Utilize SSE4.1 to skip comments at 2x speed

4 files changed, 131 insertions, 41 deletions

diff --git a/make.c b/make.c
index f96a039..598db1d 100644
--- a/make.c
+++ b/make.c
@@ -366,33 +366,78 @@ mkgmp(int nprocs)
 bool
 tagvalid(const char *file)
 {
-	if (strstr(file, "-avx2.") != NULL && (simd_flags & SIMD_AVX2) == 0)
-		return false;
-	if (strstr(file, "-neon.") != NULL && (simd_flags & SIMD_NEON) == 0)
-		return false;
-	if (strstr(file, "-sse4_1.") != NULL && (simd_flags & SIMD_SSE4_1) == 0)
-		return false;
-	return true;
+	/* No tag */
+	if (strchr(file, '-') == NULL)
+		return true;
+
+	char *want_and_have = NULL;
+	static struct pair {
+		char *tag;
+		int flag;
+	} tags[] = {
+		{"avx2",    SIMD_AVX2  },
+		{"sse4_1",  SIMD_SSE4_1},
+		{"neon",    SIMD_NEON  },
+		{"generic", 0          },
+	};
+
+	size_t len = strlen(file);
+	char *buf = malloc(len + sizeof("generic"));
+	assert(buf != NULL);
+	for (size_t i = 0; i < lengthof(tags); i++) {
+		char *sep = strrchr(file, '-');
+		char *ext = strrchr(file, '.');
+		assert(sep != NULL);
+		assert(ext != NULL);
+		sprintf(buf, "%.*s-%s%s", (int)(sep - file), file, tags[i].tag, ext);
+
+		if (fexists(buf)
+		    && ((simd_flags & tags[i].flag) != 0 || tags[i].flag == 0))
+		{
+			want_and_have = buf;
+			break;
+		}
+	}
+
+	bool ret;
+	if (want_and_have == NULL)
+		ret = false;
+	else
+		ret = strcmp(want_and_have, file) == 0;
+
+	free(buf);
+	return ret;
 }
 
 void
 chk_cpu_flags(void)
 {
+	uint32_t exx;
+	(void)exx; /* Might be unused */
+
 	if (!rflag)
 		return;
-#if __GNUC__ && __x86_64__
-	uint32_t exx;
 
+	/* Test for AVX512 */
+#if __AVX512F__
+	simd_flags |= SIMD_AVX2;
+#elif __GNUC__ && __x86_64__
 	asm volatile("cpuid" : "=b"(exx) : "a"(7), "c"(0));
-	if (exx & (1 << 5)) {
+	if (exx & (1 << 5))
 		simd_flags |= SIMD_AVX2;
-		return;
-	}
+#endif
 
+	/* Test for SSE4.1 */
+#if __SSE4_1__
+	simd_flags |= SIMD_SSE4_1;
+#elif __GNUC__ && __x86_64__
 	asm volatile("cpuid" : "=c"(exx) : "a"(1), "c"(0));
 	if (exx & (1 << 19))
 		simd_flags |= SIMD_SSE4_1;
-#elif __ARM_NEON
+#endif
+
+	/* Test for NEON */
+#if __ARM_NEON
 	simd_flags |= SIMD_NEON;
 #endif
 }
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)
 {