optimized memmove

Author: ZERICO2005

src/libc/memmove.src

@@ -0,0 +1,102 @@
+	assume	adl=1
+
+	section	.text
+
+	public	_memmove
+
+if PREFER_OS_LIBC
+
+_memmove := 0000A8h
+
+else
+
+if 1
+
+; Optimized for when src != dst
+_memmove:
+	; src > dst | LDIR | 32F + 15R + 1
+	; src < dst | LDDR | 35F + 12R + 2
+	; src = dst | LDDR | 35F + 12R + 2
+	; zero size |      | 24F + 12R + 2
+
+	ld	iy, -1
+	add	iy, sp
+	ld	bc, (iy + 10)
+	sbc	hl, hl
+	add	hl, bc
+	jr	nc, .zero
+	ld	hl, (iy + 7)
+	ld	de, (iy + 4)
+	sbc	hl, de
+	; src <= dst
+	jr	c, .copy_backwards
+; .copy_forwards:
+	add	hl, de
+	inc	hl
+	; src > dst
+	ldir
+.zero:
+	ld	hl, (iy + 4)
+	ret
+.copy_backwards:
+	; move HL and DE to the end
+	ex	de, hl
+	add	hl, bc
+	ex	de, hl	; HL = src - dst - 1, DE = dst + size
+	add	hl, de	; HL = src + size - 1
+	dec	de	; DE = dst + size - 1
+	lddr
+	ex	de, hl
+	inc	hl
+	ret
+
+else
+
+; Optimized for when src = dst
+_memmove:
+	; src > dst | LDIR | 34F + 15R + 2
+	; src < dst | LDDR | 36F + 12R + 2
+	; src = dst |      | 30F + 12R + 2
+	; zero size |      | 24F + 12R + 2
+
+	ld	iy, -1
+	add	iy, sp
+	ld	bc, (iy + 10)
+	sbc	hl, hl
+	add	hl, bc
+	jr	nc, .zero
+	ld	de, (iy + 4)
+	ld	hl, (iy + 7)
+	dec	de	; DE = dst - 1
+	sbc	hl, de
+	; src >= dst
+	jr	c, .copy_backwards
+; .copy_forwards:
+	inc	de
+	add	hl, de
+	; src == dst
+	ret	z	; <-- Optional, skips LDIR when src == dst
+	; src > dst or no-overlap
+	ldir
+.zero:
+	ld	hl, (iy + 4)
+	ret
+
+.copy_backwards:
+	; src < dst
+	; move HL and DE to the end
+	add	hl, de	; HL = src - 1
+	add	hl, bc
+	ex	de, hl
+	add	hl, bc
+	ex	de, hl
+	; HL = src + size - 1
+	; DE = dst + size - 1
+	lddr
+	ex	de, hl
+	inc	hl
+	ret
+
+end if
+
+end if

src/libc/os.src

@@ -8,8 +8,6 @@ _longjmp    := 000098h
 _memchr     := 00009Ch
 	public	_memcmp
 _memcmp     := 0000A0h
-	public	_memmove
-_memmove    := 0000A8h
 	public	_setjmp
 _setjmp     := 0000B8h
 	public	_strcat

test/standalone/asprintf_fprintf/src/main.c

@@ -33,6 +33,9 @@ extern void* NULL_ptr;
 void *T_memcpy(void *__restrict dest, const void *__restrict src, size_t n)
     __attribute__((nonnull(1, 2)));
 
+void *T_memmove(void *dest, const void *src, size_t n)
+    __attribute__((nonnull(1, 2)));
+
 void *T_memset(void *s, int c, size_t n)
     __attribute__((nonnull(1)));
 
@@ -65,6 +68,7 @@ void T_bzero(void* s, size_t n);
 #else
 
 #define T_memcpy memcpy
+#define T_memmove memmove
 #define T_memset memset
 #define T_memcmp memcmp
 #define T_memccpy memccpy
@@ -575,6 +579,24 @@ int memrchr_test(void) {
     return 0;
 }
 
+int memmove_test(void) {
+    char move_str[]        = "0123456789";
+    const char truth_str[] = "9344545689";
+    C(move_str + 5 == (char*)T_memmove(move_str + 5, move_str + 4, 3));
+    C(move_str + 3 == (char*)T_memmove(move_str + 3, move_str + 3, 0));
+    C(move_str + 1 == (char*)T_memmove(move_str + 1, move_str + 3, 4));
+    C(move_str + 2 == (char*)T_memmove(move_str + 2, move_str + 2, 6));
+    C(move_str + 9 == (char*)T_memmove(move_str + 9, move_str + 0, 0));
+    C(move_str + 0 == (char*)T_memmove(move_str + 0, move_str + 9, 1));
+    C(strcmp(move_str, truth_str) == 0);
+
+    C(NULL_ptr - 0 == (char*)T_memmove(NULL_ptr - 0, NULL_ptr - 0, 0));
+    C(NULL_ptr - 0 == (char*)T_memmove(NULL_ptr - 0, NULL_ptr - 1, 0));
+    C(NULL_ptr - 1 == (char*)T_memmove(NULL_ptr - 1, NULL_ptr - 0, 0));
+
+    return 0;
+}
+
 int run_tests(void) {
     int ret = 0;
     /* boot_asprintf */
@@ -608,10 +630,14 @@ int run_tests(void) {
         ret = strncmp_test();
         if (ret != 0) { return ret; }
 
-    /* strncmp */
+    /* memrchr */
         ret = memrchr_test();
         if (ret != 0) { return ret; }
 
+    /* memrchr */
+        ret = memmove_test();
+        if (ret != 0) { return ret; }
+
     return 0;
 }
 

test/standalone/asprintf_fprintf/src/rename.asm

@@ -2,12 +2,13 @@
 
 	section	.text
 
-	public	_T_memset, _T_memcpy, _T_memcmp, _T_memccpy, _T_mempcpy, _T_memrchr
+	public	_T_memset, _T_memcpy, _T_memmove, _T_memcmp, _T_memccpy, _T_mempcpy, _T_memrchr
 	public	_T_strlen, _T_strcmp, _T_strncmp, _T_stpcpy
 	public	_T_bzero
 
 _T_memset := _memset
 _T_memcpy := _memcpy
+_T_memmove := _memmove
 _T_memcmp := _memcmp
 _T_memccpy := _memccpy
 _T_mempcpy := _mempcpy
@@ -26,6 +27,6 @@ _T_bzero := _bzero
 _NULL_ptr:
 	db	$00, $00, $00
 
-	extern	_memset, _memcpy, _memcmp, _memccpy, _mempcpy, _memrchr
+	extern	_memset, _memcpy, _memmove, _memcmp, _memccpy, _mempcpy, _memrchr
 	extern	_strlen, _strcmp, _strncmp, _stpcpy
 	extern	_bzero