More Booleans, Conditionals

Boolean support in the compiler

We left off with a compiler that would handle the boolean values true and false. On our to-do list: compile unary boolean expressions (not e), (num? e), and (zero? e).

Let's start with not. As a reminder, not should evaluate to true (i.e., should put the runtime representation of true into rax!) when its argument is false; otherwise, it should evaluate to false.

It seems like we need a way to compare the runtime representations of values. For this, we'll use the x86-64 instruction cmp. cmp X,Y compares X to Y. It then sets processor flags based on the result. There are a bunch of flags, and we'll talk about more of them later in the class; for now, we just need to know that cmp sets the flag ZF to 1 if its arguments are the same and 0 otherwise.

Flags aren't like registers–we don't access them directly in assembly code. These flags then modify the behavior of subsequent instructions. We'll see more examples of this next lecture when we talk about conditionals. For now, we're going to use another instruction, setz, in order to access ZF. setz takes a register (*) and sets the last byte of that register to 1 (i.e., 0b00000001) if ZF is set and 0 if ZF is not set.

(*) It actually just takes the lower byte of a register, which are notated differently in assembly–for instance, the lower byte of rax is written al. Our assembly library takes care of this, so we won't talk about it too much in class.

In pseudo-assembly, how we're going to implement the not operator:

not:    
        cmp rax, 0b00011111
        mov rax, 0
        setz rax        
        shl rax, 7
        or rax, 0b0011111

So, now we can implement not:

let bool_shift = 7
let bool_mask = 0b1111111
let bool_tag = 0b0011111

let rec compile_exp (exp : s_exp) : directive list =
  match exp with
  (* some cases elided ... *)
  | Sym "true" -> [Mov (Reg Rax, Imm ((1 lsl bool_shift) lor bool_tag))]
  | Sym "false" -> [Mov (Reg Rax, Imm ((0 lsl bool_shift) lor bool_tag))]
  | Lst [Sym "not"; arg] ->
    compile_exp arg @
    [ Cmp (Reg Rax, Imm ((0 lsl bool_shift) lor bool_tag)) (* compare rax to false *)
    ; Mov (Reg Rax, Imm 0) (* zero out rax *)
    ; Setz (Reg Rax) (* 1 if ZF is set (meaning rax contained false), 0 otherwise *)
    ; Shl (Reg Rax, Imm bool_shift) (* rax << bool_shift *)
    ; Or (Reg Rax, Imm bool_tag) (* tag rax as a boolean: rax = rax | bool_tag *)
    ]

There's some duplicate logic here. We're going to make a helper function called operand_of_bool, which makes an instruction operand from a boolean using shift and or:

let operand_of_bool (b : bool) : operand =
  Imm (((if b then 1 else 0) lsl bool_shift) lor bool_tag)

We can do the same thing for numbers:

let operand_of_num (x : int) : operand =
  Imm ((x lsl num_shift) lor num_tag)

(We include lor num_tag here to be symmetric with operand_to_bool, but everything would work if we left it off–why?)

Lastly, we're going to re-use the code to convert ZF to a boolean:

let zf_to_bool: directive list =
  [Mov (Reg Rax, Imm 0) (* zero out rax *)
  ; Setz (Reg Rax) (* 1 if ZF is set, 0 otherwise *)
  ; Shl (Reg Rax, Imm bool_shift) (* rax << bool_shift *)
  ; Or (Reg Rax, Imm bool_tag) (* tag rax as a boolean: rax = rax | bool_tag *)
  ]

zf_to_bool is a list, not a function. How is that possible? Won't it depend on the value we're trying to convert? It does not! This is a list of instructions that set rax to the runtime representation of true if ZF is set and to the runtime representation of false otherwise.

Now we can implement zero? easily:

let rec compile_exp (exp : s_exp) : directive list =
  match exp with
  (* some cases elided ... *)
  | Sym "true" -> [Mov (Reg Rax, operand_of_bool true)]
  | Sym "false" -> [Mov (Reg Rax, operand_of_bool false)]
  | Lst [Sym "not"; arg] ->
    compile_exp arg @
    [ Cmp (Reg Rax, operand_of_bool false) ]
    @ zf_to_bool
  | Lst [Sym "zero?"; arg] ->
    compile_exp arg @
    [ Cmp (Reg Rax, operand_of_num 0) ]
    @ zf_to_bool

Lastly, we can implement num?. We can detect if a value is a number by looking at the last two bits and seeing if they are both zero. We can do that like this:

let rec compile_exp (exp : s_exp) : directive list =
  match exp with
  (* some cases elided ... *)
  | Lst [Sym "num?"; arg] ->
    compile_exp arg @
    [ And (Reg Rax, Imm num_mask); Cmp (Reg Rax, Imm num_tag) ]
    @ zf_to_bool

Conditionals

Now that we've implemented booleans, we can implement if. Our if form looks like this:

(if <test> <then> <else>)

An if expression evaluates to the then expression if test evaluates to a "truthy" value, and evaluates to the else expression otherwise. Remember that in our language, all values other than false are truthy!

