pythongh-126024: optimize UTF-8 decoder for short non-ASCII string (python#126025)

Author: methane

Misc/NEWS.d/next/Core_and_Builtins/2024-10-27-04-47-28.gh-issue-126024.XCQSqT.rst

@@ -0,0 +1,2 @@
+Optimize decoding of short UTF-8 sequences containing non-ASCII characters
+by approximately 15%.

Objects/unicodeobject.c

@@ -4979,39 +4979,228 @@ PyUnicode_DecodeUTF8(const char *s,
 #include "stringlib/codecs.h"
 #include "stringlib/undef.h"
 
+#if (SIZEOF_SIZE_T == 8)
 /* Mask to quickly check whether a C 'size_t' contains a
    non-ASCII, UTF8-encoded char. */
-#if (SIZEOF_SIZE_T == 8)
 # define ASCII_CHAR_MASK 0x8080808080808080ULL
+// used to count codepoints in UTF-8 string.
+# define VECTOR_0101     0x0101010101010101ULL
+# define VECTOR_00FF     0x00ff00ff00ff00ffULL
 #elif (SIZEOF_SIZE_T == 4)
 # define ASCII_CHAR_MASK 0x80808080U
+# define VECTOR_0101     0x01010101U
+# define VECTOR_00FF     0x00ff00ffU
 #else
 # error C 'size_t' size should be either 4 or 8!
 #endif
 
+#if (defined(__clang__) || defined(__GNUC__))
+#define HAVE_CTZ 1
+static inline unsigned int
+ctz(size_t v)
+{
+    return __builtin_ctzll((unsigned long long)v);
+}
+#elif defined(_MSC_VER)
+#define HAVE_CTZ 1
+static inline unsigned int
+ctz(size_t v)
+{
+    unsigned long pos;
+#if SIZEOF_SIZE_T == 4
+    _BitScanForward(&pos, v);
+#else
+    _BitScanForward64(&pos, v);
+#endif /* SIZEOF_SIZE_T */
+    return pos;
+}
+#endif
+
+#if HAVE_CTZ
+// load p[0]..p[size-1] as a little-endian size_t
+// without unaligned access nor read ahead.
+static size_t
+load_unaligned(const unsigned char *p, size_t size)
+{
+    assert(size <= SIZEOF_SIZE_T);
+    union {
+        size_t s;
+        unsigned char b[SIZEOF_SIZE_T];
+    } u;
+    u.s = 0;
+    switch (size) {
+    case 8:
+        u.b[7] = p[7];
+        _Py_FALLTHROUGH;
+    case 7:
+        u.b[6] = p[6];
+        _Py_FALLTHROUGH;
+    case 6:
+        u.b[5] = p[5];
+        _Py_FALLTHROUGH;
+    case 5:
+        u.b[4] = p[4];
+        _Py_FALLTHROUGH;
+    case 4:
+        u.b[3] = p[3];
+        _Py_FALLTHROUGH;
+    case 3:
+        u.b[2] = p[2];
+        _Py_FALLTHROUGH;
+    case 2:
+        u.b[1] = p[1];
+        _Py_FALLTHROUGH;
+    case 1:
+        u.b[0] = p[0];
+        break;
+    case 0:
+        break;
+    default:
+        Py_UNREACHABLE();
+    }
+    return u.s;
+}
+#endif
+
+/*
+ * Find the first non-ASCII character in a byte sequence.
+ *
+ * This function scans a range of bytes from `start` to `end` and returns the
+ * index of the first byte that is not an ASCII character (i.e., has the most
+ * significant bit set). If all characters in the range are ASCII, it returns
+ * `end - start`.
+ */
 static Py_ssize_t
-ascii_decode(const char *start, const char *end, Py_UCS1 *dest)
+find_first_nonascii(const unsigned char *start, const unsigned char *end)
 {
-    const char *p = start;
+    const unsigned char *p = start;
 
