Skip to content

L
linux-davinci
Commit 77a746ce authored by Jesper Nilsson's avatar Jesper Nilsson Committed by Linus Torvalds
Browse files
Options

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: default avatarJesper Nilsson <jesper.nilsson@axis.com>
Cc: Mikael Starvik <mikael.starvik@axis.com>
Cc: Adrian Bunk <bunk@stusta.de>
Signed-off-by: default avatarAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
parent 3c828e49
Show whitespace changes
Inline Side-by-side
Showing with 404 additions and 391 deletions
arch/cris/arch-v10/lib/memset.c
View file @ 77a746ce
/*#************************************************************************#*/ /* A memset for CRIS.
/*#-------------------------------------------------------------------------*/ Copyright (C) 1999-2005 Axis Communications.
/*# */ All rights reserved.
/*# FUNCTION NAME: memset() */
/*# */ Redistribution and use in source and binary forms, with or without
/*# PARAMETERS: void* dst; Destination address. */ modification, are permitted provided that the following conditions
/*# int c; Value of byte to write. */ are met:
/*# int len; Number of bytes to write. */
/*# */ 1. Redistributions of source code must retain the above copyright
/*# RETURNS: dst. */ notice, this list of conditions and the following disclaimer.
/*# */
/*# DESCRIPTION: Sets the memory dst of length len bytes to c, as standard. */ 2. Neither the name of Axis Communications nor the names of its
/*# Framework taken from memcpy. This routine is */ contributors may be used to endorse or promote products derived
/*# very sensitive to compiler changes in register allocation. */ from this software without specific prior written permission.
/*# Should really be rewritten to avoid this problem. */
/*# */ THIS SOFTWARE IS PROVIDED BY AXIS COMMUNICATIONS AND ITS CONTRIBUTORS
/*#-------------------------------------------------------------------------*/ ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
/*# */ LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
/*# HISTORY */ A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL AXIS
/*# */ COMMUNICATIONS OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
/*# DATE NAME CHANGES */ INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
/*# ---- ---- ------- */ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
/*# 990713 HP Tired of watching this function (or */ SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
/*# really, the nonoptimized generic */ HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
/*# implementation) take up 90% of simulator */ STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
/*# output. Measurements needed. */ IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
/*# */ POSSIBILITY OF SUCH DAMAGE. */
/*#-------------------------------------------------------------------------*/
/* FIXME: This file should really only be used for reference, as the
#include <linux/types.h> result is somewhat depending on gcc generating what we expect rather
than what we describe. An assembly file should be used instead. */
/* No, there's no macro saying 12*4, since it is "hard" to get it into
the asm in a good way. Thus better to expose the problem everywhere. /* Note the multiple occurrence of the expression "12*4", including the
*/ asm. It is hard to get it into the asm in a good way. Thus better to
expose the problem everywhere: no macro. */
/* Assuming 1 cycle per dword written or read (ok, not really true), and
one per instruction, then 43+3*(n/48-1) <= 24+24*(n/48-1) /* Assuming one cycle per dword written or read (ok, not really true; the
so n >= 45.7; n >= 0.9; we win on the first full 48-byte block to set. */ world is not ideal), and one cycle per instruction, then 43+3*(n/48-1)
<= 24+24*(n/48-1) so n >= 45.7; n >= 0.9; we win on the first full
#define ZERO_BLOCK_SIZE (1*12*4) 48-byte block to set. */
void *memset(void *pdst, #define MEMSET_BY_BLOCK_THRESHOLD (1 * 48)
int c,
