Blame sysdeps/s390/wcsnlen-vx.S

Packit 6c4009
/* Vector optimized 32/64 bit S/390 version of wcsnlen.
Packit 6c4009
   Copyright (C) 2015-2018 Free Software Foundation, Inc.
Packit 6c4009
   This file is part of the GNU C Library.
Packit 6c4009
Packit 6c4009
   The GNU C Library is free software; you can redistribute it and/or
Packit 6c4009
   modify it under the terms of the GNU Lesser General Public
Packit 6c4009
   License as published by the Free Software Foundation; either
Packit 6c4009
   version 2.1 of the License, or (at your option) any later version.
Packit 6c4009
Packit 6c4009
   The GNU C Library is distributed in the hope that it will be useful,
Packit 6c4009
   but WITHOUT ANY WARRANTY; without even the implied warranty of
Packit 6c4009
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
Packit 6c4009
   Lesser General Public License for more details.
Packit 6c4009
Packit 6c4009
   You should have received a copy of the GNU Lesser General Public
Packit 6c4009
   License along with the GNU C Library; if not, see
Packit 6c4009
   <http://www.gnu.org/licenses/>.  */
Packit 6c4009
Packit a537e0
#include <ifunc-wcsnlen.h>
Packit a537e0
#if HAVE_WCSNLEN_Z13
Packit 6c4009
Packit 6c4009
# include "sysdep.h"
Packit 6c4009
# include "asm-syntax.h"
Packit 6c4009
Packit 6c4009
	.text
Packit 6c4009
Packit 6c4009
/* size_t wcsnlen (const wchar_t *s, size_t maxlen)
Packit 6c4009
   Returns the number of characters in s or at most maxlen.
Packit 6c4009
Packit 6c4009
   Register usage:
Packit 6c4009
   -r1=tmp
Packit 6c4009
   -r2=address of string
Packit 6c4009
   -r3=maxlen (number of characters to be read)
Packit 6c4009
   -r4=tmp
Packit 6c4009
   -r5=current_len and return_value
Packit 6c4009
   -v16=part of s
Packit 6c4009
*/
Packit a537e0
ENTRY(WCSNLEN_Z13)
Packit 6c4009
Packit 6c4009
	.machine "z13"
Packit 6c4009
	.machinemode "zarch_nohighgprs"
Packit 6c4009
Packit 6c4009
# if !defined __s390x__
Packit 6c4009
	llgfr	%r3,%r3
Packit 6c4009
# endif /* !defined __s390x__ */
Packit 6c4009
Packit 6c4009
	clgfi	%r3,0		/* if maxlen == 0, return 0.  */
Packit 6c4009
	locgre	%r2,%r3
Packit 6c4009
	ber	%r14
Packit 6c4009
Packit 6c4009
	vlbb	%v16,0(%r2),6	/* Load s until next 4k-byte boundary.  */
Packit 6c4009
	lcbb	%r1,0(%r2),6	/* Get bytes to 4k-byte boundary or 16.  */
Packit 6c4009
	llgfr	%r1,%r1		/* Convert 32bit to 64bit.  */
Packit 6c4009
Packit 6c4009
	tmll	%r2,3		/* Test if s is 4-byte aligned?  */
Packit 6c4009
	jne	.Lfallback	/* And use common-code variant if not.  */
Packit 6c4009
Packit 6c4009
	/* Check range of maxlen and convert to byte-count.  */
Packit 6c4009
# ifdef __s390x__
Packit 6c4009
	tmhh	%r3,49152	/* Test bit 0 or 1 of maxlen.  */
Packit 6c4009
	lghi	%r4,-4		/* Max byte-count is 18446744073709551612.  */
Packit 6c4009
# else
Packit 6c4009
	tmlh	%r3,49152	/* Test bit 0 or 1 of maxlen.  */
Packit 6c4009
	llilf	%r4,4294967292	/* Max byte-count is 4294967292.  */
Packit 6c4009
# endif /* !__s390x__ */
Packit 6c4009
	sllg	%r3,%r3,2	/* Convert character-count to byte-count.  */
Packit 6c4009
	locgrne	%r3,%r4		/* Use max byte-count, if bit 0/1 was one.  */
Packit 6c4009
Packit 6c4009
	vfenezf	%v16,%v16,%v16	/* Find element not equal with zero search.  */
Packit 6c4009
	clgr	%r1,%r3
Packit 6c4009
