Introduce DispatchPrimitiveNode

to eliminate the generated code for the additional abstract `AbstractPrimitiveNode#execute(VirtualFrame)` method for each primitive.
hpi-swa · Jul 30, 2024 · e711ba7 · e711ba7 · TruffleSqueak-Bot · Jul 31, 2024
1 parent 2d4a231
commit e711ba7
Show file tree

Hide file tree

Showing 13 changed files with 109 additions and 79 deletions.
diff --git a/src/de.hpi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/model/CompiledCodeObject.java b/src/de.hpi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/model/CompiledCodeObject.java
@@ -42,7 +42,7 @@
 import de.hpi.swa.trufflesqueak.nodes.bytecodes.AbstractSqueakBytecodeDecoder;
 import de.hpi.swa.trufflesqueak.nodes.bytecodes.SqueakBytecodeSistaV1Decoder;
 import de.hpi.swa.trufflesqueak.nodes.bytecodes.SqueakBytecodeV3PlusClosuresDecoder;
-import de.hpi.swa.trufflesqueak.nodes.primitives.AbstractPrimitiveNode;
+import de.hpi.swa.trufflesqueak.nodes.primitives.DispatchPrimitiveNode;
 import de.hpi.swa.trufflesqueak.nodes.primitives.PrimitiveNodeFactory;
 import de.hpi.swa.trufflesqueak.nodes.primitives.PrimitiveNodeFactory.ArgumentsLocation;
 import de.hpi.swa.trufflesqueak.shared.SqueakLanguageConfig;
@@ -217,7 +217,7 @@ private void initializeCallTargetUnsafe() {
         final SqueakLanguage language = SqueakImageContext.getSlow().getLanguage();
         final RootNode rootNode;
         if (hasPrimitive() && PrimitiveNodeFactory.isNonFailing(this)) {
-            final AbstractPrimitiveNode primitiveNode = PrimitiveNodeFactory.getOrCreateIndexedOrNamed(this, ArgumentsLocation.IN_FRAME_ARGUMENTS);
+            final DispatchPrimitiveNode primitiveNode = PrimitiveNodeFactory.getOrCreateIndexedOrNamed(this, ArgumentsLocation.IN_FRAME_ARGUMENTS);
             assert primitiveNode != null;
             rootNode = new ExecuteNonFailingPrimitiveRootNode(language, this, primitiveNode);
         } else {

diff --git a/src/de.hpi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/ExecuteBytecodeNode.java b/src/de.hpi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/ExecuteBytecodeNode.java
@@ -27,7 +27,7 @@
 import de.hpi.swa.trufflesqueak.nodes.bytecodes.JumpBytecodes.ConditionalJumpNode;
 import de.hpi.swa.trufflesqueak.nodes.bytecodes.ReturnBytecodes.AbstractReturnNode;
 import de.hpi.swa.trufflesqueak.nodes.bytecodes.SendBytecodes.AbstractSendNode;
-import de.hpi.swa.trufflesqueak.nodes.primitives.AbstractPrimitiveNode;
+import de.hpi.swa.trufflesqueak.nodes.primitives.DispatchPrimitiveNode;
 import de.hpi.swa.trufflesqueak.nodes.primitives.PrimitiveNodeFactory;
 import de.hpi.swa.trufflesqueak.nodes.primitives.PrimitiveNodeFactory.ArgumentsLocation;
 import de.hpi.swa.trufflesqueak.util.ArrayUtils;
@@ -41,7 +41,7 @@ public final class ExecuteBytecodeNode extends AbstractExecuteContextNode implem
     private final int initialPC;
     private SourceSection section;
 
-    @Child private AbstractPrimitiveNode primitiveNode;
+    @Child private DispatchPrimitiveNode primitiveNode;
     @Child private HandlePrimitiveFailedNode handlePrimitiveFailedNode;
     @Children private AbstractBytecodeNode[] bytecodeNodes;
     @Child private HandleNonLocalReturnNode handleNonLocalReturnNode;

diff --git a/....trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/ExecuteNonFailingPrimitiveRootNode.java b/....trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/ExecuteNonFailingPrimitiveRootNode.java
@@ -15,15 +15,15 @@
 import de.hpi.swa.trufflesqueak.SqueakLanguage;
 import de.hpi.swa.trufflesqueak.exceptions.PrimitiveFailed;
 import de.hpi.swa.trufflesqueak.model.CompiledCodeObject;
-import de.hpi.swa.trufflesqueak.nodes.primitives.AbstractPrimitiveNode;
+import de.hpi.swa.trufflesqueak.nodes.primitives.DispatchPrimitiveNode;
 import de.hpi.swa.trufflesqueak.shared.SqueakLanguageConfig;
 
