A64: Clean and improve ArgumentsAdaptorTrampoline.

This patch:
 - uses named registers
 - reserves all the stack space in one go to avoid operations on csp
 - uses LDP/STP to copy the arguments

R=jochen@chromium.org

Review URL: https://codereview.chromium.org/184103003

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@19725 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/a64/builtins-a64.cc

@@ -1364,43 +1364,55 @@ void Builtins::Generate_ArgumentsAdaptorTrampoline(MacroAssembler* masm) {
   //  -- x2 : expected number of arguments
   // -----------------------------------
 
+  Register argc_actual = x0;  // Excluding the receiver.
+  Register argc_expected = x2;  // Excluding the receiver.
+  Register function = x1;
+  Register code_entry = x3;
+
   Label invoke, dont_adapt_arguments;
 
   Label enough, too_few;
-  __ Ldr(x3, FieldMemOperand(x1, JSFunction::kCodeEntryOffset));
-  __ Cmp(x0, x2);
+  __ Ldr(code_entry, FieldMemOperand(function, JSFunction::kCodeEntryOffset));
+  __ Cmp(argc_actual, argc_expected);
   __ B(lt, &too_few);
-  __ Cmp(x2, SharedFunctionInfo::kDontAdaptArgumentsSentinel);
+  __ Cmp(argc_expected, SharedFunctionInfo::kDontAdaptArgumentsSentinel);
   __ B(eq, &dont_adapt_arguments);
 
   {  // Enough parameters: actual >= expected
     EnterArgumentsAdaptorFrame(masm);
 
-    // Calculate copy start address into x10 and end address into x11.
-    // x0: actual number of arguments
-    // x1: function
-    // x2: expected number of arguments
-    // x3: code entry to call
-    __ Add(x10, fp, Operand(x0, LSL, kPointerSizeLog2));
-    // Adjust for return address and receiver
-    __ Add(x10, x10, 2 * kPointerSize);
-    __ Sub(x11, x10, Operand(x2, LSL, kPointerSizeLog2));
+    Register copy_start = x10;
+    Register copy_end = x11;
+    Register copy_to = x12;
+    Register scratch1 = x13, scratch2 = x14;
+
+    __ Lsl(argc_expected, argc_expected, kPointerSizeLog2);
+
+    // Adjust for fp, lr, and the receiver.
+    __ Add(copy_start, fp, 3 * kPointerSize);
+    __ Add(copy_start, copy_start, Operand(argc_actual, LSL, kPointerSizeLog2));
+    __ Sub(copy_end, copy_start, argc_expected);
+    __ Sub(copy_end, copy_end, kPointerSize);
+    __ Mov(copy_to, jssp);
+
+    // Claim space for the arguments, the receiver, and one extra slot.
+    // The extra slot ensures we do not write under jssp. It will be popped
+    // later.
+    __ Add(scratch1, argc_expected, 2 * kPointerSize);
+    __ Claim(scratch1, 1);
 
     // Copy the arguments (including the receiver) to the new stack frame.
-    // x0: actual number of arguments
-    // x1: function
-    // x2: expected number of arguments
-    // x3: code entry to call
-    // x10: copy start address
-    // x11: copy end address
-
-    // TODO(all): Should we push values 2 by 2?
-    Label copy;
-    __ Bind(&copy);
-    __ Cmp(x10, x11);
-    __ Ldr(x12, MemOperand(x10, -kPointerSize, PostIndex));
-    __ Push(x12);
-    __ B(gt, &copy);
+    Label copy_2_by_2;
+    __ Bind(&copy_2_by_2);
+    __ Ldp(scratch1, scratch2,
+           MemOperand(copy_start, - 2 * kPointerSize, PreIndex));
+    __ Stp(scratch1, scratch2,
+           MemOperand(copy_to, - 2 * kPointerSize, PreIndex));
+    __ Cmp(copy_start, copy_end);
+    __ B(hi, &copy_2_by_2);
+
+    // Correct the space allocated for the extra slot.
+    __ Drop(1);
 
     __ B(&invoke);
   }
@@ -1409,52 +1421,57 @@ void Builtins::Generate_ArgumentsAdaptorTrampoline(MacroAssembler* masm) {
     __ Bind(&too_few);
     EnterArgumentsAdaptorFrame(masm);
 
