sse_addslashes.diff

diff --git a/ext/standard/string.c b/ext/standard/string.c
index 97a20fb..d6b039f 100644
--- a/ext/standard/string.c
+++ b/ext/standard/string.c
@@ -34,6 +34,12 @@
 #ifdef HAVE_MONETARY_H
 # include <monetary.h>
 #endif
+
+#ifdef __SSE4_2__
+#include <nmmintrin.h>
+#include "Zend/zend_bitset.h"
+#endif
+
 /*
  * This define is here because some versions of libintl redefine setlocale
  * to point to libintl_setlocale.  That's a ridiculous thing to do as far
@@ -3872,6 +3878,11 @@ PHPAPI zend_string *php_addslashes(zend_string *str, int should_free)
 	char *end;
 	size_t offset;
 	zend_string *new_str;
+#ifdef __SSE4_2__
+	uint32_t r = 0;
+	static const char ctrs[16] = "\'\"\\\0";
+	__m128i w, s;
+#endif

 	if (!str) {
 		return ZSTR_EMPTY_ALLOC();
@@ -3880,6 +3891,37 @@ PHPAPI zend_string *php_addslashes(zend_string *str, int should_free)
 	source = ZSTR_VAL(str);
 	end = source + ZSTR_LEN(str);

+#ifdef __SSE4_2__
+	if (ZSTR_LEN(str) > 15) {
+		char *align = (char*)(((uintptr_t)source + 15) & ~15);
+		w = _mm_load_si128((__m128i *)&ctrs[0]);
+
+		if (UNEXPECTED(source != align)) {
+			do {
+				switch (*source) {
+					case '\0':
+					case '\'':
+					case '\"':
+					case '\\':
+						goto do_escape;
+					default:
+						source++;
+						break;
+				}
+			} while (source < align);
+		}
+
+		for (;end - source > 15; source += 16) {
+			s = _mm_load_si128((__m128i *)source);
+			int r = _mm_cvtsi128_si32(
+					_mm_cmpestrm(w, 4,  s, 16, _SIDD_UBYTE_OPS|_SIDD_CMP_EQUAL_ANY|_SIDD_BIT_MASK));
+			if (r) {
+				goto do_escape;
+			}
+		}
+	}
+#endif
+
 	while (source < end) {
 		switch (*source) {
 			case '\0':
@@ -3905,6 +3947,79 @@ do_escape:
 	memcpy(ZSTR_VAL(new_str), ZSTR_VAL(str), offset);
 	target = ZSTR_VAL(new_str) + offset;

+#ifdef __SSE4_2__
+	if (r) {
+		int pos = 0;
+		do {
+			int i, n = zend_ulong_ntz(r);
+			for (i = 0; i < n; i++) {
+				*target++ = source[pos + i];
+			}
+			pos += n;
+			*target++ = '\\';
+			if (source[pos] == '\0') {
+				*target++ = '0';
+			} else {
+				*target++ = source[pos];
+			}
+			pos++;
+			r = r >> (n + 1);
+		} while (r);
+
+		for (; pos < 16; pos++) {
+			*target++ = source[pos];
+		}
+	} else if (end - source > 15) {
+		char *align = (char*)(((zend_uintptr_t)source + 15) & ~15);
+		w = _mm_load_si128((__m128i *)&ctrs[0]);
+
+		if (source != align) {
+			do {
+				switch (*source) {
+					case '\0':
+						*target++ = '\\';
+						*target++ = '0';
+						break;
+					case '\'':
+					case '\"':
+					case '\\':
+						*target++ = '\\';
+						/* break is missing *intentionally* */
+					default:
+						*target++ = *source;
+						break;
+				}
+				source++;
+			} while (source < align);
+		}
+	}
+
+	for (; end - source > 15; source += 16) {
+		s = _mm_load_si128((__m128i *)source);
+		int r = _mm_cvtsi128_si32(
+				_mm_cmpestrm(w, 4, s, 16, _SIDD_UBYTE_OPS|_SIDD_CMP_EQUAL_ANY|_SIDD_BIT_MASK));
+		int pos = 0;
+		while (r) {
+			int i, n = zend_ulong_ntz(r);
+			r = r >> (n + 1);
+			for (i = 0; i < n; i++) {
+				*target++ = source[pos + i];
+			}
+			pos += n;
+			*target++ = '\\';
+			if (source[pos] == '\0') {
+				*target++ = '0';
+			} else {
+				*target++ = source[pos];
+			}
+			pos++;
+		}
+		for (; pos < 16; pos++) {
+			*target++ = source[pos];
+		}
+	}
+
+#endif
 	while (source < end) {
 		switch (*source) {
 			case '\0':
@@ -3920,11 +4035,10 @@ do_escape:
 				*target++ = *source;
 				break;
 		}
-
 		source++;
 	}

-	*target = 0;
+	*target = '\0';
 	if (should_free) {
 		zend_string_release(str);
 	}