diff options
Diffstat (limited to 'MdePkg/Library/BaseMemoryLibSse2/X64/CopyMem.S')
| -rw-r--r-- | MdePkg/Library/BaseMemoryLibSse2/X64/CopyMem.S | 26 |
1 files changed, 13 insertions, 13 deletions
diff --git a/MdePkg/Library/BaseMemoryLibSse2/X64/CopyMem.S b/MdePkg/Library/BaseMemoryLibSse2/X64/CopyMem.S index dc47510874..35db797677 100644 --- a/MdePkg/Library/BaseMemoryLibSse2/X64/CopyMem.S +++ b/MdePkg/Library/BaseMemoryLibSse2/X64/CopyMem.S @@ -44,38 +44,38 @@ ASM_PFX(InternalMemCopyMem): lea r9, [rsi + r8 - 1] # r9 <- Last byte of Source cmp rsi, rdi mov rax, rdi # rax <- Destination as return value - jae _InternalMemCopyMem_al_0000 # Copy forward if Source > Destination + jae L0 # Copy forward if Source > Destination cmp r9, rdi # Overlapped? - jae _atSym_CopyBackward # Copy backward if overlapped -_InternalMemCopyMem_al_0000: + jae L_CopyBackward # Copy backward if overlapped +L0: xor rcx, rcx sub rcx, rdi # rcx <- -rdi and rcx, 15 # rcx + rsi should be 16 bytes aligned - jz _InternalMemCopyMem_al_0001 # skip if rcx == 0 + jz L1 # skip if rcx == 0 cmp rcx, r8 cmova rcx, r8 sub r8, rcx rep movsb -_InternalMemCopyMem_al_0001: +L1: mov rcx, r8 and r8, 15 shr rcx, 4 # rcx <- # of DQwords to copy - jz _atSym_CopyBytes - movdqa [rsp + 0x18], xmm0 # save xmm0 on stack -_InternalMemCopyMem_al_0002: + jz L_CopyBytes + movdqa [rsp + 0x18], xmm0 # save xmm0 on stack +L2: movdqu xmm0, [rsi] # rsi may not be 16-byte aligned movntdq [rdi], xmm0 # rdi should be 16-byte aligned add rsi, 16 add rdi, 16 - loop _InternalMemCopyMem_al_0002 + loop L2 mfence - movdqa xmm0, [rsp + 0x18] # restore xmm0 - jmp _atSym_CopyBytes # copy remaining bytes -_atSym_CopyBackward: + movdqa xmm0, [rsp + 0x18] # restore xmm0 + jmp L_CopyBytes # copy remaining bytes +L_CopyBackward: mov rsi, r9 # rsi <- Last byte of Source lea rdi, [rdi + r8 - 1] # rdi <- Last byte of Destination std -_atSym_CopyBytes: +L_CopyBytes: mov rcx, r8 rep movsb cld |