BorrowSanitizer Benchmarks

Overview

We measure performance as a multiple of the execution time of a baseline configuration. There are three baselines:

• native — Uninstrumented native execution in debug mode.
• miri-tb — Miri with Tree Borrows enabled.
• miri-base — Miri without Tree Borrows enabled.

We disable all other forms of run-time checking for each of the Miri baselines, and we intercept invocations of Miri's interpreter to avoid including its Cargo plugin in our timing measurements. This ensures that we are only measuring the time when Miri is actually interpreting a program.

BorrowSanitizer has two configurations that we can compare against these baselines:

• full — All of our checking is enabled.
• full-no-wildcard — All of our checking is enabled, except for wildcard semantics.
• no-op — Inserts all of our run-time instrumentation, but disables the checks themselves.

To summarize the performance of a crate, we calculate relative execution time for every unit test and report the mean value. We use hyperfine to measure individual tests, and we record wall time as the mean of three iterations after a "warm-up" run.

We compile crates under BorrowSanitizer using our Cargo plugin. This plugin configures libtest as if it were being built for Miri, which disables parsing information about the user's terminal. This process is slow under both Miri and BorrowSanitizer.