size_t plen) /* No name ambiguities in this file. */
__asm__ (".syntax no_register_prefix");
void *memset(void *pdst, int c, unsigned int plen)
{ {
/* Ok. Now we want the parameters put in special registers. /* Now we want the parameters in special registers. Make sure the
Make sure the compiler is able to make something useful of this. */ compiler does something usable with this. */
register char *return_dst __asm__ ("r10") = pdst; register char *return_dst __asm__ ("r10") = pdst;
register int n __asm__ ("r12") = plen; register int n __asm__ ("r12") = plen;
register int lc __asm__ ("r11") = c; register int lc __asm__ ("r11") = c;
/* Most apps use memset sanely. Only those memsetting about 3..4 /* Most apps use memset sanely. Memsetting about 3..4 bytes or less get
bytes or less get penalized compared to the generic implementation penalized here compared to the generic implementation. */
- and that's not really sane use. */
/* Ugh. This is fragile at best. Check with newer GCC releases, if /* This is fragile performancewise at best. Check with newer GCC
they compile cascaded "x |= x << 8" sanely! */ releases, if they compile cascaded "x |= x << 8" to sane code. */
__asm__("movu.b %0,$r13\n\t" __asm__("movu.b %0,r13 \n\
"lslq 8,$r13\n\t" lslq 8,r13 \n\
"move.b %0,$r13\n\t" move.b %0,r13 \n\
"move.d $r13,%0\n\t" move.d r13,%0 \n\
"lslq 16,$r13\n\t" lslq 16,r13 \n\
"or.d $r13,%0" or.d r13,%0"
: "=r" (lc) : "0" (lc) : "r13"); : "=r" (lc) /* Inputs. */
: "0" (lc) /* Outputs. */
: "r13"); /* Trash. */
{ {
register char *dst __asm__ ("r13") = pdst; register char *dst __asm__ ("r13") = pdst;
/* This is NONPORTABLE, but since this whole routine is */
/* grossly nonportable that doesn't matter. */
if (((unsigned long) pdst & 3) != 0 if (((unsigned long) pdst & 3) != 0
/* Oops! n=0 must be a legal call, regardless of alignment. */ /* Oops! n = 0 must be a valid call, regardless of alignment. */
&& n >= 3) && n >= 3)
{ {
if ((unsigned long)dst & 1) if ((unsigned long) dst & 1)
{ {
*dst = (char) lc; *dst = (char) lc;
n--; n--;
dst++; dst++;
} }
if ((unsigned long)dst & 2) if ((unsigned long) dst & 2)
{ {
*(short *)dst = lc; *(short *) dst = lc;
n -= 2; n -= 2;
dst += 2; dst += 2;
} }
} }
/* Now the fun part. For the threshold value of this, check the equation /* Decide which setting method to use. */
above. */ if (n >= MEMSET_BY_BLOCK_THRESHOLD)
/* Decide which copying method to use. */
if (n >= ZERO_BLOCK_SIZE)
{ {
/* For large copies we use 'movem' */
/* It is not optimal to tell the compiler about clobbering any /* It is not optimal to tell the compiler about clobbering any
registers; that will move the saving/restoring of those registers registers; that will move the saving/restoring of those registers
to the function prologue/epilogue, and make non-movem sizes to the function prologue/epilogue, and make non-block sizes
suboptimal. suboptimal. */
__asm__ volatile
This method is not foolproof; it assumes that the "asm reg" ("\
declarations at the beginning of the function really are used ;; GCC does promise correct register allocations, but let's \n\
here (beware: they may be moved to temporary registers). ;; make sure it keeps its promises. \n\
This way, we do not have to save/move the registers around into .ifnc %0-%1-%4,$r13-$r12-$r11 \n\
temporaries; we can safely use them straight away. .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\" \n\
.endif \n\
If you want to check that the allocation was right; then \n\
check the equalities in the first comment. It should say
"r13=r13, r12=r12, r11=r11" */
__asm__ volatile ("\n\
;; Check that the following is true (same register names on \n\
;; both sides of equal sign, as in r8=r8): \n\
;; %0=r13, %1=r12, %4=r11 \n\
;; \n\
;; Save the registers we'll clobber in the movem process \n\ ;; Save the registers we'll clobber in the movem process \n\
;; on the stack. Don't mention them to gcc, it will only be \n\ ;; on the stack. Don't mention them to gcc, it will only be \n\