-    // Calculate copy start address into x10 and copy end address into x11.
-    // x0: actual number of arguments
-    // x1: function
-    // x2: expected number of arguments
-    // x3: code entry to call
-    // Adjust for return address.
-    __ Add(x11, fp, 1 * kPointerSize);
-    __ Add(x10, x11, Operand(x0, LSL, kPointerSizeLog2));
-    __ Add(x10, x10, 1 * kPointerSize);
+    Register copy_from = x10;
+    Register copy_end = x11;
+    Register copy_to = x12;
+    Register scratch1 = x13, scratch2 = x14;
+
+    __ Lsl(argc_expected, argc_expected, kPointerSizeLog2);
+    __ Lsl(argc_actual, argc_actual, kPointerSizeLog2);
+
+    // Adjust for fp, lr, and the receiver.
+    __ Add(copy_from, fp, 3 * kPointerSize);
+    __ Add(copy_from, copy_from, argc_actual);
+    __ Mov(copy_to, jssp);
+    __ Sub(copy_end, copy_to, 1 * kPointerSize);   // Adjust for the receiver.
+    __ Sub(copy_end, copy_end, argc_actual);
+
+    // Claim space for the arguments, the receiver, and one extra slot.
+    // The extra slot ensures we do not write under jssp. It will be popped
+    // later.
+    __ Add(scratch1, argc_expected, 2 * kPointerSize);
+    __ Claim(scratch1, 1);
 
     // Copy the arguments (including the receiver) to the new stack frame.
-    // x0: actual number of arguments
-    // x1: function
-    // x2: expected number of arguments
-    // x3: code entry to call
-    // x10: copy start address
-    // x11: copy end address
-    Label copy;
-    __ Bind(&copy);
-    __ Ldr(x12, MemOperand(x10, -kPointerSize, PostIndex));
-    __ Push(x12);
-    __ Cmp(x10, x11);  // Compare before moving to next argument.
-    __ B(ne, &copy);
+    Label copy_2_by_2;
+    __ Bind(&copy_2_by_2);
+    __ Ldp(scratch1, scratch2,
+           MemOperand(copy_from, - 2 * kPointerSize, PreIndex));
+    __ Stp(scratch1, scratch2,
+           MemOperand(copy_to, - 2 * kPointerSize, PreIndex));
+    __ Cmp(copy_to, copy_end);
+    __ B(hi, &copy_2_by_2);
+
+    __ Mov(copy_to, copy_end);
 
     // Fill the remaining expected arguments with undefined.
-    // x0: actual number of arguments
-    // x1: function
-    // x2: expected number of arguments
-    // x3: code entry to call
-    __ LoadRoot(x10, Heap::kUndefinedValueRootIndex);
-    __ Sub(x11, fp, Operand(x2, LSL, kPointerSizeLog2));
-    // Adjust for the arguments adaptor frame and already pushed receiver.
-    __ Sub(x11, x11,
-           StandardFrameConstants::kFixedFrameSizeFromFp + (2 * kPointerSize));
-
-    // TODO(all): Optimize this to use ldp?
+    __ LoadRoot(scratch1, Heap::kUndefinedValueRootIndex);
+    __ Add(copy_end, jssp, kPointerSize);
+
     Label fill;
     __ Bind(&fill);
-    __ Push(x10);
-    __ Cmp(jssp, x11);
-    __ B(ne, &fill);
+    __ Stp(scratch1, scratch1,
+           MemOperand(copy_to, - 2 * kPointerSize, PreIndex));
+    __ Cmp(copy_to, copy_end);
+    __ B(hi, &fill);
+
+    // Correct the space allocated for the extra slot.
+    __ Drop(1);
   }
 
   // Arguments have been adapted. Now call the entry point.
   __ Bind(&invoke);
-  __ Call(x3);
+  __ Call(code_entry);
 
   // Store offset of return address for deoptimizer.
   masm->isolate()->heap()->SetArgumentsAdaptorDeoptPCOffset(masm->pc_offset());
@@ -1465,7 +1482,7 @@ void Builtins::Generate_ArgumentsAdaptorTrampoline(MacroAssembler* masm) {
 
   // Call the entry point without adapting the arguments.
   __ Bind(&dont_adapt_arguments);
-  __ Jump(x3);
+  __ Jump(code_entry);
 }