This can pave the way for further transformations by exposing
identities that were previously hidden in a process to any pass that
uses SigMap. Indeed, this commit removes some ad-hoc logic from
proc_init that appears to have been tailored to the output of
genrtlil in favor of using `SigMap.apply()`. (This removal is not
optional, as the ad-hoc logic cannot cope with the result of running
proc_prune; a similar issue was fixed in proc_arst.)
The proc_prune pass is similar in nature to proc_rmdead pass: while
proc_rmdead removes branches that never become active because another
branch preempts it, proc_prune removes assignments that never become
active because another assignment preempts them.
Genrtlil contains logic similar to the proc_prune pass, but their
purpose is different: genrtlil has to prune assignments to adapt
the semantics of blocking assignments in HDLs (latest assignment
wins) to semantics of assignments in RTLIL processes (assignment in
the most specific case wins). On the other hand proc_prune is
a general purpose RTLIL simplification that benefits all frontends,
even those not using the Yosys AST library.
The proc_prune pass is added to the proc script after proc_rmdead,
since it gives better results with fewer branches.
The parser changes are slightly awkward. Consider the following IL:
process $0
<point 1>
switch \foo
<point 2>
case 1'1
assign \bar \baz
<point 3>
...
case
end
end
Before this commit, attributes are valid in <point 1>, and <point 3>
iff it is immediately followed by a `switch`. (They are essentially
attached to the switch.) But, after this commit, and because switch
cases do not have an ending delimiter, <point 3> becomes ambiguous:
the attribute could attach to either the following `case`, or to
the following `switch`. This isn't expressible in LALR(1) and results
in a reduce/reduce conflict.
To address this, attributes inside processes are now valid anywhere
inside the process: in <point 1> and <point 3> a part of case body,
and in <point 2> as a separate rule. As a consequence, attributes
can now precede `assign`s, which is made illegal in the same way it
is illegal to attach attributes to `connect`.
Attributes are tracked separately from the parser state, so this
does not affect collection of attributes at all, other than allowing
them on `case`s. The grammar change serves purely to allow attributes
in more syntactic places.
Eddie Hung
David Shah
Clifford Wolf
whitequark
Eddie Hung
Dan Ravensloft
Benedikt Tutzer