Issue #14624: UTF-16 decoding is now 3x to 4x faster on various inputs.

Patch by Serhiy Storchaka.

Misc/NEWS

@@ -10,6 +10,9 @@ What's New in Python 3.3.0 Alpha 4?
 Core and Builtins
 -----------------
 
+- Issue #14624: UTF-16 decoding is now 3x to 4x faster on various inputs.
+  Patch by Serhiy Storchaka.
+
 - asdl_seq and asdl_int_seq are now Py_ssize_t sized.
 
 - Issue #14133 (PEP 415): Implement suppression of __context__ display with an

Objects/stringlib/codecs.h

@@ -215,7 +215,6 @@ InvalidContinuation:
     goto Return;
 }
 
-#undef LONG_PTR_MASK
 #undef ASCII_CHAR_MASK
 
 
@@ -415,4 +414,152 @@ STRINGLIB(utf8_encoder)(PyObject *unicode,
 #undef MAX_SHORT_UNICHARS
 }
 
+/* The pattern for constructing UCS2-repeated masks. */
+#if SIZEOF_LONG == 8
+# define UCS2_REPEAT_MASK 0x0001000100010001ul
+#elif SIZEOF_LONG == 4
+# define UCS2_REPEAT_MASK 0x00010001ul
+#else
+# error C 'long' size should be either 4 or 8!
+#endif
+
+/* The mask for fast checking. */
+#if STRINGLIB_SIZEOF_CHAR == 1
+/* The mask for fast checking of whether a C 'long' contains a
+   non-ASCII or non-Latin1 UTF16-encoded characters. */
+# define FAST_CHAR_MASK         (UCS2_REPEAT_MASK * (0xFFFFu & ~STRINGLIB_MAX_CHAR))
+#else
+/* The mask for fast checking of whether a C 'long' may contain
+   UTF16-encoded surrogate characters. This is an efficient heuristic,
+   assuming that non-surrogate characters with a code point >= 0x8000 are
+   rare in most input.
+*/
+# define FAST_CHAR_MASK         (UCS2_REPEAT_MASK * 0x8000u)
+#endif
+/* The mask for fast byte-swapping. */
+#define STRIPPED_MASK           (UCS2_REPEAT_MASK * 0x00FFu)
+/* Swap bytes. */
+#define SWAB(value)             ((((value) >> 8) & STRIPPED_MASK) | \
+                                 (((value) & STRIPPED_MASK) << 8))
+
+Py_LOCAL_INLINE(Py_UCS4)
+STRINGLIB(utf16_decode)(const unsigned char **inptr, const unsigned char *e,
+                        STRINGLIB_CHAR *dest, Py_ssize_t *outpos,
+                        int native_ordering)
+{
+    Py_UCS4 ch;
+    const unsigned char *aligned_end =
+            (const unsigned char *) ((size_t) e & ~LONG_PTR_MASK);
+    const unsigned char *q = *inptr;
+    STRINGLIB_CHAR *p = dest + *outpos;
+    /* Offsets from q for retrieving byte pairs in the right order. */
+#ifdef BYTEORDER_IS_LITTLE_ENDIAN
+    int ihi = !!native_ordering, ilo = !native_ordering;
+#else
+    int ihi = !native_ordering, ilo = !!native_ordering;
+#endif
+    --e;
+
+    while (q < e) {
+        Py_UCS4 ch2;
+        /* First check for possible aligned read of a C 'long'. Unaligned
+           reads are more expensive, better to defer to another iteration. */
+        if (!((size_t) q & LONG_PTR_MASK)) {
+            /* Fast path for runs of in-range non-surrogate chars. */
+            register const unsigned char *_q = q;
+            while (_q < aligned_end) {
+                unsigned long block = * (unsigned long *) _q;
