adds support for compressed instructions

CMakeLists.txt

@@ -83,6 +83,10 @@ elseif(TARGET elfio::elfio)
 else()
     message(FATAL_ERROR "No elfio library found, maybe a find_package() call is missing")
 endif()
+if(TARGET lz4::lz4)
+    target_compile_definitions(${PROJECT_NAME} PUBLIC WITH_LZ4)
+    target_link_libraries(${PROJECT_NAME} PUBLIC lz4::lz4)
+endif()
 if(TARGET RapidJSON)
     target_link_libraries(${PROJECT_NAME} PUBLIC RapidJSON)
 endif()

gen_input/TGC_C.core_desc

@@ -1,6 +1,6 @@
-import "CoreDSL-Instruction-Set-Description/RV32I.core_desc"
-import "CoreDSL-Instruction-Set-Description/RVM.core_desc"
-import "CoreDSL-Instruction-Set-Description/RVC.core_desc"
+import "RV32I.core_desc"
+import "RVM.core_desc"
+import "RVC.core_desc"
 
 Core TGC_C provides RV32I, Zicsr, Zifencei, RV32M, RV32IC {
     architectural_state {

gen_input/templates/CORENAME.cpp.gtl

@@ -33,7 +33,7 @@
 def getRegisterSizes(){
 	def regs = registers.collect{it.size}
 	regs[-1]=64 // correct for NEXT_PC
-	regs+=[32, 32, 64, 64, 64] // append TRAP_STATE, PENDING_TRAP, ICOUNT, CYCLE, INSTRET
+	//regs+=[32, 32, 64, 64, 64, 32] // append TRAP_STATE, PENDING_TRAP, ICOUNT, CYCLE, INSTRET, INSTRUCTION
     return regs
 }
 %>
@@ -51,9 +51,7 @@ constexpr std::array<const char*, ${registers.size}>    iss::arch::traits<iss::a
 constexpr std::array<const uint32_t, ${getRegisterSizes().size}> iss::arch::traits<iss::arch::${coreDef.name.toLowerCase()}>::reg_bit_widths;
 constexpr std::array<const uint32_t, ${getRegisterSizes().size}> iss::arch::traits<iss::arch::${coreDef.name.toLowerCase()}>::reg_byte_offsets;
 
-${coreDef.name.toLowerCase()}::${coreDef.name.toLowerCase()}() {
-    reg.icount = 0;
-}
+${coreDef.name.toLowerCase()}::${coreDef.name.toLowerCase()}()  = default;
 
 ${coreDef.name.toLowerCase()}::~${coreDef.name.toLowerCase()}() = default;
 
@@ -64,8 +62,8 @@ void ${coreDef.name.toLowerCase()}::reset(uint64_t address) {
     reg.PC=address;
     reg.NEXT_PC=reg.PC;
     reg.PRIV=0x3;
-    reg.trap_state=0;
-    reg.icount=0;
+    trap_state=0;
+    icount=0;
 }
 
 uint8_t *${coreDef.name.toLowerCase()}::get_regs_base_ptr() {

gen_input/templates/CORENAME.h.gtl

@@ -37,7 +37,7 @@ def nativeTypeSize(int size){
 }
 def getRegisterSizes(){
     def regs = registers.collect{nativeTypeSize(it.size)}
-    regs+=[32,32, 64, 64, 64] // append TRAP_STATE, PENDING_TRAP, ICOUNT, CYCLE, INSTRET
+    // regs+=[32,32, 64, 64, 64, 32] // append TRAP_STATE, PENDING_TRAP, ICOUNT, CYCLE, INSTRET, INSTRUCTION
     return regs
 }
 def getRegisterOffsets(){
@@ -91,12 +91,7 @@ template <> struct traits<${coreDef.name.toLowerCase()}> {
     constexpr static unsigned FP_REGS_SIZE = ${constants.find {it.name=='FLEN'}?.value?:0};
 
     enum reg_e {
-        ${registers.collect{it.name}.join(', ')}, NUM_REGS,
-        TRAP_STATE=NUM_REGS,
-        PENDING_TRAP,
-        ICOUNT,
-        CYCLE,
-        INSTRET
+        ${registers.collect{it.name}.join(', ')}, NUM_REGS
     };
 
     using reg_t = uint${addrDataWidth}_t;
@@ -141,7 +136,7 @@ struct ${coreDef.name.toLowerCase()}: public arch_if {
 
     uint8_t* get_regs_base_ptr() override;
 
-    inline uint64_t get_icount() { return reg.icount; }
+    inline uint64_t get_icount() { return icount; }
 
     inline bool should_stop() { return interrupt_sim; }
 
@@ -159,7 +154,7 @@ struct ${coreDef.name.toLowerCase()}: public arch_if {
 
     virtual iss::sync_type needed_sync() const { return iss::NO_SYNC; }
 
-    inline uint32_t get_last_branch() { return reg.last_branch; }
+    inline uint32_t get_last_branch() { return last_branch; }
 
 
 #pragma pack(push, 1)
@@ -167,12 +162,14 @@ struct ${coreDef.name.toLowerCase()}: public arch_if {
         registers.each { reg -> if(reg.size>0) {%> 
         uint${byteSize(reg.size)}_t ${reg.name} = 0;<%
         }}%>
-        uint32_t trap_state = 0, pending_trap = 0;
-        uint64_t icount = 0;
-        uint64_t instret = 0;
-        uint32_t last_branch;
     } reg;
 #pragma pack(pop)
+    uint32_t trap_state = 0, pending_trap = 0;
+    uint64_t icount = 0;
+    uint64_t cycle = 0;
+    uint64_t instret = 0;
+    uint32_t instruction = 0;
+    uint32_t last_branch = 0;
     std::array<address_type, 4> addr_mode;
     
     uint64_t interrupt_sim=0;

gen_input/templates/interp/CORENAME.cpp.gtl

@@ -121,7 +121,7 @@ protected:
 
     inline void raise(uint16_t trap_id, uint16_t cause){
         auto trap_val =  0x80ULL << 24 | (cause << 16) | trap_id;
-        this->template get_reg<uint32_t>(traits::TRAP_STATE) = trap_val;
+        this->core.trap_state = trap_val;
         this->template get_reg<uint32_t>(traits::NEXT_PC) = std::numeric_limits<uint32_t>::max();
     }
 
@@ -141,39 +141,39 @@ protected:
     T& pc_assign(T& val){super::ex_info.branch_taken=true; return val;}
     inline uint8_t readSpace1(typename super::mem_type_e space, uint64_t addr){
         auto ret = super::template read_mem<uint8_t>(space, addr);
-        if(this->template get_reg<uint32_t>(traits::TRAP_STATE)) throw 0;
+        if(this->core.trap_state) throw 0;
         return ret;
     }
     inline uint16_t readSpace2(typename super::mem_type_e space, uint64_t addr){
         auto ret = super::template read_mem<uint16_t>(space, addr);
-        if(this->template get_reg<uint32_t>(traits::TRAP_STATE)) throw 0;
+        if(this->core.trap_state) throw 0;
         return ret;
     }
     inline uint32_t readSpace4(typename super::mem_type_e space, uint64_t addr){
         auto ret = super::template read_mem<uint32_t>(space, addr);
-        if(this->template get_reg<uint32_t>(traits::TRAP_STATE)) throw 0;
+        if(this->core.trap_state) throw 0;
         return ret;
     }
     inline uint64_t readSpace8(typename super::mem_type_e space, uint64_t addr){
         auto ret = super::template read_mem<uint64_t>(space, addr);
-        if(this->template get_reg<uint32_t>(traits::TRAP_STATE)) throw 0;
+        if(this->core.trap_state) throw 0;
         return ret;
     }
     inline void writeSpace1(typename super::mem_type_e space, uint64_t addr, uint8_t data){
         super::write_mem(space, addr, data);
-        if(this->template get_reg<uint32_t>(traits::TRAP_STATE)) throw 0;
+        if(this->core.trap_state) throw 0;
     }
     inline void writeSpace2(typename super::mem_type_e space, uint64_t addr, uint16_t data){
         super::write_mem(space, addr, data);
-        if(this->template get_reg<uint32_t>(traits::TRAP_STATE)) throw 0;
+        if(this->core.trap_state) throw 0;
     }
     inline void writeSpace4(typename super::mem_type_e space, uint64_t addr, uint32_t data){
         super::write_mem(space, addr, data);
-        if(this->template get_reg<uint32_t>(traits::TRAP_STATE)) throw 0;
+        if(this->core.trap_state) throw 0;
     }
     inline void writeSpace8(typename super::mem_type_e space, uint64_t addr, uint64_t data){
         super::write_mem(space, addr, data);
-        if(this->template get_reg<uint32_t>(traits::TRAP_STATE)) throw 0;
+        if(this->core.trap_state) throw 0;
     }
     template<unsigned W, typename U, typename S = typename std::make_signed<U>::type>
     inline S sext(U from) {
@@ -277,16 +277,16 @@ typename arch::traits<ARCH>::opcode_e vm_impl<ARCH>::decode_inst_id(code_word_t
 
