IssueTracker

We currently 'flatten' all conditional branches, executing both edges even when all SIMD lanes are active/inactive. We can instead perform a jump in these cases.

A heuristic should determine whether it's worth making a potentially badly predictable jump, or whether flattening is still preferable, per conditional branch.

Basic blocks which are skipped when all SIMD lanes are inactive can omit run-time tests such as the one introduced for b/153380916.

For loops which aren't unrolled we should also always jump when all lanes are inactive, since otherwise there is often an additional iteration as described in b/153380916, and there's an exit jump anyway so we don't need a heuristic for avoiding misprediction in this case.