Blame sysdeps/powerpc/powerpc64/power7/strcmp.S

Packit 6c4009
/* Optimized strcmp implementation for Power7 using 'cmpb' instruction
Packit 6c4009
   Copyright (C) 2014-2018 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 6c4009
/* The optimization is achieved here through cmpb instruction.
Packit 6c4009
   8byte aligned strings are processed with double word comparision
Packit 6c4009
   and unaligned strings are handled effectively with loop unrolling
Packit 6c4009
   technique  */
Packit 6c4009
Packit 6c4009
#include <sysdep.h>
Packit 6c4009
Packit 6c4009
#ifndef STRCMP
Packit 6c4009
# define STRCMP strcmp
Packit 6c4009
#endif
Packit 6c4009
Packit 6c4009
/* int [r3] strcmp (const char *s1 [r3], const char *s2 [r4])  */
Packit 6c4009
Packit 6c4009
	.machine	power7
Packit 6c4009
ENTRY_TOCLESS (STRCMP, 4)
Packit 6c4009
	CALL_MCOUNT 2
Packit 6c4009
Packit 6c4009
	or r9, r3, r4
Packit 6c4009
	rldicl. r10, r9, 0, 61	/* are s1 and s2 8 byte aligned..?  */
Packit 6c4009
	bne cr0, L(process_unaligned_bytes)
Packit 6c4009
	li	r5, 0
Packit 6c4009
Packit 6c4009
	.align 4
Packit 6c4009
/* process input parameters on double word aligned boundary  */
Packit 6c4009
L(unrollDword):
Packit 6c4009
	ld	r8,0(r3)
Packit 6c4009
	ld	r10,0(r4)
Packit 6c4009
	cmpb	r7,r8,r5
Packit 6c4009
	cmpdi	cr7,r7,0
Packit 6c4009
	mr	r9,r7
Packit 6c4009
	bne 	cr7,L(null_found)
Packit 6c4009
	cmpld	cr7,r8,r10
Packit 6c4009
	bne	cr7,L(different)
Packit 6c4009
Packit 6c4009
	ld	r8,8(r3)
Packit 6c4009
	ld	r10,8(r4)
Packit 6c4009
	cmpb	r7,r8,r5
Packit 6c4009
	cmpdi	cr7,r7,0
Packit 6c4009
	mr	r9,r7
Packit 6c4009
	bne 	cr7,L(null_found)
Packit 6c4009
	cmpld	cr7,r8,r10
Packit 6c4009
	bne	cr7,L(different)
Packit 6c4009
Packit 6c4009
	ld	r8,16(r3)
Packit 6c4009
	ld	r10,16(r4)
Packit 6c4009
	cmpb	r7,r8,r5
Packit 6c4009
	cmpdi	cr7,r7,0
Packit 6c4009
	mr	r9,r7
Packit 6c4009
	bne 	cr7,L(null_found)
Packit 6c4009
	cmpld	cr7,r8,r10
Packit 6c4009
	bne	cr7,L(different)
Packit 6c4009
Packit 6c4009
	ld	r8,24(r3)
Packit 6c4009
	ld	r10,24(r4)
Packit 6c4009
	cmpb	r7,r8,r5
Packit 6c4009
	cmpdi	cr7,r7,0
Packit 6c4009
	mr	r9,r7
Packit 6c4009
	bne 	cr7,L(null_found)
Packit 6c4009
	cmpld	cr7,r8,r10
Packit 6c4009
	bne	cr7,L(different)
Packit 6c4009
Packit 6c4009
	addi r3, r3, 32
Packit 6c4009
	addi r4, r4, 32
Packit 6c4009
	beq cr7, L(unrollDword)
Packit 6c4009
Packit 6c4009
	.align 4
Packit 6c4009
L(null_found):
Packit 6c4009
#ifdef __LITTLE_ENDIAN__
Packit 6c4009
	neg	r7,r9
Packit 6c4009
	and	r9,r9,r7
Packit 6c4009
	li	r7,-1
Packit 6c4009
	cntlzd	r9,r9
Packit 6c4009
	subfic	r9,r9,71
Packit 6c4009
	sld	r9,r7,r9
Packit 6c4009
#else
Packit 6c4009
	cntlzd	r9,r9
Packit 6c4009
	li	r7,-1
Packit 6c4009
	addi	r9,r9,8
Packit 6c4009
