[GR-14690] Performance optimizations mostly for single context mode

timfel · timfel · commit 8b03e0a8d4da · 2019-06-24T21:14:02.000Z
PullRequest: graalpython/545
diff --git a/ci_common/mixins.libsonnet b/ci_common/mixins.libsonnet
@@ -26,7 +26,7 @@ local const = import 'constants.libsonnet';
     linux:: linux,
 
     local linuxBench = linux + {
-        capabilities +: ["no_frequency_scaling", "tmpfs25g", "x62"],
+        capabilities +: ["no_frequency_scaling", "tmpfs25g", "x52"],
     },
     linuxBench:: linuxBench,
 
@@ -122,4 +122,4 @@ local const = import 'constants.libsonnet';
         ],
     },
     bench:: bench,
-}
+}
diff --git a/graalpython/com.oracle.graal.python.shell/src/com/oracle/graal/python/shell/GraalPythonMain.java b/graalpython/com.oracle.graal.python.shell/src/com/oracle/graal/python/shell/GraalPythonMain.java
@@ -200,6 +200,10 @@ protected List<String> preprocessArguments(List<String> givenArgs, Map<String, S
                 case "-dump":
                     if (wantsExperimental) {
                         subprocessArgs.add("Dgraal.Dump=");
+                        subprocessArgs.add("Dgraal.TraceTruffleCompilation=true");
+                        subprocessArgs.add("Dgraal.TraceTruffleInlining=true");
+                        subprocessArgs.add("Dgraal.TraceTruffleTransferToInterpreter=true");
+                        subprocessArgs.add("Dgraal.TruffleBackgroundCompilation=false");
                         inputArgs.remove("-dump");
                     } else {
                         unrecognized.add(arg);
diff --git a/graalpython/com.oracle.graal.python.test/src/com/oracle/graal/python/test/grammar/TestParserTranslator.java b/graalpython/com.oracle.graal.python.test/src/com/oracle/graal/python/test/grammar/TestParserTranslator.java
@@ -52,8 +52,6 @@
 import java.util.Set;
 import java.util.stream.Collectors;
 
-import org.junit.Test;
-
 import com.oracle.graal.python.PythonLanguage;
 import com.oracle.graal.python.builtins.objects.PNone;
 import com.oracle.graal.python.builtins.objects.complex.PComplex;
@@ -119,9 +117,11 @@
 import com.oracle.truffle.api.nodes.Node;
 import com.oracle.truffle.api.nodes.NodeUtil;
 import com.oracle.truffle.api.nodes.RootNode;
-import com.oracle.truffle.api.profiles.BranchProfile;
+import com.oracle.truffle.api.profiles.ConditionProfile;
 import com.oracle.truffle.api.source.Source;
 
+import org.junit.Test;
+
 public class TestParserTranslator {
     PythonContext context;
 
@@ -131,7 +131,7 @@ public TestParserTranslator() {
     }
 
     private static class JUnitRootNode extends PRootNode {
-        private final BranchProfile profile = BranchProfile.create();
+        private final ConditionProfile profile = ConditionProfile.createCountingProfile();
         @Child private ExpressionNode body;
         @Child private CalleeContext calleeContext = CalleeContext.create();
 
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/builtins/modules/PythonCextBuiltins.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/builtins/modules/PythonCextBuiltins.java
@@ -743,7 +743,7 @@ public TruffleObject getCallable() {
 
         @Specialization
         Object doIt(VirtualFrame frame,
-                        @Cached BranchProfile customLocalsProfile,
+                        @Cached("createCountingProfile()") ConditionProfile customLocalsProfile,
                         @Cached CExtNodes.AsPythonObjectNode asPythonObjectNode,
                         @CachedContext(PythonLanguage.class) PythonContext ctx,
                         @Cached PRaiseNode raiseNode) {
@@ -1475,7 +1475,7 @@ abstract static class MethodDescriptorRoot extends PRootNode {
         @Child protected ReadIndexedArgumentNode readSelfNode;
         @Child private CalleeContext calleeContext = CalleeContext.create();
 
-        private final BranchProfile customLocalsProfile = BranchProfile.create();
+        private final ConditionProfile customLocalsProfile = ConditionProfile.createCountingProfile();
         protected final PythonObjectFactory factory;
 