 @NodeInfo(language = SqueakLanguageConfig.ID, cost = NodeCost.NONE)
 public final class ExecuteNonFailingPrimitiveRootNode extends AbstractRootNode {
 
-    @Child private AbstractPrimitiveNode primitiveNode;
+    @Child private DispatchPrimitiveNode primitiveNode;
 
-    public ExecuteNonFailingPrimitiveRootNode(final SqueakLanguage language, final CompiledCodeObject code, final AbstractPrimitiveNode primitiveNode) {
+    public ExecuteNonFailingPrimitiveRootNode(final SqueakLanguage language, final CompiledCodeObject code, final DispatchPrimitiveNode primitiveNode) {
         super(language, code);
         this.primitiveNode = primitiveNode;
     }

diff --git a/src/de.hpi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/bytecodes/SendBytecodes.java b/src/de.hpi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/bytecodes/SendBytecodes.java
@@ -42,6 +42,7 @@
 import de.hpi.swa.trufflesqueak.nodes.dispatch.LookupClassNode;
 import de.hpi.swa.trufflesqueak.nodes.dispatch.LookupSelectorNode;
 import de.hpi.swa.trufflesqueak.nodes.primitives.AbstractPrimitiveNode;
+import de.hpi.swa.trufflesqueak.nodes.primitives.DispatchPrimitiveNode;
 import de.hpi.swa.trufflesqueak.nodes.primitives.PrimitiveNodeFactory;
 import de.hpi.swa.trufflesqueak.nodes.primitives.PrimitiveNodeFactory.ArgumentsLocation;
 import de.hpi.swa.trufflesqueak.nodes.primitives.impl.ControlPrimitives.PrimExitToDebuggerNode;
@@ -315,12 +316,14 @@ public static AbstractBytecodeNode create(final VirtualFrame frame, final Compil
                 return SendSpecialSelectorQuickPointYNodeGen.create(code, index, selectorIndex);
             }
             if (primitiveIndex > 0) {
-                final AbstractPrimitiveNode primitiveNode = PrimitiveNodeFactory.getOrCreateIndexed(primitiveIndex, 1 + numArguments, ArgumentsLocation.ON_STACK_REVERSED);
+                final int numReceiverAndArguments = 1 + numArguments;
+                final AbstractPrimitiveNode primitiveNode = PrimitiveNodeFactory.getOrCreateIndexed(primitiveIndex, numReceiverAndArguments);
                 assert primitiveNode != null;
+                final DispatchPrimitiveNode dispatchPrimitiveNode = DispatchPrimitiveNode.create(primitiveNode, ArgumentsLocation.ON_STACK_REVERSED, numReceiverAndArguments);
                 if (numArguments == 0) {
-                    return new SendSpecialSelectorQuick0ArgumentsNode(code, index, selectorIndex, primitiveNode);
+                    return new SendSpecialSelectorQuick0ArgumentsNode(code, index, selectorIndex, dispatchPrimitiveNode);
                 } else {
-                    return new SendSpecialSelectorQuick1OrMoreArgumentsNode(code, index, selectorIndex, primitiveNode);
+                    return new SendSpecialSelectorQuick1OrMoreArgumentsNode(code, index, selectorIndex, dispatchPrimitiveNode);
                 }
             } else {
                 return new SelfSendNode(code, index, 1, specialSelector, numArguments);
@@ -370,9 +373,9 @@ public final String toString() {
     }
 
     private abstract static class SendSpecialSelectorQuickNode extends AbstractSendSpecialSelectorQuickNode {
-        @Child protected AbstractPrimitiveNode primitiveNode;
+        @Child protected DispatchPrimitiveNode primitiveNode;
 
-        private SendSpecialSelectorQuickNode(final CompiledCodeObject code, final int index, final int selectorIndex, final AbstractPrimitiveNode primitiveNode) {
+        private SendSpecialSelectorQuickNode(final CompiledCodeObject code, final int index, final int selectorIndex, final DispatchPrimitiveNode primitiveNode) {
             super(code, index, selectorIndex);
             this.primitiveNode = primitiveNode;
         }
@@ -390,7 +393,7 @@ public void executeVoid(final VirtualFrame frame) {
     }
 
     private static final class SendSpecialSelectorQuick0ArgumentsNode extends SendSpecialSelectorQuickNode {
-        private SendSpecialSelectorQuick0ArgumentsNode(final CompiledCodeObject code, final int index, final int selectorIndex, final AbstractPrimitiveNode primitiveNode) {
+        private SendSpecialSelectorQuick0ArgumentsNode(final CompiledCodeObject code, final int index, final int selectorIndex, final DispatchPrimitiveNode primitiveNode) {
             super(code, index, selectorIndex, primitiveNode);
         }
 
