Add back ascii fast path for toupper/tolower

src/runtime.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 5785 matching lines @@
     // become garbage; there is no reason to keep two identical strings
     // alive.
     return s;
   }
 }
 
 
 namespace {
 
 static const uintptr_t kOneInEveryByte = kUintptrAllBitsSet / 0xFF;
-
+#ifdef ENABLE_LATIN_1
+static const uintptr_t kAsciiMask = kOneInEveryByte << 7;
+#endif
 
 // Given a word and two range boundaries returns a word with high bit
 // set in every byte iff the corresponding input byte was strictly in
 // the range (m, n). All the other bits in the result are cleared.
 // This function is only useful when it can be inlined and the
 // boundaries are statically known.
 // Requires: all bytes in the input word and the boundaries must be
 // ASCII (less than 0x7F).
 static inline uintptr_t AsciiRangeMask(uintptr_t w, char m, char n) {
   // Every byte in an ASCII string is less than or equal to 0x7F.
   ASSERT((w & (kOneInEveryByte * 0x7F)) == w);
   // Use strict inequalities since in edge cases the function could be
   // further simplified.
-  ASSERT(0 < m && m < n && n < 0x7F);
+  ASSERT(0 < m && m < n);
+#ifndef ENABLE_LATIN_1
+  ASSERT(n < 0x7F);
+#endif
   // Has high bit set in every w byte less than n.
   uintptr_t tmp1 = kOneInEveryByte * (0x7F + n) - w;
   // Has high bit set in every w byte greater than m.
   uintptr_t tmp2 = w + kOneInEveryByte * (0x7F - m);
   return (tmp1 & tmp2 & (kOneInEveryByte * 0x80));
 }
 
 
 enum AsciiCaseConversion {
   ASCII_TO_LOWER,
   ASCII_TO_UPPER
 };
 
 
 template <AsciiCaseConversion dir>
 struct FastAsciiConverter {
+#ifdef ENABLE_LATIN_1
+  static bool Convert(char* dst, char* src, int length, bool* changed_out) {
+#else
   static bool Convert(char* dst, char* src, int length) {
+#endif
 #ifdef DEBUG
     char* saved_dst = dst;
     char* saved_src = src;
 #endif
     // We rely on the distance between upper and lower case letters
     // being a known power of 2.
     ASSERT('a' - 'A' == (1 << 5));
     // Boundaries for the range of input characters than require conversion.
     const char lo = (dir == ASCII_TO_LOWER) ? 'A' - 1 : 'a' - 1;
     const char hi = (dir == ASCII_TO_LOWER) ? 'Z' + 1 : 'z' + 1;
     bool changed = false;
+#ifdef ENABLE_LATIN_1
+    uintptr_t or_acc = 0;
+#endif
     char* const limit = src + length;
 #ifdef V8_HOST_CAN_READ_UNALIGNED
     // Process the prefix of the input that requires no conversion one
     // (machine) word at a time.
     while (src <= limit - sizeof(uintptr_t)) {
       uintptr_t w = *reinterpret_cast<uintptr_t*>(src);
+#ifdef ENABLE_LATIN_1
+      or_acc |= w;
+#endif
       if (AsciiRangeMask(w, lo, hi) != 0) {
         changed = true;
         break;
       }
       *reinterpret_cast<uintptr_t*>(dst) = w;
       src += sizeof(uintptr_t);
       dst += sizeof(uintptr_t);
     }
     // Process the remainder of the input performing conversion when
     // required one word at a time.
     while (src <= limit - sizeof(uintptr_t)) {
       uintptr_t w = *reinterpret_cast<uintptr_t*>(src);
+#ifdef ENABLE_LATIN_1
+      or_acc |= w;
+#endif
       uintptr_t m = AsciiRangeMask(w, lo, hi);
       // The mask has high (7th) bit set in every byte that needs
       // conversion and we know that the distance between cases is
       // 1 << 5.
       *reinterpret_cast<uintptr_t*>(dst) = w ^ (m >> 2);
       src += sizeof(uintptr_t);
       dst += sizeof(uintptr_t);
     }
 #endif
     // Process the last few bytes of the input (or the whole input if
     // unaligned access is not supported).
     while (src < limit) {
       char c = *src;
+#ifdef ENABLE_LATIN_1
+      or_acc |= c;
+#endif
       if (lo < c && c < hi) {
         c ^= (1 << 5);
         changed = true;
       }
       *dst = c;
       ++src;
       ++dst;
     }
+#ifdef ENABLE_LATIN_1
+    if ((or_acc & kAsciiMask) != 0) {

