### Bug description: #118450 "Convert the evaluation stack to stack refs" (by @Fidget-Spinner) unfortunately seems to have introduced a [7% - 11% performance regression overall](https://github.com/faster-cpython/benchmarking-public/tree/main/results/bm-20240627-3.14.0a0-22b0de2) (though as much as 25% for some benchmarks) on Windows only. Linux (both x86_64 and aarch64) and macOS (arm) seem to have a change beneath the noise threshold. To be clear, @Fidget-Spinner's work is great and the lesson here is that for these sorts of sweeping low-level changes, we should have benchmarked on all of the Tier 1 platforms to be certain. I don't know what the cause is -- it is probably MSVC not being able to "optimize through" something and introducing unnecessary overhead, but that's just a theory. Cc: @brandtbucher, @markshannon ### CPython versions tested on: CPython main branch ### Operating systems tested on: Windows <!-- gh-linked-prs --> ### Linked PRs * gh-121270 <!-- /gh-linked-prs -->