Booleans II
Recall that we’re adding support for booleans to our interpreter and compiler. Booleans I covered adding them to our interpreter. Now we’ll turn to the compiler. Specifically, we’re going to add support for these expressions:
-
true
andfalse
, the two boolean values -
(not e)
, a unary operation which evaluates totrue
on the boolean valuefalse
andfalse
otherwise -
(num? e)
, a unary opertion which evaluates totrue
ife
is a number andfalse
otherwise -
(zero? e)
, a unary opertion which evaluates totrue
ife
is the number0
andfalse
otherwise
Types in the compiler
Now that we have an interpreter to test against, we can extend our compiler to support our new operations!
When our interpreter is executing a program, values of expressions
are instances of the value
datatype we just defined. We
won’t be able to do that in the compiler–we can’t
define new datatypes in x86-64! Remember that when our program is
executing, its values live in registers (actually, just
rax
). Registers store 64-bit integers. Right now the
values in our program are all integers, so this works fine. But how
will we add booleans? Take a second and think about how you might
implement this.
Well, we know that all of our values need to be represented, at runtime, as 64-bit integers. So instead of representing integers as themselves:
0 -> 0b00 1 -> 0b01 2 -> 0b10 3 -> 0b11 ...
We’re going to represent the integer x
as
x << 2
(shifted left by two bits):
0 -> 0b0000 1 -> 0b0100 2 -> 0b1000 3 -> 0b1100
This is exactly equivalent to representing each integer
x
as x * 4
.
This means our integers have to fit in 62 bits instead of 64. So our
minimum integer is now -(2**61)
and our maximum integer
is (2**61) - 1
.
This also means there are a bunch of 64-bit integers (how many?)
that are no longer being used to represent values! All of our
integer values now end with
00
. So anything that ends with a different pair of bits
won’t be used to represent a number. This means we can use
some of them to represent booleans, and other types!
First, though, let’s update our compiler to use this new
representation for integers. Integer constants will be
easy–we’ll just shift them left. How will we handle
add1
and sub1
? Well, remember that our
runtime representations are the values multiplied by 4. Since
multiplication distributes over addition (and subtraction), we can
just add (or subtract) 4 instead of 1! So:
let num_shift = 2 let num_mask = 0b11 let num_tag = 0b00 let rec compile_exp (exp : s_exp) : directive list = match exp with | Num n -> [Mov (Reg Rax, Imm (n lsl num_shift))] | Lst [Sym "add1"; arg] -> compile_exp arg @ [Add (Reg Rax, Imm (1 lsl num_shift))] | Lst [Sym "sub1"; arg] -> compile_exp arg @ [Sub (Reg Rax, Imm (1 lsl num_shift))] | e -> raise (BadExpression e)
(lsl
is “logical shift left.” We could also
just multiply by 4, but it’s clearer this way.)
What happens if we run a program now?
>>> compile_and_run "(add1 4)" 20
This makes sense–we’re printing out the runtime representation! We’ll need to fix that. We’ll edit our C runtime:
#include <stdio.h> #include <inttypes.h> #define num_shift 2 #define num_mask 0b11 #define num_tag 0b00 extern uint64_t entry(); void print_value(uint64_t value) { if ((value & num_mask) == num_tag) { int64_t ivalue = (int64_t)value; printf("%" PRIi64, ivalue >> num_shift); } else { printf("BAD VALUE %" PRIu64, value); } } int main(int argc, char **argv) { print_value(entry()); return 0; }
Boolean support in the runtime
While we’re editing the runtime, let’s also add support for booleans.
#include <stdio.h> #include <inttypes.h> #define num_shift 2 #define num_mask 0b11 #define num_tag 0b00 #define bool_shift 7 #define bool_mask 0b1111111 #define bool_tag 0b0011111 extern uint64_t entry(); void print_value(uint64_t value) { if ((value & num_mask) == num_tag) { int64_t ivalue = (int64_t)value; printf("%" PRIi64, ivalue >> num_shift); } else if ((value & bool_mask) == bool_tag) { if (value >> bool_shift) { printf("true"); } else { printf("false"); } } else { printf("BAD VALUE %" PRIu64, value); } } int main(int argc, char **argv) { print_value(entry()); return 0; }
We’ll need to recompile the runtime:
$ gcc -c runtime.c -o runtime.o
Boolean support in the compiler
We can now add support for true
and
false
pretty easily:
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))]
Handling our other operations will be a little trickier. 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 code1. 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 register2
and sets the last byte of that register to 1 (i.e.,
0b00000001
) if ZF
is set and 0 if
ZF
is not set.
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
Footnotes:
Actually, all of the flags are packed together in the same special RFLAGS register
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.