What makes these conditional expressions different from operations we've seen before is that we'll need to evaluate different expressions depending on the value of another expression. This is easy in the interpreter–we'll just use OCaml's if expression! In the compiler, we'll rely on a feature of x86-64 that we haven't seen yet.

Conditionals in the interpreter

This part is pretty simple! We'll just add a case to interp_exp:

interp.ml

let rec interp_exp (exp : s_exp) : value =
  match exp with
  (* some cases elided... *)
  | Lst [Sym "if"; test_exp; then_exp; else_exp] ->
    if interp_exp test_exp = Boolean false then interp_exp else_exp
    else interp_exp then_exp

And that's it! The one thing to note here is that we only evaluate one of the two expressions.

Conditionals in the compiler

x86-64 doesn't have "if" built in, but it does have a standard way of implementing conditionals: with conditional jumps.

So far, all of the assembly code we've seen is straight-line code: we start executing instructions at the entry label, and keep going until we get to ret. We can write straight-line code in higher-level languages too (and this code is generally pretty easy to compile to assembly). Higher-level languages also have various constructs to execute code conditionally, or more than once–things like conditionals and loops and functions.

x86-64 machine code, like most machine codes, really only has one way of writing non straight-line code: jumps. A jump instruction lets us start executing from a label elsewhere in our program. It's what the runtime does to start executing from our entry label.

A conditional jump lets us jump to another label depending on the flags we talked about above. We'll be using jz <label>, which jumps to a given label if and only if the ZF flag is set. So in order to compile (if test then else), we'll want something like:

    ; code for the test expression
    cmp rax, 0b00011111 ; compare to boolean false
    jz else
    ; code for the then expression
    jmp continue
else:
    ; code for the else expression
continue:

The "then" code is skipped when the test expression is false, because of the jz instruction. The "else" code is skipped whenever we evaluate the "then" code, because of the jmp instruction. Cool, right?

Our OCaml implementation follows that pseudocode:

compile.ml

let rec compile_exp (exp : s_exp) : directive list =
  match exp with
  (* some cases elided ... *)
  | Lst [Sym "if"; test_exp; then_exp; else_exp] ->
    compile_exp test_exp
    @ [Cmp (Reg Rax, operand_of_bool false); Jz "else"]
    @ compile_exp then_exp @ [Jmp "continue"]
    @ [Label "else"]
    @ compile_exp else_exp 
    @ [Label "continue"]

There's one big problem here. What if we have more than one if expression? Something like this:

(if (num? 4) (if (num? false) 1 2) 3)

Right now, our assembler is going to throw an error if we try to compile this program, something like:

program.s:17: error: label `_else' inconsistently redefined
program.s:13: note: label `_else' originally defined here
program.s:19: error: label `_continue' inconsistently redefined
program.s:15: note: label `_continue' originally defined here

We're using our label names more than once! That's not going to work. We'll need to make sure that each if expression has its own labels for else and continue. We'll use a function called gensym (short for "generated symbol") in order to generate unique label names. We can call gensym like this:

$ Util.gensym "else";;
"else__0"
$ Util.gensym "else";;
"else__1"
$ Util.gensym "continue";;
"continue__2"

This function is very different from, say, our compile_exp or interp_exp functions: it returns a different output every time we call it! (Indeed, its whole purpose is to return a different output every time we call it.) It's defined like this:

util.ml

let gensym : string -> string =
  let counter = ref 0 in
  fun s ->
    let symbol = Printf.sprintf "%s__%d" s !counter in
    counter := !counter + 1 ;
    symbol

The counter variable is what makes this function work. counter is a reference to an integer; it works sort of like a variable in a typical imperative language like Java or Python. We can update its value with counter := <new value> and read its value with !counter. This little function is a good example of idiomatic usage of references in OCaml: use references as little as possible, and hide them in functions that do a specific thing. We can't update counter from outside this function.

Using our gensym function, we can complete our if compiler:

let rec compile_exp (exp : s_exp) : directive list =
  match exp with
  (* some cases elided ... *)
  | Lst [Sym "if"; test_exp; then_exp; else_exp] ->
    let else_label = Util.gensym "else" in
    let continue_label = Util.gensym "continue" in
    compile_exp test_exp
    @ [Cmp (Reg Rax, operand_of_bool false); Jz else_label]
    @ compile_exp then_exp @ [Jmp continue_label]
    @ [Label else_label]
    @ compile_exp else_exp @ [Label continue_label]

Looking ahead

Today we introduced two new concepts in x86-64 machine code: flags and jumps. Next time we'll implement binary operations, for which we'll need one more concept: memory. After that, though, we really won't be further complicating our model of how the processor executes. We'll need a few more instructions here and there, but there won't be any more big ideas at the assembly level. This will be a blessing and a curse: the way the processor executes is relatively simple and easy to understand, which means that compiling high-level language constructs like functions is pretty challenging! It's going to be fun.