[mlir][Pass] Fix dropped statistics with nested adaptors.

[dtzWill]

Summary

When running in parallel, nesting more than once caused
statistics to be dropped.

Fix by also preparing "async" pass managers before merging,
as they may also have "async" pass managers within.

Add test checking reported statistics have expected values
with and without threading enabled.

[dtzWill]

This fixes an issue in the CIRCT project where virtually none of our statistics worked with threading enabled: https://github.com/llvm/circt/issues/3513 .

As an aside, the test was based on pipeline-stats.mlir which I do not believe still needs to check for "0" values in the reported statistics, as now they're either enabled+reported or not reported at all (and the "REQUIRES: asserts" ensures they're enabled).

mlir/lib/Pass/PassStatistics.cpp

 static void prepareStatistics(OpPassManager &pm) {
   for (Pass &pass : pm.getPasses()) {
     OpToOpPassAdaptor *adaptor = dyn_cast<OpToOpPassAdaptor>(&pass);
     if (!adaptor)
       continue;
     MutableArrayRef<OpPassManager> nestedPms = adaptor->getPassManagers();
 
     // Merge the statistics from the async pass managers into the main nested
-    // pass managers.
+    // pass managers.  Prepare recursively before merging.
     for (auto &asyncPM : adaptor->getParallelPassManagers()) {
-      for (unsigned i = 0, e = asyncPM.size(); i != e; ++i)
+      for (unsigned i = 0, e = asyncPM.size(); i != e; ++i) {
+        prepareStatistics(asyncPM[i]);
         asyncPM[i].mergeStatisticsInto(nestedPms[i]);
       }
+    }
 
     // Prepare the statistics of each of the nested passes.
     for (OpPassManager &nestedPM : nestedPms)
       prepareStatistics(nestedPM);
   }
 }
 
 /// Dump the statistics of the passes within this pass manager.

mlir/test/Pass/pipeline-stats-nested.mlir

+// REQUIRES: asserts
+// Check that statistics in nested pipelines are not ignored and that this works with and without threading.
+// RUN: mlir-opt %s -verify-each=true -pass-pipeline='builtin.module(builtin.module(func.func(test-stats-pass,test-stats-pass)))' -mlir-pass-statistics -mlir-pass-statistics-display=list -mlir-disable-threading 2>&1 | FileCheck -check-prefix=LIST %s
+// RUN: mlir-opt %s -verify-each=true -pass-pipeline='builtin.module(builtin.module(func.func(test-stats-pass,test-stats-pass)))' -mlir-pass-statistics -mlir-pass-statistics-display=list                         2>&1 | FileCheck -check-prefix=LIST %s
+
+// RUN: mlir-opt %s -verify-each=true -pass-pipeline='builtin.module(builtin.module(func.func(test-stats-pass,test-stats-pass)))' -mlir-pass-statistics -mlir-pass-statistics-display=pipeline -mlir-disable-threading 2>&1 | FileCheck -check-prefix=PIPELINE %s
+// RUN: mlir-opt %s -verify-each=true -pass-pipeline='builtin.module(builtin.module(func.func(test-stats-pass,test-stats-pass)))' -mlir-pass-statistics -mlir-pass-statistics-display=pipeline                         2>&1 | FileCheck -check-prefix=PIPELINE %s
+
+// Check for the correct statistic values.
+// Each test-stats-pass will count two ops: the func.func and the func.return .
+//      LIST: (S) 4 num-ops
+// LIST-NEXT: (S) 4 num-ops2
+//      PIPELINE: (S) 2 num-ops
+// PIPELINE-NEXT: (S) 2 num-ops2
+//      PIPELINE: (S) 2 num-ops
+// PIPELINE-NEXT: (S) 2 num-ops2
+
+module {
+  module {
+    func.func @foo() {
+      return
+    }
+  }
+}