 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){
-    // we fetch at max 4 byte, alignment is 2
-    code_word_t instr = 0;
-    auto *const data = (uint8_t *)&instr;
     auto pc=start;
-
     auto* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
     auto* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
-    auto* trap_state = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::TRAP_STATE]);
-    auto* icount = reinterpret_cast<uint64_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::ICOUNT]);
-    auto* instret = reinterpret_cast<uint64_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::INSTRET]);
+    auto& trap_state = this->core.trap_state;
+    auto& icount = this->core.icount;
+    auto& cycle = this->core.cycle;
+    auto& instret = this->core.instret;
+    auto& instr = this->core.instruction;
+    // we fetch at max 4 byte, alignment is 2
+    auto *const data = reinterpret_cast<uint8_t*>(&instr);
 
     while(!this->core.should_stop() &&
             !(is_count_limit_enabled(cond) && this->core.get_icount() >= icount_limit)){
@@ -314,11 +314,9 @@ typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e co
 		        <%}}%>// calculate next pc value
 		        *NEXT_PC = *PC + ${instr.length/8};
 		        // execute instruction
-		        try {
 		        <%instr.behavior.eachLine{%>${it}
-		        <%}%>} catch(...){}
-	    	}
-	    	break;<%}%>
+		        <%}%>TRAP_${instr.name}:break;
+	    	}// @suppress("No break at end of case")<%}%>
             default: {
                 *NEXT_PC = *PC + ((instr & 3) == 3 ? 4 : 2);
                 raise(0,  2);
@@ -328,16 +326,16 @@ typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e co
             process_spawn_blocks();
             if(this->sync_exec && POST_SYNC) this->do_sync(POST_SYNC, static_cast<unsigned>(inst_id));
             // trap check
-            if(*trap_state!=0){
-                super::core.enter_trap(*trap_state, pc.val, instr);
+            if(trap_state!=0){
+                super::core.enter_trap(trap_state, pc.val, instr);
             } else {
-                (*icount)++;
-                (*instret)++;
+                icount++;
+                instret++;
             }
-            (*reinterpret_cast<uint64_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::CYCLE]))++;
+            cycle++;
             pc.val=*NEXT_PC;
             this->core.reg.PC = this->core.reg.NEXT_PC;
-            this->core.reg.trap_state = this->core.reg.pending_trap;
+            this->core.trap_state = this->core.pending_trap;
         }
     }
     return pc;

incl/iss/arch/riscv_hart_common.h

@@ -188,7 +188,7 @@ enum {
 
 template <typename T> inline bool PTE_TABLE(T PTE) { return (((PTE) & (PTE_V | PTE_R | PTE_W | PTE_X)) == PTE_V); }
 
-enum { PRIV_U = 0, PRIV_S = 1, PRIV_M = 3 };
+enum { PRIV_U = 0, PRIV_S = 1, PRIV_M = 3, PRIV_D = 4};
 
 enum {
     ISA_A = 1,

incl/iss/arch/riscv_hart_m_p.h

@@ -203,7 +203,7 @@ public:
 
     void disass_output(uint64_t pc, const std::string instr) override {
         CLOG(INFO, disass) << fmt::format("0x{:016x}    {:40} [s:0x{:x};c:{}]",
-                pc, instr, (reg_t)state.mstatus, this->reg.icount + cycle_offset);
+                pc, instr, (reg_t)state.mstatus, this->icount + cycle_offset);
     };
 
     iss::instrumentation_if *get_instrumentation_if() override { return &instr_if; }
@@ -235,13 +235,17 @@ protected:
          */
         const std::string core_type_name() const override { return traits<BASE>::core_type; }
 
-        uint64_t get_pc() override { return arch.get_pc(); };
+        uint64_t get_pc() override { return arch.reg.PC; };
 
-        uint64_t get_next_pc() override { return arch.get_next_pc(); };
+        uint64_t get_next_pc() override { return arch.reg.NEXT_PC; };
 
-        uint64_t get_instr_count() { return arch.reg.icount; }
+        uint64_t get_instr_word() override { return arch.instruction; }
 
-        uint64_t get_total_cycles() override { return arch.reg.icount + arch.cycle_offset; }
+        uint64_t get_instr_count() override { return arch.icount; }
+
+        uint64_t get_pendig_traps() override { return arch.trap_state; }
+
+        uint64_t get_total_cycles() override { return arch.icount + arch.cycle_offset; }
 
         void set_curr_instr_cycles(unsigned cycles) override { arch.cycle_offset += cycles - 1; };
 
@@ -249,8 +253,6 @@ protected:
     };
 
     friend struct riscv_instrumentation_if;
-    addr_t get_pc() { return this->reg.PC; }
-    addr_t get_next_pc() { return this->reg.NEXT_PC; }
 
     virtual iss::status read_mem(phys_addr_t addr, unsigned length, uint8_t *const data);
     virtual iss::status write_mem(phys_addr_t addr, unsigned length, const uint8_t *const data);
@@ -565,22 +567,22 @@ iss::status riscv_hart_m_p<BASE, FEAT>::read(const address_type type, const acce
     try {
         switch (space) {
         case traits<BASE>::MEM: {
-            if (unlikely((access == iss::access_type::FETCH || access == iss::access_type::DEBUG_FETCH) && (addr & 0x1) == 1)) {
+            auto alignment = is_fetch(access)? (traits<BASE>::MISA_VAL&0x100? 2 : 4) : length;
+            if (unlikely(is_fetch(access) && (addr&(alignment-1)))) {
                 fault_data = addr;
-                if (access && iss::access_type::DEBUG) throw trap_access(0, addr);
-                this->reg.trap_state = (1 << 31); // issue trap 0
+                if (is_debug(access)) throw trap_access(0, addr);
+                this->trap_state = (1 << 31); // issue trap 0
                 return iss::Err;
             }
             try {
-                auto alignment = access == iss::access_type::FETCH? (traits<BASE>::MISA_VAL&0x100? 2 : 4) : length;
-                if(alignment>1 && (addr&(alignment-1))){
-                    this->reg.trap_state = 1<<31 | 4<<16;
+                if(!is_debug(access) && (addr&(alignment-1))){
+                    this->trap_state = 1<<31 | 4<<16;
                     fault_data=addr;
                     return iss::Err;
                 }
                 auto phys_addr = type==iss::address_type::PHYSICAL?phys_addr_t{access, space, addr}:BASE::v2p(iss::addr_t{access, type, space, addr});
                 auto res = iss::Err;
-                if(access != access_type::FETCH && memfn_range.size()){
+                if(!is_fetch(access) && memfn_range.size()){
                     auto it = std::find_if(std::begin(memfn_range), std::end(memfn_range), [phys_addr](std::tuple<uint64_t, uint64_t> const& a){
                         return std::get<0>(a)<=phys_addr.val && (std::get<0>(a)+std::get<1>(a))>phys_addr.val;
                     });
@@ -593,12 +595,12 @@ iss::status riscv_hart_m_p<BASE, FEAT>::read(const address_type type, const acce
                     res = read_mem( phys_addr, length, data);
                 }
                 if (unlikely(res != iss::Ok)){
-                    this->reg.trap_state = (1 << 31) | (5 << 16); // issue trap 5 (load access fault
+                    this->trap_state = (1 << 31) | (5 << 16); // issue trap 5 (load access fault
                     fault_data=addr;
                 }
                 return res;
             } catch (trap_access &ta) {
-                this->reg.trap_state = (1 << 31) | ta.id;
+                this->trap_state = (1 << 31) | ta.id;
                 fault_data=ta.addr;
                 return iss::Err;
             }
@@ -624,7 +626,7 @@ iss::status riscv_hart_m_p<BASE, FEAT>::read(const address_type type, const acce
         }
         return iss::Ok;
     } catch (trap_access &ta) {
-        this->reg.trap_state = (1 << 31) | ta.id;
+        this->trap_state = (1 << 31) | ta.id;
         fault_data=ta.addr;
         return iss::Err;
     }
@@ -662,12 +664,12 @@ iss::status riscv_hart_m_p<BASE, FEAT>::write(const address_type type, const acc
             if (unlikely((access && iss::access_type::FETCH) && (addr & 0x1) == 1)) {
                 fault_data = addr;
                 if (access && iss::access_type::DEBUG) throw trap_access(0, addr);
-                this->reg.trap_state = (1 << 31); // issue trap 0
+                this->trap_state = (1 << 31); // issue trap 0
                 return iss::Err;
             }
             try {
                 if(length>1 && (addr&(length-1)) && (access&access_type::DEBUG) != access_type::DEBUG){
-                    this->reg.trap_state = 1<<31 | 6<<16;
+                    this->trap_state = 1<<31 | 6<<16;
                     fault_data=addr;
                     return iss::Err;
                 }
@@ -686,12 +688,12 @@ iss::status riscv_hart_m_p<BASE, FEAT>::write(const address_type type, const acc
                     res = write_mem( phys_addr, length, data);
                 }
                 if (unlikely(res != iss::Ok)) {
-                    this->reg.trap_state = (1 << 31) | (7 << 16); // issue trap 7 (Store/AMO access fault)
+                    this->trap_state = (1 << 31) | (7 << 16); // issue trap 7 (Store/AMO access fault)
                     fault_data=addr;
                 }
                 return res;
             } catch (trap_access &ta) {
-                this->reg.trap_state = (1 << 31) | ta.id;
+                this->trap_state = (1 << 31) | ta.id;
                 fault_data=ta.addr;
                 return iss::Err;
             }
@@ -751,7 +753,7 @@ iss::status riscv_hart_m_p<BASE, FEAT>::write(const address_type type, const acc
         }
         return iss::Ok;
     } catch (trap_access &ta) {
-        this->reg.trap_state = (1 << 31) | ta.id;
+        this->trap_state = (1 << 31) | ta.id;
         fault_data=ta.addr;
         return iss::Err;
     }
@@ -797,7 +799,7 @@ template <typename BASE, features_e FEAT> iss::status riscv_hart_m_p<BASE, FEAT>
 }
 
