// The usual way to traverse a tree structure is to use a stack, often // implemented with a recursive function so you can remember where you // used to be and where to go next. This is perfectly fine, and for // most trees, the depth is going to be so small that the stack will // not ever become too large. // // But sometimes you either have to be able to handle extremely // lopsided trees, or are perhaps in an environment where you want to // traverse a tree, but do not have access to a stack. The latter may // happen when doing e.g. GPU or vector programming. For such cases // there is a nifty trick for stack-free tree walking. The only thing // it requires is being able to uniquely identify tree nodes (e.g. by // their address) and each node having a pointer to its parent. I'll // demonstrate the technique in C. #include <stdio.h> #include <stdlib.h> // We'll be working with binary trees and for simplicity the // node-level data is just an integer. struct node { int data; struct node* parent; // NULL for the root. struct node* left; struct node* right; }; // We'll need some trees to test with, so here's recursive functions // for creating and freeing trees. This creation can also be done in // a stack-free manner, but I'll leave that for another program. struct node* mk_tree(struct node *parent, int d) { if (d == 0) { return NULL; } else { struct node *x = malloc(sizeof(struct node)); x->data = 0; x->parent = parent; x->left = mk_tree(x, d-1); x->right = mk_tree(x, d-1); return x; } } void free_tree(struct node* x) { if (x) { free_tree(x->left); free_tree(x->right); free(x); } } // Time to walk the tree. For each node, let's increment its value. void visit(struct node *x) { x->data++; } // Here's the standard recursive way of walking the tree. This is a // preorder traversal, because we modify the node value before walking // down the left branch. This isn't important and could have been // done any way. void walk_tree_withstack(struct node *x) { if (x) { visit(x); walk_tree_withstack(x->left); walk_tree_withstack(x->right); } } // Okay, so how do we do this without a stack? walk_tree_withstack() // uses the C call stack to "suspend" execution while recursively // traversing the children. Since the nodes have parent pointers, we // don't need the call stack to know where to return to, but we do need // to know whether we are visiting a node for the first time, have // returned from the left child, or returned from the right child. We // can do this by keeping a pointer to the _previous_ node we visited, // and upon entering a new node, we check whether we just came from // the parent or one of the children. void walk_tree_stackfree(struct node *x) { // Current node we are visiting. Will be x->parent when we are done. struct node *cur = x; // Last node we visited. struct node *prev = x->parent; while (cur != x->parent) { // Initially, suppose the next node to visit is our parent. This // is the case if we've already visited the children. struct node *next = cur->parent; if (prev == cur->parent) { // If the _previous_ node is the parent, then we must be visiting // this node for the first time. visit(cur); // We want to visit the left node (if it exists), otherwise the // right node (if it exists). if (cur->left) { next = cur->left; } else if (cur->right) { next = cur->right; } } else if (prev == cur->left) { // We are returning from the left child, so we want to visit the // right child (if it exists). if (cur->right) { next = cur->right; } } prev = cur; cur = next; } } // The control flow is much more convoluted. If would be even more // convoluted if our tree supported an arbitrary nubmer of children // per node. But how much slower is stack-free traversal? Let's see. // First, define a way to measure performance. #include <time.h> double seconds() { struct timespec ts; clock_gettime(CLOCK_MONOTONIC, &ts); return ts.tv_sec + ts.tv_nsec/1000000000.0; } // And then a main() function that measures the time to construct, // walk, and free a tree of given depth. int main(int argc, char** argv) { if (argc != 2) { fprintf(stderr, "Usage: %s N\n", argv[0]); exit(1); } int d = atoi(argv[1]); double a,b; a = seconds(); struct node *t = mk_tree(NULL, d); b = seconds(); printf("Constructing:\t\t%fs\n", b-a); a = seconds(); walk_tree_withstack(t); b = seconds(); printf("Walking with stack:\t%fs\n", b-a); a = seconds(); walk_tree_stackfree(t); b = seconds(); printf("Walking without stack:\t%fs\n", b-a); a = seconds(); free_tree(t); b = seconds(); printf("Freeing:\t\t%fs\n", b-a); } // Here are some numbers on my computer: // // $ gcc treewalk.c -o treewalk -O2 // // $ ./treewalk 20 // Constructing: 0.035791s // Walking with stack: 0.005445s // Walking without stack: 0.006163s // Freeing: 0.009331s // // $ ./treewalk 25 // Constructing: 1.140084s // Walking with stack: 0.181610s // Walking without stack: 0.195902s // Freeing: 0.301698s // // Stack-free traversal is a bit slower, probably because of the extra // control flow, but it's not awful. Further, the stack-free // implementation can handle extremely unbalanced trees without // blowing the stack, and works in settings where a stack is not // available or practical. I have used it myself for traversing // bounding volume hierarchies in a ray tracer running on GPU: // // https://github.com/athas/raytracingthenextweekinfuthark/