[mlir] add an option to print op stats in JSON

[okkwon]

Update to use a json object instead of an array

[okkwon]

The debian failure is caused by "OSError: [Errno 28] No space left on device".

[okkwon]

I reverted the commit because there is a test failure for the python pass manager. Looking into it.

[okkwon]

update the python pass manager test

[okkwon]

rebase to fix a build error for bazel

mlir/include/mlir/Transforms/Passes.td

 def LoopInvariantCodeMotion : Pass<"loop-invariant-code-motion"> {
   let summary = "Hoist loop invariant instructions outside of the loop";
   let constructor = "mlir::createLoopInvariantCodeMotionPass()";
 }
 
 def PrintOpStats : Pass<"print-op-stats"> {
   let summary = "Print statistics of operations";
   let constructor = "mlir::createPrintOpStatsPass()";
+  let options = [
+    Option<"printAsJSON", "json", "bool", /*default=*/"false",
+           "print the stats as JSON">
+  ];
 }
 
 def SCCP : Pass<"sccp"> {
   let summary = "Sparse Conditional Constant Propagation";
   let description = [{
     This pass implements a general algorithm for sparse conditional constant
     propagation. This algorithm detects values that are known to be constant and
     optimistically propagates this throughout the IR. Any values proven to be

mlir/lib/Transforms/OpStats.cpp

struct PrintOpStatsPass : public PrintOpStatsBase<PrintOpStatsPass> {
   explicit PrintOpStatsPass(raw_ostream &os) : os(os) {}
 
   // Prints the resultant operation statistics post iterating over the module.
   void runOnOperation() override;
 
   // Print summary of op stats.
   void printSummary();
 
+  // Print symmary of op stats in JSON.
+  void printSummaryInJSON();
+
 private:
   llvm::StringMap<int64_t> opCount;
   raw_ostream &os;
 };
 } // namespace
 
 void PrintOpStatsPass::runOnOperation() {
   opCount.clear();
 
   // Compute the operation statistics for the currently visited operation.
-  getOperation()->walk([&](Operation *op) { ++opCount[op->getName().getStringRef()]; });
+  getOperation()->walk(
+      [&](Operation *op) { ++opCount[op->getName().getStringRef()]; });
+  if (printAsJSON) {
+    printSummaryInJSON();
+  } else
     printSummary();
 }
 
 void PrintOpStatsPass::printSummary() {
   os << "Operations encountered:\n";
   os << "-----------------------\n";
   SmallVector<StringRef, 64> sorted(opCount.keys());
   llvm::sort(sorted);
 
else
       os << llvm::right_justify(dialectName, maxLenDialect + 2) << '.';
 
     // Left justify the operation name.
     os << llvm::left_justify(opName, maxLenOpName) << " , " << opCount[key]
        << '\n';
   }
 }
 
+void PrintOpStatsPass::printSummaryInJSON() {
+  SmallVector<StringRef, 64> sorted(opCount.keys());
+  llvm::sort(sorted);
+
+  os << "{\n";
+
+  for (unsigned i = 0, e = sorted.size(); i != e; ++i) {
+    const auto &key = sorted[i];
+    os << "  \"" << key << "\" : " << opCount[key];
+    if (i != e - 1)
+      os << ",\n";
+    else
+      os << "\n";
+  }
+  os << "}\n";
+}
+
 std::unique_ptr<Pass> mlir::createPrintOpStatsPass(raw_ostream &os) {
   return std::make_unique<PrintOpStatsPass>(os);
 }

mlir/test/CAPI/pass.c

void testPrintPassPipeline() {
   MlirOpPassManager nestedModulePm = mlirPassManagerGetNestedUnder(
       pm, mlirStringRefCreateFromCString("builtin.module"));
   MlirOpPassManager nestedFuncPm = mlirOpPassManagerGetNestedUnder(
       nestedModulePm, mlirStringRefCreateFromCString("func.func"));
   MlirPass printOpStatPass = mlirCreateTransformsPrintOpStats();
   mlirOpPassManagerAddOwnedPass(nestedFuncPm, printOpStatPass);
 
   // Print the top level pass manager
-  // CHECK: Top-level: builtin.module(func.func(print-op-stats))
+  // CHECK: Top-level: builtin.module(func.func(print-op-stats{json=false}))
   fprintf(stderr, "Top-level: ");
   mlirPrintPassPipeline(mlirPassManagerGetAsOpPassManager(pm), printToStderr,
                         NULL);
   fprintf(stderr, "\n");
 
