allows functions in interp and updates generated

2023-09-05 10:08:00 +02:00 · 2023-09-05 10:08:00 +02:00 · e21f8dc379
parent d5763d2f36
commit e21f8dc379
5 changed files with 2452 additions and 2402 deletions
--- a/TGC_C_instr.yaml
+++ b/TGC_C_instr.yaml
@ -15,59 +15,51 @@ RV32I:
  - JAL:
    encoding: 0b00000000000000000000000001101111
    mask: 0b00000000000000000000000001111111
-    attributes: [[name:no_cont]]
    size:   32
    branch:   true
    delay:   1
  - JALR:
    encoding: 0b00000000000000000000000001100111
    mask: 0b00000000000000000111000001111111
-    attributes: [[name:no_cont]]
    size:   32
    branch:   true
    delay:   1
  - BEQ:
    encoding: 0b00000000000000000000000001100011
    mask: 0b00000000000000000111000001111111
-    attributes: [[name:no_cont], [name:cond]]
    size:   32
    branch:   true
-    delay:   [1,1]
+    delay:   1
  - BNE:
    encoding: 0b00000000000000000001000001100011
    mask: 0b00000000000000000111000001111111
-    attributes: [[name:no_cont], [name:cond]]
    size:   32
    branch:   true
-    delay:   [1,1]
+    delay:   1
  - BLT:
    encoding: 0b00000000000000000100000001100011
    mask: 0b00000000000000000111000001111111
-    attributes: [[name:no_cont], [name:cond]]
    size:   32
    branch:   true
-    delay:   [1,1]
+    delay:   1
  - BGE:
    encoding: 0b00000000000000000101000001100011
    mask: 0b00000000000000000111000001111111
-    attributes: [[name:no_cont], [name:cond]]
    size:   32
    branch:   true
-    delay:   [1,1]
+    delay:   1
  - BLTU:
    encoding: 0b00000000000000000110000001100011
    mask: 0b00000000000000000111000001111111
-    attributes: [[name:no_cont], [name:cond]]
    size:   32
    branch:   true
-    delay:   [1,1]
+    delay:   1
  - BGEU:
    encoding: 0b00000000000000000111000001100011
    mask: 0b00000000000000000111000001111111
-    attributes: [[name:no_cont], [name:cond]]
    size:   32
    branch:   true
-    delay:   [1,1]
+    delay:   1
  - LB:
    encoding: 0b00000000000000000000000000000011
    mask: 0b00000000000000000111000001111111
@ -239,14 +231,12 @@ RV32I:
  - ECALL:
    encoding: 0b00000000000000000000000001110011
    mask: 0b11111111111111111111111111111111
-    attributes: [[name:no_cont]]
    size:   32
    branch:   false
    delay:   1
  - EBREAK:
    encoding: 0b00000000000100000000000001110011
    mask: 0b11111111111111111111111111111111
-    attributes: [[name:no_cont]]
    size:   32
    branch:   false
    delay:   1
@ -391,7 +381,6 @@ RV32IC:
  - CJAL:
    encoding: 0b0010000000000001
    mask: 0b1110000000000011
-    attributes: [[name:no_cont]]
    size:   16
    branch:   true
    delay:   1
@ -458,24 +447,21 @@ RV32IC:
  - CJ:
    encoding: 0b1010000000000001
    mask: 0b1110000000000011
-    attributes: [[name:no_cont]]
    size:   16
    branch:   true
    delay:   1
  - CBEQZ:
    encoding: 0b1100000000000001
    mask: 0b1110000000000011
-    attributes: [[name:no_cont], [name:cond]]
    size:   16
    branch:   true
-    delay:   [1,1]
+    delay:   1
  - CBNEZ:
    encoding: 0b1110000000000001
    mask: 0b1110000000000011
-    attributes: [[name:no_cont], [name:cond]]
    size:   16
    branch:   true
-    delay:   [1,1]
+    delay:   1
  - CSLLI:
    encoding: 0b0000000000000010
    mask: 0b1111000000000011
@ -497,7 +483,6 @@ RV32IC:
  - CJR:
    encoding: 0b1000000000000010
    mask: 0b1111000001111111
-    attributes: [[name:no_cont]]
    size:   16
    branch:   true
    delay:   1
@ -510,14 +495,12 @@ RV32IC:
  - CJALR:
    encoding: 0b1001000000000010
    mask: 0b1111000001111111
-    attributes: [[name:no_cont]]
    size:   16
    branch:   true
    delay:   1
  - CEBREAK:
    encoding: 0b1001000000000010
    mask: 0b1111111111111111
-    attributes: [[name:no_cont]]
    size:   16
    branch:   false
    delay:   1
@ -530,7 +513,6 @@ RV32IC:
  - DII:
    encoding: 0b0000000000000000
    mask: 0b1111111111111111
