added all of the code to append obj files to pe files...

MessageBoxA - hello world doesnt work because top 32bits of registers
are not being extended...

include/devirt_utils.hpp

@@ -1,10 +1,18 @@
 #include <vmprofiler.hpp>
+
 #include <coff/image.hpp>
+#include <iostream>
+#include <sstream>
+#include <string>
 
 #define VM_ENTER_NAME "vmenter_"
 #define VM_EXIT_NAME "vmexit_"
 #define VM_RTN_NAME "rtn_"
 
+#define FIX_MAKE_ZERO_OFFSET 0x25
+#define FIX_MAKE_RELOC_OFFSET 0x30
+#define FIX_MAKE_JMP_OFFSET 0x43
+
 namespace devirt
 {
     namespace util

src/devirt_t.cpp

@@ -14,15 +14,17 @@ namespace vm
         // sub rsp, num_bytes
         auto current_rtn = vmp_rtns.back();
         auto rsp_addr = ir_builder->CreateLoad( current_rtn->stack );
+        rsp_addr->setAlignment( llvm::Align( 1 ) );
+
         auto sub_rsp_val =
             ir_builder->CreateGEP( ir_builder->getInt8Ty(), rsp_addr, ir_builder->getInt8( 0 - num_bytes ) );
 
-        ir_builder->CreateStore( sub_rsp_val, current_rtn->stack );
+        ir_builder->CreateStore( sub_rsp_val, current_rtn->stack )->setAlignment( llvm::Align( 1 ) );
 
         // mov [rsp], val
         auto resized_new_rsp_addr = ir_builder->CreateBitCast(
             sub_rsp_val, llvm::PointerType::get( llvm::IntegerType::get( *llvm_ctx, num_bytes * 8 ), 0ull ) );
-        ir_builder->CreateStore( val, resized_new_rsp_addr );
+        ir_builder->CreateStore( val, resized_new_rsp_addr )->setAlignment( llvm::Align( 1 ) );
     }
 
     llvm::Value *devirt_t::pop( std::uint8_t num_bytes )
@@ -30,14 +32,20 @@ namespace vm
         // mov rax, [rsp]
         auto current_rtn = vmp_rtns.back();
         auto rsp_addr = ir_builder->CreateLoad( current_rtn->stack );
+        rsp_addr->setAlignment( llvm::Align( 1 ) );
+
         auto new_rsp_addr =
             ir_builder->CreateGEP( ir_builder->getInt8Ty(), rsp_addr, ir_builder->getInt8( num_bytes ) );
 
         auto resized_new_rsp_addr = ir_builder->CreateBitCast(
             rsp_addr, llvm::PointerType::get( llvm::IntegerType::get( *llvm_ctx, num_bytes * 8 ), 0ull ) );
+
         auto pop_val = ir_builder->CreateLoad( resized_new_rsp_addr );
-        ir_builder->CreateStore( new_rsp_addr, current_rtn->stack );
+        pop_val->setAlignment( llvm::Align( 1 ) );
-        ir_builder->CreateStore( llvm::UndefValue::get( ir_builder->getInt8Ty() ), rsp_addr );
+
+        ir_builder->CreateStore( new_rsp_addr, current_rtn->stack )->setAlignment( llvm::Align( 1 ) );
+        ir_builder->CreateStore( llvm::UndefValue::get( ir_builder->getInt8Ty() ), rsp_addr )
+            ->setAlignment( llvm::Align( 1 ) );
         return pop_val;
     }
 
@@ -48,9 +56,14 @@ namespace vm
             auto cast_ptr = ir_builder->CreatePointerCast(
                 var, llvm::PointerType::get( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), 0 ) );
 
