Add optimized version of memcpy on ia32.

src/utils.h

 // Copyright 2006-2008 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 507 matching lines @@
  private:
   Vector<char> buffer_;
   int position_;
 
   bool is_finalized() const { return position_ < 0; }
 
   DISALLOW_IMPLICIT_CONSTRUCTORS(StringBuilder);
 };
 
 
+// Custom memcpy implementation for platforms where the standard version
+// may not be good enough.
+// TODO(lrn): Check whether some IA32 platforms should be excluded.
+#if defined(V8_TARGET_ARCH_IA32)
+
+// TODO(lrn): Extend to other platforms as needed, especially X64.

[Erik Corry, 2010/06/04 07:07:40]

I resent that "especially x64"!

[Lasse Reichstein, 2010/06/04 11:52:13]

You think it's more needed on ARM than on X64? :P

+
+typedef void (*MemCopyFunction)(void* dest, const void* src, size_t size);
+
+// Implemented in codegen-<arch>.cc.
+MemCopyFunction CreateMemCopyFunction();
+
+// Copy memory area to disjoint memory area.
+static inline void MemCopy(void* dest, const void* src, size_t size) {
+  static MemCopyFunction memcopy = CreateMemCopyFunction();
+  (*memcopy)(dest, src, size);
+}
+
+
+// Limit below which the extra overhead of the MemCopy function is likely
+// to outweigh the benefits of faster copying.
+// TODO(lrn): Try to find a more precise value.
+static const int kMinComplexMemCopy = 256;
+
+#else  // V8_TARGET_ARCH_IA32
+
+static inline void MemCopy(void* dest, const void* src, size_t size) {
+  memcpy(dest, src, size);
+}
+
+static const int kMinComplexMemCopy = 256;
+
+#endif  // V8_TARGET_ARCH_IA32
+
+
 // Copy from ASCII/16bit chars to ASCII/16bit chars.
 template <typename sourcechar, typename sinkchar>
 static inline void CopyChars(sinkchar* dest, const sourcechar* src, int chars) {
   sinkchar* limit = dest + chars;
 #ifdef V8_HOST_CAN_READ_UNALIGNED
   if (sizeof(*dest) == sizeof(*src)) {
+    if (chars >= kMinComplexMemCopy) {

[Erik Corry, 2010/06/04 07:07:40]

This should surely be chars * sizeof(*dest).

[Lasse Reichstein, 2010/06/04 11:52:13]

It would make sense to base the decision on the number of bytes moved, and not
the length of the string being copied.

+      MemCopy(dest, src, chars * sizeof(*dest));
+      return;
+    }
     // Number of characters in a uintptr_t.
     static const int kStepSize = sizeof(uintptr_t) / sizeof(*dest);  // NOLINT
     while (dest <= limit - kStepSize) {
       *reinterpret_cast<uintptr_t*>(dest) =
           *reinterpret_cast<const uintptr_t*>(src);
       dest += kStepSize;
       src += kStepSize;
     }
   }
 #endif
@@ skipping 59 matching lines @@
 // Copies data from |src| to |dst|.  The data spans MUST not overlap.
 inline void CopyWords(Object** dst, Object** src, int num_words) {
   ASSERT(Min(dst, src) + num_words <= Max(dst, src));
   ASSERT(num_words > 0);
 
   // Use block copying memcpy if the segment we're copying is
   // enough to justify the extra call/setup overhead.
   static const int kBlockCopyLimit = 16;
 
   if (num_words >= kBlockCopyLimit) {
     memcpy(dst, src, num_words * kPointerSize);

[antonm, 2010/06/03 13:51:21]

may I ask you to switch to your MemCpy here as well?

   } else {
     int remaining = num_words;
     do {
       remaining--;
       *dst++ = *src++;
     } while (remaining > 0);
   }
 }
 
 
@@ skipping 33 matching lines @@
 
   Dest dest;
   memcpy(&dest, &source, sizeof(dest));
   return dest;
 }
 
 } }  // namespace v8::internal
 
 
 #endif  // V8_UTILS_H_