Blame sysdeps/s390/memcpy-z900.S

Packit 1e2b20
/* memcpy - copy a block from source to destination.  31/64 bit S/390 version.
Packit 6c4009
   Copyright (C) 2012-2018 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
Packit 6c4009
#include <sysdep.h>
Packit 6c4009
#include "asm-syntax.h"
Packit a65cea
#include <ifunc-memcpy.h>
Packit 6c4009
Packit 6c4009
/* INPUT PARAMETERS
Packit 6c4009
     %r2 = address of destination memory area
Packit 6c4009
     %r3 = address of source memory area
Packit 6c4009
     %r4 = number of bytes to copy.  */
Packit 6c4009
Packit 6c4009
       .text
Packit 1e2b20
Packit 1e2b20
#if defined __s390x__
Packit 1e2b20
# define LTGR	ltgr
Packit 1e2b20
# define CGHI	cghi
Packit 1e2b20
# define LGR	lgr
Packit 1e2b20
# define AGHI	aghi
Packit 1e2b20
# define BRCTG	brctg
Packit 1e2b20
#else
Packit 1e2b20
# define LTGR	ltr
Packit 1e2b20
# define CGHI	chi
Packit 1e2b20
# define LGR	lr
Packit 1e2b20
# define AGHI	ahi
Packit 1e2b20
# define BRCTG	brct
Packit 1e2b20
#endif /* ! defined __s390x__  */
Packit 1e2b20
Packit a65cea
#if HAVE_MEMCPY_Z900_G5
Packit a65cea
ENTRY(MEMPCPY_Z900_G5)
Packit a65cea
# if defined __s390x__
Packit 6c4009
	.machine "z900"
Packit a65cea
# else
Packit 1e2b20
	.machine "g5"
Packit a65cea
# endif /* ! defined __s390x__  */
Packit 1e2b20
	LGR     %r1,%r2             # Use as dest
Packit 6c4009
	la      %r2,0(%r4,%r2)      # Return dest + n
Packit 1e2b20
	j	.L_Z900_G5_start
Packit a65cea
END(MEMPCPY_Z900_G5)
Packit 6c4009
Packit a65cea
ENTRY(MEMCPY_Z900_G5)
Packit a65cea
# if defined __s390x__
Packit 6c4009
	.machine "z900"
Packit a65cea
# else
Packit 1e2b20
	.machine "g5"
Packit a65cea
# endif /* ! defined __s390x__  */
Packit 1e2b20
	LGR     %r1,%r2             # r1: Use as dest ; r2: Return dest
Packit 1e2b20
.L_Z900_G5_start:
Packit 1e2b20
	LTGR    %r4,%r4
Packit 1e2b20
	je      .L_Z900_G5_4
Packit 1e2b20
	AGHI    %r4,-1
Packit a65cea
# if defined __s390x__
Packit 1e2b20
	srlg	%r5,%r4,8
Packit a65cea
# else
Packit 1e2b20
	lr	%r5,%r4
Packit 1e2b20
	srl	%r5,8
Packit a65cea
# endif /* ! defined __s390x__  */
Packit 1e2b20
	LTGR    %r5,%r5
Packit 1e2b20
	jne     .L_Z900_G5_13
Packit 1e2b20
.L_Z900_G5_3:
Packit a65cea
# if defined __s390x__
Packit 1e2b20
	larl    %r5,.L_Z900_G5_15
Packit a65cea
#  define Z900_G5_EX_D 0
Packit a65cea
# else
Packit 1e2b20
	basr    %r5,0
Packit 1e2b20
.L_Z900_G5_14:
Packit a65cea
#  define Z900_G5_EX_D .L_Z900_G5_15-.L_Z900_G5_14
Packit a65cea
# endif /* ! defined __s390x__  */
Packit 1e2b20
	ex      %r4,Z900_G5_EX_D(%r5)
Packit 1e2b20
.L_Z900_G5_4:
Packit 6c4009
	br      %r14
Packit 1e2b20
.L_Z900_G5_13:
Packit 1e2b20
	CGHI	%r5,4096            # Switch to mvcle for copies >1MB
Packit 6c4009
	jh      __memcpy_mvcle
Packit 1e2b20
.L_Z900_G5_12:
Packit 6c4009
	mvc     0(256,%r1),0(%r3)
Packit 6c4009
	la      %r1,256(%r1)
Packit 6c4009
	la      %r3,256(%r3)
Packit 1e2b20
	BRCTG   %r5,.L_Z900_G5_12
Packit 1e2b20
	j       .L_Z900_G5_3
Packit 1e2b20
.L_Z900_G5_15:
Packit 6c4009
	mvc     0(1,%r1),0(%r3)
Packit a65cea
END(MEMCPY_Z900_G5)
Packit a65cea
#endif /* HAVE_MEMCPY_Z900_G5  */
Packit 6c4009
Packit 6c4009
ENTRY(__memcpy_mvcle)
