README.md

bootstrap stage 05

This stage consists of a C compiler capable of compiling tcc (after some modifications to tcc's source code). Run

$ make

to build our C compiler and tcc. This will take some time (approx. 25 seconds on my computer). This also compiles a "Hello, world!" executable, a.out, with our compiler.

We can now compile tcc with itself. But first, you'll need to produce the header files and library files which are needed to compile (almost) any program with tcc:

$ make tcc-files

The files will be installed to ./tcc-bootstrap. If you want to change this, make sure to change both the TCCINST variable in the makefile, and the CONFIG_TCCDIR macro in tcc-0.9.27/config.h. Anyways, once this installation is done, you should be able to compile any C program with tcc-0.9.27/tcc0, including tcc itself:

$ cd tcc-0.9.27
$ ./tcc0 tcc.c -o tcc1

Now, let's try doing the same thing, but starting with gcc instead of our C compiler:

$ gcc tcc.c -o tcc0a
$ ./tcc0a tcc.c -o tcc1a

In theory, these should produce the same files, since the output of tcc shouldn't depend on which compiler it was compiled with. If they are different, then perhaps a bug was introduced in some early version of gcc, and replicated in all C compilers since then! Well, only one way to find out:

$ diff tcc1 tcc1a
Binary files tcc1 and tcc1a differ

!!! Is there some malicious code hiding in the difference between these two files? Well unfortunately (fortunately, really) the truth is more boring than that:

$ ./tcc1 tcc.c -o tcc2
$ diff tcc2 tcc1a
$

Yes, after compiling tcc with itself one more time, we get the same executable as the gcc-tcc one. I'm not sure why tcc1 differs from tcc2, but there you go. Turns out there isn't some malicious self-replicating code hiding in gcc after all.*

the C compiler

The C compiler for this stage is written in the 04 language, using the 04a preprocessor and is spread out across multiple files:

util.b         - various utilities (syscall, puts, memset, etc.)
constants.b    - numerical and string constants used by the rest of the program
idents.b       - functions for creating mappings from identifiers to arbitrary 64-bit values
preprocess.b   - preprocesses C files
tokenize.b     - turns preprocessing tokens into tokens (see explanation below)
parse.b        - turns tokens into a nice representation of the program
codegen.b      - turns parse.b's representation into CPU instructions
main.b         - puts everything together

The whole thing is ~12,000 lines of code, which is ~280KB when compiled.

It can be compiled with make or:

../04a/out04 main.b in04
../04/out03 in04 out04

the C standard

In 1989, the C programming language was standardized by the ANSI.

The C89 standard (in theory) defines which C programs are legal, and exactly what any particular legal C program does. A draft of it, which is about as good as the real thing, is available here.

Since 1989, more features have been added to C, and so more C standards have been published. To keep things simple, our compiler only supports the features from C89, except that declarations can appear after statements and // single-line comments are allowed (both of these were added in C99).

compiler high-level details

Compiling a C program involves several "translation phases" (C89 standard § 2.1.1.2). Here, I'll only be outlining the process our C compiler uses. The technical details of the standard are slightly different.

First, each time a backslash is immediately followed by a newline, both are deleted, e.g.

Hel\
lo,
wo\
rld!

becomes

Hello,
world!

Well, we actually turn this into

Hello,

world!

so that line numbers are preserved for errors (this doesn't change the meaning of any program). This feature exists so that you can spread one line of code across multiple lines, which is useful sometimes.

Then, comments are deleted (technically, replaced with spaces), and the file is split up into preprocesing tokens. A preprocessing token is one of:

• A number (e.g. 5, 10.2, 3.6.6)
• A string literal (e.g. "Hello")
• A symbol (e.g. <, {, .)
• An identifier (e.g. int, x, main)
• A character constant (e.g. 'a', '\n')
• A space character
• A newline character

Note that preprocessing tokens are just strings of characters, and aren't assigned any meaning yet; 3.6.6e-.3 is a valid "preprocessing number" even though it's gibberish.

Next, preprocessor directives are executed. These include things like

#define A_NUMBER 4

which will replace every preprocessing token consisting of the identifier A_NUMBER in the rest of the program with 4. Also in this phase,

#include "X"

is replaced with the (preprocessing tokens in the) file named X.

