| layout | toc_group | link_title | permalink |
|---|---|---|---|
docs |
truffle |
Host Optimization for Interpreter Code |
/graalvm-as-a-platform/language-implementation-framework/HostOptimization/ |
For the following document, we disambiguate between host and guest compilation.
- Host compilation is applied to the Java implementation of the interpreter. If the interpreter runs on HotSpot, this kind of compilation is applied when Truffle interpreters are JIT compiled (or dynamically compiled) as Java applications. This compilation is applied ahead of time during native image generation.
- Guest compilation is applied to guest language code. This kind of compilation uses Partial Evaluation and Futamura projections to derive optimized code from Truffle ASTs and bytecodes.
This section discusses domain-specific host compilations applied to Truffle AST and bytecode interpreters.
Truffle interpreters are written to support runtime compilation by applying the first Futamura projection. Runtime compilable code, also often referred to as partial evaluatable code, has the following characteristics:
- It is naturally designed for high-performance, as it also defines the performance after runtime compilation of a language.
- It is written to avoid recursion, as recursive code cannot quickly be partially evaluated.
- It avoids complex abstractions and third-party code as they are typically not designed for PE.
- The boundaries of partially evaluatable code are reliably defined by methods annotated with
@TruffleBoundary, blocks dominated by a call toCompilerDirectives.transferToInterpreter()or a block protected by a call toCompilerDirectives.inInterpreter().
Truffle host inlining leverages these properties and forces inlining during host compilation for runtime compilable code paths as much as possible. The general assumption is that code important for runtime compilation is also important for interpreter execution. Whenever a PE boundary is detected, the host inlining phase no longer makes any inlining decisions and defers them to later inlining phases better suited for regular Java code.
The source code for this phase can be found in HostInliningPhase.
Truffle host inlining is applied when compiling a method annotated with @HostCompilerDirectives.BytecodeInterpreterSwitch.
The maximum node cost for such methods can be configured using -H:TruffleHostInliningByteCodeInterpreterBudget=100000 for native images and -Dgraal.TruffleHostInliningByteCodeInterpreterBudget=100000 on HotSpot.
If a method that is annotated with @BytecodeInterpreterSwitch calls a method with the same annotation then the method is directly inlined as long as the cost of both methods do not exceed the budget.
In other words, any such method will be treated by the inlining phase just as if they would be part of the root bytecode switch method.
This allows bytecode interpreter switches to be composed of multiple methods if needed.
Native image, during closed world analysis, computes all methods that are reachable for runtime compilation.
Any potentially reachable method from RootNode.execute(...) is determined as runtime compilable.
For native images, in addition to bytecode interpreter switches, all runtime compilable methods are optimized using Truffle host inlining.
The maximum node cost of such an inlining pass can be configured with -H:TruffleHostInliningBaseBudget=5000.
On HotSpot the set of runtime compilable methods is unknown.
Therefore, we can only rely on regular Java method inlining for methods not annotated as bytecode interpreter switch on HotSpot.
Whenever the maximum budget for a compilation unit is reached, inlining will be stopped.
The same budget will be used for the exploration of subtrees during inlining.
If a call cannot be fully explored and inlined within the budget, then no decision is taken on the individual subtree.
For the vast majority of runtime compilable methods, this limit will not be reached, as it is prevented by natural PE boundaries as well as polymorphic calls to execute methods of @Child nodes.
If there are methods that exceed the budget limit, then the recommendation is to optimize such nodes by adding more PE boundaries.
If a method exceeds the limit, it is likely that the same code also has a high cost for runtime compilation.
The inlining decisions performed by this phase is best debugged with -H:Log=HostInliningPhase,~CanonicalizerPhase,~GraphBuilderPhase for native images or -Dgraal.Log=HostInliningPhase,~CanonicalizerPhase,~GraphBuilderPhase on HotSpot.
You can redirect the output to a file with -Dgraal.LogFile=FILE (works for both).
