1 //: The bedrock level 1 of abstraction is now done, and we're going to start
  2 //: building levels above it that make programming in x86 machine code a
  3 //: little more ergonomic.
  4 //:
  5 //: All levels will be "pass through by default". Whatever they don't
  6 //: understand they will silently pass through to lower levels.
  7 //:
  8 //: Since raw hex bytes of machine code are always possible to inject, SubX is
  9 //: not a language, and we aren't building a compiler. This is something
 10 //: deliberately leakier. Levels are more for improving auditing, checks and
 11 //: error messages rather than for hiding low-level details.
 12 
 13 //: Translator workflow: read 'source' file. Run a series of transforms on it,
 14 //: each passing through what it doesn't understand. The final program should
 15 //: be just machine code, suitable to write to an ELF binary.
 16 //:
 17 //: Higher levels usually transform code on the basis of metadata.
 18 
 19 :(before "End Main")
 20 if (is_equal(argv[1], "translate")) {
 21   START_TRACING_UNTIL_END_OF_SCOPE;
 22   assert(argc > 3);
 23   program p;
 24   ifstream fin(argv[2]);
 25   if (!fin) {
 26     cerr << "could not open " << argv[2] << '\n';
 27     return 1;
 28   }
 29   parse(fin, p);
 30   if (trace_contains_errors()) return 1;
 31   transform(p);
 32   if (trace_contains_errors()) return 1;
 33   save_elf(p, argv[3]);
 34   if (trace_contains_errors()) unlink(argv[3]);
 35   return 0;
 36 }
 37 
 38 :(code)
 39 // write out a program to a bare-bones ELF file
 40 void save_elf(const program& p, const char* filename) {
 41   ofstream out(filename, ios::binary);
 42   write_elf_header(out, p);
 43   for (size_t i = 0;  i < p.segments.size();  ++i)
 44     write_segment(p.segments.at(i), out);
 45   out.close();
 46 }
 47 
 48 void write_elf_header(ostream& out, const program& p) {
 49   char c = '\0';
 50 #define O(X)  c = (X); out.write(&c, sizeof(c))
 51 // host is required to be little-endian
 52 #define emit(X)  out.write(reinterpret_cast<const char*>(&X), sizeof(X))
 53   //// ehdr
 54   // e_ident
 55   O(0x7f); O(/*E*/0x45); O(/*L*/0x4c); O(/*F*/0x46);
 56     O(0x1);  // 32-bit format
 57     O(0x1);  // little-endian
 58     O(0x1); O(0x0);
 59   for (size_t i = 0;  i < 8;  ++i) { O(0x0); }
 60   // e_type
 61   O(0x02); O(0x00);
 62   // e_machine
 63   O(0x03); O(0x00);
 64   // e_version
 65   O(0x01); O(0x00); O(0x00); O(0x00);
 66   // e_entry
 67   int e_entry = p.segments.at(0).start;  // convention
 68   emit(e_entry);
 69   // e_phoff -- immediately after ELF header
 70   int e_phoff = 0x34;
 71   emit(e_phoff);
 72   // e_shoff; unused
 73   int dummy32 = 0;
 74   emit(dummy32);
 75   // e_flags; unused
 76   emit(dummy32);
 77   // e_ehsize
 78   uint16_t e_ehsize = 0x34;
 79   emit(e_ehsize);
 80   // e_phentsize
 81   uint16_t e_phentsize = 0x20;
 82   emit(e_phentsize);
 83   // e_phnum
 84   uint16_t e_phnum = SIZE(p.segments);
 85   emit(e_phnum);
 86   // e_shentsize
 87   uint16_t dummy16 = 0x0;
 88   emit(dummy16);
 89   // e_shnum
 90   emit(dummy16);
 91   // e_shstrndx
 92   emit(dummy16);
 93 
 94   uint32_t p_offset = /*size of ehdr*/0x34 + SIZE(p.segments)*0x20/*size of each phdr*/;
 95   for (int i = 0;  i < SIZE(p.segments);  ++i) {
 96     //// phdr
 97     // p_type
 98     uint32_t p_type = 0x1;
 99     emit(p_type);
100     // p_offset
101     emit(p_offset);
102     // p_vaddr
103     uint32_t p_start = p.segments.at(i).start;
104     emit(p_start);
105     // p_paddr
106     emit(p_start);
107     // p_filesz
108     uint32_t size = num_words(p.segments.at(i));
109     assert(p_offset + size < SEGMENT_SIZE);
110     emit(size);
111     // p_memsz
112     emit(size);
113     // p_flags
114     uint32_t p_flags = (i == 0) ? /*r-x*/0x5 : /*rw-*/0x6;  // convention: only first segment is code
115     emit(p_flags);
116 
117     // p_align
118     // "As the system creates or augments a process image, it logically copies
119     // a file's segment to a virtual memory segment.  When—and if— the system
120     // physically reads the file depends on the program's execution behavior,
121     // system load, and so on.  A process does not require a physical page
122     // unless it references the logical page during execution, and processes
123     // commonly leave many pages unreferenced. Therefore delaying physical
124     // reads frequently obviates them, improving system performance. To obtain
125     // this efficiency in practice, executable and shared object files must
126     // have segment images whose file offsets and virtual addresses are
127     // congruent, modulo the page size." -- http://refspecs.linuxbase.org/elf/elf.pdf (page 95)
128     uint32_t p_align = 0x1000;  // default page size on linux
129     emit(p_align);
130     if (p_offset % p_align != p_start % p_align) {
131       raise << "segment starting at 0x" << HEXWORD << p_start << " is improperly aligned; alignment for p_offset " << p_offset << " should be " << (p_offset % p_align) << " but is " << (p_start % p_align) << '\n' << end();
132       return;
133     }
134 
135     // prepare for next segment
136     p_offset += size;
137   }
138 #undef O
139 #undef emit
140 }
141 
142 void write_segment(const segment& s, ostream& out) {
143   for (int i = 0;  i < SIZE(s.lines);  ++i) {
144     const vector<word>& w = s.lines.at(i).words;
145     for (int j = 0;  j < SIZE(w);  ++j) {
146       uint8_t x = hex_byte(w.at(j).data);  // we're done with metadata by this point
147       out.write(reinterpret_cast<const char*>(&x), /*sizeof(byte)*/1);
148     }
149   }
150 }
151 
152 uint32_t num_words(const segment& s) {
153   uint32_t sum = 0;
154   for (int i = 0;  i < SIZE(s.lines);  ++i)
155     sum += SIZE(s.lines.at(i).words);
156   return sum;
157 }
158 
159 :(before "End Includes")
160 using std::ios;