IssueTracker

b/179668593 optimized Subzero usage by enhancing the rr::Optimizer to eliminate more loads and stores. I stopped short of using dominator analysis for a full-blown 'mem2reg' pass.

This can be further improved upon by emitting phi instructions at the Reactor level. For example prior to invoking Variable::materializeAll() prior to an If statement, we could move all unmaterialized variables in a separate list, and at the start of the 'end' basic block emit phi instructions for the ones that were written to in the true/false blocks. This would save loads and stores currently not eliminated for Subzero, while for LLVM it would benefit compile time. This becomes more important as we add actual branches for SPIR-V conditionals. It would also lessen the impact of skipping sampling on inactive paths.

Loops are more difficult since their bodies can dominate themselves: b/180131694#comment4