alpha/lib/memchr.S

0001 /* Copyright (C) 1996 Free Software Foundation, Inc.
0002    This file is part of the GNU C Library.
0003    Contributed by David Mosberger (davidm@cs.arizona.edu).
0004
0005    The GNU C Library is free software; you can redistribute it and/or
0006    modify it under the terms of the GNU Library General Public License as
0007    published by the Free Software Foundation; either version 2 of the
0008    License, or (at your option) any later version.
0009
0010    The GNU C Library is distributed in the hope that it will be useful,
0011    but WITHOUT ANY WARRANTY; without even the implied warranty of
0012    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
0013    Library General Public License for more details.
0014
0015    You should have received a copy of the GNU Library General Public
0016    License along with the GNU C Library; see the file COPYING.LIB.  If not,
0017    write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
0018    Boston, MA 02111-1307, USA.  */
0019
0020 /* Finds characters in a memory area.  Optimized for the Alpha:
0021
0022       - memory accessed as aligned quadwords only
0023       - uses cmpbge to compare 8 bytes in parallel
0024       - does binary search to find 0 byte in last
0025         quadword (HAKMEM needed 12 instructions to
0026         do this instead of the 9 instructions that
0027         binary search needs).
0028
0029 For correctness consider that:
0030
0031       - only minimum number of quadwords may be accessed
0032       - the third argument is an unsigned long
0033 */
0034 #include <asm/export.h>
0035         .set noreorder
0036         .set noat
0037
0038     .globl memchr
0039     .ent memchr
0040 memchr:
0041     .frame $30,0,$26,0
0042     .prologue 0
0043
0044     # Hack -- if someone passes in (size_t)-1, hoping to just
0045     # search til the end of the address space, we will overflow
0046     # below when we find the address of the last byte.  Given
0047     # that we will never have a 56-bit address space, cropping
0048     # the length is the easiest way to avoid trouble.
0049     zap $18, 0x80, $5   #-e0    :
0050
0051     beq $18, $not_found # .. e1 :
0052         ldq_u   $1, 0($16)  # e1    : load first quadword
0053     insbl   $17, 1, $2  # .. e0 : $2 = 000000000000ch00
0054     and $17, 0xff, $17  #-e0    : $17 = 00000000000000ch
0055     cmpult  $18, 9, $4  # .. e1 :
0056     or  $2, $17, $17    # e0    : $17 = 000000000000chch
0057         lda     $3, -1($31) # .. e1 :
0058     sll $17, 16, $2 #-e0    : $2 = 00000000chch0000
0059     addq    $16, $5, $5 # .. e1 :
0060     or  $2, $17, $17    # e1    : $17 = 00000000chchchch
0061     unop            #   :
0062     sll $17, 32, $2 #-e0    : $2 = chchchch00000000
0063     or  $2, $17, $17    # e1    : $17 = chchchchchchchch
0064     extql   $1, $16, $7 # e0    :
0065     beq $4, $first_quad # .. e1 :
0066
0067     ldq_u   $6, -1($5)  #-e1    : eight or less bytes to search
0068     extqh   $6, $16, $6 # .. e0 :
0069     mov $16, $0     # e0    :
0070     or  $7, $6, $1  # .. e1 : $1 = quadword starting at $16
0071
0072     # Deal with the case where at most 8 bytes remain to be searched
0073     # in $1.  E.g.:
0074     #   $18 = 6
0075     #   $1 = ????c6c5c4c3c2c1
0076 $last_quad:
0077     negq    $18, $6     #-e0    :
0078         xor $17, $1, $1 # .. e1 :
0079     srl $3, $6, $6  # e0    : $6 = mask of $18 bits set
0080         cmpbge  $31, $1, $2 # .. e1 :
0081     and $2, $6, $2  #-e0    :
0082         beq     $2, $not_found  # .. e1 :
0083
0084 $found_it:
0085     # Now, determine which byte matched:
0086         negq    $2, $3      # e0    :
0087         and     $2, $3, $2  # e1    :
0088
0089         and     $2, 0x0f, $1    #-e0    :
0090         addq    $0, 4, $3   # .. e1 :
0091         cmoveq  $1, $3, $0  # e0    :
0092
0093         addq    $0, 2, $3   # .. e1 :
0094         and     $2, 0x33, $1    #-e0    :
0095         cmoveq  $1, $3, $0  # .. e1 :
0096
0097         and     $2, 0x55, $1    # e0    :
0098         addq    $0, 1, $3   # .. e1 :
0099         cmoveq  $1, $3, $0  #-e0    :
0100
0101 $done:  ret         # .. e1 :
0102
0103     # Deal with the case where $18 > 8 bytes remain to be
0104     # searched.  $16 may not be aligned.
0105     .align 4
0106 $first_quad:
0107     andnot  $16, 0x7, $0    #-e1    :
0108         insqh   $3, $16, $2 # .. e0 : $2 = 0000ffffffffffff ($16<0:2> ff)
0109         xor $1, $17, $1 # e0    :
0110     or  $1, $2, $1  # e1    : $1 = ====ffffffffffff
0111         cmpbge  $31, $1, $2 #-e0    :
0112         bne     $2, $found_it   # .. e1 :
0113
0114     # At least one byte left to process.
0115
0116     ldq $1, 8($0)   # e0    :
0117     subq    $5, 1, $18  # .. e1 :
0118     addq    $0, 8, $0   #-e0    :
0119
0120     # Make $18 point to last quad to be accessed (the
0121     # last quad may or may not be partial).
0122
0123     andnot  $18, 0x7, $18   # .. e1 :
0124     cmpult  $0, $18, $2 # e0    :
0125     beq $2, $final  # .. e1 :
0126
0127     # At least two quads remain to be accessed.
0128
0129     subq    $18, $0, $4 #-e0    : $4 <- nr quads to be processed
0130     and $4, 8, $4   # e1    : odd number of quads?
0131     bne $4, $odd_quad_count # e1 :
0132
0133     # At least three quads remain to be accessed
0134
0135     mov $1, $4      # e0    : move prefetched value to correct reg
0136
0137     .align  4
0138 $unrolled_loop:
0139     ldq $1, 8($0)   #-e0    : prefetch $1
0140     xor $17, $4, $2 # .. e1 :
0141     cmpbge  $31, $2, $2 # e0    :
0142     bne $2, $found_it   # .. e1 :
0143
0144     addq    $0, 8, $0   #-e0    :
0145 $odd_quad_count:
0146     xor $17, $1, $2 # .. e1 :
0147     ldq $4, 8($0)   # e0    : prefetch $4
0148     cmpbge  $31, $2, $2 # .. e1 :
0149     addq    $0, 8, $6   #-e0    :
0150     bne $2, $found_it   # .. e1 :
0151
0152     cmpult  $6, $18, $6 # e0    :
0153     addq    $0, 8, $0   # .. e1 :
0154     bne $6, $unrolled_loop #-e1 :
0155
0156     mov $4, $1      # e0    : move prefetched value into $1
0157 $final: subq    $5, $0, $18 # .. e1 : $18 <- number of bytes left to do
0158     bne $18, $last_quad # e1    :
0159
0160 $not_found:
0161     mov $31, $0     #-e0    :
0162     ret         # .. e1 :
0163
0164         .end memchr
0165     EXPORT_SYMBOL(memchr)