[LLVMdev] GSoC - Range Analysis

[John Regehr]

> I actually disagree on the last point.  For static array bounds checking with 
> SAFECode, we want to assume that all integer operations can overflow 
> (including those with the nsw attribute).  We don't care about the other 
> integer semantic rules from C (like the promotion rules) or whether overflow 
> is language defined or not (for those who don't know, unsigned ints have 
> defined overflow in C; signed ints do not).  All we want to know is whether a 
> pointer p is within the bounds of a set of valid objects, and we just need to 
> take integer overflow into account because it can happen on the hardware.

I'd have to see some specific examples to figure out if what you're saying 
makes sense. All of my experience with integer overflows in C/C++ had lead 
me to the conclusion that playing games with the semantics usually gets 
you into hot water.

John