Effectively, the constness of the shift amount is not used.

This is intentional. Basically, if we special-case x << 1, then should we special-case x << (1 + 1)? "No constant folding is ever performed when determining the shape of a value" is a nice simple answer. Constant folding is not.

• __check_shamt prevents shifting by a signed Value, so b_signed will never be true.

That's a leftover of some very early changes. We can get rid of that.

I'm curious, do you have a case where constant folding would help you? In general shape determination in nMigen is very conservative, so I'm not sure if fixing this particular instance would help a lot.

No, I just thought it was a bit odd to see that. Truncation "just works":

x = Signal(32)
y = Signal(32)

m.d.comb += y.eq(x << 5)
m.d.comb += Assert(y == (x << 5)[:32])

so it never really matters.

Clearly if I did something like:

x = Signal(32)
y = Signal(len(x << 5))

It would make a difference, but honestly I never do that and I don't see why anyone else would.