SubX: A minimalist assembly language for a subset of the x86 ISA

SubX is a simple, minimalist stack for programming your computer.

$ git clone https://github.com/akkartik/mu
$ cd mu/subx
$ ./subx  # print out a help message

SubX is designed:

• to enable automation of arbitrary manual tests
• to be easy to implement in itself, and
• to help learn and teach the x86 instruction set.

It requires a Unix-like environment with a C++ compiler (Linux or BSD or Mac OS). Running subx will transparently compile it as necessary.

You can generate native ELF binaries with it that run on a bare Linux kernel. No other dependencies needed.

$ ./subx translate examples/ex1.subx -o examples/ex1
$ ./examples/ex1  # only on Linux
$ echo $?
42

You can run the generated binaries on an interpreter/VM for better error messages.

$ ./subx run examples/ex1  # on Linux or BSD or OS X
$ echo $?
42

Emulated runs generate a trace that permits time-travel debugging.

$ ./subx --debug translate examples/factorial.subx -o examples/factorial
saving address->label information to 'labels'
saving address->source information to 'source_lines'

$ ./subx --debug --trace run examples/factorial
saving trace to 'last_run'

$ ../browse_trace/browse_trace last_run  # text-mode debugger UI

You can write tests for your assembly programs. The entire stack is thoroughly covered by automated tests. SubX's tagline: tests before syntax.

$ ./subx test
$ ./subx run apps/factorial test

You can use SubX to translate itself. For example, running natively on Linux:

# generate translator phases using the C++ translator
$ ./subx translate 0*.subx apps/subx-common.subx apps/hex.subx    -o hex
$ ./subx translate 0*.subx apps/subx-common.subx apps/survey.subx -o survey
$ ./subx translate 0*.subx apps/subx-common.subx apps/pack.subx   -o pack
$ ./subx translate 0*.subx apps/subx-common.subx apps/assort.subx -o assort
$ ./subx translate 0*.subx apps/subx-common.subx apps/dquotes.subx -o dquotes
$ ./subx translate 0*.subx apps/subx-common.subx apps/tests.subx  -o tests
$ chmod +x hex survey pack assort dquotes tests

# use the generated translator phases to translate SubX programs
$ cat examples/ex1.subx |./tests |./dquotes |./assort |./pack |./survey |./hex > a.elf
$ chmod +x a.elf
$ ./a.elf
$ echo $?
42

# or, automating the above steps
$ ./ntranslate ex1.subx
$ chmod +x a.elf
$ ./a.elf
$ echo $?
42

Or, running in a VM on other platforms:

$ ./translate ex1.subx  # generates identical a.elf to above
$ ./subx run a.elf
$ echo $?
42

You can use it to learn about the x86 processor that (almost certainly) runs your computer. (See below.)

You can read its tiny zero-dependency internals and understand how they work. You can hack on it, and its thorough tests will raise the alarm when you break something.

Eventually you will be able to program in higher-level notations. But you'll always have tests as guardrails and traces for inspecting runs. The entire stack will always be designed for others to comprehend. You'll always be empowered to understand how things work, and change what doesn't work for you. You'll always be expected to make small changes during upgrades.

What it looks like

Here is the first example we ran above, a program that just returns 42:

bb/copy-to-EBX  0x2a/imm32  # 42 in hex
b8/copy-to-EAX  1/imm32/exit
cd/syscall  0x80/imm8

Every line contains at most one instruction. Instructions consist of words separated by whitespace. Words may be opcodes (defining the operation being performed) or arguments (specifying the data the operation acts on). Any word can have extra metadata attached to it after /. Some metadata is required (like the /imm32 and /imm8 above), but unrecognized metadata is silently skipped so you can attach comments to words (like the instruction name /copy-to-EAX above, or the /exit operand).

SubX doesn't provide much syntax (there aren't even the usual mnemonics for opcodes), but it does provide error-checking. If you miss an operand or accidentally add an extra operand you'll get a nice error. SubX won't arbitrarily interpret bytes of data as instructions or vice versa.