-    attributes: [[name:no_cont]]
    size:   16
    branch:   false
    delay:   1
--- a/gen_input/templates/interp/CORENAME.cpp.gtl
+++ b/gen_input/templates/interp/CORENAME.cpp.gtl
@ -43,6 +43,7 @@ def nativeTypeSize(int size){
 #include <sstream>
 #include <boost/coroutine2/all.hpp>
 #include <functional>
+#include <exception>

 #ifndef FMT_HEADER_ONLY
 #define FMT_HEADER_ONLY
@ -59,6 +60,10 @@ using namespace iss::arch;
 using namespace iss::debugger;
 using namespace std::placeholders;

+struct memory_access_exception : public std::exception{
+    memory_access_exception(){}
+};
+
 template <typename ARCH> class vm_impl : public iss::interp::vm_base<ARCH> {
 public:
    using traits = arch::traits<ARCH>;
@ -91,30 +96,9 @@ protected:

    inline const char *name(size_t index){return index<traits::reg_aliases.size()?traits::reg_aliases[index]:"illegal";}

-    typename arch::traits<ARCH>::opcode_e decode_inst_id(code_word_t instr);
    virt_addr_t execute_inst(finish_cond_e cond, virt_addr_t start, uint64_t icount_limit) override;

    // some compile time constants
-    // enum { MASK16 = 0b1111110001100011, MASK32 = 0b11111111111100000111000001111111 };
-    enum { MASK16 = 0b1111111111111111, MASK32 = 0b11111111111100000111000001111111 };
-    enum { EXTR_MASK16 = MASK16 >> 2, EXTR_MASK32 = MASK32 >> 2 };
-    enum {
-        LUT_SIZE = 1 << util::bit_count(static_cast<uint32_t>(EXTR_MASK32)),
-        LUT_SIZE_C = 1 << util::bit_count(static_cast<uint32_t>(EXTR_MASK16))
-    };
-
-    std::array<compile_func, LUT_SIZE> lut;
-
-    std::array<compile_func, LUT_SIZE_C> lut_00, lut_01, lut_10;
-    std::array<compile_func, LUT_SIZE> lut_11;
-
-    struct instruction_pattern {
-        uint32_t value;
-        uint32_t mask;
-        typename arch::traits<ARCH>::opcode_e id;
-    };
-
-    std::array<std::vector<instruction_pattern>, 4> qlut;

    inline void raise(uint16_t trap_id, uint16_t cause){
        auto trap_val =  0x80ULL << 24 | (cause << 16) | trap_id;
@ -314,28 +298,30 @@ typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e co
            // 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));
-            switch(inst_id){<%instructions.eachWithIndex{instr, idx -> %>
-            case arch::traits<ARCH>::opcode_e::${instr.name}: {
-                <%instr.fields.eachLine{%>${it}
-                <%}%>if(this->disass_enabled){
-                    /* generate console output when executing the command */<%instr.disass.eachLine{%>
+            try{
+                switch(inst_id){<%instructions.eachWithIndex{instr, idx -> %>
+                case arch::traits<ARCH>::opcode_e::${instr.name}: {
+                    <%instr.fields.eachLine{%>${it}
+                    <%}%>if(this->disass_enabled){
+                        /* generate console output when executing the command */<%instr.disass.eachLine{%>
+                        ${it}<%}%>
+                    }
+                    // used registers<%instr.usedVariables.each{ k,v->
+                    if(v.isArray) {%>
+                    auto* ${k} = reinterpret_cast<uint${nativeTypeSize(v.type.size)}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::${k}0]);<% }else{ %> 
+                    auto* ${k} = reinterpret_cast<uint${nativeTypeSize(v.type.size)}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::${k}]);
+                    <%}}%>// calculate next pc value
+                    *NEXT_PC = *PC + ${instr.length/8};
+                    // execute instruction<%instr.behavior.eachLine{%>
                    ${it}<%}%>
+                    break;
+                }// @suppress("No break at end of case")<%}%>
+                default: {
+                    *NEXT_PC = *PC + ((instr & 3) == 3 ? 4 : 2);
+                    raise(0,  2);
                }
-                // used registers<%instr.usedVariables.each{ k,v->
-                if(v.isArray) {%>
