Cobra

In this assignment you'll implement a small language called Cobra, which implementes an encoded binary representation of different values. It also uses C function calls to implement some user-facing operations, like printing and reporting errors.

The Cobra Language

As usual, there are a few pieces that go into defining a language for us to compile.

• A description of the concrete syntax – the text the programmer writes

• A description of the abstract syntax – how to express what the programmer wrote in a data structure our compiler uses. As in Boa, this will include a surface expr type, and a core aexpr type.

• The semantics—or description of the behavior—of the abstrac syntax, so our compiler knows what the code it generates should do.

Concrete Syntax

The concrete syntax of Boa is:

<expr> :=
  | let <bindings> in <expr>
  | if <expr>: <expr> else: <expr>
  | <binop-expr>

<binop-expr> :=
  | <identifier>
  | <number>
  | true
  | false
  | add1(<expr>)
  | sub1(<expr>)
  | isnum(<expr>)
  | isbool(<expr>)
  | print(<expr>)
  | <expr> + <expr>
  | <expr> - <expr>
  | <expr> * <expr>
  | <expr> < <expr>
  | <expr> > <expr>
  | <expr> == <expr>
  | ( <expr> )

<bindings> :=
  | <identifier> = <expr>
  | <identifier> = <expr>, <bindings>
}

Abstract Syntax

User-facing

The abstract syntax of Cobra is an OCaml datatype, and corresponds nearly one-to-one with the concrete syntax. Here, we've added E prefixes to the constructors, which will distinguish them from the ANF forms later.

type prim1 =
  | Add1
  | Sub1
  | Print
  | IsNum
  | IsBool

type prim2 =
  | Plus
  | Minus
  | Times
  | Less
  | Greater
  | Equal

type expr =
  | ELet of (string * expr) list * expr
  | EPrim1 of prim1 * expr
  | EPrim2 of prim2 * expr * expr
  | EIf of expr * expr * expr
  | ENumber of int
  | EBool of bool
  | EId of string

Compiler-facing

The compiler-facing abstract syntax of Boa splits the above expressions into three categories

type immexpr =
  | ImmNumber of int
  | ImmBool of bool
  | ImmId of string

and cexpr =
  | CPrim1 of prim1 * immexpr
  | CPrim2 of prim2 * immexpr * immexpr
  | CIf of immexpr * aexpr * aexpr
  | CImmExpr of immexpr

and aexpr =
  | ALet of string * cexpr * aexpr
  | ACExpr of cexpr

These are quite similar to the constructs for Boa – the main additions are the new primitives.

Semantics

With the addition of two types to the language, there are two main changes that ripple through the implementation:

• The representation of values
• The possibility of errors

There is one other major addition, which is the print primitive, discussed more below.

The representation of values requires a definition. We'll use the following representations for the Cobra runtime:

• true will be represented as the constant 0xFFFFFFFF
• false will be represented as the constant 0x7FFFFFFF
• numbers will be represented with a zero in the rightmost bit, as in class. So, for example, 2 is represented as 0x00000004.

You should augment the provided print function in main.c to print these values correctly: true and false should print as those words, and numbers should print out as the underlying number being represented.

You should raise errors in the following cases:

• -, +, *, <, and > should raise an error (by printing it out) with the substring "expected a number" if the operation doesn't get two numbers (you can print more than this if you like, but it must be a substring)
• add1 and sub1 should raise an error with the substring "expected a number" if the argument isn't a number
• +, -, and * should raise an error with the substring "overflow" if the result overflows, and falls outside the range representable in 31 bits. The jo instruction (not to be confused with the Joe Instructor) which jumps if the last instruction overflowed, is helpful here.
• if should raise an error with the substring "expected a boolean" if the conditional value is not a boolean.

These error messages should be printed on standard error, so use a call like:

fprintf(stderr, "Error: expected a number")

I recommend raising an error by adding some fixed code to the end of your generated code that calls into error functions you implement in main.c. For example, you might insert code like:

internal_error_non_number:
  push eax
  call error_non_number

Which will store the value in eax on the top of the stack, move esp appropriately, and perform a jump into error_non_number function, which you will write in main.c as a function of one argument.

The other operators, ==, isnum, isbool, and print, cannot raise errors, and always succeed.

The final piece of new semantics is the print operator. A print expression should pass the value of its subexpression to the print function in main.c, and evaluate to the same value (print in main.c helps out here by returning its argument). The main work you need to do here is similar to when calling an error function; evaluate the argument, push it onto the stack with push, and then call into print.

Examples

let x = 1 in
let y = print(x + 1) in
print(y + 2)

# will output

2
4
4

The first 2 comes from the first print expression. The first 4 comes from the second print expression. The final line prints the answer of the program as usual, so there's an “extra” 4.

if 54: true else: false

# prints (on standard error) something like:

Error: expected a boolean in if, got 54

Implementing Cobra

Memory Layout and Calling C Functions

In order to set up the stack properly to call C functions, like print and your error functions, it's necessary to make a few changes to what we had in Boa.

