Add accelerated (SIMD-based) implementation of mb_check_encoding for UTF-8

The new SSE2-based implementation of mb_check_encoding for UTF-8 is
about 10% faster for 0-5 byte strings, more than 3 times faster for
~100-byte strings, and just under 4 times faster for ~10,000-byte
strings.

I believe it may be possible to make this function much faster again.
Some possible directions for further performance optimization include:

• If other ISA extensions like AVX or AVX-512 are available, use a
  similar algorithm, but process text in blocks of 32 or 64 bytes
  (instead of 16 bytes).
• If other SIMD ISA extensions are available, use the greater variety
  of available instructions to make some of the checks tighter.
• Even if only SSE/SSE2 are available, find clever ways to squeeze
  instructions out of the hot path. This would probably require a lot
  of perusing instruction mauals and thinking hard about which SIMD
  instructions could be used to perform the same checks with fewer
  instructions.
• Find a better algorithm, possibly one where more checks could be
  combined (just as the current algorithm combines the checks for
  certain overlong code units and reserved codepoints).

Author: alexdowad

ext/mbstring/mbstring.c

@@ -4587,13 +4587,184 @@ MBSTRING_API bool php_mb_check_encoding(const char *input, size_t length, const
 	return true;
 }
 
+static bool mb_fast_check_utf8(zend_string *str)
+{
+	unsigned char *p = (unsigned char*)ZSTR_VAL(str), *e = p + ZSTR_LEN(str);
+
+#ifdef __SSE2__
+	e -= sizeof(__m128i);
+
+	/* For checking for illegal bytes 0xF5-FF */
+	const __m128i over_f5 = _mm_set1_epi8(-117);
+	/* For checking for overlong 3-byte code units and reserved codepoints U+D800-DFFF */
+	const __m128i over_9f = _mm_set1_epi8(-97);
+	/* For checking for overlong 4-byte code units and invalid codepoints > U+10FFFF */
+	const __m128i over_8f = _mm_set1_epi8(-113);
+	/* For checking for illegal bytes 0xC0-C1 */
+	const __m128i find_c0 = _mm_set1_epi8(-64);
+	const __m128i c0_to_c1 = _mm_set1_epi8(-126);
+	/* For checking structure of continuation bytes */
+	const __m128i find_e0 = _mm_set1_epi8(-32);
+	const __m128i find_f0 = _mm_set1_epi8(-16);
+
+	__m128i last_block = _mm_setzero_si128();
+	__m128i operand;
+
+	while (p <= e) {
+		operand = _mm_loadu_si128((__m128i*)p); /* Load 16 bytes */
+
+check_operand: ;
+		/* If all 16 bytes are single-byte characters, there is no need to check anything */
+		__m128i multibyte = _mm_cmplt_epi8(operand, _mm_setzero_si128());
+		if (_mm_movemask_epi8(multibyte)) {
+			/* Check for >= 0xF5, which are illegal byte values in UTF-8
+			 * AVX512 has instructions for vectorized unsigned compare, but SSE2 only has signed compare
+			 * So we add an offset to shift 0xF5-FF to the far low end of the signed byte range
+			 * Then a single signed compare will pick out any bad bytes
+			 * `bad` is a vector of 16 good/bad values, where 0x00 means good and 0xFF means bad */
+			__m128i bad = _mm_cmplt_epi8(_mm_add_epi8(operand, over_f5), over_f5);
+
+			/* Check for overlong 3-byte code units AND reserved codepoints U+D800-DFFF
+			 * 0xE0 followed by a byte < 0xA0 indicates an overlong 3-byte code unit, and
+			 * 0xED followed by a byte >= 0xA0 indicates a reserved codepoint
+			 * We can check for both problems at once by generating a vector where each byte < 0xA0
+			 * is mapped to 0xE0, and each byte >= 0xA0 is mapped to 0xED
+			 * Shift the original block right by one byte, and XOR the shifted block with the bitmask
+			 * Any matches will give a 0x00 byte; do a compare with a zero vector to pick out the
+			 * bad positions, and OR them into `bad` */
+			__m128i operand2 = _mm_or_si128(_mm_slli_si128(operand, 1), _mm_srli_si128(last_block, 15));
+			__m128i mask1 = _mm_or_si128(find_e0, _mm_and_si128(_mm_set1_epi8(0xD), _mm_cmpgt_epi8(operand, over_9f)));
+			bad = _mm_or_si128(bad, _mm_cmpeq_epi8(_mm_setzero_si128(), _mm_xor_si128(operand2, mask1)));
+
+			/* Check for overlong 4-byte code units AND invalid codepoints > U+10FFFF
+			 * Similar to the previous check; 0xF0 followed by < 0x90 indicates an overlong 4-byte
+			 * code unit, and 0xF4 followed by >= 0x90 indicates a codepoint over U+10FFFF
+			 * Build the bitmask, XOR it with the shifted block, check for 0x00 bytes in the result */
+			__m128i mask2 = _mm_or_si128(find_f0, _mm_and_si128(_mm_set1_epi8(0x4), _mm_cmpgt_epi8(operand, over_8f)));
+			bad = _mm_or_si128(bad, _mm_cmpeq_epi8(_mm_setzero_si128(), _mm_xor_si128(operand2, mask2)));
+
+			/* Check for overlong 2-byte code units
+			 * Any 0xC0 or 0xC1 byte can only be the first byte of an overlong 2-byte code unit
+			 * Same deal as before; add an offset to shift 0xC0-C1 to the far low end of the signed
+			 * byte range, do a signed compare to pick out any bad bytes */
+			bad = _mm_or_si128(bad, _mm_cmplt_epi8(_mm_add_epi8(operand, find_c0), c0_to_c1));
