(Analysis by Nathan Pinsker)

This is a shortest-path problem where some of the tiles follow special rules. In shortest-path problems (and dynamic programming problems in general), solving them requires you to capture the "state" of the system: some small amount of information that describes a subproblem, so that you don't have to repeat work. In a standard shortest-path algorithm, this "state" might be simply the current location. In this problem, we'll take the state to be the current location, and also a boolean that represents whether Bessie currently smells like oranges. In this way, we can run a Dijkstra algorithm with a priority queue as normal.

You could argue that the "state" could use some additional information, like whether Bessie is currenly sliding and, if so, in what direction. This is fine, and this approach can solve the problem as well! It actually ends up running faster -- you can use a BFS instead of a heap-based queue -- but at the cost of making the transitions between states a bit more complicated and harder to code.

Here is Mark Gordon's code, which takes the second approach. He uses integers between 0 and 10,000,000 to represent Bessie's current state: (1000 x 1000) for her position, times 5 for whether she is sliding, times 2 for whether she smells like oranges.

#include <iostream>
#include <vector>
#include <queue>

using namespace std;

int dr[] = {-1, 0, 1, 0};
int dc[] = {0, -1, 0, 1};

struct state {
  int r;
  int c;
  int ld;
  bool smell;

  state(int r, int c, int ld, bool smell) :
      r(r), c(c), ld(ld), smell(smell) {
  }

  int pack() {
    return (smell ? 1 : 0) + 2 * (ld + 1) + 10 * c + 10000 * r;
  }

  static state unpack(int x) {
    return state(x / 10000, (x / 10) % 1000,
                 (x / 2) % 5 - 1, x & 1);
  }
};

int getcell(const vector<vector<int>>& A, int r, int c) {
  if (r < 0 || r >= A.size() || c < 0 || c >= A[r].size()) {
    return 0;
  }
  return A[r][c];
}

int main() {
  ios_base::sync_with_stdio(false);

  int N, M;
  cin >> N >> M;
  vector<vector<int>> A(N, vector<int>(M));
  for (int i = 0; i < N; i++) {
    for (int j = 0; j < M; j++) {
      cin >> A[i][j];
    }
  }
  queue<int> q;
  vector<int> D(10000000, -1);

  int s = state(0, 0, -1, false).pack();
  q.push(s);
  D[s] = 0;

  while (!q.empty()) {
    state st = state::unpack(q.front());
    q.pop();

    if (st.r == N - 1 && st.c == M - 1) {
      cout << D[st.pack()] << '\n';
      return 0;
    }
    if (A[st.r][st.c] == 4 && st.ld != -1) {
      int col = getcell(A, st.r + dr[st.ld], st.c + dc[st.ld]);
      if (col != 0 && col != 3) {
        state nst = state(st.r + dr[st.ld], st.c + dc[st.ld], st.ld, col == 2);
        if (D[nst.pack()] != -1) {
          continue;
        }
        D[nst.pack()] = D[st.pack()] + 1;
        q.push(nst.pack());
        continue;
      }
    }
    for (int i = 0; i < 4; i++) {
      int col = getcell(A, st.r + dr[i], st.c + dc[i]);
      if (col == 0 || (col == 3 && !st.smell)) {
        continue;
      }
      state nst = state(st.r + dr[i], st.c + dc[i], i,
                        col == 2 ? true : col == 4 ? false : st.smell);
      if (D[nst.pack()] != -1) {
        continue;
      }
      D[nst.pack()] = D[st.pack()] + 1;
      q.push(nst.pack());
    }
  }
  cout << "-1\n";
  return 0;
}

Here is my code, which takes the first approach, using the Java PriorityQueue so that Bessie's slides can be processed all at once. I'm sorry for the awful butchery of implementing 4-tuples:

import java.io.BufferedReader;
import java.io.IOException;
import java.io.InputStreamReader;
import java.util.PriorityQueue;
import java.util.StringTokenizer;

public class DEC15GoldA {
	static final int N = 1005;
	
	static int[][] tile;
	static int[][][] visited;
	static int[][] dirs = { {-1, 0}, {1, 0}, {0, -1}, {0, 1} };
	static int n, m;
	
	static PriorityQueue<Pair<Integer, Pair<Integer, Pair<Integer, Integer>>>> q;
	
	public static void main(String[] args) throws IOException {
		BufferedReader br = new BufferedReader(new InputStreamReader(System.in));
		StringTokenizer st = new StringTokenizer(br.readLine());
		
		n = Integer.parseInt(st.nextToken());
		m = Integer.parseInt(st.nextToken());
		q = new PriorityQueue<>();
		
		tile = new int[N][N];
		visited = new int[N][N][2];
		
		for (int i=0; i<n; ++i) {
			st = new StringTokenizer(br.readLine());
			for (int j=0; j<m; ++j) {
				int color = Integer.parseInt(st.nextToken());
				tile[i][j] = color;
				visited[i][j][0] = visited[i][j][1] = 100000000;
			}
		}
		q.add(fromInts(0, 0, 0, 0));
		
		while (q.size() > 0) {
			Pair<Integer, Pair<Integer, Pair<Integer, Integer>>> front = q.remove();
			int dist = front.left;
			int x = front.right.left;
			int y = front.right.right.left;
			int isSmelly = front.right.right.right;
			
			if (x == m-1 && y == n-1) {
				System.out.println(dist);
				return;
			}
			if (visited[y][x][isSmelly] != 100000000) {
				continue;
			}
			visited[y][x][isSmelly] = dist;
			
			for (int i=0; i<dirs.length; ++i) {
				int nx = x + dirs[i][0];
				int ny = y + dirs[i][1];
				int nd = dist + 1;
				int nSmelly = isSmelly;

				if (!isPathable(nx, ny, isSmelly)) continue;
				
				if (tile[ny][nx] == 4) {
					while (isPathable(nx + dirs[i][0], ny + dirs[i][1], isSmelly) &&
							tile[ny][nx] == 4) {
						nx += dirs[i][0];
						ny += dirs[i][1];
						nd ++;
						nSmelly = 0;
					}
				}
				if (tile[ny][nx] == 2) {
					nSmelly = 1;
				}
				if (visited[ny][nx][nSmelly] <= nd) continue;
				q.add(fromInts(nd, nx, ny, nSmelly));
			}
		}
		System.out.println("-1");
	}
	public static boolean isPathable(int x, int y, int smellsNice) {
		if (x < 0 || x >= m || y < 0 || y >= n) return false;
		if (tile[y][x] == 0) return false;
		if (tile[y][x] == 3) return (smellsNice > 0);
		
		return true;
	}
	public static Pair<Integer, Pair<Integer, Pair<Integer, Integer>>> fromInts(int a, int b, int c, int d) {
		Pair<Integer, Integer> p1 = new Pair<>(c, d);
		Pair<Integer, Pair<Integer, Integer>> p2 = new Pair<>(b, p1);
		return new Pair<>(a, p2);
	}
}
class Pair<L extends Comparable<L>, R> implements Comparable<Pair<L, R>> {
	public L left;
	public R right;
	
	public Pair(L left, R right) {
		this.left = left;
		this.right = right;
	}
	public int compareTo(Pair<L, R> other) {
		return this.left.compareTo(other.left);
	}
}