Improve FastCloneShallowArray/ObjectStubs with VFP copying on ARM

src/arm/code-stubs-arm.cc

 // Copyright 2012 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 312 matching lines @@
   __ mov(cp, r0);
   __ add(sp, sp, Operand(2 * kPointerSize));
   __ Ret();
 
   // Need to collect. Call into runtime system.
   __ bind(&gc);
   __ TailCallRuntime(Runtime::kPushBlockContext, 2, 1);
 }
 
 
+int FastCloneShallowArrayStub::MaximumClonedLength() {
+  return CpuFeatures::IsSupported(VFP2) ? 128 : 8;

[danno, 2012/10/29 15:29:24]

By increasing this to 128, V8 will generate 128 versions of this stub rather
then 8, which will explode the footprint of clone stubs in the snapshot. 8 was
chosen because it finds the sweet-spot between amount of code generation and the
more frequent cases. Please keep it constant and fixed to 8, I think your
optimization still applies, even for shorter arrays.

+}
+
+
 static void GenerateFastCloneShallowArrayCommon(
     MacroAssembler* masm,
     int length,
     FastCloneShallowArrayStub::Mode mode,
     Label* fail) {
   // Registers on entry:
   //
   // r3: boilerplate literal array.
   ASSERT(mode != FastCloneShallowArrayStub::CLONE_ANY_ELEMENTS);
 
   // All sizes here are multiples of kPointerSize.
   int elements_size = 0;
   if (length > 0) {
     elements_size = mode == FastCloneShallowArrayStub::CLONE_DOUBLE_ELEMENTS
         ? FixedDoubleArray::SizeFor(length)
         : FixedArray::SizeFor(length);
   }
-  int size = JSArray::kSize + elements_size;
+  int elements_offset = JSArray::kSize;
+  int size = elements_offset + elements_size;
 
   // Allocate both the JS array and the elements array in one big
   // allocation. This avoids multiple limit checks.
   __ AllocateInNewSpace(size,
                         r0,
                         r1,
                         r2,
                         fail,
-                        TAG_OBJECT);
+                        NO_ALLOCATION_FLAGS);
 
   // Copy the JS array part.
-  for (int i = 0; i < JSArray::kSize; i += kPointerSize) {
-    if ((i != JSArray::kElementsOffset) || (length == 0)) {
-      __ ldr(r1, FieldMemOperand(r3, i));
-      __ str(r1, FieldMemOperand(r0, i));
+  __ sub(r3, r3, Operand(kHeapObjectTag));
+  STATIC_ASSERT(JSArray::kSize == 4 * kPointerSize);
+  RegList temps = r4.bit() | r5.bit() | r6.bit() | r7.bit();
+  __ ldm(ia_w, r3, temps);
+  if (length > 0) {
+    // Set the pointer to the elements array if it is non-empty. r0 is untagged,
+    // so we need to include the tag here.
+    STATIC_ASSERT(JSArray::kElementsOffset == 2 * kPointerSize);
+    __ add(r6, r0, Operand(elements_offset + kHeapObjectTag));
+  }
+  __ stm(ia_w, r0, temps);
+
+  // Copy the elements from the boilerplate array, if needed. At this point,
+  // r0 points to the beginning of elements.
+  int restore_size = JSArray::kSize - kHeapObjectTag;
+  if (length > 0) {
+    // Load the elements array from the boilerplate.
+    __ ldr(r3, MemOperand(r3, JSArray::kElementsOffset - JSArray::kSize));
+    ASSERT((elements_size % kPointerSize) == 0);
+    if (CpuFeatures::IsSupported(VFP2)) {
+      CpuFeatures::Scope scope(VFP2);
+      SwVfpRegister first_reg = s0;
+      SwVfpRegister last_reg = s15;
+      __ sub(r3, r3, Operand(kHeapObjectTag));
+      __ VFPCopyFields(r0, r3, elements_size / kPointerSize,
+                       first_reg, last_reg);
+      restore_size += elements_size;
+    } else {
+      __ CopyFields(r6, r3, r1.bit(), elements_size / kPointerSize);
     }
   }
 
-  if (length > 0) {
-    // Get hold of the elements array of the boilerplate and setup the
-    // elements pointer in the resulting object.
-    __ ldr(r3, FieldMemOperand(r3, JSArray::kElementsOffset));
-    __ add(r2, r0, Operand(JSArray::kSize));
-    __ str(r2, FieldMemOperand(r0, JSArray::kElementsOffset));
-
-    // Copy the elements array.
-    ASSERT((elements_size % kPointerSize) == 0);
-    __ CopyFields(r2, r3, r1.bit(), elements_size / kPointerSize);
-  }
+  // At this point, r0 points to the end of the copied region. Bring it back to
+  // the beginning and tag it.
+  __ sub(r0, r0, Operand(restore_size));
 }
 
 void FastCloneShallowArrayStub::Generate(MacroAssembler* masm) {
   // Stack layout on entry:
   //
   // [sp]: constant elements.
   // [sp + kPointerSize]: literal index.
   // [sp + (2 * kPointerSize)]: literals array.
 
