On 10/03/2020 19:24, John Fastabend wrote: > I guess I'm not opposed to supporting (a) it seems like it should > be doable. Ah ok. Indeed if we add u32 bounds we get (a) for free, I just wasn't sure it was reason enough by itself to justify them. > For (b) the primary reason is to keep symmetry between 32-bit and > 64-bit cases. But also we could have mixed signed 32-bit comparisons > which this helps with. > > Example tracking bounds with [x,y] being signed 32-bit > bounds and [x',y'] being unsigned 32-bit bounds. > > r1 = # [x,y],[x',y'] > w1 > 0 goto pc+y [x,y],[1 ,y'] > w1 s> -10 goto pc+x [-10,y],[1 ,y'] > > We can't really deduce much from that in __reg_deduce_bounds so > we get stuck with different bounds on signed and unsigned space. > Same case as 64-bit world fwiw. I guess we could do more work > and use 64-bit/32-bit together and deduce something Ah ok, problem is when you have good u32 bounds but know nothing about the high 32, so your u64 bounds don't capture those u32 bounds. I think I get it now and I agree that u32 bounds are worth doing :-) -ed