Packit 6c4009
	# Using as standalone function will result in unexpected
Packit 6c4009
	# results since the length field is incremented by 1 in order to
Packit 6c4009
	# compensate the changes already done in the functions above.
Packit 1e2b20
	LGR     %r0,%r2             # backup return dest [ + n ]
Packit 1e2b20
	AGHI    %r4,1               # length + 1
Packit 1e2b20
	LGR     %r5,%r4             # source length
Packit 1e2b20
	LGR     %r4,%r3             # source address
Packit 1e2b20
	LGR     %r2,%r1             # destination address
Packit 1e2b20
	LGR     %r3,%r5             # destination length = source length
Packit 6c4009
.L_MVCLE_1:
Packit 6c4009
	mvcle   %r2,%r4,0           # thats it, MVCLE is your friend
Packit 6c4009
	jo      .L_MVCLE_1
Packit 1e2b20
	LGR     %r2,%r0             # return destination address
Packit 6c4009
	br      %r14
Packit 6c4009
END(__memcpy_mvcle)
Packit 1e2b20
Packit 1e2b20
#undef LTGR
Packit 1e2b20
#undef CGHI
Packit 1e2b20
#undef LGR
Packit 1e2b20
#undef AGHI
Packit 1e2b20
#undef BRCTG
Packit a65cea
Packit a65cea
#if HAVE_MEMCPY_Z10
Packit a65cea
ENTRY(MEMPCPY_Z10)
Packit a65cea
	.machine "z10"
Packit a65cea
	.machinemode "zarch_nohighgprs"
Packit a65cea
	lgr     %r1,%r2         # Use as dest
Packit a65cea
	la      %r2,0(%r4,%r2)  # Return dest + n
Packit a65cea
	j	.L_Z10_start
Packit a65cea
END(MEMPCPY_Z10)
Packit a65cea
Packit a65cea
ENTRY(MEMCPY_Z10)
Packit a65cea
	.machine "z10"
Packit a65cea
	.machinemode "zarch_nohighgprs"
Packit a65cea
	lgr     %r1,%r2         # r1: Use as dest ; r2: Return dest
Packit a65cea
.L_Z10_start:
Packit a65cea
# if !defined __s390x__
Packit a65cea
	llgfr	%r4,%r4
Packit a65cea
# endif /* !defined __s390x__  */
Packit a65cea
	cgije   %r4,0,.L_Z10_4
Packit a65cea
	aghi    %r4,-1
Packit a65cea
	srlg    %r5,%r4,8
Packit a65cea
	cgijlh  %r5,0,.L_Z10_13
Packit a65cea
.L_Z10_3:
Packit a65cea
	exrl    %r4,.L_Z10_15
Packit a65cea
.L_Z10_4:
Packit a65cea
	br      %r14
Packit a65cea
.L_Z10_13:
Packit a65cea
	cgfi    %r5,65535	# Switch to mvcle for copies >16MB
Packit a65cea
	jh      __memcpy_mvcle
Packit a65cea
.L_Z10_12:
Packit a65cea
	pfd     1,768(%r3)
Packit a65cea
	pfd     2,768(%r1)
Packit a65cea
	mvc     0(256,%r1),0(%r3)
Packit a65cea
	la      %r1,256(%r1)
Packit a65cea
	la      %r3,256(%r3)
Packit a65cea
	brctg   %r5,.L_Z10_12
Packit a65cea
	j       .L_Z10_3
Packit a65cea
.L_Z10_15:
Packit a65cea
	mvc     0(1,%r1),0(%r3)
Packit a65cea
END(MEMCPY_Z10)
Packit a65cea
#endif /* HAVE_MEMCPY_Z10  */
Packit a65cea
Packit a65cea
#if HAVE_MEMCPY_Z196
