Blame sysdeps/s390/wmemset-vx.S

Packit 6c4009
/* Vector Optimized 32/64 bit S/390 version of wmemset.
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 Service 02e39a
#include <ifunc-wmemset.h>
Packit Service 02e39a
#if HAVE_WMEMSET_Z13
Packit 6c4009
Packit 6c4009
# include "sysdep.h"
Packit 6c4009
# include "asm-syntax.h"
Packit 6c4009
Packit 6c4009
	.text
Packit 6c4009
Packit 6c4009
/* wchar_t *wmemset(wchar_t *dest, wchar_t wc, size_t n)
Packit 6c4009
   Fill an array of wide-characters with a constant wide character
Packit 6c4009
   and returns dest.
Packit 6c4009
Packit 6c4009
   Register usage:
Packit 6c4009
   -r0=tmp
Packit 6c4009
   -r1=tmp
Packit 6c4009
   -r2=dest or current-pointer
Packit 6c4009
   -r3=wc
Packit 6c4009
   -r4=n
Packit 6c4009
   -r5=tmp
Packit 6c4009
   -v16=replicated wc
Packit 6c4009
   -v17,v18,v19=copy of v16 for vstm
Packit 6c4009
   -v31=saved dest for return
Packit 6c4009
*/
Packit Service 02e39a
ENTRY(WMEMSET_Z13)
Packit 6c4009
	.machine "z13"
Packit 6c4009
	.machinemode "zarch_nohighgprs"
Packit 6c4009
Packit 6c4009
# if !defined __s390x__
Packit 6c4009
	llgfr	%r4,%r4
Packit 6c4009
# endif /* !defined __s390x__ */
Packit 6c4009
Packit 6c4009
	vlvgg	%v31,%r2,0	/* Save destination pointer for return.  */
Packit 6c4009
	clgije	%r4,0,.Lend
Packit 6c4009
Packit 6c4009
	vlvgf	%v16,%r3,0	/* Generate vector with wchar_t wc.  */
Packit 6c4009
	vrepf	%v16,%v16,0
Packit 6c4009
Packit 6c4009
	/* Check range of maxlen and convert to byte-count.  */
Packit 6c4009
# ifdef __s390x__
Packit 6c4009
	tmhh	%r4,49152	/* Test bit 0 or 1 of maxlen.  */
Packit 6c4009
	lghi	%r5,-4		/* Max byte-count is 18446744073709551612.  */
Packit 6c4009
# else
Packit 6c4009
	tmlh	%r4,49152	/* Test bit 0 or 1 of maxlen.  */
Packit 6c4009
	llilf	%r5,4294967292	/* Max byte-count is 4294967292.  */
Packit 6c4009
# endif /* !__s390x__ */
Packit 6c4009
	sllg	%r4,%r4,2	/* Convert character-count to byte-count.  */
Packit 6c4009
	locgrne	%r4,%r5		/* Use max byte-count, if bit 0/1 was one.  */
Packit 6c4009
Packit 6c4009
	/* Align dest to 16 byte.  */
Packit 6c4009
	risbg	%r0,%r2,60,128+63,0 /* Test if s is aligned and
Packit 6c4009
				       %r3 = bits 60-63 'and' 15.  */
Packit 6c4009
	je	.Lpreloop	/* If s is aligned, loop aligned.  */
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
	lghi	%r1,16
Packit 6c4009
	slr	%r1,%r0		/* Compute byte count to load (16-x).  */
Packit 6c4009
	clgr	%r1,%r4
Packit 6c4009
	locgrh	%r1,%r4		/* min (byte count, n)  */
Packit 6c4009
	aghik	%r5,%r1,-1	/* vstl needs highest index.  */
Packit 6c4009
	vstl	%v16,%r5,0(%r2)	/* Store remaining bytes.  */
Packit 6c4009
	clgrje	%r1,%r4,.Lend	/* Return if n bytes where set.  */
Packit 6c4009
	slgr	%r4,%r1		/* Compute remaining byte count.  */
Packit 6c4009
	la	%r2,0(%r1,%r2)
Packit 6c4009
