Optimize some commonly used utf8 functions by:
- using counted loops instead of searching for terminating null. In
the important cases the caller already knows the length: change
the API to pass it in. Keep the old API version as well to avoid
extensive changes to non-critical debug and test code.
- ensure the common cases are at the start of if/then/else chains.
Usually 99+% of characters are ASCII even in mixed strings.
- for the "convert" functions, when both utf8 and utf16 lengths are
passed, and are equal, it means the entire string is ASCII, and a
specialized loop can be used. The compiler might then unroll or
even vectorize this.
The functions improved are (tested on Nexus 5 with a 44 character
ASCII string):
CountModifiedUtf8Chars : 20% faster
ConvertUtf16ToModifiedUtf8: 80% faster
ConvertModifiedUtf8ToUtf16: 200% faster
Also for completeness CountUtf8Bytes has been cleaned up a little, but
the speed is unchanged. Unlike CountModifiedUtf8Chars, it was already
passed the length, rather than searching for null.
Change-Id: I1c9b7dea3eda869fc9f5f6b4dd6be8cdd5bc3ac0
diff --git a/runtime/jni_internal.cc b/runtime/jni_internal.cc
index 415109f..5e3fa19 100644
--- a/runtime/jni_internal.cc
+++ b/runtime/jni_internal.cc
@@ -1689,7 +1689,8 @@
} else {
CHECK_NON_NULL_MEMCPY_ARGUMENT(length, buf);
const jchar* chars = s->GetValue();
- ConvertUtf16ToModifiedUtf8(buf, chars + start, length);
+ size_t bytes = CountUtf8Bytes(chars + start, length);
+ ConvertUtf16ToModifiedUtf8(buf, bytes, chars + start, length);
}
}
@@ -1772,7 +1773,7 @@
char* bytes = new char[byte_count + 1];
CHECK(bytes != nullptr); // bionic aborts anyway.
const uint16_t* chars = s->GetValue();
- ConvertUtf16ToModifiedUtf8(bytes, chars, s->GetLength());
+ ConvertUtf16ToModifiedUtf8(bytes, byte_count, chars, s->GetLength());
bytes[byte_count] = '\0';
return bytes;
}
diff --git a/runtime/mirror/string.cc b/runtime/mirror/string.cc
index be869d4..33aca03 100644
--- a/runtime/mirror/string.cc
+++ b/runtime/mirror/string.cc
@@ -109,12 +109,17 @@
String* String::AllocFromModifiedUtf8(Thread* self, const char* utf) {
DCHECK(utf != nullptr);
- size_t char_count = CountModifiedUtf8Chars(utf);
- return AllocFromModifiedUtf8(self, char_count, utf);
+ size_t byte_count = strlen(utf);
+ size_t char_count = CountModifiedUtf8Chars(utf, byte_count);
+ return AllocFromModifiedUtf8(self, char_count, utf, byte_count);
+}
+
+String* String::AllocFromModifiedUtf8(Thread* self, int32_t utf16_length, const char* utf8_data_in) {
+ return AllocFromModifiedUtf8(self, utf16_length, utf8_data_in, strlen(utf8_data_in));
}
String* String::AllocFromModifiedUtf8(Thread* self, int32_t utf16_length,
- const char* utf8_data_in) {
+ const char* utf8_data_in, int32_t utf8_length) {
gc::AllocatorType allocator_type = Runtime::Current()->GetHeap()->GetCurrentAllocator();
SetStringCountVisitor visitor(utf16_length);
String* string = Alloc<true>(self, utf16_length, allocator_type, visitor);
@@ -122,7 +127,7 @@
return nullptr;
}
uint16_t* utf16_data_out = string->GetValue();
- ConvertModifiedUtf8ToUtf16(utf16_data_out, utf8_data_in);
+ ConvertModifiedUtf8ToUtf16(utf16_data_out, utf16_length, utf8_data_in, utf8_length);
return string;
}
@@ -217,7 +222,7 @@
const uint16_t* chars = GetValue();
size_t byte_count = GetUtfLength();
std::string result(byte_count, static_cast<char>(0));