-                auto* ${k} = reinterpret_cast<uint${nativeTypeSize(v.type.size)}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::${k}0]);<% }else{ %> 
-                auto* ${k} = reinterpret_cast<uint${nativeTypeSize(v.type.size)}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::${k}]);
-                <%}}%>// calculate next pc value
-                *NEXT_PC = *PC + ${instr.length/8};
-                // execute instruction<%instr.behavior.eachLine{%>
-                ${it}<%}%>
-                TRAP_${instr.name}:break;
-            }// @suppress("No break at end of case")<%}%>
-            default: {
-                *NEXT_PC = *PC + ((instr & 3) == 3 ? 4 : 2);
-                raise(0,  2);
-            }
-            }
+                }
+            }catch(memory_access_exception& e){}
            // post execution stuff
            process_spawn_blocks();
            if(this->sync_exec && POST_SYNC) this->do_sync(POST_SYNC, static_cast<unsigned>(inst_id));
--- a/gen_input/templates/tcc/CORENAME.cpp.gtl
+++ b/gen_input/templates/tcc/CORENAME.cpp.gtl
@ -167,8 +167,9 @@ private:
        auto cur_pc_val = tu.constant(pc.val, traits::reg_bit_widths[traits::PC]);
        pc=pc+ ${instr.length/8};
        gen_set_pc(tu, pc, traits::NEXT_PC);
-        tu.open_scope();<%instr.behavior.eachLine{%>
-        ${it}<%}%>
+        tu.open_scope();
+        <%instr.behavior.eachLine{%>${it}
+        <%}%>
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,${idx});
        gen_trap_check(tu);
--- a/src/vm/interp/vm_tgc5c.cpp
+++ b/src/vm/interp/vm_tgc5c.cpp
--- a/src/vm/tcc/vm_tgc5c.cpp
+++ b/src/vm/tcc/vm_tgc5c.cpp
@ -120,57 +120,7 @@ protected:
        }
    }

-    // some compile time constants
-    // enum { MASK16 = 0b1111110001100011, MASK32 = 0b11111111111100000111000001111111 };
-    enum { MASK16 = 0b1111111111111111, MASK32 = 0b11111111111100000111000001111111 };
-    enum { EXTR_MASK16 = MASK16 >> 2, EXTR_MASK32 = MASK32 >> 2 };
-    enum { LUT_SIZE = 1 << util::bit_count(static_cast<uint32_t>(EXTR_MASK32)), LUT_SIZE_C = 1 << util::bit_count(static_cast<uint32_t>(EXTR_MASK16)) };
-
-    std::array<compile_func, LUT_SIZE> lut;
-
-    std::array<compile_func, LUT_SIZE_C> lut_00, lut_01, lut_10;
-    std::array<compile_func, LUT_SIZE> lut_11;
-
-    std::array<compile_func *, 4> qlut;
-
-    std::array<const uint32_t, 4> lutmasks = {{EXTR_MASK16, EXTR_MASK16, EXTR_MASK16, EXTR_MASK32}};
-
-    void expand_bit_mask(int pos, uint32_t mask, uint32_t value, uint32_t valid, uint32_t idx, compile_func lut[],
-                         compile_func f) {
-        if (pos < 0) {
-            lut[idx] = f;
-        } else {
-            auto bitmask = 1UL << pos;
-            if ((mask & bitmask) == 0) {
-                expand_bit_mask(pos - 1, mask, value, valid, idx, lut, f);
-            } else {
-                if ((valid & bitmask) == 0) {
-                    expand_bit_mask(pos - 1, mask, value, valid, (idx << 1), lut, f);
-                    expand_bit_mask(pos - 1, mask, value, valid, (idx << 1) + 1, lut, f);
-                } else {
-                    auto new_val = idx << 1;
-                    if ((value & bitmask) != 0) new_val++;
-                    expand_bit_mask(pos - 1, mask, value, valid, new_val, lut, f);
-                }
-            }
-        }
-    }
-
-    inline uint32_t extract_fields(uint32_t val) { return extract_fields(29, val >> 2, lutmasks[val & 0x3], 0); }
-
-    uint32_t extract_fields(int pos, uint32_t val, uint32_t mask, uint32_t lut_val) {
-        if (pos >= 0) {
-            auto bitmask = 1UL << pos;
-            if ((mask & bitmask) == 0) {
-                lut_val = extract_fields(pos - 1, val, mask, lut_val);
-            } else {
-                auto new_val = lut_val << 1;
-                if ((val & bitmask) != 0) new_val++;
-                lut_val = extract_fields(pos - 1, val, mask, new_val);
-            }
-        }
