Early in the linked thesis there is a one-page argument about the shortcomings of traditional approaches, which technically isn't what you asked but might still answer the side of the question that deals with human usage at least:
I’d imagine there’s some challenging judgement calls that such a tool would have to make. Like, in Go, you can reorder the members of a struct definition. In many cases this is just diff noise to reviewers. HOWEVER, it does impact the layout of the struct in memory, so it can be semantically meaningful in performance work.
A wild nitpicker appears. I understand where you're coming from & why this matters. But Go, the language spec, doesn't make any guarantees about struct layout at all. A layout difference may be meaningful, practically, but it's potentially unreliable.
PHP long stated that associative array sorting order was unstable and not guaranteed (especially when the union (+) operator or array_merge function were involved) - that doesn't mean ten bazillion websites wouldn't instantly break if they ever actually changed the ordering to be unpredictable.
Language designers need to contend with the fact that the ultimate final say in whether a thing is or not is whether that behavior is observed.
Didn't ruby actually do exactly this though? And it broke a million websites and they changed it back in the next version and have made it explicit ever since? To me that is much stronger evidence than what we think would happen if php did it.
I don't know about Ruby, but one example I can think of where a language made the instability explicit is that early on in the language Go changed the behavior of the select statement:
> If one or more of the communications can proceed, a single one that can proceed is chosen via a uniform pseudo-random selection.
In an early implementation it would pick in lexical order, IIRC (and the specification did not mention how a communication should be picked). Not only could this lead to bugs, apparently some people were relying on it and they didn't want that.
The tl;dr is that there's an almost infinite number of ways to atomize/conceptualize code into meaningful "units" (to "register" it, in my supervisor's words), and the most appropriate way to do that is largely perspectival — it depends on what you care about after the fact, and there is no single maximal way to do it up front.
I mean to have an improvement over the status quo we need to simply find a conception that works better than lines as units of code. Let’s not let perfect be the enemy of the good.
love to tell someone who literally wrote a masters thesis on a topic what we need to "simply do" to solve it lol. I almost want to admire the confidence but
>>I’d imagine there’s some challenging judgement calls that such a tool would have to make
Just thinking about it makes my head spin. I spend a lot of time working out font/color hierarchies, supplementary to coding and data viz. Arguably what you're bringing up is a case for a carefully colored diff that visually cues whether something is a true semantic change or indicative of a lower level issue. I'm comfortable with reading a plain ol' diff that just shows me what changed, superficially, and interpreting it. While I think OP's idea is awesome, it also might create more confusion than it resolves; and resolving confusion is the point of a diff.
Efficiency is not the issue at this point. My prototype diffing algorithm was linear and there have been improvements on it already (I think something called "truediff" is linear but an order of magnitude better! I could be misremembering the name, don't quote me :) ).
The real difficult part is in how you represent AST-level changes, which will limit what your merging algorithm can do. In particular, working around moving "the same subtree" into different places is difficult. Imagine the following conflict:
Both p and q move the same thing to different places so they need a human to make a decision about what's the correct merge. Depending on your choice of "what is a change", even detecting this type of conflict will be difficult. And that's because we didn't add insertions nor deletions. Because now, say p was:
([1,2,3], [4,5]) -- p --> ([1,3], [2,5])
One could argue that we can now merge, because '4' was also deleted hence the position in which we insert '2' in the second list is irrelevant.
If we extrapolate from lists of integers to arbitrary ASTs the difficulties become even worse :)
