ARM64: Use pair memory access in deoptimizer entry

BUG=
R=ulan@chromium.org

Review URL: https://codereview.chromium.org/228573003

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@20613 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/arm64/deoptimizer-arm64.cc

@@ -138,6 +138,43 @@ Code* Deoptimizer::NotifyStubFailureBuiltin() {
 }
 
 
+#define __ masm->
+
+static void CopyRegisterDumpToFrame(MacroAssembler* masm,
+                                    Register frame,
+                                    CPURegList reg_list,
+                                    Register scratch1,
+                                    Register scratch2,
+                                    int src_offset,
+                                    int dst_offset) {
+  int offset0, offset1;
+  CPURegList copy_to_input = reg_list;
+  int reg_count = reg_list.Count();
+  int reg_size = reg_list.RegisterSizeInBytes();
+  for (int i = 0; i < (reg_count / 2); i++) {
+    __ PeekPair(scratch1, scratch2, src_offset + (i * reg_size * 2));
+
+    offset0 = (copy_to_input.PopLowestIndex().code() * reg_size) + dst_offset;
+    offset1 = (copy_to_input.PopLowestIndex().code() * reg_size) + dst_offset;
+
+    if ((offset0 + reg_size) == offset1) {
+      // Registers are adjacent: store in pairs.
+      __ Stp(scratch1, scratch2, MemOperand(frame, offset0));
+    } else {
+      // Registers are not adjacent: store individually.
+      __ Str(scratch1, MemOperand(frame, offset0));
+      __ Str(scratch2, MemOperand(frame, offset1));
+    }
+  }
+  if ((reg_count & 1) != 0) {
+    __ Peek(scratch1, src_offset + (reg_count - 1) * reg_size);
+    offset0 = (copy_to_input.PopLowestIndex().code() * reg_size) + dst_offset;
+    __ Str(scratch1, MemOperand(frame, offset0));
+  }
+}
+
+#undef __
+
 #define __ masm()->
 
 void Deoptimizer::EntryGenerator::Generate() {
@@ -200,25 +237,13 @@ void Deoptimizer::EntryGenerator::Generate() {
   __ Ldr(x1, MemOperand(deoptimizer, Deoptimizer::input_offset()));
 
   // Copy core registers into the input frame.
-  CPURegList copy_to_input = saved_registers;
-  for (int i = 0; i < saved_registers.Count(); i++) {
-    // TODO(all): Look for opportunities to optimize this by using ldp/stp.
-    __ Peek(x2, i * kPointerSize);
-    CPURegister current_reg = copy_to_input.PopLowestIndex();
-    int offset = (current_reg.code() * kPointerSize) +
-        FrameDescription::registers_offset();
-    __ Str(x2, MemOperand(x1, offset));
-  }
+  CopyRegisterDumpToFrame(masm(), x1, saved_registers, x2, x4, 0,
+                          FrameDescription::registers_offset());
 
   // Copy FP registers to the input frame.
-  for (int i = 0; i < saved_fp_registers.Count(); i++) {
-    // TODO(all): Look for opportunities to optimize this by using ldp/stp.
-    int dst_offset = FrameDescription::double_registers_offset() +
-        (i * kDoubleSize);
-    int src_offset = kFPRegistersOffset + (i * kDoubleSize);
-    __ Peek(x2, src_offset);
-    __ Str(x2, MemOperand(x1, dst_offset));
-  }
+  CopyRegisterDumpToFrame(masm(), x1, saved_fp_registers, x2, x4,
+                          kFPRegistersOffset,
+                          FrameDescription::double_registers_offset());
 
   // Remove the bailout id and the saved registers from the stack.
   __ Drop(1 + (kSavedRegistersAreaSize / kXRegSize));