Patch from "David J. Fiddes" <D.J@fiddes.surfaid.org>:

I've fixed a few minor probs with the optimised version that Eric put
    together for me the other day and sent the fixes back to him. Provided he
    doesn't have a problem with it we've got a pretty solid in_cksum for the
    ColdFire as well as straight m68k. I've enclosed my updated in_cksum_m68k.c

    At the moment my own bottlenecks are elsewhere...as my driver is pulling
    16bit data chunks through a libchip-esq access routine from the chip which
    for a polled I/O device is never going to be quick.
This commit is contained in:
Joel Sherrill
1998-09-11 13:12:04 +00:00
parent 0ea9829786
commit b39563695d
5 changed files with 185 additions and 630 deletions

View File

@@ -37,7 +37,11 @@
#include <sys/param.h>
#include <sys/mbuf.h>
#if (defined(__GNUC__) && (defined(__mc68000__) || defined(__m68k__)) && (!__mcf5200__))
#if (defined (__mcf5200__))
# define IS_COLDFIRE 1
#else
# define IS_COLDFIRE 0
#endif
#define REDUCE { sum = (sum & 0xFFFF) + (sum >> 16); if (sum > 0xFFFF) sum -= 0xFFFF; }
@@ -109,15 +113,21 @@ in_cksum(m, len)
* It should work for all 68k family members.
*/
{
unsigned long tcnt = mlen, t1;
unsigned long tcnt = mlen, t1, t2;
__asm__ volatile (
"movel %2,%3\n\t"
"lsrl #6,%2 | count/64 = # loop traversals\n\t"
"andl #0x3c,%3 | Then find fractions of a chunk\n\t"
"negl %3\n\t | Each long uses 4 instruction bytes\n\t"
#if IS_COLDFIRE
"move %%cc,%4 | Move the CC into a temp reg\n\t"
"andil #0xf,%4 | Clear X (extended carry flag)\n\t"
"move %4,%%cc | Move the modified value back to the cc\n\t"
#else
"andi #0xf,%%cc | Clear X (extended carry flag)\n\t"
#endif
"jmp %%pc@(lcsum2_lbl-.-2:b,%3) | Jump into loop\n"
"lcsum1_lbl: | Begin inner loop...\n\t"
"lcsum1_lbl: | Begin inner loop...\n\t"
"movel %1@+,%3 | 0: Fetch 32-bit word\n\t"
"addxl %3,%0 | Add word + previous carry\n\t"
"movel %1@+,%3 | 1: Fetch 32-bit word\n\t"
@@ -150,14 +160,31 @@ in_cksum(m, len)
"addxl %3,%0 | Add word + previous carry\n\t"
"movel %1@+,%3 | F: Fetch 32-bit word\n\t"
"addxl %3,%0 | Add word + previous carry\n"
"lcsum2_lbl:\n\tdbf %2,lcsum1_lbl | (NB- dbra doesn't affect X)\n\t"
"movel %0,%3 | Fold 32 bit sum to 16 bits\n\t"
"swap %3 | (NB- swap doesn't affect X)\n\t"
"addxw %3,%0 |\n\t"
"moveq #0,%3 | Add in last carry\n\t"
"addxw %3,%0 |\n\t"
"lcsum2_lbl: | End of unrolled loop\n\t"
#if IS_COLDFIRE
"moveq #0,%3 | Add in last carry\n\t"
"addxl %3,%0 |\n\t"
"subql #1,%2 | Update loop count\n\t"
"bplb lcsum1_lbl | Loop (with X clear) if not done\n\t"
"movel #0xffff,%2 | Get word mask\n\t"
"movel %0,%3 | Fold 32 bit sum to 16 bits\n\t"
"swap %3 |\n\t"
"andl %2,%0 |\n\t"
"andl %2,%3 |\n\t"
"addl %3,%0 |\n\t"
"movel %0,%3 | Add in last carry\n\t"
"swap %3 |\n\t"
"addl %3,%0 |\n\t"
#else
"dbf %2,lcsum1_lbl | (NB- dbf doesn't affect X)\n\t"
"movel %0,%3 | Fold 32 bit sum to 16 bits\n\t"
"swap %3 | (NB- swap doesn't affect X)\n\t"
"addxw %3,%0 |\n\t"
"moveq #0,%3 | Add in last carry\n\t"
"addxw %3,%0 |\n\t"
#endif
"andl #0xffff,%0 | Mask to 16-bit sum\n" :
"=d" (sum), "=a" (w), "=d" (tcnt) , "=d" (t1) :
"=d" (sum), "=a" (w), "=d" (tcnt) , "=d" (t1), "=d" (t2) :
"0" (sum), "1" (w), "2" (tcnt) :
"cc", "memory");
}
@@ -193,119 +220,3 @@ in_cksum(m, len)
REDUCE;
return (~sum & 0xffff);
}
#else
/*
* Checksum routine for Internet Protocol family headers (Portable Version).
