pkg/front_end/test/compile_benchmark.dart - sdk.git - Git at Google

 import 'dart:convert';
 import 'dart:io';
 import 'dart:typed_data';

 import 'package:front_end/src/fasta/kernel/utils.dart';
 import 'package:kernel/kernel.dart';
 import 'package:kernel/binary/ast_from_binary.dart';

 import "simple_stats.dart";

 /// This pass all parameters after "--" to the compiler (via `compileEntryPoint`
 /// via the helper `compile_benchmark_helper.dart`). It is meant to "benchmark"/
 /// instrument the compiler to give insights into specific runs, i.e. specific
 /// compiles (compiling program `a` and program `b` might have different
 /// characteristics). Said another way, it will always instrument and give
 /// insights on *the compiler*, not the program it's asked to compile (not
 /// directly anyway -- if you want to know if that program successfully
 /// exercises a specific part of the compiler it _does_ do that).

 final Uri benchmarkHelper =
     Platform.script.resolve("compile_benchmark_helper.dart");

 void main(List<String> args) {
   List<String>? arguments;
   bool tryToAnnotate = false;
   bool tryToSlowDown = false;
   bool timeInsteadOfCount = false;
   for (int i = 0; i < args.length; i++) {
     if (args[i] == "--tryToAnnotate") {
       tryToAnnotate = true;
     } else if (args[i] == "--tryToSlowDown") {
       tryToSlowDown = true;
     } else if (args[i] == "--timeInsteadOfCount") {
       timeInsteadOfCount = true;
     } else if (args[i] == "--") {
       arguments = args.sublist(i + 1);
       break;
     } else {
       throw "Unknown argument '${args[i]}'";
     }
   }
   if (arguments == null || arguments.isEmpty) {
     throw "No arguments given to compiler.\n"
         "Give arguments as `dart compile_benchmark.dart -- "
         "argument1 argument2 (etc)`";
   }

   Directory tmp = Directory.systemTemp.createTempSync("benchmark");
   try {
     // Compile the helper to get a dill we can compile with.
     final Uri helperDill = tmp.uri.resolve("compile.dill");

     print("Compiling $benchmarkHelper into $helperDill");
     runXTimes(1, [
       benchmarkHelper.toString(),
       benchmarkHelper.toString(),
       "-o",
       helperDill.toString(),
     ]);
     File f = new File.fromUri(helperDill);
     if (!f.existsSync()) throw "$f doesn't exist!";

     List<int> dillData = new File.fromUri(helperDill).readAsBytesSync();
     doWork(
       tmp,
       dillData,
       arguments,
       tryToAnnotate: tryToAnnotate,
       tryToSlowDown: tryToSlowDown,
       timeInsteadOfCount: timeInsteadOfCount,
     );
   } finally {
     tmp.deleteSync(recursive: true);
   }
 }

 /// Perform asked operations on the dill data provided.
 ///
 /// Will save files into the provided tmp directory, do timing or counting
 /// instructions and run that with the arguments passed, and then do comparative
 /// runs of transformed/non-transformed versions of the dill. The possible
 /// transformations are:
 /// * [tryToAnnotate] which will add a `@pragma("vm:prefer-inline")` annotation
 ///   to each procedure one at a time.
 /// * [tryToSlowDown] which will add a busywait for approximately 0.002 ms to
 ///   each procedure one at a time. (Unsurprisingly this makes it slower
 ///   proportional to the number of times the procedure is called, so the
 ///   counting annotation is likely at least as useful).
 ///
 void doWork(Directory tmp, List<int> dillData, List<String> arguments,
     {bool tryToAnnotate: false,
     bool tryToSlowDown: false,
     bool timeInsteadOfCount: false}) {
   File f = new File.fromUri(tmp.uri.resolve("a.dill"));
   f.writeAsBytesSync(dillData);
   Uri dillOrgInTmp = f.uri;
   print("Wrote to $f");

