root/arch/sh/lib64/memset.S

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 /* Cloned and hacked for uClibc by Paul Mundt, December 2003 */
   3 /* Modified by SuperH, Inc. September 2003 */
   4 !
   5 ! Fast SH memset
   6 !
   7 ! by Toshiyasu Morita (tm@netcom.com)
   8 !
   9 ! SH5 code by J"orn Rennecke (joern.rennecke@superh.com)
  10 ! Copyright 2002 SuperH Ltd.
  11 !
  12 
  13 #if __BYTE_ORDER == __LITTLE_ENDIAN
  14 #define SHHI shlld
  15 #define SHLO shlrd
  16 #else
  17 #define SHHI shlrd
  18 #define SHLO shlld
  19 #endif
  20 
  21         .section .text..SHmedia32,"ax"
  22         .globl  memset
  23         .type   memset, @function
  24 
  25         .align 5
  26 
  27 memset:
  28         pta/l multiquad, tr0
  29         andi r2, 7, r22
  30         ptabs r18, tr2
  31         mshflo.b r3,r3,r3
  32         add r4, r22, r23
  33         mperm.w r3, r63, r3     // Fill pattern now in every byte of r3
  34 
  35         movi 8, r9
  36         bgtu/u r23, r9, tr0 // multiquad
  37 
  38         beqi/u r4, 0, tr2       // Return with size 0 - ensures no mem accesses
  39         ldlo.q r2, 0, r7
  40         shlli r4, 2, r4
  41         movi -1, r8
  42         SHHI r8, r4, r8
  43         SHHI r8, r4, r8
  44         mcmv r7, r8, r3
  45         stlo.q r2, 0, r3
  46         blink tr2, r63
  47 
  48 multiquad:
  49         pta/l lastquad, tr0
  50         stlo.q r2, 0, r3
  51         shlri r23, 3, r24
  52         add r2, r4, r5
  53         beqi/u r24, 1, tr0 // lastquad
  54         pta/l loop, tr1
  55         sub r2, r22, r25
  56         andi r5, -8, r20   // calculate end address and
  57         addi r20, -7*8, r8 // loop end address; This might overflow, so we need
  58                            // to use a different test before we start the loop
  59         bge/u r24, r9, tr1 // loop
  60         st.q r25, 8, r3
  61         st.q r20, -8, r3
  62         shlri r24, 1, r24
  63         beqi/u r24, 1, tr0 // lastquad
  64         st.q r25, 16, r3
  65         st.q r20, -16, r3
  66         beqi/u r24, 2, tr0 // lastquad
  67         st.q r25, 24, r3
  68         st.q r20, -24, r3
  69 lastquad:
  70         sthi.q r5, -1, r3
  71         blink tr2,r63
  72 
  73 loop:
  74 !!!     alloco r25, 32  // QQQ comment out for short-term fix to SHUK #3895.
  75                         // QQQ commenting out is locically correct, but sub-optimal
  76                         // QQQ Sean McGoogan - 4th April 2003.
  77         st.q r25, 8, r3
  78         st.q r25, 16, r3
  79         st.q r25, 24, r3
  80         st.q r25, 32, r3
  81         addi r25, 32, r25
  82         bgeu/l r8, r25, tr1 // loop
  83 
  84         st.q r20, -40, r3
  85         st.q r20, -32, r3
  86         st.q r20, -24, r3
  87         st.q r20, -16, r3
  88         st.q r20, -8, r3
  89         sthi.q r5, -1, r3
  90         blink tr2,r63
  91 
  92         .size   memset,.-memset