Consider the following example, which shows previously described common patterns of partial evaluatable code in Truffle interpreters:
class BytecodeNode extends Node {
@CompilationFinal(dimensions = 1) final byte[] ops;
@Children final BaseNode[] polymorphic = new BaseNode[]{new SubNode1(), new SubNode2()};
@Child SubNode1 monomorphic = new SubNode1();
BytecodeNode(byte[] ops) {
this.ops = ops;
}
@BytecodeInterpreterSwitch
@ExplodeLoop(kind = LoopExplosionKind.MERGE_EXPLODE)
public void execute() {
int bci = 0;
while (bci < ops.length) {
switch (ops[bci++]) {
case 0:
// regular operation
add(21, 21);
break;
case 1:
// complex operation in @TruffleBoundary annotated method
truffleBoundary();
break;
case 2:
// complex operation protected behind inIntepreter
if (CompilerDirectives.inInterpreter()) {
protectedByInIntepreter();
}
break;
case 3:
// complex operation dominated by transferToInterpreter
CompilerDirectives.transferToInterpreterAndInvalidate();
dominatedByTransferToInterpreter();
break;
case 4:
// first level of recursion is inlined
recursive(5);
break;
case 5:
// can be inlined is still monomorphic (with profile)
monomorphic.execute();
break;
case 6:
for (int y = 0; y < polymorphic.length; y++) {
// can no longer be inlined (no longer monomorphic)
polymorphic[y].execute();
}
break;
default:
// propagates transferToInterpeter from within the call
throw CompilerDirectives.shouldNotReachHere();
}
}
}
private static int add(int a, int b) {
return a + b;
}
private void protectedByInIntepreter() {
}
private void dominatedByTransferToInterpreter() {
}
private void recursive(int i) {
if (i == 0) {
return;
}
recursive(i - 1);
}
@TruffleBoundary
private void truffleBoundary() {
}
abstract static class BaseNode extends Node {
abstract int execute();
}
static class SubNode1 extends BaseNode {
@Override
int execute() {
return 42;
}
}
static class SubNode2 extends BaseNode {
@Override
int execute() {
return 42;
}
}
}We can run this as a unittest in the Graal repository (see class HostInliningBytecodeInterpreterExampleTest) by running the following command line in graal/compiler:
mx unittest -Dgraal.Log=HostInliningPhase,~CanonicalizerPhase,~GraphBuilderPhase -Dgraal.Dump=:3 HostInliningBytecodeInterpreterExampleTest
This prints:
[thread:1] scope: main
[thread:1] scope: main.Testing
Context: HotSpotMethod<HostInliningBytecodeInterpreterExampleTest$BytecodeNode.execute()>
Context: StructuredGraph:1{HotSpotMethod<HostInliningBytecodeInterpreterExampleTest$BytecodeNode.execute()>}
[thread:1] scope: main.Testing.EnterpriseHighTier.HostInliningPhase
Truffle host inlining completed after 2 rounds. Graph cost changed from 136 to 137 after inlining:
Root[org.graalvm.compiler.truffle.test.HostInliningBytecodeInterpreterExampleTest$BytecodeNode.execute]
INLINE org.graalvm.compiler.truffle.test.HostInliningBytecodeInterpreterExampleTest$BytecodeNode.add(int, int) [inlined 2, monomorphic false, deopt false, inInterpreter false, propDeopt false, subTreeInvokes 0, subTreeCost 8, incomplete false, reason null]
CUTOFF org.graalvm.compiler.truffle.test.HostInliningBytecodeInterpreterExampleTest$BytecodeNode.truffleBoundary() [inlined -1, monomorphic false, deopt false, inInterpreter false, propDeopt false, subTreeInvokes 1, subTreeCost 0, incomplete false, reason truffle boundary]
INLINE com.oracle.truffle.api.CompilerDirectives.inInterpreter() [inlined 0, monomorphic false, deopt false, inInterpreter false, propDeopt false, subTreeInvokes 0, subTreeCost 6, incomplete false, reason null]
CUTOFF org.graalvm.compiler.truffle.test.HostInliningBytecodeInterpreterExampleTest$BytecodeNode.protectedByInIntepreter() [inlined -1, monomorphic false, deopt false, inInterpreter true, propDeopt false, subTreeInvokes 1, subTreeCost 0, incomplete false, reason protected by inInterpreter()]
INLINE com.oracle.truffle.api.CompilerDirectives.transferToInterpreterAndInvalidate() [inlined 3, monomorphic false, deopt true, inInterpreter false, propDeopt false, subTreeInvokes 0, subTreeCost 32, incomplete false, reason null]
INLINE com.oracle.truffle.api.CompilerDirectives.inInterpreter() [inlined 3, monomorphic false, deopt true, inInterpreter false, propDeopt false, subTreeInvokes 0, subTreeCost 6, incomplete false, reason null]
CUTOFF org.graalvm.compiler.truffle.runtime.hotspot.AbstractHotSpotTruffleRuntime.traceTransferToInterpreter() [inlined -1, monomorphic false, deopt true, inInterpreter true, propDeopt false, subTreeInvokes 0, subTreeCost 0, incomplete false, reason dominated by transferToInterpreter()]
