From 55d3a6ea39c3275958d2ac9a666b7bf8ffeb6ed7 Mon Sep 17 00:00:00 2001
From: Colin Leroy-Mira <colin@colino.net>
Date: Wed, 4 Sep 2024 22:52:52 +0200
Subject: [PATCH] Optimize stpcpy's size and speed

---
 libsrc/common/stpcpy.c |  7 -------
 libsrc/common/stpcpy.s | 22 ++++++++++++++++++++++
 libsrc/common/strcpy.s |  7 +++++--
 test/val/stpcpy.c      | 13 +++++++++++--
 4 files changed, 38 insertions(+), 11 deletions(-)
 delete mode 100644 libsrc/common/stpcpy.c
 create mode 100644 libsrc/common/stpcpy.s

diff --git a/libsrc/common/stpcpy.c b/libsrc/common/stpcpy.c
deleted file mode 100644
index 12af47f2f..000000000
--- a/libsrc/common/stpcpy.c
+++ /dev/null
@@ -1,7 +0,0 @@
-#include <string.h>
-
-char * __fastcall__ stpcpy (char * dst, const char * src)
-{
-    strcpy (dst, src);
-    return dst + strlen (src);
-}
diff --git a/libsrc/common/stpcpy.s b/libsrc/common/stpcpy.s
new file mode 100644
index 000000000..c8a10db94
--- /dev/null
+++ b/libsrc/common/stpcpy.s
@@ -0,0 +1,22 @@
+;
+; Colin Leroy-Mira, 4 Sept. 2024
+;
+; char* stpcpy (char* dest, const char* src);
+;
+
+        .export         _stpcpy
+        .import         _strcpy
+
+        .importzp       tmp1, ptr2
+
+_stpcpy:
+        jsr     _strcpy
+
+        ldx     ptr2+1        ; Load dest pointer's last high byte
+        tya                   ; Get the last offset strcpy wrote to
+
+        clc
+        adc     ptr2          ; Add to low byte value
+        bcc     :+
+        inx
+:       rts                   ; Return pointer to dest's terminator
diff --git a/libsrc/common/strcpy.s b/libsrc/common/strcpy.s
index 77b39fe76..9a100f540 100644
--- a/libsrc/common/strcpy.s
+++ b/libsrc/common/strcpy.s
@@ -25,6 +25,9 @@ L1:     lda     (ptr1),y
         inc     ptr2+1
         bne     L1
 
-L9:     lda     ptr2            ; X still contains high byte
-        rts
+L9:     lda     ptr2            ; X still contains dest's original high byte
 
+        ; On exit, we want AX to be dest (as this is what strcpy returns).
+        ; We also want (ptr2),y to still point to dest's terminator, as this
+        ; is used by stpcpy().
+        rts
diff --git a/test/val/stpcpy.c b/test/val/stpcpy.c
index 8bdbfb926..1cc6458ed 100644
--- a/test/val/stpcpy.c
+++ b/test/val/stpcpy.c
@@ -8,10 +8,12 @@
 #define STR_SHORT   "Hello, World!"
 #define STR_LONG    "This is a longer test string for stpcpy."
 
+char dest[512];
+char multi_page[300];
+
 int
 main ()
 {
-    char dest[50];
     const char *src_empty;
     const char *src_short;
     const char *src_long;
@@ -38,7 +40,14 @@ main ()
     assert(end == &dest[sizeof (STR_LONG) - 1]);
     printf ("Test 3 passed.\n");
 
+    memset(multi_page, 'a', sizeof(multi_page)-1);
+    multi_page[sizeof(multi_page)-1] = '\0';
+    end = stpcpy (dest, multi_page);
+    assert(!strcmp (dest, multi_page));
+    assert(!*end);
+    assert(end == &dest[sizeof (multi_page) - 1]);
+    printf ("Test 4 passed.\n");
+
     printf ("All tests passed.\n");
     return EXIT_SUCCESS;
 }
-