implement cu_memtest in x86-64 assembly in more efficient way that C does

arch/x86_64.S

@@ -1,6 +1,7 @@
 .globl hw_bswap16
 .globl hw_bswap32
 .globl hw_bswap64
+.globl cu_memtest
 
 .text
 
@@ -19,3 +20,31 @@ hw_bswap64:
     bswap %rax
     ret
 
+cu_memtest:
+    // RDI - ptr, RSI - size
+    xchg %rsi, %rdi
+    
+    mov %rdi, %rcx
+    shr $3, %rcx
+    je .testok
+.testq:
+    lodsq
+    test %rax, %rax
+    jnz .testfail
+    loop .testq
+
+    mov %rdi, %rcx
+    and $7, %rcx
+    je .testok
+.testb:
+    lodsb
+    test %al, %al
+    jnz .testfail
+    loop .testb
+
+.testok:
+    xor %rax, %rax
+    ret
+.testfail:
+    mov %rsi, %rax
+    ret

cutil.c

@@ -25,7 +25,7 @@ void cutil_exit()
 {
 }
 
-const void* cu_memtest(const void* mem, uint size)
+/*const void* cu_memtest(const void* mem, uint size)
 {
     unsigned i;
     for (i = 0; i < size / sizeof(uword); i++)
@@ -41,4 +41,4 @@ const void* cu_memtest(const void* mem, uint size)
     }
 
     return NULL;
-}
+}*/

cutil.h

@@ -21,6 +21,6 @@ extern void* (*cu_memmove)(void*,const void*,size_t);
 void cutil_init();
 void cutil_exit();
 
-const void* cu_memtest(const void* mem, uint size);
+extern const void* cu_memtest(const void* mem, uint size);
 
 #endif