- ConvertUtf16ToModifiedUtf8(&result[0], chars, GetLength());
+ ConvertUtf16ToModifiedUtf8(&result[0], byte_count, chars, GetLength());
return result;
}
diff --git a/runtime/mirror/string.h b/runtime/mirror/string.h
index 80ebd2c..e2cfb8d 100644
--- a/runtime/mirror/string.h
+++ b/runtime/mirror/string.h
@@ -116,6 +116,10 @@
static String* AllocFromModifiedUtf8(Thread* self, const char* utf)
SHARED_REQUIRES(Locks::mutator_lock_) REQUIRES(!Roles::uninterruptible_);
+ static String* AllocFromModifiedUtf8(Thread* self, int32_t utf16_length,
+ const char* utf8_data_in, int32_t utf8_length)
+ SHARED_REQUIRES(Locks::mutator_lock_) REQUIRES(!Roles::uninterruptible_);
+
static String* AllocFromModifiedUtf8(Thread* self, int32_t utf16_length, const char* utf8_data_in)
SHARED_REQUIRES(Locks::mutator_lock_) REQUIRES(!Roles::uninterruptible_);
diff --git a/runtime/utf.cc b/runtime/utf.cc
index 10600e2..5a11698 100644
--- a/runtime/utf.cc
+++ b/runtime/utf.cc
@@ -23,28 +23,50 @@
namespace art {
+// This is used only from debugger and test code.
size_t CountModifiedUtf8Chars(const char* utf8) {
+ return CountModifiedUtf8Chars(utf8, strlen(utf8));
+}
+
+/*
+ * This does not validate UTF8 rules (nor did older code). But it gets the right answer
+ * for valid UTF-8 and that's fine because it's used only to size a buffer for later
+ * conversion.
+ *
+ * Modified UTF-8 consists of a series of bytes up to 21 bit Unicode code points as follows:
+ * U+0001 - U+007F 0xxxxxxx
+ * U+0080 - U+07FF 110xxxxx 10xxxxxx
+ * U+0800 - U+FFFF 1110xxxx 10xxxxxx 10xxxxxx
+ * U+10000 - U+1FFFFF 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+ *
+ * U+0000 is encoded using the 2nd form to avoid nulls inside strings (this differs from
+ * standard UTF-8).
+ * The four byte encoding converts to two utf16 characters.
+ */
+size_t CountModifiedUtf8Chars(const char* utf8, size_t byte_count) {
+ DCHECK_LE(byte_count, strlen(utf8));
size_t len = 0;
- int ic;
- while ((ic = *utf8++) != '\0') {
+ const char* end = utf8 + byte_count;
+ for (; utf8 < end; ++utf8) {
+ int ic = *utf8;
len++;
- if ((ic & 0x80) == 0) {
- // one-byte encoding
+ if (LIKELY((ic & 0x80) == 0)) {
+ // One-byte encoding.
continue;
}
- // two- or three-byte encoding
+ // Two- or three-byte encoding.
utf8++;
if ((ic & 0x20) == 0) {
- // two-byte encoding
+ // Two-byte encoding.
continue;
}
utf8++;
if ((ic & 0x10) == 0) {
- // three-byte encoding
+ // Three-byte encoding.
continue;
}
- // four-byte encoding: needs to be converted into a surrogate
+ // Four-byte encoding: needs to be converted into a surrogate
// pair.
utf8++;
len++;
@@ -52,6 +74,7 @@
return len;
}
+// This is used only from debugger and test code.
void ConvertModifiedUtf8ToUtf16(uint16_t* utf16_data_out, const char* utf8_data_in) {
while (*utf8_data_in != '\0') {
const uint32_t ch = GetUtf16FromUtf8(&utf8_data_in);
@@ -65,13 +88,53 @@
}
}
-void ConvertUtf16ToModifiedUtf8(char* utf8_out, const uint16_t* utf16_in, size_t char_count) {
+void ConvertModifiedUtf8ToUtf16(uint16_t* utf16_data_out, size_t out_chars,
+ const char* utf8_data_in, size_t in_bytes) {
+ const char *in_start = utf8_data_in;
+ const char *in_end = utf8_data_in + in_bytes;
+ uint16_t *out_p = utf16_data_out;
+
+ if (LIKELY(out_chars == in_bytes)) {
+ // Common case where all characters are ASCII.
