A HUMAN STUDY OF PATCH MAINTAINABILITYZachary P. Fry, Bryan Landau, Westley Weimer

University of Virginia

{zpf5a,bal2ag,weimer}@virginia.edu

Bug Fixing

Fixing bugs manually is difficult and costly. Recent techniques explore automated

patches: Evolutionary techniques – GenProg Dynamic modification – ClearView Enforcement of pre/post-conditions – AutoFix-E Program transformation via static analysis – AFix

While these techniques save developers time, there is some concern as to whether the patches produced are human-understandable and maintainable in the long run.

2

3

Questions Moving Forward

How can we concretely measure these notions of human understandability and future maintainability?

Can we automatically augment machine-generated patches to improve maintainability?

In practice, are machine-generated patches as maintainable as human-generated patches?

4

Questions Moving Forward

How can we concretely measure these notions of human understandability and future maintainability?

Can we automatically augment machine-generated patches to improve maintainability?

In practice, are machine-generated patches as maintainable as human-generated patches?

5

Measuring quality and maintainability

Functional Quality – Does the implementation match the specification? Does the code execute “correctly”?

Non-functional Quality – Is the code understandable to humans? How difficult is it to understand and alter

the code in the future?

✓

?

6

Software Functional Quality

Perfect: Implementation matches specification

Direct software quality metrics: Testing Defect density Mean time to failure

Indirect software quality metrics: Cyclomatic complexity Coupling and cohesion (CK metrics) Software readability

7

Software Non-functional Quality

Maintainability: Human-centric factors affecting the ease

with which bugs can be fixed and features can be added

Broadly related to the “understandability” of code

Not easy to concretely measure with heuristics like functional correctness

These automatically-generated patches have been shown to be of high quality functionally – what about non-functionally?

8

Patch Maintainability Defined Rather than using an approximation

to measure understandability, we will directly measure humans’ abilities to perform maintenance tasks

Task: ask human participants questions that require them to read and understand a piece of code and measure the effort required to provide correct answers

Simulate the maintenance process as closely as possible

9

Php Bug #54454

Title: “substr_compare incorrectly reports equality in some cases”

Bug description: “if main_str is shorter than str, substr_compare

[mistakenly] checks only up to the length of main_str”

substr_compare(“cat”, “catapult”) = true

10

Motivating Example

if (offset >= s1_len) { php_error_docref(NULL TSRMLS_CC,

E_WARNING, "The start

position cannot exceed string length"); RETURN_FALSE;}

if (len > s1_len - offset) {len = s1_len - offset;

}

cmp_len = (uint) (len ? len : MAX(s2_len, (s1_len - offset)));

11

Motivating Example

len--; if (mode & 2) {

for (i = len - 1; i >= 0; i--) {

if (mask[(unsigned char)c[i]]) {

len--; }

else {break; }

}}

if (return_value) { RETVAL_STRINGL(c, len, 1); } else {

12

Automatic Documentation

Intuitions suggest that patches augmented with documentation are more maintainable

Human patches can contain comments with hints as to the developer’s intention when changing code Automatic approaches cannot easily reason about

why a change is made, but can describe what was changed

Automatically Synthesized Documentation: DeltaDoc (Buse et al. ASE 2010) Measures semantic program changes Outputs natural language descriptions of changes

13

Automatic Documentation

if (!con->conditional_is_valid[dc->comp]) {if (con->conf.log_condition_handling) {

TRACE("cond[%d] is valid: %d", dc->comp, con->conditional_is_valid[dc-

>comp]); } /* If not con->conditional_is_valid[dc->comp]

No longer return COND_RESULT_UNSET; */return COND_RESULT_UNSET;

} /* pass the rules */ switch (dc->comp) {

case COMP_HTTP_HOST: {char *ck_colon = NULL, *val_colon = NULL;

14

Questions Moving Forward

How can we concretely measure these notions of human understandability and future maintainability?

Can we automatically augment machine-generated patches to improve maintainability?

In practice, are machine-generated patches as maintainable as human-generated patches?

15

Evaluation

Focused research questions to answer: 1) How do different types of patches affect

maintainability? 2) Which source code characteristics are

predictive of our maintainability measurements? 3) Do participants’ intuitions about

maintainability and its causes agree with measured maintainability?

To answer these questions directly we performed a human study using over 150 participants with real patches from existing systems

16

Experiment - Subject Patches

Program LOC Defects Patches

gzip 491,083 1 2libtiff 77,258 7 14lighttpd 61,528 3 4php 1,046,421 9 17python 407,917 1 2wireshark 2,812,340 11 11

Total: 4,896,547 32 50

We used patches from six benchmarks over a variety subject domains

17

Experiment - Subject Patches Original – the defective, un-patched

code used as a baseline for measuring relative changes

Human-Accepted – human patches that have not been reverted to date

Human-Reverted – human-created patches that were later reverted

Machine – automatically-generated patches created by the GenProg tool

Machine+Doc – the same patches as above, but augmented with automatically synthesized documentation

18

Experiment – Maintenance Task

Sillito et al. – “Questions programmers ask during software evolution tasks”

Recorded and categorized the questions developers actually asked while performing real maintenance tasks

“What is the value of the variable “y” on line X?”

Not: “Does this type have any siblings in the type hierarchy?”