+                if (native_ordering) {
+                    /* Can use buffer directly */
+                    if (block & FAST_CHAR_MASK)
+                        break;
+                }
+                else {
+                    /* Need to byte-swap */
+                    if (block & SWAB(FAST_CHAR_MASK))
+                        break;
+#if STRINGLIB_SIZEOF_CHAR == 1
+                    block >>= 8;
+#else
+                    block = SWAB(block);
+#endif
+                }
+#ifdef BYTEORDER_IS_LITTLE_ENDIAN
+# if SIZEOF_LONG == 4
+                p[0] = (STRINGLIB_CHAR)(block & 0xFFFFu);
+                p[1] = (STRINGLIB_CHAR)(block >> 16);
+# elif SIZEOF_LONG == 8
+                p[0] = (STRINGLIB_CHAR)(block & 0xFFFFu);
+                p[1] = (STRINGLIB_CHAR)((block >> 16) & 0xFFFFu);
+                p[2] = (STRINGLIB_CHAR)((block >> 32) & 0xFFFFu);
+                p[3] = (STRINGLIB_CHAR)(block >> 48);
+# endif
+#else
+# if SIZEOF_LONG == 4
+                p[0] = (STRINGLIB_CHAR)(block >> 16);
+                p[1] = (STRINGLIB_CHAR)(block & 0xFFFFu);
+# elif SIZEOF_LONG == 8
+                p[0] = (STRINGLIB_CHAR)(block >> 48);
+                p[1] = (STRINGLIB_CHAR)((block >> 32) & 0xFFFFu);
+                p[2] = (STRINGLIB_CHAR)((block >> 16) & 0xFFFFu);
+                p[3] = (STRINGLIB_CHAR)(block & 0xFFFFu);
+# endif
+#endif
+                _q += SIZEOF_LONG;
+                p += SIZEOF_LONG / 2;
+            }
+            q = _q;
+            if (q >= e)
+                break;
+        }
+
+        ch = (q[ihi] << 8) | q[ilo];
+        q += 2;
+        if (!Py_UNICODE_IS_SURROGATE(ch)) {
+#if STRINGLIB_SIZEOF_CHAR < 2
+            if (ch > STRINGLIB_MAX_CHAR)
+                /* Out-of-range */
+                goto Return;
+#endif
+            *p++ = (STRINGLIB_CHAR)ch;
+            continue;
+        }
+
+        /* UTF-16 code pair: */
+        if (q >= e)
+            goto UnexpectedEnd;
+        if (!Py_UNICODE_IS_HIGH_SURROGATE(ch))
+            goto IllegalEncoding;
+        ch2 = (q[ihi] << 8) | q[ilo];
+        q += 2;
+        if (!Py_UNICODE_IS_LOW_SURROGATE(ch2))
+            goto IllegalSurrogate;
+        ch = Py_UNICODE_JOIN_SURROGATES(ch, ch2);
+#if STRINGLIB_SIZEOF_CHAR < 4
+        /* Out-of-range */
+        goto Return;
+#else
+        *p++ = (STRINGLIB_CHAR)ch;
+#endif
+    }
+    ch = 0;
+Return:
+    *inptr = q;
+    *outpos = p - dest;
+    return ch;
+UnexpectedEnd:
+    ch = 1;
+    goto Return;
+IllegalEncoding:
+    ch = 2;
+    goto Return;
+IllegalSurrogate:
+    ch = 3;
+    goto Return;
+}
+#undef UCS2_REPEAT_MASK
+#undef FAST_CHAR_MASK
+#undef STRIPPED_MASK
+#undef SWAB
+#undef LONG_PTR_MASK
 #endif /* STRINGLIB_IS_UNICODE */

Objects/unicodeobject.c

@@ -5195,25 +5195,6 @@ PyUnicode_DecodeUTF16(const char *s,
     return PyUnicode_DecodeUTF16Stateful(s, size, errors, byteorder, NULL);
 }
 