 template <typename BASE, features_e FEAT> iss::status riscv_hart_m_p<BASE, FEAT>::read_cycle(unsigned addr, reg_t &val) {
-    auto cycle_val = this->reg.icount + cycle_offset;
+    auto cycle_val = this->icount + cycle_offset;
     if (addr == mcycle) {
         val = static_cast<reg_t>(cycle_val);
     } else if (addr == mcycleh) {
@@ -819,16 +821,16 @@ template <typename BASE, features_e FEAT> iss::status riscv_hart_m_p<BASE, FEAT>
             mcycle_csr = (static_cast<uint64_t>(val)<<32) + (mcycle_csr & 0xffffffff);
         }
     }
-    cycle_offset = mcycle_csr-this->reg.icount; // TODO: relying on wrap-around
+    cycle_offset = mcycle_csr-this->icount; // TODO: relying on wrap-around
     return iss::Ok;
 }
 
 template <typename BASE, features_e FEAT> iss::status riscv_hart_m_p<BASE, FEAT>::read_instret(unsigned addr, reg_t &val) {
     if ((addr&0xff) == (minstret&0xff)) {
-        val = static_cast<reg_t>(this->reg.instret);
+        val = static_cast<reg_t>(this->instret);
     } else if ((addr&0xff) == (minstreth&0xff)) {
         if (sizeof(typename traits<BASE>::reg_t) != 4) return iss::Err;
-        val = static_cast<reg_t>(this->reg.instret >> 32);
+        val = static_cast<reg_t>(this->instret >> 32);
     }
     return iss::Ok;
 }
@@ -837,20 +839,20 @@ template <typename BASE, features_e FEAT> iss::status riscv_hart_m_p<BASE, FEAT>
     if (sizeof(typename traits<BASE>::reg_t) != 4) {
         if ((addr&0xff) == (minstreth&0xff))
             return iss::Err;
-        this->reg.instret = static_cast<uint64_t>(val);
+        this->instret = static_cast<uint64_t>(val);
     } else {
         if ((addr&0xff) == (minstret&0xff)) {
-            this->reg.instret = (this->reg.instret & 0xffffffff00000000) + val;
+            this->instret = (this->instret & 0xffffffff00000000) + val;
         } else  {
-            this->reg.instret = (static_cast<uint64_t>(val)<<32) + (this->reg.instret & 0xffffffff);
+            this->instret = (static_cast<uint64_t>(val)<<32) + (this->instret & 0xffffffff);
         }
     }
-    this->reg.instret--;
+    this->instret--;
     return iss::Ok;
 }
 
 template <typename BASE, features_e FEAT> iss::status riscv_hart_m_p<BASE, FEAT>::read_time(unsigned addr, reg_t &val) {
-    uint64_t time_val = this->reg.icount / (100000000 / 32768 - 1); //-> ~3052;
+    uint64_t time_val = this->icount / (100000000 / 32768 - 1); //-> ~3052;
     if (addr == time) {
         val = static_cast<reg_t>(time_val);
     } else if (addr == timeh) {
@@ -969,7 +971,7 @@ iss::status riscv_hart_m_p<BASE, FEAT>::read_mem(phys_addr_t paddr, unsigned len
         const mem_type::page_type &p = mem(paddr.val / mem.page_size);
         uint64_t offs = paddr.val & mem.page_addr_mask;
         std::copy(p.data() + offs, p.data() + offs + length, data);
-        if (this->reg.icount > 30000) data[3] |= 0x80;
+        if (this->icount > 30000) data[3] |= 0x80;
     } break;
     default: {
         for(auto offs=0U; offs<length; ++offs) {
@@ -1029,7 +1031,7 @@ iss::status riscv_hart_m_p<BASE, FEAT>::write_mem(phys_addr_t paddr, unsigned le
                             LOG(INFO) << "tohost value is 0x" << std::hex << hostvar << std::dec << " (" << hostvar
                                       << "), stopping simulation";
                         }
-                        this->reg.trap_state=std::numeric_limits<uint32_t>::max();
+                        this->trap_state=std::numeric_limits<uint32_t>::max();
                         this->interrupt_sim=hostvar;
                         break;
                         //throw(iss::simulation_stopped(hostvar));
@@ -1117,7 +1119,7 @@ template <typename BASE, features_e FEAT> void riscv_hart_m_p<BASE, FEAT>::check
         	enabled_interrupts >>= 1;
         	res++;
         }
-        this->reg.pending_trap = res << 16 | 1; // 0x80 << 24 | (cause << 16) | trap_id
+        this->pending_trap = res << 16 | 1; // 0x80 << 24 | (cause << 16) | trap_id
     }
 }
 
@@ -1127,6 +1129,7 @@ template <typename BASE, features_e FEAT> uint64_t riscv_hart_m_p<BASE, FEAT>::e
     auto trap_id = bit_sub<0, 16>(flags);
     auto cause = bit_sub<16, 15>(flags);
     // calculate effective privilege level
+    unsigned new_priv = PRIV_M;
     if (trap_id == 0) { // exception
         if (cause == 11) cause = 0x8 + PRIV_M; // adjust environment call cause
         // store ret addr in xepc register
@@ -1146,10 +1149,13 @@ template <typename BASE, features_e FEAT> uint64_t riscv_hart_m_p<BASE, FEAT>::e
             csr[mtval] = (instr & 0x3)==3?instr:instr&0xffff;
             break;
         case 3:
-            //TODO: implement debug mode behavior
-            // csr[dpc] = addr;
-            // csr[dcsr] = (csr[dcsr] & ~0x1c3) | (1<<6) | PRIV_M; //FIXME: cause should not be 4 (stepi)
+            if((FEAT & FEAT_DEBUG) && (csr[dcsr] & 0x8000)) {
+                this->reg.DPC = addr;
+                csr[dcsr] = (csr[dcsr] & ~0x1c3) | (1<<6) | PRIV_M; //FIXME: cause should not be 4 (stepi)
+                new_priv = this->reg.PRIV | PRIV_D;
+            } else {
                 csr[mtval] = addr;
+            }
             break;
         case 4:
         case 6:
@@ -1161,7 +1167,7 @@ template <typename BASE, features_e FEAT> uint64_t riscv_hart_m_p<BASE, FEAT>::e
         fault_data = 0;
     } else {
         csr[mepc] = this->reg.NEXT_PC & get_pc_mask(); // store next address if interrupt
-        this->reg.pending_trap = 0;
+        this->pending_trap = 0;
     }
     csr[mcause] = (trap_id << (traits<BASE>::XLEN-1)) + cause;
     // update mstatus
@@ -1182,8 +1188,8 @@ template <typename BASE, features_e FEAT> uint64_t riscv_hart_m_p<BASE, FEAT>::e
     this->reg.NEXT_PC = ivec & ~0x3UL;
     if ((ivec & 0x1) == 1 && trap_id != 0) this->reg.NEXT_PC += 4 * cause;
     // reset trap state
-    this->reg.PRIV = PRIV_M;
-    this->reg.trap_state = 0;
+    this->reg.PRIV = new_priv;
+    this->trap_state = 0;
     std::array<char, 32> buffer;
 #if defined(_MSC_VER)
     sprintf(buffer.data(), "0x%016llx", addr);

incl/iss/arch/riscv_hart_msu_vp.h

@@ -307,7 +307,7 @@ public:
 
     void disass_output(uint64_t pc, const std::string instr) override {
         CLOG(INFO, disass) << fmt::format("0x{:016x}    {:40} [p:{};s:0x{:x};c:{}]",
-                pc, instr, lvl[this->reg.PRIV], (reg_t)state.mstatus, this->reg.ccount + cycle_offset);
+                pc, instr, lvl[this->reg.PRIV], (reg_t)state.mstatus, this->icount + cycle_offset);
     };
 
     iss::instrumentation_if *get_instrumentation_if() override { return &instr_if; }
@@ -340,9 +340,13 @@ protected:
 
         virtual uint64_t get_next_pc() { return arch.get_next_pc(); };
 
-        uint64_t get_instr_count() { return arch.reg.icount; }
+        uint64_t get_instr_word() override { return arch.instruction; }
 