   // Print the pipeline nested one level down
-  // CHECK: Nested Module: func.func(print-op-stats)
+  // CHECK: Nested Module: func.func(print-op-stats{json=false})
   fprintf(stderr, "Nested Module: ");
   mlirPrintPassPipeline(nestedModulePm, printToStderr, NULL);
   fprintf(stderr, "\n");
 
   // Print the pipeline nested two levels down
   // CHECK: Nested Module>Func: print-op-stats
   fprintf(stderr, "Nested Module>Func: ");
   mlirPrintPassPipeline(nestedFuncPm, printToStderr, NULL);
   fprintf(stderr, "\n");
 
   mlirPassManagerDestroy(pm);
   mlirContextDestroy(ctx);
 }
 
 void testParsePassPipeline() {
   MlirContext ctx = mlirContextCreate();
   MlirPassManager pm = mlirPassManagerCreate(ctx);
   // Try parse a pipeline.
   MlirLogicalResult status = mlirParsePassPipeline(
       mlirPassManagerGetAsOpPassManager(pm),
-      mlirStringRefCreateFromCString("builtin.module(func.func(print-op-stats),"
-                                     " func.func(print-op-stats))"));
+      mlirStringRefCreateFromCString(
+          "builtin.module(func.func(print-op-stats{json=false}),"
+          " func.func(print-op-stats{json=false}))"));
   // Expect a failure, we haven't registered the print-op-stats pass yet.
   if (mlirLogicalResultIsSuccess(status)) {
     fprintf(
         stderr,
         "Unexpected success parsing pipeline without registering the pass\n");
     exit(EXIT_FAILURE);
   }
   // Try again after registrating the pass.
   mlirRegisterTransformsPrintOpStats();
   status = mlirParsePassPipeline(
       mlirPassManagerGetAsOpPassManager(pm),
-      mlirStringRefCreateFromCString("builtin.module(func.func(print-op-stats),"
-                                     " func.func(print-op-stats))"));
+      mlirStringRefCreateFromCString(
+          "builtin.module(func.func(print-op-stats{json=false}),"
+          " func.func(print-op-stats{json=false}))"));
   // Expect a failure, we haven't registered the print-op-stats pass yet.
   if (mlirLogicalResultIsFailure(status)) {
     fprintf(stderr,
             "Unexpected failure parsing pipeline after registering the pass\n");
     exit(EXIT_FAILURE);
   }
 
-  // CHECK: Round-trip: builtin.module(func.func(print-op-stats),
-  // func.func(print-op-stats))
+  // CHECK: Round-trip: builtin.module(func.func(print-op-stats{json=false}),
+  // func.func(print-op-stats{json=false}))
   fprintf(stderr, "Round-trip: ");
   mlirPrintPassPipeline(mlirPassManagerGetAsOpPassManager(pm), printToStderr,
                         NULL);
   fprintf(stderr, "\n");
   mlirPassManagerDestroy(pm);
   mlirContextDestroy(ctx);
 }
 

mlir/test/IR/op-stats-json.mlir

+// RUN: mlir-opt -allow-unregistered-dialect -print-op-stats=json %s -o=/dev/null 2>&1 | FileCheck %s
+
+func.func @main(tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> {
+^bb0(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>):
+  %0 = arith.addf %arg0, %arg1 : tensor<4xf32>
+  %1 = arith.addf %arg0, %arg1 : tensor<4xf32>
+  %2 = arith.addf %arg0, %arg1 : tensor<4xf32>
+  %3 = arith.addf %arg0, %arg1 : tensor<4xf32>
+  %4 = arith.addf %arg0, %arg1 : tensor<4xf32>
+  %5 = arith.addf %arg0, %arg1 : tensor<4xf32>
+  %10 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %11 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %12 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %13 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %14 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %15 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %16 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %17 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %18 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %19 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %20 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %21 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %22 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %23 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %24 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %25 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %26 = "xla.add"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  %30 = "long_op_name"(%0, %arg1) : (tensor<4xf32>,tensor<4xf32>)-> tensor<4xf32>
+  return %1 : tensor<4xf32>
+}
+
+// CHECK: {
+// CHECK:   "arith.addf" : 6,
+// CHECK:   "func.return" : 1,
+// CHECK:   "long_op_name" : 1,
+// CHECK:   "xla.add" : 17
+// CHECK: }