updates templates for decoder in seperate class, adds again generated templates

gen_input/templates/asmjit/CORENAME.cpp.gtl

@@ -37,6 +37,7 @@
 #include <iss/asmjit/vm_base.h>
 #include <asmjit/asmjit.h>
 #include <util/logging.h>
+#include <vm/instruction_decoder.h>
 
 #ifndef FMT_HEADER_ONLY
 #define FMT_HEADER_ONLY
@@ -115,20 +116,11 @@ private:
      * start opcode definitions
      ****************************************************************************/
     struct instruction_descriptor {
-        size_t length;
+        uint32_t length;
         uint32_t value;
         uint32_t mask;
         compile_func op;
     };
-    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, ${instructions.size()}> instr_descr = {{
          /* entries are: size, valid value, valid mask, function ptr */<%instructions.each{instr -> %>
@@ -136,6 +128,9 @@ private:
         {${instr.length}, ${instr.encoding}, ${instr.mask}, &this_class::__${generator.functionName(instr.name)}},<%}%>
     }};
 
+    //needs to be declared after instr_descr
+    decoder instr_decoder;
+
     /* instruction definitions */<%instructions.eachWithIndex{instr, idx -> %>
     /* instruction ${idx}: ${instr.name} */
     continuation_e __${generator.functionName(instr.name)}(virt_addr_t& pc, code_word_t instr, jit_holder& jh){
@@ -200,72 +195,22 @@ private:
         gen_instr_epilogue(jh);
         return BRANCH;
     }
-     
-    //decoding functionality
-
-    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;
-            }); 
-        }
-    }
-    compile_func decode_instr(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 decode_instr(child, word);
-                }  
-            }  
-        }
-        return nullptr;
-    }
 };
 
 template <typename ARCH> vm_impl<ARCH>::vm_impl() { this(new ARCH()); }
 
 template <typename ARCH>
 vm_impl<ARCH>::vm_impl(ARCH &core, unsigned core_id, unsigned cluster_id)
-: vm_base<ARCH>(core, core_id, cluster_id) {
+: vm_base<ARCH>(core, core_id, cluster_id)
-    root = new decoding_tree_node(std::numeric_limits<uint32_t>::max());
+, instr_decoder([this]() {
-    for(auto instr: instr_descr){
+        std::vector<generic_instruction_descriptor> g_instr_descr;
-        root->instrs.push_back(instr);
+        g_instr_descr.reserve(instr_descr.size());
-    }
+        for (uint32_t i = 0; i < instr_descr.size(); ++i) {
-    populate_decoding_tree(root);
+            generic_instruction_descriptor new_instr_descr {instr_descr[i].value, instr_descr[i].mask, i};
-}
+            g_instr_descr.push_back(new_instr_descr);
+        }
+        return std::move(g_instr_descr);
+    }()) {}
 
 template <typename ARCH>
 continuation_e vm_impl<ARCH>::gen_single_inst_behavior(virt_addr_t &pc, unsigned int &inst_cnt, jit_holder& jh) {
@@ -281,7 +226,10 @@ continuation_e vm_impl<ARCH>::gen_single_inst_behavior(virt_addr_t &pc, unsigned
     if (instr == 0x0000006f || (instr&0xffff)==0xa001)
         throw simulation_stopped(0); // 'J 0' or 'C.J 0'
     ++inst_cnt;
-    auto f = decode_instr(root, instr);
+    uint32_t inst_index = instr_decoder.decode_instr(instr);
+    compile_func f = nullptr;
+    if(inst_index < instr_descr.size())
+        f = instr_descr[inst_index].op;
     if (f == nullptr) 
         f = &this_class::illegal_instruction;
     return (this->*f)(pc, instr, jh);

gen_input/templates/llvm/CORENAME.cpp.gtl

@@ -36,6 +36,7 @@
 #include <iss/iss.h>
 #include <iss/llvm/vm_base.h>
 #include <util/logging.h>
+#include <vm/instruction_decoder.h>
 