-        uint64_t get_total_cycles() override { return arch.reg.icount + arch.cycle_offset; }
+        uint64_t get_instr_count() { return arch.icount; }
+
+        uint64_t get_pendig_traps() override { return arch.trap_state; }
+
+        uint64_t get_total_cycles() override { return arch.icount + arch.cycle_offset; }
 
         virtual void set_curr_instr_cycles(unsigned cycles) { arch.cycle_offset += cycles - 1; };
 
@@ -607,13 +611,19 @@ iss::status riscv_hart_msu_vp<BASE>::read(const address_type type, const access_
     try {
         switch (space) {
         case traits<BASE>::MEM: {
-            if (unlikely((access == iss::access_type::FETCH || access == iss::access_type::DEBUG_FETCH) && (addr & 0x1) == 1)) {
+            auto alignment = is_fetch(access)? (traits<BASE>::MISA_VAL&0x100? 2 : 4) : length;
+            if (unlikely(is_fetch(access) && (addr&(alignment-1)))) {
                 fault_data = addr;
                 if (access && iss::access_type::DEBUG) throw trap_access(0, addr);
-                this->reg.trap_state = (1 << 31); // issue trap 0
+                this->trap_state = (1 << 31); // issue trap 0
                 return iss::Err;
             }
             try {
+                if(!is_debug(access)  && (addr&(alignment-1))){
+                    this->trap_state = 1<<31 | 4<<16;
+                    fault_data=addr;
+                    return iss::Err;
+                }
                 if (unlikely((addr & ~PGMASK) != ((addr + length - 1) & ~PGMASK))) { // we may cross a page boundary
                     vm_info vm = hart_state_type::decode_vm_info(this->reg.PRIV, state.satp);
                     if (vm.levels != 0) { // VM is active
@@ -629,12 +639,12 @@ iss::status riscv_hart_msu_vp<BASE>::read(const address_type type, const access_
                         read_mem( BASE::v2p(phys_addr_t{access, space, addr}), length, data):
                         read_mem( BASE::v2p(iss::addr_t{access, type, space, addr}), length, data);
                 if (unlikely(res != iss::Ok)){
-                	this->reg.trap_state = (1 << 31) | (5 << 16); // issue trap 5 (load access fault
+                	this->trap_state = (1 << 31) | (5 << 16); // issue trap 5 (load access fault
                     fault_data=addr;
                 }
                 return res;
             } catch (trap_access &ta) {
-                this->reg.trap_state = (1 << 31) | ta.id;
+                this->trap_state = (1 << 31) | ta.id;
                 fault_data=ta.addr;
                 return iss::Err;
             }
@@ -650,7 +660,7 @@ iss::status riscv_hart_msu_vp<BASE>::read(const address_type type, const access_
             case 3: { // SFENCE:VMA upper
                 auto tvm = state.mstatus.TVM;
                 if (this->reg.PRIV == PRIV_S & tvm != 0) {
-                    this->reg.trap_state = (1 << 31) | (2 << 16);
+                    this->trap_state = (1 << 31) | (2 << 16);
                     this->fault_data = this->reg.PC;
                     return iss::Err;
                 }
@@ -671,7 +681,7 @@ iss::status riscv_hart_msu_vp<BASE>::read(const address_type type, const access_
         }
         return iss::Ok;
     } catch (trap_access &ta) {
-        this->reg.trap_state = (1 << 31) | ta.id;
+        this->trap_state = (1 << 31) | ta.id;
         fault_data=ta.addr;
         return iss::Err;
     }
@@ -709,7 +719,7 @@ iss::status riscv_hart_msu_vp<BASE>::write(const address_type type, const access
             if (unlikely((access && iss::access_type::FETCH) && (addr & 0x1) == 1)) {
                 fault_data = addr;
                 if (access && iss::access_type::DEBUG) throw trap_access(0, addr);
-                this->reg.trap_state = (1 << 31); // issue trap 0
+                this->trap_state = (1 << 31); // issue trap 0
                 return iss::Err;
             }
             try {
@@ -728,12 +738,12 @@ iss::status riscv_hart_msu_vp<BASE>::write(const address_type type, const access
                         write_mem(phys_addr_t{access, space, addr}, length, data):
                         write_mem(BASE::v2p(iss::addr_t{access, type, space, addr}), length, data);
                 if (unlikely(res != iss::Ok)) {
-                    this->reg.trap_state = (1 << 31) | (7 << 16); // issue trap 7 (Store/AMO access fault)
+                    this->trap_state = (1 << 31) | (7 << 16); // issue trap 7 (Store/AMO access fault)
                     fault_data=addr;
                 }
                 return res;
             } catch (trap_access &ta) {
-                this->reg.trap_state = (1 << 31) | ta.id;
+                this->trap_state = (1 << 31) | ta.id;
                 fault_data=ta.addr;
                 return iss::Err;
             }
@@ -782,7 +792,7 @@ iss::status riscv_hart_msu_vp<BASE>::write(const address_type type, const access
                 ptw.clear();
                 auto tvm = state.mstatus.TVM;
                 if (this->reg.PRIV == PRIV_S & tvm != 0) {
-                    this->reg.trap_state = (1 << 31) | (2 << 16);
+                    this->trap_state = (1 << 31) | (2 << 16);
                     this->fault_data = this->reg.PC;
                     return iss::Err;
                 }
@@ -798,7 +808,7 @@ iss::status riscv_hart_msu_vp<BASE>::write(const address_type type, const access
         }
         return iss::Ok;
     } catch (trap_access &ta) {
-        this->reg.trap_state = (1 << 31) | ta.id;
+        this->trap_state = (1 << 31) | ta.id;
         fault_data=ta.addr;
         return iss::Err;
     }
@@ -844,7 +854,7 @@ template <typename BASE> iss::status riscv_hart_msu_vp<BASE>::write_reg(unsigned
 }
 
 template <typename BASE> iss::status riscv_hart_m_p<BASE>::read_cycle(unsigned addr, reg_t &val) {
-    auto cycle_val = this->reg.icount + cycle_offset;
+    auto cycle_val = this->icount + cycle_offset;
     if (addr == mcycle) {
         val = static_cast<reg_t>(cycle_val);
     } else if (addr == mcycleh) {
@@ -866,7 +876,7 @@ template <typename BASE> iss::status riscv_hart_m_p<BASE>::write_cycle(unsigned
             mcycle_csr = (static_cast<uint64_t>(val)<<32) + (mcycle_csr & 0xffffffff);
         }
     }
-    cycle_offset = mcycle_csr-this->reg.icount; // TODO: relying on wrap-around
+    cycle_offset = mcycle_csr-this->icount; // TODO: relying on wrap-around
     return iss::Ok;
 }
 
@@ -897,7 +907,7 @@ template <typename BASE> iss::status riscv_hart_m_p<BASE>::write_instret(unsigne
 }
 
 template <typename BASE> iss::status riscv_hart_m_p<BASE>::read_time(unsigned addr, reg_t &val) {
-    uint64_t time_val = this->reg.icount / (100000000 / 32768 - 1); //-> ~3052;
+    uint64_t time_val = this->icount / (100000000 / 32768 - 1); //-> ~3052;
     if (addr == time) {
         val = static_cast<reg_t>(time_val);
     } else if (addr == timeh) {
@@ -966,7 +976,7 @@ template <typename BASE> iss::status riscv_hart_msu_vp<BASE>::write_epc(unsigned
 template <typename BASE> iss::status riscv_hart_msu_vp<BASE>::read_satp(unsigned addr, reg_t &val) {
     reg_t tvm = state.mstatus.TVM;
     if (this->reg.PRIV == PRIV_S & tvm != 0) {
-        this->reg.trap_state = (1 << 31) | (2 << 16);
+        this->trap_state = (1 << 31) | (2 << 16);
         this->fault_data = this->reg.PC;
         return iss::Err;
     }
@@ -977,7 +987,7 @@ template <typename BASE> iss::status riscv_hart_msu_vp<BASE>::read_satp(unsigned
 template <typename BASE> iss::status riscv_hart_msu_vp<BASE>::write_satp(unsigned addr, reg_t val) {
     reg_t tvm = state.mstatus.TVM;
     if (this->reg.PRIV == PRIV_S & tvm != 0) {
-        this->reg.trap_state = (1 << 31) | (2 << 16);
+        this->trap_state = (1 << 31) | (2 << 16);
         this->fault_data = this->reg.PC;
         return iss::Err;
     }
@@ -1033,7 +1043,7 @@ iss::status riscv_hart_msu_vp<BASE>::read_mem(phys_addr_t paddr, unsigned length
         const mem_type::page_type &p = mem(paddr.val / mem.page_size);
         uint64_t offs = paddr.val & mem.page_addr_mask;
         std::copy(p.data() + offs, p.data() + offs + length, data);
-        if (this->reg.icount > 30000) data[3] |= 0x80;
+        if (this->icount > 30000) data[3] |= 0x80;
     } break;
     default: {
         for(auto offs=0U; offs<length; ++offs) {
@@ -1093,7 +1103,7 @@ iss::status riscv_hart_msu_vp<BASE>::write_mem(phys_addr_t paddr, unsigned lengt
                             LOG(INFO) << "tohost value is 0x" << std::hex << hostvar << std::dec << " (" << hostvar
                                       << "), stopping simulation";
                         }
-                        this->reg.trap_state=std::numeric_limits<uint32_t>::max();
+                        this->trap_state=std::numeric_limits<uint32_t>::max();
                         this->interrupt_sim=hostvar;
                         break;
                         //throw(iss::simulation_stopped(hostvar));
@@ -1162,7 +1172,7 @@ template <typename BASE> void riscv_hart_msu_vp<BASE>::check_interrupt() {
     if (enabled_interrupts != 0) {
         int res = 0;
         while ((enabled_interrupts & 1) == 0) enabled_interrupts >>= 1, res++;
-        this->reg.pending_trap = res << 16 | 1; // 0x80 << 24 | (cause << 16) | trap_id
+        this->pending_trap = res << 16 | 1; // 0x80 << 24 | (cause << 16) | trap_id
     }
 }
 
