aboutsummaryrefslogtreecommitdiffstats
path: root/arch/cris
diff options
context:
space:
mode:
authorJesper Nilsson <jesper.nilsson@axis.com>2008-02-14 22:31:22 -0500
committerLinus Torvalds <torvalds@woody.linux-foundation.org>2008-02-14 23:58:04 -0500
commit77a746cec58801208818ee19115da0e4d41f9002 (patch)
tree6508eb39880feaeecb1c2e01563110269066c026 /arch/cris
parent3c828e49453c4cb750b231d7116b8721c12b8663 (diff)
cris: import memset.c from newlib: fixes compile error with newer (pre4.3) gcc
Adrian Bunk reported the following compile error with a SVN head GCC: ... CC arch/cris/arch-v10/lib/memset.o /home/bunk/linux/kernel-2.6/git/linux-2.6/arch/cris/arch-v10/lib/memset.c: In function 'memset': /home/bunk/linux/kernel-2.6/git/linux-2.6/arch/cris/arch-v10/lib/memset.c:164: error: lvalue required as increment operand /home/bunk/linux/kernel-2.6/git/linux-2.6/arch/cris/arch-v10/lib/memset.c:165: error: lvalue required as increment operand /home/bunk/linux/kernel-2.6/git/linux-2.6/arch/cris/arch-v10/lib/memset.c:166: error: lvalue required as increment operand /home/bunk/linux/kernel-2.6/git/linux-2.6/arch/cris/arch-v10/lib/memset.c:167: error: lvalue required as increment operand /home/bunk/linux/kernel-2.6/git/linux-2.6/arch/cris/arch-v10/lib/memset.c:185: error: lvalue required as increment operand /home/bunk/linux/kernel-2.6/git/linux-2.6/arch/cris/arch-v10/lib/memset.c:189: error: lvalue required as increment operand /home/bunk/linux/kernel-2.6/git/linux-2.6/arch/cris/arch-v10/lib/memset.c:192: error: lvalue required as increment operand ... etc ... This is due to the use of the construct: *((long*)dst)++ = lc; Which is no longer legal since casts don't return an lvalue. The solution is to import the implementation from newlib, which is continually autotested together with GCC mainline, and uses the construct: *(long *) dst = lc; dst += 4; With this change, the generated code actually shrinks 76 bytes since gcc notices that it can use autoincrement for the move instruction in CRIS. text data bss dec hex filename 304 0 0 304 130 memset.old.o text data bss dec hex filename 228 0 0 228 e4 memset.o Since this is an import of a file from newlib, I'm not touching the formatting or correcting any checkpatch errors. Note also that even if the two files for the CRIS v10 and CRIS v32 are identical at the moment, it might be possible to tweak the CRIS v32 version. Thus, I'm not yet folding them into the same file, at least not until we've done some research on it. Signed-off-by: Jesper Nilsson <jesper.nilsson@axis.com> Cc: Mikael Starvik <mikael.starvik@axis.com> Cc: Adrian Bunk <bunk@stusta.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'arch/cris')
-rw-r--r--arch/cris/arch-v10/lib/memset.c397
-rw-r--r--arch/cris/arch-v32/lib/memset.c398
2 files changed, 404 insertions, 391 deletions
diff --git a/arch/cris/arch-v10/lib/memset.c b/arch/cris/arch-v10/lib/memset.c
index 42c1101043a3..c94ea9b3ec29 100644
--- a/arch/cris/arch-v10/lib/memset.c
+++ b/arch/cris/arch-v10/lib/memset.c
@@ -1,252 +1,259 @@
1/*#************************************************************************#*/ 1/* A memset for CRIS.
2/*#-------------------------------------------------------------------------*/ 2 Copyright (C) 1999-2005 Axis Communications.
3/*# */ 3 All rights reserved.
4/*# FUNCTION NAME: memset() */ 4
5/*# */ 5 Redistribution and use in source and binary forms, with or without
6/*# PARAMETERS: void* dst; Destination address. */ 6 modification, are permitted provided that the following conditions
7/*# int c; Value of byte to write. */ 7 are met:
8/*# int len; Number of bytes to write. */ 8
9/*# */ 9 1. Redistributions of source code must retain the above copyright
10/*# RETURNS: dst. */ 10 notice, this list of conditions and the following disclaimer.
11/*# */ 11
12/*# DESCRIPTION: Sets the memory dst of length len bytes to c, as standard. */ 12 2. Neither the name of Axis Communications nor the names of its
13/*# Framework taken from memcpy. This routine is */ 13 contributors may be used to endorse or promote products derived
14/*# very sensitive to compiler changes in register allocation. */ 14 from this software without specific prior written permission.