         @TruffleBoundary
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/builtins/objects/cext/PThreadState.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/builtins/objects/cext/PThreadState.java
@@ -84,7 +84,7 @@
 import com.oracle.truffle.api.library.CachedLibrary;
 import com.oracle.truffle.api.library.ExportLibrary;
 import com.oracle.truffle.api.library.ExportMessage;
-import com.oracle.truffle.api.profiles.BranchProfile;
+import com.oracle.truffle.api.profiles.ConditionProfile;
 import com.oracle.truffle.llvm.spi.NativeTypeLibrary;
 
 @ExportLibrary(InteropLibrary.class)
@@ -481,7 +481,7 @@ protected GetTracebackRootNode(TruffleLanguage<?> language) {
         @Child private GetTracebackNode getTracebackNode;
         @Child private CalleeContext calleeContext = CalleeContext.create();
 
-        private final BranchProfile profile = BranchProfile.create();
+        private final ConditionProfile profile = ConditionProfile.createCountingProfile();
 
         @Override
         public Object execute(VirtualFrame frame) {
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/builtins/objects/function/PFunction.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/builtins/objects/function/PFunction.java
@@ -53,7 +53,7 @@ public class PFunction extends PythonObject {
     private final Assumption codeStableAssumption;
     private final Assumption defaultsStableAssumption;
     private final PythonObject globals;
-    private final PCell[] closure;
+    @CompilationFinal(dimensions = 1) private final PCell[] closure;
     private final boolean isStatic;
     @CompilationFinal private PCode code;
     @CompilationFinal(dimensions = 1) private Object[] defaultValues;
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/ModuleRootNode.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/ModuleRootNode.java
@@ -38,14 +38,14 @@
 import com.oracle.truffle.api.frame.FrameDescriptor;
 import com.oracle.truffle.api.frame.FrameSlot;
 import com.oracle.truffle.api.frame.VirtualFrame;
-import com.oracle.truffle.api.profiles.BranchProfile;
+import com.oracle.truffle.api.profiles.ConditionProfile;
 import com.oracle.truffle.api.source.SourceSection;
 
 public class ModuleRootNode extends PClosureRootNode {
     private static final Signature SIGNATURE = new Signature(false, -1, false, new String[0], new String[0]);
     private final String name;
     private final String doc;
-    private final BranchProfile customLocalsProfile = BranchProfile.create();
+    private final ConditionProfile customLocalsProfile = ConditionProfile.createCountingProfile();
     @Child private ExpressionNode body;
     @Child private WriteGlobalNode writeModuleDoc;
     @Child private CalleeContext calleeContext = CalleeContext.create();
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/PClosureFunctionRootNode.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/PClosureFunctionRootNode.java
@@ -40,11 +40,11 @@
  */
 package com.oracle.graal.python.nodes;
 
+import com.oracle.graal.python.PythonLanguage;
 import com.oracle.graal.python.builtins.objects.function.Signature;
 import com.oracle.graal.python.parser.ExecutionCellSlots;
 import com.oracle.truffle.api.Assumption;
 import com.oracle.truffle.api.CompilerDirectives;
-import com.oracle.truffle.api.TruffleLanguage;
 import com.oracle.truffle.api.frame.FrameDescriptor;
 import com.oracle.truffle.api.frame.FrameSlot;
 
@@ -53,7 +53,7 @@ public abstract class PClosureFunctionRootNode extends PClosureRootNode {
     @CompilerDirectives.CompilationFinal(dimensions = 1) protected final Assumption[] cellEffectivelyFinalAssumptions;
     private final Signature signature;
 
-    protected PClosureFunctionRootNode(TruffleLanguage<?> language, FrameDescriptor frameDescriptor, ExecutionCellSlots executionCellSlots, Signature signature) {
+    protected PClosureFunctionRootNode(PythonLanguage language, FrameDescriptor frameDescriptor, ExecutionCellSlots executionCellSlots, Signature signature) {
         super(language, frameDescriptor, executionCellSlots.getFreeVarSlots());
         this.cellVarSlots = executionCellSlots.getCellVarSlots();
         this.cellEffectivelyFinalAssumptions = executionCellSlots.getCellVarAssumptions();
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/PClosureRootNode.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/PClosureRootNode.java
@@ -40,49 +40,69 @@
  */
 package com.oracle.graal.python.nodes;
 
