implement x86-64 assembly-optimized functions _cu_memcpy, _cu_memset, cu_memzero

arch/x86_64.S

@@ -1,6 +1,9 @@
 .globl hw_bswap16
 .globl hw_bswap32
 .globl hw_bswap64
+.globl _cu_memcpy
+.globl _cu_memset
+.globl cu_memzero
 .globl cu_memtest
 
 .text
@@ -20,6 +23,45 @@ hw_bswap64:
     bswap %rax
     ret
 
+
+_cu_memcpy:
+    // RDI - dst, RSI - src, RDX - size
+    mov %rdx, %rcx
+    shr $3, %rcx
+    rep movsq
+
+    mov %rdx, %rcx
+    and $7, %rcx
+    rep movsb
+    ret
+
+_cu_memset:
+    // RDI - ptr, RSI - val, RDX - size
+    mov %rdx, %rax
+
+    mov %rsi, %rcx
+    shr $3, %rcx
+    rep stosq
+
+    mov %rsi, %rcx
+    and $7, %rcx
+    rep stosb
+
+    ret
+
+cu_memzero:
+    xor %rax, %rax
+
+    mov %rsi, %rcx
+    shr $3, %rcx
+    rep stosq
+
+    mov %rsi, %rcx
+    and $7, %rcx
+    rep stosb
+
+    ret
+
 cu_memtest:
     // RDI - ptr, RSI - size
     xchg %rsi, %rdi

cutil.h

@@ -4,6 +4,9 @@
 #include "cutypes.h"
 #include <stdlib.h>
 
+extern void* _cu_memset(void* dst, int val, size_t size);
+extern void* _cu_memcpy(void* dst, const void* src, size_t size);
+
 #define CUTIL_MALLOC malloc
 #define CUTIL_REALLOC realloc
 #define CUTIL_FREE free
@@ -22,5 +25,6 @@ void cutil_init();
 void cutil_exit();
 
 extern const void* cu_memtest(const void* mem, uint size);
+extern void cu_memzero(void* dst, size_t size);
 
 #endif

test.c

@@ -109,6 +109,17 @@ 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];
+
+    _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++)
+        printf("\t%u\n", val4[i]);
+    printf("\n");
+
     printf("[endian]\n");
     printf("cutil_endian\t%u\n", cu_endian);
     printf("%x\t%x\n", 0x1234, cu_bswap16(0x1234));