1 files changed, 9 insertions, 8 deletions
diff --git a/zap/source/amd64/mem/cp.s b/zap/source/amd64/mem/cp.s
index 2860acf..3c9f2e1 100644
--- a/zap/source/amd64/mem/cp.s
+++ b/zap/source/amd64/mem/cp.s
@@ -19,8 +19,8 @@ zap_cp:
 vmovdqu ymm0,[rsi] # val02 = *(unsigned int256_t *)src;
 vmovdqu [rdi],ymm0 # *(unsigned int256_t *)dest = val02;
 
- add rsi,0x20 # dest += 0x20u;
- add rdi,0x20 # src += 0x20u;
+ add rdi,0x20 # dest += 0x20u;
+ add rsi,0x20 # src += 0x20u;
 sub rdx,0x20 # num -= 0x20u;
 jmp short .big02cp # goto big02cp;
 
@@ -31,8 +31,8 @@ zap_cp:
 movdqu xmm0,[rsi] # val01 = *(unsigned int128_t *)src;
 movdqu [rdi],xmm0 # *(unsigned int128_t *)dest = val01;
 
- add rsi,0x10 # dest += 0x10u;
- add rdi,0x10 # src += 0x10u;
+ add rdi,0x10 # dest += 0x10u;
+ add rsi,0x10 # src += 0x10u;
 sub rdx,0x10 # num -= 0x10u;
 jmp short .big01cp # goto big01cp;
 
@@ -43,8 +43,8 @@ zap_cp:
 mov rcx,[rsi] # val8 = *(zap_i04 *)src;
 mov [rdi],rcx # *(zap_i04 *)dest = val8;
 
- add rdi,0x8 # dest += 0x8u;
- add rsi,0x8 # src += 0x8u;
+ add rsi,0x8 # dest += 0x8u;
+ add rdi,0x8 # src += 0x8u;
 sub rdx,0x8 # num -= 0x8u;
 jmp short .wrdcp # goto wrdcp
 
@@ -53,7 +53,7 @@ zap_cp:
 jz short .done # goto done
 
 mov cl,[rsi] # val1 = *(zap_i8 *)src;
- mov [rsi],cl # *(zap_i8 *)dest = val1;
+ mov [rdi],cl # *(zap_i8 *)dest = val1;
 
 inc rdi # ++dest;
 inc rsi # ++src;
@@ -61,4 +61,5 @@ zap_cp:
 jmp short .bytecp # goto bytecp;
 
 .done:
- ret # return
+ mov rax,rdi
+ ret # return dest;