+import com.oracle.graal.python.PythonLanguage;
 import com.oracle.graal.python.builtins.objects.cell.PCell;
 import com.oracle.graal.python.builtins.objects.function.PArguments;
+import com.oracle.truffle.api.Assumption;
 import com.oracle.truffle.api.CompilerDirectives;
-import com.oracle.truffle.api.TruffleLanguage;
 import com.oracle.truffle.api.frame.Frame;
 import com.oracle.truffle.api.frame.FrameDescriptor;
 import com.oracle.truffle.api.frame.FrameSlot;
 import com.oracle.truffle.api.frame.VirtualFrame;
 import com.oracle.truffle.api.nodes.ExplodeLoop;
 
 public abstract class PClosureRootNode extends PRootNode {
+    private static final PCell[] NO_CLOSURE = new PCell[0];
+    private final Assumption singleContextAssumption;
     @CompilerDirectives.CompilationFinal(dimensions = 1) protected final FrameSlot[] freeVarSlots;
+    @CompilerDirectives.CompilationFinal(dimensions = 1) protected PCell[] closure;
     private final int length;
 
-    protected PClosureRootNode(TruffleLanguage<?> language, FrameDescriptor frameDescriptor, FrameSlot[] freeVarSlots) {
+    protected PClosureRootNode(PythonLanguage language, FrameDescriptor frameDescriptor, FrameSlot[] freeVarSlots) {
         super(language, frameDescriptor);
+        this.singleContextAssumption = language.singleContextAssumption;
         this.freeVarSlots = freeVarSlots;
         this.length = freeVarSlots != null ? freeVarSlots.length : 0;
     }
 
     protected void addClosureCellsToLocals(Frame frame) {
-        PCell[] closure = PArguments.getClosure(frame);
-        if (closure != null) {
+        PCell[] frameClosure = PArguments.getClosure(frame);
+        if (frameClosure != null) {
+            if (singleContextAssumption.isValid() && closure == null) {
+                CompilerDirectives.transferToInterpreterAndInvalidate();
+                closure = frameClosure;
+            } else if (closure != NO_CLOSURE && ((!singleContextAssumption.isValid() && closure != null) || closure != frameClosure)) {
+                CompilerDirectives.transferToInterpreterAndInvalidate();
+                closure = NO_CLOSURE;
+            }
             assert freeVarSlots != null : "closure root node: the free var slots cannot be null when the closure is not null";
-            assert closure.length == freeVarSlots.length : "closure root node: the closure must have the same length as the free var slots array";
-            if (freeVarSlots.length < 32) {
-                addClosureCellsToLocalsExploded(frame, closure);
+            assert frameClosure.length == freeVarSlots.length : "closure root node: the closure must have the same length as the free var slots array";
+            if (closure != null && closure != NO_CLOSURE) {
+                if (freeVarSlots.length < 32) {
+                    addClosureCellsToLocalsExploded(frame, closure);
+                } else {
+                    addClosureCellsToLocalsLoop(frame, closure);
+                }
             } else {
-                addClosureCellsToLocalsLoop(frame, closure);
+                if (freeVarSlots.length < 32) {
+                    addClosureCellsToLocalsExploded(frame, frameClosure);
+                } else {
+                    addClosureCellsToLocalsLoop(frame, frameClosure);
+                }
             }
         }
     }
 
-    protected void addClosureCellsToLocalsLoop(Frame frame, PCell[] closure) {
+    protected void addClosureCellsToLocalsLoop(Frame frame, PCell[] frameClosure) {
         for (int i = 0; i < length; i++) {
-            frame.setObject(freeVarSlots[i], closure[i]);
+            frame.setObject(freeVarSlots[i], frameClosure[i]);
         }
     }
 
     @ExplodeLoop
-    protected void addClosureCellsToLocalsExploded(Frame frame, PCell[] closure) {
+    protected void addClosureCellsToLocalsExploded(Frame frame, PCell[] frameClosure) {
         for (int i = 0; i < length; i++) {
-            frame.setObject(freeVarSlots[i], closure[i]);
+            frame.setObject(freeVarSlots[i], frameClosure[i]);
         }
     }
 
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/attributes/ReadAttributeFromDynamicObjectNode.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/attributes/ReadAttributeFromDynamicObjectNode.java
@@ -122,7 +122,8 @@ protected Object readDirectFinal(DynamicObject dynamicObject, Object key,
                     }, //
                     assumptions = {
                                     "layoutAssumption"
-                    })
+                    }, //
+                    replaces = "readDirectFinal")
     protected Object readDirect(DynamicObject dynamicObject, Object key,
                     @Cached("key") Object cachedKey,
                     @Cached("attrKey(cachedKey)") Object attrKey,
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/function/BuiltinFunctionRootNode.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/function/BuiltinFunctionRootNode.java
@@ -48,15 +48,15 @@
 import com.oracle.truffle.api.dsl.NodeFactory;
 import com.oracle.truffle.api.frame.VirtualFrame;
 import com.oracle.truffle.api.nodes.Node;
-import com.oracle.truffle.api.profiles.BranchProfile;
+import com.oracle.truffle.api.profiles.ConditionProfile;
 
