Princeton University
COS 217: Introduction to Programming Systems

Assignment 4: Assembly Language Programming and Testing


Purpose

The purpose of this assignment is to help you learn about computer architecture, assembly language programming, and testing strategies. It also will give you the opportunity to learn more about the GNU/Unix programming tools, especially bash, emacs, gcc, gdb, and gprof.

The assignment consists of two parts, each of which has subparts. We encourage you to complete Part 1 by the end of the first week of the assignment period.


Rules

Part 2f, as defined below, is the "extra challenge" part of this assignment. While doing the "extra challenge" part of the assignment, you are bound to observe the course policies regarding assignment conduct as given in the course Policies web page, plus one additional policy: you may not use any "human" sources of information. That is, you may not consult with the course's staff members, the lab teaching assistants, other current students via Piazza, or any other people while working on the "extra challenge" part of an assignment, except for clarification of requirements.

The extra challenge part is worth 12 percent of this assignment. So if you don't do any of the "extra challenge" part and all other parts of your assignment solution are perfect and submitted on time, then your grade for the assignment will be 88 percent.


Part 1: A Word Counting Program in Assembly Language

Part 1a: Translate to Assembly Language

The Unix operating system has a command named wc (word count). In its simplest form, wc reads characters from stdin until end-of-file, and writes to stdout a count of how many lines, words, and characters it has read. A word is a sequence of characters that is delimited by one or more white space characters.

Consider some examples. In the following, a space is shown as s and a newline character as n.

If the file named proverb contains these characters:

Learningsissan
treasureswhichn
accompaniessitsn
ownerseverywhere.n
--sChinesesproverbn

then the command:

$ wc < proverb

writes this line to stdout:

  5 12 82

If the file proverb2 contains these characters:

Learningsissan
treasureswhichn
accompaniessitsn
ownerseverywhere.n
--sssChinesesproverb

(note that the last "line" does not end with a newline character) then the command:

$ wc < proverb2

writes this line to stdout:

  4 12 83

The file mywc.c in the /u/cos217/Assignment4 directory contains a C program that implements the subset of the wc command described above. Translate that program into assembly language, thus creating a file named mywc.s. It is acceptable to use global (i.e. bss section or data section) variables in mywc.s, but we encourage you to use local (i.e. stack) variables instead. Your assembly language program must have exactly the same behavior (i.e. must write exactly the same characters to stdout) as the given C program.

Part 1b: Test

Design a test plan for your mywc program. Your test plan must include tests in three categories: (1) boundary testing, (2) statement testing, and (3) stress testing.

Create text files to test your programs. Name each such file such that its prefix is mywc and its suffix is .txt. The command ls mywc*.txt must display the names of all mywc test files, and only those files.

Describe your mywc test plan in your readme file. Your description must have this structure:

You may submit as many test files as you want. However at most three of your test files may be large, and a large test file must contain no more than 50000 characters and no more than 1000 lines. It would be difficult for your grader to scroll through a test file that exceeds those limits.

The nobel /u/cos217/Assignment4 directory contains bash shell scripts named testmywc and testmywcdiff that automate your testing. Comments at the beginning of those files describe how to use them. After copying the scripts to your project directory, you may need to execute the commands chmod 700 testmywc and chmod 700 testmywcdiff to give them "executable" permissions.


Part 2: Beat the Compiler

Background

Many programming environments contain modules to handle high-precision integer arithmetic. For example, the Java Development Kit (JDK) contains the BigDecimal and BigInteger classes.

The Fibonacci numbers are used often in computer science. See http://en.wikipedia.org/wiki/Fibonacci_numbers for some background information. Note in particular that Fibonacci numbers can be very large integers.

This part of the assignment asks you to use assembly language to compose a minimal but fast high-precision integer arithmetic module, and use it to compute large Fibonacci numbers.

Part 2a: Add BigInt Objects Using C Code

Suppose you must compute Fibonacci number 500000, that is, fib(500000)...

The /u/cos217/Assignment4 directory contains a C program that computes Fibonacci numbers. It consists of two modules: a client and a BigInt ADT.

The client consists of the file fib.c. The client accepts an integer x as a command-line argument, where x must be a non-negative integer. The client performs some boundary condition and stress tests, writing the results to stdout. Then it computes and writes fib(x) to stdout as a hexadecimal number. It writes to stderr the amount of CPU time consumed while performing the computation. The client module delegates most of the work to BigInt objects.