@@ -1306,7 +1316,7 @@ template <typename BASE> uint64_t riscv_hart_msu_vp<BASE>::enter_trap(uint64_t f
         if (cur_priv != PRIV_M && ((csr[mideleg] >> cause) & 0x1) != 0)
             new_priv = (csr[sideleg] >> cause) & 0x1 ? PRIV_U : PRIV_S;
         csr[uepc | (new_priv << 8)] = this->reg.NEXT_PC; // store next address if interrupt
-        this->reg.pending_trap = 0;
+        this->pending_trap = 0;
     }
     size_t adr = ucause | (new_priv << 8);
     csr[adr] = (trap_id << 31) + cause;
@@ -1351,7 +1361,7 @@ template <typename BASE> uint64_t riscv_hart_msu_vp<BASE>::enter_trap(uint64_t f
                        << lvl[cur_priv] << " to " << lvl[new_priv];
     // reset trap state
     this->reg.PRIV = new_priv;
-    this->reg.trap_state = 0;
+    this->trap_state = 0;
     update_vm_info();
     return this->reg.NEXT_PC;
 }
@@ -1363,7 +1373,7 @@ template <typename BASE> uint64_t riscv_hart_msu_vp<BASE>::leave_trap(uint64_t f
 
     auto tsr = state.mstatus.TSR;
     if (cur_priv == PRIV_S && inst_priv == PRIV_S && tsr != 0) {
-        this->reg.trap_state = (1 << 31) | (2 << 16);
+        this->trap_state = (1 << 31) | (2 << 16);
         this->fault_data = this->reg.PC;
         return this->reg.PC;
     }
@@ -1402,7 +1412,7 @@ template <typename BASE> void riscv_hart_msu_vp<BASE>::wait_until(uint64_t flags
     auto status = state.mstatus;
     auto tw = status.TW;
     if (this->reg.PRIV == PRIV_S && tw != 0) {
-        this->reg.trap_state = (1 << 31) | (2 << 16);
+        this->trap_state = (1 << 31) | (2 << 16);
         this->fault_data = this->reg.PC;
     }
 }

incl/iss/arch/riscv_hart_mu_p.h

@@ -218,7 +218,7 @@ public:
 
     void disass_output(uint64_t pc, const std::string instr) override {
         CLOG(INFO, disass) << fmt::format("0x{:016x}    {:40} [p:{};s:0x{:x};c:{}]",
-                pc, instr, lvl[this->reg.PRIV], (reg_t)state.mstatus, this->reg.icount + cycle_offset);
+                pc, instr, lvl[this->reg.PRIV], (reg_t)state.mstatus, this->icount + cycle_offset);
     };
 
     iss::instrumentation_if *get_instrumentation_if() override { return &instr_if; }
@@ -254,9 +254,13 @@ protected:
 
         virtual uint64_t get_next_pc() { return arch.get_next_pc(); };
 
-        uint64_t get_instr_count() { return arch.reg.icount; }
+        uint64_t get_instr_word() override { return arch.instruction; }
 
-        uint64_t get_total_cycles() override { return arch.reg.icount + arch.cycle_offset; }
+        uint64_t get_instr_count() { return arch.icount; }
+
+        uint64_t get_pendig_traps() override { return arch.trap_state; }
+
+        uint64_t get_total_cycles() override { return arch.icount + arch.cycle_offset; }
 
         virtual void set_curr_instr_cycles(unsigned cycles) { arch.cycle_offset += cycles - 1; };
 
@@ -701,23 +705,23 @@ iss::status riscv_hart_mu_p<BASE, FEAT>::read(const address_type type, const acc
         switch (space) {
         case traits<BASE>::MEM: {
             if(FEAT & FEAT_PMP){
-                if(!pmp_check(access, addr, length) && (access&access_type::DEBUG) != access_type::DEBUG) {
+                if(!pmp_check(access, addr, length) && !is_debug(access)) {
                     fault_data = addr;
-                    if (access && iss::access_type::DEBUG) throw trap_access(0, addr);
-                    this->reg.trap_state = (1 << 31) | ((access==access_type::FETCH?1:5) << 16); // issue trap 1
+                    if (is_debug(access)) throw trap_access(0, addr);
+                    this->trap_state = (1 << 31) | ((access==access_type::FETCH?1:5) << 16); // issue trap 1
                     return iss::Err;
                 }
             }
-            if (unlikely((access == iss::access_type::FETCH || access == iss::access_type::DEBUG_FETCH) && (addr & 0x1) == 1)) {
+            auto alignment = is_fetch(access)? (traits<BASE>::MISA_VAL&0x100? 2 : 4) : length;
+            if (unlikely(is_fetch(access) && (addr&(alignment-1)))) {
                 fault_data = addr;
-                if (access && iss::access_type::DEBUG) throw trap_access(0, addr);
-                this->reg.trap_state = (1 << 31); // issue trap 0
+                if (is_debug(access)) throw trap_access(0, addr);
+                this->trap_state = (1 << 31); // issue trap 0
                 return iss::Err;
             }
             try {
-                auto alignment = access == iss::access_type::FETCH? (traits<BASE>::MISA_VAL&0x100? 2 : 4) : length;
-                if(alignment>1 && (addr&(alignment-1))){
-                    this->reg.trap_state = 1<<31 | 4<<16;
+                if(!is_debug(access)  && (addr&(alignment-1))){
+                    this->trap_state = 1<<31 | 4<<16;
                     fault_data=addr;
                     return iss::Err;
                 }
@@ -736,12 +740,12 @@ iss::status riscv_hart_mu_p<BASE, FEAT>::read(const address_type type, const acc
                     res = read_mem( phys_addr, length, data);
                 }
                 if (unlikely(res != iss::Ok)){
-                    this->reg.trap_state = (1 << 31) | (5 << 16); // issue trap 5 (load access fault
+                    this->trap_state = (1 << 31) | (5 << 16); // issue trap 5 (load access fault
                     fault_data=addr;
                 }
                 return res;
             } catch (trap_access &ta) {
-                this->reg.trap_state = (1 << 31) | ta.id;
+                this->trap_state = (1 << 31) | ta.id;
                 fault_data=ta.addr;
                 return iss::Err;
             }
@@ -767,7 +771,7 @@ iss::status riscv_hart_mu_p<BASE, FEAT>::read(const address_type type, const acc
         }
         return iss::Ok;
     } catch (trap_access &ta) {
-        this->reg.trap_state = (1 << 31) | ta.id;
+        this->trap_state = (1 << 31) | ta.id;
         fault_data=ta.addr;
         return iss::Err;
     }
@@ -806,19 +810,19 @@ iss::status riscv_hart_mu_p<BASE, FEAT>::write(const address_type type, const ac
                 if(!pmp_check(access, addr, length) && (access&access_type::DEBUG) != access_type::DEBUG) {
                     fault_data = addr;
                     if (access && iss::access_type::DEBUG) throw trap_access(0, addr);
-                    this->reg.trap_state = (1 << 31) | (7 << 16); // issue trap 1
+                    this->trap_state = (1 << 31) | (7 << 16); // issue trap 1
                     return iss::Err;
                 }
             }
             if (unlikely((access && iss::access_type::FETCH) && (addr & 0x1) == 1)) {
                 fault_data = addr;
                 if (access && iss::access_type::DEBUG) throw trap_access(0, addr);
-                this->reg.trap_state = (1 << 31); // issue trap 0
+                this->trap_state = (1 << 31); // issue trap 0
                 return iss::Err;
             }
             try {
                 if(length>1 && (addr&(length-1)) && (access&access_type::DEBUG) != access_type::DEBUG){
-                    this->reg.trap_state = 1<<31 | 6<<16;
+                    this->trap_state = 1<<31 | 6<<16;
                     fault_data=addr;
                     return iss::Err;
                 }
@@ -837,12 +841,12 @@ iss::status riscv_hart_mu_p<BASE, FEAT>::write(const address_type type, const ac
                     res = write_mem( phys_addr, length, data);
                 }
                 if (unlikely(res != iss::Ok)) {
-                    this->reg.trap_state = (1 << 31) | (7 << 16); // issue trap 7 (Store/AMO access fault)
+                    this->trap_state = (1 << 31) | (7 << 16); // issue trap 7 (Store/AMO access fault)
                     fault_data=addr;
                 }
                 return res;
             } catch (trap_access &ta) {
-                this->reg.trap_state = (1 << 31) | ta.id;
+                this->trap_state = (1 << 31) | ta.id;
                 fault_data=ta.addr;
                 return iss::Err;
             }
@@ -902,7 +906,7 @@ iss::status riscv_hart_mu_p<BASE, FEAT>::write(const address_type type, const ac
         }
         return iss::Ok;
     } catch (trap_access &ta) {
-        this->reg.trap_state = (1 << 31) | ta.id;
+        this->trap_state = (1 << 31) | ta.id;
         fault_data=ta.addr;
         return iss::Err;
     }
@@ -948,7 +952,7 @@ template <typename BASE, features_e FEAT> iss::status riscv_hart_mu_p<BASE, FEAT
 }
 