-/* Two masks for fast checking of whether a C 'long' may contain
-   UTF16-encoded surrogate characters. This is an efficient heuristic,
-   assuming that non-surrogate characters with a code point >= 0x8000 are
-   rare in most input.
-   FAST_CHAR_MASK is used when the input is in native byte ordering,
-   SWAPPED_FAST_CHAR_MASK when the input is in byteswapped ordering.
-*/
-#if (SIZEOF_LONG == 8)
-# define FAST_CHAR_MASK         0x8000800080008000L
-# define SWAPPED_FAST_CHAR_MASK 0x0080008000800080L
-# define STRIPPED_MASK          0x00FF00FF00FF00FFL
-#elif (SIZEOF_LONG == 4)
-# define FAST_CHAR_MASK         0x80008000L
-# define SWAPPED_FAST_CHAR_MASK 0x00800080L
-# define STRIPPED_MASK          0x00FF00FFL
-#else
-# error C 'long' size should be either 4 or 8!
-#endif
-
 PyObject *
 PyUnicode_DecodeUTF16Stateful(const char *s,
                               Py_ssize_t size,
@@ -5226,30 +5207,15 @@ PyUnicode_DecodeUTF16Stateful(const char *s,
     Py_ssize_t endinpos;
     Py_ssize_t outpos;
     PyObject *unicode;
-    const unsigned char *q, *e, *aligned_end;
+    const unsigned char *q, *e;
     int bo = 0;       /* assume native ordering by default */
-    int native_ordering = 0;
+    int native_ordering;
     const char *errmsg = "";
-    /* Offsets from q for retrieving byte pairs in the right order. */
-#ifdef BYTEORDER_IS_LITTLE_ENDIAN
-    int ihi = 1, ilo = 0;
-#else
-    int ihi = 0, ilo = 1;
-#endif
     PyObject *errorHandler = NULL;
     PyObject *exc = NULL;
 
-    /* Note: size will always be longer than the resulting Unicode
-       character count */
-    unicode = PyUnicode_New(size, 127);
-    if (!unicode)
-        return NULL;
-    if (size == 0)
-        return unicode;
-    outpos = 0;
-
     q = (unsigned char *)s;
-    e = q + size - 1;
+    e = q + size;
 
     if (byteorder)
         bo = *byteorder;
@@ -5258,155 +5224,98 @@ PyUnicode_DecodeUTF16Stateful(const char *s,
        byte order setting accordingly. In native mode, the leading BOM
        mark is skipped, in all other modes, it is copied to the output
        stream as-is (giving a ZWNBSP character). */
-    if (bo == 0) {
+    if (bo == 0 && size >= 2) {
-        if (size >= 2) {
+        const Py_UCS4 bom = (q[1] << 8) | q[0];
-            const Py_UCS4 bom = (q[ihi] << 8) | q[ilo];
+        if (bom == 0xFEFF) {
-#ifdef BYTEORDER_IS_LITTLE_ENDIAN
+            q += 2;
-            if (bom == 0xFEFF) {
+            bo = -1;
-                q += 2;
-                bo = -1;
-            }
-            else if (bom == 0xFFFE) {
-                q += 2;
-                bo = 1;
-            }
-#else
-            if (bom == 0xFEFF) {
-                q += 2;
-                bo = 1;
-            }
-            else if (bom == 0xFFFE) {
-                q += 2;
-                bo = -1;
-            }
-#endif
         }
+        else if (bom == 0xFFFE) {
+            q += 2;
+            bo = 1;
+        }
+        if (byteorder)
+            *byteorder = bo;
     }
 
-    if (bo == -1) {
+    if (q == e) {
-        /* force LE */
+        if (consumed)
-        ihi = 1;
+            *consumed = size;
-        ilo = 0;
+        Py_INCREF(unicode_empty);
-    }
+        return unicode_empty;
-    else if (bo == 1) {
-        /* force BE */
-        ihi = 0;
-        ilo = 1;
     }
+
 #ifdef BYTEORDER_IS_LITTLE_ENDIAN
-    native_ordering = ilo < ihi;
+    native_ordering = bo <= 0;
 #else
-    native_ordering = ilo > ihi;
+    native_ordering = bo >= 0;
 #endif
 
