Add minor improvements and cleanup

libc/str/stpcpy.c

@@ -17,20 +17,45 @@
 │ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA                │
 │ 02110-1301 USA                                                               │
 ╚─────────────────────────────────────────────────────────────────────────────*/
+#include "libc/intrin/pcmpeqb.h"
+#include "libc/intrin/pmovmskb.h"
 #include "libc/str/str.h"
 
 /**
- * Copies string and advances destination pointer.
+ * Copies bytes from 𝑠 to 𝑑 until a NUL is encountered.
+ *
+ * @param 𝑑 is destination memory
+ * @param 𝑠 is a NUL-terminated string
+ * @note 𝑑 and 𝑠 can't overlap
+ * @return pointer to nul byte
+ * @see strcpy(), memccpy()
  * @asyncsignalsafe
  */
-char *stpcpy(char *dst, const char *src) {
-  char c;
-  for (;;) {
-    c = *src;
-    *dst = c;
-    if (!c) break;
-    ++src;
-    ++dst;
+char *stpcpy(char *d, const char *s) {
+  size_t i;
+  uint8_t v1[16], v2[16], vz[16];
+  i = 0;
+  while (((uintptr_t)(s + i) & 15)) {
+    if (!(d[i] = s[i])) {
+      return d + i;
+    }
+    ++i;
+  }
+  for (;;) {
+    memset(vz, 0, 16);
+    memcpy(v1, s + i, 16);
+    pcmpeqb(v2, v1, vz);
+    if (!pmovmskb(v2)) {
+      memcpy(d + i, v1, 16);
+      i += 16;
+    } else {
+      break;
+    }
+  }
+  for (;;) {
+    if (!(d[i] = s[i])) {
+      return d + i;
+    }
+    ++i;
   }
-  return dst;
 }