sysdeps/s390/multiarch/wcsncat-vx.S

   1 /* Vector optimized 32/64 bit S/390 version of wcsncat.
   2    Copyright (C) 2015-2018 Free Software Foundation, Inc.
   3    This file is part of the GNU C Library.
   4
   5    The GNU C Library is free software; you can redistribute it and/or
   6    modify it under the terms of the GNU Lesser General Public
   7    License as published by the Free Software Foundation; either
   8    version 2.1 of the License, or (at your option) any later version.
   9
  10    The GNU C Library is distributed in the hope that it will be useful,
  11    but WITHOUT ANY WARRANTY; without even the implied warranty of
  12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13    Lesser General Public License for more details.
  14
  15    You should have received a copy of the GNU Lesser General Public
  16    License along with the GNU C Library; if not, see
  17    <http://www.gnu.org/licenses/>.  */
  18
  19 #if defined HAVE_S390_VX_ASM_SUPPORT && IS_IN (libc)
  20
  21 # include "sysdep.h"
  22 # include "asm-syntax.h"
  23
  24         .text
  25
  26 /* wchar_t * wcsncat (wchar_t *dest, const wchar_t *src, size_t n)
  27    Concatenate two strings - at most n characters of src.
  28
  29    Register usage:
  30    -r0=saved dest pointer for return
  31    -r1=tmp
  32    -r2=dest
  33    -r3=src
  34    -r4=n
  35    -r5=current_len
  36    -r6=tmp
  37    -r7=tmp
  38    -v16=part of src
  39    -v17=index of zero
  40    -v18=part of src
  41    -v31=register save area for r6, r7
  42 */
  43 ENTRY(__wcsncat_vx)
  44         .machine "z13"
  45         .machinemode "zarch_nohighgprs"
  46
  47 # if !defined __s390x__
  48         llgfr   %r4,%r4
  49 # endif /* !defined __s390x__ */
  50
  51         clgfi   %r4,0
  52         ber     %r14            /* Nothing to do, if n == 0.  */
  53
  54         vlbb    %v16,0(%r2),6   /* Load s until next 4k-byte boundary.  */
  55         lcbb    %r1,0(%r2),6    /* Get bytes to 4k-byte boundary or 16.  */
  56
  57         /* If either src or dest is not 4byte aligned, use __wcsncat_c.  */
  58         tmll    %r2,3           /* Test if s is 4-byte aligned?   */
  59         jne     .Lfallback      /* And use common-code variant if not.  */
  60         tmll    %r3,3           /* Test if src is 4-byte aligned?   */
  61         jne     .Lfallback      /* And use common-code variant if not.  */
  62
  63         lgr     %r0,%r2         /* Save destination pointer for return.  */
  64         vlvgp   %v31,%r6,%r7    /* Save registers.  */
  65
  66         /* WCSLEN
  67            %r1 = loaded bytes (tmp)
  68            %r6 = zero byte index (tmp)
  69            %r2 = dst
  70         */
  71         vfenezf %v16,%v16,%v16  /* Find element not equal with zero search.  */
  72         vlgvb   %r5,%v16,7      /* Load zero index or 16 if not found.  */
  73         clrjl   %r5,%r1,.Llen_end /* Found zero within loaded bytes, end.  */
  74
  75         /* Align s to 16 byte.  */
  76         risbgn  %r1,%r2,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15.  */
  77         lghi    %r5,16          /* current_len = 16.  */
  78         slr     %r5,%r1         /* Compute bytes to 16bytes boundary.  */
  79
  80         /* Find zero in 16byte aligned loop.  */
  81 .Llen_loop:
  82         vl      %v16,0(%r5,%r2) /* Load s.  */
  83         vfenezfs %v16,%v16,%v16 /* Find element not equal with zero search.  */
  84         je      .Llen_found     /* Jump away if zero was found.  */
  85         vl      %v16,16(%r5,%r2)
  86         vfenezfs %v16,%v16,%v16
  87         je      .Llen_found16
  88         vl      %v16,32(%r5,%r2)
  89         vfenezfs %v16,%v16,%v16
  90         je      .Llen_found32
  91         vl      %v16,48(%r5,%r2)
  92         vfenezfs %v16,%v16,%v16
  93         je      .Llen_found48
  94
  95         aghi    %r5,64
  96         j       .Llen_loop      /* No zero -> loop.  */
  97
  98 .Llen_found48:
  99         aghi    %r5,16
 100 .Llen_found32:
 101         aghi    %r5,16
