When compilers surprise you
It has been reported that Matt Godbolt — the compiler-savvy engineer behind Compiler Explorer — posted a small marvel: a routine that sums integers up to a value, written as a straightforward loop, was transformed by compilers in very different and surprising ways. GCC stuck with loop-y tricks: pairwise accumulation, clever use of lea, and at -O3 even vectorised parallel adds. Neat, predictable, efficient. But then Clang tossed the loop out entirely. Who doesn’t love a plot twist?
The trick
Godbolt walked through the assembly and the math. GCC’s approach folds x and x+1 into x*2+1 so it can add two numbers per iteration; ramp up optimisation and it even parallelises the work. Clang, however, allegedly recognised the algebraic identity beneath the loop and emitted code equivalent to the closed-form v(v - 1) / 2 — an O(1) solution replacing the original O(n) loop. Why exactly Clang emits the specific instruction sequence it does — and not a “simpler” form — is unclear; it’s been suggested the choice helps avoid overflow and follows from how Clang tracks induction variables.
Why this matters
This is a tiny, dazzling reminder that compilers are not merely translators. They’re reasoners trained by decades of engineering. It’s still possible, after twenty years of working with compilers, to be surprised — to nearly fall off your chair. The post is day 24 of an Advent of Compiler Optimisations 2025 series and, amusingly, was written by Godbolt and proof‑read by LLMs and humans. Read it, watch the accompanying video, and remember: sometimes optimisation is math masquerading as magic.
Comments