1/* A memset for CRIS.
2   Copyright (C) 1999-2005 Axis Communications.
3   All rights reserved.
4
5   Redistribution and use in source and binary forms, with or without
6   modification, are permitted provided that the following conditions
7   are met:
8
9   1. Redistributions of source code must retain the above copyright
10      notice, this list of conditions and the following disclaimer.
11
12   2. Neither the name of Axis Communications nor the names of its
13      contributors may be used to endorse or promote products derived
14      from this software without specific prior written permission.
15
16   THIS SOFTWARE IS PROVIDED BY AXIS COMMUNICATIONS AND ITS CONTRIBUTORS
17   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL AXIS
20   COMMUNICATIONS OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
21   INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
22   (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
23   SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24   HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
25   STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
26   IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27   POSSIBILITY OF SUCH DAMAGE.  */
28
29/* FIXME: This file should really only be used for reference, as the
30   result is somewhat depending on gcc generating what we expect rather
31   than what we describe.  An assembly file should be used instead.  */
32
33/* Note the multiple occurrence of the expression "12*4", including the
34   asm.  It is hard to get it into the asm in a good way.  Thus better to
35   expose the problem everywhere: no macro.  */
36
37/* Assuming one cycle per dword written or read (ok, not really true; the
38   world is not ideal), and one cycle per instruction, then 43+3*(n/48-1)
39   <= 24+24*(n/48-1) so n >= 45.7; n >= 0.9; we win on the first full
40   48-byte block to set.  */
41
42#define MEMSET_BY_BLOCK_THRESHOLD (1 * 48)
43
44/* No name ambiguities in this file.  */
45__asm__ (".syntax no_register_prefix");
46
47void *memset(void *pdst, int c, unsigned int plen)
48{
49  /* Now we want the parameters in special registers.  Make sure the
50     compiler does something usable with this.  */
51
52  register char *return_dst __asm__ ("r10") = pdst;
53  register int n __asm__ ("r12") = plen;
54  register int lc __asm__ ("r11") = c;
55
56  /* Most apps use memset sanely.  Memsetting about 3..4 bytes or less get
57     penalized here compared to the generic implementation.  */
58
59  /* This is fragile performancewise at best.  Check with newer GCC
60     releases, if they compile cascaded "x |= x << 8" to sane code.  */
61  __asm__("movu.b %0,r13						\n\
62	   lslq 8,r13							\n\
63	   move.b %0,r13						\n\
64	   move.d r13,%0						\n\
65	   lslq 16,r13							\n\
66	   or.d r13,%0"
67          : "=r" (lc)		/* Inputs.  */
68	  : "0" (lc)		/* Outputs.  */
69	  : "r13");		/* Trash.  */
70
71  {
72    register char *dst __asm__ ("r13") = pdst;
73
74    if (((unsigned long) pdst & 3) != 0
75	/* Oops! n = 0 must be a valid call, regardless of alignment.  */
76	&& n >= 3)
77      {
78	if ((unsigned long) dst & 1)
79	  {
80	    *dst = (char) lc;
81	    n--;
82	    dst++;
83	  }
84
85	if ((unsigned long) dst & 2)
86	  {
87	    *(short *) dst = lc;
88	    n -= 2;
89	    dst += 2;
90	  }
91      }
92
93    /* Decide which setting method to use.  */
94    if (n >= MEMSET_BY_BLOCK_THRESHOLD)
95      {
96	/* It is not optimal to tell the compiler about clobbering any
97	   registers; that will move the saving/restoring of those registers
98	   to the function prologue/epilogue, and make non-block sizes
99	   suboptimal.  */
100	__asm__ volatile
101	  ("\
102	   ;; GCC does promise correct register allocations, but let's	\n\
103	   ;; make sure it keeps its promises.				\n\
104	   .ifnc %0-%1-%4,$r13-$r12-$r11				\n\
105	   .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\"	\n\
106	   .endif							\n\