15/*# Should really be rewritten to avoid this problem. */ 15
16/*# */ 16 THIS SOFTWARE IS PROVIDED BY AXIS COMMUNICATIONS AND ITS CONTRIBUTORS
17/*#-------------------------------------------------------------------------*/ 17 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18/*# */ 18 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19/*# HISTORY */ 19 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL AXIS
20/*# */ 20 COMMUNICATIONS OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
21/*# DATE NAME CHANGES */ 21 INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
22/*# ---- ---- ------- */ 22 (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
23/*# 990713 HP Tired of watching this function (or */ 23 SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24/*# really, the nonoptimized generic */ 24 HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
25/*# implementation) take up 90% of simulator */ 25 STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
26/*# output. Measurements needed. */ 26 IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27/*# */ 27 POSSIBILITY OF SUCH DAMAGE. */
28/*#-------------------------------------------------------------------------*/ 28
29 29/* FIXME: This file should really only be used for reference, as the
30#include <linux/types.h> 30 result is somewhat depending on gcc generating what we expect rather
31 31 than what we describe. An assembly file should be used instead. */
32/* No, there's no macro saying 12*4, since it is "hard" to get it into 32
33 the asm in a good way. Thus better to expose the problem everywhere. 33/* Note the multiple occurrence of the expression "12*4", including the
34 */ 34 asm. It is hard to get it into the asm in a good way. Thus better to
35 35 expose the problem everywhere: no macro. */
36/* Assuming 1 cycle per dword written or read (ok, not really true), and 36
37 one per instruction, then 43+3*(n/48-1) <= 24+24*(n/48-1) 37/* Assuming one cycle per dword written or read (ok, not really true; the
38 so n >= 45.7; n >= 0.9; we win on the first full 48-byte block to set. */ 38 world is not ideal), and one cycle per instruction, then 43+3*(n/48-1)
39 39 <= 24+24*(n/48-1) so n >= 45.7; n >= 0.9; we win on the first full
40#define ZERO_BLOCK_SIZE (1*12*4) 40 48-byte block to set. */
41 41
42void *memset(void *pdst, 42#define MEMSET_BY_BLOCK_THRESHOLD (1 * 48)
43 int c, 43
44 size_t plen) 44/* No name ambiguities in this file. */
45__asm__ (".syntax no_register_prefix");
46
47void *memset(void *pdst, int c, unsigned int plen)
45{ 48{
46 /* Ok. Now we want the parameters put in special registers. 49 /* Now we want the parameters in special registers. Make sure the
47 Make sure the compiler is able to make something useful of this. */ 50 compiler does something usable with this. */
48 51
49 register char *return_dst __asm__ ("r10") = pdst; 52 register char *return_dst __asm__ ("r10") = pdst;
50 register int n __asm__ ("r12") = plen; 53 register int n __asm__ ("r12") = plen;
51 register int lc __asm__ ("r11") = c; 54 register int lc __asm__ ("r11") = c;
52 55
53 /* Most apps use memset sanely. Only those memsetting about 3..4 56 /* Most apps use memset sanely. Memsetting about 3..4 bytes or less get
54 bytes or less get penalized compared to the generic implementation 57 penalized here compared to the generic implementation. */
55 - and that's not really sane use. */
56 58
57 /* Ugh. This is fragile at best. Check with newer GCC releases, if 59 /* This is fragile performancewise at best. Check with newer GCC
58 they compile cascaded "x |= x << 8" sanely! */ 60 releases, if they compile cascaded "x |= x << 8" to sane code. */
59 __asm__("movu.b %0,$r13\n\t" 61 __asm__("movu.b %0,r13 \n\
60 "lslq 8,$r13\n\t" 62 lslq 8,r13 \n\
61 "move.b %0,$r13\n\t" 63 move.b %0,r13 \n\
62 "move.d $r13,%0\n\t" 64 move.d r13,%0 \n\
63 "lslq 16,$r13\n\t" 65 lslq 16,r13 \n\
64 "or.d $r13,%0" 66 or.d r13,%0"
65 : "=r" (lc) : "0" (lc) : "r13"); 67 : "=r" (lc) /* Inputs. */
68 : "0" (lc) /* Outputs. */
69 : "r13"); /* Trash. */
66 70
67 { 71 {
68 register char *dst __asm__ ("r13") = pdst; 72 register char *dst __asm__ ("r13") = pdst;
69 73
70 /* This is NONPORTABLE, but since this whole routine is */ 74 if (((unsigned long) pdst & 3) != 0