 #ifndef FMT_HEADER_ONLY
 #define FMT_HEADER_ONLY
@@ -136,20 +137,11 @@ private:
      * start opcode definitions
      ****************************************************************************/
     struct instruction_descriptor {
-        size_t length;
+        uint32_t length;
         uint32_t value;
         uint32_t mask;
         compile_func op;
     };
-    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, ${instructions.size()}> instr_descr = {{
          /* entries are: size, valid value, valid mask, function ptr */<%instructions.each{instr -> %>
@@ -157,6 +149,9 @@ private:
         {${instr.length}, ${instr.encoding}, ${instr.mask}, &this_class::__${generator.functionName(instr.name)}},<%}%>
     }};
 
+    //needs to be declared after instr_descr
+    decoder instr_decoder;
+
     /* instruction definitions */<%instructions.eachWithIndex{instr, idx -> %>
     /* instruction ${idx}: ${instr.name} */
     std::tuple<continuation_e, BasicBlock*> __${generator.functionName(instr.name)}(virt_addr_t& pc, code_word_t instr, BasicBlock* bb){
@@ -219,58 +214,6 @@ private:
         this->builder.CreateBr(bb);
         return std::make_tuple(BRANCH, nullptr);
     }    
-    //decoding functionality
-
-    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;
-            }); 
-        }
-    }
-    compile_func decode_instr(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 decode_instr(child, word);
-                }  
-            }  
-        }
-        return nullptr;
-    }
 };
 
 template <typename CODE_WORD> void debug_fn(CODE_WORD instr) {
@@ -282,13 +225,16 @@ template <typename ARCH> vm_impl<ARCH>::vm_impl() { this(new ARCH()); }
 
 template <typename ARCH>
 vm_impl<ARCH>::vm_impl(ARCH &core, unsigned core_id, unsigned cluster_id)
-: vm_base<ARCH>(core, core_id, cluster_id) {
+: vm_base<ARCH>(core, core_id, cluster_id)
-    root = new decoding_tree_node(std::numeric_limits<uint32_t>::max());
+, instr_decoder([this]() {
-    for(auto instr:instr_descr){
+        std::vector<generic_instruction_descriptor> g_instr_descr;
-        root->instrs.push_back(instr);
+        g_instr_descr.reserve(instr_descr.size());
-    }
+        for (uint32_t i = 0; i < instr_descr.size(); ++i) {
-    populate_decoding_tree(root);
+            generic_instruction_descriptor new_instr_descr {instr_descr[i].value, instr_descr[i].mask, i};
-}
+            g_instr_descr.push_back(new_instr_descr);
+        }
+        return std::move(g_instr_descr);
+    }()) {}
 
 template <typename ARCH>
 std::tuple<continuation_e, BasicBlock *>
@@ -309,13 +255,16 @@ vm_impl<ARCH>::gen_single_inst_behavior(virt_addr_t &pc, unsigned int &inst_cnt,
 //            res = this->core.read(this->core.v2p(pc + 2), 2, data + 2);
 //        }
 //    } else {
-        auto res = this->core.read(paddr, 4, data);
+    auto res = this->core.read(paddr, 4, data);
-        if (res != iss::Ok) throw trap_access(TRAP_ID, pc.val);
+    if (res != iss::Ok) throw trap_access(TRAP_ID, pc.val);
 //    }
     if (instr == 0x0000006f || (instr&0xffff)==0xa001) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
     // curr pc on stack
     ++inst_cnt;
-    auto f = decode_instr(root, instr);
+    uint32_t inst_index = instr_decoder.decode_instr(instr);
+    compile_func f = nullptr;
+    if(inst_index < instr_descr.size())
+        f = instr_descr[inst_index].op;
     if (f == nullptr) {
         f = &this_class::illegal_instruction;
     }

gen_input/templates/tcc/CORENAME.cpp.gtl

@@ -37,6 +37,7 @@
 #include <iss/tcc/vm_base.h>
 #include <util/logging.h>
 #include <sstream>
+#include <vm/instruction_decoder.h>
 
