[personal profile] codedot
I am currently working on implementing needed reduction for interaction nets. To do that, I first needed to refactor a lot of somewhat ugly fast-written code in inet-lib. At some point, I changed retrieving an element from an array to .pop() from .shift(), just because in JavaScript the former happens to be a cheaper operation than the latter.

Many commits later, I decided to play with the program a little bit and compare performance between .shift()ing and .pop()ing. Boom! The program appeared to be broken. Even worse, invariance of the queue that is represented by that array with respect to the order in which it is processed is the whole point of interaction nets, namely the property of strong confluence also known as the one-step diamond property. I thought I fucked up hard.

First, I took a look at git-blame(1) for the line of code that calls .pop(), and found the corresponding commit. Then, I marked its parent commit as good with git-bisect(1). After a few steps, git-bisect(1) found the first bad commit.

Evidently, the problem had something to do with indirection applied by non-deterministic extension of interaction nets. And it did not take more than a couple of minutes to figure out a simple one-liner fix.

Overall, it took less than half an hour from finding a bug to fixing it which I first thought would take hours if not days. To me, it looks like yet another evidence that the idea of git-bisect(1) is totally genius. So, thanks again, Linus!

P. S. Free advice: when making commits, it is always useful to keep in mind 1) a possible need to git-grep(1) some lines of code later, and 2) almost inevitable need to deal with bugs which is a lot easier when commits are suitable for git-bisect(1).
Anonymous( )Anonymous You may post here only if codedot has given you access; posting by non-Access List accounts has been disabled.
Identity URL: 
Account name:
If you don't have an account you can create one now.
HTML doesn't work in the subject.


Notice: This account is set to log the IP addresses of everyone who comments.
Links will be displayed as unclickable URLs to help prevent spam.

Most Popular Tags

August 2017

20212223 2425 26
Powered by Dreamwidth Studios