;; upset. \n\ ;; upset. \n\
subq 11*4,$sp \n\ subq 11*4,sp \n\
movem $r10,[$sp] \n\ movem r10,[sp] \n\
\n\ \n\
move.d $r11,$r0 \n\ move.d r11,r0 \n\
move.d $r11,$r1 \n\ move.d r11,r1 \n\
move.d $r11,$r2 \n\ move.d r11,r2 \n\
move.d $r11,$r3 \n\ move.d r11,r3 \n\
move.d $r11,$r4 \n\ move.d r11,r4 \n\
move.d $r11,$r5 \n\ move.d r11,r5 \n\
move.d $r11,$r6 \n\ move.d r11,r6 \n\
move.d $r11,$r7 \n\ move.d r11,r7 \n\
move.d $r11,$r8 \n\ move.d r11,r8 \n\
move.d $r11,$r9 \n\ move.d r11,r9 \n\
move.d $r11,$r10 \n\ move.d r11,r10 \n\
\n\ \n\
;; Now we've got this: \n\ ;; Now we've got this: \n\
;; r13 - dst \n\ ;; r13 - dst \n\
;; r12 - n \n\ ;; r12 - n \n\
\n\ \n\
;; Update n for the first loop \n\ ;; Update n for the first loop \n\
subq 12*4,$r12 \n\ subq 12*4,r12 \n\
0: \n\ 0: \n\
subq 12*4,$r12 \n\ "
#ifdef __arch_common_v10_v32
/* Cater to branch offset difference between v32 and v10. We
assume the branch below has an 8-bit offset. */
" setf\n"
#endif
" subq 12*4,r12 \n\
bge 0b \n\ bge 0b \n\
movem $r11,[$r13+] \n\ movem r11,[r13+] \n\
\n\ \n\
addq 12*4,$r12 ;; compensate for last loop underflowing n \n\ ;; Compensate for last loop underflowing n. \n\
addq 12*4,r12 \n\
\n\ \n\
;; Restore registers from stack \n\ ;; Restore registers from stack. \n\
movem [$sp+],$r10" movem [sp+],r10"
/* Outputs */ : "=r" (dst), "=r" (n) /* Outputs. */
/* Inputs */ : "0" (dst), "1" (n), "r" (lc)); : "=r" (dst), "=r" (n)
/* Inputs. */
: "0" (dst), "1" (n), "r" (lc));
} }
/* Either we directly starts copying, using dword copying /* An ad-hoc unroll, used for 4*12-1..16 bytes. */
in a loop, or we copy as much as possible with 'movem' while (n >= 16)
and then the last block (<44 bytes) is copied here.
This will work since 'movem' will have updated src,dst,n. */
while ( n >= 16 )
{ {
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
n -= 16; n -= 16;
} }
/* A switch() is definitely the fastest although it takes a LOT of code.
* Particularly if you inline code this.
*/
switch (n) switch (n)
{ {
case 0: case 0:
break; break;
case 1: case 1:
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 2: case 2:
*(short*)dst = (short) lc; *(short *) dst = (short) lc;
break; break;
case 3: case 3:
*((short*)dst)++ = (short) lc; *(short *) dst = (short) lc; dst += 2;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 4: case 4:
*((long*)dst)++ = lc; *(long *) dst = lc;
break; break;
case 5: case 5:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 6: case 6:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(short*)dst = (short) lc; *(short *) dst = (short) lc;
break; break;
case 7: case 7:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((short*)dst)++ = (short) lc; *(short *) dst = (short) lc; dst += 2;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 8: case 8:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc;
break; break;
case 9: case 9:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 10: case 10:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(short*)dst = (short) lc; *(short *) dst = (short) lc;
break; break;
case 11: case 11:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((short*)dst)++ = (short) lc; *(short *) dst = (short) lc; dst += 2;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 12: case 12:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc;
break; break;
case 13: case 13:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 14: case 14:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(short*)dst = (short) lc; *(short *) dst = (short) lc;
break; break;
case 15: case 15:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((short*)dst)++ = (short) lc; *(short *) dst = (short) lc; dst += 2;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
} }
} }
return return_dst; /* destination pointer. */ return return_dst;
} /* memset() */ }
arch/cris/arch-v32/lib/memset.c
View file @ 77a746ce
/*#************************************************************************#*/ /* A memset for CRIS.