*
* This routine is very heavily used in the network
* code and should be modified for each CPU to be as fast as possible.
*/
#define ADDCARRY(x) (x > 65535 ? x -= 65535 : x)
#define REDUCE \
{l_util.l = sum; sum = l_util.s[0] + l_util.s[1]; ADDCARRY(sum);}
int
in_cksum(m, len)
register struct mbuf *m;
register int len;
{
register u_short *w;
register int sum = 0;
register int mlen = 0;
int byte_swapped = 0;
union {
char c[2];
u_short s;
} s_util;
union {
u_short s[2];
long l;
} l_util;
for (;m && len; m = m->m_next) {
if (m->m_len == 0)
continue;
w = mtod(m, u_short *);
if (mlen == -1) {
/*
* The first byte of this mbuf is the continuation
* of a word spanning between this mbuf and the
* last mbuf.
*
* s_util.c[0] is already saved when scanning previous
* mbuf.
*/
s_util.c[1] = *(char *)w;
sum += s_util.s;
w = (u_short *)((char *)w + 1);
mlen = m->m_len - 1;
len--;
} else
mlen = m->m_len;
if (len < mlen)
mlen = len;
len -= mlen;
/*
* Force to even boundary.
*/
if ((1 & (int) w) && (mlen > 0)) {
REDUCE;
sum <<= 8;
s_util.c[0] = *(u_char *)w;
w = (u_short *)((char *)w + 1);
mlen--;
byte_swapped = 1;
}
/*
* Unroll the loop to make overhead from
* branches &c small.
*/
while ((mlen -= 32) >= 0) {
sum += w[0]; sum += w[1]; sum += w[2]; sum += w[3];
sum += w[4]; sum += w[5]; sum += w[6]; sum += w[7];
sum += w[8]; sum += w[9]; sum += w[10]; sum += w[11];
sum += w[12]; sum += w[13]; sum += w[14]; sum += w[15];
w += 16;
}
mlen += 32;
while ((mlen -= 8) >= 0) {
sum += w[0]; sum += w[1]; sum += w[2]; sum += w[3];
w += 4;
}
mlen += 8;
if (mlen == 0 && byte_swapped == 0)
continue;
REDUCE;
while ((mlen -= 2) >= 0) {
sum += *w++;
}
if (byte_swapped) {
REDUCE;
sum <<= 8;
byte_swapped = 0;
if (mlen == -1) {
s_util.c[1] = *(char *)w;
sum += s_util.s;
mlen = 0;
} else
mlen = -1;
} else if (mlen == -1)
s_util.c[0] = *(char *)w;
}
if (len)
printf("cksum: out of data\n");
if (mlen == -1) {
/* The last mbuf has odd # of bytes. Follow the
standard (the odd byte may be shifted left by 8 bits
or not as determined by endian-ness of the machine) */
s_util.c[1] = 0;
sum += s_util.s;
}
REDUCE;
return (~sum & 0xffff);
}
#endif

View File

@@ -37,7 +37,11 @@
#include <sys/param.h>
#include <sys/mbuf.h>
#if (defined(__GNUC__) && (defined(__mc68000__) || defined(__m68k__)) && (!__mcf5200__))
#if (defined (__mcf5200__))
# define IS_COLDFIRE 1
#else
# define IS_COLDFIRE 0
#endif
#define REDUCE { sum = (sum & 0xFFFF) + (sum >> 16); if (sum > 0xFFFF) sum -= 0xFFFF; }
@@ -109,15 +113,21 @@ in_cksum(m, len)
* It should work for all 68k family members.
