Two studies are reshaping assumptions about compiler optimisations, finding that higher optimisation levels do not reliably produce faster software and that carefully targeted techniques often outperform blanket settings. Examining common toolchains across a mix of real-world workloads, the research highlights wide variability in results from flags such as -O3 and -Ofast, with code size growth, cache effects and platform differences frequently erasing or reversing expected gains.
One study maps performance across compiler versions and architectures, showing that the same project can swing between speedups and slowdowns solely from toolchain updates. The work identifies profile-guided optimisation and link-time optimisation as the most consistently beneficial approaches, especially when guided by representative input data. It also flags brittle interactions among individual flags—such as aggressive inlining combined with vectorisation—that can inflate binaries and degrade branch prediction, underscoring the need to test configurations rather than rely on defaults.
A second study focuses on auto-vectorisation and memory behaviour, reporting that compute-bound kernels often benefit from SIMD transforms while memory-bound workloads see modest or negative returns unless layout and prefetching are addressed. The analysis tracks energy use alongside performance, finding that faster isn’t always greener: some optimisations reduce runtime but raise power draw enough to negate efficiency gains, particularly on mobile and edge-class processors with dynamic frequency scaling.
The findings point to practical guidance for developers and vendors. Teams are urged to evaluate optimisations with production-like data, monitor binary size and cache metrics, and pin toolchains to reduce regression risk in continuous integration. For compiler authors, the studies call for clearer diagnostics about why transformations trigger, more stable flag semantics across versions and expanded benchmark suites that reflect modern, memory-intensive applications.