-            return ir_builder->CreateLoad( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), cast_ptr );
+            auto result = ir_builder->CreateLoad( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), cast_ptr );
+            result->setAlignment( llvm::Align( 1 ) );
+            return result;
         }
-        return ir_builder->CreateLoad( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), var );
+
+        auto result = ir_builder->CreateLoad( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), var );
+        result->setAlignment( llvm::Align( 1 ) );
+        return result;
     }
 
     llvm::Value *devirt_t::load_value( std::uint8_t byte_size, llvm::AllocaInst *var )
@@ -60,9 +73,14 @@ namespace vm
             auto cast_ptr = ir_builder->CreatePointerCast(
                 var, llvm::PointerType::get( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), 0 ) );
 
-            return ir_builder->CreateLoad( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), cast_ptr );
+            auto result = ir_builder->CreateLoad( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), cast_ptr );
+            result->setAlignment( llvm::Align( 1 ) );
+            return result;
         }
-        return ir_builder->CreateLoad( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), var );
+
+        auto result = ir_builder->CreateLoad( llvm::IntegerType::get( *llvm_ctx, byte_size * 8 ), var );
+        result->setAlignment( llvm::Align( 1 ) );
+        return result;
     }
 
     bool devirt_t::compile( std::vector< std::uint8_t > &obj )
@@ -170,7 +188,7 @@ namespace vm
         auto or3 = ir_builder->CreateOr( or2, shifted_sf );
         auto or4 = ir_builder->CreateOr( or3, shifted_af );
         auto or5 = ir_builder->CreateOr( or4, shifted_pf );
-        return ir_builder->CreateXor( or5, llvm::ConstantInt::get( llvm::IntegerType::get( *llvm_ctx, 64 ), 514 ) );
+        return ir_builder->CreateOr( or5, llvm::ConstantInt::get( llvm::IntegerType::get( *llvm_ctx, 64 ), 514 ) );
     }
 
 } // namespace vm

src/devirt_utils.cpp

@@ -57,12 +57,169 @@ namespace devirt
         if ( obj.empty() || bin.empty() )
             return false;
 