@@ -408,7 +411,7 @@ protected void popArgumentAndPush(final VirtualFrame frame, final Object result)
     private static final class SendSpecialSelectorQuick1OrMoreArgumentsNode extends SendSpecialSelectorQuickNode {
         @CompilationFinal private int stackPointer;
 
-        private SendSpecialSelectorQuick1OrMoreArgumentsNode(final CompiledCodeObject code, final int index, final int selectorIndex, final AbstractPrimitiveNode primitiveNode) {
+        private SendSpecialSelectorQuick1OrMoreArgumentsNode(final CompiledCodeObject code, final int index, final int selectorIndex, final DispatchPrimitiveNode primitiveNode) {
             super(code, index, selectorIndex, primitiveNode);
         }
 
@@ -429,7 +432,7 @@ private abstract static class SendSpecialSelectorQuickWithClassCheckNode extends
         @CompilationFinal private ClassObject cachedReceiverClass;
         @CompilationFinal private Assumption cachedCallTargetStableAssumption;
 
-        @Child protected AbstractPrimitiveNode primitiveNode;
+        @Child protected DispatchPrimitiveNode primitiveNode;
         @Child private FrameStackReadNode peekAtReceiverNode;
         @Child private LookupClassNode lookupClassNode = LookupClassNode.create();
 
@@ -460,7 +463,7 @@ private boolean doesNotMatchClassOrMethodInvalidated(final ClassObject actualCla
                 final Object lookupResult = actualClass.lookupInMethodDictSlow(findSelector());
                 if (lookupResult instanceof final CompiledCodeObject primitiveMethod && primitiveMethod.hasPrimitive()) {
                     assert primitiveMethod.getNumArgs() == findNumArguments();
-                    final AbstractPrimitiveNode node = PrimitiveNodeFactory.getOrCreateIndexedOrNamed(primitiveMethod, ArgumentsLocation.ON_STACK_REVERSED);
+                    final DispatchPrimitiveNode node = PrimitiveNodeFactory.getOrCreateIndexedOrNamed(primitiveMethod, ArgumentsLocation.ON_STACK_REVERSED);
                     if (node == null) {
                         return true; // primitive not found / supported
                     }

diff --git a/...hpi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/dispatch/CachedDispatchNode.java b/...hpi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/dispatch/CachedDispatchNode.java
@@ -27,7 +27,7 @@
 import de.hpi.swa.trufflesqueak.nodes.context.frame.GetOrCreateContextNode;
 import de.hpi.swa.trufflesqueak.nodes.dispatch.CreateFrameArgumentNodes.CreateFrameArgumentsForDNUNode;
 import de.hpi.swa.trufflesqueak.nodes.dispatch.CreateFrameArgumentNodes.CreateFrameArgumentsForOAMNode;
-import de.hpi.swa.trufflesqueak.nodes.primitives.AbstractPrimitiveNode;
+import de.hpi.swa.trufflesqueak.nodes.primitives.DispatchPrimitiveNode;
 import de.hpi.swa.trufflesqueak.nodes.primitives.PrimitiveNodeFactory;
 import de.hpi.swa.trufflesqueak.nodes.primitives.PrimitiveNodeFactory.ArgumentsLocation;
 import de.hpi.swa.trufflesqueak.util.FrameAccess;
@@ -48,7 +48,7 @@ protected static final CachedDispatchNode create(final VirtualFrame frame, final
             return createDNUNode(frame, selector, argumentCount, image, receiverClass);
         } else if (lookupResult instanceof final CompiledCodeObject lookupMethod) {
             if (lookupMethod.hasPrimitive()) {
-                final AbstractPrimitiveNode primitiveNode = PrimitiveNodeFactory.getOrCreateIndexedOrNamed(lookupMethod, ArgumentsLocation.ON_STACK);
+                final DispatchPrimitiveNode primitiveNode = PrimitiveNodeFactory.getOrCreateIndexedOrNamed(lookupMethod, ArgumentsLocation.ON_STACK);
                 if (primitiveNode != null) {
                     return new CachedDispatchPrimitiveNode(argumentCount, lookupMethod, primitiveNode);
                 }
@@ -94,9 +94,9 @@ protected static final class CachedDispatchPrimitiveNode extends CachedDispatchN
         private final int argumentCount;
         private final PrimitiveFailedCounter failureCounter;
 