Packit 6c4009
.Lpreloop:
Packit 6c4009
	/* Now we are 16-byte aligned.  */
Packit 6c4009
	clgijl	%r4,17,.Lremaining
Packit 6c4009
	srlg	%r1,%r4,8	/* Split into 256byte blocks */
Packit 6c4009
	clgije	%r1,0,.Lpreloop64
Packit 6c4009
	vlr	%v17,%v16
Packit 6c4009
	vlr	%v18,%v16
Packit 6c4009
	vlr	%v19,%v16
Packit 6c4009
Packit 6c4009
.Lloop256:
Packit 6c4009
	vstm	%v16,%v19,0(%r2)
Packit 6c4009
	vstm	%v16,%v19,64(%r2)
Packit 6c4009
	vstm	%v16,%v19,128(%r2)
Packit 6c4009
	vstm	%v16,%v19,192(%r2)
Packit 6c4009
	la	%r2,256(%r2)
Packit 6c4009
	brctg	%r1,.Lloop256	/* Loop until all blocks are processed.  */
Packit 6c4009
Packit 6c4009
	llgfr	%r4,%r4
Packit 6c4009
	nilf	%r4,255		/* Get remaining bytes */
Packit 6c4009
	je	.Lend		/* Skip store remaining bytes if zero.  */
Packit 6c4009
Packit 6c4009
.Lpreloop64:
Packit 6c4009
	clgijl	%r4,17,.Lremaining
Packit 6c4009
	clgijl	%r4,33,.Lpreloop16
Packit 6c4009
	srlg	%r1,%r4,5	/* Split into 32byte blocks */
Packit 6c4009
Packit 6c4009
.Lloop32:
Packit 6c4009
	vst	%v16,0(%r2)
Packit 6c4009
	vst	%v16,16(%r2)
Packit 6c4009
	la	%r2,32(%r2)
Packit 6c4009
	brctg	%r1,.Lloop32	/* Loop until all blocks are processed.  */
Packit 6c4009
Packit 6c4009
	llgfr	%r4,%r4
Packit 6c4009
	nilf	%r4,31		/* Get remaining bytes */
Packit 6c4009
	je	.Lend		/* Skip store remaining bytes if zero.  */
Packit 6c4009
Packit 6c4009
.Lpreloop16:
Packit 6c4009
	clgijl	%r4,17,.Lremaining
Packit 6c4009
	srlg	%r1,%r4,4	/* Split into 16byte blocks */
Packit 6c4009
Packit 6c4009
.Lloop16:
Packit 6c4009
	vst	%v16,0(%r2)
Packit 6c4009
	la	%r2,16(%r2)
Packit 6c4009
	brctg	%r1,.Lloop16	/* Loop until all blocks are processed.  */
Packit 6c4009
Packit 6c4009
	llgfr	%r4,%r4
Packit 6c4009
	nilf	%r4,15		/* Get remaining bytes */
Packit 6c4009
	je	.Lend		/* Skip store remaining bytes if zero.  */
Packit 6c4009
Packit 6c4009
.Lremaining:
Packit 6c4009
	aghi	%r4,-1		/* vstl needs highest index.  */
Packit 6c4009
	vstl	%v16,%r4,0(%r2)
Packit 6c4009
Packit 6c4009
.Lend:
Packit 6c4009
	vlgvg	%r2,%v31,0	/* Load saved dest for return value.  */
Packit 6c4009
	br	%r14
Packit 6c4009
.Lfallback:
Packit 6c4009
	srlg	%r4,%r4,2	/* Convert byte-count to character-count.  */
Packit Service 02e39a
	jg	WMEMSET_C
Packit Service 02e39a
END(WMEMSET_Z13)
Packit Service 02e39a
Packit Service 02e39a
# if ! HAVE_WMEMSET_IFUNC
Packit Service 02e39a
strong_alias (WMEMSET_Z13, __wmemset)
Packit Service 02e39a
weak_alias (__wmemset, wmemset)
Packit Service 02e39a
# endif
Packit Service 02e39a
Packit Service 02e39a
# if defined HAVE_S390_MIN_Z13_ZARCH_ASM_SUPPORT \
Packit Service 02e39a
	&& defined SHARED && IS_IN (libc)
Packit Service 02e39a
strong_alias (WMEMSET_Z13, __GI___wmemset)
Packit Service 02e39a
weak_alias (WMEMSET_Z13, __GI_wmemset)
Packit Service 02e39a
# endif
Packit Service 02e39a
#endif