-    aligned_end = (const unsigned char *) ((size_t) e & ~LONG_PTR_MASK);
+    /* Note: size will always be longer than the resulting Unicode
-    while (q < e) {
+       character count */
-        Py_UCS4 ch;
+    unicode = PyUnicode_New((e - q + 1) / 2, 127);
-        /* First check for possible aligned read of a C 'long'. Unaligned
+    if (!unicode)
-           reads are more expensive, better to defer to another iteration. */
+        return NULL;
-        if (!((size_t) q & LONG_PTR_MASK)) {
+
-            /* Fast path for runs of non-surrogate chars. */
+    outpos = 0;
-            register const unsigned char *_q = q;
+    while (1) {
+        Py_UCS4 ch = 0;
+        if (e - q >= 2) {
             int kind = PyUnicode_KIND(unicode);
-            void *data = PyUnicode_DATA(unicode);
+            if (kind == PyUnicode_1BYTE_KIND) {
-            while (_q < aligned_end) {
+                if (PyUnicode_IS_ASCII(unicode))
-                unsigned long block = * (unsigned long *) _q;
+                    ch = asciilib_utf16_decode(&q, e,
-                Py_UCS4 maxch;
+                            PyUnicode_1BYTE_DATA(unicode), &outpos,
-                if (native_ordering) {
+                            native_ordering);
-                    /* Can use buffer directly */
+                else
-                    if (block & FAST_CHAR_MASK)
+                    ch = ucs1lib_utf16_decode(&q, e,
-                        break;
+                            PyUnicode_1BYTE_DATA(unicode), &outpos,
-                }
+                            native_ordering);
-                else {
+            } else if (kind == PyUnicode_2BYTE_KIND) {
-                    /* Need to byte-swap */
+                ch = ucs2lib_utf16_decode(&q, e,
-                    if (block & SWAPPED_FAST_CHAR_MASK)
+                        PyUnicode_2BYTE_DATA(unicode), &outpos,
-                        break;
+                        native_ordering);
-                    block = ((block >> 8) & STRIPPED_MASK) |
+            } else {
-                            ((block & STRIPPED_MASK) << 8);
+                assert(kind == PyUnicode_4BYTE_KIND);
-                }
+                ch = ucs4lib_utf16_decode(&q, e,
-                maxch = (Py_UCS2)(block & 0xFFFF);
+                        PyUnicode_4BYTE_DATA(unicode), &outpos,
-#if SIZEOF_LONG == 8
+                        native_ordering);
-                ch = (Py_UCS2)((block >> 16) & 0xFFFF);
-                maxch = MAX_MAXCHAR(maxch, ch);
-                ch = (Py_UCS2)((block >> 32) & 0xFFFF);
-                maxch = MAX_MAXCHAR(maxch, ch);
-                ch = (Py_UCS2)(block >> 48);
-                maxch = MAX_MAXCHAR(maxch, ch);
-#else
-                ch = (Py_UCS2)(block >> 16);
-                maxch = MAX_MAXCHAR(maxch, ch);
-#endif
-                if (maxch > PyUnicode_MAX_CHAR_VALUE(unicode)) {
-                    if (unicode_widen(&unicode, outpos, maxch) < 0)
-                        goto onError;
-                    kind = PyUnicode_KIND(unicode);
-                    data = PyUnicode_DATA(unicode);
-                }
-#ifdef BYTEORDER_IS_LITTLE_ENDIAN
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)(block & 0xFFFF));
-#if SIZEOF_LONG == 8
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 16) & 0xFFFF));
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 32) & 0xFFFF));
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 48)));
-#else
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)(block >> 16));
-#endif
-#else
-#if SIZEOF_LONG == 8
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 48)));
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 32) & 0xFFFF));
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)((block >> 16) & 0xFFFF));
-#else
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)(block >> 16));
-#endif
-                PyUnicode_WRITE(kind, data, outpos++, (Py_UCS2)(block & 0xFFFF));
-#endif
-                _q += SIZEOF_LONG;
             }
-            q = _q;
-            if (q >= e)
-                break;
         }
-        ch = (q[ihi] << 8) | q[ilo];
 