CUTOFF org.graalvm.compiler.truffle.test.HostInliningBytecodeInterpreterExampleTest$BytecodeNode.dominatedByTransferToInterpreter() [inlined -1, monomorphic false, deopt true, inInterpreter false, propDeopt false, subTreeInvokes 0, subTreeCost 0, incomplete false, reason dominated by transferToInterpreter()]
INLINE org.graalvm.compiler.truffle.test.HostInliningBytecodeInterpreterExampleTest$BytecodeNode.recursive(int) [inlined 4, monomorphic false, deopt false, inInterpreter false, propDeopt false, subTreeInvokes 1, subTreeCost 20, incomplete false, reason null]
CUTOFF org.graalvm.compiler.truffle.test.HostInliningBytecodeInterpreterExampleTest$BytecodeNode.recursive(int) [inlined -1, monomorphic false, deopt false, inInterpreter false, propDeopt false, subTreeInvokes 1, subTreeCost 0, incomplete false, reason recursive]
INLINE org.graalvm.compiler.truffle.test.HostInliningBytecodeInterpreterExampleTest$BytecodeNode$SubNode1.execute() [inlined 1, monomorphic false, deopt false, inInterpreter false, propDeopt false, subTreeInvokes 0, subTreeCost 6, incomplete false, reason null]
CUTOFF org.graalvm.compiler.truffle.test.HostInliningBytecodeInterpreterExampleTest$BytecodeNode$BaseNode.execute() [inlined -1, monomorphic false, deopt false, inInterpreter false, propDeopt false, subTreeInvokes 1, subTreeCost 0, incomplete false, reason not direct call: no type profile]
CUTOFF com.oracle.truffle.api.CompilerDirectives.shouldNotReachHere() [inlined -1, monomorphic false, deopt false, inInterpreter false, propDeopt true, subTreeInvokes 0, subTreeCost 98, incomplete false, reason propagates transferToInterpreter]
Note that we have also used the -Dgraal.Dump=:3 option, which sends the graphs to any running IdealGraphVisualizer instance for further inspection.
On Native Image, use -H:Dump=:2 -H:MethodFilter=... to dump host compilation graphs of a given method.
To debug CUTOFF decisions for incomplete exploration (entries with incomplete true) use the -Dgraal.TruffleHostInliningPrintExplored=true option to see all incomplete subtrees in the log.
After learning how to debug and trace host inlining decisions, it is time to look at some ways to tune it.
As a first step, it is necessary to identify compilation units essential for good interpreter performance.
To do this, a Truffle interpreter can be executed in interpreter-only mode by setting the engine.Compilation flag to false.
After that, a Java profiler may be used to identify hot spots in the execution.
For further details on profiling, see Profiling.md
If you are looking for advice on how and when to optimize Truffle interpreters, see Optimizing.md
After identifying a hot method, for example, the bytecode dispatch loop in a Truffle bytecode interpreter, we can further investigate using host inlining logging as described in the previous section.
Interesting entries are prefixed with CUTOFF and have a reason that explains the reason for the individual cutoff.
Common reasons for CUTOFF entries are:
dominated by transferToInterpreter()orprotected by inInterpreter(): This means that the is call performed in a slow-path. Host inlining will not decide on such calls and just mark them as CUTOFF.target method not inlinablethis happens for host VM methods that cannot be inlined. There is typically not much we can do about that.Out of budgetwe ran out of budget for inlining this method. This happens if the cost of the method becomes too high.
Additionally, to avoid the explosion of code size, host inlining has a built-in heuristic to detect call subtrees that are considered too complex to inline. For example, the tracing may print the following:
CUTOFF com.oracle.truffle.espresso.nodes.BytecodeNode.putPoolConstant(VirtualFrame, int, char, int) [inlined -1, explored 0, monomorphic false, deopt false, inInterpreter false, propDeopt false, graphSize 1132, subTreeCost 5136, invokes 1, subTreeInvokes 12, forced false, incomplete false, reason call has too many fast-path invokes - too complex, please optimize, see truffle/docs/HostOptimization.md
This indicates that there are too many fast-path invokes (by default 10) in the subtree, it also stops exploring after that number.
The -Dgraal.TruffleHostInliningPrintExplored=true flag may be provided to see the entire subtree for the decision.