/*#-------------------------------------------------------------------------*/ Copyright (C) 1999-2005 Axis Communications.
/*# */ All rights reserved.
/*# FUNCTION NAME: memset() */
/*# */ Redistribution and use in source and binary forms, with or without
/*# PARAMETERS: void* dst; Destination address. */ modification, are permitted provided that the following conditions
/*# int c; Value of byte to write. */ are met:
/*# int len; Number of bytes to write. */
/*# */ 1. Redistributions of source code must retain the above copyright
/*# RETURNS: dst. */ notice, this list of conditions and the following disclaimer.
/*# */
/*# DESCRIPTION: Sets the memory dst of length len bytes to c, as standard. */ 2. Neither the name of Axis Communications nor the names of its
/*# Framework taken from memcpy. This routine is */ contributors may be used to endorse or promote products derived
/*# very sensitive to compiler changes in register allocation. */ from this software without specific prior written permission.
/*# Should really be rewritten to avoid this problem. */
/*# */ THIS SOFTWARE IS PROVIDED BY AXIS COMMUNICATIONS AND ITS CONTRIBUTORS
/*#-------------------------------------------------------------------------*/ ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
/*# */ LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
/*# HISTORY */ A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL AXIS
/*# */ COMMUNICATIONS OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
/*# DATE NAME CHANGES */ INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
/*# ---- ---- ------- */ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
/*# 990713 HP Tired of watching this function (or */ SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
/*# really, the nonoptimized generic */ HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
/*# implementation) take up 90% of simulator */ STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
/*# output. Measurements needed. */ IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
/*# */ POSSIBILITY OF SUCH DAMAGE. */
/*#-------------------------------------------------------------------------*/
/* FIXME: This file should really only be used for reference, as the
#include <linux/types.h> result is somewhat depending on gcc generating what we expect rather
than what we describe. An assembly file should be used instead. */
/* No, there's no macro saying 12*4, since it is "hard" to get it into
the asm in a good way. Thus better to expose the problem everywhere. /* Note the multiple occurrence of the expression "12*4", including the
*/ asm. It is hard to get it into the asm in a good way. Thus better to
expose the problem everywhere: no macro. */
/* Assuming 1 cycle per dword written or read (ok, not really true), and
one per instruction, then 43+3*(n/48-1) <= 24+24*(n/48-1) /* Assuming one cycle per dword written or read (ok, not really true; the
so n >= 45.7; n >= 0.9; we win on the first full 48-byte block to set. */ world is not ideal), and one cycle per instruction, then 43+3*(n/48-1)
<= 24+24*(n/48-1) so n >= 45.7; n >= 0.9; we win on the first full
#define ZERO_BLOCK_SIZE (1*12*4) 48-byte block to set. */
void *memset(void *pdst, #define MEMSET_BY_BLOCK_THRESHOLD (1 * 48)
int c,
size_t plen) /* No name ambiguities in this file. */
__asm__ (".syntax no_register_prefix");
void *memset(void *pdst, int c, unsigned int plen)
{ {
/* Ok. Now we want the parameters put in special registers. /* Now we want the parameters in special registers. Make sure the
Make sure the compiler is able to make something useful of this. */ compiler does something usable with this. */
register char *return_dst __asm__ ("r10") = pdst; register char *return_dst __asm__ ("r10") = pdst;