-        @Child private AbstractPrimitiveNode primitiveNode;
+        @Child private DispatchPrimitiveNode primitiveNode;
 
-        private CachedDispatchPrimitiveNode(final int argumentCount, final CompiledCodeObject method, final AbstractPrimitiveNode primitiveNode) {
+        private CachedDispatchPrimitiveNode(final int argumentCount, final CompiledCodeObject method, final DispatchPrimitiveNode primitiveNode) {
             super(method);
             this.argumentCount = argumentCount;
             this.primitiveNode = primitiveNode;

diff --git a/...pi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/dispatch/DispatchEagerlyNode.java b/...pi.swa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/dispatch/DispatchEagerlyNode.java
@@ -50,7 +50,7 @@ public static DispatchEagerlyNode create() {
                     limit = "INLINE_CACHE_SIZE", assumptions = {"cachedMethod.getCallTargetStable()", "failureCounter.getAssumption()"}, rewriteOn = PrimitiveFailed.class)
     protected static final Object doPrimitiveEagerly(final VirtualFrame frame, @SuppressWarnings("unused") final CompiledCodeObject method, final Object[] receiverAndArguments,
                     @Cached("method") final CompiledCodeObject cachedMethod,
-                    @Cached("getOrCreateIndexedOrNamed(cachedMethod, PROVIDED_ON_EXECUTE)") final AbstractPrimitiveNode primitiveNode,
+                    @Cached("getOrCreateIndexedOrNamed(cachedMethod)") final AbstractPrimitiveNode primitiveNode,
                     @Cached("create(primitiveNode)") final PrimitiveFailedCounter failureCounter) {
         try {
             return primitiveNode.executeWithArguments(frame, receiverAndArguments);

diff --git a/...wa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/plugins/ffi/PrimExternalCallNode.java b/...wa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/plugins/ffi/PrimExternalCallNode.java
@@ -91,15 +91,10 @@ private static Object lookupModuleLibrary(final SqueakImageContext context, fina
         return moduleLibrary;
     }
 
-    @Override
-    public Object execute(final VirtualFrame frame) {
-        return doExternalCall(frame.materialize());
-    }
-
     @Override
     public Object executeWithArguments(final VirtualFrame frame, final Object... receiverAndArguments) {
         // arguments are handled via manipulation of the stack pointer, see below
-        return execute(frame);
+        return doExternalCall(frame.materialize());
     }
 
     @TruffleBoundary

diff --git a/...wa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/primitives/AbstractPrimitiveNode.java b/...wa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/primitives/AbstractPrimitiveNode.java
@@ -7,18 +7,13 @@
 package de.hpi.swa.trufflesqueak.nodes.primitives;
 
 import com.oracle.truffle.api.CompilerAsserts;
-import com.oracle.truffle.api.dsl.NodeChild;
 import com.oracle.truffle.api.frame.VirtualFrame;
 
 import de.hpi.swa.trufflesqueak.model.CompiledCodeObject;
 import de.hpi.swa.trufflesqueak.nodes.AbstractNode;
-import de.hpi.swa.trufflesqueak.nodes.context.ArgumentNodes.AbstractArgumentNode;
 
-@NodeChild(value = "arguments", type = AbstractArgumentNode[].class)
 public abstract class AbstractPrimitiveNode extends AbstractNode {
 
-    public abstract Object execute(VirtualFrame frame);
-
     public abstract Object executeWithArguments(VirtualFrame frame, Object... receiverAndArguments);
 
     public boolean acceptsMethod(@SuppressWarnings("unused") final CompiledCodeObject method) {

diff --git a/...esqueak/src/de/hpi/swa/trufflesqueak/nodes/primitives/AbstractSingletonPrimitiveNode.java b/...esqueak/src/de/hpi/swa/trufflesqueak/nodes/primitives/AbstractSingletonPrimitiveNode.java
@@ -30,11 +30,6 @@ public final Object executeWithArguments(final VirtualFrame frame, final Object.
         return execute();
     }
 
-    @Override
-    public final Object execute(final VirtualFrame frame) {
-        return execute();
-    }
-
     protected abstract Object execute();
 