	locgrh	%r1,%r3		/* loaded_byte_count
Packit 6c4009
				   = min (loaded_byte_count, maxlen)  */
Packit 6c4009
Packit 6c4009
	vlgvb	%r5,%v16,7	/* Load zero index or 16 if not found.  */
Packit 6c4009
	clrjl	%r5,%r1,.Lend	/* Found zero within loaded bytes -> return.  */
Packit 6c4009
Packit 6c4009
	clgr	%r1,%r3		/* If loaded_byte_count == maxlen -> end.  */
Packit 6c4009
	locgre	%r5,%r3
Packit 6c4009
	je	.Lend
Packit 6c4009
Packit 6c4009
	/* Align s to 16 byte.  */
Packit 6c4009
	risbgn	%r4,%r2,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15.  */
Packit 6c4009
	lghi	%r5,16		/* current_len = 16.  */
Packit 6c4009
	slr	%r5,%r4		/* Compute bytes to 16bytes boundary.  */
Packit 6c4009
Packit 6c4009
	lgr	%r1,%r5		/* If %r5 + 64 < maxlen? -> loop64.  */
Packit 6c4009
	aghi	%r1,64
Packit 6c4009
	clgrjl	%r1,%r3,.Lloop64
Packit 6c4009
Packit 6c4009
	/* Find zero in max 64byte with aligned s.  */
Packit 6c4009
.Llt64:
Packit 6c4009
	vl	%v16,0(%r5,%r2)	/* Load s.  */
Packit 6c4009
	vfenezfs %v16,%v16,%v16	/* Find element not equal with zero search.  */
Packit 6c4009
	je	.Lfound		/* Jump away if zero was found.  */
Packit 6c4009
	aghi	%r5,16
Packit 6c4009
	clgrjhe	%r5,%r3,.Lfound /* If current_len >= maxlen -> end.  */
Packit 6c4009
	vl	%v16,0(%r5,%r2)
Packit 6c4009
	vfenezfs %v16,%v16,%v16
Packit 6c4009
	je	.Lfound
Packit 6c4009
	aghi	%r5,16
Packit 6c4009
	clgrjhe	%r5,%r3,.Lfound
Packit 6c4009
	vl	%v16,0(%r5,%r2)
Packit 6c4009
	vfenezfs %v16,%v16,%v16
Packit 6c4009
	je	.Lfound
Packit 6c4009
	aghi	%r5,16
Packit 6c4009
	clgrjhe	%r5,%r3,.Lfound
Packit 6c4009
	vl	%v16,0(%r5,%r2)
Packit 6c4009
	vfenezfs %v16,%v16,%v16
Packit 6c4009
	j	.Lfound
Packit 6c4009
Packit 6c4009
.Lfound48:
Packit 6c4009
	aghi	%r5,16
Packit 6c4009
.Lfound32:
Packit 6c4009
	aghi	%r5,16
Packit 6c4009
.Lfound16:
Packit 6c4009
	aghi	%r5,16
Packit 6c4009
.Lfound:
Packit 6c4009
	vlgvb	%r4,%v16,7	/* Load byte index of zero or 16 if no zero.  */
Packit 6c4009
	algr	%r5,%r4
Packit 6c4009
Packit 6c4009
	clgr	%r5,%r3
Packit 6c4009
	locgrh	%r5,%r3		/* Return min (current_len, maxlen).  */
Packit 6c4009
.Lend:
Packit 6c4009
	srlg	%r2,%r5,2	/* Convert byte-count to character-count.  */
Packit 6c4009
	br	%r14
Packit 6c4009
Packit 6c4009
	/* Find zero in 16byte aligned loop.  */
Packit 6c4009
.Lloop64:
Packit 6c4009
	vl	%v16,0(%r5,%r2)	/* Load s.  */
Packit 6c4009
	vfenezfs %v16,%v16,%v16	/* Find element not equal with zero search.  */
Packit 6c4009
	je	.Lfound		/* Jump away if zero was found.  */
Packit 6c4009
	vl	%v16,16(%r5,%r2)
Packit 6c4009
	vfenezfs %v16,%v16,%v16
Packit 6c4009
	je	.Lfound16
Packit 6c4009
	vl	%v16,32(%r5,%r2)
Packit 6c4009
	vfenezfs %v16,%v16,%v16
Packit 6c4009
	je	.Lfound32
Packit 6c4009
	vl	%v16,48(%r5,%r2)
Packit 6c4009
	vfenezfs %v16,%v16,%v16
Packit 6c4009
	je	.Lfound48
Packit 6c4009
Packit 6c4009
	aghi	%r5,64
Packit 6c4009
	lgr	%r1,%r5		/* If %r5 + 64 < maxlen? -> loop64.  */
Packit 6c4009
	aghi	%r1,64
Packit 6c4009
	clgrjl	%r1,%r3,.Lloop64
Packit 6c4009
Packit 6c4009
	j	.Llt64
Packit 6c4009
Packit 6c4009
.Lfallback:
Packit a537e0
	jg	WCSNLEN_C
Packit a537e0
END(WCSNLEN_Z13)
Packit a537e0
Packit a537e0
# if ! HAVE_WCSNLEN_IFUNC
Packit a537e0
strong_alias (WCSNLEN_Z13, __wcsnlen)
Packit a537e0
weak_alias (__wcsnlen, wcsnlen)
Packit a537e0
# endif
Packit a537e0
#endif