   List<Procedure> sortedProcedures;
   if (timeInsteadOfCount) {
     sortedProcedures = doTimingInstrumentation(dillData, tmp, arguments);
   } else {
     sortedProcedures = doCountingInstrumentation(dillData, tmp, arguments);
   }
   print("\n\n");

   bool didSomething = false;

   if (tryToAnnotate) {
     didSomething = true;
     for (Procedure p in sortedProcedures) {
       print("Prefer inline $p (${p.location})");
       Uri preferredInlined = preferInlineProcedure(
           dillData,
           tmp.uri,
           (lib) => lib.importUri == p.enclosingLibrary.importUri,
           p.enclosingClass?.name,
           p.name.text);

       print("\nOriginal runs:");
       List<int> runtimesA =
           runXTimes(5, [dillOrgInTmp.toString(), ...arguments]);

       print("\nModified runs:");
       List<int> runtimesB =
           runXTimes(5, [preferredInlined.toString(), ...arguments]);

       print(SimpleTTestStat.ttest(runtimesB, runtimesA));
       print("\n------------\n");
     }
   }

   if (tryToSlowDown) {
     didSomething = true;
     for (Procedure p in sortedProcedures) {
       Uri? busyWaiting = busyWaitProcedure(
           dillData,
           tmp.uri,
           (lib) => lib.importUri == p.enclosingLibrary.importUri,
           p.enclosingClass?.name,
           p.name.text);
       if (busyWaiting == null) continue;

       print("Slow down $p (${p.location})");

       print("\nOriginal runs:");
       List<int> runtimesA =
           runXTimes(2, [dillOrgInTmp.toString(), ...arguments]);

       print("\nModified runs:");
       List<int> runtimesB =
           runXTimes(2, [busyWaiting.toString(), ...arguments]);

       print(SimpleTTestStat.ttest(runtimesB, runtimesA));
       print("\n------------\n");
     }
   }

   if (!didSomething) {
     runXTimes(10, [dillOrgInTmp.toString(), ...arguments]);
   }
 }

 /// Instrument the [dillData] so that each procedure-call can be registered
 /// (in package:front_end) and we find out how many times each procedure is
 /// called for a specific run, then run it, print the result and return the
 /// procedures in sorted order (most calls first).
 List<Procedure> doCountingInstrumentation(
     List<int> dillData, Directory tmp, List<String> arguments) {
   Instrumented instrumented = instrumentCallsCount(dillData, tmp.uri);
   List<dynamic> stdout = [];
   runXTimes(1, [instrumented.dill.toString(), ...arguments], stdout);
   List<int> procedureCountsTmp = new List<int>.from(jsonDecode(stdout.single));
   List<IntPair> procedureCounts = [];
   for (int i = 0; i < procedureCountsTmp.length; i += 2) {
     procedureCounts
         .add(new IntPair(procedureCountsTmp[i], procedureCountsTmp[i + 1]));
   }
   // Sort highest call-count first.
   procedureCounts.sort((a, b) => b.value - a.value);
   List<Procedure> sortedProcedures = [];
   for (IntPair p in procedureCounts) {
     if (p.value > 1000) {
       Procedure procedure = instrumented.procedures[p.key];
       String location = procedure.location.toString();
       if (location.length > 50) {
         location = location.substring(location.length - 50);
       }
       print("Called $procedure ${p.value} times ($location)");
       sortedProcedures.add(procedure);
     }
   }
   return sortedProcedures;
 }

 /// Instrument the [dillData] so that each (sync) procedure-call can be timed
 /// (time on stack, i.e. not only the procedure itself, but also the
 /// procedure-calls it makes) (in package:front_end) and we find out how long
 /// each procedure is on the stack for a specific run, then run it, print the
 /// result and return the procedures in sorted order (most time on stack first).
 List<Procedure> doTimingInstrumentation(
     List<int> dillData, Directory tmp, List<String> arguments) {
   Instrumented instrumented = instrumentCallsTiming(dillData, tmp.uri);
   List<dynamic> stdout = [];
   runXTimes(1, [instrumented.dill.toString(), ...arguments], stdout);
   List<int> procedureTimeTmp = new List<int>.from(jsonDecode(stdout.single));
   List<IntPair> procedureTime = [];
   for (int i = 0; i < procedureTimeTmp.length; i += 2) {
     procedureTime
         .add(new IntPair(procedureTimeTmp[i], procedureTimeTmp[i + 1]));
   }
   // Sort highest time-on-stack first.
   procedureTime.sort((a, b) => b.value - a.value);
   List<Procedure> sortedProcedures = [];
   for (IntPair p in procedureTime) {
     if (p.value > 1000) {
       Procedure procedure = instrumented.procedures[p.key];
       String location = procedure.location.toString();
       if (location.length > 50) {
         location = location.substring(location.length - 50);
       }
       print("$procedure was on stack for ${p.value} microseconds ($location)");
       sortedProcedures.add(procedure);
     }
   }
   return sortedProcedures;
 }

 class IntPair {
   final int key;
   final int value;

