Make polymorhic bitvector

[cdonovick]

Addresses #122

Adds polymorphic types which allows "run-time polymorphism"

Enables the following:

val = BitVector[size](...)
cond = Bit(...) # may be symbolic
s_val = SIntVector[size](val)
u_val = UIntVector[size](val)
c_val = cond.ite(s_val, u_val)
res = c_val < 1

where the final statement is effectively transformed to following by the type system

m0 = type(s_val).__lt__
m1 = type(u_val).__lt__ 
if m0 is m1:
    res = m0(c_val, 1) # doesn't matter which is called
else:
    res = cond.ite(m0(c_val, 1), m1(c_val, 1))

[cdonovick]

Discovered some bugs closing for now

[rdaly525]

Looks good, but could you add the following chained ite's to your tests?

ite(Poly[A ,B], A)
ite(Poly[A, B], B)
ite(Poly[A, B], C)
ite(A, Poly[A, B])
ite(B, Poly[A, B])
ite(C, Poly[A, B])
ite(Poly[A, B], Poly[A,B])
ite(Poly[A, B], Poly[A, B, C])
ite(Poly[A, B, C], Poly[A, B])

[cdonovick]

None of those are valid but I'll add tests for chaining.

[rdaly525]

Poly[A,B] is just indicating the psuedo-code type of the argument. These should all be valid.

[leonardt]

Could use T.random(width) here

[leonardt]

(and above)

[leonardt]

This seems to be the magic, a PolyType store's it's select condition, operators are dispatched based on that condition when invoked.

[leonardt]

In theory this might work with magma types?

[leonardt]

Actually not quite, Poly would need to implement the magma protocol, but then I think it could work, maybe just by concatenating the bits of the two subtypes together as the bits representation?

[cdonovick]

I plan on handling magma with my MagmaBV proposal. phanrahan/magma#587

MagmaBV will use the same machinery to build ites as BitVector / SMTBitVector and implement the magma protocol.

[leonardt]

LGTM

[cdonovick]

Worth noting that this is going to generate a lot of terms of the form:

cond.ite(
    f(cond.ite(a,  b), *fargs),
    g(cond.ite(a,  b), *gargs),
)
# as an optimization on RTL we will want to transform this to:
cond.ite(
    f(a, *fargs),
    g(b, *gargs),
)

This transformation really shouldn't be too hard as its so structured. Further simplifying things fargs will almost always be exactly the same gargs.

[rdaly525]

Before this can be merged, this should be run with lassen. I think there is a bug somewhere in the ALU that this will address.

[cdonovick]

@rdaly525 is that bug in rtl generation? Cause this won't fix anything with magma.

[rdaly525]

Just make sure that lassen unit tests still pass before merging this PR.

[cdonovick]

Passes all tests on lassen StanfordAHA/lassen#166