Avoid syncing locals back to frame if not accessed

Author: msimacek

graalpython/com.oracle.graal.python/src/com/oracle/graal/python/builtins/objects/frame/FrameBuiltins.java

@@ -306,7 +306,9 @@ Object getUpdating(VirtualFrame frame, PFrame self,
                 PFrame pyFrame = materializeNode.executeOnStack(false, true, frame);
                 assert pyFrame == self;
             }
-            return getFrameLocalsNode.execute(inliningTarget, self);
+            Object locals = getFrameLocalsNode.execute(inliningTarget, self);
+            self.setLocalsAccessed(true);
+            return locals;
         }
     }
 

graalpython/com.oracle.graal.python/src/com/oracle/graal/python/builtins/objects/frame/PFrame.java

@@ -86,6 +86,7 @@ public final class PFrame extends PythonBuiltinObject {
     public static final int NO_JUMP = -2;
     private int jumpDestLine = DISALLOW_JUMPS;
     private Object localTraceFun = null;
+    private boolean localsAccessed;
 
     private boolean traceLine = true;
 
@@ -119,6 +120,14 @@ public void setJumpDestLine(int jumpDestLine) {
         this.jumpDestLine = jumpDestLine;
     }
 
+    public boolean localsAccessed() {
+        return localsAccessed;
+    }
+
+    public void setLocalsAccessed(boolean localsAccessed) {
+        this.localsAccessed = localsAccessed;
+    }
+
     // TODO: frames: this is a large object, think about how to make this
     // smaller
     public static final class Reference {

graalpython/com.oracle.graal.python/src/com/oracle/graal/python/nodes/bytecode/PBytecodeRootNode.java

@@ -615,6 +615,7 @@ public final class PBytecodeRootNode extends PRootNode implements BytecodeOSRNod
     private static final byte TRACE_PROFILE_LINE = 1;
     private static final byte TRACE_PROFILE_NEW_FRAME = 1 << 1;
     private static final byte TRACE_PROFILE_EXISTING_FRAME = 1 << 2;
+    private static final byte TRACE_PROFILE_SYNC_LOCALS_BACK = 1 << 3;
     @CompilationFinal(dimensions = 1) byte[] traceProfileData;
 
     @CompilationFinal private Object osrMetadata;
@@ -3254,7 +3255,7 @@ private void invokeTraceFunction(VirtualFrame virtualFrame, Object arg, PythonCo
                 pyFrame.setLineLock(line);
             }
             Object result = doInvokeTraceFunction(event, pyFrame, traceFn, nonNullArg);
-            syncLocalsBackToFrame(virtualFrame, pyFrame);
+            syncLocalsBackToFrame(virtualFrame, pyFrame, bci);
             // https://github.com/python/cpython/issues/104232
             if (useLocalFn) {
                 Object realResult = result == PNone.NONE ? traceFn : result;
@@ -3279,15 +3280,19 @@ private void invokeTraceFunction(VirtualFrame virtualFrame, Object arg, PythonCo
     private static Object doInvokeTraceFunction(PythonContext.TraceEvent event, PFrame pyFrame, Object traceFn, Object nonNullArg) {
         // Force locals dict sync, so that we can sync them back later
         GetFrameLocalsNode.executeUncached(pyFrame);
+        pyFrame.setLocalsAccessed(false);
         return CallTernaryMethodNode.getUncached().execute(null, traceFn, pyFrame, event.pythonName, nonNullArg);
     }
 
-    private void syncLocalsBackToFrame(VirtualFrame virtualFrame, PFrame pyFrame) {
+    private void syncLocalsBackToFrame(VirtualFrame virtualFrame, PFrame pyFrame, int bci) {
         Frame localFrame = virtualFrame;
         if (co.isGeneratorOrCoroutine()) {
             localFrame = PArguments.getGeneratorFrame(virtualFrame);
         }
-        GetFrameLocalsNode.syncLocalsBackToFrame(co, this, pyFrame, localFrame);
+        if (pyFrame.localsAccessed()) {
+            enterTraceProfile(bci, TRACE_PROFILE_SYNC_LOCALS_BACK);
+            GetFrameLocalsNode.syncLocalsBackToFrame(co, this, pyFrame, localFrame);
+        }
     }
 
     private void profileCEvent(VirtualFrame virtualFrame, Object callable, PythonContext.ProfileEvent event, MutableLoopData mutableData, byte tracingOrProfilingEnabled, int bci) {
@@ -3327,7 +3332,7 @@ private void invokeProfileFunction(VirtualFrame virtualFrame, Object arg, Python
 
         try {
             Object result = doInvokeProfileFunction(arg, event, pyFrame, profileFun);
-            syncLocalsBackToFrame(virtualFrame, pyFrame);
+            syncLocalsBackToFrame(virtualFrame, pyFrame, bci);
             Object realResult = result == PNone.NONE ? null : result;
             pyFrame.setLocalTraceFun(realResult);
         } catch (Throwable e) {
@@ -3342,6 +3347,7 @@ private void invokeProfileFunction(VirtualFrame virtualFrame, Object arg, Python
     private static Object doInvokeProfileFunction(Object arg, PythonContext.ProfileEvent event, PFrame pyFrame, Object profileFun) {
         // Force locals dict sync, so that we can sync them back later
         GetFrameLocalsNode.executeUncached(pyFrame);
+        pyFrame.setLocalsAccessed(false);
         return CallTernaryMethodNode.getUncached().execute(null, profileFun, pyFrame, event.name, arg == null ? PNone.NONE : arg);
     }