Blame sysdeps/powerpc/powerpc32/476/memset.S

Packit 6c4009
/* Optimized memset for PowerPC476 (128-byte cacheline).
Packit 6c4009
   Copyright (C) 2010-2018 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
#include <sysdep.h>
Packit 6c4009
Packit 6c4009
/* memset
Packit 6c4009
Packit 6c4009
       r3:destination address and return address
Packit 6c4009
       r4:source integer to copy
Packit 6c4009
       r5:byte count
Packit 6c4009
       r11:sources integer to copy in all 32 bits of reg
Packit 6c4009
       r12:temp return address
Packit 6c4009
Packit 6c4009
       Save return address in r12
Packit 6c4009
       If destinationn is unaligned and count is greater tha 255 bytes
Packit 6c4009
       set 0-3 bytes to make destination aligned
Packit 6c4009
       If count is greater tha 255 bytes and setting zero to memory
Packit 6c4009
       use dbcz to set memeory when we can
Packit 6c4009
       otherwsie do the follwoing
Packit 6c4009
       If 16 or more words to set we use 16 word copy loop.
Packit 6c4009
       Finaly we set 0-15 extra bytes with string store. */
Packit 6c4009
Packit 6c4009
EALIGN (memset, 5, 0)
Packit 6c4009
       rlwinm  r11,r4,0,24,31
Packit 6c4009
       rlwimi  r11,r4,8,16,23
Packit 6c4009
       rlwimi  r11,r11,16,0,15
Packit 6c4009
       addi    r12,r3,0
Packit 6c4009
       cmpwi   r5,0x00FF
Packit 6c4009
       ble     L(preword8_count_loop)
Packit 6c4009
       cmpwi   r4,0x00
Packit 6c4009
       beq     L(use_dcbz)
Packit 6c4009
       neg     r6,r3
Packit 6c4009
       clrlwi. r6,r6,30
Packit 6c4009
       beq     L(preword8_count_loop)
Packit 6c4009
       addi    r8,0,1
Packit 6c4009
       mtctr   r6
Packit 6c4009
       subi    r3,r3,1
Packit 6c4009
Packit 6c4009
L(unaligned_bytecopy_loop):
Packit 6c4009
       stbu    r11,0x1(r3)
Packit 6c4009
       subf.   r5,r8,r5
Packit 6c4009
       beq     L(end_memset)
Packit 6c4009
       bdnz    L(unaligned_bytecopy_loop)
Packit 6c4009
       addi    r3,r3,1
Packit 6c4009
Packit 6c4009
L(preword8_count_loop):
Packit 6c4009
       srwi.   r6,r5,4
Packit 6c4009
       beq     L(preword2_count_loop)
Packit 6c4009
       mtctr   r6
Packit 6c4009
       addi    r3,r3,-4
Packit 6c4009
       mr      r8,r11
Packit 6c4009
       mr      r9,r11
Packit 6c4009
       mr      r10,r11
Packit 6c4009
Packit 6c4009
L(word8_count_loop_no_dcbt):
Packit 6c4009
       stwu    r8,4(r3)
Packit 6c4009
       stwu    r9,4(r3)
Packit 6c4009
       subi    r5,r5,0x10
Packit 6c4009
       stwu    r10,4(r3)
Packit 6c4009
       stwu    r11,4(r3)
Packit 6c4009
       bdnz    L(word8_count_loop_no_dcbt)
Packit 6c4009
       addi    r3,r3,4
Packit 6c4009
Packit 6c4009
L(preword2_count_loop):
Packit 6c4009
       clrlwi. r7,r5,28
Packit 6c4009
       beq     L(end_memset)
Packit 6c4009
       mr      r8,r11
Packit 6c4009
       mr      r9,r11
Packit 6c4009
       mr      r10,r11
Packit 6c4009
       mtxer   r7
Packit 6c4009
       stswx   r8,0,r3
Packit 6c4009
Packit 6c4009
L(end_memset):
Packit 6c4009
       addi    r3,r12,0
Packit 6c4009
       blr
Packit 6c4009
Packit 6c4009
L(use_dcbz):
Packit 6c4009
       neg     r6,r3
Packit 6c4009
       clrlwi. r7,r6,28
Packit 6c4009
       beq     L(skip_string_loop)
Packit 6c4009
       mr      r8,r11
Packit 6c4009
       mr      r9,r11
Packit 6c4009
       mr      r10,r11
Packit 6c4009
       subf    r5,r7,r5
Packit 6c4009
       mtxer   r7
Packit 6c4009
       stswx   r8,0,r3
Packit 6c4009
       add     r3,r3,r7
Packit 6c4009
Packit 6c4009
L(skip_string_loop):
Packit 6c4009
       clrlwi  r8,r6,25
Packit 6c4009
       srwi.   r8,r8,4
Packit 6c4009
       beq     L(dcbz_pre_loop)
Packit 6c4009
       mtctr   r8
Packit 6c4009
Packit 6c4009
L(word_loop):
Packit 6c4009
       stw     r11,0(r3)
Packit 6c4009
       subi    r5,r5,0x10
Packit 6c4009
       stw     r11,4(r3)
Packit 6c4009
       stw     r11,8(r3)
Packit 6c4009
       stw     r11,12(r3)
Packit 6c4009
       addi    r3,r3,0x10
Packit 6c4009
       bdnz    L(word_loop)
Packit 6c4009
Packit 6c4009
L(dcbz_pre_loop):
Packit 6c4009
       srwi    r6,r5,7
Packit 6c4009
       mtctr   r6
Packit 6c4009
       addi    r7,0,0
Packit 6c4009
Packit 6c4009
L(dcbz_loop):
Packit 6c4009
       dcbz    r3,r7
Packit 6c4009
       addi    r3,r3,0x80
Packit 6c4009
       subi    r5,r5,0x80
Packit 6c4009
       bdnz    L(dcbz_loop)
Packit 6c4009
       srwi.   r6,r5,4
Packit 6c4009
       beq     L(postword2_count_loop)
Packit 6c4009
       mtctr   r6
Packit 6c4009
Packit 6c4009
L(postword8_count_loop):
Packit 6c4009
       stw     r11,0(r3)
Packit 6c4009
       subi    r5,r5,0x10
Packit 6c4009
       stw     r11,4(r3)
Packit 6c4009
       stw     r11,8(r3)
Packit 6c4009
       stw     r11,12(r3)
Packit 6c4009
       addi    r3,r3,0x10
Packit 6c4009
       bdnz    L(postword8_count_loop)
Packit 6c4009
Packit 6c4009
L(postword2_count_loop):
Packit 6c4009
       clrlwi. r7,r5,28
Packit 6c4009
       beq     L(end_memset)
Packit 6c4009
       mr      r8,r11
Packit 6c4009
       mr      r9,r11
Packit 6c4009
       mr      r10,r11
Packit 6c4009
       mtxer   r7
Packit 6c4009
       stswx   r8,0,r3
Packit 6c4009
       b       L(end_memset)
Packit 6c4009
END (memset)
Packit 6c4009
libc_hidden_builtin_def (memset)