71 /* grossly nonportable that doesn't matter. */ 75 /* Oops! n = 0 must be a valid call, regardless of alignment. */
76 && n >= 3)
77 {
78 if ((unsigned long) dst & 1)
79 {
80 *dst = (char) lc;
81 n--;
82 dst++;
83 }
72 84
73 if (((unsigned long) pdst & 3) != 0 85 if ((unsigned long) dst & 2)
74 /* Oops! n=0 must be a legal call, regardless of alignment. */ 86 {
75 && n >= 3) 87 *(short *) dst = lc;
76 { 88 n -= 2;
77 if ((unsigned long)dst & 1) 89 dst += 2;
78 { 90 }
79 *dst = (char) lc; 91 }
80 n--;
81 dst++;
82 }
83
84 if ((unsigned long)dst & 2)
85 {
86 *(short *)dst = lc;
87 n -= 2;
88 dst += 2;
89 }
90 }
91 92
92 /* Now the fun part. For the threshold value of this, check the equation 93 /* Decide which setting method to use. */
93 above. */ 94 if (n >= MEMSET_BY_BLOCK_THRESHOLD)
94 /* Decide which copying method to use. */ 95 {
95 if (n >= ZERO_BLOCK_SIZE) 96 /* It is not optimal to tell the compiler about clobbering any
96 { 97 registers; that will move the saving/restoring of those registers
97 /* For large copies we use 'movem' */ 98 to the function prologue/epilogue, and make non-block sizes
98 99 suboptimal. */
99 /* It is not optimal to tell the compiler about clobbering any 100 __asm__ volatile
100 registers; that will move the saving/restoring of those registers 101 ("\
101 to the function prologue/epilogue, and make non-movem sizes 102 ;; GCC does promise correct register allocations, but let's \n\
102 suboptimal. 103 ;; make sure it keeps its promises. \n\
103 104 .ifnc %0-%1-%4,$r13-$r12-$r11 \n\
104 This method is not foolproof; it assumes that the "asm reg" 105 .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\" \n\
105 declarations at the beginning of the function really are used 106 .endif \n\
106 here (beware: they may be moved to temporary registers). 107 \n\
107 This way, we do not have to save/move the registers around into 108 ;; Save the registers we'll clobber in the movem process \n\
108 temporaries; we can safely use them straight away. 109 ;; on the stack. Don't mention them to gcc, it will only be \n\
109 110 ;; upset. \n\
110 If you want to check that the allocation was right; then 111 subq 11*4,sp \n\
111 check the equalities in the first comment. It should say 112 movem r10,[sp] \n\
112 "r13=r13, r12=r12, r11=r11" */
113 __asm__ volatile ("\n\
114 ;; Check that the following is true (same register names on \n\
115 ;; both sides of equal sign, as in r8=r8): \n\
116 ;; %0=r13, %1=r12, %4=r11 \n\
117 ;; \n\
118 ;; Save the registers we'll clobber in the movem process \n\
119 ;; on the stack. Don't mention them to gcc, it will only be \n\
120 ;; upset. \n\
121 subq 11*4,$sp \n\
122 movem $r10,[$sp] \n\
123 \n\ 113 \n\
124 move.d $r11,$r0 \n\ 114 move.d r11,r0 \n\
125 move.d $r11,$r1 \n\ 115 move.d r11,r1 \n\
126 move.d $r11,$r2 \n\ 116 move.d r11,r2 \n\
127 move.d $r11,$r3 \n\ 117 move.d r11,r3 \n\
128 move.d $r11,$r4 \n\ 118 move.d r11,r4 \n\
129 move.d $r11,$r5 \n\ 119 move.d r11,r5 \n\
130 move.d $r11,$r6 \n\ 120 move.d r11,r6 \n\
131 move.d $r11,$r7 \n\ 121 move.d r11,r7 \n\
132 move.d $r11,$r8 \n\ 122 move.d r11,r8 \n\
133 move.d $r11,$r9 \n\ 123 move.d r11,r9 \n\
134 move.d $r11,$r10 \n\ 124 move.d r11,r10 \n\
135 \n\ 125 \n\
136 ;; Now we've got this: \n\ 126 ;; Now we've got this: \n\
137 ;; r13 - dst \n\ 127 ;; r13 - dst \n\
138 ;; r12 - n \n\ 128 ;; r12 - n \n\
139 \n\ 129 \n\
140 ;; Update n for the first loop \n\ 130 ;; Update n for the first loop \n\
141 subq 12*4,$r12 \n\ 131 subq 12*4,r12 \n\
1420: \n\ 1320: \n\
143 subq 12*4,$r12 \n\ 133"
144 bge 0b \n\ 134#ifdef __arch_common_v10_v32
145 movem $r11,[$r13+] \n\ 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\
146 \n\ 142 \n\
147 addq 12*4,$r12 ;; compensate for last loop underflowing n \n\ 143 ;; Compensate for last loop underflowing n. \n\
144 addq 12*4,r12 \n\
148 \n\ 145 \n\
149 ;; Restore registers from stack \n\ 146 ;; Restore registers from stack. \n\
150 movem [$sp+],$r10" 147 movem [sp+],r10"
151 148
152 /* Outputs */ : "=r" (dst), "=r" (n) 149 /* Outputs. */
153 /* Inputs */ : "0" (dst), "1" (n), "r" (lc)); 150 : "=r" (dst), "=r" (n)
154 151
155 } 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 }
156 165
157 /* Either we directly starts copying, using dword copying
158 in a loop, or we copy as much as possible with 'movem'
159 and then the last block (<44 bytes) is copied here.