     @Override

diff --git a/...wa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/primitives/DispatchPrimitiveNode.java b/...wa.trufflesqueak/src/de/hpi/swa/trufflesqueak/nodes/primitives/DispatchPrimitiveNode.java
@@ -0,0 +1,48 @@
+/*
+ * Copyright (c) 2024 Software Architecture Group, Hasso Plattner Institute
+ * Copyright (c) 2024 Oracle and/or its affiliates
+ *
+ * Licensed under the MIT License.
+ */
+package de.hpi.swa.trufflesqueak.nodes.primitives;
+
+import com.oracle.truffle.api.frame.VirtualFrame;
+import com.oracle.truffle.api.nodes.ExplodeLoop;
+
+import de.hpi.swa.trufflesqueak.nodes.AbstractNode;
+import de.hpi.swa.trufflesqueak.nodes.context.ArgumentNodes.AbstractArgumentNode;
+import de.hpi.swa.trufflesqueak.nodes.primitives.PrimitiveNodeFactory.ArgumentsLocation;
+
+public final class DispatchPrimitiveNode extends AbstractNode {
+    @Child protected AbstractPrimitiveNode primitiveNode;
+    @Children protected AbstractArgumentNode[] argumentNodes;
+
+    private DispatchPrimitiveNode(final AbstractPrimitiveNode primitiveNode, final AbstractArgumentNode[] argumentNodes) {
+        this.primitiveNode = primitiveNode;
+        this.argumentNodes = argumentNodes;
+    }
+
+    public static DispatchPrimitiveNode create(final AbstractPrimitiveNode primitiveNode, final ArgumentsLocation location, final int numReceiverAndArguments) {
+        return new DispatchPrimitiveNode(primitiveNode, createArgumentNodes(location, numReceiverAndArguments));
+    }
+
+    private static AbstractArgumentNode[] createArgumentNodes(final ArgumentsLocation location, final int numReceiverAndArguments) {
+        final AbstractArgumentNode[] argumentNodes = new AbstractArgumentNode[numReceiverAndArguments];
+        final boolean useStack = location == ArgumentsLocation.ON_STACK || location == ArgumentsLocation.ON_STACK_REVERSED;
+        final int offset = location == ArgumentsLocation.ON_STACK_REVERSED ? numReceiverAndArguments : 0;
+        for (int i = 0; i < numReceiverAndArguments; i++) {
+            argumentNodes[i] = AbstractArgumentNode.create(i - offset, useStack);
+        }
+        return argumentNodes;
+    }
+
+    @ExplodeLoop
+    public Object execute(final VirtualFrame frame) {
+        final int numArguments = argumentNodes.length;
+        final Object[] arguments = new Object[numArguments];
+        for (int i = 0; i < numArguments; i++) {
+            arguments[i] = argumentNodes[i].execute(frame);
+        }
+        return primitiveNode.executeWithArguments(frame, arguments);
+    }
+}
Benchmark Name	Min	Geomean	Median	Mean	Max	Total (ms)	Total (min)
Bounce	552	560	554.33	553	554.33	110866	1.85
CD	492	504	495.7	493	495.68	99139	1.65
DeltaBlue	290	488	420.74	417	419.51	84147	1.4
Havlak	1171	1232	1200.15	1204	1200.08	240029	4
Json	368	377	370.93	369	370.91	74185	1.24
List	379	389	379.75	379	379.74	75949	1.27
Mandelbrot	223	234	227.12	226.5	227.09	45423	0.76
NBody	256	270	259.26	258	259.24	51851	0.86
Permute	154	168	155.7	155	155.68	31139	0.52
Queens	231	246	232.89	232	232.88	46578	0.78
Richards	1242	1291	1246.99	1245.5	1246.98	249398	4.16
Sieve	270	286	271.62	271	271.61	54324	0.91
Storage	140	149	141.84	141	141.82	28368	0.47
Towers	201	219	205.05	205	205.03	41009	0.68
	5969	6413	6162.03	6149	6160.59	1232405	20.54
Benchmark Name	Min	Geomean	Median	Mean	Max	Total (ms)	Total (min)
Bounce	552	561	554.64	553	554.63	110928	1.85
CD	489	499	492.54	491	492.53	98508	1.64
DeltaBlue	286	503	422.18	417	420.64	84435	1.41
Havlak	1142	1269	1174.47	1177	1174.39	234894	3.91
Json	385	394	387.72	386	387.71	77544	1.29
List	378	388	378.76	378	378.75	75751	1.26
Mandelbrot	225	235	229.34	231	229.32	45867	0.76
NBody	256	271	259.57	258	259.55	51914	0.87
Permute	154	167	156.01	155	156	31203	0.52
Queens	231	244	233.62	234	233.61	46724	0.78
Richards	1264	1307	1268.54	1269	1268.53	253708	4.23
Sieve	270	282	271.58	271	271.57	54315	0.91
Storage	142	154	143.99	143	143.96	28797	0.48
Towers	196	212	198.11	198	198.1	39622	0.66
	5970	6486	6171.05	6161	6169.3	1234210	20.57