 public final class BuiltinFunctionRootNode extends PRootNode {
     private final Signature signature;
     private final Builtin builtin;
     private final String name;
     private final NodeFactory<? extends PythonBuiltinBaseNode> factory;
     private final boolean declaresExplicitSelf;
-    private final BranchProfile customLocalsProfile = BranchProfile.create();
+    private final ConditionProfile customLocalsProfile = ConditionProfile.createCountingProfile();
     @Child private BuiltinCallNode body;
     @Child private CalleeContext calleeContext = CalleeContext.create();
 
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/function/FunctionRootNode.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/function/FunctionRootNode.java
@@ -44,7 +44,7 @@
 import com.oracle.truffle.api.frame.VirtualFrame;
 import com.oracle.truffle.api.nodes.ExplodeLoop;
 import com.oracle.truffle.api.nodes.NodeUtil;
-import com.oracle.truffle.api.profiles.BranchProfile;
+import com.oracle.truffle.api.profiles.ConditionProfile;
 import com.oracle.truffle.api.profiles.ValueProfile;
 import com.oracle.truffle.api.source.SourceSection;
 
@@ -59,7 +59,7 @@ public class FunctionRootNode extends PClosureFunctionRootNode {
     private final SourceSection sourceSection;
     private final boolean isGenerator;
     private final ValueProfile generatorFrameProfile;
-    private final BranchProfile customLocalsProfile = BranchProfile.create();
+    private final ConditionProfile customLocalsProfile = ConditionProfile.createCountingProfile();
 
     @Child private ExpressionNode body;
     @Child private CalleeContext calleeContext = CalleeContext.create();
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/util/ChannelNodes.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/util/ChannelNodes.java
@@ -298,6 +298,8 @@ public abstract static class WriteToChannelNode extends WriteToChannelBaseNode {
 
         @Specialization
         int writeSeekable(SeekableByteChannel channel, SequenceStorage s, int len,
+                        @Cached BranchProfile limitProfile,
+                        @Cached("createBinaryProfile()") ConditionProfile maxSizeProfile,
                         @Cached PRaiseNode raise) {
             long availableSize;
             try {
@@ -306,31 +308,34 @@ int writeSeekable(SeekableByteChannel channel, SequenceStorage s, int len,
                 gotException.enter();
                 throw raise.raise(OSError, e);
             }
-            if (availableSize > MAX_WRITE) {
+            if (maxSizeProfile.profile(availableSize > MAX_WRITE)) {
                 availableSize = MAX_WRITE;
             }
             int sz = (int) Math.min(availableSize, len);
-            return writeWritable(channel, s, sz, raise);
-
+            return writeWritable(channel, s, sz, limitProfile, raise);
         }
 
         @Specialization
         int writeWritable(WritableByteChannel channel, SequenceStorage s, int len,
+                        @Cached BranchProfile limitProfile,
                         @Cached PRaiseNode raise) {
             ByteBuffer src = allocateBuffer(getBytes(s));
             if (src.remaining() > len) {
+                limitProfile.enter();
                 src.limit(len);
             }
             return writeFromBuffer(channel, src, raise);
         }
 
         @Specialization
         int writeGeneric(Channel channel, SequenceStorage s, int len,
+                        @Cached BranchProfile limitProfile,
+                        @Cached("createBinaryProfile()") ConditionProfile maxSizeProfile,
                         @Cached PRaiseNode raise) {
             if (channel instanceof SeekableByteChannel) {
-                return writeSeekable((SeekableByteChannel) channel, s, len, raise);
+                return writeSeekable((SeekableByteChannel) channel, s, len, limitProfile, maxSizeProfile, raise);
             } else if (channel instanceof ReadableByteChannel) {
-                return writeWritable((WritableByteChannel) channel, s, len, raise);
+                return writeWritable((WritableByteChannel) channel, s, len, limitProfile, raise);
             } else {
                 throw raise.raise(OSError, "file not opened for reading");
             }
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/runtime/AsyncHandler.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/runtime/AsyncHandler.java
@@ -67,7 +67,7 @@
 import com.oracle.truffle.api.frame.VirtualFrame;
 import com.oracle.truffle.api.nodes.Node;
 import com.oracle.truffle.api.nodes.Node.Child;
-import com.oracle.truffle.api.profiles.BranchProfile;
+import com.oracle.truffle.api.profiles.ConditionProfile;
 
