adds faster instruction decoding

This commit is contained in:
Eyck-Alexander Jentzsch 2023-07-23 08:05:15 +02:00
parent 7b31b8ca8e
commit c78026b720
1 changed files with 66 additions and 27 deletions

View File

@ -153,14 +153,22 @@ private:
/****************************************************************************
* start opcode definitions
****************************************************************************/
struct InstructionDesriptor {
struct instruction_descriptor {
size_t length;
uint32_t value;
uint32_t mask;
typename arch::traits<ARCH>::opcode_e op;
};
const std::array<InstructionDesriptor, 87> instr_descr = {{
struct decoding_tree_node{
std::vector<instruction_descriptor> instrs;
std::vector<decoding_tree_node*> children;
uint32_t submask = std::numeric_limits<uint32_t>::max();
uint32_t value;
decoding_tree_node(uint32_t value) : value(value){}
};
decoding_tree_node* root {nullptr};
const std::array<instruction_descriptor, 87> instr_descr = {{
/* entries are: size, valid value, valid mask, function ptr */
{32, 0b00000000000000000000000000110111, 0b00000000000000000000000001111111, arch::traits<ARCH>::opcode_e::LUI},
{32, 0b00000000000000000000000000010111, 0b00000000000000000000000001111111, arch::traits<ARCH>::opcode_e::AUIPC},
@ -251,18 +259,61 @@ private:
{16, 0b0000000000000000, 0b1111111111111111, arch::traits<ARCH>::opcode_e::DII},
}};
//static constexpr typename traits::addr_t upper_bits = ~traits::PGMASK;
iss::status fetch_ins(virt_addr_t pc, uint8_t * data){
auto phys_pc = this->core.v2p(pc);
//if ((pc.val & upper_bits) != ((pc.val + 2) & upper_bits)) { // we may cross a page boundary
// if (this->core.read(phys_pc, 2, data) != iss::Ok) return iss::Err;
// if ((data[0] & 0x3) == 0x3) // this is a 32bit instruction
// if (this->core.read(this->core.v2p(pc + 2), 2, data + 2) != iss::Ok) return iss::Err;
//} else {
if (this->core.read(phys_pc, 4, data) != iss::Ok) return iss::Err;
//}
return iss::Ok;
}
void populate_decoding_tree(decoding_tree_node* root){
//create submask
for(auto instr: root->instrs){
root->submask &= instr.mask;
}
//put each instr according to submask&encoding into children
for(auto instr: root->instrs){
bool foundMatch = false;
for(auto child: root->children){
//use value as identifying trait
if(child->value == (instr.value&root->submask)){
child->instrs.push_back(instr);
foundMatch = true;
}
}
if(!foundMatch){
decoding_tree_node* child = new decoding_tree_node(instr.value&root->submask);
child->instrs.push_back(instr);
root->children.push_back(child);
}
}
root->instrs.clear();
//call populate_decoding_tree for all children
if(root->children.size() >1)
for(auto child: root->children){
populate_decoding_tree(child);
}
else{
//sort instrs by value of the mask, this works bc we want to have the least restrictive one last
std::sort(root->children[0]->instrs.begin(), root->children[0]->instrs.end(), [](const instruction_descriptor& instr1, const instruction_descriptor& instr2) {
return instr1.mask > instr2.mask;
});
}
}
typename arch::traits<ARCH>::opcode_e decodeInstr(decoding_tree_node* node, code_word_t word){
if(!node->children.size()){
if(node->instrs.size() == 1) return node->instrs[0].op;
for(auto instr : node->instrs){
if((instr.mask&word) == instr.value) return instr.op;
}
}
else{
for(auto child : node->children){
if (child->value == (node->submask&word)){
return decodeInstr(child, word);
}
}
}
return arch::traits<ARCH>::opcode_e::MAX_OPCODE;
}
};
template <typename CODE_WORD> void debug_fn(CODE_WORD insn) {
@ -290,15 +341,11 @@ template <typename ARCH>
vm_impl<ARCH>::vm_impl(ARCH &core, unsigned core_id, unsigned cluster_id)
: vm_base<ARCH>(core, core_id, cluster_id) {
unsigned id=0;
for (auto instr : instr_descr) {
auto quadrant = instr.value & 0x3;
qlut[quadrant].push_back(instruction_pattern{instr.value, instr.mask, instr.op});
}
for(auto& lut: qlut){
std::sort(std::begin(lut), std::end(lut), [](instruction_pattern const& a, instruction_pattern const& b){
return bit_count(a.mask) > bit_count(b.mask);
});
root = new decoding_tree_node(std::numeric_limits<uint32_t>::max());
for(auto instr:instr_descr){
root->instrs.push_back(instr);
}
populate_decoding_tree(root);
}
inline bool is_count_limit_enabled(finish_cond_e cond){
@ -309,14 +356,6 @@ inline bool is_jump_to_self_enabled(finish_cond_e cond){
return (cond & finish_cond_e::JUMP_TO_SELF) == finish_cond_e::JUMP_TO_SELF;
}
template <typename ARCH>
typename arch::traits<ARCH>::opcode_e vm_impl<ARCH>::decode_inst_id(code_word_t instr){
for(auto& e: qlut[instr&0x3]){
if(!((instr&e.mask) ^ e.value )) return e.id;
}
return arch::traits<ARCH>::opcode_e::MAX_OPCODE;
}
template <typename ARCH>
typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e cond, virt_addr_t start, uint64_t icount_limit){
auto pc=start;
@ -338,7 +377,7 @@ typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e co
} else {
if (is_jump_to_self_enabled(cond) &&
(instr == 0x0000006f || (instr&0xffff)==0xa001)) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
auto inst_id = decode_inst_id(instr);
auto inst_id = decodeInstr(root, instr);
// pre execution stuff
this->core.reg.last_branch = 0;
if(this->sync_exec && PRE_SYNC) this->do_sync(PRE_SYNC, static_cast<unsigned>(inst_id));