• Allocating stack space ahead of time: At the start of our generated code (which we now recognize is a function body participating in a C runtime stack), we need to make sure we make enough stack space for all variables we create, and reserve that space ahead of time. To do this, we move esp to point to a location that is N words away (so N * 4 bytes for us), where N is the greatest number of variables we need at once. This is actually tricky to compute to be fully optimal (teaser for later in the semester: by “tricky” I mean NP-hard), but it's easy to get an OK heuristic – we can compute the maximum depth of nested definitions.

  To do this, we need the count_vars function, which we implemented in class, so I've provided it. You need to add instructions to the provided spot in stack_setup in order to make sure the correct space is allocated on the stack by subtracting the right amount from esp.

• Using the Base Pointer: In addition, this means that all variable references need to happen from ebp rather than from esp. This is because esp can change while we are pushing arguments onto the stack for other function calls, so ebp is the place we should trust for consistent offsets.

• Participating in the C stack: As a C function callee (from main) and caller (of error functions and print), our code has some responsibilities. First, we need to store the old base pointer upon entry, and update the base pointer to hold the current top of the stack (which includes the return pointer into main, for example). This is why the typical top two lines of most C functions are:

  push ebp
  mov ebp, esp
  

  Similarly, when we're done with the function, we need to restore the stack pointer to its old location, and put the old base pointer value back. This is why the last lines before a ret in a C function are often:

  mov esp, ebp
  pop ebp
  

• Other Responsibilities: If we were using registers beyond eax, ebp, and esp, we'd be responsible for storing some of them as callee, and some as caller. But we're not going to go into those details for this assignment. Since we aren't using those registers, it has no effect on our code's behavior.

New Assembly Constructs

• Sized

  You may run into errors that report that the size of an operation is ambiguous. This could happen if you write, for example:

  cmp [ebp-8], 0
  

  Because the assembler doesn't know if the program should move a four-byte zero, a one-byte zero, or something in between into memory starting at [ebp-8]. To solve this, you can supply a size:

  cmp [ebp-8], DWORD 0
  

  This tells the assembler to use the “double word” size for 0, which corresponds to 32 bits. A WORD corresponds to 16 bits, and a BYTE corresponds to 16 bits. To get a sized argument, you can use the Sized constructor from arg.

• HexConst

  Sometimes it's nice to read things in hex notation as opposed to decimal constants. I've provided a new HexConst arg that's useful for this case.

• IPush, IPop

  These two instructions manage values on the stack. push adds a value at the current location of esp, and increments esp to point past the added value. pop decrements esp and moves the value at the location esp was pointing to into the provided arg.

• ICall

  A call does two things:

  - Pushes the next _code_ location onto the stack (just like a `push`),
    which becomes the return pointer
  - Performs an unconditional `jmp` to the provided label
  

  call does not affect ebp, which the program must maintain on its own.

• IShr, IShl: Bit shifting operations

• IAnd, IOr, IXor: Bit masking operations

• IJo, IJno: Jump to the provided label if the last arithmetic operation did/did not overflow

As usual, full summaries of the instructions we use are at this assembly guide.

Testing Functions

These are the same as they were for Boa. ANF is provided, and hasn't changed aside from the addition of new primitives and EBool. So your tests should focus on te and t tests.

An old friend is helpful here, too: valgrind. You can run valgrind output/some_test.run in order to get a little more feedback on tests that fail with -10 as their exit code (which usually indicates a segfault). This can sometimes tip you off quite well as to how memory is off, since sometimes you'll see code trying to jump to a constant that's in your code, or other obvious tells that there's something off in the stack. Also, if you've done all your stack management correctly, valgrind will report a clean run for your program!

Recommended TODO List

Here's an order in which you could consider tackling the implementation:

1. Fix the print function in main.c so that it prints out the right output. It will need to check for the tag using C bitwise operators, and use printf to print the right value.
2. Take a first shot at figuring out how to increase the stack appropriately by using count_vars and adding to stack_setup.
3. Fill in the CPrim1 case for everything but print, and figure out how to check for errors, and call the "non-number" error reporting function. Test as you go. Be aware that if the function call segfaults, it may be because you need to refine step 2.
4. Implement print by compiling it to with a call to the print after pushing appropriate arguments. Be aware that if the call doesn't work, it may be because of step 2 again. Test as you go; be aware that you should test interesting sequences of print expressions and let-bindings to make sure your stack integrity is good before and after calls.
5. Fill in all of the CPrim2 cases, using the error-reporting from the last step. Test as you go.
6. Complete the if case and test as you go.

Handing In

A complete implementation is due by Thursday, February 25 at 11:59pm.