   IntPair(this.key, this.value);

   @override
   String toString() {
     return "IntPair[$key: $value]";
   }
 }

 /// Adds the annotation `@pragma("vm:prefer-inline")` to the specified procedure
 /// and serialize the resulting dill into `b.dill` (return uri).
 ///
 /// The annotation is copied from the [preferInlineMe] method in the helper.
 Uri preferInlineProcedure(List<int> dillData, Uri tmp,
     bool libraryMatcher(Library lib), String? className, String procedureName) {
   Component component = new Component();
   new BinaryBuilder(dillData, disableLazyReading: true)
       .readComponent(component);
   Procedure preferInlineMeProcedure = getProcedure(component,
       (lib) => lib.fileUri == benchmarkHelper, null, "preferInlineMe");
   ConstantExpression annotation =
       preferInlineMeProcedure.annotations.single as ConstantExpression;
   Procedure markProcedure =
       getProcedure(component, libraryMatcher, className, procedureName);
   markProcedure.addAnnotation(
       new ConstantExpression(annotation.constant, annotation.type));

   Uint8List newDillData = serializeComponent(component);
   File f = new File.fromUri(tmp.resolve("b.dill"));
   f.writeAsBytesSync(newDillData);
   return f.uri;
 }

 /// Makes the procedure specified call [busyWait] from the helper and serialize
 /// the resulting dill into `c.dill` (return uri).
 ///
 /// This will make the procedure busy-wait approximately 0.002 ms for each
 /// invocation (+ whatever overhead and imprecision).
 Uri? busyWaitProcedure(List<int> dillData, Uri tmp,
     bool libraryMatcher(Library lib), String? className, String procedureName) {
   Component component = new Component();
   new BinaryBuilder(dillData, disableLazyReading: true)
       .readComponent(component);
   Procedure busyWaitProcedure = getProcedure(
       component, (lib) => lib.fileUri == benchmarkHelper, null, "busyWait");

   Procedure markProcedure =
       getProcedure(component, libraryMatcher, className, procedureName);
   if (markProcedure.function.body == null) return null;

   Statement orgBody = markProcedure.function.body as Statement;
   markProcedure.function.body = new Block([
     new ExpressionStatement(new StaticInvocation(
         busyWaitProcedure, new Arguments([new IntLiteral(2 /* 0.002 ms */)]))),
     orgBody
   ])
     ..parent = markProcedure.function;

   Uint8List newDillData = serializeComponent(component);
   File f = new File.fromUri(tmp.resolve("c.dill"));
   f.writeAsBytesSync(newDillData);
   return f.uri;
 }

 /// Instrument the [dillData] so that each procedure-call can be registered
 /// (in package:front_end) and we find out how many times each procedure is
 /// called for a specific run.
 ///
 /// Uses the [registerCall] in the helper.
 /// Numbers each procedure, saves the instrumented dill and returns both the
 /// dill and the list of procedures so that procedure i in the list will be
 /// annotated with a call to `registerCall(i)`.
 Instrumented instrumentCallsCount(List<int> dillData, Uri tmp) {
   Component component = new Component();
   new BinaryBuilder(dillData, disableLazyReading: true)
       .readComponent(component);
   Procedure registerCallProcedure = getProcedure(
       component, (lib) => lib.fileUri == benchmarkHelper, null, "registerCall");
   RegisterCallTransformer registerCallTransformer =
       new RegisterCallTransformer(registerCallProcedure);
   component.accept(registerCallTransformer);

   Uint8List newDillData = serializeComponent(component);
   File f = new File.fromUri(tmp.resolve("counting.dill"));
   f.writeAsBytesSync(newDillData);

   return new Instrumented(f.uri, registerCallTransformer.procedures);
 }

 /// Instrument the [dillData] so that each (sync) procedure-call can be timed
 /// (time on stack, i.e. not only the procedure itself, but also the
 /// procedure-calls it makes) (in package:front_end) and we find out how long
 /// each procedure is on the stack for a specific run.
 ///
 /// Uses [registerCallStart] and [registerCallEnd] from the helper.
 /// Numbers each sync procedure, saves the instrumented dill and returns both
 /// the dill and the list of procedures so that procedure i in the list will be
 /// annotated with a call-pair to `registerCallStart(i)` and
 /// `registerCallEnd(i)`.
 Instrumented instrumentCallsTiming(List<int> dillData, Uri tmp) {
   Component component = new Component();
   new BinaryBuilder(dillData, disableLazyReading: true)
       .readComponent(component);
   Procedure registerCallStartProcedure = getProcedure(component,
       (lib) => lib.fileUri == benchmarkHelper, null, "registerCallStart");
   Procedure registerCallEndProcedure = getProcedure(component,
       (lib) => lib.fileUri == benchmarkHelper, null, "registerCallEnd");
   RegisterTimeTransformer registerTimeTransformer = new RegisterTimeTransformer(
       registerCallStartProcedure, registerCallEndProcedure);
   component.accept(registerTimeTransformer);

