This pass breaks large SSA2 functions into zones, which are connected subgraphs of the dominator tree. For each zone, at the node that dominates the zone (the "zone root"), it places a tuple collecting all of the live variables at that node. It replaces any variables used in that zone with offsets from the tuple. The goal is to decrease the liveness information in large SSA functions.
Details and Notes
Compute strongly-connected components to avoid put tuple constructions in loops.
There are two (expert) flags that govern the use of this pass
Zone splitting only works when the number of basic blocks in a function is greater than n. The n used to cut the dominator tree is set by -zone-cut-depth.
There is currently no attempt to be safe-for-space. That is, the tuples are not restricted to containing only "small" values.
In the HOL program, the particular problem is the main function, which has 161,783 blocks and 257,519 variables — the product of those two numbers being about 41 billion. Now, we’re not likely going to need that much space since we use a sparse representation. But even 1/100th would really hurt. And of course this rules out bit vectors.