ARM: Small copy optimization. Copying 64bits at a time.

BUG=none
TEST=none

Review URL: https://chromiumcodereview.appspot.com/14121006

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@14254 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/arm/assembler-arm.h

@@ -212,6 +212,7 @@ const Register pc  = { kRegister_pc_Code };
 
 // Single word VFP register.
 struct SwVfpRegister {
+  static const int kSizeInBytes = 4;
   bool is_valid() const { return 0 <= code_ && code_ < 32; }
   bool is(SwVfpRegister reg) const { return code_ == reg.code_; }
   int code() const {
@@ -242,6 +243,7 @@ struct DwVfpRegister {
   static const int kNumReservedRegisters = 2;
   static const int kMaxNumAllocatableRegisters = kMaxNumRegisters -
       kNumReservedRegisters;
+  static const int kSizeInBytes = 8;
 
   // Note: the number of registers can be different at snapshot and run-time.
   // Any code included in the snapshot must be able to run both with 16 or 32

src/arm/code-stubs-arm.cc

@@ -4448,7 +4448,7 @@ void ArgumentsAccessStub::GenerateNewStrict(MacroAssembler* masm) {
       Context::STRICT_MODE_ARGUMENTS_BOILERPLATE_INDEX)));
 
   // Copy the JS object part.
-  __ CopyFields(r0, r4, r3.bit(), JSObject::kHeaderSize / kPointerSize);
+  __ CopyFields(r0, r4, d0, s0, JSObject::kHeaderSize / kPointerSize);
 
   // Get the length (smi tagged) and set that as an in-object property too.
   STATIC_ASSERT(Heap::kArgumentsLengthIndex == 0);

src/arm/full-codegen-arm.cc

@@ -1563,7 +1563,7 @@ void FullCodeGenerator::VisitRegExpLiteral(RegExpLiteral* expr) {
   // r0: Newly allocated regexp.
   // r5: Materialized regexp.
   // r2: temp.
-  __ CopyFields(r0, r5, r2.bit(), size / kPointerSize);
+  __ CopyFields(r0, r5, d0, s0, size / kPointerSize);
   context()->Plug(r0);
 }
 

src/arm/lithium-codegen-arm.cc

@@ -5614,17 +5614,8 @@ void LCodeGen::DoRegExpLiteral(LRegExpLiteral* instr) {
 
   __ bind(&allocated);
   // Copy the content into the newly allocated memory.
-  // (Unroll copy loop once for better throughput).
-  for (int i = 0; i < size - kPointerSize; i += 2 * kPointerSize) {
-    __ ldr(r3, FieldMemOperand(r1, i));
-    __ ldr(r2, FieldMemOperand(r1, i + kPointerSize));
-    __ str(r3, FieldMemOperand(r0, i));
-    __ str(r2, FieldMemOperand(r0, i + kPointerSize));
-  }
-  if ((size % (2 * kPointerSize)) != 0) {
-    __ ldr(r3, FieldMemOperand(r1, size - kPointerSize));
-    __ str(r3, FieldMemOperand(r0, size - kPointerSize));
-  }
+  __ CopyFields(r0, r1, double_scratch0(), double_scratch0().low(),
+                size / kPointerSize);
 }
 
 

src/arm/macro-assembler-arm.cc

@@ -3160,27 +3160,24 @@ void MacroAssembler::AllocateHeapNumberWithValue(Register result,
 // Copies a fixed number of fields of heap objects from src to dst.
 void MacroAssembler::CopyFields(Register dst,
                                 Register src,
-                                RegList temps,
+                                DwVfpRegister double_scratch,
+                                SwVfpRegister single_scratch,
                                 int field_count) {
-  // At least one bit set in the first 15 registers.
-  ASSERT((temps & ((1 << 15) - 1)) != 0);
-  ASSERT((temps & dst.bit()) == 0);
-  ASSERT((temps & src.bit()) == 0);
-  // Primitive implementation using only one temporary register.
-
-  Register tmp = no_reg;
-  // Find a temp register in temps list.
-  for (int i = 0; i < 15; i++) {
-    if ((temps & (1 << i)) != 0) {
-      tmp.set_code(i);
-      break;
-    }
+  int double_count = field_count / (DwVfpRegister::kSizeInBytes / kPointerSize);
+  for (int i = 0; i < double_count; i++) {
+    vldr(double_scratch, FieldMemOperand(src, i * DwVfpRegister::kSizeInBytes));
+    vstr(double_scratch, FieldMemOperand(dst, i * DwVfpRegister::kSizeInBytes));
   }
-  ASSERT(!tmp.is(no_reg));
 
-  for (int i = 0; i < field_count; i++) {
-    ldr(tmp, FieldMemOperand(src, i * kPointerSize));
-    str(tmp, FieldMemOperand(dst, i * kPointerSize));
+  STATIC_ASSERT(SwVfpRegister::kSizeInBytes == kPointerSize);
+  STATIC_ASSERT(2 * SwVfpRegister::kSizeInBytes == DwVfpRegister::kSizeInBytes);
+
+  int remain = field_count % (DwVfpRegister::kSizeInBytes / kPointerSize);
+  if (remain != 0) {
+    vldr(single_scratch,
+         FieldMemOperand(src, (field_count - 1) * kPointerSize));
+    vstr(single_scratch,
+         FieldMemOperand(dst, (field_count - 1) * kPointerSize));
   }
 }
 

src/arm/macro-assembler-arm.h

@@ -743,7 +743,11 @@ class MacroAssembler: public Assembler {
                                    Label* gc_required);
 
   // Copies a fixed number of fields of heap objects from src to dst.
-  void CopyFields(Register dst, Register src, RegList temps, int field_count);
+  void CopyFields(Register dst,
+                  Register src,
+                  DwVfpRegister double_scratch,
+                  SwVfpRegister single_scratch,
+                  int field_count);
 
   // Copies a number of bytes from src to dst. All registers are clobbered. On
   // exit src and dst will point to the place just after where the last byte was