register int n __asm__ ("r12") = plen; register int n __asm__ ("r12") = plen;
register int lc __asm__ ("r11") = c; register int lc __asm__ ("r11") = c;
/* Most apps use memset sanely. Only those memsetting about 3..4 /* Most apps use memset sanely. Memsetting about 3..4 bytes or less get
bytes or less get penalized compared to the generic implementation penalized here compared to the generic implementation. */
- and that's not really sane use. */
/* Ugh. This is fragile at best. Check with newer GCC releases, if /* This is fragile performancewise at best. Check with newer GCC
they compile cascaded "x |= x << 8" sanely! */ releases, if they compile cascaded "x |= x << 8" to sane code. */
__asm__("movu.b %0,$r13 \n\ __asm__("movu.b %0,r13 \n\
lslq 8,$r13 \n\ lslq 8,r13 \n\
move.b %0,$r13 \n\ move.b %0,r13 \n\
move.d $r13,%0 \n\ move.d r13,%0 \n\
lslq 16,$r13 \n\ lslq 16,r13 \n\
or.d $r13,%0" or.d r13,%0"
: "=r" (lc) : "0" (lc) : "r13"); : "=r" (lc) /* Inputs. */
: "0" (lc) /* Outputs. */
: "r13"); /* Trash. */
{ {
register char *dst __asm__ ("r13") = pdst; register char *dst __asm__ ("r13") = pdst;
/* This is NONPORTABLE, but since this whole routine is */
/* grossly nonportable that doesn't matter. */
if (((unsigned long) pdst & 3) != 0 if (((unsigned long) pdst & 3) != 0
/* Oops! n=0 must be a legal call, regardless of alignment. */ /* Oops! n = 0 must be a valid call, regardless of alignment. */
&& n >= 3) && n >= 3)
{ {
if ((unsigned long)dst & 1) if ((unsigned long) dst & 1)
{ {
*dst = (char) lc; *dst = (char) lc;
n--; n--;
dst++; dst++;
} }
if ((unsigned long)dst & 2) if ((unsigned long) dst & 2)
{ {
*(short *)dst = lc; *(short *) dst = lc;
n -= 2; n -= 2;
dst += 2; dst += 2;
} }
} }
/* Now the fun part. For the threshold value of this, check the equation /* Decide which setting method to use. */
above. */ if (n >= MEMSET_BY_BLOCK_THRESHOLD)
/* Decide which copying method to use. */
if (n >= ZERO_BLOCK_SIZE)
{ {
/* For large copies we use 'movem' */
/* It is not optimal to tell the compiler about clobbering any /* It is not optimal to tell the compiler about clobbering any
registers; that will move the saving/restoring of those registers registers; that will move the saving/restoring of those registers
to the function prologue/epilogue, and make non-movem sizes to the function prologue/epilogue, and make non-block sizes
suboptimal. suboptimal. */
__asm__ volatile
This method is not foolproof; it assumes that the "asm reg" ("\
declarations at the beginning of the function really are used ;; GCC does promise correct register allocations, but let's \n\
here (beware: they may be moved to temporary registers). ;; make sure it keeps its promises. \n\
This way, we do not have to save/move the registers around into
temporaries; we can safely use them straight away.
If you want to check that the allocation was right; then
check the equalities in the first comment. It should say
"r13=r13, r12=r12, r11=r11" */
__asm__ volatile (" \n\
;; Check that the register asm declaration got right. \n\
;; The GCC manual says it will work, but there *has* been bugs. \n\
.ifnc %0-%1-%4,$r13-$r12-$r11 \n\ .ifnc %0-%1-%4,$r13-$r12-$r11 \n\
.err \n\ .error \"GCC reg alloc bug: %0-%1-%4 != $r13-$r12-$r11\" \n\
.endif \n\ .endif \n\
\n\ \n\
;; Save the registers we'll clobber in the movem process \n\ ;; Save the registers we'll clobber in the movem process \n\
;; on the stack. Don't mention them to gcc, it will only be \n\ ;; on the stack. Don't mention them to gcc, it will only be \n\
;; upset. \n\ ;; upset. \n\
subq 11*4,$sp \n\ subq 11*4,sp \n\
movem $r10,[$sp] \n\ movem r10,[sp] \n\
\n\ \n\
move.d $r11,$r0 \n\ move.d r11,r0 \n\
move.d $r11,$r1 \n\ move.d r11,r1 \n\