 template <typename BASE, features_e FEAT> iss::status riscv_hart_mu_p<BASE, FEAT>::read_cycle(unsigned addr, reg_t &val) {
-    auto cycle_val = this->reg.icount + cycle_offset;
+    auto cycle_val = this->icount + cycle_offset;
     if (addr == mcycle) {
         val = static_cast<reg_t>(cycle_val);
     } else if (addr == mcycleh) {
@@ -970,16 +974,16 @@ template <typename BASE, features_e FEAT> iss::status riscv_hart_mu_p<BASE, FEAT
             mcycle_csr = (static_cast<uint64_t>(val)<<32) + (mcycle_csr & 0xffffffff);
         }
     }
-    cycle_offset = mcycle_csr-this->reg.icount; // TODO: relying on wrap-around
+    cycle_offset = mcycle_csr-this->icount; // TODO: relying on wrap-around
     return iss::Ok;
 }
 
 template <typename BASE, features_e FEAT> iss::status riscv_hart_mu_p<BASE, FEAT>::read_instret(unsigned addr, reg_t &val) {
     if ((addr&0xff) == (minstret&0xff)) {
-        val = static_cast<reg_t>(this->reg.instret);
+        val = static_cast<reg_t>(this->instret);
     } else if ((addr&0xff) == (minstreth&0xff)) {
         if (sizeof(typename traits<BASE>::reg_t) != 4) return iss::Err;
-        val = static_cast<reg_t>(this->reg.instret >> 32);
+        val = static_cast<reg_t>(this->instret >> 32);
     }
     return iss::Ok;
 }
@@ -988,20 +992,20 @@ template <typename BASE, features_e FEAT> iss::status riscv_hart_mu_p<BASE, FEAT
     if (sizeof(typename traits<BASE>::reg_t) != 4) {
         if ((addr&0xff) == (minstreth&0xff))
             return iss::Err;
-        this->reg.instret = static_cast<uint64_t>(val);
+        this->instret = static_cast<uint64_t>(val);
     } else {
         if ((addr&0xff) == (minstret&0xff)) {
-            this->reg.instret = (this->reg.instret & 0xffffffff00000000) + val;
+            this->instret = (this->instret & 0xffffffff00000000) + val;
         } else  {
-            this->reg.instret = (static_cast<uint64_t>(val)<<32) + (this->reg.instret & 0xffffffff);
+            this->instret = (static_cast<uint64_t>(val)<<32) + (this->instret & 0xffffffff);
         }
     }
-    this->reg.instret--;
+    this->instret--;
     return iss::Ok;
 }
 
 template <typename BASE, features_e FEAT> iss::status riscv_hart_mu_p<BASE, FEAT>::read_time(unsigned addr, reg_t &val) {
-    uint64_t time_val = this->reg.icount / (100000000 / 32768 - 1); //-> ~3052;
+    uint64_t time_val = this->icount / (100000000 / 32768 - 1); //-> ~3052;
     if (addr == time) {
         val = static_cast<reg_t>(time_val);
     } else if (addr == timeh) {
@@ -1137,7 +1141,7 @@ iss::status riscv_hart_mu_p<BASE, FEAT>::read_mem(phys_addr_t paddr, unsigned le
         const mem_type::page_type &p = mem(paddr.val / mem.page_size);
         uint64_t offs = paddr.val & mem.page_addr_mask;
         std::copy(p.data() + offs, p.data() + offs + length, data);
-        if (this->reg.icount > 30000) data[3] |= 0x80;
+        if (this->icount > 30000) data[3] |= 0x80;
     } break;
     default: {
         for(auto offs=0U; offs<length; ++offs) {
@@ -1196,7 +1200,7 @@ iss::status riscv_hart_mu_p<BASE, FEAT>::write_mem(phys_addr_t paddr, unsigned l
                             LOG(INFO) << "tohost value is 0x" << std::hex << hostvar << std::dec << " (" << hostvar
                                       << "), stopping simulation";
                         }
-                        this->reg.trap_state=std::numeric_limits<uint32_t>::max();
+                        this->trap_state=std::numeric_limits<uint32_t>::max();
                         this->interrupt_sim=hostvar;
                         break;
                         //throw(iss::simulation_stopped(hostvar));
@@ -1284,19 +1288,19 @@ template <typename BASE, features_e FEAT> void riscv_hart_mu_p<BASE, FEAT>::chec
         	enabled_interrupts >>= 1;
         	res++;
         }
-        this->reg.pending_trap = res << 16 | 1; // 0x80 << 24 | (cause << 16) | trap_id
+        this->pending_trap = res << 16 | 1; // 0x80 << 24 | (cause << 16) | trap_id
     }
 }
 
 template <typename BASE, features_e FEAT> uint64_t riscv_hart_mu_p<BASE, FEAT>::enter_trap(uint64_t flags, uint64_t addr, uint64_t instr) {
     // flags are ACTIVE[31:31], CAUSE[30:16], TRAPID[15:0]
     // calculate and write mcause val
-    if(flags==std::numeric_limits<uint64_t>::max()) flags=this->reg.trap_state;
+    if(flags==std::numeric_limits<uint64_t>::max()) flags=this->trap_state;
     auto trap_id = bit_sub<0, 16>(flags);
     auto cause = bit_sub<16, 15>(flags);
     if (trap_id == 0 && cause == 11) cause = 0x8 + this->reg.PRIV; // adjust environment call cause
     // calculate effective privilege level
-    auto new_priv = PRIV_M;
+    unsigned new_priv = PRIV_M;
     if (trap_id == 0) { // exception
         if (this->reg.PRIV != PRIV_M && ((csr[medeleg] >> cause) & 0x1) != 0)
             new_priv = PRIV_U;
@@ -1317,10 +1321,13 @@ template <typename BASE, features_e FEAT> uint64_t riscv_hart_mu_p<BASE, FEAT>::
             csr[utval | (new_priv << 8)] = (instr & 0x3)==3?instr:instr&0xffff;
             break;
         case 3:
-            //TODO: implement debug mode behavior
-            // csr[dpc] = addr;
-            // csr[dcsr] = (csr[dcsr] & ~0x1c3) | (1<<6) | PRIV_M; //FIXME: cause should not be 4 (stepi)
+            if((FEAT & FEAT_DEBUG) && (csr[dcsr] & 0x8000)) {
+                this->reg.DPC = addr;
+                csr[dcsr] = (csr[dcsr] & ~0x1c3) | (1<<6) | PRIV_M; //FIXME: cause should not be 4 (stepi)
+                new_priv = this->reg.PRIV | PRIV_D;
+            } else {
                 csr[utval | (new_priv << 8)] = addr;
+            }
             break;
         case 4:
         case 6:
@@ -1335,7 +1342,7 @@ template <typename BASE, features_e FEAT> uint64_t riscv_hart_mu_p<BASE, FEAT>::
         if (this->reg.PRIV != PRIV_M && ((csr[mideleg] >> cause) & 0x1) != 0)
             new_priv = PRIV_U;
         csr[uepc | (new_priv << 8)] = this->reg.NEXT_PC; // store next address if interrupt
-        this->reg.pending_trap = 0;
+        this->pending_trap = 0;
     }
     size_t adr = ucause | (new_priv << 8);
     csr[adr] = (trap_id << (traits<BASE>::XLEN-1)) + cause;
@@ -1375,7 +1382,7 @@ template <typename BASE, features_e FEAT> uint64_t riscv_hart_mu_p<BASE, FEAT>::
                        << lvl[this->reg.PRIV] << " to " << lvl[new_priv];
     // reset trap state
     this->reg.PRIV = new_priv;
-    this->reg.trap_state = 0;
+    this->trap_state = 0;
     return this->reg.NEXT_PC;
 }
 
