[intl] Avoid a temporary allocation while converting to ICU string

If we need to build an icu::UnicodeString for a string that is currently
in one-byte representation, we first have to expand the string's content
into a two-byte representation. Doing so involves allocating an array,
which is slow. With this change, we can convert short strings on the
stack instead to save time. The cutoff length for what counts as "short"
is pretty arbitrary, but we believe many strings fit into an 80-column
line. This increases the score of cdjs in JetStream 2 by 35% on my
machine, because cdjs is basically a test of localeCompare throughput.

Bug: v8:9305

Change-Id: Iba081ac5a8fa7659edf06ac97ba8acf3f8328d59
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1630848
Commit-Queue: Seth Brenith <seth.brenith@microsoft.com>
Reviewed-by: Jakob Kummerow <jkummerow@chromium.org>
Reviewed-by: Benedikt Meurer <bmeurer@chromium.org>
Cr-Commit-Position: refs/heads/master@{#61901}

src/objects/intl-objects.cc

@@ -196,10 +196,20 @@ icu::UnicodeString Intl::ToICUUnicodeString(Isolate* isolate,
   DCHECK(string->IsFlat());
   DisallowHeapAllocation no_gc;
   std::unique_ptr<uc16[]> sap;
-  return icu::UnicodeString(
-      GetUCharBufferFromFlat(string->GetFlatContent(no_gc), &sap,
-                             string->length()),
-      string->length());
+  // Short one-byte strings can be expanded on the stack to avoid allocating a
+  // temporary buffer.
+  constexpr int kShortStringSize = 80;
+  UChar short_string_buffer[kShortStringSize];
+  const UChar* uchar_buffer = nullptr;
+  const String::FlatContent& flat = string->GetFlatContent(no_gc);
+  int32_t length = string->length();
+  if (flat.IsOneByte() && length <= kShortStringSize) {
+    CopyChars(short_string_buffer, flat.ToOneByteVector().begin(), length);
+    uchar_buffer = short_string_buffer;
+  } else {
+    uchar_buffer = GetUCharBufferFromFlat(flat, &sap, length);
+  }
+  return icu::UnicodeString(uchar_buffer, length);
 }
 
 namespace {