So much for syntax. What do all these numbers actually mean? SubX supports a small subset of the 32-bit x86 instruction set that likely runs on your computer. (Think of the name as short for "sub-x86".) Instructions operate on a few registers:

• Six general-purpose 32-bit registers: EAX, EBX, ECX, EDX, ESI and EDI
• Two additional 32-bit registers: ESP and EBP (I suggest you only use these to manage the call stack.)
• Four 1-bit flag registers for conditional branching:
  • zero/equal flag ZF
  • sign flag SF
  • overflow flag OF
  • carry flag CF

SubX programs consist of instructions like 89/copy, 01/add, 3d/compare and 52/push-ECX which modify these registers as well as a byte-addressable memory. For a complete list of supported instructions, run subx help opcodes.

(SubX doesn't support floating-point registers yet. Intel processors support an 8-bit mode, 16-bit mode and 64-bit mode. SubX will never support them. There are other flags. SubX will never support them. There are also many more instructions that SubX will never support.)

It's worth distinguishing between an instruction's operands and its arguments. Arguments are provided directly in instructions. Operands are pieces of data in register or memory that are operated on by instructions. Intel processors determine operands from arguments in fairly complex ways.

Lengthy interlude: How x86 instructions compute operands

The Intel processor manual is the final source of truth on the x86 instruction set, but it can be forbidding to make sense of, so here's a quick orientation. You will need familiarity with binary numbers, and maybe a few other things. Email me any time if something isn't clear. I love explaining this stuff for as long as it takes. The bad news is that it takes some getting used to. The good news is that internalizing the next 500 words will give you a significantly deeper understanding of your computer.

Most instructions operate on an operand in register or memory ('reg/mem'), and a second operand in a register. The register operand is specified fairly directly using the 3-bit /r32 argument:

• 0 means register EAX
• 1 means register ECX
• 2 means register EDX
• 3 means register EBX
• 4 means register ESP
• 5 means register EBP
• 6 means register ESI
• 7 means register EDI

The reg/mem operand, however, gets complex. It can be specified by 1-7 arguments, each ranging in size from 2 bits to 4 bytes.

The key argument that's always present for reg/mem operands is /mod, the addressing mode. This is a 2-bit argument that can take 4 possible values, and it determines what other arguments are required, and how to interpret them.

• If /mod is 3: the operand is in the register described by the 3-bit /rm32 argument similarly to /r32 above.

• If /mod is 0: the operand is in the address provided in the register described by /rm32. That's *rm32 in C syntax.

• If /mod is 1: the operand is in the address provided by adding the register in /rm32 with the (1-byte) displacement. That's *(rm32 + /disp8) in C syntax.

• If /mod is 2: the operand is in the address provided by adding the register in /rm32 with the (4-byte) displacement. That's *(/rm32 + /disp32) in C syntax.

In the last three cases, one exception occurs when the /rm32 argument contains 4. Rather than encoding register ESP, it means the address is provided by three whole new arguments (/base, /index and /scale) in a totally different way:

reg/mem = *(/base + /index * (2 ^ /scale))

(There are a couple more exceptions ☹; see Table 2-2 and Table 2-3 of the Intel manual for the complete story.)

Phew, that was a lot to take in. Some examples to work through as you reread and digest it:

1. To read directly from the EAX register, /mod must be 3 (direct mode), and /rm32 must be 0. There must be no /base, /index or /scale arguments.

2. To read from *EAX (in C syntax), /mod must be 0 (indirect mode), and the /rm32 argument must be 0. There must be no /base, /index or /scale arguments.

3. To read from *(EAX+4), /mod must be 1 (indirect + disp8 mode), /rm32 must be 0, there must be no SIB byte, and there must be a single displacement byte containing 4.

4. To read from *(EAX+ECX+4), one approach would be to set /mod to 1 as above, /rm32 to 4 (SIB byte next), /base to 0, /index to 1 (ECX) and a single displacement byte to 4. (What should the scale bits be? Can you think of another approach?)

5. To read from *(EAX+ECX+1000), one approach would be:

   • /mod: 2 (indirect + disp32)
   • /rm32: 4 (/base, /index and /scale arguments required)
   • /base: 0 (EAX)
   • /index: 1 (ECX)
   • /disp32: 4 bytes containing 1000

Putting it all together

Here's a more meaty example:

This program sums the first 10 natural numbers. By convention I use horizontal tabstops to help read instructions, dots to help follow the long lines, comments before groups of instructions to describe their high-level purpose, and comments at the end of complex instructions to state the low-level operation they perform. Numbers are always in hexadecimal (base 16) and must start with a digit ('0'..'9'); use the '0x' prefix when a number starts with a letter ('a'..'f'). I tend to also include it as a reminder when numbers look like decimal numbers.