-        return lut_val;
-    }
+    
    template<unsigned W, typename U, typename S = typename std::make_signed<U>::type>
    inline S sext(U from) {
        auto mask = (1ULL<<W) - 1;
@ -182,14 +132,23 @@ private:
    /****************************************************************************
     * start opcode definitions
     ****************************************************************************/
-    struct InstructionDesriptor {
+    struct instruction_descriptor {
        size_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){}
+    };

-    const std::array<InstructionDesriptor, 87> instr_descr = {{
+    decoding_tree_node* root {nullptr};
+
+    const std::array<instruction_descriptor, 87> instr_descr = {{
         /* entries are: size, valid value, valid mask, function ptr */
        /* instruction LUI, encoding '0b00000000000000000000000000110111' */
        {32, 0b00000000000000000000000000110111, 0b00000000000000000000000001111111, &this_class::__lui},
@ -394,6 +353,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,0);
        gen_trap_check(tu);
@ -426,6 +386,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,1);
        gen_trap_check(tu);
@ -465,6 +426,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,2);
        gen_trap_check(tu);
@ -506,6 +468,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,3);
        gen_trap_check(tu);
@ -545,6 +508,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,4);
        gen_trap_check(tu);
@ -584,6 +548,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,5);
        gen_trap_check(tu);
@ -623,6 +588,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,6);
        gen_trap_check(tu);
@ -662,6 +628,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,7);
        gen_trap_check(tu);
@ -701,6 +668,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,8);
        gen_trap_check(tu);
@ -740,6 +708,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,9);
        gen_trap_check(tu);
@ -775,6 +744,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,10);
        gen_trap_check(tu);
@ -810,6 +780,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,11);
        gen_trap_check(tu);
@ -845,6 +816,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,12);
        gen_trap_check(tu);
@ -880,6 +852,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,13);
        gen_trap_check(tu);
@ -915,6 +888,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,14);
        gen_trap_check(tu);
@ -947,6 +921,7 @@ private:
        	tu.write_mem(traits::MEM, store_address, tu.ext(tu.load(rs2+ traits::X0, 0),8,true));
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,15);
        gen_trap_check(tu);
@ -979,6 +954,7 @@ private:
        	tu.write_mem(traits::MEM, store_address, tu.ext(tu.load(rs2+ traits::X0, 0),16,true));
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,16);
        gen_trap_check(tu);
@ -1011,6 +987,7 @@ private:
        	tu.write_mem(traits::MEM, store_address, tu.ext(tu.load(rs2+ traits::X0, 0),32,true));
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,17);
        gen_trap_check(tu);
@ -1044,6 +1021,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,18);
        gen_trap_check(tu);
@ -1077,6 +1055,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,19);
        gen_trap_check(tu);
@ -1110,6 +1089,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,20);
        gen_trap_check(tu);
@ -1143,6 +1123,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,21);
        gen_trap_check(tu);
@ -1176,6 +1157,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,22);
        gen_trap_check(tu);
@ -1209,6 +1191,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,23);
        gen_trap_check(tu);
@ -1242,6 +1225,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,24);
        gen_trap_check(tu);
@ -1275,6 +1259,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,25);