107									\n\
108	   ;; Save the registers we'll clobber in the movem process	\n\
109	   ;; on the stack.  Don't mention them to gcc, it will only be	\n\
110	   ;; upset.							\n\
111	   subq	   11*4,sp						\n\
112	   movem   r10,[sp]						\n\
113									\n\
114	   move.d  r11,r0						\n\
115	   move.d  r11,r1						\n\
116	   move.d  r11,r2						\n\
117	   move.d  r11,r3						\n\
118	   move.d  r11,r4						\n\
119	   move.d  r11,r5						\n\
120	   move.d  r11,r6						\n\
121	   move.d  r11,r7						\n\
122	   move.d  r11,r8						\n\
123	   move.d  r11,r9						\n\
124	   move.d  r11,r10						\n\
125									\n\
126	   ;; Now we've got this:					\n\
127	   ;; r13 - dst							\n\
128	   ;; r12 - n							\n\
129									\n\
130	   ;; Update n for the first loop				\n\
131	   subq	   12*4,r12						\n\
1320:									\n\
133"
134#ifdef __arch_common_v10_v32
135	   /* Cater to branch offset difference between v32 and v10.  We
136	      assume the branch below has an 8-bit offset.  */
137"	   setf\n"
138#endif
139"	   subq	  12*4,r12						\n\
140	   bge	   0b							\n\
141	   movem	r11,[r13+]					\n\
142									\n\
143	   ;; Compensate for last loop underflowing n.			\n\
144	   addq	  12*4,r12						\n\
145									\n\
146	   ;; Restore registers from stack.				\n\
147	   movem [sp+],r10"
148
149	   /* Outputs.	*/
150	   : "=r" (dst), "=r" (n)
151
152	   /* Inputs.  */
153	   : "0" (dst), "1" (n), "r" (lc));
154      }
155
156    /* An ad-hoc unroll, used for 4*12-1..16 bytes. */
157    while (n >= 16)
158      {
159	*(long *) dst = lc; dst += 4;
160	*(long *) dst = lc; dst += 4;
161	*(long *) dst = lc; dst += 4;
162	*(long *) dst = lc; dst += 4;
163	n -= 16;
164      }
165
166    switch (n)
167      {
168      case 0:
169        break;
170
171      case 1:
172        *dst = (char) lc;
173        break;
174
175      case 2:
176        *(short *) dst = (short) lc;
177        break;
178
179      case 3:
180        *(short *) dst = (short) lc; dst += 2;
181        *dst = (char) lc;
182        break;
183
184      case 4:
185        *(long *) dst = lc;
186        break;
187
188      case 5:
189        *(long *) dst = lc; dst += 4;
190        *dst = (char) lc;
191        break;
192
193      case 6:
194        *(long *) dst = lc; dst += 4;
195        *(short *) dst = (short) lc;
196        break;
197
198      case 7:
199        *(long *) dst = lc; dst += 4;
200        *(short *) dst = (short) lc; dst += 2;
201        *dst = (char) lc;
202        break;
203
204      case 8:
205        *(long *) dst = lc; dst += 4;
206        *(long *) dst = lc;
207        break;
208
209      case 9:
210        *(long *) dst = lc; dst += 4;
211        *(long *) dst = lc; dst += 4;
212        *dst = (char) lc;
213        break;
214
215      case 10:
216        *(long *) dst = lc; dst += 4;
217        *(long *) dst = lc; dst += 4;
218        *(short *) dst = (short) lc;
219        break;
220
221      case 11:
222        *(long *) dst = lc; dst += 4;
223        *(long *) dst = lc; dst += 4;
224        *(short *) dst = (short) lc; dst += 2;
225        *dst = (char) lc;
226        break;
227
228      case 12:
229        *(long *) dst = lc; dst += 4;
230        *(long *) dst = lc; dst += 4;
231        *(long *) dst = lc;
232        break;
233
234      case 13:
235        *(long *) dst = lc; dst += 4;
236        *(long *) dst = lc; dst += 4;
237        *(long *) dst = lc; dst += 4;
238        *dst = (char) lc;
239        break;
240
241      case 14:
242        *(long *) dst = lc; dst += 4;
243        *(long *) dst = lc; dst += 4;
244        *(long *) dst = lc; dst += 4;
245        *(short *) dst = (short) lc;
246        break;
247
248      case 15:
249        *(long *) dst = lc; dst += 4;
250        *(long *) dst = lc; dst += 4;
251        *(long *) dst = lc; dst += 4;
252        *(short *) dst = (short) lc; dst += 2;
253        *dst = (char) lc;
254        break;
255      }
256  }
257
258  return return_dst;
259}
260