Spurious performance test failure due to additional stack allocation
In !10446 (closed) we observed that there were two failures on i386 due to additional bytes allocated.
Further investigation revealed two things:
- The difference in the test is due to more stack allocations, specifically more stack underflow.
- On x86_64 the test reported significantly more stack allocations due to many more underflow events (due to the large word size)
Therefore the proposal is to run the performance tests with a larger stack chunk size (say -kc128k) which will produce more stable benchmarks which are more sensitive to actual changes in compiler performance (stack allocations don't really affect runtime)