Oh, it's not entirely incorrect. If you mean "up to k" by "few" :D

What I did was: take all free cells neighbouring blocked cells, try to move only among them; if the top and left border of the rectangle formed by cells and is free, the shortest path between them is the obvious choice and can be easily computed; otherwise, just don't try it at all and try moving to cells that can lead through top and left borders. Dijkstra'd.

I also proved it and it works, but I forgot about the possibility that the path LD->UR corners doesn't lead from to (it can be LR->UD as well). So just checking the whole border of that rectangle instead passes systests. Strangely, my original approach passes 70something-ish systests out of 90, while it should've failed quite quickly (it ignores half the cases, although the checked condition is pretty strong...).

Maybe it's an issue I have with ad-hoc problems, but I truly believe this 600 pt problem was easier to get right than the 250 pt problem from the previos SRM.

After reading the statement of this problem, I immediately thought "coordinates compression + dijkstra". The I coded it and got it accepted with 360 points. Obviously, I had to debug some stuff and be careful with a couple of things, like always including 0 and K as coordinates, handling "intermediate steps" inside the dijkstra, etc., but the thing is I got it accepted.

The same didn't happen with the 250 pt problem from SRM 645. I had many approaches, coded all of them and they all turned out to be wrong. In fact, I still don't know what the correct greedy algorithm is.

user evima (no such handle on CF)

Check evima, for me looks very similar to evima :)

in 31st place gained 25 rating points (2350->2375)

Sort contestants by rating and you'll see that he is 34th among them by rating, and exactly 31st by rating among those who actually opened at least one problem. It gives us a rough approximation for his expected place. When there are only 30 contestants better than you, and your actual rank is 31st — you think that this should be rewarded by more points? This works only in Good Bye 2014 :)

Right, I know it.

But even while comparing it with other SRM's at this time slot — number of participants is decreasing. Rounds 623, 630, 634, 638, 642 all were on this bad time slot — and number of participants during this sequence of rounds is going down.

Maybe we'll see more participants when TCO will be about to start; but TC is clearly losing popularity. Maybe it is because of recent problems (few failed rounds, late announcement of January schedule etc.). Or maybe guys just switched to CodeForces, because it is better than TC right now:)

DFS does not look for a shortest path. You are not trying to optimize it in any way, you are just looking for any path, and when you found it — you don't try to optimize it anymore.

It's a common mistake. You'll find some long path to one of key cells, call recursion from it with small value of k, and don't reach target cell as a result. BTW, here is an editorial.

Guys, come on, why not to use ideone, where I can fork it and run it easily?

I used BFS, read here.

It seems, that you tejas.pandey misunderstood the question, you are checking max x only at the final position... Problem with both DFS I see, that you are marking the field visited even if it is not visited in shortest path — so you cannot continue from that field, when sorter path is found and you can miss some case...