#1: Course Overview

Sankha Narayan Guria

EECS 700: Introduction to Program Synthesis

1
Instructor
Sankha Narayan Guria

• Assistant Professor starting this Fall

• Previously: PhD @ University of
Maryland
• Research Areas: Program synthesis and
program analysis

2
Logistics
• Lectures:
• Mon Wed Fri: 3:00 - 3:50pm LEEP2 G411 But you already knew that!

• Office Hours:
• Wed: 4:00 - 5:00pm Eaton 2034

• Course Website:
• https://sankhs.com/eecs700/
• Communications through Canvas and Discord

3
 Objectives
1. Understand what program • Lectures
synthesis can do and how • Read and discuss research papers

2. Use existing synthesis

tools

3. Contribute to synthesis • Project

techniques and tools
towards a publication in an
academic conference

4
Evaluation
• Class Participation
• Ask/answer questions in class
• Participate in discussions on Discord
• Paper Reviews
• 9 papers, 5% each
• Final Project
• Team formed by deadline: 5%
• 1 page project proposal: 15%
• Project presentation: 15%
• Final report: 15%

5
Paper Reviews
• Due on Thursday, by end of day
• First review is due next week!
• Will be posted on website a week before due date
• Reviews submitted via Canvas
• Link on website
• Website has details of review guidelines
• Discussion:
• Before review is due: discuss on Discord
• After review is due: discuss in class (Friday)

6
Project
• Kinds of projects:
• Re-implement techniques from a paper
• Apply existing synthesis framework to a new domain
• Extend/improve existing synthesis algorithm or tool
• Develop a new synthesis algorithm or tool
• …
• Judged in terms of
• Quality of execution
• Originality
• Scope

7
Project
Team forming • Teams of 2-3
• Pick a project:
• List of suggested projects coming soon on
website
• Please talk to me!
• One page: explain what you plan to do
Proposal and give some evidence that you’ve
started to work on it
• Presentations in last few classes
Presentation • ~10-15 min per project
• 3-8 pages, structured like a research
Report paper
8
Now to the good stuff …

9
The goal: automate programming
What is program synthesis?
 append:
push ebp
mov ebp, esp
push eax
push ebx
push len
“Any
call malloc
s ufficie
mov ebx, [ebp + 12]
ntly a
d
mov [eax + info], ebx
mov dword [eax + next], 0 vance
from d compiler
mov ebx, [ebp + 8]
cmp dword [ebx], 0

a synt
hesize is indisting
je null_pointer void insert(node *xs, int x) {
mov ebx, [ebx] node *new;
node *temp;
r” ui s ha
b
next_element:
le
node *prev;
cmp dword [ebx + next], 0
je found_last new = (node *)malloc(sizeof(node));
mov ebx, [ebx + next] if(new == NULL) {
jmp next_element printf("Insufficient memory.");
return;
found_last: }
push eax new->val = x;
push addMes new->next = NULL;
call puts if (xs == NULL) {
add esp, 4 xs = new;
pop eax } else if(x < xs->val) {
mov [ebx + next], eax new->next = xs;
xs = new;
go_out: } else {
pop ebx prev = xs;
pop eax temp = xs->next;
mov esp, ebp while(temp != NULL && x > temp->val) {
pop ebp prev = temp;
ret 8 temp = temp->next;
}
null_pointer: if(temp == NULL) {

?
push eax prev->next = new;
push nullMes } else {
call puts new->next = temp;
add esp, 4 prev->next = new; insert x [] = [x]
pop eax } insert x (y:ys)
mov [ebx], eax } | x ≤ y = x:y:ys
jmp go_out } | otherwise = y:(insert x ys)

Assembly C Haskell modern program

synthesis
Modern program synthesis: FlashFill

[Gulwani 2011]
FlashFill: a feature of Excel 2013
FlashFill: a feature of Excel 2013
Modern program synthesis: Sketch
• Problem: isolate the least significant zero bit in a word
• example: 0010 0101 à 0000 0010
• Easy to implement with a loop
int W = 32;
bit[W] isolate0 (bit[W] x) { // W: word size
bit[W] ret = 0;
for (int i = 0; i < W; i++)
if (!x[i]) { ret[i] = 1; return ret; }
}

• Can this be done more efficiently with bit manipulation?