The following calls are considered fast-path invokes:
- Invokes where the target method is annotated by
@TruffleBoundary. - Invokes that are polymorphic or where no monomorphic profiling feedback is available. For example, a call to a subexpression's execute method.
- Invokes that are recursive.
- Invokes that are too complex themselves. For example, invokes that have too many fast-path invokes.
The following calls are not considered fast-path invokes:
- Invokes that can be inlined using the host inlining heuristic.
- Invokes in a slow-path, like any invoke that is dominated by
transferToInterpreter()or protected byisInterpreter(). - Invokes that cannot be inlined due to limitations of the host VM, like calls to
Throwable.fillInStackTrace(). - Invokes that are no longer reachable.
It is impossible to avoid fast-path invokes entirely, as, for example, child nodes need to be executed in an AST.
It is theoretically possible to avoid all fast-path invokes in a bytecode interpreter.
In practice, languages will rely on @TruffleBoundary to the runtime to implement more complex bytecodes.
In the following sections, we discuss techniques on how to improve host interpreter code:
As mentioned in the previous section, a heuristic automatically cuts inlining subtrees with too many calls in them. One way to optimize this is by using the @InliningCutoff annotation.
Consider the following example:
abstract class AddNode extends Node {
abstract Object execute(Object a, Object b);
@Specialization int doInt(int a, int b) { return a + b; }
@Specialization double doDouble(double a, double b) { return a + b; }
@Specialization double doGeneric(Object a, Object b, @Cached LookupAndCallNode callNode) {
return callNode.execute("__add", a, b);
}
}
In this example, the specializations doInt and doDouble are very simple, but there is also the doGeneric specialization, which calls into a complex lookup chain.
Assuming that the LookupAndCallNode.execute is a very complex method with more than ten fast-path subtree calls, we could not expect the execute method to get inlined.
Host inlining currently does not support automatic component analysis; though it can be specified manually using the @InliningCutoff annotation:
abstract class AddNode extends Node {
abstract Object execute(Object a, Object b);
@Specialization int doInt(int a, int b) { return a + b; }
@Specialization double doDouble(double a, double b) { return a + b; }
@HostCompilerDirectives.InliningCutoff
@Specialization double doGeneric(Object a, Object b, @Cached LookupAndCallNode callNode) {
return callNode.execute("__add__", a, b);
}
}
After changing the code, host inlining may now decide to inline the execute method of the AddNode if it fits into the host inlining budget, but force a CUTOFF at the doGeneric(...) method call.
Please see the javadoc on other use-cases for using this annotation.
The following is an example where the code is efficient for compilation using partial evaluation but is not ideal for host compilation.
@Child HelperNode helperNode;
final boolean negate;
// ....
int execute(int argument) {
if (negate) {
return helperNode.execute(-argument);
} else {
return helperNode.execute(argument);
}
}
When this code is compiled using partial evaluation, this code is efficient as the condition is guaranteed to fold to a single case, as the negate field is compilation final.
During host optimization, the negate field is not compilation final, and the compiler would either inline the code twice or decide not to inline the execute method.
In order to avoid this the code can be rewritten as follows:
@Child HelperNode helperNode;
final boolean negate;
// ....
int execute(int argument) {
int negatedArgument;
if (negate) {
negatedArgument = -argument;
} else {
negatedArgument = argument;
}
return helperNode.execute(negatedArgument);
}
Similar code patterns can arise indirectly through code generation if many specializations with the same method body are used. Host compilers typically have a hard time optimizing such patterns automatically.
Consider the following example:
int execute(int argument) {
if (argument == 0) {
CompilerDirectives.transferToInterpeterAndInvalidate();
throw new RuntimeException("Invalid zero argument " + argument);
}
return argument;
}
The Java compiler generates bytecode equivalent to the following code:
int execute(int argument) {
if (argument == 0) {
CompilerDirectives.transferToInterpeterAndInvalidate();
throw new RuntimeException(new StringBuilder("Invalid zero argument ").append(argument).build());
}
return argument;
}
While this code is efficient for partial evaluation, this code takes up unnecessary space during host inlining. It is therefore recommended to extract a single method for the slow-path part of the code:
int execute(int argument) {
if (argument == 0) {
CompilerDirectives.transferToInterpeterAndInvalidate();
throw invalidZeroArgument(argument);
}
return argument;
}
RuntimeException invalidZeroArgument(int argument) {
throw new RuntimeException("Invalid zero argument " + argument);
}