+    if (end - start >= SIZEOF_SIZE_T) {
+        const unsigned char *p2 = _Py_ALIGN_UP(p, SIZEOF_SIZE_T);
+        if (p < p2) {
+#if HAVE_CTZ
+#if defined(_M_AMD64) || defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
+            // x86 and amd64 are little endian and can load unaligned memory.
+            size_t u = *(const size_t*)p & ASCII_CHAR_MASK;
+#else
+            size_t u = load_unaligned(p, p2 - p) & ASCII_CHAR_MASK;
+#endif
+            if (u) {
+                return p - start + (ctz(u) - 7) / 8;
+            }
+            p = p2;
+        }
+#else
+        while (p < p2) {
+            if (*p & 0x80) {
+                return p - start;
+            }
+            p++;
+        }
+#endif
+        const unsigned char *e = end - SIZEOF_SIZE_T;
+        while (p <= e) {
+            size_t u = (*(const size_t *)p) & ASCII_CHAR_MASK;
+            if (u) {
+#if PY_LITTLE_ENDIAN && HAVE_CTZ
+                return p - start + (ctz(u) - 7) / 8;
+#else
+                // big endian and minor compilers are difficult to test.
+                // fallback to per byte check.
+                break;
+#endif
+            }
+            p += SIZEOF_SIZE_T;
+        }
+    }
+#if HAVE_CTZ
+    // we can not use *(const size_t*)p to avoid buffer overrun.
+    size_t u = load_unaligned(p, end - p) & ASCII_CHAR_MASK;
+    if (u) {
+        return p - start + (ctz(u) - 7) / 8;
+    }
+    return end - start;
+#else
+    while (p < end) {
+        if (*p & 0x80) {
+            break;
+        }
+        p++;
+    }
+    return p - start;
+#endif
+}
+
+static inline int
+scalar_utf8_start_char(unsigned int ch)
+{
+    // 0xxxxxxx or 11xxxxxx are first byte.
+    return (~ch >> 7 | ch >> 6) & 1;
+}
+
+static inline size_t
+vector_utf8_start_chars(size_t v)
+{
+    return ((~v >> 7) | (v >> 6)) & VECTOR_0101;
+}
+
+
+// Count the number of UTF-8 code points in a given byte sequence.
+static Py_ssize_t
+utf8_count_codepoints(const unsigned char *s, const unsigned char *end)
+{
+    Py_ssize_t len = 0;
+
+    if (end - s >= SIZEOF_SIZE_T) {
+        while (!_Py_IS_ALIGNED(s, ALIGNOF_SIZE_T)) {
+            len += scalar_utf8_start_char(*s++);
+        }
+
+        while (s + SIZEOF_SIZE_T <= end) {
+            const unsigned char *e = end;
+            if (e - s > SIZEOF_SIZE_T * 255) {
+                e = s + SIZEOF_SIZE_T * 255;
+            }
+            Py_ssize_t vstart = 0;
+            while (s + SIZEOF_SIZE_T <= e) {
+                size_t v = *(size_t*)s;
+                size_t vs = vector_utf8_start_chars(v);
+                vstart += vs;
+                s += SIZEOF_SIZE_T;
+            }
+            vstart = (vstart & VECTOR_00FF) + ((vstart >> 8) & VECTOR_00FF);
+            vstart += vstart >> 16;
+#if SIZEOF_SIZE_T == 8
+            vstart += vstart >> 32;
+#endif
+            len += vstart & 0x7ff;
+        }
+    }
+    while (s < end) {
+        len += scalar_utf8_start_char(*s++);
+    }
+    return len;
+}
+
+static Py_ssize_t
+ascii_decode(const char *start, const char *end, Py_UCS1 *dest)
+{
 #if SIZEOF_SIZE_T <= SIZEOF_VOID_P
-    if (_Py_IS_ALIGNED(p, ALIGNOF_SIZE_T)
+    if (_Py_IS_ALIGNED(start, ALIGNOF_SIZE_T)
         && _Py_IS_ALIGNED(dest, ALIGNOF_SIZE_T))
     {
         /* Fast path, see in STRINGLIB(utf8_decode) for
            an explanation. */
-        /* Help allocation */
-        const char *_p = p;
-        Py_UCS1 * q = dest;
-        while (_p + SIZEOF_SIZE_T <= end) {
-            size_t value = *(const size_t *) _p;
+        const char *p = start;
+        Py_UCS1 *q = dest;
+        while (p + SIZEOF_SIZE_T <= end) {
+            size_t value = *(const size_t *) p;
             if (value & ASCII_CHAR_MASK)
                 break;
             *((size_t *)q) = value;
-            _p += SIZEOF_SIZE_T;
+            p += SIZEOF_SIZE_T;
             q += SIZEOF_SIZE_T;
         }
-        p = _p;
         while (p < end) {
             if ((unsigned char)*p & 0x80)
                 break;
@@ -5020,31 +5209,12 @@ ascii_decode(const char *start, const char *end, Py_UCS1 *dest)
         return p - start;
     }
 #endif
-    while (p < end) {
-        /* Fast path, see in STRINGLIB(utf8_decode) in stringlib/codecs.h
-           for an explanation. */
-        if (_Py_IS_ALIGNED(p, ALIGNOF_SIZE_T)) {
-            /* Help allocation */
-            const char *_p = p;
-            while (_p + SIZEOF_SIZE_T <= end) {
-                size_t value = *(const size_t *) _p;
-                if (value & ASCII_CHAR_MASK)
-                    break;
-                _p += SIZEOF_SIZE_T;
-            }
-            p = _p;
-            if (_p == end)
-                break;
-        }
-        if ((unsigned char)*p & 0x80)
-            break;
-        ++p;
-    }
-    memcpy(dest, start, p - start);
-    return p - start;
+    Py_ssize_t pos = find_first_nonascii((const unsigned char*)start,
+                                         (const unsigned char*)end);
+    memcpy(dest, start, pos);
+    return pos;
 }
 