@@ -1384,7 +1391,7 @@ template <typename BASE, features_e FEAT> uint64_t riscv_hart_mu_p<BASE, FEAT>::
     auto inst_priv = (flags & 0x3)? 3:0;
     if(inst_priv>cur_priv){
         auto trap_val =  0x80ULL << 24 | (2 << 16); // illegal instruction
-        this->reg.trap_state = trap_val;
+        this->trap_state = trap_val;
         this->reg.NEXT_PC = std::numeric_limits<uint32_t>::max();
     } else {
         auto status = state.mstatus;

incl/iss/arch/tgc_c.h

@@ -58,12 +58,7 @@ template <> struct traits<tgc_c> {
     constexpr static unsigned FP_REGS_SIZE = 0;
 
     enum reg_e {
-        X0, X1, X2, X3, X4, X5, X6, X7, X8, X9, X10, X11, X12, X13, X14, X15, X16, X17, X18, X19, X20, X21, X22, X23, X24, X25, X26, X27, X28, X29, X30, X31, PC, NEXT_PC, PRIV, DPC, NUM_REGS,
-        TRAP_STATE=NUM_REGS,
-        PENDING_TRAP,
-        ICOUNT,
-        CYCLE,
-        INSTRET
+        X0, X1, X2, X3, X4, X5, X6, X7, X8, X9, X10, X11, X12, X13, X14, X15, X16, X17, X18, X19, X20, X21, X22, X23, X24, X25, X26, X27, X28, X29, X30, X31, PC, NEXT_PC, PRIV, DPC, NUM_REGS
     };
 
     using reg_t = uint32_t;
@@ -76,11 +71,11 @@ template <> struct traits<tgc_c> {
 
     using phys_addr_t = iss::typed_addr_t<iss::address_type::PHYSICAL>;
 
-    static constexpr std::array<const uint32_t, 41> reg_bit_widths{
-        {32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,8,32,32,32,64,64,64}};
+    static constexpr std::array<const uint32_t, 36> reg_bit_widths{
+        {32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,8,32}};
 
-    static constexpr std::array<const uint32_t, 41> reg_byte_offsets{
-        {0,4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,116,120,124,128,132,136,137,141,145,149,157,165}};
+    static constexpr std::array<const uint32_t, 36> reg_byte_offsets{
+        {0,4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,116,120,124,128,132,136,137}};
 
     static const uint64_t addr_mask = (reg_t(1) << (XLEN - 1)) | ((reg_t(1) << (XLEN - 1)) - 1);
 
@@ -197,7 +192,7 @@ struct tgc_c: public arch_if {
 
     uint8_t* get_regs_base_ptr() override;
 
-    inline uint64_t get_icount() { return reg.icount; }
+    inline uint64_t get_icount() { return icount; }
 
     inline bool should_stop() { return interrupt_sim; }
 
@@ -215,7 +210,7 @@ struct tgc_c: public arch_if {
 
     virtual iss::sync_type needed_sync() const { return iss::NO_SYNC; }
 
-    inline uint32_t get_last_branch() { return reg.last_branch; }
+    inline uint32_t get_last_branch() { return last_branch; }
 
 
 #pragma pack(push, 1)
@@ -256,12 +251,14 @@ struct tgc_c: public arch_if {
         uint32_t NEXT_PC = 0; 
         uint8_t PRIV = 0; 
         uint32_t DPC = 0;
-        uint32_t trap_state = 0, pending_trap = 0;
-        uint64_t icount = 0;
-        uint64_t instret = 0;
-        uint32_t last_branch;
     } reg;
 #pragma pack(pop)
+    uint32_t trap_state = 0, pending_trap = 0;
+    uint64_t icount = 0;
+    uint64_t cycle = 0;
+    uint64_t instret = 0;
+    uint32_t instruction = 0;
+    uint32_t last_branch = 0;
     std::array<address_type, 4> addr_mode;
     
     uint64_t interrupt_sim=0;

incl/iss/plugin/cycle_estimate.h

@@ -81,7 +81,7 @@ public:
     void callback(instr_info_t instr_info, exec_info const&) override;
 
 private:
-    iss::instrumentation_if *arch_instr;
+    iss::instrumentation_if *instr_if;
     std::vector<instr_desc> delays;
     struct pair_hash {
         size_t operator()(const std::pair<uint64_t, uint64_t> &p) const {

incl/iss/plugin/pctrace.h

@@ -44,7 +44,7 @@
 
 namespace iss {
 namespace plugin {
-
+class lz4compress_steambuf;
 class cov : public iss::vm_plugin {
     struct instr_delay {
         std::string instr_name;
@@ -88,10 +88,13 @@ public:
 private:
     iss::instrumentation_if *instr_if  {nullptr};
     std::ofstream output;
+#ifdef WITH_LZ4
+    std::unique_ptr<lz4compress_steambuf> strbuf;
+    std::ostream ostr;
+#endif
     std::string filename;
     std::vector<instr_desc> delays;
-    bool jumped, first;
-
+    bool jumped{false}, first{true};
 };
 }
 }

src/iss/tgc_c.cpp

@@ -41,12 +41,10 @@ using namespace iss::arch;
 
 constexpr std::array<const char*, 36>    iss::arch::traits<iss::arch::tgc_c>::reg_names;
 constexpr std::array<const char*, 36>    iss::arch::traits<iss::arch::tgc_c>::reg_aliases;
-constexpr std::array<const uint32_t, 41> iss::arch::traits<iss::arch::tgc_c>::reg_bit_widths;
-constexpr std::array<const uint32_t, 41> iss::arch::traits<iss::arch::tgc_c>::reg_byte_offsets;
+constexpr std::array<const uint32_t, 36> iss::arch::traits<iss::arch::tgc_c>::reg_bit_widths;
+constexpr std::array<const uint32_t, 36> iss::arch::traits<iss::arch::tgc_c>::reg_byte_offsets;
 
-tgc_c::tgc_c() {
-    reg.icount = 0;
-}
+tgc_c::tgc_c()  = default;
 
 tgc_c::~tgc_c() = default;
 
@@ -57,8 +55,8 @@ void tgc_c::reset(uint64_t address) {
     reg.PC=address;
     reg.NEXT_PC=reg.PC;
     reg.PRIV=0x3;
-    reg.trap_state=0;
-    reg.icount=0;
+    trap_state=0;
+    icount=0;
 }
 
 uint8_t *tgc_c::get_regs_base_ptr() {

src/main.cpp

@@ -60,13 +60,13 @@ int main(int argc, char *argv[]) {
     desc.add_options()
         ("help,h", "Print help message")
         ("verbose,v", po::value<int>()->implicit_value(0), "Sets logging verbosity")
-        ("logfile,f", po::value<std::string>(), "Sets default log file.")
+        ("logfile,l", po::value<std::string>(), "Sets default log file.")
         ("disass,d", po::value<std::string>()->implicit_value(""), "Enables disassembly")
         ("gdb-port,g", po::value<unsigned>()->default_value(0), "enable gdb server and specify port to use")
         ("instructions,i", po::value<uint64_t>()->default_value(std::numeric_limits<uint64_t>::max()), "max. number of instructions to simulate")
         ("reset,r", po::value<std::string>(), "reset address")
         ("dump-ir", "dump the intermediate representation")
-        ("elf", po::value<std::vector<std::string>>(), "ELF file(s) to load")
+        ("elf,f", po::value<std::vector<std::string>>(), "ELF file(s) to load")
         ("mem,m", po::value<std::string>(), "the memory input file")
         ("plugin,p", po::value<std::vector<std::string>>(), "plugin to activate")
         ("backend", po::value<std::string>()->default_value("interp"), "the memory input file")

src/plugin/cycle_estimate.cpp

@@ -48,7 +48,7 @@ using namespace rapidjson;
 using namespace std;
 
 iss::plugin::cycle_estimate::cycle_estimate(string const& config_file_name)
-: arch_instr(nullptr)
+: instr_if(nullptr)
 , config_file_name(config_file_name)
 {
 }
@@ -57,9 +57,9 @@ iss::plugin::cycle_estimate::~cycle_estimate() {
 }
 
 bool iss::plugin::cycle_estimate::registration(const char* const version, vm_if& vm) {
-    arch_instr = vm.get_arch()->get_instrumentation_if();
-    if(!arch_instr) return false;
-    const string  core_name = arch_instr->core_type_name();
+    instr_if = vm.get_arch()->get_instrumentation_if();
+    if(!instr_if) return false;
+    const string  core_name = instr_if->core_type_name();
     if (config_file_name.length() > 0) {
         ifstream is(config_file_name);
         if (is.is_open()) {
@@ -108,11 +108,11 @@ bool iss::plugin::cycle_estimate::registration(const char* const version, vm_if&
 }
 
 void iss::plugin::cycle_estimate::callback(instr_info_t instr_info, exec_info const& exc_info) {
-    assert(arch_instr && "No instrumentation interface available but callback executed");
+    assert(instr_if && "No instrumentation interface available but callback executed");
     auto entry = delays[instr_info.instr_id];
     bool taken = exc_info.branch_taken;
     if (exc_info.branch_taken && (entry.taken > 1))
-        arch_instr->set_curr_instr_cycles(entry.taken);
+        instr_if->set_curr_instr_cycles(entry.taken);
     else if (entry.not_taken > 1)
-        arch_instr->set_curr_instr_cycles(entry.not_taken);
+        instr_if->set_curr_instr_cycles(entry.not_taken);
 }

src/plugin/pctrace.cpp

@@ -1,6 +1,7 @@
 #include <iss/arch_if.h>
 #include <iss/plugin/pctrace.h>
 #include <util/logging.h>
+#include <util/ities.h>
 #include <rapidjson/document.h>
 #include <rapidjson/istreamwrapper.h>
 #include "rapidjson/writer.h"
@@ -8,27 +9,99 @@
 #include <rapidjson/ostreamwrapper.h>
 #include <rapidjson/error/en.h>
 #include <fstream>
-
 #include <iostream>
+#ifdef WITH_LZ4
+#include <lz4frame.h>
+#endif
 