Packit a65cea
ENTRY(MEMPCPY_Z196)
Packit a65cea
	.machine "z196"
Packit a65cea
	.machinemode "zarch_nohighgprs"
Packit a65cea
	lgr     %r1,%r2         # Use as dest
Packit a65cea
	la      %r2,0(%r4,%r2)  # Return dest + n
Packit a65cea
	j	.L_Z196_start
Packit a65cea
END(MEMPCPY_Z196)
Packit a65cea
Packit a65cea
ENTRY(MEMCPY_Z196)
Packit a65cea
	.machine "z196"
Packit a65cea
	.machinemode "zarch_nohighgprs"
Packit a65cea
	lgr     %r1,%r2         # r1: Use as dest ; r2: Return dest
Packit a65cea
.L_Z196_start:
Packit a65cea
# if !defined __s390x__
Packit a65cea
	llgfr	%r4,%r4
Packit a65cea
# endif /* !defined __s390x__  */
Packit a65cea
	ltgr    %r4,%r4
Packit a65cea
	je      .L_Z196_4
Packit 61cd83
.L_Z196_start2:
Packit a65cea
	aghi    %r4,-1
Packit a65cea
	srlg    %r5,%r4,8
Packit a65cea
	ltgr    %r5,%r5
Packit a65cea
	jne     .L_Z196_5
Packit a65cea
.L_Z196_3:
Packit a65cea
	exrl    %r4,.L_Z196_14
Packit a65cea
.L_Z196_4:
Packit a65cea
	br      %r14
Packit a65cea
.L_Z196_5:
Packit a65cea
	cgfi    %r5,262144      # Switch to mvcle for copies >64MB
Packit a65cea
	jh      __memcpy_mvcle
Packit a65cea
.L_Z196_2:
Packit a65cea
	pfd     1,768(%r3)
Packit a65cea
	pfd     2,768(%r1)
Packit a65cea
	mvc     0(256,%r1),0(%r3)
Packit a65cea
	aghi    %r5,-1
Packit a65cea
	la      %r1,256(%r1)
Packit a65cea
	la      %r3,256(%r3)
Packit a65cea
	jne     .L_Z196_2
Packit a65cea
	j       .L_Z196_3
Packit a65cea
.L_Z196_14:
Packit a65cea
	mvc     0(1,%r1),0(%r3)
Packit a65cea
END(MEMCPY_Z196)
Packit a65cea
#endif /* HAVE_MEMCPY_Z196  */
Packit a65cea
Packit 61cd83
#if HAVE_MEMMOVE_Z13
Packit 61cd83
ENTRY(MEMMOVE_Z13)
Packit 61cd83
	.machine "z13"
Packit 61cd83
	.machinemode "zarch_nohighgprs"
Packit 61cd83
# if !defined __s390x__
Packit 61cd83
	/* Note: The 31bit dst and src pointers are prefixed with zeroes.  */
Packit 61cd83
	llgfr	%r4,%r4
Packit 61cd83
	llgfr	%r3,%r3
Packit 61cd83
	llgfr	%r2,%r2
Packit 61cd83
# endif /* !defined __s390x__ */
Packit 61cd83
	sgrk	%r0,%r2,%r3
Packit 61cd83
	clgijh	%r4,16,.L_MEMMOVE_Z13_LARGE
Packit 61cd83
	aghik	%r5,%r4,-1
Packit 61cd83
.L_MEMMOVE_Z13_SMALL:
Packit 61cd83
	jl .L_MEMMOVE_Z13_END		/* Jump away if len was zero.  */
Packit 61cd83
	/* Store up to 16 bytes with vll/vstl which needs the index
Packit 61cd83
	   instead of lengths.  */
Packit 61cd83
	vll	%v16,%r5,0(%r3)
Packit 61cd83
	vstl	%v16,%r5,0(%r2)
Packit 61cd83
.L_MEMMOVE_Z13_END:
Packit 61cd83
	br      %r14
Packit 61cd83
.L_MEMMOVE_Z13_LARGE:
Packit 61cd83
	lgr     %r1,%r2			/* For memcpy: r1: Use as dest ;