   Uint8List newDillData = serializeComponent(component);
   File f = new File.fromUri(tmp.resolve("timing.dill"));
   f.writeAsBytesSync(newDillData);

   return new Instrumented(f.uri, registerTimeTransformer.procedures);
 }

 /// Class holding both the uri of a saved dill file and a list of procedures (in
 /// order) that has some reference to the annotation added to the dill file.
 class Instrumented {
   final Uri dill;
   final List<Procedure> procedures;

   Instrumented(this.dill, this.procedures);
 }

 class RegisterCallTransformer extends RecursiveVisitor {
   final Procedure registerCallProcedure;
   RegisterCallTransformer(this.registerCallProcedure);
   List<Procedure> procedures = [];

   @override
   void visitLibrary(Library node) {
     if (node.importUri.isScheme("package") &&
         node.importUri.pathSegments.first == "front_end") {
       super.visitLibrary(node);
     }
   }

   @override
   void visitProcedure(Procedure node) {
     if (node.function.body == null) return;
     int procedureNum = procedures.length;
     procedures.add(node);
     Statement orgBody = node.function.body as Statement;
     node.function.body = new Block([
       new ExpressionStatement(new StaticInvocation(registerCallProcedure,
           new Arguments([new IntLiteral(procedureNum)]))),
       orgBody
     ]);
     node.function.body!.parent = node.function;
   }
 }

 class RegisterTimeTransformer extends RecursiveVisitor {
   final Procedure registerCallStartProcedure;
   final Procedure registerCallEndProcedure;

   RegisterTimeTransformer(
       this.registerCallStartProcedure, this.registerCallEndProcedure);

   List<Procedure> procedures = [];

   @override
   void visitLibrary(Library node) {
     if (node.importUri.isScheme("package") &&
         node.importUri.pathSegments.first == "front_end") {
       super.visitLibrary(node);
     }
   }

   @override
   void visitProcedure(Procedure node) {
     if (node.function.body == null) return;
     if (node.function.dartAsyncMarker != AsyncMarker.Sync) return;
     int procedureNum = procedures.length;
     procedures.add(node);
     Statement orgBody = node.function.body as Statement;
     // Rewrite as
     // {
     //    registerCallStartProcedure(x);
     //    try {
     //      originalBody
     //    } finally {
     //      registerCallEndProcedure(x);
     //    }
     // }
     Block block = new Block([
       new ExpressionStatement(new StaticInvocation(registerCallStartProcedure,
           new Arguments([new IntLiteral(procedureNum)]))),
       new TryFinally(
         orgBody,
         new ExpressionStatement(new StaticInvocation(registerCallEndProcedure,
             new Arguments([new IntLiteral(procedureNum)]))),
       )
     ]);
     node.function.body = block;
     node.function.body!.parent = node.function;
   }
 }

 Procedure getProcedure(Component component, bool libraryMatcher(Library lib),
     String? className, String procedureName) {
   Library lib = component.libraries.where(libraryMatcher).single;
   List<Procedure> procedures = lib.procedures;
   if (className != null) {
     Class cls = lib.classes.where((c) => c.name == className).single;
     procedures = cls.procedures;
   }
   // TODO: This will fail for getter/setter pairs. Fix that.
   return procedures.where((p) => p.name.text == procedureName).single;
 }

 List<int> runXTimes(int x, List<String> arguments, [List<dynamic>? stdout]) {
   List<int> result = [];
   Stopwatch stopwatch = new Stopwatch()..start();
   for (int i = 0; i < x; i++) {
     stopwatch.reset();
     ProcessResult run = Process.runSync(Platform.resolvedExecutable, arguments,
         runInShell: true);
     int ms = stopwatch.elapsedMilliseconds;
     result.add(ms);
     print(ms);
     if (run.exitCode != 0) throw "Got exit code ${run.exitCode}";
     if (stdout != null) {
       stdout.add(run.stdout);
     }
   }
   return result;
 }
	import 'dart:convert';
	import 'dart:io';
	import 'dart:typed_data';

	import 'package:front_end/src/fasta/kernel/utils.dart';
	import 'package:kernel/kernel.dart';
	import 'package:kernel/binary/ast_from_binary.dart';

	import "simple_stats.dart";

	/// This pass all parameters after "--" to the compiler (via `compileEntryPoint`
	/// via the helper `compile_benchmark_helper.dart`). It is meant to "benchmark"/
	/// instrument the compiler to give insights into specific runs, i.e. specific
	/// compiles (compiling program `a` and program `b` might have different
	/// characteristics). Said another way, it will always instrument and give
	/// insights on the compiler, not the program it's asked to compile (not
	/// directly anyway -- if you want to know if that program successfully
	/// exercises a specific part of the compiler it _does_ do that).

	final Uri benchmarkHelper =
	Platform.script.resolve("compile_benchmark_helper.dart");

	void main(List<String> args) {
	List<String>? arguments;
	bool tryToAnnotate = false;
	bool tryToSlowDown = false;
	bool timeInsteadOfCount = false;
	for (int i = 0; i < args.length; i++) {
	if (args[i] == "--tryToAnnotate") {
	tryToAnnotate = true;
	} else if (args[i] == "--tryToSlowDown") {
	tryToSlowDown = true;
	} else if (args[i] == "--timeInsteadOfCount") {
	timeInsteadOfCount = true;
	} else if (args[i] == "--") {
	arguments = args.sublist(i + 1);
	break;
	} else {
	throw "Unknown argument '${args[i]}'";
	}
	}
	if (arguments == null \|\| arguments.isEmpty) {
	throw "No arguments given to compiler.\n"
	"Give arguments as `dart compile_benchmark.dart -- "
	"argument1 argument2 (etc)`";
	}

	Directory tmp = Directory.systemTemp.createTempSync("benchmark");
	try {
	// Compile the helper to get a dill we can compile with.
	final Uri helperDill = tmp.uri.resolve("compile.dill");

	print("Compiling $benchmarkHelper into $helperDill");
	runXTimes(1, [
	benchmarkHelper.toString(),
	benchmarkHelper.toString(),
	"-o",
	helperDill.toString(),
	]);
	File f = new File.fromUri(helperDill);
	if (!f.existsSync()) throw "$f doesn't exist!";

	List<int> dillData = new File.fromUri(helperDill).readAsBytesSync();
	doWork(
	tmp,
	dillData,
	arguments,
	tryToAnnotate: tryToAnnotate,
	tryToSlowDown: tryToSlowDown,
	timeInsteadOfCount: timeInsteadOfCount,
	);
	} finally {
	tmp.deleteSync(recursive: true);
	}
	}

	/// Perform asked operations on the dill data provided.
	///
	/// Will save files into the provided tmp directory, do timing or counting
	/// instructions and run that with the arguments passed, and then do comparative
	/// runs of transformed/non-transformed versions of the dill. The possible
	/// transformations are:
	/// * [tryToAnnotate] which will add a `@pragma("vm:prefer-inline")` annotation
	/// to each procedure one at a time.
	/// * [tryToSlowDown] which will add a busywait for approximately 0.002 ms to
	/// each procedure one at a time. (Unsurprisingly this makes it slower
	/// proportional to the number of times the procedure is called, so the
	/// counting annotation is likely at least as useful).
	///
	void doWork(Directory tmp, List<int> dillData, List<String> arguments,
	{bool tryToAnnotate: false,
	bool tryToSlowDown: false,
	bool timeInsteadOfCount: false}) {
	File f = new File.fromUri(tmp.uri.resolve("a.dill"));
	f.writeAsBytesSync(dillData);
	Uri dillOrgInTmp = f.uri;
	print("Wrote to $f");

	List<Procedure> sortedProcedures;
	if (timeInsteadOfCount) {
	sortedProcedures = doTimingInstrumentation(dillData, tmp, arguments);
	} else {
	sortedProcedures = doCountingInstrumentation(dillData, tmp, arguments);
	}
	print("\n\n");

	bool didSomething = false;

	if (tryToAnnotate) {
	didSomething = true;
	for (Procedure p in sortedProcedures) {
	print("Prefer inline $p (${p.location})");
	Uri preferredInlined = preferInlineProcedure(
	dillData,
	tmp.uri,
	(lib) => lib.importUri == p.enclosingLibrary.importUri,
	p.enclosingClass?.name,
	p.name.text);

	print("\nOriginal runs:");
	List<int> runtimesA =
	runXTimes(5, [dillOrgInTmp.toString(), ...arguments]);

	print("\nModified runs:");
	List<int> runtimesB =
	runXTimes(5, [preferredInlined.toString(), ...arguments]);

	print(SimpleTTestStat.ttest(runtimesB, runtimesA));
	print("\n------------\n");
	}
	}

	if (tryToSlowDown) {
	didSomething = true;
	for (Procedure p in sortedProcedures) {
	Uri? busyWaiting = busyWaitProcedure(
	dillData,
	tmp.uri,
	(lib) => lib.importUri == p.enclosingLibrary.importUri,
	p.enclosingClass?.name,
	p.name.text);
	if (busyWaiting == null) continue;

	print("Slow down $p (${p.location})");

	print("\nOriginal runs:");
	List<int> runtimesA =
	runXTimes(2, [dillOrgInTmp.toString(), ...arguments]);

	print("\nModified runs:");
	List<int> runtimesB =
	runXTimes(2, [busyWaiting.toString(), ...arguments]);

	print(SimpleTTestStat.ttest(runtimesB, runtimesA));
	print("\n------------\n");
	}
	}

	if (!didSomething) {
	runXTimes(10, [dillOrgInTmp.toString(), ...arguments]);
	}
	}

	/// Instrument the [dillData] so that each procedure-call can be registered
	/// (in package:front_end) and we find out how many times each procedure is
	/// called for a specific run, then run it, print the result and return the
	/// procedures in sorted order (most calls first).
	List<Procedure> doCountingInstrumentation(
	List<int> dillData, Directory tmp, List<String> arguments) {
	Instrumented instrumented = instrumentCallsCount(dillData, tmp.uri);
	List<dynamic> stdout = [];
	runXTimes(1, [instrumented.dill.toString(), ...arguments], stdout);
	List<int> procedureCountsTmp = new List<int>.from(jsonDecode(stdout.single));
	List<IntPair> procedureCounts = [];
	for (int i = 0; i < procedureCountsTmp.length; i += 2) {
	procedureCounts
	.add(new IntPair(procedureCountsTmp[i], procedureCountsTmp[i + 1]));
	}
	// Sort highest call-count first.
	procedureCounts.sort((a, b) => b.value - a.value);
	List<Procedure> sortedProcedures = [];
	for (IntPair p in procedureCounts) {
	if (p.value > 1000) {
	Procedure procedure = instrumented.procedures[p.key];
	String location = procedure.location.toString();
	if (location.length > 50) {
	location = location.substring(location.length - 50);
	}
	print("Called $procedure ${p.value} times ($location)");
	sortedProcedures.add(procedure);
	}
	}
	return sortedProcedures;
	}

	/// Instrument the [dillData] so that each (sync) procedure-call can be timed
	/// (time on stack, i.e. not only the procedure itself, but also the
	/// procedure-calls it makes) (in package:front_end) and we find out how long
	/// each procedure is on the stack for a specific run, then run it, print the
	/// result and return the procedures in sorted order (most time on stack first).
	List<Procedure> doTimingInstrumentation(
	List<int> dillData, Directory tmp, List<String> arguments) {
	Instrumented instrumented = instrumentCallsTiming(dillData, tmp.uri);
	List<dynamic> stdout = [];
	runXTimes(1, [instrumented.dill.toString(), ...arguments], stdout);
	List<int> procedureTimeTmp = new List<int>.from(jsonDecode(stdout.single));
	List<IntPair> procedureTime = [];
	for (int i = 0; i < procedureTimeTmp.length; i += 2) {
	procedureTime
	.add(new IntPair(procedureTimeTmp[i], procedureTimeTmp[i + 1]));
	}
	// Sort highest time-on-stack first.
	procedureTime.sort((a, b) => b.value - a.value);
	List<Procedure> sortedProcedures = [];
	for (IntPair p in procedureTime) {
	if (p.value > 1000) {
	Procedure procedure = instrumented.procedures[p.key];
	String location = procedure.location.toString();
	if (location.length > 50) {
	location = location.substring(location.length - 50);
	}
	print("$procedure was on stack for ${p.value} microseconds ($location)");
	sortedProcedures.add(procedure);
	}
	}
	return sortedProcedures;
	}

	class IntPair {
	final int key;
	final int value;

	IntPair(this.key, this.value);

	@override
	String toString() {
	return "IntPair[$key: $value]";
	}
	}

	/// Adds the annotation `@pragma("vm:prefer-inline")` to the specified procedure
	/// and serialize the resulting dill into `b.dill` (return uri).
	///
	/// The annotation is copied from the [preferInlineMe] method in the helper.
	Uri preferInlineProcedure(List<int> dillData, Uri tmp,
	bool libraryMatcher(Library lib), String? className, String procedureName) {
	Component component = new Component();
	new BinaryBuilder(dillData, disableLazyReading: true)
	.readComponent(component);
	Procedure preferInlineMeProcedure = getProcedure(component,
	(lib) => lib.fileUri == benchmarkHelper, null, "preferInlineMe");
	ConstantExpression annotation =
	preferInlineMeProcedure.annotations.single as ConstantExpression;
	Procedure markProcedure =
	getProcedure(component, libraryMatcher, className, procedureName);
	markProcedure.addAnnotation(
	new ConstantExpression(annotation.constant, annotation.type));

	Uint8List newDillData = serializeComponent(component);
	File f = new File.fromUri(tmp.resolve("b.dill"));
	f.writeAsBytesSync(newDillData);
	return f.uri;
	}

	/// Makes the procedure specified call [busyWait] from the helper and serialize
	/// the resulting dill into `c.dill` (return uri).
	///
	/// This will make the procedure busy-wait approximately 0.002 ms for each
	/// invocation (+ whatever overhead and imprecision).
	Uri? busyWaitProcedure(List<int> dillData, Uri tmp,
	bool libraryMatcher(Library lib), String? className, String procedureName) {
	Component component = new Component();
	new BinaryBuilder(dillData, disableLazyReading: true)
	.readComponent(component);
	Procedure busyWaitProcedure = getProcedure(
	component, (lib) => lib.fileUri == benchmarkHelper, null, "busyWait");

	Procedure markProcedure =
	getProcedure(component, libraryMatcher, className, procedureName);
	if (markProcedure.function.body == null) return null;

	Statement orgBody = markProcedure.function.body as Statement;
	markProcedure.function.body = new Block([
	new ExpressionStatement(new StaticInvocation(
	busyWaitProcedure, new Arguments([new IntLiteral(2 /* 0.002 ms */)]))),
	orgBody
	])
	..parent = markProcedure.function;

	Uint8List newDillData = serializeComponent(component);
	File f = new File.fromUri(tmp.resolve("c.dill"));
	f.writeAsBytesSync(newDillData);
	return f.uri;
	}

	/// Instrument the [dillData] so that each procedure-call can be registered
	/// (in package:front_end) and we find out how many times each procedure is
	/// called for a specific run.
	///
	/// Uses the [registerCall] in the helper.
	/// Numbers each procedure, saves the instrumented dill and returns both the
	/// dill and the list of procedures so that procedure i in the list will be
	/// annotated with a call to `registerCall(i)`.
	Instrumented instrumentCallsCount(List<int> dillData, Uri tmp) {
	Component component = new Component();
	new BinaryBuilder(dillData, disableLazyReading: true)
	.readComponent(component);
	Procedure registerCallProcedure = getProcedure(
	component, (lib) => lib.fileUri == benchmarkHelper, null, "registerCall");
	RegisterCallTransformer registerCallTransformer =
	new RegisterCallTransformer(registerCallProcedure);
	component.accept(registerCallTransformer);

	Uint8List newDillData = serializeComponent(component);
	File f = new File.fromUri(tmp.resolve("counting.dill"));
	f.writeAsBytesSync(newDillData);

	return new Instrumented(f.uri, registerCallTransformer.procedures);
	}

	/// Instrument the [dillData] so that each (sync) procedure-call can be timed
	/// (time on stack, i.e. not only the procedure itself, but also the
	/// procedure-calls it makes) (in package:front_end) and we find out how long
	/// each procedure is on the stack for a specific run.
	///
	/// Uses [registerCallStart] and [registerCallEnd] from the helper.
	/// Numbers each sync procedure, saves the instrumented dill and returns both
	/// the dill and the list of procedures so that procedure i in the list will be
	/// annotated with a call-pair to `registerCallStart(i)` and
	/// `registerCallEnd(i)`.
	Instrumented instrumentCallsTiming(List<int> dillData, Uri tmp) {
	Component component = new Component();
	new BinaryBuilder(dillData, disableLazyReading: true)
	.readComponent(component);
	Procedure registerCallStartProcedure = getProcedure(component,
	(lib) => lib.fileUri == benchmarkHelper, null, "registerCallStart");
	Procedure registerCallEndProcedure = getProcedure(component,
	(lib) => lib.fileUri == benchmarkHelper, null, "registerCallEnd");
	RegisterTimeTransformer registerTimeTransformer = new RegisterTimeTransformer(
	registerCallStartProcedure, registerCallEndProcedure);
	component.accept(registerTimeTransformer);

	Uint8List newDillData = serializeComponent(component);
	File f = new File.fromUri(tmp.resolve("timing.dill"));
	f.writeAsBytesSync(newDillData);

	return new Instrumented(f.uri, registerTimeTransformer.procedures);
	}

	/// Class holding both the uri of a saved dill file and a list of procedures (in
	/// order) that has some reference to the annotation added to the dill file.
	class Instrumented {
	final Uri dill;
	final List<Procedure> procedures;

	Instrumented(this.dill, this.procedures);
	}

	class RegisterCallTransformer extends RecursiveVisitor {
	final Procedure registerCallProcedure;
	RegisterCallTransformer(this.registerCallProcedure);
	List<Procedure> procedures = [];

	@override
	void visitLibrary(Library node) {
	if (node.importUri.isScheme("package") &&
	node.importUri.pathSegments.first == "front_end") {
	super.visitLibrary(node);
	}
	}

	@override
	void visitProcedure(Procedure node) {
	if (node.function.body == null) return;
	int procedureNum = procedures.length;
	procedures.add(node);
	Statement orgBody = node.function.body as Statement;
	node.function.body = new Block([
	new ExpressionStatement(new StaticInvocation(registerCallProcedure,
	new Arguments([new IntLiteral(procedureNum)]))),
	orgBody
	]);
	node.function.body!.parent = node.function;
	}
	}

	class RegisterTimeTransformer extends RecursiveVisitor {
	final Procedure registerCallStartProcedure;
	final Procedure registerCallEndProcedure;

	RegisterTimeTransformer(
	this.registerCallStartProcedure, this.registerCallEndProcedure);

	List<Procedure> procedures = [];

	@override
	void visitLibrary(Library node) {
	if (node.importUri.isScheme("package") &&
	node.importUri.pathSegments.first == "front_end") {
	super.visitLibrary(node);
	}
	}

	@override
	void visitProcedure(Procedure node) {
	if (node.function.body == null) return;
	if (node.function.dartAsyncMarker != AsyncMarker.Sync) return;
	int procedureNum = procedures.length;
	procedures.add(node);
	Statement orgBody = node.function.body as Statement;
	// Rewrite as
	// {
	// registerCallStartProcedure(x);
	// try {
	// originalBody
	// } finally {
	// registerCallEndProcedure(x);
	// }
	// }
	Block block = new Block([
	new ExpressionStatement(new StaticInvocation(registerCallStartProcedure,
	new Arguments([new IntLiteral(procedureNum)]))),
	new TryFinally(
	orgBody,
	new ExpressionStatement(new StaticInvocation(registerCallEndProcedure,
	new Arguments([new IntLiteral(procedureNum)]))),
	)
	]);
	node.function.body = block;
	node.function.body!.parent = node.function;
	}
	}

	Procedure getProcedure(Component component, bool libraryMatcher(Library lib),
	String? className, String procedureName) {
	Library lib = component.libraries.where(libraryMatcher).single;
	List<Procedure> procedures = lib.procedures;
	if (className != null) {
	Class cls = lib.classes.where((c) => c.name == className).single;
	procedures = cls.procedures;
	}
	// TODO: This will fail for getter/setter pairs. Fix that.
	return procedures.where((p) => p.name.text == procedureName).single;
	}

	List<int> runXTimes(int x, List<String> arguments, [List<dynamic>? stdout]) {
	List<int> result = [];
	Stopwatch stopwatch = new Stopwatch()..start();
	for (int i = 0; i < x; i++) {
	stopwatch.reset();
	ProcessResult run = Process.runSync(Platform.resolvedExecutable, arguments,
	runInShell: true);
	int ms = stopwatch.elapsedMilliseconds;
	result.add(ms);
	print(ms);
	if (run.exitCode != 0) throw "Got exit code ${run.exitCode}";
	if (stdout != null) {
	stdout.add(run.stdout);
	}
	}
	return result;
	}