shunshou
commented
6 years ago @azidar What do you expect should happen in these cases?
1) Directly assign a <= b where a has known range that's smaller than b. Should that be an error without explicit b.wrap(a), b.clip(a), or b.assignInterval(a)? Also, re: back-propagation of ranges, if b has some smaller bound and b <= x - y, I assume you can't make any assumptions about the ranges of x and y. However, if you've constrained x and y to be only positive and use only multiply and add ops i.e. b <= x + y, you can further optimize x and y ranges -- should be min(x_range, b_range). That's not something the tool should be able to figure out, maybe? But it'd be nice to have some optional Firrtl pass go and trim widths that way, maybe (unless synthesis tools are smart about that). More generally, does a always keep its range if known or will it take on b's range if b's range is smaller?
2) Directly assign a <= b where a's bp = 5 and b's bp = 12 (both known). You're losing precision, which means that your interval also changes. In that case, before range inference, it might make sense to explicitly add b.bpset(a)? -- Otherwise, you're propagating the wrong ranges, which potentially accumulates and becomes probelamtic.
Looks like a always keeps a's range...
Branches