-        // TODO: append a new section to the pe file...
+        std::vector< std::uint8_t > map_buff;
-        // put the entire obj file into this new section...
+        auto bin_img = reinterpret_cast< win::image_t<> * >( bin.data() );
-        // fix up the vmenter functions to jmp to their respective devirtualized routines...
+
-        //
+        // increase the size_image by the size of the obj as the entire obj will be in the .devirt section...
-        // move relocation directory to the end of the module and append new relocations...
+        map_buff.resize(
-        // export all of the new devirtualized/llvm generated functions...
+            reinterpret_cast< win::image_t<> * >( bin.data() )->get_nt_headers()->optional_header.size_image +=
+            obj.size() );
+
+        // copy over dos headers, pe headers (section headers, optional header etc)...
+        std::memcpy( map_buff.data(), bin.data(), bin_img->get_nt_headers()->optional_header.size_headers );
+
+        // map sections into map_img... also make image offset == virtual offset...
+        auto map_img = reinterpret_cast< win::image_t<> * >( map_buff.data() );
+        std::for_each( bin_img->get_nt_headers()->get_sections(),
+                       bin_img->get_nt_headers()->get_sections() + bin_img->get_file_header()->num_sections,
+                       [ & ]( coff::section_header_t &section ) {
+                           if ( !section.virtual_address || !section.ptr_raw_data )
+                               return;
+
+                           std::memcpy( map_buff.data() + section.virtual_address, bin.data() + section.ptr_raw_data,
+                                        section.size_raw_data );
+                       } );
+
+        std::for_each( map_img->get_nt_headers()->get_sections(),
+                       map_img->get_nt_headers()->get_sections() + bin_img->get_file_header()->num_sections,
+                       [ & ]( coff::section_header_t &section ) {
+                           section.ptr_raw_data = section.virtual_address;
+                           section.size_raw_data = section.virtual_size;
+                       } );
+
+        auto pe_sections = map_img->get_nt_headers()->get_sections();
+        auto num_sections = map_img->get_file_header()->num_sections;
+
+        // append .devirt section to the original binary...
+        strcpy( pe_sections[ num_sections ].name.short_name, ".devirt" );
+        pe_sections[ num_sections ].virtual_size = obj.size();
+        pe_sections[ num_sections ].size_raw_data = obj.size();
+        pe_sections[ num_sections ].virtual_address = ( pe_sections[ num_sections - 1 ].virtual_address +
+                                                        pe_sections[ num_sections - 1 ].virtual_size + 0x1000 ) &
+                                                      ~0xFFFull;
+        pe_sections[ num_sections ].ptr_raw_data = pe_sections[ num_sections ].virtual_address;
+
+        coff::section_characteristics_t prots;
+        prots.flags = 0ull;
+        prots.mem_execute = true;
+        prots.mem_read = true;
+        pe_sections[ num_sections ].characteristics = prots;
+        ++map_img->get_file_header()->num_sections;
+        std::memcpy( map_buff.data() + pe_sections[ num_sections ].virtual_address, obj.data(), obj.size() );
+
+        auto obj_img_addr = map_buff.data() + pe_sections[ num_sections ].virtual_address;
+        auto obj_img = reinterpret_cast< coff::image_t * >( obj_img_addr );
+        auto str_table = obj_img->get_strings();
+        auto symbols = obj_img->get_symbols();
+        auto obj_sections = obj_img->get_sections();
+        auto symbol_cnt = obj_img->file_header.num_symbols;
+
+        static const auto get_symbol = [ & ]( std::string symbol_name ) -> auto
+        {
+            return std::find_if( symbols, symbols + symbol_cnt, [ & ]( coff::symbol_t &symbol ) {
+                if ( symbol.derived_type != coff::derived_type_id::function )
+                    return false;
+
+                return !strcmp( symbol.name.to_string( str_table ).data(), symbol_name.c_str() );
+            } );
+        };
+
+        std::vector< std::pair< std::uint32_t, std::uint16_t > > new_relocs;
+        std::for_each( symbols, symbols + symbol_cnt, [ & ]( coff::symbol_t &symbol ) {
+            if ( symbol.derived_type != coff::derived_type_id::function )
+                return;
+
+            auto symbol_name = symbol.name.to_string( str_table );
+            if ( strstr( symbol_name.data(), VM_ENTER_NAME ) )
+            {
+                // fix 0x1000000000000000 to 0x0000000000000000
+                // also push back a new relocation for this entry...
+                auto symbol_addr = reinterpret_cast< std::uintptr_t >( obj_img_addr ) + symbol.value +