        gen_trap_check(tu);
@ -1308,6 +1293,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,26);
        gen_trap_check(tu);
@ -1341,6 +1327,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,27);
        gen_trap_check(tu);
@ -1374,6 +1361,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,28);
        gen_trap_check(tu);
@ -1407,6 +1395,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,29);
        gen_trap_check(tu);
@ -1440,6 +1429,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,30);
        gen_trap_check(tu);
@ -1473,6 +1463,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,31);
        gen_trap_check(tu);
@ -1506,6 +1497,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,32);
        gen_trap_check(tu);
@ -1539,6 +1531,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,33);
        gen_trap_check(tu);
@ -1572,6 +1565,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,34);
        gen_trap_check(tu);
@ -1605,6 +1599,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,35);
        gen_trap_check(tu);
@ -1638,6 +1633,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,36);
        gen_trap_check(tu);
@ -1666,6 +1662,7 @@ private:
        tu.open_scope();
        tu.write_mem(traits::FENCE, static_cast<uint32_t>(traits:: fence), tu.constant((uint8_t)pred<< 4|succ,8));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,37);
        gen_trap_check(tu);
@ -1686,6 +1683,7 @@ private:
        tu.open_scope();
        this->gen_raise_trap(tu, 0,  11);
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,38);
        gen_trap_check(tu);
@ -1706,6 +1704,7 @@ private:
        tu.open_scope();
        this->gen_raise_trap(tu, 0,  3);
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,39);
        gen_trap_check(tu);
@ -1726,6 +1725,7 @@ private:
        tu.open_scope();
        this->gen_leave_trap(tu, 3);
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,40);
        gen_trap_check(tu);
@ -1746,6 +1746,7 @@ private:
        tu.open_scope();
        this->gen_wait(tu, 1);
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,41);
        gen_trap_check(tu);
@ -1784,6 +1785,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,42);
        gen_trap_check(tu);
@ -1822,6 +1824,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,43);
        gen_trap_check(tu);
@ -1860,6 +1863,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,44);
        gen_trap_check(tu);
@ -1895,6 +1899,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,45);
        gen_trap_check(tu);
@ -1932,6 +1937,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,46);
        gen_trap_check(tu);
@ -1969,6 +1975,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,47);
        gen_trap_check(tu);
@ -1995,6 +2002,7 @@ private:
        tu.open_scope();
        tu.write_mem(traits::FENCE, static_cast<uint32_t>(traits:: fencei), tu.constant(imm,16));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,48);
        gen_trap_check(tu);
@ -2029,6 +2037,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,49);
        gen_trap_check(tu);
@ -2063,6 +2072,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,50);
        gen_trap_check(tu);
@ -2097,6 +2107,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,51);
        gen_trap_check(tu);
@ -2131,6 +2142,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,52);
        gen_trap_check(tu);
@ -2174,6 +2186,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,53);
        gen_trap_check(tu);
@ -2213,6 +2226,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,54);
        gen_trap_check(tu);
@ -2259,6 +2273,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,55);
        gen_trap_check(tu);
@ -2298,6 +2313,7 @@ private:
        	tu.close_scope();
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,56);
        gen_trap_check(tu);
@ -2328,6 +2344,7 @@ private:
        	this->gen_raise_trap(tu, 0,  2);