The BigInt ADT performs high precision integer arithmetic. It is a minimal ADT; essentially it implements only an "add" operation. The BigInt ADT consists of four files:

Study the given code. Then build a fib program consisting of the files fib.c, bigint.c, and bigintadd.c, with no optimizations. Run the program to compute fib(500000). In your readme file note the amount of CPU time consumed.

Part 2b: Add BigInt Objects Using C Code Built with Compiler Optimization

Suppose you decide that the amount of CPU time consumed is unacceptably large. You decide to command the compiler to optimize the code that it produces...

Build the fib program using optimization. Specifically, build with the -D NDEBUG option so the preprocessor disables the assert macro, and the -O3 (that's an upper case "oh" followed by "3") option so the compiler generates optimized code. Run the resulting program to compute fib(500000). In your readme file note the amount of CPU time consumed.

Part 2c: Profile the Code

Suppose you decide that the amount of CPU time consumed still is too large. You decide to investigate by doing a gprof analysis to determine which functions are consuming the most time...

Perform a gprof analysis of the code from Part 2b. Save the textual report in a file named gprofreport. Don't delete the file; as described later in this document, you must submit that file.

Part 2d: Add BigInt Objects Using Assembly Language Code

Suppose, not surprisingly, your gprof analysis shows that most CPU time is spent executing the BigInt_add function. In an attempt to gain speed, you decide to code the BigInt_add function manually in assembly language...

Manually translate the C code in the bigintadd.c file into assembly language, thus creating the file bigintadd.s. You should not translate the code in other files into assembly language.

Your assembly language code must store all local variables defined in the BigInt_larger and BigInt_add functions in memory, on the stack.

Note that assert is a parameterized macro, not a function. (See Section 14.3 of the King book for a description of parameterized macros.) So assembly language code cannot call assert. When translating bigintadd.c to assembly language, simply pretend that the calls of assert are not in the C code.

Build a fib program consisting of the files fib.c, bigint.c, and bigintadd.s using the -D NDEBUG and -O3 options. Run the program to compute fib(x) for various values of x, and make sure it writes the same output to stdout as the program built from C code does. Finally, run the program to compute fib(500000). In your readme file note the amount of CPU time consumed.

Part 2e: Add BigInt Objects Using Optimized Assembly Language Code

Suppose, to your horror, you discover that you have taken a step backward: the CPU time consumed by your assembly language code is approximately (within a few seconds) the same as that of the non-optimized compiler-generated code! So you decide to optimize your assembly language code...

Manually optimize your assembly language code in bigintadd.s, thus creating the file bigintaddopt.s. Specifically, perform these optimizations:

  1. Store local variables (but not parameters) in registers instead of in memory. Remember that EBX, ESI, and EDI are callee-save registers, and that EAX, ECX, and EDX are caller-save registers.

  2. "Inline" the call of the BigInt_larger function. That is, eliminate the BigInt_larger function, placing its code within the BigInt_add function.

Build a fib program consisting of the files fib.c, bigint.c, and bigintaddopt.s using the -D NDEBUG and -O3 options. Run the program to compute fib(x) for various values of x, and make sure it writes the same output to stdout as the program built from C code does. Finally, run the program to compute fib(500000). In your readme file note the amount of CPU time consumed.

Can you write assembly language code that is approximately as fast as the optimized code that the compiler generates? That is, can you approximately tie the compiler?

Part 2f: Add BigInt Objects Using Highly Optimized Assembly Language Code

Finally, suppose you decide to optimize your assembly language code even further, moving away from a statement-by-statement translation of C code into assembly language...

Further optimize your assembly language code in bigintaddopt.s, thus creating the file bigintaddoptopt.s. Specifically, perform these optimizations:

  1. Factor as much code as possible out of the loop.

  2. Use the assembly language loop patterns described in Section 3.6.5 of the Bryant & O'Hallaron book instead of the simpler but less efficient patterns described in precepts.

  3. Use the adcl ("add with carry long") instruction effectively. The adcl instruction computes the sum of its source operand, its destination operand, and the "carry" flag from the EFLAGS register, places the sum in the destination operand, and assigns 1 (or 0) to the "carry" flag if a carry occurred (or did not occur) during the addition. Effective use of the adcl instruction will use the "carry" flag in the EFLAGS register instead of a uiCarry variable to keep track of carries during addition.

