SIMD fast path for skipping preprocessing chunks

main.c

@@ -8,6 +8,8 @@
 #include <time.h>
 #include <fcntl.h>
 
+#include <immintrin.h>
+
 enum token_kind {
     
     TOKEN_KEYWORD,
@@ -58,7 +60,7 @@ nondigit(struct str s)
 {
     if (s.size) {
         char c = s.text[0];
-        return((c == '_') || ('A' <= c && c <= 'Z') || ('a' <= c && c <= 'z'));
+        return(('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || (c == '_'));
     }
     return(0);
 }
@@ -1035,9 +1037,11 @@ comment(struct str s)
                 advance(&s, 1);
             }
             
+            if (s.size) {
                 return(start - s.size);
             }
         }
+    }
     
     return(0);
 }
@@ -1087,14 +1091,15 @@ lex(char *text, int size)
 }
 
 static void
-preprocess(char *data, int size)
+preprocess_scalar(char *data, int chunk_size, int full_size)
 {
-    for (int i = 0; i < size; ++i) {
+    for (int i = 0; i < chunk_size; ++i) {
         if (data[i] == '\\') {
-            for (int j = i + 1; j < size; ++j) {
+            for (int j = i + 1; j < full_size; ++j) {
                 if (data[j] == '\n') {
                     data[j] = ' ';
                     data[i] = ' ';
+                    break;
                 } else if (data[j] != ' ' && data[j] != '\t') {
                     break;
                 }
@@ -1103,6 +1108,26 @@ preprocess(char *data, int size)
     }
 }
 
+static void
+preprocess(char *data, int size)
+{
+    int chunk_size = 16;
+    int whole = size & (~(chunk_size - 1));
+    
+    __m128i mask = _mm_set1_epi8('\\');
+    
+    for (int i = 0; i < whole; i += chunk_size) {
+        __m128i chunk = _mm_loadu_si128((__m128i *) (data + i));
+        __m128i match = _mm_cmpeq_epi8(chunk, mask);
+        
+        if (!_mm_testz_si128(match, match)) {
+            preprocess_scalar(data + i, chunk_size, size - i);
+        }
+    }
+    
+    preprocess_scalar(data + whole, size - whole, size - whole);
+}
+
 static void
 run(char *data, int size)
 {