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,57);
        gen_trap_check(tu);
@ -2355,6 +2372,7 @@ private:
        auto offs = tu.assignment(tu.ext((tu.add(tu.load(rs1+ 8+ traits::X0, 0),tu.constant(uimm,8))),32,true),32);
        tu.store(rd+ 8 + traits::X0,tu.ext(tu.ext(tu.read_mem(traits::MEM, offs, 32),32,false),32,true));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,58);
        gen_trap_check(tu);
@ -2382,6 +2400,7 @@ private:
        auto offs = tu.assignment(tu.ext((tu.add(tu.load(rs1+ 8+ traits::X0, 0),tu.constant(uimm,8))),32,true),32);
        tu.write_mem(traits::MEM, offs, tu.ext(tu.load(rs2+ 8+ traits::X0, 0),32,true));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,59);
        gen_trap_check(tu);
@ -2414,6 +2433,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,60);
        gen_trap_check(tu);
@ -2434,6 +2454,7 @@ private:
        gen_set_pc(tu, pc, traits::NEXT_PC);
        tu.open_scope();
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,61);
        gen_trap_check(tu);
@ -2461,6 +2482,7 @@ private:
        tu.store(traits::NEXT_PC, PC_val_v);
        tu.store(traits::LAST_BRANCH, tu.constant(2U, 2));
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,62);
        gen_trap_check(tu);
@ -2493,6 +2515,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,63);
        gen_trap_check(tu);
@ -2523,6 +2546,7 @@ private:
            tu.store(rd + traits::X0,tu.constant((uint32_t)((int32_t)sext<18>(imm)),32));
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,64);
        gen_trap_check(tu);
@ -2552,6 +2576,7 @@ private:
        	this->gen_raise_trap(tu, 0,  2);
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,65);
        gen_trap_check(tu);
@ -2573,6 +2598,7 @@ private:
        tu.open_scope();
        this->gen_raise_trap(tu, 0,  2);
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,66);
        gen_trap_check(tu);
@ -2598,6 +2624,7 @@ private:
        tu.open_scope();
        tu.store(rs1+ 8 + traits::X0,tu.lshr(tu.load(rs1+ 8+ traits::X0, 0),tu.constant(shamt,8)));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,67);
        gen_trap_check(tu);
@ -2628,6 +2655,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,68);
        gen_trap_check(tu);
@ -2653,6 +2681,7 @@ private:
        tu.open_scope();
        tu.store(rs1+ 8 + traits::X0,tu.ext((tu.bitwise_and(tu.load(rs1+ 8+ traits::X0, 0),tu.constant((int8_t)sext<6>(imm),8))),32,true));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,69);
        gen_trap_check(tu);
@ -2678,6 +2707,7 @@ private:
        tu.open_scope();
        tu.store(rd+ 8 + traits::X0,tu.ext((tu.sub(tu.load(rd+ 8+ traits::X0, 0),tu.load(rs2+ 8+ traits::X0, 0))),32,true));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,70);
        gen_trap_check(tu);
@ -2703,6 +2733,7 @@ private:
        tu.open_scope();
        tu.store(rd+ 8 + traits::X0,tu.bitwise_xor(tu.load(rd+ 8+ traits::X0, 0),tu.load(rs2+ 8+ traits::X0, 0)));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,71);
        gen_trap_check(tu);
@ -2728,6 +2759,7 @@ private:
        tu.open_scope();
        tu.store(rd+ 8 + traits::X0,tu.bitwise_or(tu.load(rd+ 8+ traits::X0, 0),tu.load(rs2+ 8+ traits::X0, 0)));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,72);
        gen_trap_check(tu);
@ -2753,6 +2785,7 @@ private:
        tu.open_scope();
        tu.store(rd+ 8 + traits::X0,tu.bitwise_and(tu.load(rd+ 8+ traits::X0, 0),tu.load(rs2+ 8+ traits::X0, 0)));
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,73);
        gen_trap_check(tu);
@ -2779,6 +2812,7 @@ private:
        tu.store(traits::NEXT_PC, PC_val_v);
        tu.store(traits::LAST_BRANCH, tu.constant(2U, 2));
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,74);
        gen_trap_check(tu);
@ -2808,6 +2842,7 @@ private:
        tu.store(traits::LAST_BRANCH, tu.constant(2U, 2));
        tu.close_scope();
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,75);
