This is an experiment to reduce overhead of the shadow stack pass. The pass is currently maintaining more stack slots than necessary (anywhere something managed naively goes to the stack) to be sure, but there is room for improvement.

The relatively simple changes in this PR reduce optimized binary size by about 10% and increase throughput by a little less than 10% (testing with the compiler). Optimizations here are risky, though, because any mistake, not only in the shadow stack pass itself but also in compiler code, for instance where an actually necessary (pre-)instrumentation is missing, can lead to unforeseen consequences that only show themselves randomly.

As such it may make sense to test with various programs and various GC configurations first to minimize the risk.

• I've read the contributing guidelines