Try running this example now:

$ ./subx translate examples/ex3.subx -o examples/ex3
$ ./subx run examples/ex3
$ echo $?
55

If you're on Linux you can also run it natively:

$ ./examples/ex3
$ echo $?
55

Use it now to follow along for a more complete tour of SubX syntax.

The syntax of SubX programs

SubX programs map to the same ELF binaries that a conventional Linux system uses. Linux ELF binaries consist of a series of segments. In particular, they distinguish between code and data. Correspondingly, SubX programs consist of a series of segments, each starting with a header line: == followed by a name and approximate starting address.

All code must lie in a segment called 'code'. Execution begins at the start of the code segment by default.

You can reuse segment names:

== code
...A...

== data
...B...

== code
...C...

The code segment now contains the instructions of A as well as C.

Within the code segment, each line contains a comment, label or instruction. Comments start with a # and are ignored. Labels should always be the first word on a line, and they end with a :.

Instruction arguments must specify their type, from:

• /mod
• /rm32
• /r32
• /subop (sometimes the /r32 bits in an instruction are used as an extra opcode)
• displacement: /disp8 or /disp32
• immediate: /imm8 or /imm32

Different instructions (opcodes) require different arguments. SubX will validate each instruction in your programs, and raise an error anytime you miss or spuriously add an argument.

I recommend you order arguments consistently in your programs. SubX allows arguments in any order, but only because that's simplest to explain/implement. Switching order from instruction to instruction is likely to add to the reader's burden. Here's the order I've been using after opcodes:

        |<--------- reg/mem --------->|        |<- reg/mem? ->|
/subop  /mod /rm32  /base /index /scale  /r32   /displacement   /immediate

Instructions can refer to labels in displacement or immediate arguments, and they'll obtain a value based on the address of the label: immediate arguments will contain the address directly, while displacement arguments will contain the difference between the address and the address of the current instruction. The latter is mostly useful for jump and call instructions.

Functions are defined using labels. By convention, labels internal to functions (that must only be jumped to) start with a $. Any other labels must only be called, never jumped to. All labels must be unique.

A special label is Entry, which can be used to specify/override the entry point of the program. It doesn't have to be unique, and the latest definition will override earlier ones.

(The Entry label, along with duplicate segment headers, allows programs to be built up incrementally out of multiple layers.)

The data segment consists of labels as before and byte values. Referring to data labels in either code segment instructions or data segment values (using the imm32 metadata either way) yields their address.

Automatic tests are an important part of SubX, and there's a simple mechanism to provide a test harness: all functions that start with test- are called in turn by a special, auto-generated function called run-tests. How you choose to call it is up to you.

I try to keep things simple so that there's less work to do when I eventually implement SubX in SubX. But there is one convenience: instructions can provide a string literal surrounded by quotes (") in an imm32 argument. SubX will transparently copy it to the data segment and replace it with its address. Strings are the only place where a SubX word is allowed to contain spaces.

That should be enough information for writing SubX programs. The examples/ directory provides some fodder for practice, giving a more gradual introduction to SubX features. This repo includes the binary for all examples. At any commit, an example's binary should be identical bit for bit with the result of translating the corresponding .subx file. The binary should also be natively runnable on a Linux system running on Intel x86 processors, either 32- or 64-bit. If either of these invariants is broken it's a bug on my part.

Roadmap and status

• Self-hosting. (✓)

  tests |dquotes |assort |pack |survey |hex

• A script to package SubX together with a minimal Linux kernel image (compiled from source, of course).

• Testable, dependency-injected vocabulary of primitives

  • Streams: read(), write(). (✓)
  • exit() (✓)
  • Sockets
  • Files
  • Concurrency, and a framework for testing blocking code

• Higher-level notations. Like programming languages, but with thinner implementations that you can -- and are expected to! -- modify.

  • syntax for addressing modes: %reg, *reg, *(reg+disp), *(reg+reg+disp), *(reg+reg<<n + disp)
  • function calls in a single line, using addressing modes for arguments
  • syntax for controlling a type checker, akin to the top-level Mu language
  • a register allocation verifier. Programmer provides registers for variables; verifier checks that register reads are for the same type that was last written -- across all control flow paths.