160 This will work since 'movem' will have updated src,dst,n. */
161
162 while ( n >= 16 )
163 {
164 *((long*)dst)++ = lc;
165 *((long*)dst)++ = lc;
166 *((long*)dst)++ = lc;
167 *((long*)dst)++ = lc;
168 n -= 16;
169 }
170
171 /* A switch() is definitely the fastest although it takes a LOT of code.
172 * Particularly if you inline code this.
173 */
174 switch (n) 166 switch (n)
175 { 167 {
176 case 0: 168 case 0:
177 break; 169 break;
170
178 case 1: 171 case 1:
179 *(char*)dst = (char) lc; 172 *dst = (char) lc;
180 break; 173 break;
174
181 case 2: 175 case 2:
182 *(short*)dst = (short) lc; 176 *(short *) dst = (short) lc;
183 break; 177 break;
178
184 case 3: 179 case 3:
185 *((short*)dst)++ = (short) lc; 180 *(short *) dst = (short) lc; dst += 2;
186 *(char*)dst = (char) lc; 181 *dst = (char) lc;
187 break; 182 break;
183
188 case 4: 184 case 4:
189 *((long*)dst)++ = lc; 185 *(long *) dst = lc;
190 break; 186 break;
187
191 case 5: 188 case 5:
192 *((long*)dst)++ = lc; 189 *(long *) dst = lc; dst += 4;
193 *(char*)dst = (char) lc; 190 *dst = (char) lc;
194 break; 191 break;
192
195 case 6: 193 case 6:
196 *((long*)dst)++ = lc; 194 *(long *) dst = lc; dst += 4;
197 *(short*)dst = (short) lc; 195 *(short *) dst = (short) lc;
198 break; 196 break;
197
199 case 7: 198 case 7:
200 *((long*)dst)++ = lc; 199 *(long *) dst = lc; dst += 4;
201 *((short*)dst)++ = (short) lc; 200 *(short *) dst = (short) lc; dst += 2;
202 *(char*)dst = (char) lc; 201 *dst = (char) lc;
203 break; 202 break;
203
204 case 8: 204 case 8:
205 *((long*)dst)++ = lc; 205 *(long *) dst = lc; dst += 4;
206 *((long*)dst)++ = lc; 206 *(long *) dst = lc;
207 break; 207 break;
208
208 case 9: 209 case 9:
209 *((long*)dst)++ = lc; 210 *(long *) dst = lc; dst += 4;
210 *((long*)dst)++ = lc; 211 *(long *) dst = lc; dst += 4;
211 *(char*)dst = (char) lc; 212 *dst = (char) lc;
212 break; 213 break;
214
213 case 10: 215 case 10:
214 *((long*)dst)++ = lc; 216 *(long *) dst = lc; dst += 4;
215 *((long*)dst)++ = lc; 217 *(long *) dst = lc; dst += 4;
216 *(short*)dst = (short) lc; 218 *(short *) dst = (short) lc;
217 break; 219 break;
220
218 case 11: 221 case 11:
219 *((long*)dst)++ = lc; 222 *(long *) dst = lc; dst += 4;
220 *((long*)dst)++ = lc; 223 *(long *) dst = lc; dst += 4;
221 *((short*)dst)++ = (short) lc; 224 *(short *) dst = (short) lc; dst += 2;
222 *(char*)dst = (char) lc; 225 *dst = (char) lc;
223 break; 226 break;
227
224 case 12: 228 case 12:
225 *((long*)dst)++ = lc; 229 *(long *) dst = lc; dst += 4;
226 *((long*)dst)++ = lc; 230 *(long *) dst = lc; dst += 4;
227 *((long*)dst)++ = lc; 231 *(long *) dst = lc;
228 break; 232 break;
233
229 case 13: 234 case 13:
230 *((long*)dst)++ = lc; 235 *(long *) dst = lc; dst += 4;
231 *((long*)dst)++ = lc; 236 *(long *) dst = lc; dst += 4;
232 *((long*)dst)++ = lc; 237 *(long *) dst = lc; dst += 4;
233 *(char*)dst = (char) lc; 238 *dst = (char) lc;
234 break; 239 break;
240
235 case 14: 241 case 14:
236 *((long*)dst)++ = lc; 242 *(long *) dst = lc; dst += 4;
237 *((long*)dst)++ = lc; 243 *(long *) dst = lc; dst += 4;
238 *((long*)dst)++ = lc; 244 *(long *) dst = lc; dst += 4;
239 *(short*)dst = (short) lc; 245 *(short *) dst = (short) lc;
240 break; 246 break;
247
241 case 15: 248 case 15:
242 *((long*)dst)++ = lc; 249 *(long *) dst = lc; dst += 4;
243 *((long*)dst)++ = lc; 250 *(long *) dst = lc; dst += 4;
244 *((long*)dst)++ = lc; 251 *(long *) dst = lc; dst += 4;
245 *((short*)dst)++ = (short) lc; 252 *(short *) dst = (short) lc; dst += 2;
246 *(char*)dst = (char) lc; 253 *dst = (char) lc;
247 break; 254 break;
248 } 255 }
249 } 256 }
250 257
251 return return_dst; /* destination pointer. */ 258 return return_dst;
252} /* memset() */ 259}
