Utilize SSE4.1 to skip comments at 2x speed

3 files changed, 73 insertions, 28 deletions

diff --git a/src/lexer-generic.c b/src/lexer-generic.c
new file mode 100644
index 0000000..b841886
--- /dev/null
+++ b/src/lexer-generic.c
@@ -0,0 +1,28 @@
+#include <stdbool.h>
+
+#include "common.h"
+#include "types.h"
+
+bool
+skpcmnt(const uchar **ptr, const uchar *end)
+{
+	int nst = 1;
+	const uchar *p = *ptr;
+
+	for (p++; likely(p < end); p++) {
+		if (p + 1 < end) {
+			if (p[0] == '*' && p[1] == '/') {
+				p++;
+				if (--nst == 0) {
+					*ptr = ++p;
+					return true;
+				}
+			} else if (p[0] == '/' && p[1] == '*') {
+				p++;
+				nst++;
+			}
+		}
+	}
+
+	return false;
+}
diff --git a/src/lexer-sse4_1.c b/src/lexer-sse4_1.c
new file mode 100644
index 0000000..16df370
--- /dev/null
+++ b/src/lexer-sse4_1.c
@@ -0,0 +1,43 @@
+#include <stdbool.h>
+#include <stddef.h>
+#include <x86intrin.h>
+
+#include "common.h"
+#include "types.h"
+
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+
+bool
+skpcmnt(const uchar **ptr, const uchar *end)
+{
+	int nst = 1;
+	const uchar *p = *ptr, needles[] = {'/', '*', 0, 0, 0, 0, 0, 0,
+	                                    0,   0,   0, 0, 0, 0, 0, 0};
+	const __m128i set = _mm_loadu_si128((const __m128i *)needles);
+
+	while (likely(p < end)) {
+		ptrdiff_t len = end - p;
+		size_t blksz = MIN(len, 16);
+		__m128i blk = _mm_loadu_si128((const __m128i *)p);
+		int off = _mm_cmpestri(set, 2, blk, blksz, _SIDD_CMP_EQUAL_ANY);
+
+		if (off == 16) {
+			p += 16;
+			continue;
+		}
+
+		if (p[off] == '*' && p[off + 1] == '/') {
+			p += off + 2;
+			if (--nst == 0) {
+				*ptr = p;
+				return true;
+			}
+		} else if (p[off] == '/' && p[off + 1] == '*') {
+			p += off + 2;
+			nst++;
+		} else
+			p += off + 1;
+	}
+
+	return false;
+}
diff --git a/src/lexer.c b/src/lexer.c
index ed2414a..0ed057d 100644
--- a/src/lexer.c
+++ b/src/lexer.c
@@ -32,7 +32,7 @@ static void lexemesresz(lexemes_t *toks)
 /* Advance PTR (which points to the start of a comment) to the end of the
    comment, or END.  Returns true if the comment was well-formed and
    false if the comment was unterminated.  Handles nested comments. */
-static bool skip_comment(const uchar **ptr, const uchar *end)
+bool skpcmnt(const uchar **ptr, const uchar *end)
 	__attribute__((nonnull));
 
 static const bool is_numeric_lookup[UCHAR_MAX + 1] = {
@@ -88,7 +88,7 @@ lexstring(const uchar *code, size_t codesz)
 		/* Single- or double-byte literals */
 		case '/':
 			if (code < end && code[0] == '*') {
-				if (!skip_comment(&code, end))
+				if (!skpcmnt(&code, end))
 					err("Unterminated comment at byte %td", code - start);
 				continue;
 			}
@@ -172,32 +172,6 @@ fallback:
 	return data;
 }
 
-bool
-skip_comment(const uchar **ptr, const uchar *end)
-{
-	int nst = 1;
-	const uchar *p = *ptr;
-
-	for (p++; likely(p < end); p++) {
-		if (p + 1 < end) {
-			if (p[0] == '*' && p[1] == '/') {
-				p++;
-				if (--nst == 0)
-					goto out;
-			} else if (p[0] == '/' && p[1] == '*') {
-				p++;
-				nst++;
-			}
-		}
-	}
-
-	return false;
-
-out:
-	*ptr = ++p;
-	return true;
-}
-
 lexemes_t
 mklexemes(void)
 {