 102 .Llen_found16:
 103         aghi    %r5,16
 104 .Llen_found:
 105         vlgvb   %r1,%v16,7      /* Load byte index of zero.  */
 106         algr    %r5,%r1
 107
 108 .Llen_end:
 109         /* WCSNCPY
 110            %r1 = zero byte index (tmp)
 111            %r6 = loaded bytes (tmp)
 112            %r3 = curr src pointer
 113            %r2 = curr dst pointer
 114            %r7 = border, tmp
 115         */
 116         la      %r2,0(%r5,%r2)  /* strcpy at end of dst-string.  */
 117
 118         vlbb    %v16,0(%r3),6   /* Load s until next 4k-byte boundary.  */
 119         lcbb    %r6,0(%r3),6    /* Get bytes to 4k-byte boundary or 16.  */
 120         llgfr   %r6,%r6         /* Convert 32bit to 64bit.  */
 121
 122         lghi    %r5,0           /* current_len = 0.  */
 123
 124         /* Check range of maxlen and convert to byte-count.  */
 125 # ifdef __s390x__
 126         tmhh    %r4,49152       /* Test bit 0 or 1 of maxlen.  */
 127         lghi    %r1,-4          /* Max byte-count is 18446744073709551612.  */
 128 # else
 129         tmlh    %r4,49152       /* Test bit 0 or 1 of maxlen.  */
 130         llilf   %r1,4294967292  /* Max byte-count is 4294967292.  */
 131 # endif /* !__s390x__ */
 132         sllg    %r4,%r4,2       /* Convert character-count to byte-count.  */
 133         locgrne %r4,%r1         /* Use max byte-count, if bit 0/1 was one.  */
 134
 135         clgrjle %r4,%r6,.Lcpy_remaining_v16 /* If n <= loaded-bytes
 136                                                -> process remaining.  */
 137
 138         /* n > loaded-byte-count.  */
 139         vfenezf %v17,%v16,%v16  /* Find element not equal with zero search.  */
 140         vlgvb   %r1,%v17,7      /* Load zero index or 16 if not found.  */
 141         clrjl   %r1,%r6,.Lcpy_found_v16_store /* Found zero within loaded bytes,
 142                                                  copy and return.  */
 143
 144         /* Align s to 16 byte.  */
 145         risbgn  %r1,%r3,60,128+63,0 /* %r3 = bits 60-63 of %r2 'and' 15.  */
 146         lghi    %r5,15          /* current_len = 15.  */
 147         slr     %r5,%r1         /* Compute highest index to 16byte boundary.  *
 148
 149         /* Zero not found and maxlen > loaded-byte-count.  */
 150         vstl    %v16,%r5,0(%r2) /* Copy loaded characters - no zero.  */
 151         ahi     %r5,1           /* Start loop at next character.  */
 152
 153         /*
 154           Now we are 16byte aligned, so we can load a full vreg
 155           without page fault.
 156          */
 157         lgr     %r1,%r5         /* If %r5 + 64 < maxlen? -> loop64.  */
 158         aghi    %r1,64
 159         clgrjl  %r1,%r4,.Lcpy_loop64
 160
 161         vl      %v16,0(%r5,%r3) /* Load s.  */
 162         clgijl  %r4,17,.Lcpy_remaining_v16 /* If n <=16,
 163                                                process remaining bytes.  */
 164 .Lcpy_lt64:
 165         lgr     %r7,%r4
 166         slgfi   %r7,16          /* border_len = n - 16.  */
 167
 168         clgrjhe %r5,%r7,.Lcpy_remaining_v16
 169         vfenezfs %v17,%v16,%v16 /* Find element not equal with zero search.  */
 170         je      .Lcpy_found_v16 /* Jump away if zero was found.  */
 171         vl      %v18,16(%r5,%r3) /* Load next part of s.  */
 172         vst     %v16,0(%r5,%r2) /* Save previous part without zero to dst.  */
 173         aghi    %r5,16
 174
 175         clgrjhe %r5,%r7,.Lcpy_remaining_v18
 176         vfenezfs %v17,%v18,%v18
 177         je      .Lcpy_found_v18
 178         vl      %v16,16(%r5,%r3)
 179         vst     %v18,0(%r5,%r2)
 180         aghi    %r5,16
 181
 182         clgrjhe %r5,%r7,.Lcpy_remaining_v16
 183         vfenezfs %v17,%v16,%v16
 184         je      .Lcpy_found_v16
 185         vl      %v18,16(%r5,%r3)
 186         vst     %v16,0(%r5,%r2)
 187         aghi    %r5,16
 188
 189 .Lcpy_remaining_v18:
 190         vlr     %v16,%v18
 191 .Lcpy_remaining_v16:
 192         /* v16 contains the remaining bytes [1...16].
