[LLHD] Update Deseq pass to work with process results

Add an updated version of the desequentialization pass. This new version
expects all probes and drives to have been hoisted out of processes, and
for the processes to produce the interesting signals as a result. Resets
can now be distinguished from clocks more reliably. The previous version
of the pass is left untouched for now, since it is used in the
circt-verilog pipeline. Once process lowering also works with process
results, we can get rid of the old pass implementations in one go.

This pass is an absolute headache to implement because of the fuzzy and
ill-defined Verilog and VHDL semantics it has to match. I expect us to
iterate heavily on this in the future, along the definition of
`llhd.process` and potentially process flavors that are more convenient
to analyze.

In particular, the pass has to do a lot of things in one go: it checks
whether a process has only a single wait, if the wait carries past
values into the present as block arguments, whether the IR expresses a
posedge/negedge detection based on past and present values, whether the
process properly observes the necessary triggers values, and it has to
match the detected triggers against a very narrow list of things that a
register can represent.

Life would be easier if we had a more restricted form of processes that
already express the basic structure of a loop through a single wait,
observed triggers, and which past values are required. This would then
also allow us to detect and generate latches, which are missing
completely right now.

The pass currently performs a data flow analysis by propagating DNFs and
tables of potential result values across a lattice. I'm not too happy
with this implementation, and doing a depth-first traversal of the IR
may be a better approach in the future. We should probably revisit this
once we have improved the process op a bit and update this pass again.

This commit also adds an implementation of the Disjunctive Normal Form
as a way of expressing and manipulating boolean functions. This is
necessary to reliably detect if a process computes the posedge or
negedge of a value. In the future we may want to switch this to a more
compact representation of the underlying truth table, which can be done
fairly efficiently using APInt. The current implementation of
`DNF::optimize()` already does this internally. Note that DNFs are very
inefficient at representing arbitrary boolean functions, as they have a
tendency to grow exponentially. Therefore care must be taken that only
operations on potential trigger values of sequential processes are
captured in the DNF, and all other values are kept as opaque values.
This keeps the complexity bounded.

Author: fabianschuiki

include/circt/Dialect/LLHD/IR/LLHDStructureOps.td

@@ -91,6 +91,7 @@ def FinalOp : LLHDOp<"final", [
 def WaitOp : LLHDOp<"wait", [
   AttrSizedOperandSegments,
   HasParent<"ProcessOp">,
+  Pure,
   Terminator,
 ]> {
   let summary = "Suspend execution of a process";
@@ -131,6 +132,7 @@ def WaitOp : LLHDOp<"wait", [
 
 def HaltOp : LLHDOp<"halt", [
   ParentOneOf<["ProcessOp", "FinalOp"]>,
+  Pure,
   Terminator,
 ]> {
   let summary = "Terminate execution of a process";

include/circt/Dialect/LLHD/Transforms/LLHDPasses.td

@@ -162,4 +162,14 @@ def LowerProcessesPass : Pass<"llhd-lower-processes", "hw::HWModuleOp"> {
   ];
 }
 
+def DeseqPass : Pass<"llhd-deseq", "hw::HWModuleOp"> {
+  let summary = "Convert sequential processes to registers";
+  let dependentDialects = [
+    "comb::CombDialect",
+    "hw::HWDialect",
+    "mlir::scf::SCFDialect",
+    "seq::SeqDialect",
+  ];
+}
+
 #endif // CIRCT_DIALECT_LLHD_TRANSFORMS_PASSES

lib/Dialect/LLHD/Transforms/CMakeLists.txt

@@ -1,5 +1,7 @@
 add_circt_dialect_library(CIRCTLLHDTransforms
+  Deseq.cpp
   DesequentializationPass.cpp
+  DNF.cpp
   EarlyCodeMotionPass.cpp
   FunctionEliminationPass.cpp
   HoistSignals.cpp