Hacker News Reader

Target a virtual architecture (LLVM IR). Run the original code in an emulator for the target system (QEMU); give warnings where the two differ, with a switch for the virtual architecture to mimic the behavior of whatever architecture you need.

You get friendly C, a new way to think about warnings, and your old code can still be compatible with the new virtual architecture.

Next up, world hunger.

View on HN

You can't simply run code for all possible inputs.

No, but you can certainly exercise it well. It might be a degree of confidence measure. Combined with property verification (so, formal verification), you'd get a damn good idea they'd be close.

This solution relies on some undecidable problems as being decidable

How so? I don't see the connection to undeciability.

If you have undefined behaviour after an infinite loop that breaks on some condition, then that undefined behaviour will not necessarily have been triggered. Thus, no warning.

Regehr himself "disproved" Fermat's Last Theorem by abusing a non-terminating loop and over-aggressive compilers.

http://blog.regehr.org/archives/140

Yes, this I a great example of the consequences of undecidability. Was taught to me in my algorithms and proofs class

Reply on news.ycombinator.com