        gen_trap_check(tu);
@ -2837,6 +2872,7 @@ private:
        tu.store(traits::LAST_BRANCH, tu.constant(2U, 2));
        tu.close_scope();
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,76);
        gen_trap_check(tu);
@ -2869,6 +2905,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,77);
        gen_trap_check(tu);
@ -2900,6 +2937,7 @@ private:
        	tu.store(rd + traits::X0,tu.ext(tu.ext(tu.read_mem(traits::MEM, offs, 32),32,false),32,true));
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,78);
        gen_trap_check(tu);
@ -2932,6 +2970,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,79);
        gen_trap_check(tu);
@ -2963,6 +3002,7 @@ private:
        	this->gen_raise_trap(tu, 0, 2);
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,80);
        gen_trap_check(tu);
@ -2983,6 +3023,7 @@ private:
        tu.open_scope();
        this->gen_raise_trap(tu, 0, 2);
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,81);
        gen_trap_check(tu);
@ -3015,6 +3056,7 @@ private:
        	}
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,82);
        gen_trap_check(tu);
@ -3048,6 +3090,7 @@ private:
        	tu.store(traits::LAST_BRANCH, tu.constant(2U, 2));
        }
        auto returnValue = std::make_tuple(BRANCH);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,83);
        gen_trap_check(tu);
@ -3068,6 +3111,7 @@ private:
        tu.open_scope();
        this->gen_raise_trap(tu, 0,  3);
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,84);
        gen_trap_check(tu);
@ -3099,6 +3143,7 @@ private:
        	tu.write_mem(traits::MEM, offs, tu.ext(tu.load(rs2+ traits::X0, 0),32,true));
        }
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,85);
        gen_trap_check(tu);
@ -3119,6 +3164,7 @@ private:
        tu.open_scope();
        this->gen_raise_trap(tu, 0,  2);
        auto returnValue = std::make_tuple(CONT);
+        
        tu.close_scope();
        vm_base<ARCH>::gen_sync(tu, POST_SYNC,86);
        gen_trap_check(tu);
@ -3136,6 +3182,59 @@ 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) {
@ -3148,14 +3247,11 @@ 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) {
-    qlut[0] = lut_00.data();
-    qlut[1] = lut_01.data();
-    qlut[2] = lut_10.data();
-    qlut[3] = lut_11.data();
-    for (auto instr : instr_descr) {
-        auto quantrant = instr.value & 0x3;
-        expand_bit_mask(29, lutmasks[quantrant], instr.value >> 2, instr.mask >> 2, 0, qlut[quantrant], instr.op);
+    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);
 }

 template <typename ARCH>
@ -3171,7 +3267,7 @@ vm_impl<ARCH>::gen_single_inst_behavior(virt_addr_t &pc, unsigned int &inst_cnt,
 //    if ((pc.val & upper_bits) != ((pc.val + 2) & upper_bits)) { // we may cross a page boundary
 //        auto res = this->core.read(paddr, 2, data);
 //        if (res != iss::Ok) throw trap_access(TRAP_ID, pc.val);
-//        if ((instr & 0x3) == 0x3) { // this is a 32bit instruction
+//        if ((insn & 0x3) == 0x3) { // this is a 32bit instruction
 //            res = this->core.read(this->core.v2p(pc + 2), 2, data + 2);
 //        }
 //    } else {
@ -3181,8 +3277,7 @@ vm_impl<ARCH>::gen_single_inst_behavior(virt_addr_t &pc, unsigned int &inst_cnt,
    if (instr == 0x0000006f || (instr&0xffff)==0xa001) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
    // curr pc on stack
    ++inst_cnt;
-    auto lut_val = extract_fields(instr);
-    auto f = qlut[instr & 0x3][lut_val];
+    auto f = decode_instr(root, instr);
    if (f == nullptr) {
        f = &this_class::illegal_intruction;
    }
@ -3218,7 +3313,7 @@ std::unique_ptr<vm_if> create<arch::tgc5c>(arch::tgc5c *core, unsigned short por
    if (port != 0) debugger::server<debugger::gdb_session>::run_server(ret, port);
    return std::unique_ptr<vm_if>(ret);
 }
-} // namespace tcc
+} // namesapce tcc
 } // namespace iss

 #include <iss/factory.h>