+                                   obj_sections[ symbol.section_index - 1 ].ptr_raw_data;
+
+                *reinterpret_cast< std::uintptr_t * >( symbol_addr + FIX_MAKE_ZERO_OFFSET ) = 0ull;
+                *reinterpret_cast< std::uintptr_t * >( symbol_addr + FIX_MAKE_RELOC_OFFSET ) = 0ull;
+
+                auto page = ( symbol.value + obj_sections[ symbol.section_index - 1 ].ptr_raw_data +
+                              FIX_MAKE_RELOC_OFFSET + pe_sections[ num_sections ].virtual_address ) &
+                            ~0xFFFull;
+
+                auto offset = ( symbol.value + obj_sections[ symbol.section_index - 1 ].ptr_raw_data +
+                                FIX_MAKE_RELOC_OFFSET + pe_sections[ num_sections ].virtual_address ) &
+                              0xFFFull;
+
+                new_relocs.push_back( { page, offset } );
+
+                // look up the rtn_xxxxxx function symbol for this vm enter and make it jmp to it...
+                auto rtn_offset = std::stoull( symbol_name.data() + sizeof( VM_ENTER_NAME ) - 1, nullptr, 16 );
+
+                std::stringstream rtn_name;
+                rtn_name << "rtn_" << std::hex << rtn_offset;
+                auto rtn_sym = get_symbol( rtn_name.str() );
+                auto relocs = reinterpret_cast< coff::reloc_t * >(
+                    obj_sections[ rtn_sym->section_index - 1 ].ptr_relocs + obj_img_addr );
+
+                auto rtn_rva = rtn_offset - map_img->get_nt_headers()->optional_header.image_base;
+                std::int32_t devirt_rtn_rva = symbol_addr - reinterpret_cast< std::uintptr_t >( map_buff.data() );
+
+                *reinterpret_cast< std::int32_t * >( rtn_rva + map_buff.data() + 1 ) = devirt_rtn_rva - ( rtn_rva + 5 );
+                *reinterpret_cast< std::uint8_t * >( rtn_rva + map_buff.data() ) = 0xE9;
+
+                // apply relocations to the rtn_xxxxx...
+                for ( auto reloc_idx = 0u; reloc_idx < obj_sections[ rtn_sym->section_index - 1 ].num_relocs;
+                      ++reloc_idx )
+                {
+                    coff::reloc_t &reloc = relocs[ reloc_idx ];
+
+                    auto vmexit_sym =
+                        get_symbol( std::string( symbols[ reloc.symbol_index ].name.to_string( str_table ) ) );
+
+                    auto vmexit_offset = vmexit_sym->value + obj_sections[ vmexit_sym->section_index - 1 ].ptr_raw_data;
+
+                    auto reloc_file_offset =
+                        reloc.virtual_address + obj_sections[ rtn_sym->section_index - 1 ].ptr_raw_data;
+
+                    std::int32_t rva = vmexit_offset - ( std::int32_t )( reloc_file_offset + 4 );
+                    *reinterpret_cast< std::int32_t * >( obj_img_addr + reloc_file_offset ) = rva;
+                }
+
+                auto rtn_addr = reinterpret_cast< std::uintptr_t >(
+                    obj_img_addr + obj_sections[ rtn_sym->section_index - 1 ].ptr_raw_data + rtn_sym->value );
+
+                // create jmp to the rtn_xxxxx....
+                std::int32_t rva = rtn_addr - ( std::int64_t )( symbol_addr + FIX_MAKE_JMP_OFFSET + 5 );
+                *reinterpret_cast< std::uint8_t * >( symbol_addr + FIX_MAKE_JMP_OFFSET ) = 0xE9;
+                *reinterpret_cast< std::int32_t * >( symbol_addr + FIX_MAKE_JMP_OFFSET + 1 ) = rva;
+            }
+        } );
+
+        auto resize_cnt = new_relocs.size() * ( sizeof( win::reloc_entry_t ) + sizeof( win::reloc_block_t ) );
+        map_buff.resize( map_img->get_nt_headers()->optional_header.size_image += resize_cnt );
+
+        map_img = reinterpret_cast< win::image_t<> * >( map_buff.data() );
+        auto basereloc_dir = map_img->get_directory( win::directory_id::directory_entry_basereloc );
+        auto reloc_dir = reinterpret_cast< win::reloc_directory_t * >( basereloc_dir->rva + map_buff.data() );
+
+        basereloc_dir->size += resize_cnt;
+        for ( const auto &[ reloc_rva, reloc_offset ] : new_relocs )
+        {
+            win::reloc_block_t *reloc_block = &reloc_dir->first_block;
+            while ( reloc_block->base_rva && reloc_block->size_block )
+                reloc_block = reloc_block->next();
+
+            reloc_block->base_rva = reloc_rva;
+            reloc_block->size_block = sizeof( win::reloc_entry_t ) + sizeof uint64_t;
+
+            reloc_block->next()->base_rva = 0ull;
+            reloc_block->next()->size_block = 0ull;
+
+            reloc_block->entries[ 0 ].type = win::reloc_type_id::rel_based_dir64;
+            reloc_block->entries[ 0 ].offset = reloc_offset;
+        }
+
+        // replace bin vector with map_buff vector...
+        bin.clear();
+        bin.insert( bin.begin(), map_buff.begin(), map_buff.end() );
         return true;
     }
 } // namespace devirt