19

Human Study…15 if (dc->prev) {16 if (con->conf.log_condition_handling) { 17 log_error_write(srv, __FILE__, __LINE__, "sb", "go prev", dc->prev->key);18 }19 /* make sure prev is checked first */20 config_check_cond_cached(srv, con, dc->prev);21 /* one of prev set me to FALSE */22 if (COND_RESULT_FALSE == con->cond_cache[dc->context_ndx].result) {23 return COND_RESULT_FALSE;24 }25 26 }27 28 if (!con->conditional_is_valid[dc->comp]) {29 if (con->conf.log_condition_handling) {30 TRACE("cond[%d] is valid: %d", dc->comp, con->conditional_is_valid[dc->comp]);31 }32 33 return COND_RESULT_UNSET;34 }…

20

Human Study

Question presentation

Question: What is the value of the variable "con->conditional_is_valid[dc->comp]" on line 33? (recall, you can use inequality symbols in your answer)

Answer to the Question Above:

21

Human Study…15 if (dc->prev) {16 if (con->conf.log_condition_handling) { 17 log_error_write(srv, __FILE__, __LINE__, "sb", "go prev", dc->prev->key);18 }19 /* make sure prev is checked first */20 config_check_cond_cached(srv, con, dc->prev);21 /* one of prev set me to FALSE */22 if (COND_RESULT_FALSE == con->cond_cache[dc->context_ndx].result) {23 return COND_RESULT_FALSE;24 }25 26 }27 28 if (!con->conditional_is_valid[dc->comp]) {29 if (con->conf.log_condition_handling) {30 TRACE("cond[%d] is valid: %d", dc->comp, con->conditional_is_valid[dc->comp]);31 }32 33 return COND_RESULT_UNSET;34 }…

22

Human Study

Question presentation

Question: What is the value of the variable "con->conditional_is_valid[dc->comp]" on line 33? (recall, you can use inequality symbols in your answer)

Answer to the Question Above:

False

23

Evaluation Metrics

Correctness – is the right answer reported? Time – what is the “maintenance effort”

associated with understanding this code? We favor correctness over time

Participants were instructed to spend as much time as they deemed necessary to correctly answer the questions

The percentages of correct answers over all types of patches were not different in a statistically significant way

We focus on time, as it is an analog for the software engineering effort associated with program understanding

24

Type of Patch vs. Maintainability

Effort = average number of minutes it took participants to report a correct answer for all patches of a given type relative to the original code

25

Type of Patch vs. Maintainability

Effort = average number of minutes it took participants to report a correct answer for all patches of a given type relative to the original code

26

Characteristics of Maintainability We measured various code features for

all patches used in the human study Using a logistic regression model, we

can predict human accuracy when answering the questions in the study 73.16% of the time

A Principle Component Analysis shows that 17 features account for 90% of the variance in the data Modeling maintainability is a complex

problem

27

Characteristics of MaintainabilityCode Feature Predictive

Power

Ratio of variable uses per assignment 0.178

Code readability 0.157

Ratio of variables declared out of scope vs. in scope 0.146

Number of total tokens 0.097

Number of non-whitespace characters 0.090

Number of macro uses 0.080

Average token length 0.078

Average line length 0.072

Number of conditionals 0.070

Number of variable declarations or assignments 0.056

Maximum conditional clauses on any path 0.055

Number of blank lines 0.054

28

Human Intuition vs. Measurement

After completing the study, participants were asked to report which code features they thought increased maintainability the mostHuman Reported Feature Vote

sPredictive Power

Descriptive variable names 35 *0.000

Clear whitespace and indentation 25 *0.003

Presence of comments 25 0.022

Shorter function 8 *0.000

Presence of nested conditionals 8 0.033

Presence of compiler directives / macros 7 0.080

Presence of global variables 5 0.146

Use of goto statements 5 *0.000

Lack of conditional complexity 5 0.055

Uniform use and format of curly braces

5 0.014

29

Conclusions

From conducting a human study involving over 150 participants and patches fixing high-priority defects from real systems we conclude: The fact that humans take less time, on

average, to answer questions about machine-generated patches with automated documentation than human-created patches validates the possibility of using automatic patch generation techniques in practice

There is a strong disparity between human intuitions about maintainability and our measurements and thus we think further study is merited in this area

30

Questions?

31

Modified DeltaDoc

We modify DeltaDoc in the following ways: Include all changes, regardless of length of output Ignore all internal optimizations that lead to loss

of information (e.g. ignore suspected unrelated statements)

Include all relevant programmatic information (e.g. function arguments)

Ignore all high-level output optimizations Favor comprehensive explanations over

brevity Insert output directly above patches as

comments

32

Experiment - Participants

Over 150 participants 27 fourth-year undergraduate CS

students 14 CS graduate students 116 Mechanical Turk internet participants

Accuracy cutoff imposed Ensuring people don’t try to “game the

system” requires special consideration Any participant who failed to answer all

questions or scored below one standard deviation of the average undergraduate student’s score was removed

33

Experiment - Questions

What conditions must hold to always reach line X during normal execution?

What is the value of the variable “y” on line X?

What conditions must be true for the function “z()” to be called on line X?

At line X, which variables must be in scope?

Given the following values for relevant variables, what lines are executed by beginning at line X? Y=5 && Z=True.