diff options
| -rw-r--r-- | arch/cris/arch-v10/lib/memset.c | 397 | ||||
| -rw-r--r-- | arch/cris/arch-v32/lib/memset.c | 398 |
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 | ||
| 42 | void *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 | |||
| 47 | void *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\ | ||
