Control Flow Analysis
Subsequent branch-free instructions make up segments of linear code. The first instruction of the program is a leader, as is every target of a branch, and every instruction immediately following a branch (including conditional branches and procedure returns) is a leader. A basic block is the maximum segment of linear code associated with each leader -- it ends with either the program's last instruction or the first branch following a leader. Points exist before and after each statement in a basic block (there are thus a minimum of 2 points per block, and s + 1 for a block with s statements). A basic block a flows to b if and only if:
- either b immediately follows a, and a does not end in an unconditional branch,
- or a ends in a branch, of which b is a potential target.
Note that an indirect branch, without context information, trivializes all blocks (every instruction becomes a leader) and flows to them all from at least that point (an arborescence is formed)!
The directed multigraph defined by interpreting basic blocks as vertices, and flow relationships as edges, yields its control flow graph (CFG). A start node exists for each CFG, corresponding to the basic block whose header is the first instruction of the program.
The antisymmetric, transitive, reflexive domination relation is defined on vertices of a CFG (and thus basic blocks of the underlying program). A vertex a dominates b (a <= b) if every path from the start node s to b passes through a. A vertex a properly dominates b (a < b) if a dominates and is not equal to b. A vertex a directly/immediately dominates b (a <d b) if a properly dominates b, and a dominates no vertex c that dominates b. This relation induces the dominator tree, where nodes dominate all descendents in the tree. The start node s dominates all nodes, properly dominates all nodes but itself, and roots the dominator tree.
- It should be obvious that a's preceding of b in the CFG is not necessary for even immediate dominance of b by a.
Paths p1,p2...pn are defined as sequences of points such that either pi+1 immediately follows pi in a basic block, or pi+1 begins a successor block to a block terminated by pi. It is on these points that dataflow equations are evaluated.
Dataflow analysis is most usefully performed into and out of regions, subsets of the nodes such that a header exists which dominates all nodes in the region, and all edges between nodes in the region are themselves in the region. A loop is a region which is strongly connected, where all back-edges to the header are themselves within the region FIXME -- unclear.
Loops can be discovered via domination analysis (it is important to note that this refers to loops in the generated code, not loop constructs of the source language, and furthermore that all possible loops will be found (ie, unstructured loops constructed from C gotos)). Discover all strongly-connected subgraphs (SCCs) of the CFG (subgraphs where, for each vertex, a path (not necessarily an edge) exists from that vertex to all other nodes of the subgraph); if a subgraph contains a node dominating all that subgraph's nodes, the subgraph is a loop. The trivial case is, of course, a statement which jumps to itself, ala the BASIC program 10 GOTO 10. Implementation via Kosaraju's Algorithm is simple, with O(|V|+|E|) time complexity using graph encoding and O(N2) time complexity using adjacency matrices:
- Perform a recursive depth-first traversal of the graph starting from s. Each time you return, add that node onto an auxiliary vector. Upon the traversal's completion, this vector sorts the nodes topologically.
- Until the the vector is empty, use the last node of the vector to begin traversing the transpose graph. Remove the path from the vector; these paths partition the graph into SCCs.
Kosaraju's algorithm is improved upon by Tarjan's Algorithm and Gabow's Algorithm. Natural loop identification proceeds via identification of back edges (edges from a node b to a node a, where a dominates b). A loop is associated with every such back edge; if a backedge exists from b to a, the associated loop is entered at a, and consists additionally of all nodes which can reach b without going through a. Similarly, a loop is associated with the target of every back edge, this being the union of all such backedges' associated natural loops.
If a statement S writes to a variable V, S is said to define V. If statement S reads from a variable V, S is said to use V (the two are not mutually exclusive). A definition of v is killed between p1 and p2 if every path between them contains a definition of v; conversely, if a path exists from p1 to p2 which does not redefine v, and v has been defined on input to p1, it reaches p2 from p1.
- Primary references: Allen/Kennedy ch. 2-3, Muchnick ch. 14
- Statement order within a loop is never a factor in preserving loop-carried dependencies
- "On the perfect accuracy of an approximate subscript analysis test" (Klappholz, Psarris, Kong, 1990) analyzes the GCD and Banerjee inequalities, explaining the crappiness of the former and general robustness of the latter. "On the Accuracy of the Banerjee Test" (same authors, 1991) suggests improvements on the Banerjee test.
- "Eliminating False Data Dependencies using the Omega Test" (Pugh, Wonnacott, 1992) moves from integer programming-based (Diophantine) solutions to a subclass of the Presburger formulae.
- Primary references: Allen/Kennedy ch. 8, Muchnick ch. 16
- Take advantage of temporal locality and, in a sense, spatial locality at the subword level.
- Loop fusion merges loops in order to enhance temporal locality
- Especially relevant for machine-generated (preprocessed) input
- Primary references: Allen/Kennedy ch. 9, Muchnick ch. 20. See also Caches on the Architecture page.
- The benefits of spatial locality can now be taken advantage of at the more natural word level.
- Loop interchange -- are we striding through nested loops optimally? If not, we can interchange levels, provided all dependencies are preserved. Especially powerful when combined with hardware- or software-based prefetching and further augmented with non-blocking caches. Optimal interpositioning is NP-hard in general but excellent heuristics exist. See Allen/Kennedy p. 472-474.
- This can be valuable enough that we perform small transformations to accommodate it: loop alignment, loop peeling, etc
- Loop blocking/tiling/strip-mining -- breaking out subloops to work on a cache-friendly block at once
- Especially good for parallelizing across coherent processing elements with local, coherent caches -- coherency traffic is minimized
- Also well-suited for cases of reuse between iterations of a non-innermost loop
- Loop safety is closed under blocking operations, but not necessarily under interchange.
- Blocking can, in some cases, make subsequent interchange safe! "If the strip size is less than or equal to the threshold of the dependence that might prevent interchange, then the strip-mine-and-interchange is legal, even though the interchange would not be." (Allen/Kennedy p. 480)