• Trick: adding 1 to a string of ones turns the next zero to a 1
• i.e. 000111 + 1 = 001000

[Solar-Lezama 2013]
Sketch: space of possible
implementations
/**
* Generate the set of all bit-vector expressions
* involving +, &, xor and bitwise negation (~).
*/

generator bit[W] gen(bit[W] x){

if(??) return x;
if(??) return ??;
if(??) return ~gen(x);
if(??){
return {| gen(x) (+ | & | ^) gen(x) |};
}
}
Sketch: synthesis goal
generator bit[W] gen(bit[W] x, int depth){
assert depth > 0;
if(??) return x;
if(??) return ??;
if(??) return ~gen(x, depth-1);
if(??){
return {| gen(x, depth-1) (+ | & | ^) gen(x, depth-1) |};
}
}

bit[W] isolate0fast (bit[W] x) implements isolate0 {

return gen(x, 3);
}
Sketch: output

bit[W] isolate0fast (bit[W] x) {

return (~x) & (x + 1);
}
~0010 0101 0010 0101 + 1
= 1101 1010 = 0010 0110

&
0000 0010
Modern program synthesis: Synquid
• Problem: intersection of sets represented as strictly sorted
lists
• example: intersect [4, 8, 15, 16, 23, 42] [8, 16, 32, 64] à [8, 16]
• Also: we want a guarantee that it’s correct on all inputs!

[Polikarpova et al. 2016]

Synquid: synthesis goal and
components
• Step 1: define synthesis goal as a type sorted
list

intersect
intersect ::
:: xs:List
xs:SLista a→ ys:List a
→ ys:SList a
v:SList
→→ {List a a | elems v = elems xs ∩ elems ys}
the set of
elements

• Step 2: define a set of components

• Which primitive operations is our function likely to use?
• Here: {Nil, Cons, <}
Synquid: output
xs ys result
intersection = \xs . \ys .
match xs with [4, 8, 15, 16, 23, 42] [8, 16, 32, 64]
Nil -> xs [8, 15, 16, 23, 42] [8, 16, 32, 64] [8]
Cons x xt ->
match ys with [15, 16, 23, 42] [16, 32, 64]
Nil -> ys [16, 23, 42] [16, 32, 64] [8, 16]
Cons y yt ->
if x < y [23, 42] [32, 64]
then intersection xt ys
[42] [32, 64]
else
if y < x [42] [64]
then intersection xs yt
[] [64]
else Cons x (intersection xt yt)
 Modern program synthesis: GitHub
Copilot
// find all images
// and add a green border around them input
// and add class “githubCopilot” to them
function go() {

• var images = document.getElementByTagName(‘img’);

• for (var i = 0; i < images.length; i++) {
• if (images[i].className.indexOf(‘githubCopilot’) == -1) { output
• images[i].className += ‘ githubCopilot’;
• images[i].style.border = ‘1px solid green’;
• }
• }
•
What is program synthesis?

specification search program

program
space

24
Dimensions in program synthesis
Behavioral spec:
what should the program should do?

Search strategy:
How does the system find Structural spec:
the program you want? what is the space of programs to
explore?

[Gulwani 2010]
Behavioral spec
• How do you tell the system what the program
should do?
• What is the input language / format?
• What is the interaction model?
• What happens when the intent is ambiguous?

• Q: What did the behavioral spec look like in

FlashFill / Sketch / Synquid / Copilot?
Behavioral spec: examples
• Input/output examples
• Reference implementation
• Formal specifications (pre/post
conditions, types, ...)
• Natural language
• Context
Structural spec
• What is the space of programs to explore?
• Large enough to contain interesting programs, yet
small enough to exclude garbage and enable
efficient search
• Built-in or user defined?
• Can we extract domain knowledge from existing
code?

• Q: What did the structural spec look like in

FlashFill / Sketch / Synquid / Copilot?
Structural spec: examples
• Built-in DSL
• User-defined DSL (grammar)
• User-provided components
• Languages with synthesis constructs
• e.g. generators in Sketch
• Learned language model
Search strategies
• Synthesis is search:
• Find a program in the space defined by structural
constraints that satisfies behavioral constraints
• Challenge: the space is astronomically large
• The search algorithm is the heart of a synthesis
technique
• How does the system find the program you want?
• How does it know it’s the program you want?
• How can it leverage structural constraints to guide the
search?
• How can it leverage behavioral constraints to guide the
search?
Search strategies: examples
• Enumerative (explicit) search
• exhaustively enumerate all programs in the language in
the order of increasing size
• Stochastic search
• random exploration of the search space guided by a
fitness function
• Representation-based search
• use a data structure to represent a large set of
programs
• Constraint-based search
• translate to constraints and use a solver