+ for (const char *p = in_start; p < in_end;) {
+ // Safe even if char is signed because ASCII characters always have
+ // the high bit cleared.
+ *out_p++ = dchecked_integral_cast<uint16_t>(*p++);
+ }
+ return;
+ }
+
+ // String contains non-ASCII characters.
+ for (const char *p = in_start; p < in_end;) {
+ const uint32_t ch = GetUtf16FromUtf8(&p);
+ const uint16_t leading = GetLeadingUtf16Char(ch);
+ const uint16_t trailing = GetTrailingUtf16Char(ch);
+
+ *out_p++ = leading;
+ if (trailing != 0) {
+ *out_p++ = trailing;
+ }
+ }
+}
+
+void ConvertUtf16ToModifiedUtf8(char* utf8_out, size_t byte_count,
+ const uint16_t* utf16_in, size_t char_count) {
+ if (LIKELY(byte_count == char_count)) {
+ // Common case where all characters are ASCII.
+ const uint16_t *utf16_end = utf16_in + char_count;
+ for (const uint16_t *p = utf16_in; p < utf16_end;) {
+ *utf8_out++ = dchecked_integral_cast<char>(*p++);
+ }
+ return;
+ }
+
+ // String contains non-ASCII characters.
while (char_count--) {
const uint16_t ch = *utf16_in++;
if (ch > 0 && ch <= 0x7f) {
*utf8_out++ = ch;
} else {
- // char_count == 0 here implies we've encountered an unpaired
+ // Char_count == 0 here implies we've encountered an unpaired
// surrogate and we have no choice but to encode it as 3-byte UTF
// sequence. Note that unpaired surrogates can occur as a part of
// "normal" operation.
@@ -161,34 +224,31 @@
size_t CountUtf8Bytes(const uint16_t* chars, size_t char_count) {
size_t result = 0;
- while (char_count--) {
+ const uint16_t *end = chars + char_count;
+ while (chars < end) {
const uint16_t ch = *chars++;
- if (ch > 0 && ch <= 0x7f) {
- ++result;
- } else if (ch >= 0xd800 && ch <= 0xdbff) {
- if (char_count > 0) {
+ if (LIKELY(ch != 0 && ch < 0x80)) {
+ result++;
+ continue;
+ }
+ if (ch < 0x800) {
+ result += 2;
+ continue;
+ }
+ if (ch >= 0xd800 && ch < 0xdc00) {
+ if (chars < end) {
const uint16_t ch2 = *chars;
// If we find a properly paired surrogate, we emit it as a 4 byte
// UTF sequence. If we find an unpaired leading or trailing surrogate,
// we emit it as a 3 byte sequence like would have done earlier.
- if (ch2 >= 0xdc00 && ch2 <= 0xdfff) {
+ if (ch2 >= 0xdc00 && ch2 < 0xe000) {
chars++;
- char_count--;
-
result += 4;
- } else {
- result += 3;
+ continue;
}
- } else {
- // This implies we found an unpaired trailing surrogate at the end
- // of a string.
- result += 3;
}
- } else if (ch > 0x7ff) {
- result += 3;
- } else {
- result += 2;
}
+ result += 3;
}
return result;
}
diff --git a/runtime/utf.h b/runtime/utf.h
index 1193d29..03158c4 100644
--- a/runtime/utf.h
+++ b/runtime/utf.h
@@ -40,6 +40,7 @@
* Returns the number of UTF-16 characters in the given modified UTF-8 string.
