MarcelGarus / soil

A byte code specification and a JIT-compiler written in Assembly.

Soil

Soil is a virtual machine specification that is designed to be easy to implement on typical machines.

To get started, run make. This creates some executables:

• assemble: can turn .recipe files (Soil assembly) into .soil files (Soil binaries).
• soil-c: Reference implementation of a Soil interpreter written in C. This is slow.
• soil-asm: A Soil JIT compiler written in Assembly. Some syscalls are missing and there are some bugs.
• soil-rust-compiler: A Soil compiler to FASM written in Rust.

For example, to run the hello.recipe, you can run this:

cat hello.recipe | ./assemble | ./soil-c

The Anatomy of Soil

Soil consists of three parts of state: registers, memory, and byte code.

Soil is not a von Neumann machine – byte code and memory live in separate worlds. Byte code can only read/write the memory, not byte code itself. You can't reflect on the byte code itself, for example, to store pointers to instructions. This gives Soil implementations the freedom to JIT-compile the byte code on startup.

Soil binaries are files that contain byte code and initial memory.

Registers

Soil has 8 registers, all of which hold 64 bits.

name	description
sp	stack pointer
st	status register
a	general-purpose register
b	general-purpose register
c	general-purpose register
d	general-purpose register
e	general-purpose register
f	general-purpose register

Initially, sp is the memory size. All other registers are zero.

Memory

It also has byte-addressed memory. For now, the size of the memory is hardcoded to something big.

Byte Code

Byte code consists of a sequence of instructions.

Soil runs the instructions in sequence, starting from the first. Some instructions alter control flow by jumping to other instructions.

All instructions start with a byte containing the opcode, followed by the arguments to the operation. The following instructions are available:

opcode	mnemonic	arg 0	arg 1	description
00	nop	-	-	Does nothing.
e0	panic	-	-	Panics.
e1	trystart	catch: word	-	If a panic occurs, catches it, resets sp, and jumps to the catch address.
e2	tryend		-	Ends a scope started by trystart.
d0	move	to: reg	from: reg	Sets to to from.
d1	movei	to: reg	value: word	Sets to to value.
d2	moveib	to: reg	value: byte	Sets to to value, zeroing the upper bits.
d3	load	to: reg	from: reg	Interprets from as an address and sets to to the 64 bits at that address in memory.
d4	loadb	to: reg	from: reg	Interprets from as an address and sets to to the 8 bits at that address in memory.
d5	store	to: reg	from: reg	Interprets to as an address and sets the 64 bits at that address in memory to from.
d6	storeb	to: reg	from: reg	Interprets to as an address and sets the 8 bits at that address in memory to from.
d7	push	reg: reg	-	Decreases sp by 8, then runs store sp reg.
d8	pop	reg: reg	-	Runs load reg sp, then increases sp by 8.
f0	jump	to: word	-	Continues executing at the toth byte.
f1	cjump	to: word	-	Runs jump to if st is not 0.
f2	call	target: word	-	Runs jump target. Saves the formerly next instruction on an internal stack so that ret returns.
f3	ret	-	-	Returns to the instruction after the matching call.
f4	syscall	number: byte	-	Performs a syscall. Behavior depends on the syscall. The syscall can access all registers and memory.
c0	cmp	left: reg	right: reg	Saves left - right in st.
c1	isequal	-	-	If st is 0, sets st to 1, otherwise to 0.
c2	isless	-	-	If st is less than 0, sets st to 1, otherwise to 0.
c3	isgreater	-	-	If st is greater than 0, sets st to 1, otherwise to 0.
c4	islessequal	-	-	If st is 0 or less, sets st to 1, otherwise to 0.
c5	isgreaterequal	-	-	If st is 0 or greater, sets st to 1, otherwise to 0.
c6	isnotequal	-	-	If st is 0, sets st to 0, otherwise to 1.
c7	fcmp	left: reg	right: reg	Saves left - right in st.
c8	fisequal	-	-	If st is 0, sets st to 1, otherwise to 0.
c9	fisless	-	-	If st is less than 0, sets st to 1, otherwise to 0.
ca	fisgreater	-	-	If st is greater than 0, sets st to 1, otherwise to 0.
cb	fislessequal	-	-	If st is 0 or less, sets st to 1, otherwise to 0.
cc	fisgreaterequal	-	-	If st is 0 or greater, sets st to 1, otherwise to 0.
cd	fisnotequal	-	-	If st is 0, sets st to 0, otherwise to 1.
ce	inttofloat	reg: reg	-	Interprets reg as an int and sets it to a float of about the same value. TODO: specify edge cases
cf	floattoint	reg: reg	-	Interprets reg as a float and sets it to its int, rounded down. TODO: specify edge cases
a0	add	to: reg	from: reg	Adds from to to.
a1	sub	to: reg	from: reg	Subtracts from from to.
a2	mul	to: reg	from: reg	Multiplies from and to. Saves the result in to.
a3	div	dividend: reg	divisor: reg	Divides dividend by divisor. Saves the quotient in dividend.
a4	rem	dividend: reg	divisor: reg	Divides dividend by divisor. Saves the remainder in dividend.
a5	fadd	to: reg	from: reg	Adds from to to, interpreted as floats.
a6	fsub	to: reg	from: reg	Subtracts from from to, interpreted as floats.
a7	fmul	to: reg	from: reg	Multiplies from and to, interpreted as floats. Saves the result in to.
a8	fdiv	dividend: reg	divisor: reg	Divides dividend by divisor, interpreted as floats. Saves the quotient in dividend.
b0	and	to: reg	from: reg	Binary-ands to and from. Saves the result in to.
b1	or	to: reg	from: reg	Binary-ors to and from. Saves the result in to.
b2	xor	to: reg	from: reg	Binary-xors to and from. Saves the result in to.
b3	not	to: reg	-	Inverts the bits of to.

To make memorization easier, the first characters of the instruction hex opcodes describe what kind of instruction it is:

• 00: nop
• a*: arithmetic
• b*: binary
• c*: comparisons / conversions
• d*: data operations
• e*: error
• f*: control flow

Binaries

Soil binaries are stuctured like this:

• magic bytes soil (4 bytes)
• the only thing following are sections, each of which has:
  • type (1 byte)
  • length (8 byte), useful for skipping sections
  • content (length parsed above)
• byte code
  • section type 0
  • length (8 bytes)
  • byte code (length parsed above)
• initial memory
  • section type 1
  • length (8 bytes)
  • content (length parsed above)
• name
  • section type 2
  • length (8 bytes)
  • content (length parsed above)
• labels
  • section type 3
  • length (8 bytes)
  • number of labels (8 bytes)
  • for each label:
    • position in the byte code (8 bytes)
    • label length (8 bytes)
    • label (length parsed above)
• description
  • section type 4
  • length (8 bytes)
  • content (length parsed above)