Dominator (graph theory)

Applications

Dominators, and dominance frontiers particularly, have applications in compilers for computing static single assignment form. A number of compiler optimizations can also benefit from dominators. The flow graph in this case comprises basic blocks.

Dominators play a crucial role in control flow analysis by identifying the program behaviors that are relevant to a specific statement or operation, which helps in optimizing and simplifying the control flow of programs for analysis.

Automatic parallelization benefits from postdominance frontiers. This is an efficient method of computing control dependence, which is critical to the analysis.

Memory usage analysis can benefit from the dominator tree to easily find leaks and identify high memory usage.

In hardware systems, dominators are used for computing signal probabilities for test generation, estimating switching activities for power and noise analysis, and selecting cut points in equivalence checking. In software systems, they are used for reducing the size of the test set in structural testing techniques such as statement and branch coverage.

Algorithms

Let ${\displaystyle n_0}$ be the source node on the Control-flow graph. The dominators of a node ${\displaystyle n}$ are given by the maximal solution to the following data-flow equations:

${\displaystyle \mathrm{Dom}(n)=\{\begin{array}{ll}\left\{n\right\}& \text{ if }n=n_0\\ \left\{n\right\}\cup \left(\bigcap _{p\in \text{preds}(n)}^{}\mathrm{Dom}(p)\right)& \text{ if }n\ne n_0\end{array}}$

The dominator of the start node is the start node itself. The set of dominators for any other node ${\displaystyle n}$ is the intersection of the set of dominators for all predecessors ${\displaystyle p}$ of ${\displaystyle n}$. The node ${\displaystyle n}$ is also in the set of dominators for ${\displaystyle n}$.

An algorithm for the direct solution is:

// dominator of the start node is the start itself Dom(n<sub id="mw0w">0</sub>) = {n<sub id="mw1A">0</sub>} // for all other nodes, set all nodes as the dominators <b id="mw1Q">for each</b> n <b id="mw1g">in</b> N - {n<sub id="mw1w">0</sub>} Dom(n) = N; // iteratively eliminate nodes that are not dominators <b id="mw2A">while</b> changes in any Dom(n) <b id="mw2Q">for each</b> n <b id="mw2g">in</b> N - {n<sub id="mw2w">0</sub>}: Dom(n) = {n} union with intersection over Dom(p) for all p in pred(n)

The direct solution is quadratic in the number of nodes, or O(n²). Lengauer and Tarjan developed an algorithm which is almost linear, and in practice, except for a few artificial graphs, the algorithm and a simplified version of it are as fast or faster than any other known algorithm for graphs of all sizes and its advantage increases with graph size.

Keith D. Cooper, Timothy J. Harvey, and Ken Kennedy of Rice University describe an algorithm that essentially solves the above data flow equations but uses well engineered data structures to improve performance.

Postdominance

Analogous to the definition of dominance above, a node z is said to post-dominate a node n if all paths to the exit node of the graph starting at n must go through z. Similarly, the immediate post-dominator of a node n is the postdominator of n that doesn't strictly postdominate any other strict postdominators of n.