*/
size_t CountModifiedUtf8Chars(const char* utf8);
+size_t CountModifiedUtf8Chars(const char* utf8, size_t byte_count);
/*
* Returns the number of modified UTF-8 bytes needed to represent the given
@@ -51,6 +52,8 @@
* Convert from Modified UTF-8 to UTF-16.
*/
void ConvertModifiedUtf8ToUtf16(uint16_t* utf16_out, const char* utf8_in);
+void ConvertModifiedUtf8ToUtf16(uint16_t* utf16_out, size_t out_chars,
+ const char* utf8_in, size_t in_bytes);
/*
* Compare two modified UTF-8 strings as UTF-16 code point values in a non-locale sensitive manner
@@ -71,7 +74,8 @@
* this anyway, so if you want a NUL-terminated string, you know where to
* put the NUL byte.
*/
-void ConvertUtf16ToModifiedUtf8(char* utf8_out, const uint16_t* utf16_in, size_t char_count);
+void ConvertUtf16ToModifiedUtf8(char* utf8_out, size_t byte_count,
+ const uint16_t* utf16_in, size_t char_count);
/*
* The java.lang.String hashCode() algorithm.
diff --git a/runtime/utf_test.cc b/runtime/utf_test.cc
index 94a6ea5..64e4eb7 100644
--- a/runtime/utf_test.cc
+++ b/runtime/utf_test.cc
@@ -48,7 +48,7 @@
};
// A test string that contains a UTF-8 encoding of a surrogate pair
-// (code point = U+10400)
+// (code point = U+10400).
static const uint8_t kSurrogateEncoding[] = {
0xed, 0xa0, 0x81,
0xed, 0xb0, 0x80,
@@ -66,13 +66,13 @@
EXPECT_EQ(0, GetTrailingUtf16Char(pair));
EXPECT_ARRAY_POSITION(1, ptr, start);
- // Two byte sequence
+ // Two byte sequence.
pair = GetUtf16FromUtf8(&ptr);
EXPECT_EQ(0xa2, GetLeadingUtf16Char(pair));
EXPECT_EQ(0, GetTrailingUtf16Char(pair));
EXPECT_ARRAY_POSITION(3, ptr, start);
- // Three byte sequence
+ // Three byte sequence.
pair = GetUtf16FromUtf8(&ptr);
EXPECT_EQ(0x20ac, GetLeadingUtf16Char(pair));
EXPECT_EQ(0, GetTrailingUtf16Char(pair));
@@ -84,7 +84,7 @@
EXPECT_EQ(0xdfe0, GetTrailingUtf16Char(pair));
EXPECT_ARRAY_POSITION(10, ptr, start);
- // Null terminator
+ // Null terminator.
pair = GetUtf16FromUtf8(&ptr);
EXPECT_EQ(0, GetLeadingUtf16Char(pair));
EXPECT_EQ(0, GetTrailingUtf16Char(pair));
@@ -117,7 +117,8 @@
ASSERT_EQ(expected.size(), CountUtf8Bytes(&input[0], input.size()));
std::vector<uint8_t> output(expected.size());
- ConvertUtf16ToModifiedUtf8(reinterpret_cast<char*>(&output[0]), &input[0], input.size());
+ ConvertUtf16ToModifiedUtf8(reinterpret_cast<char*>(&output[0]), expected.size(),
+ &input[0], input.size());
EXPECT_EQ(expected, output);
}
@@ -139,10 +140,10 @@
AssertConversion({ 'h', 'e', 'l', 'l', 'o' }, { 0x68, 0x65, 0x6c, 0x6c, 0x6f });
AssertConversion({
- 0xd802, 0xdc02, // Surrogate pair
- 0xdef0, 0xdcff, // Three byte encodings
- 0x0101, 0x0000, // Two byte encodings
- 'p' , 'p' // One byte encoding
+ 0xd802, 0xdc02, // Surrogate pair.
+ 0xdef0, 0xdcff, // Three byte encodings.
+ 0x0101, 0x0000, // Two byte encodings.