Then preprocessing tokens are turned into tokens. Tokens are one of:

• A keyword (e.g. int, while)
• A symbol (e.g. <, -, {)
• An identifier (e.g. main, f, x_3)
• An integer literal (e.g. 77, 0x123)
• A character literal (e.g. 'a', '\n')
• A floating-point literal (e.g. 3.6, 5e10)

Next, an internal representation of the program is constructed in memory. This is where we read the tokens if ( a ) printf ( "Hello!\n" ) ; and interpret it as an if statement, whose condition is the variable a, and whose body consists of the single statement calling the printf function with the argument "Hello!\n".

Finally, we turn this internal representation into code for every function.

executable format

This compiler's executables are much more sophisticated than the previous ones'. Instead of storing code and data all in one segment, we have three segments: one 6MB segment for code (the program's functions are only allowed to use up 4MB of that, though), one 4MB segment for read-only data (strings), and one 4MB segment for read-write data.

Well, it should only be read-write, but unfortunately it also has to be executable...

syscalls

Of course, we need some way of making system calls in C. We do this with a macro, __syscall, which you'll find in stdc_common.h:

static unsigned char __syscall_data[] = {
	// mov rax, [rsp+24]
	0x48, 0x8b, 0x84, 0x24, 24, 0, 0, 0,
	// mov rdi, rax
	0x48, 0x89, 0xc7,
	// mov rax, [rsp+32]
	0x48, 0x8b, 0x84, 0x24, 32, 0, 0, 0,
	// mov rsi, rax
	0x48, 0x89, 0xc6,
	// mov rax, [rsp+40]
	0x48, 0x8b, 0x84, 0x24, 40, 0, 0, 0,
	// mov rdx, rax
	0x48, 0x89, 0xc2,
	// mov rax, [rsp+48]
	0x48, 0x8b, 0x84, 0x24, 48, 0, 0, 0,
	// mov r10, rax
	0x49, 0x89, 0xc2,
	// mov rax, [rsp+56]
	0x48, 0x8b, 0x84, 0x24, 56, 0, 0, 0,
	// mov r8, rax
	0x49, 0x89, 0xc0,
	// mov rax, [rsp+64]
	0x48, 0x8b, 0x84, 0x24, 64, 0, 0, 0,
	// mov r9, rax
	0x49, 0x89, 0xc1,
	// mov rax, [rsp+16]
	0x48, 0x8b, 0x84, 0x24, 16, 0, 0, 0,
	// syscall
	0x0f, 0x05,
	// mov [rsp+8], rax
	0x48, 0x89, 0x84, 0x24, 8, 0, 0, 0,
	// ret
	0xc3
};

#define __syscall(no, arg1, arg2, arg3, arg4, arg5, arg6)\
	(((unsigned long (*)(unsigned long, unsigned long, unsigned long,\
	   unsigned long, unsigned long, unsigned long, unsigned long))__syscall_data)\
		(no, arg1, arg2, arg3, arg4, arg5, arg6))

The __syscall_data array contains machine language instructions which perform a system call, and the __syscall macro "calls" the array as if it were a function. This is why we need a read-write-executable data segment -- otherwise we'd need to implement system calls in the compiler.

C standard library

The C89 standard specifies a bunch of "standard library" functions which any implementation has to make available, e.g. printf(), atoi(), exit(). Fortunately, we don't have to write these functions in the 04 language; we can write them in C.

To use a particular function, a C program needs to include the appropriate header file, e.g. #include <stdio.h> lets you use printf() and other I/O-related functions. Normally, these header files just declare what types the parameters to the functions should be, but we actually put the function implementations there.

Let's take a look at the contents of ctype.h, which provides the functions islower, isupper, etc.:

#ifndef _CTYPE_H
#define _CTYPE_H

#include <stdc_common.h>

int islower(int c) {
	return c >= 'a' && c <= 'z';
}

int isupper(int c) {
	return c >= 'A' && c <= 'Z';
}

int isalpha(int c) {
	return isupper(c) || islower(c);
}

int isalnum(int c) {
	return isalpha(c) || isdigit(c);
}

...