*/
{
unsigned long tcnt = mlen, t1;
unsigned long tcnt = mlen, t1, t2;
__asm__ volatile (
"movel %2,%3\n\t"
"lsrl #6,%2 | count/64 = # loop traversals\n\t"
"andl #0x3c,%3 | Then find fractions of a chunk\n\t"
"negl %3\n\t | Each long uses 4 instruction bytes\n\t"
#if IS_COLDFIRE
"move %%cc,%4 | Move the CC into a temp reg\n\t"
"andil #0xf,%4 | Clear X (extended carry flag)\n\t"
"move %4,%%cc | Move the modified value back to the cc\n\t"
#else
"andi #0xf,%%cc | Clear X (extended carry flag)\n\t"
#endif
"jmp %%pc@(lcsum2_lbl-.-2:b,%3) | Jump into loop\n"
"lcsum1_lbl: | Begin inner loop...\n\t"
"lcsum1_lbl: | Begin inner loop...\n\t"
"movel %1@+,%3 | 0: Fetch 32-bit word\n\t"
"addxl %3,%0 | Add word + previous carry\n\t"
"movel %1@+,%3 | 1: Fetch 32-bit word\n\t"
@@ -150,14 +160,31 @@ in_cksum(m, len)
"addxl %3,%0 | Add word + previous carry\n\t"
"movel %1@+,%3 | F: Fetch 32-bit word\n\t"
"addxl %3,%0 | Add word + previous carry\n"
"lcsum2_lbl:\n\tdbf %2,lcsum1_lbl | (NB- dbra doesn't affect X)\n\t"
"movel %0,%3 | Fold 32 bit sum to 16 bits\n\t"
"swap %3 | (NB- swap doesn't affect X)\n\t"
"addxw %3,%0 |\n\t"
"moveq #0,%3 | Add in last carry\n\t"
"addxw %3,%0 |\n\t"
"lcsum2_lbl: | End of unrolled loop\n\t"
#if IS_COLDFIRE
"moveq #0,%3 | Add in last carry\n\t"
"addxl %3,%0 |\n\t"
"subql #1,%2 | Update loop count\n\t"
"bplb lcsum1_lbl | Loop (with X clear) if not done\n\t"
"movel #0xffff,%2 | Get word mask\n\t"
"movel %0,%3 | Fold 32 bit sum to 16 bits\n\t"
"swap %3 |\n\t"
"andl %2,%0 |\n\t"
"andl %2,%3 |\n\t"
"addl %3,%0 |\n\t"
"movel %0,%3 | Add in last carry\n\t"
"swap %3 |\n\t"
"addl %3,%0 |\n\t"
#else
"dbf %2,lcsum1_lbl | (NB- dbf doesn't affect X)\n\t"
"movel %0,%3 | Fold 32 bit sum to 16 bits\n\t"
"swap %3 | (NB- swap doesn't affect X)\n\t"
"addxw %3,%0 |\n\t"
"moveq #0,%3 | Add in last carry\n\t"
"addxw %3,%0 |\n\t"
#endif
"andl #0xffff,%0 | Mask to 16-bit sum\n" :
"=d" (sum), "=a" (w), "=d" (tcnt) , "=d" (t1) :
"=d" (sum), "=a" (w), "=d" (tcnt) , "=d" (t1), "=d" (t2) :
"0" (sum), "1" (w), "2" (tcnt) :
"cc", "memory");
}
@@ -193,119 +220,3 @@ in_cksum(m, len)
REDUCE;
return (~sum & 0xffff);
}
#else
/*
* Checksum routine for Internet Protocol family headers (Portable Version).