+ 'p' , 'p' // One byte encoding.
}, {
0xf0, 0x90, 0xa0, 0x82,
0xed, 0xbb, 0xb0, 0xed, 0xb3, 0xbf,
@@ -160,4 +161,186 @@
AssertConversion({ 'h', 0xdc00, 0xdc00, 'e' }, { 'h', 0xed, 0xb0, 0x80, 0xed, 0xb0, 0x80, 'e' });
}
+// Old versions of functions, here to compare answers with optimized versions.
+
+size_t CountModifiedUtf8Chars_reference(const char* utf8) {
+ size_t len = 0;
+ int ic;
+ while ((ic = *utf8++) != '\0') {
+ len++;
+ if ((ic & 0x80) == 0) {
+ // one-byte encoding
+ continue;
+ }
+ // two- or three-byte encoding
+ utf8++;
+ if ((ic & 0x20) == 0) {
+ // two-byte encoding
+ continue;
+ }
+ utf8++;
+ if ((ic & 0x10) == 0) {
+ // three-byte encoding
+ continue;
+ }
+
+ // four-byte encoding: needs to be converted into a surrogate
+ // pair.
+ utf8++;
+ len++;
+ }
+ return len;
+}
+
+static size_t CountUtf8Bytes_reference(const uint16_t* chars, size_t char_count) {
+ size_t result = 0;
+ while (char_count--) {
+ const uint16_t ch = *chars++;
+ if (ch > 0 && ch <= 0x7f) {
+ ++result;
+ } else if (ch >= 0xd800 && ch <= 0xdbff) {
+ if (char_count > 0) {
+ const uint16_t ch2 = *chars;
+ // If we find a properly paired surrogate, we emit it as a 4 byte
+ // UTF sequence. If we find an unpaired leading or trailing surrogate,
+ // we emit it as a 3 byte sequence like would have done earlier.
+ if (ch2 >= 0xdc00 && ch2 <= 0xdfff) {
+ chars++;
+ char_count--;
+
+ result += 4;
+ } else {
+ result += 3;
+ }
+ } else {
+ // This implies we found an unpaired trailing surrogate at the end
+ // of a string.
+ result += 3;
+ }
+ } else if (ch > 0x7ff) {
+ result += 3;
+ } else {
+ result += 2;
+ }
+ }
+ return result;
+}
+
+static void ConvertUtf16ToModifiedUtf8_reference(char* utf8_out, const uint16_t* utf16_in,
+ size_t char_count) {
+ while (char_count--) {
+ const uint16_t ch = *utf16_in++;
+ if (ch > 0 && ch <= 0x7f) {
+ *utf8_out++ = ch;
+ } else {
+ // Char_count == 0 here implies we've encountered an unpaired
+ // surrogate and we have no choice but to encode it as 3-byte UTF
+ // sequence. Note that unpaired surrogates can occur as a part of
+ // "normal" operation.
+ if ((ch >= 0xd800 && ch <= 0xdbff) && (char_count > 0)) {
+ const uint16_t ch2 = *utf16_in;
+
+ // Check if the other half of the pair is within the expected
+ // range. If it isn't, we will have to emit both "halves" as
+ // separate 3 byte sequences.
+ if (ch2 >= 0xdc00 && ch2 <= 0xdfff) {
+ utf16_in++;
+ char_count--;
+ const uint32_t code_point = (ch << 10) + ch2 - 0x035fdc00;
+ *utf8_out++ = (code_point >> 18) | 0xf0;
+ *utf8_out++ = ((code_point >> 12) & 0x3f) | 0x80;
+ *utf8_out++ = ((code_point >> 6) & 0x3f) | 0x80;
+ *utf8_out++ = (code_point & 0x3f) | 0x80;
+ continue;
+ }
+ }
+
+ if (ch > 0x07ff) {
+ // Three byte encoding.
