[Arc] Add dominance-aware pass to sink ops and merge scf.if ops (#7702)

Add the `MergeIfs` pass to the Arc dialect. This pass covers a handful
of control flow optimizations that are valuable to pick up after
`hw.module`s have been linearized and lowered into `arc.model` ops:

- It moves operations closer to their earliest user, if possible sinking
  them into blocks if all uses are nested in the same block.
- It merges adjacent `scf.if` operations with the same condition.
- It moves operations in between two `scf.if` operations ahead of the
  first if op to allow them to be merged.

The `MergeIfs` pass can operate on SSACFG regions. It assigns an integer
order to each operation and considers that order to determine up to
which point operations can be moved without moving beyond their first
use and without crossing interfering side-effecting ops. The pass is
aware of side-effects, and in particular uses the non-aliasing between
`arc.state` and `arc.memory` to track read/write side-effects at a
per-state level, which allows for fairly aggressive optimization. Other
side-effecting ops act as a hard barrier and will not be moved or moved
over.

This pass supersedes the very effective `GroupResetsAndEnables` pass
we have been using until now. The latter relies on the
`LegalizeStateUpdate` pass to run at a later point however, which will
be removed in a future PR, thus making this new pass necessary.

This is a preparatory step for a later PR that overhauls the LowerState
pass. That rewrite will make the `arc.clock_tree` and `arc.passthrough`
ops obsolete, which is why they are not present in the tests.

include/circt/Dialect/Arc/ArcPasses.h

@@ -31,7 +31,6 @@ std::unique_ptr<mlir::Pass> createAllocateStatePass();
 std::unique_ptr<mlir::Pass> createArcCanonicalizerPass();
 std::unique_ptr<mlir::Pass> createDedupPass();
 std::unique_ptr<mlir::Pass> createFindInitialVectorsPass();
-std::unique_ptr<mlir::Pass> createGroupResetsAndEnablesPass();
 std::unique_ptr<mlir::Pass>
 createInferMemoriesPass(const InferMemoriesOptions &options = {});
 std::unique_ptr<mlir::Pass> createInlineArcsPass();

include/circt/Dialect/Arc/ArcPasses.td

@@ -92,13 +92,6 @@ def FindInitialVectors : Pass<"arc-find-initial-vectors", "mlir::ModuleOp"> {
   ];
 }
 
-def GroupResetsAndEnables : Pass<"arc-group-resets-and-enables",
-                                 "mlir::ModuleOp"> {
-  let summary = "Group reset and enable conditions of lowered states";
-  let constructor = "circt::arc::createGroupResetsAndEnablesPass()";
-  let dependentDialects = ["arc::ArcDialect", "mlir::scf::SCFDialect"];
-}
-
 def InferMemories : Pass<"arc-infer-memories", "mlir::ModuleOp"> {
   let summary = "Convert `FIRRTL_Memory` instances to dedicated memory ops";
   let constructor = "circt::arc::createInferMemoriesPass()";
@@ -282,6 +275,31 @@ def MakeTables : Pass<"arc-make-tables", "mlir::ModuleOp"> {
   let dependentDialects = ["arc::ArcDialect"];
 }
 
+def MergeIfsPass : Pass<"arc-merge-ifs"> {
+  let summary = "Merge control flow structures";
+  let description = [{
+    This pass optimizes control flow in a few ways. It moves operations closer
+    to their earliest user, if possible sinking them into blocks if all uses are
+    nested in the same block. It merges adjacent `scf.if` operations with the
+    same condition. And it moves operations in between two `scf.if` operations
+    ahead of the first if op to allow them to be merged. The pass runs on any
+    SSACFG regions nested under the operation it is applied to.
+
+    Note that this pass assumes that `!arc.state` and `!arc.memory` values can
+    never alias. That is, different values are assumed to never point to the
+    same storage location in simulation memory.
+  }];
+  let statistics = [
+    Statistic<"numOpsSunk", "sunk", "Ops sunk into blocks">,
+    Statistic<"numOpsMovedToUser", "moved-to-user", "Ops moved to first user">,
+    Statistic<"numIfsMerged", "ifs-merged", "Adjacent scf.if ops merged">,
+    Statistic<"numOpsMovedFromBetweenIfs", "moved-from-between-ifs",
+      "Ops moved from between ifs to enable merging">,
+    Statistic<"numIterations", "iterations",
+      "Number of iterations until no more ops were sunk/merged">,
+  ];
+}
+
 def MuxToControlFlow : Pass<"arc-mux-to-control-flow", "mlir::ModuleOp"> {
   let summary = "Convert muxes with large independent fan-ins to if-statements";
   let constructor = "circt::arc::createMuxToControlFlowPass()";

lib/Dialect/Arc/Transforms/CMakeLists.txt

@@ -4,7 +4,6 @@ add_circt_dialect_library(CIRCTArcTransforms
   ArcCanonicalizer.cpp
   Dedup.cpp
   FindInitialVectors.cpp
-  GroupResetsAndEnables.cpp
   InferMemories.cpp
   InferStateProperties.cpp
   InlineArcs.cpp
@@ -17,6 +16,7 @@ add_circt_dialect_library(CIRCTArcTransforms
   LowerState.cpp
   LowerVectorizations.cpp
   MakeTables.cpp
+  MergeIfs.cpp
   MuxToControlFlow.cpp
   PrintCostModel.cpp
   SimplifyVariadicOps.cpp