#endif

The first two lines and last line prevent problems when the file is included multiple times. We begin by including stdc_common.h, which has a bunch of functions and type definitions which all our header files use, and then we define each of the necessary C standard library functions.

limitations

There are various minor ways in which this compiler doesn't actually handle all of C89. Here is a (probably incomplete) list of things we do wrong:

• trigraphs are not handled
• char[] string literal initializers can't contain null characters (e.g. char x[] = "a\0b"; doesn't work)
• you can only access members of l-values (e.g. int x = function_which_returns_struct().member; doesn't work)
• no default-int (this is a legacy feature of C, e.g. main() {} can technically stand in for int main() {})
• the keyword auto is not handled (again, a legacy feature of C)
• default: must come after all case labels in a switch statement.
• external variable declarations are ignored, and global variables can only be declared once (e.g. extern int x; int main() { return x; } int x = 5; doesn't work)
• typedefs, and struct/union/enum definitions aren't allowed inside functions
• conditional expressions aren't allowed inside case (horribly, switch (x) { case 5 ? 6 : 3: ; } is legal C).
• bit-fields aren't handled
• Technically, 1[array] is equivalent to array[1], but we don't handle that.
• C89 has very weird typing rules about void*/non-void* inside conditional expressions. We don't handle that properly.
• C89 allows calling functions without declaring them, for legacy reasons. We don't handle that.
• Floating-point constant expressions are very limited. Only double literals and 0 are supported.
• In floating-point literals, the numbers before and after the decimal point must be less than 2⁶⁴.
• The only "address constants" we allow are string literals, e.g. int y, x = &y; is not allowed as a global declaration.
• Redefining a macro is always an error, even if it's the same definition.
• You can't have a variable/function/etc. called defined.
• Various little things about when macros are evaluated in some contexts.
• The horrible, horrible function setjmp, which surely no one uses, is not properly supported. Oh wait, tcc uses it. Fortunately it's not critically important to tcc.
• Wide characters and wide character strings are not supported.
• The localtime() function assumes you are in the UTC+0 timezone.
• mktime() always fails.
• The keywords signed, volatile, register, and const are all ignored, but this should almost never have an effect on a legal C program.
• Converting unsigned long to double/float treats the number as signed (this is incorrect for unsigned long values above 2⁶³).

anecdotes

Making this C compiler took over a month. Here are some interesting things which happened along the way:

• Writing code to parse floating-point numbers in a language which doesn't have floats turned out to be quite a fun challenge! Not all decimal numbers have a perfect floating point representation. You could round 0.1 up to ~0.1000000000000000056, or down to ~0.0999999999999999917. This stage's C compiler should be correct up to rounding (which is all that the C standard requires). But typically C compilers will round to whichever is closest to the decimal value. Implementing this correctly is a lot harder than you might expect. For example,

0.09999999999999999861222121921855432447046041488647460937499
rounds down, but
0.09999999999999999861222121921855432447046041488647460937501
rounds up.

Good luck writing code which handles that!

• Originally, there was a bug where negative powers of 2 were being interpreted as half of their actual value, e.g. x = 0.25; would set x to 0.125, but x = 4;, x = 0.3;, etc. would all work just fine.
• Writing the functions in math.h, although probably not necessary for compiling tcc, was fun! There are quite a few interesting optimizations you can make, and little tricks for avoiding losses in floating-point accuracy.
• A very difficult to track down bug happened the first time I ran tcc: there was a declaration along the lines of char x[] = "a\0b\0c"; but it got compiled as char x[] = "a";!
• Originally, I was just treating labels the same as any other statements, but tcc actually has code like:

...
goto lbl;
...
if (some_condition)
    lbl: do_something();

so the do_something(); was not being considered as part of the if statement.

• The first time I compiled tcc with itself (and then with itself again), I actually got a different executable from the gcc one. After spending a long time looking at disassemblies, I found the culprit:

# if defined(__linux__)
    tcc_define_symbol(s, "__linux__", NULL);
    tcc_define_symbol(s, "__linux", NULL);
# endif

If the __linux__ macro is defined (to indicate that the OS is linux), tcc will also define the __linux__ macro in any programs it compiles. Unlike gcc, our compiler doesn't define the __linux__ macro, so when it's used to compile tcc, tcc won't define it either, no matter how many times you compile it with itself!

modifications of tcc's source code

Some modifications were needed to bring tcc's source code in line with what our compiler expects.

You can find a full list of modifications in diffs.txt, but I'll provide an overview (and explanation) here.

