diff --git a/CMakeLists.txt b/CMakeLists.txt
index aec662a..ac3bf7f 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -7,6 +7,7 @@ project(dbt-rise-tgc VERSION 1.0.0)
 include(GNUInstallDirs)
 
 find_package(elfio)
+find_package(Boost COMPONENTS coroutine)
 
 if(WITH_LLVM)
     if(DEFINED ENV{LLVM_HOME})
@@ -66,7 +67,7 @@ elseif("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
     target_compile_options(${PROJECT_NAME} PRIVATE /wd4293)
 endif()
 target_include_directories(${PROJECT_NAME} PUBLIC incl)
-target_link_libraries(${PROJECT_NAME} PUBLIC softfloat scc-util jsoncpp)
+target_link_libraries(${PROJECT_NAME} PUBLIC softfloat scc-util jsoncpp Boost::coroutine)
 if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
     target_link_libraries(${PROJECT_NAME} PUBLIC -Wl,--whole-archive dbt-core -Wl,--no-whole-archive)
 else()
diff --git a/gen_input/templates/CORENAME.h.gtl b/gen_input/templates/CORENAME.h.gtl
index 2105569..fa69704 100644
--- a/gen_input/templates/CORENAME.h.gtl
+++ b/gen_input/templates/CORENAME.h.gtl
@@ -161,7 +161,7 @@ struct ${coreDef.name.toLowerCase()}: public arch_if {
 
     inline uint32_t get_last_branch() { return reg.last_branch; }
 
-protected:
+
 #pragma pack(push, 1)
     struct ${coreDef.name}_regs {<%
         registers.each { reg -> if(reg.size>0) {%> 
diff --git a/gen_input/templates/interp/CORENAME.cpp.gtl b/gen_input/templates/interp/CORENAME.cpp.gtl
index 7cf5c47..26db3d0 100644
--- a/gen_input/templates/interp/CORENAME.cpp.gtl
+++ b/gen_input/templates/interp/CORENAME.cpp.gtl
@@ -45,6 +45,8 @@ def nativeTypeSize(int size){
 #include <iss/interp/vm_base.h>
 #include <util/logging.h>
 #include <sstream>
+#include <boost/coroutine2/all.hpp>
+#include <functional>
 
 #ifndef FMT_HEADER_ONLY
 #define FMT_HEADER_ONLY
@@ -59,6 +61,7 @@ namespace interp {
 namespace ${coreDef.name.toLowerCase()} {
 using namespace iss::arch;
 using namespace iss::debugger;
+using namespace std::placeholders;
 
 template <typename ARCH> class vm_impl : public iss::interp::vm_base<ARCH> {
 public:
@@ -130,6 +133,10 @@ protected:
         this->core.wait_until(type);
     }
 
+    using yield_t = boost::coroutines2::coroutine<void>::push_type;
+    using coro_t = boost::coroutines2::coroutine<void>::pull_type;
+    std::vector<coro_t> spawn_blocks;
+
     template<typename T>
     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){
@@ -174,7 +181,9 @@ protected:
         auto sign_mask = 1ULL<<(W-1);
         return (from & mask) | ((from & sign_mask) ? ~mask : 0);
     }
-
+<%functions.each{ it.eachLine { %>
+    ${it}<%}%>
+<%}%>
 private:
     /****************************************************************************
      * start opcode definitions
@@ -196,11 +205,6 @@ private:
     /* instruction ${idx}: ${instr.name} */
     compile_ret_t __${generator.functionName(instr.name)}(virt_addr_t& pc, code_word_t instr){
         // pre execution stuff
-        auto* PC = reinterpret_cast<uint${addrDataWidth}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
-        auto NEXT_PC = reinterpret_cast<uint${addrDataWidth}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
-        *PC=*NEXT_PC;
-        auto* trap_state = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::TRAP_STATE]);
-        *trap_state = *reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PENDING_TRAP]);
         if(this->sync_exec && PRE_SYNC) this->do_sync(PRE_SYNC, ${idx});
         <%instr.fields.eachLine{%>${it}
         <%}%>if(this->disass_enabled){
@@ -208,6 +212,9 @@ private:
             <%instr.disass.eachLine{%>${it}
             <%}%>
         }
+        auto* PC = reinterpret_cast<uint${addrDataWidth}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
+        auto* NEXT_PC = reinterpret_cast<uint${addrDataWidth}_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]);
         // 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>::${k}0]);<% }else{ %> 
@@ -219,6 +226,7 @@ private:
         <%instr.behavior.eachLine{%>${it}
         <%}%>} catch(...){}
         // post execution stuff
+        for(auto& spawn_block:spawn_blocks) spawn_block();
         if(this->sync_exec && POST_SYNC) this->do_sync(POST_SYNC, ${idx});
         // trap check
         if(*trap_state!=0){
@@ -335,6 +343,8 @@ typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e co
             auto f = decode_inst(insn);
             auto old_pc = pc.val;
             pc = (this->*f)(pc, insn);
+            this->core.reg.PC = this->core.reg.NEXT_PC;
+            this->core.reg.trap_state = this->core.reg.pending_trap;
         }
     }
     return pc;