src/lifters/lconst.cpp

@@ -3,50 +3,50 @@
 namespace vm
 {
     lifters_t::lifter_callback_t lifters_t::lconstq =
-        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block,
+        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block, const vm::instrs::virt_instr_t &vinstr,
-               const vm::instrs::virt_instr_t &vinstr, llvm::IRBuilder<> *ir_builder ) {
+               llvm::IRBuilder<> *ir_builder ) {
             rtn->push( 8, llvm::ConstantInt::get( ir_builder->getInt64Ty(), vinstr.operand.imm.u ) );
         };
 
     lifters_t::lifter_callback_t lifters_t::lconstdwsxq =
-        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block,
+        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block, const vm::instrs::virt_instr_t &vinstr,
-               const vm::instrs::virt_instr_t &vinstr, llvm::IRBuilder<> *ir_builder ) {
+               llvm::IRBuilder<> *ir_builder ) {
             rtn->push( 8, llvm::ConstantInt::get( ir_builder->getInt64Ty(), vinstr.operand.imm.u ) );
         };
 
     lifters_t::lifter_callback_t lifters_t::lconstwsxq =
-        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block,
+        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block, const vm::instrs::virt_instr_t &vinstr,
-               const vm::instrs::virt_instr_t &vinstr, llvm::IRBuilder<> *ir_builder ) {
+               llvm::IRBuilder<> *ir_builder ) {
             rtn->push( 8, llvm::ConstantInt::get( ir_builder->getInt64Ty(), vinstr.operand.imm.u ) );
         };
 
     lifters_t::lifter_callback_t lifters_t::lconstbsxq =
-        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block,
+        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block, const vm::instrs::virt_instr_t &vinstr,
-               const vm::instrs::virt_instr_t &vinstr, llvm::IRBuilder<> *ir_builder ) {
+               llvm::IRBuilder<> *ir_builder ) {
             rtn->push( 8, llvm::ConstantInt::get( ir_builder->getInt64Ty(), vinstr.operand.imm.u ) );
         };
 
     lifters_t::lifter_callback_t lifters_t::lconstdw =
-        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block,
+        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block, const vm::instrs::virt_instr_t &vinstr,
-               const vm::instrs::virt_instr_t &vinstr, llvm::IRBuilder<> *ir_builder ) {
+               llvm::IRBuilder<> *ir_builder ) {
             rtn->push( 4, llvm::ConstantInt::get( ir_builder->getInt32Ty(), vinstr.operand.imm.u ) );
         };
 
     lifters_t::lifter_callback_t lifters_t::lconstwsxdw =
-        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block,
+        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block, const vm::instrs::virt_instr_t &vinstr,
-               const vm::instrs::virt_instr_t &vinstr, llvm::IRBuilder<> *ir_builder ) {
+               llvm::IRBuilder<> *ir_builder ) {
             rtn->push( 4, llvm::ConstantInt::get( ir_builder->getInt32Ty(), vinstr.operand.imm.u ) );
         };
 
     lifters_t::lifter_callback_t lifters_t::lconstbsxdw =
-        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block,
+        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block, const vm::instrs::virt_instr_t &vinstr,
-               const vm::instrs::virt_instr_t &vinstr, llvm::IRBuilder<> *ir_builder ) {
+               llvm::IRBuilder<> *ir_builder ) {
             rtn->push( 4, llvm::ConstantInt::get( ir_builder->getInt32Ty(), vinstr.operand.imm.u ) );
         };
 
     lifters_t::lifter_callback_t lifters_t::lconstbzxw =
-        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block,
+        [ & ]( vm::devirt_t *rtn, const vm::instrs::code_block_t &vm_code_block, const vm::instrs::virt_instr_t &vinstr,
-               const vm::instrs::virt_instr_t &vinstr, llvm::IRBuilder<> *ir_builder ) {
+               llvm::IRBuilder<> *ir_builder ) {
             rtn->push( 2, llvm::ConstantInt::get( ir_builder->getInt16Ty(), vinstr.operand.imm.u ) );
         };
 } // namespace vm

