void CNodeCodeGenVisitor::Visit(CSet* s) {
IRBuilder<> builder = builders_.top();
Value* offset_ptr = builder.CreateGEP(ptr_, GetPtrOffset(s->GetOffset()));
Value* set_val = GetDataOffset(s->GetAmt());
builder.CreateStore(set_val, offset_ptr);
VisitNextCNode(s);
}
Value*
code_emitter::emit_sizeof (Type *type)
{
Value *size = builder->CreateGEP(ConstantPointerNull::get(PointerType::getUnqual(type)),
make_int_const(1));
return builder->CreatePtrToInt(size, sizeof(gpointer) == 4 ? Type::Int32Ty : Type::Int64Ty);
}
void CNodeCodeGenVisitor::Visit(COutput* s) {
IRBuilder<> builder = builders_.top();
Value* offset_ptr = builder.CreateGEP(ptr_, GetPtrOffset(s->GetOffset()));
Value* ptr_value = builder.CreateLoad(offset_ptr);
builder.CreateCall(put_char_, ptr_value);
VisitNextCNode(s);
}
void CNodeCodeGenVisitor::Visit(CInput* s) {
IRBuilder<> builder = builders_.top();
Value* ptr_offset = builder.CreateGEP(ptr_, GetPtrOffset(s->GetOffset()));
Value* input = builder.CreateCall(get_char_);
builder.CreateStore(input, ptr_offset);
VisitNextCNode(s);
}
void CNodeCodeGenVisitor::Visit(CMul* s) {
IRBuilder<> builder = builders_.top();
int op_offset = s->GetOpOffset();
int target_offset = s->GetTargetOffset();
int amt = s->GetAmt();
Value* op_offset_ptr = builder.CreateGEP(ptr_, GetPtrOffset(op_offset));
Value* target_offset_ptr =
builder.CreateGEP(ptr_, GetPtrOffset(target_offset));
Value* mul_val = GetDataOffset(amt);
Value* op_val = builder.CreateLoad(op_offset_ptr);
Value* target_val = builder.CreateLoad(target_offset_ptr);
Value* mul_result = builder.CreateMul(op_val, mul_val);
Value* add_result = builder.CreateAdd(target_val, mul_result);
builder.CreateStore(add_result, target_offset_ptr);
VisitNextCNode(s);
}
void CNodeCodeGenVisitor::Visit(CAdd* s) {
IRBuilder<> builder = builders_.top();
Value* offset_ptr = builder.CreateGEP(ptr_, GetPtrOffset(s->GetOffset()));
Value* offset_val = builder.CreateLoad(offset_ptr);
Value* add_val = GetDataOffset(s->GetAmt());
Value* result = builder.CreateAdd(offset_val, add_val);
builder.CreateStore(result, offset_ptr);
VisitNextCNode(s);
}
Value*
code_emitter::emit_const_value_addr (value_t *value)
{
g_assert(const_value_index_map.find(value) != const_value_index_map.end());
int index = const_value_index_map[value];
vector<Value*> indices;
indices.push_back(make_int_const(0));
indices.push_back(make_int_const(index));
Value *const_var = NULL;
if ((value->const_type | CONST_T) == (CONST_X | CONST_Y | CONST_T))
const_var = xy_vars_var;
else if ((value->const_type | CONST_T) == (CONST_X | CONST_T))
const_var = x_vars_var;
else if ((value->const_type | CONST_T) == (CONST_Y | CONST_T))
const_var = y_vars_var;
else
g_assert_not_reached();
return builder->CreateGEP(const_var, indices.begin(), indices.end());
}
void CNodeCodeGenVisitor::Visit(CPtrMov* s) {
IRBuilder<> builder = builders_.top();
ptr_ = builder.CreateGEP(ptr_, GetPtrOffset(s->GetAmt()));
VisitNextCNode(s);
}
void ASTCodeGenVisitor::Visit(DecrPtr* s) {
IRBuilder<> builder = builders_.top();
ptr_ = builder.CreateGEP(ptr_, neg_one);
VisitNextASTNode(s);
}
JITFunction*
CompilePipeline(Pipeline *pipeline, Thread *thread) {
size_t i = 0;
size_t size = 0;
std::unique_ptr<Module> owner = make_unique<Module>("PipelineFunction", thread->context);
Module *module = owner.get();
std::string fname = std::string("PipelineFunction") + std::to_string(thread->functions++);
size_t input_count = pipeline->inputData->objects.size();
size_t output_count = pipeline->outputData->objects.size();
size_t function_arg_count = 6;
size_t arg_count = input_count + output_count + (function_arg_count - 2);
size_t start_addr = input_count + output_count;
size_t end_addr = start_addr + 1;
size_t result_sizes_addr = end_addr + 1;
size_t thread_nr_addr = result_sizes_addr + 1;
IRBuilder<> *builder = &thread->builder;
auto passmanager = CreatePassManager(module, thread->jit.get());
module->setDataLayout(thread->jit->getTargetMachine().createDataLayout());
Type *int8_tpe = Type::getInt8Ty(thread->context);
Type *int8ptr_tpe = PointerType::get(int8_tpe, 0);