+namespace iss {
+namespace plugin {
 
 using namespace rapidjson;
 using namespace std;
 
-iss::plugin::cov::cov(std::string const &filename)
+#ifdef WITH_LZ4
+class lz4compress_steambuf: public std::streambuf {
+public:
+    lz4compress_steambuf(const lz4compress_steambuf&) = delete;
+    lz4compress_steambuf& operator=(const lz4compress_steambuf&) = delete;
+    lz4compress_steambuf(std::ostream &sink, size_t buf_size)
+    : sink(sink)
+    , src_buf(buf_size)
+    , dest_buf(LZ4F_compressBound(buf_size, nullptr))
+    {
+        auto errCode = LZ4F_createCompressionContext(&ctx, LZ4F_VERSION);
+        if (LZ4F_isError(errCode) != 0)
+            throw std::runtime_error(std::string("Failed to create LZ4 context: ") + LZ4F_getErrorName(errCode));
+        size_t ret = LZ4F_compressBegin(ctx, &dest_buf.front(), dest_buf.capacity(), nullptr);
+        if (LZ4F_isError(ret) != 0)
+            throw std::runtime_error(std::string("Failed to start LZ4 compression: ") + LZ4F_getErrorName(ret));
+        setp(src_buf.data(), src_buf.data() + src_buf.size() - 1);
+        sink.write(dest_buf.data(), ret);
+    }
+
+    ~lz4compress_steambuf() {
+        close();
+    }
+
+    void close() {
+        if (closed)
+            return;
+        sync();
+        auto ret = LZ4F_compressEnd(ctx, dest_buf.data(), dest_buf.capacity(), nullptr);
+        if (LZ4F_isError(ret) != 0)
+            throw std::runtime_error(std::string("Failed to finish LZ4 compression: ") + LZ4F_getErrorName(ret));
+        sink.write(dest_buf.data(), ret);
+        LZ4F_freeCompressionContext(ctx);
+        closed = true;
+    }
+
+private:
+    int_type overflow(int_type ch) override {
+        compress_and_write();
+        *pptr() = static_cast<char_type>(ch);
+        pbump(1);
+        return ch;
+    }
+
+    int_type sync() override {
+        compress_and_write();
+        return 0;
+    }
+
+    void compress_and_write() {
+        if (closed)
+            throw std::runtime_error("Cannot write to closed stream");
+        if(auto orig_size = pptr() - pbase()){
+            auto ret = LZ4F_compressUpdate(ctx, dest_buf.data(), dest_buf.capacity(), pbase(), orig_size, nullptr);
+            if (LZ4F_isError(ret) != 0)
+                throw std::runtime_error(std::string("LZ4 compression failed: ") + LZ4F_getErrorName(ret));
+            if(ret) sink.write(dest_buf.data(), ret);
+            pbump(-orig_size);
+        }
+    }
+
+    std::ostream &sink;
+    std::vector<char> src_buf;
+    std::vector<char> dest_buf;
+    LZ4F_compressionContext_t ctx{ nullptr };
+    bool closed{ false };
+};
+#endif
+
+cov::cov(std::string const &filename)
 : instr_if(nullptr)
 , filename(filename)
-{
-    output.open("output.trc");
-    jumped = false;
-    first = true;
-}
+, output("output.trc")
+#ifdef WITH_LZ4
+, strbuf(new lz4compress_steambuf(output, 4096))
+, ostr(strbuf.get())
+#endif
+{ }
 
-iss::plugin::cov::~cov() {
-    output.close();
-}
+cov::~cov() { }
 
-bool iss::plugin::cov::registration(const char *const version, vm_if& vm) {
+bool cov::registration(const char *const version, vm_if& vm) {
     instr_if = vm.get_arch()->get_instrumentation_if();
     if(!instr_if) return false;
     const string  core_name = instr_if->core_type_name();
@@ -77,57 +150,31 @@ bool iss::plugin::cov::registration(const char *const version, vm_if& vm) {
         }
     }
     return true;
-
-
-
 }
 
-inline string formatPC(uint64_t pc) {
-    stringstream stream;
-    stream << "0x" << std::hex << pc;
-    return stream.str();
-}
-
-void iss::plugin::cov::callback(instr_info_t iinfo, const exec_info& einfo) {
-//    auto delay = 0;
-//    auto entry = delays[iinfo.instr_id];
-//    bool taken = einfo.branch_taken;
-//    if (einfo.branch_taken)
-//        delay = entry.taken;
-//    else
-//        delay = entry.not_taken;
-//
-//    if (first){
-//        output << formatPC(instr_if->get_pc()) << "," << delay;
-//        first = false;
-//    }
-//    if(instr_if->get_next_pc()-instr_if->get_pc() != delays[iinfo.instr_id].size/8){
-//        //The goal is to keep the output in start-target pairs, so after a jump the target address needs to get written
-//        //to the output. If the target happens to also be a start, we keep the pairing by adding a 0-delay entry.
-//        if (jumped)
-//            output <<"\n" <<formatPC(instr_if->get_pc()) << "," << 0;
-//        output <<"\n" << formatPC(instr_if->get_pc()) << "," << delay;
-//        jumped = true;
-//    }
-//    else{
-//        if (jumped){
-//            output <<"\n" << formatPC(instr_if->get_pc()) << "," << delay;
-//            jumped = false;
-//        }
-//        else if(delay!=1){
-//            output <<"\n" << formatPC(instr_if->get_pc()) << "," << delay;
-//            output <<"\n" << formatPC(instr_if->get_pc()) << "," << 0;
-//        }
-//
-//    }
-
-//source code for the full output
+void cov::callback(instr_info_t iinfo, const exec_info& einfo) {
     auto delay = 0;
-    auto entry = delays[iinfo.instr_id];
+    size_t id = iinfo.instr_id;
+    auto entry = delays[id];
+    auto instr = instr_if->get_instr_word();
+    auto call = id==65 || id ==86 || ((id==2 || id==3) && bit_sub<7,5>(instr)!=0) ;//not taking care of tail calls (jalr with loading x6)
     bool taken = einfo.branch_taken;
-    if (einfo.branch_taken)
+    bool compressed = (instr&0x3)!=0x3;
+    if (einfo.branch_taken) {
         delay = entry.taken;
-    else
+        if(entry.taken > 1)
+            instr_if->set_curr_instr_cycles(entry.taken);
+    } else {
         delay = entry.not_taken;
-    output<<std::hex <<"0x" << instr_if->get_pc() <<"," << delay << "\n";
+        if (entry.not_taken > 1)
+            instr_if->set_curr_instr_cycles(entry.not_taken);
+    }
+#ifndef WITH_LZ4
+    output<<std::hex <<"0x" << instr_if->get_pc() <<"," << delay <<"," << call<<","<<(compressed?2:4) <<"\n";
+#else
+    auto rdbuf=ostr.rdbuf();
+    ostr<<std::hex <<"0x" << instr_if->get_pc() <<"," << delay <<"," << call<<","<<(compressed?2:4) <<"\n";
+#endif
+}
+}
 }

src/sysc/core_complex.cpp

@@ -105,7 +105,7 @@ public:
     heart_state_t &get_state() { return this->state; }
 
     void notify_phase(iss::arch_if::exec_phase p) override {
-        if (p == iss::arch_if::ISTART) owner->sync(this->reg.icount);
+        if (p == iss::arch_if::ISTART) owner->sync(this->icount);
     }
 
     sync_type needed_sync() const override { return PRE_SYNC; }
@@ -115,7 +115,7 @@ public:
             std::stringstream s;
             s << "[p:" << lvl[this->reg.PRIV] << ";s:0x" << std::hex << std::setfill('0')
               << std::setw(sizeof(reg_t) * 2) << (reg_t)this->state.mstatus << std::dec << ";c:"
-              << this->reg.icount + this->cycle_offset << "]";
+              << this->icount + this->cycle_offset << "]";
             SCCDEBUG(owner->name())<<"disass: "
                 << "0x" << std::setw(16) << std::right << std::setfill('0') << std::hex << pc << "\t\t" << std::setw(40)
                 << std::setfill(' ') << std::left << instr << s.str();
@@ -175,9 +175,9 @@ public:
 
     void wait_until(uint64_t flags) override {
         SCCDEBUG(owner->name()) << "Sleeping until interrupt";
-        do {
+        while(this->pending_trap == 0 && (this->csr[arch::mip] & this->csr[arch::mie]) == 0) {
             sc_core::wait(wfi_evt);
-        } while (this->reg.pending_trap == 0);
+        }
         PLAT::wait_until(flags);
     }
 
@@ -204,7 +204,7 @@ public:
             this->csr[arch::mip] &= ~mask;
         this->check_interrupt();
         if(value)
-            SCCTRACE(owner->name()) << "Triggering interrupt " << id << " Pending trap: " << this->reg.pending_trap;
+            SCCTRACE(owner->name()) << "Triggering interrupt " << id << " Pending trap: " << this->pending_trap;
     }
 
 private: