond;
   LLVMBasicBlockRef after_block;

   if (!step)
      step = LLVMConstInt(LLVMTypeOf(end), 1, 0);

   next = LLVMBuildAdd(builder, state->counter, step, "");

   LLVMBuildStore(builder, next, state->counter_var);

   cond = LLVMBuildICmp(builder, llvm_cond, next, end, "");

   after_block = lp_build_insert_new_block(state->gallivm, "loop_end");

   LLVMBuildCondBr(builder, cond, after_block, state->block);

   LLVMPositionBuilderAtEnd(builder, after_block);

   state->counter = LLVMBuildLoad(builder, state->counter_var, "");
}

/**
 * Check if the mask predicate is zero.  If so, jump to the end of the block.
 */
void
lp_build_mask_check(struct lp_build_mask_context *mask)
{
   LLVMBuilderRef builder = mask->skip.gallivm->builder;
   LLVMValueRef value;
   LLVMValueRef cond;

   value = lp_build_mask_value(mask);

   /*
    * XXX this doesn't quite generate the most efficient code possible, if
    * the masks are vectors which have all bits set to the same value
    * in each element.
    * movmskps/pmovmskb would be more efficient to get the required value
    * into ordinary reg (certainly with 8 floats).
    * Not sure if llvm could figure that out on its own.
    */

   /* cond = (mask == 0) */
   cond = LLVMBuildICmp(builder,
                        LLVMIntEQ,
                        LLVMBuildBitCast(builder, value, mask->reg_type, ""),
                        LLVMConstNull(mask->reg_type),
                        "");

   /* if cond, goto end of block */
   lp_build_flow_skip_cond_break(&mask->skip, cond);
}

/* Find the last bit set. */
static void emit_umsb(const struct lp_build_tgsi_action *action,
		      struct lp_build_tgsi_context *bld_base,
		      struct lp_build_emit_data *emit_data)
{
	struct gallivm_state *gallivm = bld_base->base.gallivm;
	LLVMBuilderRef builder = gallivm->builder;
	LLVMValueRef args[2] = {
		emit_data->args[0],
		/* Don't generate code for handling zero: */
		LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
	};

	LLVMValueRef msb =
		lp_build_intrinsic(builder, "llvm.ctlz.i32",
				emit_data->dst_type, args, ARRAY_SIZE(args),
				LLVMReadNoneAttribute);

	/* The HW returns the last bit index from MSB, but TGSI wants
	 * the index from LSB. Invert it by doing "31 - msb". */
	msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
			   msb, "");

	/* Check for zero: */
	emit_data->output[emit_data->chan] =
		LLVMBuildSelect(builder,
				LLVMBuildICmp(builder, LLVMIntEQ, args[0],
					      bld_base->uint_bld.zero, ""),
				lp_build_const_int32(gallivm, -1), msb, "");
}

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;
}

void
JITImpl::checkReturnValue(LLVMValueRef call, InstructionProperties &properties)
{
  if (!mayReturnEarly(properties))
    return;
  LLVMValueRef cmp =
    LLVMBuildICmp(builder, LLVMIntNE, call,
                  LLVMConstInt(LLVMTypeOf(call), 0, JIT_RETURN_CONTINUE), "");
  emitCondEarlyReturn(cmp, call);
}