+ *utf8_out++ = (ch >> 12) | 0xe0;
+ *utf8_out++ = ((ch >> 6) & 0x3f) | 0x80;
+ *utf8_out++ = (ch & 0x3f) | 0x80;
+ } else /*(ch > 0x7f || ch == 0)*/ {
+ // Two byte encoding.
+ *utf8_out++ = (ch >> 6) | 0xc0;
+ *utf8_out++ = (ch & 0x3f) | 0x80;
+ }
+ }
+ }
+}
+
+// Exhaustive test of converting a single code point to UTF-16, then UTF-8, and back again.
+
+static void codePointToSurrogatePair(uint32_t code_point, uint16_t &first, uint16_t &second) {
+ first = (code_point >> 10) + 0xd7c0;
+ second = (code_point & 0x03ff) + 0xdc00;
+}
+
+static void testConversions(uint16_t *buf, int char_count) {
+ char bytes_test[8], bytes_reference[8];
+ uint16_t out_buf_test[4], out_buf_reference[4];
+ int byte_count_test, byte_count_reference;
+ int char_count_test, char_count_reference;
+
+ // Calculate the number of utf-8 bytes for the utf-16 chars.
+ byte_count_reference = CountUtf8Bytes_reference(buf, char_count);
+ byte_count_test = CountUtf8Bytes(buf, char_count);
+ EXPECT_EQ(byte_count_reference, byte_count_test);
+
+ // Convert the utf-16 string to utf-8 bytes.
+ ConvertUtf16ToModifiedUtf8_reference(bytes_reference, buf, char_count);
+ ConvertUtf16ToModifiedUtf8(bytes_test, byte_count_test, buf, char_count);
+ for (int i = 0; i < byte_count_test; ++i) {
+ EXPECT_EQ(bytes_reference[i], bytes_test[i]);
+ }
+
+ // Calculate the number of utf-16 chars from the utf-8 bytes.
+ bytes_reference[byte_count_reference] = 0; // Reference function needs null termination.
+ char_count_reference = CountModifiedUtf8Chars_reference(bytes_reference);
+ char_count_test = CountModifiedUtf8Chars(bytes_test, byte_count_test);
+ EXPECT_EQ(char_count, char_count_reference);
+ EXPECT_EQ(char_count, char_count_test);
+
+ // Convert the utf-8 bytes back to utf-16 chars.
+ // Does not need copied _reference version of the function because the original
+ // function with the old API is retained for debug/testing code.
+ ConvertModifiedUtf8ToUtf16(out_buf_reference, bytes_reference);
+ ConvertModifiedUtf8ToUtf16(out_buf_test, char_count_test, bytes_test, byte_count_test);
+ for (int i = 0; i < char_count_test; ++i) {
+ EXPECT_EQ(buf[i], out_buf_reference[i]);
+ EXPECT_EQ(buf[i], out_buf_test[i]);
+ }
+}
+
+TEST_F(UtfTest, ExhaustiveBidirectionalCodePointCheck) {
+ for (int codePoint = 0; codePoint <= 0x10ffff; ++codePoint) {
+ uint16_t buf[4];
+ if (codePoint <= 0xffff) {
+ if (codePoint >= 0xd800 && codePoint <= 0xdfff) {
+ // According to the Unicode standard, no character will ever
+ // be assigned to these code points, and they can not be encoded
+ // into either utf-16 or utf-8.
+ continue;
+ }
+ buf[0] = 'h';
+ buf[1] = codePoint;
+ buf[2] = 'e';
+ testConversions(buf, 2);
+ testConversions(buf, 3);
+ testConversions(buf + 1, 1);
+ testConversions(buf + 1, 2);
+ } else {
+ buf[0] = 'h';
+ codePointToSurrogatePair(codePoint, buf[1], buf[2]);
+ buf[3] = 'e';
+ testConversions(buf, 2);
+ testConversions(buf, 3);
+ testConversions(buf, 4);
+ testConversions(buf + 1, 1);
+ testConversions(buf + 1, 2);
+ testConversions(buf + 1, 3);
+ }
+ }
+}
+
} // namespace art