Running

subx currently has the following sub-commands:

• subx help: some helpful documentation to have at your fingertips.

• subx test: runs all automated tests.

• subx translate <input files> -o <output ELF binary>: translates .subx files into an executable ELF binary.

• subx run <ELF binary>: simulates running the ELF binaries emitted by subx translate. Useful for debugging, and also enables more thorough testing of translate.

  Remember, not all 32-bit Linux binaries are guaranteed to run. I'm not building general infrastructure here for all of the x86 instruction set. SubX is about programming with a small, regular subset of 32-bit x86.

A few hints for debugging

Writing programs in SubX is surprisingly pleasant and addictive. Reading programs is a work in progress, and hopefully the extensive unit tests help. However, debugging programs is where one really faces up to the low-level nature of SubX. Even the smallest modifications need testing to make sure they work. In my experience, there is no modification so small that I get it working on the first attempt. And when it doesn't work, there are no clear error messages. Machine code is too simple-minded for that. You can't use a debugger, since SubX's simplistic ELF binaries contain no debugging information. So debugging requires returning to basics and practicing with a new, more rudimentary but hopefully still workable toolkit:

• Start by nailing down a concrete set of steps for reproducibly obtaining the error or erroneous behavior.

• If possible, turn the steps into a failing test. It's not always possible, but SubX's primary goal is to keep improving the variety of tests one can write.

• Start running the single failing test alone. This involves modifying the top of the program (or the final .subx file passed in to subx translate) by replacing the call to run-tests with a call to the appropriate test- function.

• Generate a trace for the failing test while running your program in emulated mode (subx run):

  $ ./subx translate input.subx -o binary
  $ ./subx --trace run binary arg1 arg2  2>trace
  

  The ability to generate a trace is the essential reason for the existence of subx run mode. It gives far better visibility into program internals than running natively.

• As a further refinement, it is possible to render label names in the trace by adding a second flag to both the translate and run commands:

  $ ./subx --debug translate input.subx -o binary
  $ ./subx --debug --trace run binary arg1 arg2  2>trace
  

  subx --debug translate emits a mapping from label to address in a file called labels. subx --debug --trace run reads in the labels file at the start and prints out any matching label name as it traces each instruction executed.

  Here's a sample of what a trace looks like, with a few boxes highlighted:

  

  Each of the green boxes shows the trace emitted for a single instruction. It starts with a line of the form run: inst: ___ followed by the opcode for the instruction, the state of registers before the instruction executes, and various other facts deduced during execution. Some instructions first print a matching label. In the above screenshot, the red boxes show that address 0x0900005e maps to label $loop and presumably marks the start of some loop. Function names get similar run: == label lines.

• One trick when emitting traces with labels:

  $ grep label trace
  

  This is useful for quickly showing you the control flow for the run, and the function executing when the error occurred. I find it useful to start with this information, only looking at the complete trace after I've gotten oriented on the control flow. Did it get to the loop I just modified? How many times did it go through the loop?

• Once you have SubX displaying labels in traces, it's a short step to modify the program to insert more labels just to gain more insight. For example, consider the following function:

  

  This function contains a series of jump instructions. If a trace shows is-hex-lowercase-byte? being encountered, and then $is-hex-lowercase-byte?:end being encountered, it's still ambiguous what happened. Did we hit an early exit, or did we execute all the way through? To clarify this, add temporary labels after each jump:

  

  Now the trace should have a lot more detail on which of these labels was reached, and precisely when the exit was taken.

• If you find yourself wondering, "when did the contents of this memory address change?", subx run has some rudimentary support for watch points. Just insert a label starting with $watch- before an instruction that writes to the address, and its value will start getting dumped to the trace after every instruction thereafter.

• Once we have a sense for precisely which instructions we want to look at, it's time to look at the trace as a whole. Key is the state of registers before each instruction. If a function is receiving bad arguments it becomes natural to inspect what values were pushed on the stack before calling it, tracing back further from there, and so on.

  I occasionally want to see the precise state of the stack segment, in which case I uncomment a commented-out call to dump_stack() in the vm.cc layer. It makes the trace a lot more verbose and a lot less dense, necessitating a lot more scrolling around, so I keep it turned off most of the time.

• If the trace seems overwhelming, try browsing it in the 'time-travel debugger'.