• First, we (and C89) don't allow a comma after the last member in an initializer. In several places, the last comma in an initializer/enum definition was removed, or an irrelevant entry was added to the end.
• Global variables were sometimes declared twice, which we don't support. So a bunch of duplicate declarations were removed.
• The # if defined(__linux__) and # endif mentioned above were removed.
• In a bunch of places, ELFW(something) had to be replaced with ELF64_something due to subtleties of how we evaluate macros.
• offsetof(type, member) isn't considered a constant expression by our compiler, so some initializers were replaced by functions called at the top of main.
• In several places, default: had to be moved to after every case label.
• In two places, -some_long_double_expression had to be replaced with a function call to negate_ld (a function I wrote for negating long doubles). This is because tcc only supports negating long doubles if the compiler which compiled it has an 80-bit long double type, and our compiler doesn't.
• \0 was replaced with \n as a separator for keyword names in the tcc_keywords global variable.
• Forced tcc to use R_X86_64_PC32 relocations, because its plt code doesn't seem to work for static executables.
• Lastly, there's the config.h file, which is normally produced by tcc's configure script, but it's easy to write one manually (see tcc-0.9.27/config.h) The last line causes the inline keyword (added in C99) to be ignored.
• Some stuff was added to the bottom of tcc.c to deal with problems when compiling with musl.

*libc

If you look in tcc's source code, you will not find implementations of any of the C standard library functions. So how can programs compiled with tcc use those functions?

When a program compiled with tcc (under default settings) calls printf, say, it actually gets the instructions for printf from a separate library file (called something like /usr/lib/x86_64-linux-gnu/libc-2.31.so). There are very good reasons for this: for example, if there a security bug were found in printf, it would be much easier to replace the library file than re-compile every program which uses printf.

Now this library file is itself compiled from C source files (typically glibc). So, we can't really say that the self-compiled tcc was built from scratch, and there could be malicious self-replicating code in glibc.

I have to thank Dawid33 for coming up with the idea to try this...

Compiling glibc is difficult (see below), but it's possible to build an old version of musl, an alternate libc, with just our tcc0 executable. This can be done with

make musl

in this directory. Now you can run

cd tcc-0.9.27
./tcc0 -nostdinc -nostdlib -B ../tcc-boostrap -I ../musl-bootstrap/include tcc.c ../musl-bootstrap/lib/*.[oa] -o tcc1

to get a tcc executable that is fully independent of any libc installed on your system. We can do the same with gcc

./tcc0a -nostdinc -nostdlib -B ../tcc-boostrap -I ../musl-bootstrap/include tcc.c ../musl-bootstrap/lib/*.[oa] -o tcc1a

And once more the files tcc1a and tcc1 differ. But with one more round:

./tcc1 -nostdinc -nostdlib -B ../tcc-boostrap -I ../musl-bootstrap/include tcc.c ../musl-bootstrap/lib/*.[oa] -o tcc2

we find that tcc1a and tcc2 do not differ. So glibc has not betrayed us, and we now have a fully-functioning self-compiling compiler built from nothing but human-readable source code.

But hardly anyone uses tcc to compile anything important. For all we know the hypothetical malicious code in gcc or glibc only replicates itself if the compiler is sufficiently advanced (e.g. gcc or clang). Ideally we would build up to a fully bootstrapped build of gcc. Unfortunately that turns out to be quite a challenge...

compiling glibc

You should be able to use musl alongside tcc to build an old version of gcc. This also requires building several tools needed to compile gcc. You should then be able to build an old version of glibc, and with that, a modern version of gcc.

Well, I tried this. And it is an absolute nightmare. GNU has created a horrible web of programs that all depend on each other. According to the recommended build process, you need awk to build awk, sed to build sed, sed to build grep, etc. Here was a "guide" I was starting to write for how to get to glibc:

• install tcc, musl
• build mrsh, make, basic utilities
• chroot
• build & install coreutils
• build & install dash
• build & install sed-4.2
• build & install ld, as (from binutils)
• build & install gcc
• build & install grep-3.7
• build & install awk
• build & install bash
• build & install glibc (didn't work)

Each of these programs uses a ./configure script to set up the code and Makefiles. These scripts are basically impossible to use without already having most of these programs. So, I resorted to configuring the build with the ordinary binary versions of sed, etc. I had on my machine. This made broken Makefiles which I spent hours editing by hand — and is it really compiled from scratch if it's built from computer-generated source files and Makefiles? And although the developers at GNU refrain from declaring variables after statements, they can't help but use bizarre gcc-specific extensions. After hours and hours of fixing compiler errors, I decided to give up.

I'll just say, as a reminder to my future self, and a warning to anyone else who wants to compile glibc from scratch: THIS WAY LIES MADNESS.