   // Load boilerplate object into r3 and check if we need to create a
@@ skipping 59 matching lines @@
 
   // Return and remove the on-stack parameters.
   __ add(sp, sp, Operand(3 * kPointerSize));
   __ Ret();
 
   __ bind(&slow_case);
   __ TailCallRuntime(Runtime::kCreateArrayLiteralShallow, 3, 1);
 }
 
 
+int FastCloneShallowObjectStub::MaximumClonedProperties() {
+  return CpuFeatures::IsSupported(VFP2) ? 96 : 6;

[danno, 2012/10/29 15:29:24]

Similarly here, please keep at constant 6, unless you have code that
demonstrates that cases >6 are frequent enough to justify this.

+}
+
+
 void FastCloneShallowObjectStub::Generate(MacroAssembler* masm) {
   // Stack layout on entry:
   //
   // [sp]: object literal flags.
   // [sp + kPointerSize]: constant properties.
   // [sp + (2 * kPointerSize)]: literal index.
   // [sp + (3 * kPointerSize)]: literals array.
 
   // Load boilerplate object into r3 and check if we need to create a
   // boilerplate.
   Label slow_case;
   __ ldr(r3, MemOperand(sp, 3 * kPointerSize));
   __ ldr(r0, MemOperand(sp, 2 * kPointerSize));
   __ add(r3, r3, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
   __ ldr(r3, MemOperand(r3, r0, LSL, kPointerSizeLog2 - kSmiTagSize));
   __ CompareRoot(r3, Heap::kUndefinedValueRootIndex);
   __ b(eq, &slow_case);
 
   // Check that the boilerplate contains only fast properties and we can
   // statically determine the instance size.
   int size = JSObject::kHeaderSize + length_ * kPointerSize;
   __ ldr(r0, FieldMemOperand(r3, HeapObject::kMapOffset));
   __ ldrb(r0, FieldMemOperand(r0, Map::kInstanceSizeOffset));
   __ cmp(r0, Operand(size >> kPointerSizeLog2));
   __ b(ne, &slow_case);
 
   // Allocate the JS object and copy header together with all in-object
   // properties from the boilerplate.
-  __ AllocateInNewSpace(size, r0, r1, r2, &slow_case, TAG_OBJECT);
-  for (int i = 0; i < size; i += kPointerSize) {
-    __ ldr(r1, FieldMemOperand(r3, i));
-    __ str(r1, FieldMemOperand(r0, i));
+  if (CpuFeatures::IsSupported(VFP2)) {
+    CpuFeatures::Scope scope(VFP2);
+    __ AllocateInNewSpace(size, r0, r1, r2, &slow_case, NO_ALLOCATION_FLAGS);
+    SwVfpRegister first_reg = s0;
+    SwVfpRegister last_reg = s15;
+    __ sub(r3, r3, Operand(kHeapObjectTag));
+    __ VFPCopyFields(r0, r3, size / kPointerSize, first_reg, last_reg);
+    __ sub(r0, r0, Operand(size - kHeapObjectTag));
+  } else {
+    __ AllocateInNewSpace(size, r0, r1, r2, &slow_case, TAG_OBJECT);
+    for (int i = 0; i < size; i += kPointerSize) {
+      __ ldr(r1, FieldMemOperand(r3, i));
+      __ str(r1, FieldMemOperand(r0, i));
+    }
   }
 
   // Return and remove the on-stack parameters.
   __ add(sp, sp, Operand(4 * kPointerSize));
   __ Ret();
 
   __ bind(&slow_case);
   __ TailCallRuntime(Runtime::kCreateObjectLiteralShallow, 4, 1);
 }
 
@@ skipping 7112 matching lines @@
 
   __ Pop(lr, r5, r1);
   __ Ret();
 }
 
 #undef __
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_ARM