Use generated Neon version of MemCopy() on ARM, if platform supports it.

src/arm/codegen-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 96 matching lines @@
   OS::ProtectCode(buffer, actual_size);
 
 #if !defined(USE_SIMULATOR)
   return FUNCTION_CAST<UnaryMathFunction>(buffer);
 #else
   fast_exp_arm_machine_code = buffer;
   return &fast_exp_simulator;
 #endif
 }
 
+static void MemCopyWrapper(void* dest, const void* src, size_t size) {
+  memcpy(dest, src, size);
+}
+
+// Based on Bionic's memcpy.s
+OS::MemCopyFunction CreateMemCopyFunction() {
+  size_t actual_size;
+  static const int kCacheLineSize = 64;
+  static const int kPrefetchDistance = kCacheLineSize * 4;
+  // Allocate buffer in executable space.
+  byte* buffer = static_cast<byte*>(OS::Allocate(1 * KB,
+                                                 &actual_size,
+                                                 true));
+  if (buffer == NULL) return &MemCopyWrapper;
+  if (!CpuFeatures::IsSupported(NEON)) return &MemCopyWrapper;
+
+  MacroAssembler masm(NULL, buffer, static_cast<int>(actual_size));
+
+  CpuFeatureScope use_neon(&masm, NEON);
+  Label less16, aligned16, aligned8, skip_copy8, skip_copy4,
+        fix_remainder, main_loop, has32, less32;

[danno, 2013/03/27 08:40:07]

coding style: 4-char indent on line continuation.

[Nike, 2013/03/29 08:48:52]

Done.

+
+  // ----------- S t a t e -------------
+  //  -- r0    : dest
+  //  -- r1    : src
+  //  -- r2    : count
+  // -----------------------------------
+
+  __ push(lr);
+
+  // Start preloading as early as possible
+  __ pld(r1, kCacheLineSize * 0);
+  __ pld(r1, kCacheLineSize * 1);
+
+  // Do we have at least 16-bytes to copy (needed for alignment below)
+  __ cmp(r2, Operand(16));
+  __ b(&less16, lo);

[danno, 2013/03/27 08:40:07]

Note that the per-OS generated mem copy doesn't need to support copies of less
than a specific size (OS::kMinComplexMemCopy). You should take that into account
here (i.e. not handle the smaller sizes at all) to save complexity of this code.
Feel free to add a debug-only assert that detects copy sizes that are too small.

[Nike, 2013/03/29 08:48:52]

Done.