-        q += 2;
+        switch (ch)
-
+        {
-        if (!Py_UNICODE_IS_SURROGATE(ch)) {
+        case 0:
+            /* remaining byte at the end? (size should be even) */
+            if (q == e || consumed)
+                goto End;
+            errmsg = "truncated data";
+            startinpos = ((const char *)q) - starts;
+            endinpos = ((const char *)e) - starts;
+            break;
+            /* The remaining input chars are ignored if the callback
+               chooses to skip the input */
+        case 1:
+            errmsg = "unexpected end of data";
+            startinpos = ((const char *)q) - 2 - starts;
+            endinpos = ((const char *)e) - starts;
+            break;
+        case 2:
+            errmsg = "illegal encoding";
+            startinpos = ((const char *)q) - 2 - starts;
+            endinpos = startinpos + 2;
+            break;
+        case 3:
+            errmsg = "illegal UTF-16 surrogate";
+            startinpos = ((const char *)q) - 4 - starts;
+            endinpos = startinpos + 2;
+            break;
+        default:
             if (unicode_putchar(&unicode, &outpos, ch) < 0)
                 goto onError;
             continue;
         }
 
-        /* UTF-16 code pair: */
-        if (q > e) {
-            errmsg = "unexpected end of data";
-            startinpos = (((const char *)q) - 2) - starts;
-            endinpos = ((const char *)e) + 1 - starts;
-            goto utf16Error;
-        }
-        if (Py_UNICODE_IS_HIGH_SURROGATE(ch)) {
-            Py_UCS4 ch2 = (q[ihi] << 8) | q[ilo];
-            q += 2;
-            if (Py_UNICODE_IS_LOW_SURROGATE(ch2)) {
-                if (unicode_putchar(&unicode, &outpos,
-                                    Py_UNICODE_JOIN_SURROGATES(ch, ch2)) < 0)
-                    goto onError;
-                continue;
-            }
-            else {
-                errmsg = "illegal UTF-16 surrogate";
-                startinpos = (((const char *)q)-4)-starts;
-                endinpos = startinpos+2;
-                goto utf16Error;
-            }
-
-        }
-        errmsg = "illegal encoding";
-        startinpos = (((const char *)q)-2)-starts;
-        endinpos = startinpos+2;
-        /* Fall through to report the error */
-
-      utf16Error:
         if (unicode_decode_call_errorhandler(
                 errors,
                 &errorHandler,
@@ -5421,33 +5330,8 @@ PyUnicode_DecodeUTF16Stateful(const char *s,
                 &outpos))
             goto onError;
     }
-    /* remaining byte at the end? (size should be even) */
-    if (e == q) {
-        if (!consumed) {
-            errmsg = "truncated data";
-            startinpos = ((const char *)q) - starts;
-            endinpos = ((const char *)e) + 1 - starts;
-            if (unicode_decode_call_errorhandler(
-                    errors,
-                    &errorHandler,
-                    "utf16", errmsg,
-                    &starts,
-                    (const char **)&e,
-                    &startinpos,
-                    &endinpos,
-                    &exc,
-                    (const char **)&q,
-                    &unicode,
-                    &outpos))
-                goto onError;
-            /* The remaining input chars are ignored if the callback
-               chooses to skip the input */
-        }
-    }
-
-    if (byteorder)
-        *byteorder = bo;
 
+End:
     if (consumed)
         *consumed = (const char *)q-starts;
 
@@ -5466,9 +5350,6 @@ PyUnicode_DecodeUTF16Stateful(const char *s,
     return NULL;
 }
 
-#undef FAST_CHAR_MASK
-#undef SWAPPED_FAST_CHAR_MASK
-
 PyObject *
 _PyUnicode_EncodeUTF16(PyObject *str,
                        const char *errors,