Type *int8ptrptr_tpe = PointerType::get(int8ptr_tpe, 0);
Type *int64_tpe = Type::getInt64Ty(thread->context);
Type *int64ptr_tpe = PointerType::get(int64_tpe, 0);
JITInformation info;
// arguments of the function
// the arguments are (void **result, void** inputs, size_t start, size_t end);
// note that we don't actually use void**, we use int8**, because LLVM does not support void pointers
std::vector<Type*> arguments(function_arg_count);
i = 0;
arguments[i++] = int8ptrptr_tpe; // void** results
arguments[i++] = int8ptrptr_tpe; // void** inputs
arguments[i++] = int64_tpe; // size_t start
arguments[i++] = int64_tpe; // size_t end
arguments[i++] = int64ptr_tpe; // size_t* result_sizes
arguments[i++] = int64_tpe; // size_t thread_nr
assert(i == function_arg_count);
/*for(auto inputs = pipeline->inputData->objects.begin(); inputs != pipeline->inputData->objects.end(); inputs++, i++) {
arguments[i] = PointerType::get(getLLVMType(thread->context, inputs->type), 0);
}
for(auto outputs = pipeline->outputData->objects.begin(); outputs != pipeline->outputData->objects.end(); outputs++, i++) {
arguments[i] = PointerType::get(getLLVMType(thread->context, outputs->type), 0);
}*/
// create the LLVM function
FunctionType *prototype = FunctionType::get(int64_tpe, arguments, false);
Function *function = Function::Create(prototype, GlobalValue::ExternalLinkage, fname, module);
function->setCallingConv(CallingConv::C);
// create the basic blocks
BasicBlock *loop_entry = BasicBlock::Create(thread->context, "entry", function, 0);
BasicBlock *loop_cond = BasicBlock::Create(thread->context, "for.cond", function, 0);
BasicBlock *loop_body = BasicBlock::Create(thread->context, "for.body", function, 0);
BasicBlock *loop_inc = BasicBlock::Create(thread->context, "for.inc", function, 0);
BasicBlock *loop_end = BasicBlock::Create(thread->context, "for.end", function, 0);
info.builder = &thread->builder;
info.context = &thread->context;
info.function = function;
info.loop_entry = loop_entry;
info.loop_cond = loop_cond;
info.loop_body = loop_body;
info.loop_inc = loop_inc;
info.loop_end = loop_end;
info.current = loop_body;
#ifndef _NOTDEBUG
// argument names (for debug purposes only)
std::vector<std::string> argument_names(arg_count);
i = 0;
for(auto inputs = pipeline->inputData->objects.begin(); inputs != pipeline->inputData->objects.end(); inputs++, i++) {
argument_names[i] = std::string("inputs") + std::to_string(i);
}
for(auto outputs = pipeline->outputData->objects.begin(); outputs != pipeline->outputData->objects.end(); outputs++, i++) {
argument_names[i] = std::string("outputs") + std::to_string(i - input_count);
}
argument_names[i++] = "start";
argument_names[i++] = "end";
argument_names[i++] = "result_sizes";
argument_names[i++] = "thread_nr";
#endif
std::vector<AllocaInst*> argument_addresses(arg_count);
builder->SetInsertPoint(loop_entry);
{
// allocate space for the arguments
auto args = function->arg_begin();
i = 0;
for(auto outputs = pipeline->outputData->objects.begin(); outputs != pipeline->outputData->objects.end(); outputs++, i++) {
Type *column_type = PointerType::get(getLLVMType(thread->context, outputs->source->type), 0);
Value *voidptrptr = builder->CreateGEP(int8ptr_tpe, &*args, ConstantInt::get(int64_tpe, i, true));
Value *voidptr = builder->CreateLoad(voidptrptr, "voidptr");
Value *columnptr = builder->CreatePointerCast(voidptr, column_type);
argument_addresses[i] = builder->CreateAlloca(column_type, nullptr, argument_names[i]);
builder->CreateStore(columnptr, argument_addresses[i]);
outputs->alloca_address = (void*) argument_addresses[i];
if (size == 0 || size == 1) {
assert(outputs->source->size >= 0);
size = outputs->source->size;
}
assert(size == outputs->source->size || outputs->source->size == 1);
}
args++;
for(auto inputs = pipeline->inputData->objects.begin(); inputs != pipeline->inputData->objects.end(); inputs++, i++) {
Type *column_type = PointerType::get(getLLVMType(thread->context, inputs->source->type), 0);
Value *voidptrptr = builder->CreateGEP(int8ptr_tpe, &*args, ConstantInt::get(int64_tpe, i - output_count, true));
