162306a36Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0 */ 262306a36Sopenharmony_ci/* 362306a36Sopenharmony_ci * strlen.S (c) 1995 David Mosberger (davidm@cs.arizona.edu) 462306a36Sopenharmony_ci * 562306a36Sopenharmony_ci * Finds length of a 0-terminated string. Optimized for the 662306a36Sopenharmony_ci * Alpha architecture: 762306a36Sopenharmony_ci * 862306a36Sopenharmony_ci * - memory accessed as aligned quadwords only 962306a36Sopenharmony_ci * - uses bcmpge to compare 8 bytes in parallel 1062306a36Sopenharmony_ci * - does binary search to find 0 byte in last 1162306a36Sopenharmony_ci * quadword (HAKMEM needed 12 instructions to 1262306a36Sopenharmony_ci * do this instead of the 9 instructions that 1362306a36Sopenharmony_ci * binary search needs). 1462306a36Sopenharmony_ci */ 1562306a36Sopenharmony_ci#include <linux/export.h> 1662306a36Sopenharmony_ci .set noreorder 1762306a36Sopenharmony_ci .set noat 1862306a36Sopenharmony_ci 1962306a36Sopenharmony_ci .align 3 2062306a36Sopenharmony_ci 2162306a36Sopenharmony_ci .globl strlen 2262306a36Sopenharmony_ci .ent strlen 2362306a36Sopenharmony_ci 2462306a36Sopenharmony_cistrlen: 2562306a36Sopenharmony_ci ldq_u $1, 0($16) # load first quadword ($16 may be misaligned) 2662306a36Sopenharmony_ci lda $2, -1($31) 2762306a36Sopenharmony_ci insqh $2, $16, $2 2862306a36Sopenharmony_ci andnot $16, 7, $0 2962306a36Sopenharmony_ci or $2, $1, $1 3062306a36Sopenharmony_ci cmpbge $31, $1, $2 # $2 <- bitmask: bit i == 1 <==> i-th byte == 0 3162306a36Sopenharmony_ci bne $2, found 3262306a36Sopenharmony_ci 3362306a36Sopenharmony_ciloop: ldq $1, 8($0) 3462306a36Sopenharmony_ci addq $0, 8, $0 # addr += 8 3562306a36Sopenharmony_ci nop # helps dual issue last two insns 3662306a36Sopenharmony_ci cmpbge $31, $1, $2 3762306a36Sopenharmony_ci beq $2, loop 3862306a36Sopenharmony_ci 3962306a36Sopenharmony_cifound: blbs $2, done # make aligned case fast 4062306a36Sopenharmony_ci negq $2, $3 4162306a36Sopenharmony_ci and $2, $3, $2 4262306a36Sopenharmony_ci 4362306a36Sopenharmony_ci and $2, 0x0f, $1 4462306a36Sopenharmony_ci addq $0, 4, $3 4562306a36Sopenharmony_ci cmoveq $1, $3, $0 4662306a36Sopenharmony_ci 4762306a36Sopenharmony_ci and $2, 0x33, $1 4862306a36Sopenharmony_ci addq $0, 2, $3 4962306a36Sopenharmony_ci cmoveq $1, $3, $0 5062306a36Sopenharmony_ci 5162306a36Sopenharmony_ci and $2, 0x55, $1 5262306a36Sopenharmony_ci addq $0, 1, $3 5362306a36Sopenharmony_ci cmoveq $1, $3, $0 5462306a36Sopenharmony_ci 5562306a36Sopenharmony_cidone: subq $0, $16, $0 5662306a36Sopenharmony_ci ret $31, ($26) 5762306a36Sopenharmony_ci 5862306a36Sopenharmony_ci .end strlen 5962306a36Sopenharmony_ci EXPORT_SYMBOL(strlen) 60