From 4dc9fe3a550f66ccfca225f59c62d8abe9fc685a Mon Sep 17 00:00:00 2001
From: mykola2312 <49044616+mykola2312@users.noreply.github.com>
Date: Sun, 15 May 2022 21:19:38 +0300
Subject: [PATCH] try assembly memmove, doesn't work because granularity of
 move

---
 arch/x86_64.S | 50 ++++++++++++++++++++++++++++++++++++++++++++++++--
 cutil.c       |  2 +-
 cutil.h       |  4 ++--
 test.c        |  8 ++++----
 4 files changed, 55 insertions(+), 9 deletions(-)

diff --git a/arch/x86_64.S b/arch/x86_64.S
index 8f3a74f..55f7b79 100644
--- a/arch/x86_64.S
+++ b/arch/x86_64.S
@@ -2,6 +2,7 @@
 .globl hw_bswap32
 .globl hw_bswap64
 .globl _cu_memcpy
+.globl _cu_memmove
 .globl _cu_memset
 .globl cu_memzero
 .globl cu_memtest
@@ -28,21 +29,66 @@ _cu_memcpy:
     // RDI - dst, RSI - src, RDX - size
     mov %rdx, %rcx
     shr $3, %rcx
+    je .cpy1
     rep movsq
-
+.cpy1:
     mov %rdx, %rcx
     and $7, %rcx
     rep movsb
     ret
 
+_cu_memmove:
+    // RDI - dst, RSI - src, RDX - size
+    cmp %rsi, %rdi
+    je .retmove
+    jl .lessmove
+.greatermove:   // dst > src
+    add %rdx, %rdi
+    add %rdx, %rsi
+    std
+    mov %rdx, %rcx
+    shr $3, %rcx
+    je .greater1_
+    sub $8, %rdi
+    sub $8, %rsi
+    rep movsq
+.greater1:
+    add $8, %rdi
+    add $8, %rsi
+    mov %rdx, %rcx
+    and $7, %rcx
+    rep movsb
+    cld
+    jmp .retmove
+.greater1_:
+    dec %rdi
+    dec %rsi
+    mov %rdx, %rcx
+    and $7, %rcx
+    rep movsb
+    cld
+    jmp .retmove
+.lessmove:      // dst < src
+    mov %rdx, %rcx
+    shr $3, %rcx
+    je .less1
+    rep movsq
+.less1:
+    mov %rdx, %rcx
+    and $7, %rcx
+    rep movsb
+.retmove:
+    ret
+
 _cu_memset:
     // RDI - ptr, RSI - val, RDX - size
     mov %rdx, %rax
 
     mov %rsi, %rcx
     shr $3, %rcx
+    je .set1
     rep stosq
-
+.set1:
     mov %rsi, %rcx
     and $7, %rcx
     rep stosb
diff --git a/cutil.c b/cutil.c
index b8170cd..0128301 100644
--- a/cutil.c
+++ b/cutil.c
@@ -25,7 +25,7 @@ void cutil_exit()
 {
 }
 
-void _cu_memmove(void* dst, void* src, size_t size)
+void __cu_memmove(void* dst, void* src, size_t size)
 {
     if (dst > src)
     {
diff --git a/cutil.h b/cutil.h
index c212a49..01cdc8d 100644
--- a/cutil.h
+++ b/cutil.h
@@ -6,6 +6,8 @@
 
 extern void* _cu_memset(void* dst, int val, size_t size);
 extern void* _cu_memcpy(void* dst, const void* src, size_t size);
+extern void* _cu_memmove(void* dst, void* src, size_t size);
+void __cu_memmove(void* dst, void* src, size_t size);
 
 #define CUTIL_MALLOC malloc
 #define CUTIL_REALLOC realloc
@@ -27,6 +29,4 @@ void cutil_exit();
 extern const void* cu_memtest(const void* mem, uint size);
 extern void cu_memzero(void* dst, size_t size);
 
-void _cu_memmove(void* dst, void* src, size_t size);
-
 #endif
\ No newline at end of file
diff --git a/test.c b/test.c
index f9f3556..0e63860 100644
--- a/test.c
+++ b/test.c
@@ -110,20 +110,20 @@ int main()
     char val2[] = {0,0,0,1,0,0,0,0,0,0,4,0,0};
     printf("val2\tcu_memtest\t%p\n", cu_memtest(val2, sizeof(val2)));
 
-    uint val3[4] = {1, 2, 3, 4};
-    uint val4[4];
+    uint val3[5] = {1, 2, 3, 4, 0};
+    uint val4[5];
 
     _cu_memcpy(val4, val3, sizeof(val3));
     cu_memzero(val3, sizeof(val3));
     printf("cu_memzero test\t%p\n", cu_memtest(val3, sizeof(val3)));
     printf("val4 values\n");
-    for (unsigned i = 0; i < 4; i++)
+    for (unsigned i = 0; i < 5; i++)
         printf("\t%u\n", val4[i]);
     printf("\n");
     
     _cu_memmove(&val4[1], &val4[0], sizeof(uint) * 3);
     printf("val4 memmove ->\n");
-    for (unsigned i = 0; i < 4; i++)
+    for (unsigned i = 0; i < 5; i++)
         printf("\t%u\n", val4[i]);
     printf("\n");