Feel free to implement any additional optimizations you want. However, your BigInt_add function must be a general-purpose function for adding two given BigInt objects; the function cannot be specific to the task of adding two Fibonacci numbers to generate a third Fibonacci number. In other words, your function must work with the given fib.c client.

This part is difficult; we will not think unkindly of you if you decide not to do it. To do it properly you will need to learn about the adcl instruction, and about which instructions affect and do not affect the "carry" flag in the EFLAGS register.

Hint: When writing bigintaddoptopt.s, the problem is this: How can you preserve the value of the "carry" flag between executions of the adcl instruction? One solution is to save and then restore the value of the EFLAGS register. Another solution is to express the logic such that only instructions that do not affect the "carry" flag are executed between each execution of the adcl instruction.

Build a fib program consisting of the files fib.c, bigint.c, and bigintaddoptopt.s using the -D NDEBUG and -O3 options. Run the program to compute fib(x) for various values of x, and make sure it writes the same output to stdout as the program built from C code does. Finally, run the program to compute fib(500000). In your readme file note the amount of CPU time consumed.

Can you beat the compiler?


Logistics

Develop on nobel. Use emacs to create source code. Use gdb to debug.

Do not use a C compiler to produce any of your assembly language code. Doing so would be considered academically dishonest. Instead produce your assembly language code manually.

We encourage you to develop "flattened" C code (as described in lectures and precepts) to bridge the gap between the given "normal" C code and your assembly language code. Using flattened C code as a bridge can eliminate logic errors from your assembly language code, leaving only the possibility of translation errors.

We also encourage you to use your flattened C code as comments in your assembly language code. Such comments can clarify your assembly language code substantially.

Create a readme file by copying the file /u/cos217/Assignment4/readme to your project directory, and editing the copy by replacing each area marked "<Complete this section.>" as appropriate.

One of the sections of the readme file requires you to list the authorized sources of information that you used to complete the assignment. Another section requires you to list the unauthorized sources of information that you used to complete the assignment. Your grader will not grade your submission unless you have completed those sections. To complete the "authorized sources" section of your readme file, copy the list of authorized sources given in the "Policies" web page to that section, and edit it as appropriate.

Submit your work electronically on nobel using these commands:

submit 4 mywc.s
submit 4 mywc*.txt
submit 4 gprofreport
submit 4 bigintadd.s
submit 4 bigintaddopt.s
submit 4 bigintaddoptopt.s
submit 4 readme

Grading

Minimal requirement to receive credit for the "Translate to Assembly Language" (Part 1a) implementation:

Minimal requirement to receive credit for the "Add BigInt Objects Using Assembly Language Code" (Part 2d) implementation:

Minimal requirement to receive credit for the "Add BigInt Objects Using Optimized Assembly Language Code" (Part 2e) implementation:

Minimal requirement to receive credit for the "Add BigInt Objects Using Highly Optimized Assembly Language Code" (Part 2f) implementation:

We will grade your code on quality from the user's and programmer's points of view. From the user's point of view, your code has quality if it behaves as it should. The correct behavior of your code is defined by the previous sections of this assignment specification.

From the programmer's point of view, your code has quality if it is well-styled and thereby easy to maintain. Comments in your assembly language code are especially important. Each assembly language function -- especially the main function -- must have a comment that describes what the function does. Local comments within your assembly language functions are equally important. Comments copied from corresponding flattened C code are particularly helpful.

Your assembly language code must use .equ directives to avoid "magic numbers." In particular, you must use .equ directives to give meaningful names to:

Also, your bigintaddopt.s file must use .equ directives to give meaningful names to registers that store local variables, for example, .equ REG_UICARRY, %edx. Those names must begin with REG_. (Otherwise they appear to be ordinary labels.)

To encourage good coding practices, we will deduct points if gcc217 generates warning messages.

Testing is a substantial aspect of the assignment. Approximately 10% of the grade will be based upon your mywc test plan as described in your readme file and as implemented by your data files.

Your grade for Part 2f will be based upon:


This assignment was written by Robert M. Dondero, Jr.,
based in part upon suggestions from Andrew Appel,
and with contributions by other faculty members.