1*4882a593Smuzhiyun/* SPDX-License-Identifier: GPL-2.0 */ 2*4882a593Smuzhiyun// Copyright (C) 2018 Hangzhou C-SKY Microsystems co.,ltd. 3*4882a593Smuzhiyun 4*4882a593Smuzhiyun#include <linux/linkage.h> 5*4882a593Smuzhiyun#include "sysdep.h" 6*4882a593Smuzhiyun 7*4882a593SmuzhiyunENTRY(__memcpy) 8*4882a593SmuzhiyunENTRY(memcpy) 9*4882a593Smuzhiyun /* Test if len less than 4 bytes. */ 10*4882a593Smuzhiyun mov r12, r0 11*4882a593Smuzhiyun cmplti r2, 4 12*4882a593Smuzhiyun bt .L_copy_by_byte 13*4882a593Smuzhiyun 14*4882a593Smuzhiyun andi r13, r0, 3 15*4882a593Smuzhiyun movi r19, 4 16*4882a593Smuzhiyun /* Test if dest is not 4 bytes aligned. */ 17*4882a593Smuzhiyun bnez r13, .L_dest_not_aligned 18*4882a593Smuzhiyun 19*4882a593Smuzhiyun/* Hardware can handle unaligned access directly. */ 20*4882a593Smuzhiyun.L_dest_aligned: 21*4882a593Smuzhiyun /* If dest is aligned, then copy. */ 22*4882a593Smuzhiyun zext r18, r2, 31, 4 23*4882a593Smuzhiyun 24*4882a593Smuzhiyun /* Test if len less than 16 bytes. */ 25*4882a593Smuzhiyun bez r18, .L_len_less_16bytes 26*4882a593Smuzhiyun movi r19, 0 27*4882a593Smuzhiyun 28*4882a593Smuzhiyun LABLE_ALIGN 29*4882a593Smuzhiyun.L_len_larger_16bytes: 30*4882a593Smuzhiyun#if defined(__CK860__) 31*4882a593Smuzhiyun ldw r3, (r1, 0) 32*4882a593Smuzhiyun stw r3, (r0, 0) 33*4882a593Smuzhiyun ldw r3, (r1, 4) 34*4882a593Smuzhiyun stw r3, (r0, 4) 35*4882a593Smuzhiyun ldw r3, (r1, 8) 36*4882a593Smuzhiyun stw r3, (r0, 8) 37*4882a593Smuzhiyun ldw r3, (r1, 12) 38*4882a593Smuzhiyun addi r1, 16 39*4882a593Smuzhiyun stw r3, (r0, 12) 40*4882a593Smuzhiyun addi r0, 16 41*4882a593Smuzhiyun#else 42*4882a593Smuzhiyun ldw r20, (r1, 0) 43*4882a593Smuzhiyun ldw r21, (r1, 4) 44*4882a593Smuzhiyun ldw r22, (r1, 8) 45*4882a593Smuzhiyun ldw r23, (r1, 12) 46*4882a593Smuzhiyun stw r20, (r0, 0) 47*4882a593Smuzhiyun stw r21, (r0, 4) 48*4882a593Smuzhiyun stw r22, (r0, 8) 49*4882a593Smuzhiyun stw r23, (r0, 12) 50*4882a593Smuzhiyun PRE_BNEZAD (r18) 51*4882a593Smuzhiyun addi r1, 16 52*4882a593Smuzhiyun addi r0, 16 53*4882a593Smuzhiyun#endif 54*4882a593Smuzhiyun BNEZAD (r18, .L_len_larger_16bytes) 55*4882a593Smuzhiyun 56*4882a593Smuzhiyun.L_len_less_16bytes: 57*4882a593Smuzhiyun zext r18, r2, 3, 2 58*4882a593Smuzhiyun bez r18, .L_copy_by_byte 59*4882a593Smuzhiyun.L_len_less_16bytes_loop: 60*4882a593Smuzhiyun ldw r3, (r1, 0) 61*4882a593Smuzhiyun PRE_BNEZAD (r18) 62*4882a593Smuzhiyun addi r1, 4 63*4882a593Smuzhiyun stw r3, (r0, 0) 64*4882a593Smuzhiyun addi r0, 4 65*4882a593Smuzhiyun BNEZAD (r18, .L_len_less_16bytes_loop) 66*4882a593Smuzhiyun 67*4882a593Smuzhiyun/* Test if len less than 4 bytes. */ 68*4882a593Smuzhiyun.L_copy_by_byte: 69*4882a593Smuzhiyun zext r18, r2, 1, 0 70*4882a593Smuzhiyun bez r18, .L_return 71*4882a593Smuzhiyun.L_copy_by_byte_loop: 72*4882a593Smuzhiyun ldb r3, (r1, 0) 73*4882a593Smuzhiyun PRE_BNEZAD (r18) 74*4882a593Smuzhiyun addi r1, 1 75*4882a593Smuzhiyun stb r3, (r0, 0) 76*4882a593Smuzhiyun addi r0, 1 77*4882a593Smuzhiyun BNEZAD (r18, .L_copy_by_byte_loop) 78*4882a593Smuzhiyun 79*4882a593Smuzhiyun.L_return: 80*4882a593Smuzhiyun mov r0, r12 81*4882a593Smuzhiyun rts 82*4882a593Smuzhiyun 83*4882a593Smuzhiyun/* 84*4882a593Smuzhiyun * If dest is not aligned, just copying some bytes makes the 85*4882a593Smuzhiyun * dest align. 86*4882a593Smuzhiyun */ 87*4882a593Smuzhiyun.L_dest_not_aligned: 88*4882a593Smuzhiyun sub r13, r19, r13 89*4882a593Smuzhiyun sub r2, r13 90*4882a593Smuzhiyun 91*4882a593Smuzhiyun/* Makes the dest align. */ 92*4882a593Smuzhiyun.L_dest_not_aligned_loop: 93*4882a593Smuzhiyun ldb r3, (r1, 0) 94*4882a593Smuzhiyun PRE_BNEZAD (r13) 95*4882a593Smuzhiyun addi r1, 1 96*4882a593Smuzhiyun stb r3, (r0, 0) 97*4882a593Smuzhiyun addi r0, 1 98*4882a593Smuzhiyun BNEZAD (r13, .L_dest_not_aligned_loop) 99*4882a593Smuzhiyun cmplti r2, 4 100*4882a593Smuzhiyun bt .L_copy_by_byte 101*4882a593Smuzhiyun 102*4882a593Smuzhiyun /* Check whether the src is aligned. */ 103*4882a593Smuzhiyun jbr .L_dest_aligned 104*4882a593SmuzhiyunENDPROC(__memcpy) 105