diff --git a/arch/cris/arch-v32/lib/memset.c b/arch/cris/arch-v32/lib/memset.c
index ffca1214674e..c94ea9b3ec29 100644
--- a/arch/cris/arch-v32/lib/memset.c
+++ b/arch/cris/arch-v32/lib/memset.c
@@ -1,253 +1,259 @@
1/*#************************************************************************#*/ 1/* A memset for CRIS.
2/*#-------------------------------------------------------------------------*/ 2 Copyright (C) 1999-2005 Axis Communications.
3/*# */ 3 All rights reserved.
4/*# FUNCTION NAME: memset() */ 4
5/*# */ 5 Redistribution and use in source and binary forms, with or without
6/*# PARAMETERS: void* dst; Destination address. */ 6 modification, are permitted provided that the following conditions
7/*# int c; Value of byte to write. */ 7 are met:
8/*# int len; Number of bytes to write. */ 8
9/*# */ 9 1. Redistributions of source code must retain the above copyright
10/*# RETURNS: dst. */ 10 notice, this list of conditions and the following disclaimer.
11/*# */ 11
12/*# DESCRIPTION: Sets the memory dst of length len bytes to c, as standard. */ 12 2. Neither the name of Axis Communications nor the names of its
13/*# Framework taken from memcpy. This routine is */ 13 contributors may be used to endorse or promote products derived
14/*# very sensitive to compiler changes in register allocation. */ 14 from this software without specific prior written permission.
15/*# Should really be rewritten to avoid this problem. */ 15
16/*# */ 16 THIS SOFTWARE IS PROVIDED BY AXIS COMMUNICATIONS AND ITS CONTRIBUTORS
17/*#-------------------------------------------------------------------------*/ 17 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18/*# */ 18 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19/*# HISTORY */ 19 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL AXIS
20/*# */ 20 COMMUNICATIONS OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
21/*# DATE NAME CHANGES */ 21 INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
22/*# ---- ---- ------- */ 22 (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
23/*# 990713 HP Tired of watching this function (or */ 23 SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24/*# really, the nonoptimized generic */ 24 HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
25/*# implementation) take up 90% of simulator */ 25 STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
26/*# output. Measurements needed. */ 26 IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27/*# */ 27 POSSIBILITY OF SUCH DAMAGE. */
28/*#-------------------------------------------------------------------------*/ 28
29 29/* FIXME: This file should really only be used for reference, as the
30#include <linux/types.h> 30 result is somewhat depending on gcc generating what we expect rather
31 31 than what we describe. An assembly file should be used instead. */
32/* No, there's no macro saying 12*4, since it is "hard" to get it into 32
33 the asm in a good way. Thus better to expose the problem everywhere. 33/* Note the multiple occurrence of the expression "12*4", including the
34 */ 34 asm. It is hard to get it into the asm in a good way. Thus better to
35 35 expose the problem everywhere: no macro. */
36/* Assuming 1 cycle per dword written or read (ok, not really true), and 36
37 one per instruction, then 43+3*(n/48-1) <= 24+24*(n/48-1) 37/* Assuming one cycle per dword written or read (ok, not really true; the
38 so n >= 45.7; n >= 0.9; we win on the first full 48-byte block to set. */ 38 world is not ideal), and one cycle per instruction, then 43+3*(n/48-1)
39 39 <= 24+24*(n/48-1) so n >= 45.7; n >= 0.9; we win on the first full
40#define ZERO_BLOCK_SIZE (1*12*4) 40 48-byte block to set. */
41 41
