1*4882a593Smuzhiyun/* SPDX-License-Identifier: GPL-2.0 */ 2*4882a593Smuzhiyun/* 3*4882a593Smuzhiyun * memscan.S: Optimized memscan for the Sparc. 4*4882a593Smuzhiyun * 5*4882a593Smuzhiyun * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu) 6*4882a593Smuzhiyun */ 7*4882a593Smuzhiyun 8*4882a593Smuzhiyun#include <asm/export.h> 9*4882a593Smuzhiyun 10*4882a593Smuzhiyun/* In essence, this is just a fancy strlen. */ 11*4882a593Smuzhiyun 12*4882a593Smuzhiyun#define LO_MAGIC 0x01010101 13*4882a593Smuzhiyun#define HI_MAGIC 0x80808080 14*4882a593Smuzhiyun 15*4882a593Smuzhiyun .text 16*4882a593Smuzhiyun .align 4 17*4882a593Smuzhiyun .globl __memscan_zero, __memscan_generic 18*4882a593Smuzhiyun .globl memscan 19*4882a593SmuzhiyunEXPORT_SYMBOL(__memscan_zero) 20*4882a593SmuzhiyunEXPORT_SYMBOL(__memscan_generic) 21*4882a593Smuzhiyun__memscan_zero: 22*4882a593Smuzhiyun /* %o0 = addr, %o1 = size */ 23*4882a593Smuzhiyun cmp %o1, 0 24*4882a593Smuzhiyun bne,a 1f 25*4882a593Smuzhiyun andcc %o0, 3, %g0 26*4882a593Smuzhiyun 27*4882a593Smuzhiyun retl 28*4882a593Smuzhiyun nop 29*4882a593Smuzhiyun 30*4882a593Smuzhiyun1: 31*4882a593Smuzhiyun be mzero_scan_word 32*4882a593Smuzhiyun sethi %hi(HI_MAGIC), %g2 33*4882a593Smuzhiyun 34*4882a593Smuzhiyun ldsb [%o0], %g3 35*4882a593Smuzhiyunmzero_still_not_word_aligned: 36*4882a593Smuzhiyun cmp %g3, 0 37*4882a593Smuzhiyun bne 1f 38*4882a593Smuzhiyun add %o0, 1, %o0 39*4882a593Smuzhiyun 40*4882a593Smuzhiyun retl 41*4882a593Smuzhiyun sub %o0, 1, %o0 42*4882a593Smuzhiyun 43*4882a593Smuzhiyun1: 44*4882a593Smuzhiyun subcc %o1, 1, %o1 45*4882a593Smuzhiyun bne,a 1f 46*4882a593Smuzhiyun andcc %o0, 3, %g0 47*4882a593Smuzhiyun 48*4882a593Smuzhiyun retl 49*4882a593Smuzhiyun nop 50*4882a593Smuzhiyun 51*4882a593Smuzhiyun1: 52*4882a593Smuzhiyun bne,a mzero_still_not_word_aligned 53*4882a593Smuzhiyun ldsb [%o0], %g3 54*4882a593Smuzhiyun 55*4882a593Smuzhiyun sethi %hi(HI_MAGIC), %g2 56*4882a593Smuzhiyunmzero_scan_word: 57*4882a593Smuzhiyun or %g2, %lo(HI_MAGIC), %o3 58*4882a593Smuzhiyun sethi %hi(LO_MAGIC), %g3 59*4882a593Smuzhiyun or %g3, %lo(LO_MAGIC), %o2 60*4882a593Smuzhiyunmzero_next_word: 61*4882a593Smuzhiyun ld [%o0], %g2 62*4882a593Smuzhiyunmzero_next_word_preloaded: 63*4882a593Smuzhiyun sub %g2, %o2, %g2 64*4882a593Smuzhiyunmzero_next_word_preloaded_next: 65*4882a593Smuzhiyun andcc %g2, %o3, %g0 66*4882a593Smuzhiyun