src/lifters/sreg.cpp

@@ -8,7 +8,7 @@ namespace vm
             auto t1 = rtn->pop( 8 );
             auto &vmp_rtn = rtn->vmp_rtns.back();
             auto vreg = vmp_rtn->virtual_registers[ vinstr.operand.imm.u ? vinstr.operand.imm.u / 8 : 0 ];
-            ir_builder->CreateStore( t1, vreg );
+            ir_builder->CreateStore( t1, vreg )->setAlignment( llvm::Align( 8 ) );
         };
 
     lifters_t::lifter_callback_t lifters_t::sregdw =
@@ -19,6 +19,6 @@ namespace vm
             auto vreg = vmp_rtn->virtual_registers[ vinstr.operand.imm.u ? vinstr.operand.imm.u / 8 : 0 ];
             auto vregdw = ir_builder->CreatePointerCast(
                 vreg, llvm::PointerType::get( llvm::IntegerType::get( *rtn->llvm_ctx, 32 ), 0 ) );
-            ir_builder->CreateStore( t1, vregdw );
+            ir_builder->CreateStore( t1, vregdw )->setAlignment( llvm::Align( 4 ) );
         };
 } // namespace vm

src/main.cpp

@@ -8,6 +8,8 @@ int main( int argc, const char *argv[] )
 {
     argparse::argument_parser_t parser( "vmdevirt", "virtual instruction pseudo code generator" );
     parser.add_argument().name( "--vmp2file" ).required( true ).description( "path to .vmp2 file..." );
+    parser.add_argument().name( "--out" ).required( true ).description( "output devirtualized binary name..." );
+    parser.add_argument().name( "--genobj" ).description( "write the generated obj file to disk..." );
     parser.add_argument().name( "--bin" ).required( true ).description(
         "path to the image in which to apply devirtualized code too...\n" );
 
@@ -69,6 +71,8 @@ int main( int argc, const char *argv[] )
                      vmp2_code_blocks.size() );
     }
 
+    llvm_module.print( llvm::outs(), nullptr );
+
     llvm::LLVMInitializeX86TargetInfo();
     llvm::LLVMInitializeX86Target();
     llvm::LLVMInitializeX86TargetMC();
@@ -77,9 +81,12 @@ int main( int argc, const char *argv[] )
 
     vmp_devirt.compile( compiled_obj );
     std::printf( "> compiled all routines... compiled obj size = %d\n", compiled_obj.size() );
-    std::ofstream( "devirt.o", std::ios::binary )
+
-        .write( reinterpret_cast< const char * >( compiled_obj.data() ), compiled_obj.size() );
+    if ( parser.exists( "genobj" ) )
+        std::ofstream( parser.get< std::string >( "out" ).append( ".o" ), std::ios::binary )
+            .write( reinterpret_cast< const char * >( compiled_obj.data() ), compiled_obj.size() );
 
     devirt::append( compiled_obj, bin );
-    std::ofstream( "devirt.exe", std::ios::binary ).write( reinterpret_cast< const char * >( bin.data() ), bin.size() );
+    std::ofstream( parser.get< std::string >( "out" ), std::ios::binary )
+        .write( reinterpret_cast< const char * >( bin.data() ), bin.size() );
 }

src/vmp_rtn_t.cpp

@@ -79,7 +79,7 @@ namespace vm
         }
 
         rtn_name.str( "" );
-        asm_str.append( "mov rcx, rsp; sub rsp, 0xA00; " );
+        asm_str.append( "mov rcx, rsp; sub rsp, 0xA00; int 3; int 3; int 3; int 3; int 3;" );
 
         rtn_name << "vmenter_" << std::hex << rtn_begin;
         auto entry_func = llvm::Function::Create( llvm::FunctionType::get( ir_builder->getVoidTy(), false ),