[Yang, 2013/01/14 15:56:51]

Would it be noticeably more expensive to do this check for every word instead of
accumulating the words (bitwise-and, compare and jump as opposed to bitwise-or)?
The tradeoff is that we can terminate early for the non-ascii case.

+      return false;
+    }
+#endif
 #ifdef DEBUG
     CheckConvert(saved_dst, saved_src, length, changed);
 #endif
+#ifdef ENABLE_LATIN_1
+    *changed_out = changed;
+    return true;
+#else
     return changed;
+#endif
   }
 
 #ifdef DEBUG
   static void CheckConvert(char* dst, char* src, int length, bool changed) {
     bool expected_changed = false;
     for (int i = 0; i < length; i++) {
       if (dst[i] == src[i]) continue;
       expected_changed = true;
       if (dir == ASCII_TO_LOWER) {
         ASSERT('A' <= src[i] && src[i] <= 'Z');
@@ skipping 32 matching lines @@
     Isolate* isolate,
     unibrow::Mapping<typename ConvertTraits::UnibrowConverter, 128>* mapping) {
   NoHandleAllocation ha;
   CONVERT_ARG_CHECKED(String, s, 0);
   s = s->TryFlattenGetString();
 
   const int length = s->length();
   // Assume that the string is not empty; we need this assumption later
   if (length == 0) return s;
 
-#ifndef ENABLE_LATIN_1
   // Simpler handling of ASCII strings.
   //
   // NOTE: This assumes that the upper/lower case of an ASCII
   // character is also ASCII.  This is currently the case, but it
   // might break in the future if we implement more context and locale
   // dependent upper/lower conversions.
   if (s->IsSeqOneByteString()) {
     Object* o;
     { MaybeObject* maybe_o = isolate->heap()->AllocateRawOneByteString(length);
       if (!maybe_o->ToObject(&o)) return maybe_o;
     }
     SeqOneByteString* result = SeqOneByteString::cast(o);
+#ifndef ENABLE_LATIN_1
     bool has_changed_character = ConvertTraits::AsciiConverter::Convert(
         reinterpret_cast<char*>(result->GetChars()),
         reinterpret_cast<char*>(SeqOneByteString::cast(s)->GetChars()),
         length);
     return has_changed_character ? result : s;
+#else
+    bool has_changed_character;
+    bool is_ascii = ConvertTraits::AsciiConverter::Convert(
+        reinterpret_cast<char*>(result->GetChars()),
+        reinterpret_cast<char*>(SeqOneByteString::cast(s)->GetChars()),
+        length,
+        &has_changed_character);
+    // If not ASCII, we discard the result and take the 2 byte path.
+    if (is_ascii) {

[Yang, 2013/01/14 15:56:51]

It looks like we could save time if Convert failed early if the result is not
ascii.

+      return has_changed_character ? result : s;
+    }
+#endif
   }
-#endif
 
   Object* answer;
   { MaybeObject* maybe_answer =
         ConvertCaseHelper(isolate, s, length, length, mapping);
     if (!maybe_answer->ToObject(&answer)) return maybe_answer;
   }
   if (answer->IsSmi()) {
     // Retry with correct length.
     { MaybeObject* maybe_answer =
           ConvertCaseHelper(isolate,
@@ skipping 7669 matching lines @@
     // Handle last resort GC and make sure to allow future allocations
     // to grow the heap without causing GCs (if possible).
     isolate->counters()->gc_last_resort_from_js()->Increment();
     isolate->heap()->CollectAllGarbage(Heap::kNoGCFlags,
                                        "Runtime::PerformGC");
   }
 }
 
 
 } }  // namespace v8::internal