-
 static int
 unicode_decode_utf8_impl(_PyUnicodeWriter *writer,
                          const char *starts, const char *s, const char *end,
@@ -5188,27 +5358,69 @@ unicode_decode_utf8(const char *s, Py_ssize_t size,
         return get_latin1_char((unsigned char)s[0]);
     }
 
-    // fast path: try ASCII string.
-    const char *starts = s;
-    const char *end = s + size;
-    PyObject *u = PyUnicode_New(size, 127);
-    if (u == NULL) {
+    // I don't know this check is necessary or not. But there is a test
+    // case that requires size=PY_SSIZE_T_MAX cause MemoryError.
+    if (PY_SSIZE_T_MAX - sizeof(PyCompactUnicodeObject) < (size_t)size) {
+        PyErr_NoMemory();
         return NULL;
     }
-    Py_ssize_t decoded = ascii_decode(s, end, PyUnicode_1BYTE_DATA(u));
-    if (decoded == size) {
+
+    const char *starts = s;
+    const char *end = s + size;
+
+    Py_ssize_t pos = find_first_nonascii((const unsigned char*)starts, (const unsigned char*)end);
+    if (pos == size) {  // fast path: ASCII string.
+        PyObject *u = PyUnicode_New(size, 127);
+        if (u == NULL) {
+            return NULL;
+        }
+        memcpy(PyUnicode_1BYTE_DATA(u), s, size);
         if (consumed) {
             *consumed = size;
         }
         return u;
     }
-    s += decoded;
-    size -= decoded;
+
+    int maxchr = 127;
+    Py_ssize_t maxsize = size;
+
+    unsigned char ch = (unsigned char)(s[pos]);
+    // error handler other than strict may remove/replace the invalid byte.
+    // consumed != NULL allows 1~3 bytes remainings.
+    // 0x80 <= ch < 0xc2 is invalid start byte that cause UnicodeDecodeError.
+    // otherwise: check the input and decide the maxchr and maxsize to reduce
+    // reallocation and copy.
+    if (error_handler == _Py_ERROR_STRICT && !consumed && ch >= 0xc2) {
+        // we only calculate the number of codepoints and don't determine the exact maxchr.
+        // This is because writing fast and portable SIMD code to find maxchr is difficult.
+        // If reallocation occurs for a larger maxchar, knowing the exact number of codepoints
+        // means that it is no longer necessary to allocate several times the required amount
+        // of memory.
+        maxsize = utf8_count_codepoints((const unsigned char *)s, (const unsigned char *)end);
+        if (ch < 0xc4) { // latin1
+            maxchr = 0xff;
+        }
+        else if (ch < 0xf0) { // ucs2
+            maxchr = 0xffff;
+        }
+        else { // ucs4
+            maxchr = 0x10ffff;
+        }
+    }
+    PyObject *u = PyUnicode_New(maxsize, maxchr);
+    if (!u) {
+        return NULL;
+    }
 
     // Use _PyUnicodeWriter after fast path is failed.
     _PyUnicodeWriter writer;
     _PyUnicodeWriter_InitWithBuffer(&writer, u);
-    writer.pos = decoded;
+    if (maxchr <= 255) {
+        memcpy(PyUnicode_1BYTE_DATA(u), s, pos);
+        s += pos;
+        size -= pos;
+        writer.pos = pos;
+    }
 
     if (unicode_decode_utf8_impl(&writer, starts, s, end,
                                  error_handler, errors,
@@ -5268,7 +5480,9 @@ PyUnicode_DecodeUTF8Stateful(const char *s,
                              const char *errors,
                              Py_ssize_t *consumed)
 {
-    return unicode_decode_utf8(s, size, _Py_ERROR_UNKNOWN, errors, consumed);
+    return unicode_decode_utf8(s, size,
+                               errors ? _Py_ERROR_UNKNOWN : _Py_ERROR_STRICT,
+                               errors, consumed);
 }