Hopefully these hints are enough to get you started. The main thing to remember is to not be afraid of modifying the sources. A good debugging session gets into a nice rhythm of generating a trace, staring at it for a while, modifying the sources, regenerating the trace, and so on. Email me if you'd like another pair of eyes to stare at a trace, or if you have questions or complaints.

Reference documentation on available primitives

Data Structures

• Kernel strings: null-terminated arrays of bytes. Unsafe and to be avoided, but needed for interacting with the kernel.

• Strings: length-prefixed arrays of bytes. String contents are preceded by 4 bytes (32 bytes) containing the length of the array.

• Slices: a pair of 32-bit addresses denoting a half-open

  `start`, `end`) interval to live memory with a consistent lifetime.
  
  Invariant: `start` <= `end`
  
  

• Streams: strings prefixed by 32-bit write and read indexes that the next write or read goes to, respectively.

  • offset 0: write index
  • offset 4: read index
  • offset 8: length of array (in bytes)
  • offset 12: start of array data

  Invariant: 0 <= read <= write <= length

• File descriptors (fd): Low-level 32-bit integers that the kernel uses to track files opened by the program.

• File: 32-bit value containing either a fd or an address to a stream (fake file).

• Buffered files (buffered-file): Contain a file descriptor and a stream for buffering reads/writes. Each buffered-file must exclusively perform either reads or writes.

'system calls'

As I said at the top, a primary design goal of SubX (and Mu more broadly) is to explore ways to turn arbitrary manual tests into reproducible automated tests. SubX aims for this goal by baking testable interfaces deep into the stack, at the OS syscall level. The idea is that every syscall that interacts with hardware (and so the environment) should be dependency injected so that it's possible to insert fake hardware in tests.

But those are big goals. Here are the syscalls I have so far:

• write: takes two arguments, a file f and an address to array s.

  Comparing this interface with the Unix write() syscall shows two benefits:

  1. SubX can handle 'fake' file descriptors in tests.

  2. write() accepts buffer and its length in separate arguments, which requires callers to manage the two separately and so can be error-prone. SubX's wrapper keeps the two together to increase the chances that we never accidentally go out of array bounds.

• read: takes two arguments, a file f and an address to stream s. Reads as much data from f as can fit in (the free space of) s.

  Like with write(), this wrapper around the Unix read() syscall adds the ability to handle 'fake' file descriptors in tests, and reduces the chances of clobbering outside array bounds.

  One bit of weirdness here: in tests we do a redundant copy from one stream to another. See the comments before the implementation for a discussion of alternative interfaces.

• stop: takes two arguments:

  • ed is an address to an exit descriptor. Exit descriptors allow us to exit() the program in production, but return to the test harness within tests. That allows tests to make assertions about when exit() is called.
  • value is the status code to exit() with.

  For more details on exit descriptors and how to create one, see the comments before the implementation.

• new-segment

  Allocates a whole new segment of memory for the program, discontiguous with both existing code and data (heap) segments. Just a more opinionated form of mmap.

• allocate: takes two arguments, an address to allocation-descriptor ad and an integer n

  Allocates a contiguous range of memory that is guaranteed to be exclusively available to the caller. Returns the starting address to the range in EAX.

  An allocation descriptor tracks allocated vs available addresses in some contiguous range of memory. The int specifies the number of bytes to allocate.

  Explicitly passing in an allocation descriptor allows for nested memory management, where a sub-system gets a chunk of memory and further parcels it out to individual allocations. Particularly helpful for (surprise) tests.

• ... (to be continued)

I will continue to import syscalls over time from the old Mu VM in the parent directory, which has experimented with interfaces for the screen, keyboard, mouse, disk and network.

primitives built atop system calls

(Compound arguments are usually passed in by reference. Where the results are compound objects that don't fit in a register, the caller usually passes in allocated memory for it.)

assertions for tests

• check-ints-equal: fails current test if given ints aren't equal
• check-stream-equal: fails current test if stream doesn't match string
• check-next-stream-line-equal: fails current test if next line of stream until newline doesn't match string

error handling

• error: takes three arguments, an exit-descriptor, a file and a string (message)

  Prints out the message to the file and then exits using the provided exit-descriptor.

