Skip to main content

Integers that don't overflow

· 3 min read
Mike Anderson
Mike Anderson
Hacker, Convex Foundation
Claude
Claude
AI Assistant, Anthropic

Convex 0.7.10 starts landing arbitrary-precision integers, native in the CVM. Multiply two enormous numbers and you get the right answer. Not a wrapped answer, not a reverted transaction, not a number shaved to fit a 256-bit word. The right answer.

(* 99999999999999999999 99999999999999999999)
=> 9999999999999999999800000000000000000001

Small integers stay exactly what they were — fast 64-bit longs, which cover almost everything real programs do. When a result outgrows 64 bits, the CVM promotes it to a big integer automatically, currently up to 32,768 bits. One integer type, no ceremony at the boundary.

Overflow is a policy, and most platforms chose badly​

Every platform has to decide what happens when arithmetic exceeds the machine word. The mainstream blockchain answer is a 256-bit word that either wraps silently (the source of a long series of token-supply exploits) or reverts. A whole defensive-programming culture — SafeMath and friends — exists because the VM's default answer to "what is a + b?" was sometimes wrong.

We find that backwards. Arithmetic on money should not have sharp edges at powers of two. A financial platform's integers should behave like the integers you learned at school: closed under addition and multiplication, no surprises at 2^63 or 2^256. If the maths is correct by construction, an entire category of exploit simply doesn't exist.

No free supercomputing​

The obvious objection: doesn't arbitrary precision let someone burden peers with thousand-digit multiplications? This is where juice accounting earns its keep. Operations on big integers cost juice in proportion to the work — the bigger the operands, the more you pay. Honest users doing money maths in long-sized numbers pay long-sized costs; someone who wants to multiply 4,000-byte integers can do so, at a price that reflects what they're asking several thousand peers to compute. Deterministic costs for deterministic work.

Why 32,768 bits is more interesting than it sounds​

Beyond not-overflowing, large integers make serious number theory possible on-chain. Modular exponentiation over 2048- or 4096-bit numbers is the core operation of RSA-style cryptography, accumulators and a fair amount of the verifiable-computation literature. Whether or not those land on Convex soon, we'd rather the numeric foundations didn't rule them out.

This first release is part one — the type, promotion rules and core operations — with follow-up work refining behaviour and costs across the full numeric tower in 0.7.11. It's the least glamorous kind of feature: one that removes an entire class of bugs by making the correct thing the only thing.