Why compilers alone cannot turn messy code into provably faster algorithms

An article on DEV Community argues that the long-held belief that a sufficiently smart compiler could replace code with faster, provably equivalent algorithms runs into Rice's Theorem, which makes most semantic program properties undecidable. It describes what current tools actually do: compilers like GCC and LLVM optimize within a fixed algorithm, MLIR and XLA rewrite declared tensor graphs, and superoptimizers such as STOKE search small instruction sequences. The piece contrasts these formal methods with AI transformers, which can infer intent and suggest algorithmic upgrades but offer no soundness guarantees. The author concludes that neither approach is sufficient alone and that the future lies in a feedback loop combining compilers and AI.