move.d $r11,$r2 \n\ move.d r11,r2 \n\
move.d $r11,$r3 \n\ move.d r11,r3 \n\
move.d $r11,$r4 \n\ move.d r11,r4 \n\
move.d $r11,$r5 \n\ move.d r11,r5 \n\
move.d $r11,$r6 \n\ move.d r11,r6 \n\
move.d $r11,$r7 \n\ move.d r11,r7 \n\
move.d $r11,$r8 \n\ move.d r11,r8 \n\
move.d $r11,$r9 \n\ move.d r11,r9 \n\
move.d $r11,$r10 \n\ move.d r11,r10 \n\
\n\ \n\
;; Now we've got this: \n\ ;; Now we've got this: \n\
;; r13 - dst \n\ ;; r13 - dst \n\
;; r12 - n \n\ ;; r12 - n \n\
\n\ \n\
;; Update n for the first loop \n\ ;; Update n for the first loop \n\
subq 12*4,$r12 \n\ subq 12*4,r12 \n\
0: \n\ 0: \n\
subq 12*4,$r12 \n\ "
#ifdef __arch_common_v10_v32
/* Cater to branch offset difference between v32 and v10. We
assume the branch below has an 8-bit offset. */
" setf\n"
#endif
" subq 12*4,r12 \n\
bge 0b \n\ bge 0b \n\
movem $r11,[$r13+] \n\ movem r11,[r13+] \n\
\n\ \n\
addq 12*4,$r12 ;; compensate for last loop underflowing n \n\ ;; Compensate for last loop underflowing n. \n\
addq 12*4,r12 \n\
\n\ \n\
;; Restore registers from stack \n\ ;; Restore registers from stack. \n\
movem [$sp+],$r10" movem [sp+],r10"
/* Outputs */ : "=r" (dst), "=r" (n) /* Outputs. */
/* Inputs */ : "0" (dst), "1" (n), "r" (lc)); : "=r" (dst), "=r" (n)
}
/* Either we directly starts copying, using dword copying /* Inputs. */
in a loop, or we copy as much as possible with 'movem' : "0" (dst), "1" (n), "r" (lc));
and then the last block (<44 bytes) is copied here. }
This will work since 'movem' will have updated src,dst,n. */
while ( n >= 16 ) /* An ad-hoc unroll, used for 4*12-1..16 bytes. */
while (n >= 16)
{ {
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
n -= 16; n -= 16;
} }
/* A switch() is definitely the fastest although it takes a LOT of code.
* Particularly if you inline code this.
*/
switch (n) switch (n)
{ {
case 0: case 0:
break; break;
case 1: case 1:
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 2: case 2:
*(short*)dst = (short) lc; *(short *) dst = (short) lc;
break; break;
case 3: case 3:
*((short*)dst)++ = (short) lc; *(short *) dst = (short) lc; dst += 2;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 4: case 4:
*((long*)dst)++ = lc; *(long *) dst = lc;
break; break;
case 5: case 5:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 6: case 6:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(short*)dst = (short) lc; *(short *) dst = (short) lc;
break; break;
case 7: case 7:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((short*)dst)++ = (short) lc; *(short *) dst = (short) lc; dst += 2;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 8: case 8:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc;
break; break;
case 9: case 9:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 10: case 10:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(short*)dst = (short) lc; *(short *) dst = (short) lc;
break; break;
case 11: case 11:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((short*)dst)++ = (short) lc; *(short *) dst = (short) lc; dst += 2;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 12: case 12:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc;
break; break;
case 13: case 13:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
case 14: case 14:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*(short*)dst = (short) lc; *(short *) dst = (short) lc;
break; break;
case 15: case 15:
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((long*)dst)++ = lc; *(long *) dst = lc; dst += 4;
*((short*)dst)++ = (short) lc; *(short *) dst = (short) lc; dst += 2;
*(char*)dst = (char) lc; *dst = (char) lc;
break; break;
} }
} }
return return_dst; /* destination pointer. */ return return_dst;
} /* memset() */ }
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment