• Home
• Jdk17
• Graalvm As A Platform
• Language Implementation Framework

Truffle DSL Guidelines

This document describes some Truffle DSL guidelines. Keep in mind that those are only guidelines that do not have to be strictly followed in every case. The most important part is the reasoning behind every guideline - use that to assess the trade-offs for the specific situation at hand and to chose appropriate solution.

One of the general high-level guidelines for partially evaluated (PE) code is to minimize code duplication during the PE process. This not only helps the Truffle compilation that uses partial evaluation, but also host inlining, which follows similar rules to PE when compiling the interpreter.

Another general high-level guideline for any Truffle interpreter code is to have as little code as possible in order to minimize the native-image size. This applies to runtime code and PE code, but is even more important for PE code, since native-image also AOT compiles it, but the host-inlining greatly increases the amount of code the AOT compilation produces, and on top of that native-image also needs to retain serialized Graal IR graphs of all PE code for runtime compilation.

Avoid subclassing for minor changes #

Example:

abstract class MyBaseNode extends Node {
  abstract int execute(Object o);

  @Specialization(guards = "arg == 0")
  int doZero(int arg) { /* ... */ }

  @Specialization(guards = "arg != 0")
  int doInt(int arg) { /* ... */ }

  @Specialization
  int doOther(Object o) { throw new AbstractMethodError(); }
}

abstract class Node1 extends MyBaseNode {
  @Specialization
  @Override
  final int doOther(Object o) { return 42; }
}

abstract class Node2 extends MyBaseNode {
  @Specialization
  @Override
  final int doOther(Object o) { return -1; }
}

Why: native-image binary size

Truffle DSL will generate multiple execute and executeAndSpecialize methods, which will contain the same code. Native-image does not deduplicate the code.

Solution: use delegation to a child node instead of inheritance or, if the code is simple enough, inline all the implementations and switch on a field value or argument to the execute method. Example of the former:

abstract class MyNodeHandleOther extends Node {
    abstract int execute(Object o);
}

abstract class MyNode extends Node {
  @Child MyNodeHandleOther otherHandler;

  MyNode(MyNodeHandleOther otherHandler) {
      this.otherHandler = otherHandler;
  }

  abstract int execute(Object o);

  @Specialization(guards = "arg == 0")
  int doZero(int arg) { /* ... */ }

  @Specialization(guards = "arg != 0")
  int doInt(int arg) { /* ... */ }

  @Specialization
  int doOther(Object o) {
      return otherHandler.execute(o);
  }
}

Alternative solution: if the common code is only one @Specialization or there is a simple and efficient guard that captures all the common @Specializations, then move the common @Specializations to an inline node and in all the former subclasses add one @Specialization that delegates to the common inline node.

@GenerateCached(false)
@GenerateInline
abstract class MyCommonNode extends Node {
  abstract int execute(Node node, Object o);

  @Specialization(guards = "arg == 0")
  int doZero(int arg) { /* ... */ }

  @Specialization(guards = "arg != 0")
  int doInt(int arg) { /* ... */ }
}

abstract class Node1 extends Node {
    abstract int execute(Object o);

    @Specialization
    int doInts(int o,
               @Cached MyCommonNode node) {
        return node.execute(this, o);
    }

    @Specialization
    int doOther(Object o) { return 42; }
}

// analogically for Node2

Avoid duplicated Specializations #

Avoid specializations with (almost) the same method body.

When two or more @Specializations differ only in guards or in some minor detail. This includes @Specializations that delegate to the same helper method. Example:

abstract class MyNode extends Node {
  abstract void execute(Object o);

  @Specialization
  void doObj1(MyObject1 o) { helper(o); }

  @Specialization
  void doObj2(MyObject2 o) { helper(o); }

  void helper(Object o) { /* some code */ }

  // ... more @Specializations
}

Why:

• code duplication during PE

We want to reduce code duplication as seen by PE process and not by the developer. Refactoring the code to a helper method does not help PE, because it explores every call separately.

For instance, with host inlining and our code example, the cost of fully inlining Node#execute will be (omitting some details):

size(Node#execute) + size(Node#doIt1) + size(Node#doIt2) + 2 * size(Node#helper)

Solution: refactor the code to avoid the duplication. The concrete approach differs depending on concrete situation. There is no one-size-fits-all solution. General advice is to try to merge the @Specializations that contain the code duplication. One can create an inline node that profiles the now merged conditions that used to be implicitly profiled by being separate @Specializations. For example:

@GenerateInline
@GenerateCached(false)
abstract class GuardNode extends Node {
  abstract boolean execute(Node inliningTarget, Object o);

  @Specialization
  static boolean doObj1(MyObject1 o) { return true; }

  @Specialization
  static boolean doObj2(MyObject2 o) { return true; }

  @Fallback
  static boolean doFallback(Object o) { return false; }
}

abstract class MyNode extends Node {
  abstract void execute(Object o);

  @Specialization(guards = "guardNode.execute(this, o)", limit = "1")
  void doObj(Object o,
             @Cached GuardNode guardNode) { helper(o); }

  // ...other @Specializations
}

Note that if the guard needs to be used for multiple specializations, or will be used by generated fallback guard, we are duplicating the guard logic in the same way as we were duplicating the logic inside the specializations. This may be acceptable as guards tend to be simple, but the user needs to assess if that is a good trade-off.

Avoid duplicated calls to helper methods/nodes #

Example:

@Specialization
void doDefault(boolean b, Object o) {
  if (b) {
    helper.execute(o, 42);
  } else {
    helper.execute(o, -1);
  }
}

Why: code duplication during PE

The PE process has to explore each call separately and only in later phases the Graal compiler may deduplicate the code.

Solution: common-out the calls if possible

  int num = b ? 42 : -1;
  helper.execute(o, num);

Mixing @Shared and non-@Shared inline nodes/profiles #

Avoid mixing @Shared and non-@Shared inline nodes/profiles in one @Specialization if Truffle DSL generates “data-class” for the @Specialization.

Example:

@GenerateInline(false)
class MyNode extends Node {
  // ...
  @Specialization
  void doIt(...,
      @Bind("this") Node node,
      /* more @Cached arguments such that data-class is generated */
      @Exclusive @Cached InlinedBranchProfile b1,
      @Shared @Cached InlinedBranchProfile b2)

Why: Truffle DSL generates code that is less efficient in the interpreter.

In our example: non-shared inline profile has its data stored in the data-class object, but the shared inline profile has its data stored in the instance of MyNode. However, both profiles receive the same node argument, which will be an instance of the generated data-class, so the shared profile must call node.getParent() to access its data stored in MyNode. In general, such inline nodes/profiles may need to traverse multiple parent pointers until they reach their data.

Note: this does not concern any non-inline nodes, it is OK to mix those, and it is OK to mix them with inline nodes, however, inline nodes used in one @Specialization should be either all shared or all exclusive.

Solution: change the @Shared nodes/profiles to @Exclusive or refactor the code such that sharing is not necessary anymore. Usage of @Shared (not only inline nodes/profiles) can be a sign of duplicated @Specializations, and refactoring the @Specializations will resolve the problem. If the footprint benefit outweighs the possible interpreter performance degradation, this guideline can be ignored.

Generic solution that trades off some code readability for good interpreter performance and lower footprint at the same time is to split the code into two nodes, where the outer node takes only the shared inline nodes/profiles and the inner node takes only the non-shared. The outer node can execute some common logic and also forward the shared nodes along with their inlining target node to the inner node.

Avoid unused large inline nodes #

Avoid inlining large nodes that are used only on rarely executed code-paths.

Example:

@Specialization
void doObject(Object arg,
    @Bind("this") Node inliningTarget,
    @Cached LargeInlineNode n) {

    if (arg == null) {
        // unlikely situation
        n.execute(inliningTarget, ...);
    }
}

Why: runtime memory footprint

All the fields of LargeInlineNode node will be inlined into the caller node (or Specialization data-class) increasing its memory footprint significantly.

For code-paths that are not performance sensitive in the interpreter, better alternative is the lazily initialized nodes.

For Code-paths that are performance sensitive in the interpreter:

• The footprint increase may be justified by the performance
• Use the handwritten lazily initialized @Child field pattern if applicable
• If possible restructure the code to avoid such situation

Avoid generating cached, uncached, and inline variant of one node #

If possible avoid having all three variants, because it increases the PE code footprint.