[RFC] More Consistent Notion of "Infinite Width" Values

Discussion originated in https://github.com/freechipsproject/chisel3/pull/857. The motivating example is bit extraction for literals of unspecified width. For example:

for (i <- 0 until 4) {
  io.out(i) := io.in(i) && 5.U(i)
}

Here I'm using the literal 5.U as a numerical value, I don't care about the width. However, without the special-case support in #857 this would result in an out-of-bounds indexing error. A more consistent notion would be helpful because the existing implementation that works for literals does not work for other Chisel values like Wires, Registers, nor Ports.

Sketch of Proposal

As discussed in the Chisel developers meeting on July 27th, we could create a new Chisel API (and corresponding FIRRTL primop) that does something like padToInfinity or padUnbounded. The full implications of such an operation are not clear, but at least in the case of bit select, the argument would be treated as if it were zero- or sign-extended to infinity. I imagine we need to think through exactly how this influences width inference (if at all).

PR Template

Type of issue: feature request

Impact: API modification*

It's hard for me to see how we do this in a principled way without some sort of API modification. The extend of it remains to be seen.

Development Phase: request