+
+			/* Check structure of continuation bytes (the grand finale!)
+			 * A UTF-8 byte should be a continuation byte if, and only if, it is:
+			 * 1) 1 byte after the start of a 2-byte, 3-byte, or 4-byte character
+			 * 2) 2 bytes after the start of a 3-byte or 4-byte character
+			 * 3) 3 bytes after the start of a 4-byte character
+			 * We build 3 bitmasks with 0xFF in each such position, and OR them together to
+			 * get a single bitmask with 0xFF in each position where a continuation byte should be */
+			__m128i cont_mask = _mm_cmpeq_epi8(_mm_and_si128(operand2, find_c0), find_c0);
+			__m128i operand3 = _mm_or_si128(_mm_slli_si128(operand, 2), _mm_srli_si128(last_block, 14));
+			cont_mask = _mm_or_si128(cont_mask, _mm_cmpeq_epi8(_mm_and_si128(operand3, find_e0), find_e0));
+			__m128i operand4 = _mm_or_si128(_mm_slli_si128(operand, 3), _mm_srli_si128(last_block, 13));
+			cont_mask = _mm_or_si128(cont_mask, _mm_cmpeq_epi8(_mm_and_si128(operand4, find_f0), find_f0));
+
+			/* Now, use a signed comparison to get another bitmask with 0xFF in each position where
+			 * a continuation byte actually is
+			 * XOR those two bitmasks together; if everything is good, the result should be zero
+			 * However, if a byte which should have been a continuation wasn't, or if a byte which
+			 * shouldn't have been a continuation was, we will get 0xFF in that position */
+			__m128i continuation = _mm_cmplt_epi8(operand, find_c0);
+			bad = _mm_or_si128(bad, _mm_xor_si128(continuation, cont_mask));
+
+			/* Pick out the high bit of each byte in `bad` as a 16-bit value (into a scalar register)
+			 * If that value is non-zero, then we found a bad byte somewhere! */
+			if (_mm_movemask_epi8(bad)) {
+				return false;
+			}
+		}
+
+		last_block = operand;
+		p += sizeof(__m128i);
+	}
+
+	/* Finish up 1-15 remaining bytes */
+	uint8_t shift_dist = p - e;
+	if (shift_dist < 16) {
+		/* Crazy hack here... we want to use the above vectorized code to check a block of less than 16
+		 * bytes, but there is no good way to read a variable number of bytes into an XMM register
+		 * However, we know that these bytes are part of a zend_string, and a zend_string has some
+		 * 'header' fields which occupy the memory just before its content
+		 * And, those header fields occupy more than 16 bytes...
+		 * So if we go back 16 bytes from the end of the zend_string content, and load 16 bytes from there,
+		 * we may pick up some 'junk' bytes from the zend_string header fields, but we will get the 1-15
+		 * bytes we wanted in the tail end of our XMM register, and this will never cause a segfault.
+		 * Then, we do a left shift to get rid of the unwanted bytes
+		 * Conveniently, the same left shift also zero-fills the tail end of the XMM register */
+		operand = _mm_loadu_si128((__m128i*)e);
+
+		/* This looks useless, but it's not
+		 * The PSRLDQ instruction used for this 128-bit left shift requires an immediate (literal)
+		 * shift distance, so the compiler will choke on _mm_srli_si128(operand, shift_dist) */
+		switch (shift_dist) {
+			case 0:
+				goto check_operand;
+			case 1:
+				operand = _mm_srli_si128(operand, 1);
+				goto check_operand;
+			case 2:
+				operand = _mm_srli_si128(operand, 2);
+				goto check_operand;
+			case 3:
+				operand = _mm_srli_si128(operand, 3);
+				goto check_operand;
+			case 4:
+				operand = _mm_srli_si128(operand, 4);
+				goto check_operand;
+			case 5:
+				operand = _mm_srli_si128(operand, 5);
+				goto check_operand;
+			case 6:
+				operand = _mm_srli_si128(operand, 6);
+				goto check_operand;
+			case 7:
+				operand = _mm_srli_si128(operand, 7);
+				goto check_operand;
+			case 8:
+				operand = _mm_srli_si128(operand, 8);
+				goto check_operand;
+			case 9:
+				operand = _mm_srli_si128(operand, 9);
+				goto check_operand;
+			case 10:
+				operand = _mm_srli_si128(operand, 10);
+				goto check_operand;
+			case 11:
+				operand = _mm_srli_si128(operand, 11);
+				goto check_operand;
+			case 12:
+				operand = _mm_srli_si128(operand, 12);
+				goto check_operand;
+			case 13:
+				operand = _mm_srli_si128(operand, 13);
+				goto check_operand;
+			case 14:
+				operand = _mm_srli_si128(operand, 14);
+				goto check_operand;
+			case 15:
+				operand = _mm_srli_si128(operand, 15);
+				goto check_operand;
+		}
+
+		ZEND_UNREACHABLE();
+	}
+
+	return true;
+#else
+	return php_mb_check_encoding(ZSTR_VAL(str), ZSTR_LEN(str), &mbfl_encoding_utf8);
+#endif
+}
+
 static bool mb_check_str_encoding(zend_string *str, const mbfl_encoding *encoding)
 {
 	if (encoding == &mbfl_encoding_utf8) {
 		if (GC_FLAGS(str) & IS_STR_VALID_UTF8) {
 			return true;
 		}
-		bool result = php_mb_check_encoding(ZSTR_VAL(str), ZSTR_LEN(str), encoding);
+		bool result = mb_fast_check_utf8(str);
 		if (result && !ZSTR_IS_INTERNED(str)) {
 			GC_ADD_FLAGS(str, IS_STR_VALID_UTF8);
 		}