CS164 WA4

CS164 Programming Languages and Compilers Fall 2024
Assigned: October 18
Due: October 31 at 11:59pm
Written Assignment 4
Instructions: This assignment asks you to prepare written answers to questions on code generation. Each of the questions has a short answer. You may discuss this assignment with other students and work on the problems together. However, your write-up should be your own individual work.
Please write your name, email address, and discussion section on your homework. Please start each question on a new page. All written assignments must be submitted as a PDF via Gradescope: https: // gradescope. com . Instructions for how to submit assignments to Gradescope can be found at the following links: https://gradescope.com/get_started#student-submission

1. Suppose f is a function with a call to g somewhere in the body of f: def f(…):
… g(…) …
We say that this particular call to g is a tail call if the call is the last thing f does before returning. For
example, consider the following ChocoPy functions for computing positive powers of 2:
def f(x : int, acc : int) -> int:
return f(x-1, acc*2)
return acc
def g(x : int) -> int:
return 2 * g(x-1)
Here f(x, 1) = g(x) = 2x for x ≥ 0. The recursive call to f is a tail call, while the recursive call to g is not. A function in which all recursive calls are tail calls is called tail recursive.
(a) Here is a non-tail recursive function for computing the number of odd digits of a non-negative integer.
def num_odd(n : int) -> int:
if n < 10: if n % 2 == 1: if n % 2 == 1: return num_odd(n // 10) + 1 return num_odd(n // 10) + 0 Write a tail recursive function num_odd2 that computes the same result. (Hint: Your function will most likely need two arguments, or it may need to invoke a function of two arguments.) Computer Science Tutoring
(b) Recall from lecture that function calls are usually implemented using a stack of activation records.
i. Trace the execution of num_odd and num_odd2 computing the result for 123456, writing out the stack of activation records at each step i.e., draw the stack of activation records. You don’t need to draw everything in each activation frame. Just label each activation frame with the function name and argument(s) for that frame. An example is given below, where foo(3) calls foo(2) which calls foo(1).
AR for foo(1) AR for foo(2) AR for foo(3)
ii. Explain the computation done before each activation record is removed for both num_odd and num_odd2.
iii. Is there any place where you can see potential for making the execution of the tail-recursive num_odd2 more time-efficient or space-efficient than num_odd (without changing num_odd2’s source code)? What could you do?

