Rust Without Shortcuts: What Safety Actually Costs
Abstract
Memory-unsafe languages remain the dominant root cause of severe software vulnerabilities, and national cybersecurity guidance now treats migration to memory-safe languages as a structural requirement. Automated C-to-Rust translation reports four to five percent runtime overhead, but on small programs that avoid dynamic allocation and complex aliasing. This article measures what memory safety costs as a function of the mechanism that enforces it, across twelve build configurations of ten languages and fourteen workloads, on bare-metal ARM64 and x86-64 Linux hosts and an Apple laptop. Seven metrics are captured, including hardware performance counters and a compiler-enforced unsafe-construct count of zero throughout. Across the controlled twelve benchmarks, hand-written safe Rust runs at a median of 0.99 to 1.03 times C and within 1.03 to 1.10 times its peak memory, while every configuration enforcing safety at run time—tracing collection under a just-in-time or ahead-of-time compiler, or automatic reference counting—costs 1.7 to 2.9 times Rust’s aggregate wall time under the cold-process measurement model used here. Their memory cost is uneven: 1.6 to 1.9 times Rust for Go and C# NativeAOT, 5.7 times for Swift, and an order of magnitude or more for the four just-in-time configurations. Toggling Zig’s runtime bounds and overflow checks on one source isolates that cost at 16 to 17 percent on every platform, while still leaving use-after-free reachable. The overhead is not a language constant either: on one numerical kernel, identical Rust source retires half of C’s instructions on ARM64 and two-thirds more on x86-64, an inversion no single-platform study can detect and that the generated code shows is not explained by vectorisation.