bne mzero_byte_zero 67*4882a593Smuzhiyun add %o0, 4, %o0 68*4882a593Smuzhiyun 69*4882a593Smuzhiyunmzero_check_out_of_fuel: 70*4882a593Smuzhiyun subcc %o1, 4, %o1 71*4882a593Smuzhiyun bg,a 1f 72*4882a593Smuzhiyun ld [%o0], %g2 73*4882a593Smuzhiyun 74*4882a593Smuzhiyun retl 75*4882a593Smuzhiyun nop 76*4882a593Smuzhiyun 77*4882a593Smuzhiyun1: 78*4882a593Smuzhiyun b mzero_next_word_preloaded_next 79*4882a593Smuzhiyun sub %g2, %o2, %g2 80*4882a593Smuzhiyun 81*4882a593Smuzhiyun /* Check every byte. */ 82*4882a593Smuzhiyunmzero_byte_zero: 83*4882a593Smuzhiyun ldsb [%o0 - 4], %g2 84*4882a593Smuzhiyun cmp %g2, 0 85*4882a593Smuzhiyun bne mzero_byte_one 86*4882a593Smuzhiyun sub %o0, 4, %g3 87*4882a593Smuzhiyun 88*4882a593Smuzhiyun retl 89*4882a593Smuzhiyun mov %g3, %o0 90*4882a593Smuzhiyun 91*4882a593Smuzhiyunmzero_byte_one: 92*4882a593Smuzhiyun ldsb [%o0 - 3], %g2 93*4882a593Smuzhiyun cmp %g2, 0 94*4882a593Smuzhiyun bne,a mzero_byte_two_and_three 95*4882a593Smuzhiyun ldsb [%o0 - 2], %g2 96*4882a593Smuzhiyun 97*4882a593Smuzhiyun retl 98*4882a593Smuzhiyun sub %o0, 3, %o0 99*4882a593Smuzhiyun 100*4882a593Smuzhiyunmzero_byte_two_and_three: 101*4882a593Smuzhiyun cmp %g2, 0 102*4882a593Smuzhiyun bne,a 1f 103*4882a593Smuzhiyun ldsb [%o0 - 1], %g2 104*4882a593Smuzhiyun 105*4882a593Smuzhiyun retl 106*4882a593Smuzhiyun sub %o0, 2, %o0 107*4882a593Smuzhiyun 108*4882a593Smuzhiyun1: 109*4882a593Smuzhiyun cmp %g2, 0 110*4882a593Smuzhiyun bne,a mzero_next_word_preloaded 111*4882a593Smuzhiyun ld [%o0], %g2 112*4882a593Smuzhiyun 113*4882a593Smuzhiyun retl 114*4882a593Smuzhiyun sub %o0, 1, %o0 115*4882a593Smuzhiyun 116*4882a593Smuzhiyunmzero_found_it: 117*4882a593Smuzhiyun retl 118*4882a593Smuzhiyun sub %o0, 2, %o0 119*4882a593Smuzhiyun 120*4882a593Smuzhiyunmemscan: 121*4882a593Smuzhiyun__memscan_generic: 122*4882a593Smuzhiyun /* %o0 = addr, %o1 = c, %o2 = size */ 123*4882a593Smuzhiyun cmp %o2, 0 124*4882a593Smuzhiyun bne,a 0f 125*4882a593Smuzhiyun ldub [%o0], %g2 126*4882a593Smuzhiyun 127*4882a593Smuzhiyun b,a 2f 128*4882a593Smuzhiyun1: 129*4882a593Smuzhiyun ldub [%o0], %g2 130*4882a593Smuzhiyun0: 131*4882a593Smuzhiyun cmp %g2, %o1 132*4882a593Smuzhiyun be 2f 133*4882a593Smuzhiyun addcc %o2, -1, %o2 134*4882a593Smuzhiyun bne 1b 135*4882a593Smuzhiyun add %o0, 1, %o0 136*4882a593Smuzhiyun2: 137*4882a593Smuzhiyun retl 138*4882a593Smuzhiyun nop 139