*
* This routine is very heavily used in the network
* code and should be modified for each CPU to be as fast as possible.
*/
#define ADDCARRY(x) (x > 65535 ? x -= 65535 : x)
#define REDUCE \
{l_util.l = sum; sum = l_util.s[0] + l_util.s[1]; ADDCARRY(sum);}
int
in_cksum(m, len)
register struct mbuf *m;
register int len;
{
register u_short *w;
register int sum = 0;
register int mlen = 0;
int byte_swapped = 0;
union {
char c[2];
u_short s;
} s_util;
union {
u_short s[2];
long l;
} l_util;
for (;m && len; m = m->m_next) {
if (m->m_len == 0)
continue;
w = mtod(m, u_short *);
if (mlen == -1) {
/*
* The first byte of this mbuf is the continuation
* of a word spanning between this mbuf and the
* last mbuf.
*
* s_util.c[0] is already saved when scanning previous
* mbuf.
*/
s_util.c[1] = *(char *)w;
sum += s_util.s;
w = (u_short *)((char *)w + 1);
mlen = m->m_len - 1;
len--;
} else
mlen = m->m_len;
if (len < mlen)
mlen = len;
len -= mlen;
/*
* Force to even boundary.
*/
if ((1 & (int) w) && (mlen > 0)) {
REDUCE;
sum <<= 8;
s_util.c[0] = *(u_char *)w;
w = (u_short *)((char *)w + 1);
mlen--;
byte_swapped = 1;
}
/*
* Unroll the loop to make overhead from
* branches &c small.
*/
while ((mlen -= 32) >= 0) {
sum += w[0]; sum += w[1]; sum += w[2]; sum += w[3];
sum += w[4]; sum += w[5]; sum += w[6]; sum += w[7];
sum += w[8]; sum += w[9]; sum += w[10]; sum += w[11];
sum += w[12]; sum += w[13]; sum += w[14]; sum += w[15];
w += 16;
}
mlen += 32;
while ((mlen -= 8) >= 0) {
sum += w[0]; sum += w[1]; sum += w[2]; sum += w[3];
w += 4;
}
mlen += 8;
if (mlen == 0 && byte_swapped == 0)
continue;
REDUCE;
while ((mlen -= 2) >= 0) {
sum += *w++;
}
if (byte_swapped) {
REDUCE;
sum <<= 8;
byte_swapped = 0;
if (mlen == -1) {
s_util.c[1] = *(char *)w;
sum += s_util.s;
mlen = 0;
} else
mlen = -1;
} else if (mlen == -1)
s_util.c[0] = *(char *)w;
}
if (len)
printf("cksum: out of data\n");
if (mlen == -1) {
/* The last mbuf has odd # of bytes. Follow the
standard (the odd byte may be shifted left by 8 bits
or not as determined by endian-ness of the machine) */
s_util.c[1] = 0;
sum += s_util.s;
}
REDUCE;
return (~sum & 0xffff);
}
#endif

View File

@@ -37,7 +37,11 @@
#include <sys/param.h>
#include <sys/mbuf.h>
#if (defined(__GNUC__) && (defined(__mc68000__) || defined(__m68k__)) && (!__mcf5200__))
#if (defined (__mcf5200__))
# define IS_COLDFIRE 1
#else
# define IS_COLDFIRE 0
#endif
#define REDUCE { sum = (sum & 0xFFFF) + (sum >> 16); if (sum > 0xFFFF) sum -= 0xFFFF; }
@@ -109,15 +113,21 @@ in_cksum(m, len)
* It should work for all 68k family members.
