Optimize case conversion with icu_case_mapping

src/runtime/runtime-utils.cc

+// Copyright 2016 the V8 project authors. All rights reserved.

[Dan Ehrenberg, 2016/12/02 23:37:39]

Rather than moving this to src/runtime/runtime-utils.cc, what if you just leave
it in runtime-strings.cc and made a runtime-strings.h to export the function, so
it can be used in runtime-i18n.js? Or, to be even cleaner, it could be in a
location like src/string-case.cc.

+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "src/runtime/runtime-utils.h"
+
+#include "src/base/logging.h"
+#include "src/utils.h"
+
+namespace v8 {
+namespace internal {
+
+#ifdef DEBUG
+bool CheckFastAsciiConvert(char* dst, const char* src, int length, bool changed,
+                           bool is_to_lower) {
+  bool expected_changed = false;
+  for (int i = 0; i < length; i++) {
+    if (dst[i] == src[i]) continue;
+    expected_changed = true;
+    if (is_to_lower) {
+      DCHECK('A' <= src[i] && src[i] <= 'Z');
+      DCHECK(dst[i] == src[i] + ('a' - 'A'));
+    } else {
+      DCHECK('a' <= src[i] && src[i] <= 'z');
+      DCHECK(dst[i] == src[i] - ('a' - 'A'));
+    }
+  }
+  return (expected_changed == changed);
+}
+#endif
+
+template <bool is_lower>

[jungshik at Google, 2016/12/02 06:54:41]

Initially, |is_lower| is a function argument, but I switched to a template param
hoping that it may help improve performance (it does seem to improve perf) at
the cost of increased code size. 

[Dan Ehrenberg, 2016/12/02 23:35:08]

Sounds good to me! What kind of effect does it have on performance?

+bool FastAsciiConvert(char* dst, const char* src, int length,
+                      bool* changed_out) {
+#ifdef DEBUG
+  char* saved_dst = dst;
+  const char* saved_src = src;
+#endif
+  DisallowHeapAllocation no_gc;
+  // We rely on the distance between upper and lower case letters
+  // being a known power of 2.
+  DCHECK('a' - 'A' == (1 << 5));
+  // Boundaries for the range of input characters than require conversion.
+  static const char lo = is_lower ? 'A' - 1 : 'a' - 1;
+  static const char hi = is_lower ? 'Z' + 1 : 'z' + 1;
+  bool changed = false;
+  uintptr_t or_acc = 0;
+  const char* const limit = src + length;
+
+  // dst is newly allocated and always aligned.
+  DCHECK(IsAligned(reinterpret_cast<intptr_t>(dst), sizeof(uintptr_t)));
+  // Only attempt processing one word at a time if src is also aligned.
+  if (IsAligned(reinterpret_cast<intptr_t>(src), sizeof(uintptr_t))) {
+    // Process the prefix of the input that requires no conversion one aligned
+    // (machine) word at a time.
+    while (src <= limit - sizeof(uintptr_t)) {
+      const uintptr_t w = *reinterpret_cast<const uintptr_t*>(src);
+      or_acc |= w;
+      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)) {
+      const uintptr_t w = *reinterpret_cast<const uintptr_t*>(src);
+      or_acc |= w;
+      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);
+    }
+  }
+  // Process the last few bytes of the input (or the whole input if
+  // unaligned access is not supported).
+  while (src < limit) {
+    char c = *src;
+    or_acc |= c;
+    if (lo < c && c < hi) {
+      c ^= (1 << 5);
+      changed = true;
+    }
+    *dst = c;
+    ++src;
+    ++dst;
+  }
+
+  if ((or_acc & kAsciiMask) != 0) return false;
+
+  DCHECK(
+      CheckFastAsciiConvert(saved_dst, saved_src, length, changed, is_lower));
+
+  *changed_out = changed;
+  return true;
+}
+
+template bool FastAsciiConvert<false>(char* dst, const char* src, int length,
+                                      bool* changed_out);
+template bool FastAsciiConvert<true>(char* dst, const char* src, int length,
+                                     bool* changed_out);
+
+}  // namespace internal
+}  // namespace v8