Lines Matching +full:8 +full:- +full:byte

1 /*-
4  * SPDX-License-Expression: BSD-2-Clause
29  * This stpcpy implementation copies a byte at a time until the
31  * words until it finds a word containing a zero byte, and finally
36  * it is still faster than either byte copies or the overhead of
65 	addq	$8,%rdi
69 	addq	$8,%rsi
80 	testb	%dl,%dl		/* 1st byte == 0? */
84 	shrq	$8,%rdx
86 	testb	%dl,%dl		/* 2nd byte == 0? */
90 	shrq	$8,%rdx
92 	testb	%dl,%dl		/* 3rd byte == 0? */
96 	shrq	$8,%rdx
98 	testb	%dl,%dl		/* 4th byte == 0? */
102 	shrq	$8,%rdx
104 	testb	%dl,%dl		/* 5th byte == 0? */
108 	shrq	$8,%rdx
110 	testb	%dl,%dl		/* 6th byte == 0? */
114 	shrq	$8,%rdx
116 	testb	%dl,%dl		/* 7th byte == 0? */
120 	shrq	$8,%rdx
123 	testb	%dl,%dl		/* 8th byte == 0? */
136 	and	$~0xf, %rsi		# align source to 16 byte
140 	pcmpeqb	%xmm1, %xmm0		# NUL byte present?
149 	pcmpeqb	%xmm0, %xmm1		# NUL byte present?
162 	pcmpeqb	%xmm2, %xmm1		# NUL byte present?
170 	pcmpeqb	%xmm0, %xmm1		# NUL byte present?
178 					# (behaves as bsf on pre-x86-64-v3 CPUs)
179 	add	%rsi, %rax		# point to NUL byte
180 	movdqu	-15(%rax), %xmm0	# last 16 bytes of string
181 	movdqu	%xmm0, -15(%rax, %rdi, 1) # copied to destination
182 	add	%rdi, %rax		# point to destination's NUL byte
192 .Lrunt:	lea	1(%rax), %edi		# string length including NUL byte
194 	add	%rdx, %rax		# point to NUL byte
196 	/* transfer 16--32 bytes */
200 	movdqu	-16(%rsi, %rdi, 1), %xmm0 # load last 16 bytes
202 	movdqu	%xmm0, -15(%rax)	# store last 16 bytes
205 	/* transfer 8--15 bytes */
206 .L0815:	cmp	$8, %edi
209 	mov	(%rsi), %rcx		# load first 8 bytes
210 	mov	-8(%rsi, %rdi, 1), %rdi	# load last 8 bytes
212 	mov	%rdi, -7(%rax)		# dito
215 	/* transfer 4--7 bytes */
220 	mov	-4(%rsi, %rdi, 1), %edi
222 	mov	%edi, -3(%rax)
225 	/* transfer 2--3 bytes */
232 	/* transfer 0 bytes (last byte is always NUL) */
233 .L0101:	movb	$0, (%rax)		# store terminating NUL byte
237 	.section .note.GNU-stack,"",%progbits