Packit 61cd83
					   r2: Return dest  */
Packit 61cd83
	/* The unsigned comparison (dst - src >= len) determines if we can
Packit 61cd83
	   execute the forward case with memcpy.  */
Packit 61cd83
#if ! HAVE_MEMCPY_Z196
Packit 61cd83
# error The z13 variant of memmove needs the z196 variant of memcpy!
Packit 61cd83
#endif
Packit 61cd83
	clgrjhe %r0,%r4,.L_Z196_start2
Packit 61cd83
	risbgn	%r5,%r4,4,128+63,60	/* r5 = r4 / 16  */
Packit 61cd83
	aghi	%r4,-16
Packit 61cd83
	clgijhe	%r5,8,.L_MEMMOVE_Z13_LARGE_64B
Packit 61cd83
.L_MEMMOVE_Z13_LARGE_16B_LOOP:
Packit 61cd83
	/* Store at least 16 bytes with vl/vst. The number of 16byte blocks
Packit 61cd83
	   is stored in r5.  */
Packit 61cd83
	vl	%v16,0(%r4,%r3)
Packit 61cd83
	vst	%v16,0(%r4,%r2)
Packit 61cd83
	aghi	%r4,-16
Packit 61cd83
	brctg	%r5,.L_MEMMOVE_Z13_LARGE_16B_LOOP
Packit 61cd83
	aghik	%r5,%r4,15
Packit 61cd83
	j	.L_MEMMOVE_Z13_SMALL
Packit 61cd83
.L_MEMMOVE_Z13_LARGE_64B:
Packit 61cd83
	/* Store at least 128 bytes with 4x vl/vst. The number of 64byte blocks
Packit 61cd83
	   will be stored in r0.  */
Packit 61cd83
	aghi	%r4,-48
Packit 61cd83
	srlg	%r0,%r5,2		/* r5 = %r0 / 4
Packit 61cd83
					   => Number of 64byte blocks.  */
Packit 61cd83
.L_MEMMOVE_Z13_LARGE_64B_LOOP:
Packit 61cd83
	vl	%v20,48(%r4,%r3)
Packit 61cd83
	vl	%v19,32(%r4,%r3)
Packit 61cd83
	vl	%v18,16(%r4,%r3)
Packit 61cd83
	vl	%v17,0(%r4,%r3)
Packit 61cd83
	vst	%v20,48(%r4,%r2)
Packit 61cd83
	vst	%v19,32(%r4,%r2)
Packit 61cd83
	vst	%v18,16(%r4,%r2)
Packit 61cd83
	vst	%v17,0(%r4,%r2)
Packit 61cd83
	aghi	%r4,-64
Packit 61cd83
	brctg	%r0,.L_MEMMOVE_Z13_LARGE_64B_LOOP
Packit 61cd83
	aghi	%r4,48
Packit 61cd83
	/* Recalculate the number of 16byte blocks.  */
Packit 61cd83
	risbg	%r5,%r5,62,128+63,0	/* r5 = r5 & 3
Packit 61cd83
					   => Remaining 16byte blocks.  */
Packit 61cd83
	jne	.L_MEMMOVE_Z13_LARGE_16B_LOOP
Packit 61cd83
	aghik	%r5,%r4,15
Packit 61cd83
	j	.L_MEMMOVE_Z13_SMALL
Packit 61cd83
END(MEMMOVE_Z13)
Packit 61cd83
#endif /* HAVE_MEMMOVE_Z13  */
Packit 61cd83
Packit a65cea
#if ! HAVE_MEMCPY_IFUNC
Packit a65cea
/* If we don't use ifunc, define an alias for mem[p]cpy here.
Packit a65cea
   Otherwise see sysdeps/s390/mem[p]cpy.c.  */
Packit a65cea
strong_alias (MEMCPY_DEFAULT, memcpy)
Packit a65cea
strong_alias (MEMPCPY_DEFAULT, __mempcpy)
Packit a65cea
weak_alias (__mempcpy, mempcpy)
Packit a65cea
#endif
Packit a65cea
Packit 61cd83
#if ! HAVE_MEMMOVE_IFUNC
Packit 61cd83
/* If we don't use ifunc, define an alias for memmove here.
Packit 61cd83
   Otherwise see sysdeps/s390/memmove.c.  */
Packit 61cd83
# if ! HAVE_MEMMOVE_C
Packit 61cd83
/* If the c variant is needed, then sysdeps/s390/memmove-c.c
Packit 61cd83
   defines memmove.
Packit 61cd83
   Otherwise MEMMOVE_DEFAULT is implemented here and we have to define it.  */
Packit 61cd83
strong_alias (MEMMOVE_DEFAULT, memmove)
Packit 61cd83
# endif
Packit 61cd83
#endif
Packit 61cd83
Packit a65cea
#if defined SHARED && IS_IN (libc)
Packit a65cea
/* Defines the internal symbols.
Packit a65cea
   Compare to libc_hidden_[builtin_]def (mem[p]cpy) in string/mem[p]cpy.c.  */
Packit a65cea
strong_alias (MEMCPY_DEFAULT, __GI_memcpy)
Packit a65cea
strong_alias (MEMPCPY_DEFAULT, __GI_mempcpy)
Packit a65cea
strong_alias (MEMPCPY_DEFAULT, __GI___mempcpy)
Packit 61cd83
# if ! HAVE_MEMMOVE_C
Packit 61cd83
/* If the c variant is needed, then sysdeps/s390/memmove-c.c
Packit 61cd83
   defines the internal symbol.
Packit 61cd83
   Otherwise MEMMOVE_DEFAULT is implemented here and we have to define it.  */
Packit 61cd83
strong_alias (MEMMOVE_DEFAULT, __GI_memmove)
Packit 61cd83
# endif
Packit a65cea
#endif