2. Consider the following ChocoPy classes:
class A(object):
x : int = 0
y : int = 1
def getX(self:”A”) -> int:
return self.x
def getY(self:”A”) -> int:
return self.y
class B(A):
z : int = 2
def getX(self:”B”) -> int:
return self.x + 1
def getZ(self:”B”) -> int:
return self.z
class C(B):
s : str = “c”
def getY(self:”C”) -> int:
return self.y – 1
def toString(self:”C”) -> str:
return self.s
(a) Draw a diagram that illustrates the layout of objects of type A, B and C, including their dispatch tables. You can assume the existence of labels that point to the code containing method definitions (e.g. label A.getX for the method getX defined in class A).
(b) Let obj be a variable whose static type is A. Assume that the contents of obj have been loaded into register a0. The contents may be an address of an object in memory or the address 0, which represents the None value. Write RISC-V code for the method invocation obj.getY(). You may assume the existence of a label error dispatch None that contains logic for aborting the program with an error due to a dispatch on a None value (you just need to jump to the label appropriately). You may use temporary registers such as t1 if you wish. Use ChocoPy’s calling convention (the caller should push and pop arguments on the stack). As an example of this convention, here is the RISC-V code for the function invocation g(1) assuming there already exists the label g that contains callee code for the corresponding function:
# load argument 1
# push argument on stack
# jump to function
# pop argument on stack
(c) Explain what happens in part (b) if obj references an object that has dynamic type B.
(d) Explain what happens in part (b) if obj references an object that has dynamic type C.
Programming Help
3. (a) Consider the following ChocoPy program with nested function definitions:
def exp(x: int, y: int) -> int:
a: int = 1
def f(i: int) -> int:
nonlocal a
def geta() -> int:
if i <= 0: return geta() else: a=a*x return f(i-1) return f(y) This language feature causes some complications in code generation because nested functions may need to use variables defined in enclosing functions/methods. Consider the above example. The function f not only needs access to its own activation record (for variable i) but also the activation record for exp (for variables x and a). One way to implement this feature is to use a different type of activation record for nested functions (as opposed to the activation record used for global functions and methods). This new activation record contains an extra entry known as a static link, that is passed as an extra argument when calling a nested function. The static link is a pointer to the activation record of the latest dynamic instance of the nearest statically enclosing function/method. The first two activation records for a method call to exp(2,2) are given below (one for exp(2,2) and one for the function call to f(2)). In the diagram, we have noted with an arrow that the static link for f(2) points to the word below the return address in the activation record for exp(2,2). Complete the stack of activation records at the time of the call to geta() for exp(2,2) (i.e., having three calls to f). Include the activation record for geta(). Then, draw arrows to show where the static links point. (b) Generate RISC-V code to store the result of the assignment a = a * x in function f. Assume that the code for the multiplication has already been generated with the result in a0. (Hint: refer to the diagram given on the previous page. You can use temporaries to generate code.) ... # code to compute a * x leaving the result in a0 (c) Complete the generated RISC-V code for geta(). Return the result in register a0. The function prologue and epilogue has been provided for you. (Hint: use your completion of the above diagram.) exp.f.geta: addi sp, sp, -8 sw fp, 0(sp) sw ra, 4(sp) addi fp, sp, 8 lw ra, -4(fp) lw fp, -8(fp) addi sp, sp, 8 (d) Complete the generated RISC-V code for the call to geta() from f. exp_f: jal exp.f.geta (e) Complete the generated RISC-V code for the call to f(i - 1) from f. Assume that the code for the subtraction is generated in the printed line {code to compute i - 1 in a0}, after which the result of subtraction will be present in register a0. exp.f: ... {code to compute i - 1 in a0} jal exp.f ... 4. Consider the following function defined in the small language from our Simple Code Generation lectures (suitably extended with multiplication, division, and unary negation operations): def func(a, b, c): return sqrt(a*b) / (a+b) * -c We want to produce RISC-V assembly code for this solver function which uses strictly fp-relative accesses to and from the stack when evaluating the body of the function. That is, the assembly code of the function will use sp only on entry and when calling other functions (such as sqrt), and all temporary values will be stored in the stack and read from the stack at fixed offsets from fp. For this question, please use the calling convention for this language as specified in lecture. (a) Give a definition for the function cgen(f(e1, . . . , en), nt). This function should generate code which evaluates the expression f(e_1, ..., e_n) while only using temporaries whose addresses are fp - nt or lower. Ensure that the generated code writes the current frame pointer and the arguments for f to the stack in such a way that f can properly read its arguments and correctly restore the frame pointer as it returns. Note that this generated code must properly set sp just before calling f, i.e. just before the jal instruction. No other accesses to sp are necessary or allowed. (Hint: modify the function cgen(f(e1, . . . , en)) from lecture slides as appropriate.) (b) Fill in RISC-V code for the function func so that your code uses a fixed (fp-relative) location in the stack for each stored temporary. The stack pointer should not be accessed anywhere in your assembly code except to set the stack pointer just before calling sqrt with the jal instruction, as mentioned in part (a). In particular, you may not use the macros push reg, pop, or ra <- top, since these macros all access the sp register. You will need to use the following RISC-V instructions: • mul r1, r2, r3 multiplies registers r2 and r3 and stores the result in r1. • div r1, r2, r3 divides the register r2 by r3 and stores the result in r1. Don’t worry about division by zero. • sub r1, x0, r2 may be used to compute −r2 and store the result in register r1. (Recall that x0 is the always-zero register.) (Hint: use the cgen(e,nt) function from lecture, along with your implementation for cgen(f (e1 , . . . , en ), nt). You may need to determine how to implement cgen(e1 ∗ e2 , nt), cgen(e1/e2, nt), and cgen(−e, nt) as well.) Fill in your code between the body and exit comments under func. (You may need more room than is given here.) sqrt_entry: sw ra, 0(fp) # note that even this uses fp-relative addressing! # body: reads argument at 4(fp), and places result in a0 lw ra, 0(fp) lw fp, 8(fp) sw ra, 0(fp) lw ra, 0(fp) lw fp, 16(fp) 程序代写 CS代考 加QQ: 749389476