kit

commit 325a68b5234a664320fcfb1d28031b87bc98e7de
parent de8c3b62f7bba21c192e5a973d5f73ae94c2a0f8
Author: Ryan Sepassi <rsepassi@gmail.com>
Date:   Sat, 13 Jun 2026 09:24:15 -0700

perf(lex): memchr-accelerate the splice pre-scan — skip non-backslash spans at NEON speed, no scalar whole-buffer pass

Diffstat:
M
lang/cpp/lex/lex.c
 | 
17
+++++++++++++++--

1 file changed, 15 insertions(+), 2 deletions(-)
diff --git a/lang/cpp/lex/lex.c b/lang/cpp/lex/lex.c
@@ -153,8 +153,21 @@ static void lex_fold_splices(Lexer* l, const char* src, size_t len) {
   u32* spl;
   size_t w, s, wcap;
   l->next_splice = 0;
-  for (i = 0; i + 1 < len; ++i)
-    if (src[i] == '\\' && src[i + 1] == '\n') ++nspl;
+  /* Count `\<newline>` splices without a scalar whole-buffer pass: backslash is
+   * rare in C, so a memchr-driven sweep skips non-backslash spans at glibc-NEON
+   * speed (~1 instruction per 16-32 bytes) instead of testing every byte. The
+   * `-1` keeps bs[1] in range, so a trailing lone `\\` at src[len-1] is never
+   * counted as a splice — bit-exact with the old scalar `i + 1 < len` loop. */
+  if (len) {
+    const char* p = src;
+    const char* e = src + len;
+    while (p + 1 < e) {
+      const char* bs = (const char*)memchr(p, '\\', (size_t)(e - p) - 1);
+      if (!bs) break;
+      if (bs[1] == '\n') ++nspl;
+      p = bs + 1;
+    }
+  }
   if (nspl == 0) {
     l->src = src;
     l->len = len;