	srd	r9,r7,r9
Packit 6c4009
#endif
Packit 6c4009
	or	r8,r8,r9
Packit 6c4009
	or	r10,r10,r9
Packit 6c4009
Packit 6c4009
L(different):
Packit 6c4009
	cmpb	r9,r8,r10
Packit 6c4009
#ifdef __LITTLE_ENDIAN__
Packit 6c4009
	addi	r7,r9,1
Packit 6c4009
	andc	r9,r7,r9
Packit 6c4009
	cntlzd	r9,r9
Packit 6c4009
	subfic	r9,r9,63
Packit 6c4009
#else
Packit 6c4009
	not	r9,r9
Packit 6c4009
	cntlzd	r9,r9
Packit 6c4009
	subfic	r9,r9,56
Packit 6c4009
#endif
Packit 6c4009
	srd	r3,r8,r9
Packit 6c4009
	srd	r10,r10,r9
Packit 6c4009
	rldicl	r10,r10,0,56
Packit 6c4009
	rldicl	r3,r3,0,56
Packit 6c4009
	subf	r3,r10,r3
Packit 6c4009
	blr
Packit 6c4009
Packit 6c4009
	.align 4
Packit 6c4009
L(process_unaligned_bytes):
Packit 6c4009
	lbz r9, 0(r3)		/* load byte from s1  */
Packit 6c4009
	lbz r10, 0(r4)		/* load byte from s2  */
Packit 6c4009
	cmpdi cr7, r9, 0	/* compare *s1 with NULL  */
Packit 6c4009
	beq cr7, L(diffOfNULL)	/* if *s1 is NULL , return *s1 - *s2  */
Packit 6c4009
	cmplw cr7, r9, r10	/* compare *s1 and *s2  */
Packit 6c4009
	bne cr7, L(ComputeDiff)	/* branch to compute difference and return  */
Packit 6c4009
Packit 6c4009
	lbz r9, 1(r3)		/* load next byte from s1  */
Packit 6c4009
	lbz r10, 1(r4)		/* load next byte from s2  */
Packit 6c4009
	cmpdi cr7, r9, 0	/* compare *s1 with NULL  */
Packit 6c4009
	beq cr7, L(diffOfNULL)	/* if *s1 is NULL , return *s1 - *s2  */
Packit 6c4009
	cmplw cr7, r9, r10	/* compare *s1 and *s2  */
Packit 6c4009
	bne cr7, L(ComputeDiff)	/* branch to compute difference and return  */
Packit 6c4009
Packit 6c4009
	lbz r9, 2(r3)		/* unroll 3rd byte here  */
Packit 6c4009
	lbz r10, 2(r4)
Packit 6c4009
	cmpdi cr7, r9, 0
Packit 6c4009
	beq cr7, L(diffOfNULL)
Packit 6c4009
	cmplw cr7, r9, r10
Packit 6c4009
	bne 7, L(ComputeDiff)
Packit 6c4009
Packit 6c4009
	lbz r9, 3(r3)		/* unroll 4th byte now  */
Packit 6c4009
	lbz r10, 3(r4)
Packit 6c4009
	addi r3, r3, 4		/* increment s1 by unroll factor  */
Packit 6c4009
	cmpdi cr7, r9, 0
Packit 6c4009
	cmplw cr6, 9, r10
Packit 6c4009
	beq cr7, L(diffOfNULL)
Packit 6c4009
	addi r4, r4, 4		/* increment s2 by unroll factor  */
Packit 6c4009
	beq cr6, L(process_unaligned_bytes)	/* unroll byte processing  */
Packit 6c4009
Packit 6c4009
	.align 4
Packit 6c4009
L(ComputeDiff):
Packit 6c4009
	extsw r9, r9
Packit 6c4009
	subf r10, r10, r9	/* compute s1 - s2  */
Packit 6c4009
	extsw r3, r10
Packit 6c4009
	blr			/* return  */
Packit 6c4009
Packit 6c4009
	.align 4
Packit 6c4009
L(diffOfNULL):
Packit 6c4009
	li r9, 0
Packit 6c4009
	subf r10, r10, r9	/* compute s1 - s2  */
Packit 6c4009
	extsw r3, r10		/* sign extend result  */
Packit 6c4009
	blr			/* return  */
Packit 6c4009
Packit 6c4009
END (STRCMP)
Packit 6c4009
libc_hidden_builtin_def (strcmp)