 193            Store remaining bytes and append string-termination.  */
 194         vfenezf %v17,%v16,%v16  /* Find element not equal with zero search.  */
 195         slgrk   %r7,%r4,%r5     /* Remaining bytes = maxlen - current_len.  */
 196         aghi    %r7,-1          /* vstl needs highest index.  */
 197         vlgvb   %r1,%v17,7      /* Load zero index or 16 if not found.  */
 198         la      %r2,0(%r5,%r2)  /* vstl has no index register.  */
 199         /* Zero-index within remaining-bytes, store up to zero and end.  */
 200         clgrjle %r1,%r7,.Lcpy_found_v16_store
 201         vstl    %v16,%r7,0(%r2) /* Store remaining bytes.  */
 202         lghi    %r1,0
 203         st      %r1,1(%r7,%r2)  /* Store string-null-termination beyond n.  */
 204 .Lcpy_end:
 205         /* Restore saved registers.  */
 206         vlgvg   %r6,%v31,0
 207         vlgvg   %r7,%v31,1
 208         lgr     %r2,%r0         /* Load saved dest-ptr.  */
 209         br      %r14
 210
 211 .Lcpy_found_v16_32:
 212         aghi    %r5,32
 213         j       .Lcpy_found_v16
 214 .Lcpy_found_v18_48:
 215         aghi    %r5,32
 216 .Lcpy_found_v18_16:
 217         aghi    %r5,16
 218 .Lcpy_found_v18:
 219         vlr     %v16,%v18
 220 .Lcpy_found_v16:
 221         /* v16 contains a zero. Store remaining bytes to zero. current_len
 222            has not reached border, thus checking for n is not needed!  */
 223         vlgvb   %r1,%v17,7      /* Load byte index of zero.  */
 224         la      %r2,0(%r5,%r2)
 225 .Lcpy_found_v16_store:
 226         aghi    %r1,3           /* Also copy remaining bytes of zero.  */
 227         vstl    %v16,%r1,0(%r2) /* Copy characters including zero.  */
 228         j       .Lcpy_end
 229
 230         /* Find zero in 16byte aligned loop.  */
 231 .Lcpy_loop2:
 232         vl      %v16,16(%r5,%r3)
 233         vst     %v18,0(%r5,%r2)
 234         aghi    %r5,16
 235
 236 .Lcpy_loop64:
 237         vl      %v16,0(%r5,%r3)
 238         vfenezfs %v17,%v16,%v16 /* Find element not equal with zero search.  */
 239         je      .Lcpy_found_v16 /* Jump away if zero was found.  */
 240         vl      %v18,16(%r5,%r3) /* Load next part of s.  */
 241         vst     %v16,0(%r5,%r2) /* Save previous part without zero to dst.  */
 242         vfenezfs %v17,%v18,%v18
 243         je      .Lcpy_found_v18_16
 244         vl      %v16,32(%r5,%r3)
 245         vst     %v18,16(%r5,%r2)
 246         vfenezfs %v17,%v16,%v16
 247         je      .Lcpy_found_v16_32
 248         vl      %v18,48(%r5,%r3)
 249         vst     %v16,32(%r5,%r2)
 250         vfenezfs %v17,%v18,%v18
 251         je      .Lcpy_found_v18_48
 252         vst     %v18,48(%r5,%r2)
 253
 254         aghi    %r5,64
 255         lgr     %r1,%r5         /* If %r5 + 64 < maxlen? -> loop64.  */
 256         aghi    %r1,64
 257         clgrjl  %r1,%r4,.Lcpy_loop64
 258
 259         vl      %v16,0(%r5,%r3) /* Load s.  */
 260         j       .Lcpy_lt64
 261
 262 .Lfallback:
 263         jg      __wcsncat_c
 264 END(__wcsncat_vx)
 265 #endif /* HAVE_S390_VX_ASM_SUPPORT && IS_IN (libc) */