 #ifndef FMT_HEADER_ONLY
 #define FMT_HEADER_ONLY
@@ -137,20 +138,11 @@ private:
      * start opcode definitions
      ****************************************************************************/
     struct instruction_descriptor {
-        size_t length;
+        uint32_t length;
         uint32_t value;
         uint32_t mask;
         compile_func op;
     };
-    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, ${instructions.size()}> instr_descr = {{
          /* entries are: size, valid value, valid mask, function ptr */<%instructions.each{instr -> %>
@@ -158,6 +150,9 @@ private:
         {${instr.length}, ${instr.encoding}, ${instr.mask}, &this_class::__${generator.functionName(instr.name)}},<%}%>
     }};
 
+    //needs to be declared after instr_descr
+    decoder instr_decoder;
+
     /* instruction definitions */<%instructions.eachWithIndex{instr, idx -> %>
     /* instruction ${idx}: ${instr.name} */
     compile_ret_t __${generator.functionName(instr.name)}(virt_addr_t& pc, code_word_t instr, tu_builder& tu){
@@ -198,59 +193,6 @@ private:
         vm_impl::gen_trap_check(tu);
         return BRANCH;
     }
-    
-    //decoding functionality
-
-    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;
-            }); 
-        }
-    }
-    compile_func decode_instr(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 decode_instr(child, word);
-                }  
-            }  
-        }
-        return nullptr;
-    }
 };
 
 template <typename CODE_WORD> void debug_fn(CODE_WORD instr) {
@@ -262,13 +204,16 @@ template <typename ARCH> vm_impl<ARCH>::vm_impl() { this(new ARCH()); }
 
 template <typename ARCH>
 vm_impl<ARCH>::vm_impl(ARCH &core, unsigned core_id, unsigned cluster_id)
-: vm_base<ARCH>(core, core_id, cluster_id) {
+: vm_base<ARCH>(core, core_id, cluster_id)
-    root = new decoding_tree_node(std::numeric_limits<uint32_t>::max());
+, instr_decoder([this]() {
-    for(auto instr:instr_descr){
+        std::vector<generic_instruction_descriptor> g_instr_descr;
-        root->instrs.push_back(instr);
+        g_instr_descr.reserve(instr_descr.size());
-    }
+        for (uint32_t i = 0; i < instr_descr.size(); ++i) {
-    populate_decoding_tree(root);
+            generic_instruction_descriptor new_instr_descr {instr_descr[i].value, instr_descr[i].mask, i};
-}
+            g_instr_descr.push_back(new_instr_descr);
+        }
+        return std::move(g_instr_descr);
+    }()) {}
 
 template <typename ARCH>
 std::tuple<continuation_e>
@@ -287,13 +232,16 @@ vm_impl<ARCH>::gen_single_inst_behavior(virt_addr_t &pc, unsigned int &inst_cnt,
 //            res = this->core.read(this->core.v2p(pc + 2), 2, data + 2);
 //        }
 //    } else {
-        auto res = this->core.read(paddr, 4, reinterpret_cast<uint8_t*>(&instr));
+    auto res = this->core.read(paddr, 4, reinterpret_cast<uint8_t*>(&instr));
-        if (res != iss::Ok) throw trap_access(TRAP_ID, pc.val);
+    if (res != iss::Ok) throw trap_access(TRAP_ID, pc.val);
 //    }
     if (instr == 0x0000006f || (instr&0xffff)==0xa001) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
     // curr pc on stack
     ++inst_cnt;
-    auto f = decode_instr(root, instr);
+    uint32_t inst_index = instr_decoder.decode_instr(instr);
+    compile_func f = nullptr;
+    if(inst_index < instr_descr.size())
+        f = instr_descr[inst_index].op;
     if (f == nullptr) {
         f = &this_class::illegal_instruction;
     }