Optimize rb_str_set_len()

* Avoid walking the characters, as this is called repeatedly during Zlib
  decompression and causes Bundler to be very slow when downloading the
  list of gems.
* The CodeRange needs to be CR_VALID, otherwise zlib breaks.
* A better solution would be to make CodeRange and character length lazy
  for NativeRope.

Author: eregon

src/main/java/org/truffleruby/cext/CExtNodes.java

@@ -28,6 +28,7 @@
 import com.oracle.truffle.api.profiles.ConditionProfile;
 import com.oracle.truffle.api.source.SourceSection;
 import org.jcodings.Encoding;
+import org.jcodings.specific.ASCIIEncoding;
 import org.jcodings.specific.UTF8Encoding;
 import org.truffleruby.Layouts;
 import org.truffleruby.RubyContext;
@@ -63,6 +64,7 @@
 import org.truffleruby.core.string.StringSupport;
 import org.truffleruby.interop.ToJavaStringNodeGen;
 import org.truffleruby.language.LexicalScope;
+import org.truffleruby.language.NotOptimizedWarningNode;
 import org.truffleruby.language.NotProvided;
 import org.truffleruby.language.RubyBaseNode;
 import org.truffleruby.language.RubyGuards;
@@ -443,15 +445,43 @@ public long capacity(DynamicObject string,
     @CoreMethod(names = "rb_str_set_len", onSingleton = true, required = 2, lowerFixnum = 2)
     public abstract static class RbStrSetLenNode extends CoreMethodArrayArgumentsNode {
 
+        @Child NotOptimizedWarningNode notOptimizedWarningNode;
+
         @Specialization
-        public DynamicObject strSetLen(DynamicObject string, int len,
-                @Cached("create()") StringToNativeNode stringToNativeNode) {
+        public DynamicObject strSetLen(DynamicObject string, int newByteLength,
+                @Cached("create()") StringToNativeNode stringToNativeNode,
+                @Cached("createBinaryProfile()") ConditionProfile binaryProfile) {
             final NativeRope nativeRope = stringToNativeNode.executeToNative(string);
-            final NativeRope newNativeRope = nativeRope.withByteLength(len);
+            final NativeRope newNativeRope;
+
+            if (binaryProfile.profile(nativeRope.getEncoding() == ASCIIEncoding.INSTANCE)) {
+                // TODO (eregon, 17 Jan 2019): We use CR_VALID here as zlib relies on it. Computing
+                // the coderange here on every call is too slow. A proper fix is to compute the
+                // CodeRange lazily on NativeRope.
+                newNativeRope = nativeRope.withByteLength(newByteLength, newByteLength, CodeRange.CR_VALID);
+            } else {
+                performanceWarn("calling rb_str_set_len() on non-binary string is not yet optimized");
+
+                final byte[] bytes = new byte[newByteLength];
+                nativeRope.getNativePointer().readBytes(0, bytes, 0, newByteLength);
+                final long packedLengthAndCodeRange = RopeOperations.calculateCodeRangeAndLength(nativeRope.getEncoding(), bytes, 0, newByteLength);
+                final CodeRange codeRange = CodeRange.fromInt(StringSupport.unpackArg(packedLengthAndCodeRange));
+                final int characterLength = StringSupport.unpackResult(packedLengthAndCodeRange);
+                newNativeRope = nativeRope.withByteLength(newByteLength, characterLength, codeRange);
+            }
+
             StringOperations.setRope(string, newNativeRope);
             return string;
         }
 
+        private void performanceWarn(String message) {
+            if (notOptimizedWarningNode == null) {
+                CompilerDirectives.transferToInterpreterAndInvalidate();
+                notOptimizedWarningNode = insert(new NotOptimizedWarningNode());
+            }
+            notOptimizedWarningNode.warn(message);
+        }
+
     }
 
     @CoreMethod(names = "rb_str_resize", onSingleton = true, required = 2, lowerFixnum = 2)

src/main/java/org/truffleruby/core/rope/NativeRope.java

@@ -11,7 +11,6 @@
 
 import org.jcodings.Encoding;
 import org.truffleruby.core.FinalizationService;
-import org.truffleruby.core.string.StringSupport;
 
 import com.oracle.truffle.api.CompilerDirectives.TruffleBoundary;
 import org.truffleruby.extra.ffi.Pointer;
@@ -44,18 +43,9 @@ private NativeRope(Pointer pointer, int byteLength, Encoding encoding, int chara
         this.pointer = pointer;
     }
 
-    @TruffleBoundary
-    public NativeRope withByteLength(int newByteLength) {
-        final byte[] bytes = new byte[newByteLength];
-        pointer.readBytes(0, bytes, 0, newByteLength);
-
-        final long packedLengthAndCodeRange = RopeOperations.calculateCodeRangeAndLength(getEncoding(), bytes, 0, newByteLength);
-        final CodeRange codeRange = CodeRange.fromInt(StringSupport.unpackArg(packedLengthAndCodeRange));
-        final int characterLength = StringSupport.unpackResult(packedLengthAndCodeRange);
-
-        getNativePointer().writeByte(newByteLength, (byte) 0); // Like MRI
-
-        return new NativeRope(getNativePointer(), newByteLength, getEncoding(), characterLength, codeRange);
+    public NativeRope withByteLength(int newByteLength, int characterLength, CodeRange codeRange) {
+        pointer.writeByte(newByteLength, (byte) 0); // Like MRI
+        return new NativeRope(pointer, newByteLength, getEncoding(), characterLength, codeRange);
     }
 
     public NativeRope makeCopy(FinalizationService finalizationService) {