42void *memset(void *pdst, 42#define MEMSET_BY_BLOCK_THRESHOLD (1 * 48)
43 int c, 43
44 size_t plen) 44/* No name ambiguities in this file. */
45__asm__ (".syntax no_register_prefix");
46
47void *memset(void *pdst, int c, unsigned int plen)
45{ 48{
46 /* Ok. Now we want the parameters put in special registers. 49 /* Now we want the parameters in special registers. Make sure the
47 Make sure the compiler is able to make something useful of this. */ 50 compiler does something usable with this. */
48 51
49 register char *return_dst __asm__ ("r10") = pdst; 52 register char *return_dst __asm__ ("r10") = pdst;
50 register int n __asm__ ("r12") = plen; 53 register int n __asm__ ("r12") = plen;
51 register int lc __asm__ ("r11") = c; 54 register int lc __asm__ ("r11") = c;
52 55
53 /* Most apps use memset sanely. Only those memsetting about 3..4 56 /* Most apps use memset sanely. Memsetting about 3..4 bytes or less get
54 bytes or less get penalized compared to the generic implementation 57 penalized here compared to the generic implementation. */
55 - and that's not really sane use. */
56 58
57 /* Ugh. This is fragile at best. Check with newer GCC releases, if 59 /* This is fragile performancewise at best. Check with newer GCC
58 they compile cascaded "x |= x << 8" sanely! */ 60 releases, if they compile cascaded "x |= x << 8" to sane code. */
59 __asm__("movu.b %0,$r13 \n\ 61 __asm__("movu.b %0,r13 \n\
60 lslq 8,$r13 \n\ 62 lslq 8,r13 \n\
61 move.b %0,$r13 \n\ 63 move.b %0,r13 \n\
62 move.d $r13,%0 \n\ 64 move.d r13,%0 \n\
63 lslq 16,$r13 \n\ 65 lslq 16,r13 \n\
64 or.d $r13,%0" 66 or.d r13,%0"
65 : "=r" (lc) : "0" (lc) : "r13"); 67 : "=r" (lc) /* Inputs. */
68 : "0" (lc) /* Outputs. */
69 : "r13"); /* Trash. */
66 70
67 { 71 {
68 register char *dst __asm__ ("r13") = pdst; 72 register char *dst __asm__ ("r13") = pdst;
69 73
70 /* This is NONPORTABLE, but since this whole routine is */ 74 if (((unsigned long) pdst & 3) != 0
71 /* grossly nonportable that doesn't matter. */ 75 /* Oops! n = 0 must be a valid call, regardless of alignment. */
76 && n >= 3)
77 {
78 if ((unsigned long) dst & 1)
79 {
80 *dst = (char) lc;
81 n--;
82 dst++;
83 }
72 84
73 if (((unsigned long) pdst & 3) != 0 85 if ((unsigned long) dst & 2)
74 /* Oops! n=0 must be a legal call, regardless of alignment. */ 86 {
75 && n >= 3) 87 *(short *) dst = lc;
76 { 88 n -= 2;
77 if ((unsigned long)dst & 1) 89 dst += 2;
78 { 90 }
79 *dst = (char) lc; 91 }
80 n--;
81 dst++;
82 }
83
84 if ((unsigned long)dst & 2)
85 {
86 *(short *)dst = lc;
87 n -= 2;
88 dst += 2;
89 }
90 }
91 92
92 /* Now the fun part. For the threshold value of this, check the equation 93 /* Decide which setting method to use. */
93 above. */ 94 if (n >= MEMSET_BY_BLOCK_THRESHOLD)
94 /* Decide which copying method to use. */ 95 {
95 if (n >= ZERO_BLOCK_SIZE) 96 /* It is not optimal to tell the compiler about clobbering any
96 { 97 registers; that will move the saving/restoring of those registers
97 /* For large copies we use 'movem' */ 98 to the function prologue/epilogue, and make non-block sizes
98 99 suboptimal. */
99 /* It is not optimal to tell the compiler about clobbering any 100 __asm__ volatile
100 registers; that will move the saving/restoring of those registers 101 ("\
101 to the function prologue/epilogue, and make non-movem sizes 102 ;; GCC does promise correct register allocations, but let's \n\
102 suboptimal. 103 ;; make sure it keeps its promises. \n\
103 104 .ifnc %0-%1-%4,$r13-$r12-$r11 \n\
104 This method is not foolproof; it assumes that the "asm reg" 105 .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\" \n\
105 declarations at the beginning of the function really are used 106 .endif \n\
106 here (beware: they may be moved to temporary registers).