Value *voidptr = builder->CreateLoad(voidptrptr, "voidptr");
Value *columnptr = builder->CreatePointerCast(voidptr, column_type);
argument_addresses[i] = builder->CreateAlloca(column_type, nullptr, argument_names[i]);
builder->CreateStore(columnptr, argument_addresses[i]);
inputs->alloca_address = (void*) argument_addresses[i];
if (size == 0 || size == 1) {
assert(inputs->source->size >= 0);
size = inputs->source->size;
}
assert(size == inputs->source->size || inputs->source->size == 1);
}
args++;
argument_addresses[i] = builder->CreateAlloca(arguments[2], nullptr, argument_names[i]);
builder->CreateStore(&*args, argument_addresses[i]);
args++; i++;
argument_addresses[i] = builder->CreateAlloca(arguments[3], nullptr, argument_names[i]);
builder->CreateStore(&*args, argument_addresses[i]);
args++; i++;
argument_addresses[i] = builder->CreateAlloca(arguments[4], nullptr, argument_names[i]);
builder->CreateStore(&*args, argument_addresses[i]);
args++; i++;
argument_addresses[i] = builder->CreateAlloca(arguments[5], nullptr, argument_names[i]);
builder->CreateStore(&*args, argument_addresses[i]);
args++; i++;
assert(args == function->arg_end());
assert(i == arg_count);
info.index_addr = argument_addresses[start_addr];
info.thread_addr = argument_addresses[thread_nr_addr];
PerformInitialization(info, pipeline->operation);
builder->CreateBr(loop_cond);
}
// for loop condition: index < end
builder->SetInsertPoint(loop_cond);
{
LoadInst *index = builder->CreateLoad(argument_addresses[start_addr], "index");
LoadInst *end = builder->CreateLoad(argument_addresses[end_addr], "end");
Value *condition = builder->CreateICmpSLT(index, end, "index < end");
builder->CreateCondBr(condition, loop_body, loop_end);
}
// loop body: perform the computation
builder->SetInsertPoint(loop_body);
{
LoadInst *index = builder->CreateLoad(argument_addresses[start_addr], "index");
info.index = index;
info.index_addr = argument_addresses[start_addr];
// perform the computation over the given index
// we don't use the return value because the final assignment has already taken place
Value *v = PerformOperation(info, thread->builder, thread->context, pipeline->operation, pipeline->inputData, pipeline->outputData);
if (v == NULL) {
// failed to perform operation
printf("Failed to compile pipeline %s\n", pipeline->name);
return NULL;
}
builder->CreateBr(loop_inc);
}
// loop increment: index++
builder->SetInsertPoint(loop_inc);
{
LoadInst *index = builder->CreateLoad(argument_addresses[start_addr], "index");
Value *incremented_index = builder->CreateAdd(index, ConstantInt::get(int64_tpe, 1, true), "index++");
builder->CreateStore(incremented_index, argument_addresses[start_addr]);
builder->CreateBr(loop_cond);
}
// loop end: return; (nothing happens here because we have no return value)
builder->SetInsertPoint(loop_end);
{
// return the output size of each of the columns
int i = 0;
Value *result_sizes = builder->CreateLoad(argument_addresses[result_sizes_addr], "result_sizes[]");
for(auto it = pipeline->outputData->objects.begin(); it != pipeline->outputData->objects.end(); it++) {
Value* output_count;
if (it->index_addr) {
output_count = builder->CreateLoad((Value*) it->index_addr, "count");
} else {
output_count = ConstantInt::get(int64_tpe, 1, true);
}
Value *output_addr = builder->CreateGEP(int64_tpe, result_sizes, ConstantInt::get(int64_tpe, i, true));
builder->CreateStore(output_count, output_addr);
i++;
}
builder->CreateRet(ConstantInt::get(int64_tpe, 0, true));
}
#ifndef _NOTDEBUG
verifyFunction(*function);
verifyModule(*module);
#endif
//printf("LLVM for pipeline %s\n", pipeline->name);
module->dump();
passmanager->run(*function);
// dump generated LLVM code
//module->dump();
auto handle = thread->jit->addModule(std::move(owner));
jit_function compiled_function = (jit_function) thread->jit->findSymbol(fname).getAddress();
if (!compiled_function) {
printf("Error creating function.\n");
return NULL;
}
JITFunction *jf = CreateJITFunction(thread, pipeline);
jf->size = size;
jf->function = compiled_function;
jf->jit = thread->jit.get();
jf->handle = handle;
assert(jf->function);
return jf;
}