Vl-expr-svex-translation
Compilation from (sized) vl expressions into sv::svex expressions.
There are several top-level functions for converting a VL
expression into a sv::svex expression, including vl-expr-to-svex-untyped, vl-expr-to-svex-selfdet, and vl-expr-to-svex-maybe-typed.
We assume that the expressions we are dealing with are sized.
The Garbage Convention. Our goal is to produce svex expressions that
are "correct" at least for this size. We don't care about the bits that are
"past" the expression's final width.
For example, if we are converting 3'b100 into an svex expression, it
would be valid to just use (concat 3 #b100 anything).
This has widespread consequences.
In many cases, we don't have to worry about the upper bits, because we know,
for instance, that after sizing, in a VL expression like a & b, that the
widths of a and b agree with the final width of the whole a & b
expression. So any "garbage" in the upper bits can only affect the upper bits
of the result.
However, in certain cases we have to be more careful. For instance, if we
are translating a < b, then the svex semantics are that we do an infinite
width comparison, so we need to be careful to properly sign/zero extend the two
arguments to their appropriate, final value.
Why do it this way? Some possible alternatives:
- Force everything past the final width to X.
- Sign/zero extend everything so that every expression is correct out to
infinite width.
Forcing the high bits to X has some appeal: it would be nice to know that
any accidental use of the upper bits would be conservative. In most ways it
would be very similar to what we are doing now: we would still have to be
careful in cases like a < b to mask out those Xes. But we would also have
to do more to ensure that we were following this convention. For instance, by
not caring about the upper bits, we don't have to do anything special when
translating sums: the carry from a + b can go into the upper bits and it's
just garbage. If we wanted all the upper bits to be X, we'd need to do extra
work to mask them out in this case.
The other alternative seems sort of more proactive: we would essentially be
fixing everything up to be the right size whether it needs to be fixed up or
not. For instance, if we are translating something like (a + b) & c, we
don't really need to know the true upper bits of a + b or c. We
suspect that arranging for every expression to be correct out to infinite width
would therefore incur some overhead.
Subtopics
- Sv::vl-expr.lisp
- Vttree
- A data structure for collecting warnings and constraints
while translating VL expressions to svex expressions.