107 This way, we do not have to save/move the registers around into
108 temporaries; we can safely use them straight away.
109
110 If you want to check that the allocation was right; then
111 check the equalities in the first comment. It should say
112 "r13=r13, r12=r12, r11=r11" */
113 __asm__ volatile (" \n\
114 ;; Check that the register asm declaration got right. \n\
115 ;; The GCC manual says it will work, but there *has* been bugs. \n\
116 .ifnc %0-%1-%4,$r13-$r12-$r11 \n\
117 .err \n\
118 .endif \n\
119 \n\ 107 \n\
120 ;; Save the registers we'll clobber in the movem process \n\ 108 ;; Save the registers we'll clobber in the movem process \n\
121 ;; on the stack. Don't mention them to gcc, it will only be \n\ 109 ;; on the stack. Don't mention them to gcc, it will only be \n\
122 ;; upset. \n\ 110 ;; upset. \n\
123 subq 11*4,$sp \n\ 111 subq 11*4,sp \n\
124 movem $r10,[$sp] \n\ 112 movem r10,[sp] \n\
125 \n\ 113 \n\
126 move.d $r11,$r0 \n\ 114 move.d r11,r0 \n\
127 move.d $r11,$r1 \n\ 115 move.d r11,r1 \n\
128 move.d $r11,$r2 \n\ 116 move.d r11,r2 \n\
129 move.d $r11,$r3 \n\ 117 move.d r11,r3 \n\
130 move.d $r11,$r4 \n\ 118 move.d r11,r4 \n\
131 move.d $r11,$r5 \n\ 119 move.d r11,r5 \n\
132 move.d $r11,$r6 \n\ 120 move.d r11,r6 \n\
133 move.d $r11,$r7 \n\ 121 move.d r11,r7 \n\
134 move.d $r11,$r8 \n\ 122 move.d r11,r8 \n\
135 move.d $r11,$r9 \n\ 123 move.d r11,r9 \n\
136 move.d $r11,$r10 \n\ 124 move.d r11,r10 \n\
137 \n\ 125 \n\
138 ;; Now we've got this: \n\ 126 ;; Now we've got this: \n\
139 ;; r13 - dst \n\ 127 ;; r13 - dst \n\
140 ;; r12 - n \n\ 128 ;; r12 - n \n\
141 \n\ 129 \n\
142 ;; Update n for the first loop \n\ 130 ;; Update n for the first loop \n\
143 subq 12*4,$r12 \n\ 131 subq 12*4,r12 \n\
1440: \n\ 1320: \n\
145 subq 12*4,$r12 \n\ 133"
146 bge 0b \n\ 134#ifdef __arch_common_v10_v32
147 movem $r11,[$r13+] \n\ 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\
148 \n\ 142 \n\
149 addq 12*4,$r12 ;; compensate for last loop underflowing n \n\ 143 ;; Compensate for last loop underflowing n. \n\
144 addq 12*4,r12 \n\
150 \n\ 145 \n\
151 ;; Restore registers from stack \n\ 146 ;; Restore registers from stack. \n\
152 movem [$sp+],$r10" 147 movem [sp+],r10"
153 148
154 /* Outputs */ : "=r" (dst), "=r" (n) 149 /* Outputs. */
155 /* Inputs */ : "0" (dst), "1" (n), "r" (lc)); 150 : "=r" (dst), "=r" (n)
156 } 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 }
157 165
158 /* Either we directly starts copying, using dword copying
159 in a loop, or we copy as much as possible with 'movem'
160 and then the last block (<44 bytes) is copied here.
161 This will work since 'movem' will have updated src,dst,n. */
162
163 while ( n >= 16 )
164 {
165 *((long*)dst)++ = lc;
166 *((long*)dst)++ = lc;
167 *((long*)dst)++ = lc;
168 *((long*)dst)++ = lc;
169 n -= 16;
170 }
171
172 /* A switch() is definitely the fastest although it takes a LOT of code.
173 * Particularly if you inline code this.
