projects / mirror_edk2.git / blame
| Commit | Line | Data |
|---|---|---|
| e8758b6a JJ |
1 | ;------------------------------------------------------------------------------\r |
| 2 | ;\r | |
| 3 | ; Copyright (c) 2006, Intel Corporation. All rights reserved.<BR>\r | |
| 9344f092 | 4 | ; SPDX-License-Identifier: BSD-2-Clause-Patent\r |
| e8758b6a JJ |
5 | ;\r |
| 6 | ; Module Name:\r | |
| 7 | ;\r | |
| 8 | ; SetMem.nasm\r | |
| 9 | ;\r | |
| 10 | ; Abstract:\r | |
| 11 | ;\r | |
| 12 | ; SetMem function\r | |
| 13 | ;\r | |
| 14 | ; Notes:\r | |
| 15 | ;\r | |
| 16 | ;------------------------------------------------------------------------------\r | |
| 17 | \r | |
| 18 | DEFAULT REL\r | |
| 19 | SECTION .text\r | |
| 20 | \r | |
| 21 | ;------------------------------------------------------------------------------\r | |
| 22 | ; VOID *\r | |
| 23 | ; InternalMemSetMem (\r | |
| 24 | ; IN VOID *Buffer,\r | |
| 25 | ; IN UINTN Count,\r | |
| 26 | ; IN UINT8 Value\r | |
| 27 | ; )\r | |
| 28 | ;------------------------------------------------------------------------------\r | |
| 29 | global ASM_PFX(InternalMemSetMem)\r | |
| 30 | ASM_PFX(InternalMemSetMem):\r | |
| 31 | push rdi\r | |
| 32 | mov rdi, rcx ; rdi <- Buffer\r | |
| 33 | mov al, r8b ; al <- Value\r | |
| 34 | mov r9, rdi ; r9 <- Buffer as return value\r | |
| 35 | xor rcx, rcx\r | |
| 36 | sub rcx, rdi\r | |
| 37 | and rcx, 15 ; rcx + rdi aligns on 16-byte boundary\r | |
| 38 | jz .0\r | |
| 39 | cmp rcx, rdx\r | |
| 40 | cmova rcx, rdx\r | |
| 41 | sub rdx, rcx\r | |
| 42 | rep stosb\r | |
| 43 | .0:\r | |
| 44 | mov rcx, rdx\r | |
| 45 | and rdx, 15\r | |
| 46 | shr rcx, 4\r | |
| 47 | jz @SetBytes\r | |
| 48 | mov ah, al ; ax <- Value repeats twice\r | |
| 49 | movdqa [rsp + 0x10], xmm0 ; save xmm0\r | |
| 50 | movd xmm0, eax ; xmm0[0..16] <- Value repeats twice\r | |
| 51 | pshuflw xmm0, xmm0, 0 ; xmm0[0..63] <- Value repeats 8 times\r | |
| 52 | movlhps xmm0, xmm0 ; xmm0 <- Value repeats 16 times\r | |
| 53 | .1:\r | |
| 54 | movntdq [rdi], xmm0 ; rdi should be 16-byte aligned\r | |
| 55 | add rdi, 16\r | |
| 56 | loop .1\r | |
| 57 | mfence\r | |
| 58 | movdqa xmm0, [rsp + 0x10] ; restore xmm0\r | |
| 59 | @SetBytes:\r | |
| 60 | mov ecx, edx ; high 32 bits of rcx are always zero\r | |
| 61 | rep stosb\r | |
| 62 | mov rax, r9 ; rax <- Return value\r | |
| 63 | pop rdi\r | |
| 64 | ret\r | |
| 65 | \r |