+
+  // Align destination to half cache-line for the write-buffer
+  __ rsb(r3, r0, Operand(0));
+  __ and_(r3, r3, Operand(0xf), SetCC);
+  __ b(&aligned16, eq);
+
+  // Copy up to 15-bytes (count in r3)
+  __ sub(r2, r2, r3);
+  __ mov(ip, Operand(r3, LSL, 31), SetCC);
+  __ ldrb(lr, MemOperand(r1, 1, PostIndex), mi);
+  __ strb(lr, MemOperand(r0, 1, PostIndex), mi);
+  __ ldrb(ip, MemOperand(r1, 1, PostIndex), cs);
+  __ ldrb(lr, MemOperand(r1, 1, PostIndex), cs);
+  __ strb(ip, MemOperand(r0, 1, PostIndex), cs);
+  __ strb(lr, MemOperand(r0, 1, PostIndex), cs);
+  __ mov(ip, Operand(r3, LSL, 29), SetCC);
+  __ b(&aligned8, ge);
+  // Copies 4 bytes, destination 32-bits aligned
+  __ vld4(8, r1, d0, element_0, Writeback);
+  __ vst4(8, r0, d0, element_0, Writeback, 32 / 8);
+__ bind(&aligned8);
+  __ b(&aligned16, cc);
+  // Copies 8 bytes, destination 64-bits aligned
+  __ vld1(8, r1, d0, d0, Writeback);
+  __ vst1(8, r0, d0, d0, Writeback, 64 / 8);
+
+__ bind(&aligned16);
+  // Preload immediately the next cache line, which we may need
+  __ pld(r1, kCacheLineSize * 0);
+  __ pld(r1, kCacheLineSize * 1);
+
+  // Make sure we have at least 64 bytes to copy
+  __ sub(r2, r2, Operand(64), SetCC);
+  __ b(&fix_remainder, lo);
+
+  // Preload all the cache lines we need.
+  // NOTE: the number of pld below depends on PREFETCH_DISTANCE,
+  // ideally would would increase the distance in the main loop to
+  // avoid the goofy code below. In practice this doesn't seem to make
+  // a big difference.
+  __ pld(r1, kCacheLineSize * 2);
+  __ pld(r1, kCacheLineSize * 3);
+  __ pld(r1, kPrefetchDistance);
+
+  // The main loop copies 64 bytes at a time
+__ bind(&main_loop);
+  __ vld1(8, r1, d0, d3, Writeback);
+  __ vld1(8, r1, d4, d7, Writeback);
+  __ pld(r1, kPrefetchDistance);
+  __ sub(r2, r2, Operand(64), SetCC);
+  __ vst1(8, r0, d0, d3, Writeback, 128 / 8);
+  __ vst1(8, r0, d4, d7, Writeback, 128 / 8);
+  __ b(&main_loop, hs);
+
+  // Fix-up the remaining count and make sure we have >= 32 bytes left
+__ bind(&fix_remainder);
+  __ add(r2, r2, Operand(64));
+  __ sub(r2, r2, Operand(32), SetCC);
+  __ b(&less32, lo);
+
+  // 32 bytes at a time. These cache lines were already preloaded
+__ bind(&has32);
+  __ vld1(8, r1, d0, d3, Writeback);
+  __ sub(r2, r2, Operand(32), SetCC);
+  __ vst1(8, r0, d0, d3, Writeback, 128 / 8);
+  __ b(&has32, hs);
+
+  // Less than 32 left
+__ bind(&less32);
+  __ add(r2, r2, Operand(32));
+  __ tst(r2, Operand(0x10));
+  __ b(&less16, eq);
+  // Copies 16 bytes, 128-bits aligned
+  __ vld1(8, r1, d0, d1, Writeback);
+  __ vst1(8, r0, d0, d1, Writeback, 128 / 8);
+
+  // copy up to 15-bytes (count in r2)
+__ bind(&less16);
+  __ mov(ip, Operand(r2, LSL, 29), SetCC);
+  __ b(&skip_copy8, cc);
+  __ vld1(8, r1, d0, d0, Writeback);
+  __ vst1(8, r0, d0, d0, Writeback);
+__ bind(&skip_copy8);
+  __ b(&skip_copy4, ge);
+  __ vld4(8, r1, d0, element_0, Writeback);
+  __ vst4(8, r0, d0, element_0, Writeback);
+__ bind(&skip_copy4);
+  __ mov(ip, Operand(r2, LSL, 31), SetCC);
+  __ ldrb(r3, MemOperand(r1, 1, PostIndex), mi);
+  __ ldrb(ip, MemOperand(r1, 1, PostIndex), cs);
+  __ ldrb(lr, MemOperand(r1, 1, PostIndex), cs);
+  __ strb(r3, MemOperand(r0, 1, PostIndex), mi);
+  __ strb(ip, MemOperand(r0, 1, PostIndex), cs);
+  __ strb(lr, MemOperand(r0, 1, PostIndex), cs);
+
+  __ pop(lr);
+  __ bx(lr);
+
+  CodeDesc desc;
+  masm.GetCode(&desc);
+  ASSERT(!RelocInfo::RequiresRelocation(desc));
+
+  CPU::FlushICache(buffer, actual_size);
+  OS::ProtectCode(buffer, actual_size);
+  return FUNCTION_CAST<OS::MemCopyFunction>(buffer);
+}
 
 #undef __
 
-

[danno, 2013/03/27 08:40:07]

Please remove this extraneous whitespace change.

 UnaryMathFunction CreateSqrtFunction() {
   return &sqrt;
 }
 
 // -------------------------------------------------------------------------
 // Platform-specific RuntimeCallHelper functions.
 
 void StubRuntimeCallHelper::BeforeCall(MacroAssembler* masm) const {
   masm->EnterFrame(StackFrame::INTERNAL);
   ASSERT(!masm->has_frame());
@@ skipping 568 matching lines @@
     patcher.masm()->add(r0, pc, Operand(-8));
     patcher.masm()->ldr(pc, MemOperand(pc, -4));
     patcher.masm()->dd(reinterpret_cast<uint32_t>(stub->instruction_start()));
   }
 }
 
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_ARM