Re: Weird revision walk behaviour

From: Kevin Bracey <kevin@bracey.fi>
Date: May 30, 2018, 08:20 UTC
Message-ID: <97644280-2187-d314-37ce-2c79935a63bc@bracey.fi>
In-Reply-To: <20180529210434.GA3857@sigill.intra.peff.net>

On 30/05/2018 00:04, Jeff King wrote:

Show 6 quoted lines

>
> Do we even need to do the parent rewriting here? By definition those
> parents aren't interesting, and we're TREESAME to whatever is in
> treesame_parents. So conceptually it seems like we just need a flag "I
> found a treesame parent", but we only convert that into a TREESAME flag
> if there are no relevant parents.

I think it's necessary to make the rules consistent. To mark the commit as TREESAME here when it's not TREESAME to all its parents would be inconsistent with the definition of the TREESAME flag used everywhere else:

* Original definition: "A commit is TREESAME if it is treesame to any 
parent"
* d0af66 definition: "A commit is TREESAME if it is treesame to all parents"
* Current 4d8266 definition: "A commit is TREESAME if it is treesame to 
all relevant parents; if no relevant parents then if it is treesame to 
all (irrelevant) parents."

The current problem is that the node is not marked TREESAME, but that's consistent with the definition. I think we do have to rewrite the commit so it is TREESAME as per the definition. Not flag it as TREESAME in violation of it.

It's possible you *could* get away with just flagging, because we never recompute the TREESAME flag in simple history mode. But it would be a cheat, and it may have other side effects. It means this node would remain a special rare case for others to trip up on later. And I don't think it simplifies the scan. Remembering "pointer-to-first-treesame-parent" (not a list) for the rewrite is no more complex than remembering "bool-there-was-a-treesame-parent". (A bool is what earlier code did - it worked for the original TREESAME definition. My patch series dropped that bool without replacement - missing this all-irrelevant case).

In the simple history mode, the assumption is we're "simplifying away merges up-front" here; we won't (and can't) rewrite parents later in a way that needs to recompute TREESAME. In the initial scan when all parents are relevant and we matched one, the commit became TREESAME as per the new definition immediately because of the rewrite. This applies the equivalent rewrite when no relevant parents, consistent with the general concept, and without changing the TREESAME definition.

Kevin