• error-byte: like error but takes an extra byte value that it prints out at the end of the message.

predicates

• kernel-string-equal?: compares a kernel string with a string

• string-equal?: compares two strings

• stream-data-equal?: compares a stream with a string

• next-stream-line-equal?: compares with string the next line in a stream, from read index to newline

• slice-empty?: checks if the start and end of a slice are equal

• slice-equal?: compares a slice with a string

• slice-starts-with?: compares the start of a slice with a string

• slice-ends-with?: compares the end of a slice with a string

writing to disk

• write: string -> file
  • Can also be used to cat a string into a stream.
  • Will abort the entire program if destination is a stream and doesn't have enough room.
• write-stream: stream -> file
  • Can also be used to cat one stream into another.
  • Will abort the entire program if destination is a stream and doesn't have enough room.
• write-slice: slice -> stream
  • Will abort the entire program if there isn't enough room in the destination stream.
• append-byte: int -> stream
  • Will abort the entire program if there isn't enough room in the destination stream.
• append-byte-hex: int -> stream
  • textual representation in hex, no '0x' prefix
  • Will abort the entire program if there isn't enough room in the destination stream.
• print-int32: int -> stream
  • textual representation in hex, including '0x' prefix
  • Will abort the entire program if there isn't enough room in the destination stream.
• write-buffered: string -> buffered-file
• write-slice-buffered: slice -> buffered-file
• flush: buffered-file
• write-byte-buffered: int -> buffered-file
• print-byte-buffered: int -> buffered-file
  • textual representation in hex, no '0x' prefix
• print-int32-buffered: int -> buffered-file
  • textual representation in hex, including '0x' prefix

reading from disk

• read: file -> stream
  • Can also be used to cat one stream into another.
  • Will silently stop reading when destination runs out of space.
• read-byte-buffered: buffered-file -> byte
• read-line-buffered: buffered-file -> stream
  • Will abort the entire program if there isn't enough room.

non-IO operations on streams

• new-stream: allocates space for a stream of n elements, each occupying b bytes.
  • Will abort the entire program if n*b requires more than 32 bits.
• clear-stream: resets everything in the stream to 0 (except its length).
• rewind-stream: resets the read index of the stream to 0 without modifying its contents.

reading/writing hex representations of integers

• is-hex-int?: takes a slice argument, returns boolean result in EAX
• parse-hex-int: takes a slice argument, returns int result in EAX
• is-hex-digit?: takes a 32-bit word containing a single byte, returns boolean result in EAX.
• from-hex-char: takes a hexadecimal digit character in EAX, returns its numeric value in EAX
• to-hex-char: takes a single-digit numeric value in EAX, returns its corresponding hexadecimal character in EAX

tokenization

from a stream:

• next-token: stream, delimiter byte -> slice
• skip-chars-matching: stream, delimiter byte
• skip-chars-not-matching: stream, delimiter byte

from a slice:

• next-token-from-slice: start, end, delimiter byte -> slice

  • Given a slice and a delimiter byte, returns a new slice inside the input that ends at the delimiter byte.

• skip-chars-matching-in-slice: curr, end, delimiter byte -> new-curr (in EAX)

• skip-chars-not-matching-in-slice: curr, end, delimiter byte -> new-curr (in EAX)

Conclusion

The hypothesis of Mu and SubX is that designing the entire system to be testable from day 1 and from the ground up would radically impact the culture of the eco-system in a way that no bolted-on tool or service at higher levels can replicate:

• Tests would make it easier to write programs that can be easily understood by newcomers.

• More broad-based understanding would lead to more forks.

• Tests would make it easy to share code across forks. Copy the tests over, and then copy code over and polish it until the tests pass. Manual work, but tractable and without major risks.

• The community would gain a diversified portfolio of forks for each program, a “wavefront” of possible combinations of features and alternative implementations of features. Application writers who wrote thorough tests for their apps (something they just can’t do today) would be able to bounce around between forks more easily without getting locked in to a single one as currently happens.

• There would be a stronger culture of reviewing the code for programs you use or libraries you depend on. More eyeballs would make more bugs shallow.

Resources

• Single-page cheatsheet for the x86 ISA (pdf; cached local copy)
• Concise reference for the x86 ISA
• Intel processor manual (pdf)
• Some details on the unconventional organization of this project.

Inspirations

• “Creating tiny ELF executables”
• “Bootstrapping a compiler from nothing”
• Forth implementations like StoneKnifeForth