174 */
175 switch (n) 166 switch (n)
176 { 167 {
177 case 0: 168 case 0:
178 break; 169 break;
170
179 case 1: 171 case 1:
180 *(char*)dst = (char) lc; 172 *dst = (char) lc;
181 break; 173 break;
174
182 case 2: 175 case 2:
183 *(short*)dst = (short) lc; 176 *(short *) dst = (short) lc;
184 break; 177 break;
178
185 case 3: 179 case 3:
186 *((short*)dst)++ = (short) lc; 180 *(short *) dst = (short) lc; dst += 2;
187 *(char*)dst = (char) lc; 181 *dst = (char) lc;
188 break; 182 break;
183
189 case 4: 184 case 4:
190 *((long*)dst)++ = lc; 185 *(long *) dst = lc;
191 break; 186 break;
187
192 case 5: 188 case 5:
193 *((long*)dst)++ = lc; 189 *(long *) dst = lc; dst += 4;
194 *(char*)dst = (char) lc; 190 *dst = (char) lc;
195 break; 191 break;
192
196 case 6: 193 case 6:
197 *((long*)dst)++ = lc; 194 *(long *) dst = lc; dst += 4;
198 *(short*)dst = (short) lc; 195 *(short *) dst = (short) lc;
199 break; 196 break;
197
200 case 7: 198 case 7:
201 *((long*)dst)++ = lc; 199 *(long *) dst = lc; dst += 4;
202 *((short*)dst)++ = (short) lc; 200 *(short *) dst = (short) lc; dst += 2;
203 *(char*)dst = (char) lc; 201 *dst = (char) lc;
204 break; 202 break;
203
205 case 8: 204 case 8:
206 *((long*)dst)++ = lc; 205 *(long *) dst = lc; dst += 4;
207 *((long*)dst)++ = lc; 206 *(long *) dst = lc;
208 break; 207 break;
208
209 case 9: 209 case 9:
210 *((long*)dst)++ = lc; 210 *(long *) dst = lc; dst += 4;
211 *((long*)dst)++ = lc; 211 *(long *) dst = lc; dst += 4;
212 *(char*)dst = (char) lc; 212 *dst = (char) lc;
213 break; 213 break;
214
214 case 10: 215 case 10:
215 *((long*)dst)++ = lc; 216 *(long *) dst = lc; dst += 4;
216 *((long*)dst)++ = lc; 217 *(long *) dst = lc; dst += 4;
217 *(short*)dst = (short) lc; 218 *(short *) dst = (short) lc;
218 break; 219 break;
220
219 case 11: 221 case 11:
220 *((long*)dst)++ = lc; 222 *(long *) dst = lc; dst += 4;
221 *((long*)dst)++ = lc; 223 *(long *) dst = lc; dst += 4;
222 *((short*)dst)++ = (short) lc; 224 *(short *) dst = (short) lc; dst += 2;
223 *(char*)dst = (char) lc; 225 *dst = (char) lc;
224 break; 226 break;
227
225 case 12: 228 case 12:
226 *((long*)dst)++ = lc; 229 *(long *) dst = lc; dst += 4;
227 *((long*)dst)++ = lc; 230 *(long *) dst = lc; dst += 4;
228 *((long*)dst)++ = lc; 231 *(long *) dst = lc;
229 break; 232 break;
233
230 case 13: 234 case 13:
231 *((long*)dst)++ = lc; 235 *(long *) dst = lc; dst += 4;
232 *((long*)dst)++ = lc; 236 *(long *) dst = lc; dst += 4;
233 *((long*)dst)++ = lc; 237 *(long *) dst = lc; dst += 4;
234 *(char*)dst = (char) lc; 238 *dst = (char) lc;
235 break; 239 break;
240
236 case 14: 241 case 14:
237 *((long*)dst)++ = lc; 242 *(long *) dst = lc; dst += 4;
238 *((long*)dst)++ = lc; 243 *(long *) dst = lc; dst += 4;
239 *((long*)dst)++ = lc; 244 *(long *) dst = lc; dst += 4;
240 *(short*)dst = (short) lc; 245 *(short *) dst = (short) lc;
241 break; 246 break;
247
242 case 15: 248 case 15:
243 *((long*)dst)++ = lc; 249 *(long *) dst = lc; dst += 4;
244 *((long*)dst)++ = lc; 250 *(long *) dst = lc; dst += 4;
245 *((long*)dst)++ = lc; 251 *(long *) dst = lc; dst += 4;
246 *((short*)dst)++ = (short) lc; 252 *(short *) dst = (short) lc; dst += 2;
247 *(char*)dst = (char) lc; 253 *dst = (char) lc;
248 break; 254 break;
249 } 255 }
250 } 256 }
251 257
252 return return_dst; /* destination pointer. */ 258 return return_dst;
253} /* memset() */ 259}
generated by cgit v1.2.2 (git 2.25.0) at 2024-09-28 15:35:18 -0400