 /**
  * A handler for asynchronous actions events that need to be handled on a main thread of execution,
@@ -142,7 +142,7 @@ private static class CallRootNode extends PRootNode {
         @Child private ReadCallerFrameNode readCallerFrameNode = ReadCallerFrameNode.create();
         @Child private CalleeContext calleeContext = CalleeContext.create();
 
-        private final BranchProfile profile = BranchProfile.create();
+        private final ConditionProfile profile = ConditionProfile.createCountingProfile();
 
         protected CallRootNode(TruffleLanguage<?> language) {
             super(language);
diff --git a/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/runtime/ExecutionContext.java b/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/runtime/ExecutionContext.java
@@ -58,7 +58,6 @@
 import com.oracle.truffle.api.frame.FrameInstance;
 import com.oracle.truffle.api.frame.VirtualFrame;
 import com.oracle.truffle.api.nodes.Node;
-import com.oracle.truffle.api.profiles.BranchProfile;
 import com.oracle.truffle.api.profiles.ConditionProfile;
 
 /**
@@ -179,16 +178,15 @@ public static final class CalleeContext extends Node {
         /**
          * Wrap the execution of a Python callee called from a Python frame.
          */
-        public static void enter(VirtualFrame frame, BranchProfile profile) {
+        public static void enter(VirtualFrame frame, ConditionProfile profile) {
             // tfel: Create our frame reference here and store it so that
             // there's no reference to it from the caller side.
             PFrame.Reference thisFrameRef = new PFrame.Reference(PArguments.getCallerFrameInfo(frame));
             Object customLocals = PArguments.getCustomLocals(frame);
             PArguments.setCurrentFrameInfo(frame, thisFrameRef);
             // tfel: If there are custom locals, write them into an (incomplete)
             // PFrame here
-            if (customLocals != null && !(customLocals instanceof PFrame.Reference)) {
-                profile.enter();
+            if (profile.profile(customLocals != null && !(customLocals instanceof PFrame.Reference))) {
                 thisFrameRef.setCustomLocals(customLocals);
             }
         }
diff --git a/mx.graalpython/mx_graalpython.py b/mx.graalpython/mx_graalpython.py
@@ -43,7 +43,7 @@
 from mx_gate import Task
 from mx_graalpython_bench_param import PATH_MESO, BENCHMARKS
 from mx_graalpython_benchmark import PythonBenchmarkSuite, python_vm_registry, CPythonVm, PyPyVm, GraalPythonVm, \
-    CONFIGURATION_DEFAULT, CONFIG_EXPERIMENTAL_SPLITTING, CONFIGURATION_SANDBOXED, CONFIGURATION_NATIVE
+    CONFIGURATION_DEFAULT, CONFIG_EXPERIMENTAL_SPLITTING, CONFIGURATION_SANDBOXED, CONFIGURATION_NATIVE, CONFIGURATION_DUMP
 
 SUITE = mx.suite('graalpython')
 SUITE_COMPILER = mx.suite("compiler", fatalIfMissing=False)
@@ -1003,6 +1003,9 @@ def _register_vms(namespace):
     python_vm_registry.add_vm(GraalPythonVm(config_name=CONFIGURATION_NATIVE, extra_polyglot_args=[
         "--llvm.sandboxed=false"
     ]), SUITE, 10)
+    python_vm_registry.add_vm(GraalPythonVm(config_name=CONFIGURATION_DUMP, extra_polyglot_args=[
+        "--experimental-options", "-dump"
+    ]), SUITE, 10)
 
 
 def _register_bench_suites(namespace):
diff --git a/mx.graalpython/mx_graalpython_benchmark.py b/mx.graalpython/mx_graalpython_benchmark.py
@@ -58,6 +58,7 @@
 CONFIGURATION_NATIVE = "native"
 CONFIG_EXPERIMENTAL_SPLITTING = "experimental_splitting"
 CONFIGURATION_SANDBOXED = "sandboxed"
+CONFIGURATION_DUMP = "dump"
 
 DEFAULT_ITERATIONS = 10
 
@@ -311,7 +312,7 @@ def rules(self, output, benchmarks, bm_suite_args):
         ]
 
     def runAndReturnStdOut(self, benchmarks, bmSuiteArgs):
-        # host-vm rewrite rules 
+        # host-vm rewrite rules
         ret_code, out, dims = super(PythonBenchmarkSuite, self).runAndReturnStdOut(benchmarks, bmSuiteArgs)
 
         def _replace_host_vm(key):