*/
{
unsigned long tcnt = mlen, t1;
unsigned long tcnt = mlen, t1, t2;
__asm__ volatile (
"movel %2,%3\n\t"
"lsrl #6,%2 | count/64 = # loop traversals\n\t"
"andl #0x3c,%3 | Then find fractions of a chunk\n\t"
"negl %3\n\t | Each long uses 4 instruction bytes\n\t"
#if IS_COLDFIRE
"move %%cc,%4 | Move the CC into a temp reg\n\t"
"andil #0xf,%4 | Clear X (extended carry flag)\n\t"
"move %4,%%cc | Move the modified value back to the cc\n\t"
#else
"andi #0xf,%%cc | Clear X (extended carry flag)\n\t"
#endif
"jmp %%pc@(lcsum2_lbl-.-2:b,%3) | Jump into loop\n"
"lcsum1_lbl: | Begin inner loop...\n\t"
"lcsum1_lbl: | Begin inner loop...\n\t"
"movel %1@+,%3 | 0: Fetch 32-bit word\n\t"
"addxl %3,%0 | Add word + previous carry\n\t"
"movel %1@+,%3 | 1: Fetch 32-bit word\n\t"
@@ -150,14 +160,31 @@ in_cksum(m, len)
"addxl %3,%0 | Add word + previous carry\n\t"
"movel %1@+,%3 | F: Fetch 32-bit word\n\t"
"addxl %3,%0 | Add word + previous carry\n"
"lcsum2_lbl:\n\tdbf %2,lcsum1_lbl | (NB- dbra doesn't affect X)\n\t"
"movel %0,%3 | Fold 32 bit sum to 16 bits\n\t"
"swap %3 | (NB- swap doesn't affect X)\n\t"
"addxw %3,%0 |\n\t"
"moveq #0,%3 | Add in last carry\n\t"
"addxw %3,%0 |\n\t"
"lcsum2_lbl: | End of unrolled loop\n\t"
#if IS_COLDFIRE
"moveq #0,%3 | Add in last carry\n\t"
"addxl %3,%0 |\n\t"
"subql #1,%2 | Update loop count\n\t"
"bplb lcsum1_lbl | Loop (with X clear) if not done\n\t"
"movel #0xffff,%2 | Get word mask\n\t"
"movel %0,%3 | Fold 32 bit sum to 16 bits\n\t"
"swap %3 |\n\t"
"andl %2,%0 |\n\t"
"andl %2,%3 |\n\t"
"addl %3,%0 |\n\t"
"movel %0,%3 | Add in last carry\n\t"
"swap %3 |\n\t"
"addl %3,%0 |\n\t"
#else
"dbf %2,lcsum1_lbl | (NB- dbf doesn't affect X)\n\t"
"movel %0,%3 | Fold 32 bit sum to 16 bits\n\t"
"swap %3 | (NB- swap doesn't affect X)\n\t"
"addxw %3,%0 |\n\t"
"moveq #0,%3 | Add in last carry\n\t"
"addxw %3,%0 |\n\t"
#endif
"andl #0xffff,%0 | Mask to 16-bit sum\n" :
"=d" (sum), "=a" (w), "=d" (tcnt) , "=d" (t1) :
"=d" (sum), "=a" (w), "=d" (tcnt) , "=d" (t1), "=d" (t2) :
"0" (sum), "1" (w), "2" (tcnt) :
"cc", "memory");
}
@@ -193,119 +220,3 @@ in_cksum(m, len)
REDUCE;
return (~sum & 0xffff);
}
#else
/*
* Checksum routine for Internet Protocol family headers (Portable Version).
*
* This routine is very heavily used in the network
* code and should be modified for each CPU to be as fast as possible.
*/
#define ADDCARRY(x) (x > 65535 ? x -= 65535 : x)
#define REDUCE \
{l_util.l = sum; sum = l_util.s[0] + l_util.s[1]; ADDCARRY(sum);}
int
in_cksum(m, len)
register struct mbuf *m;
register int len;
{
register u_short *w;
register int sum = 0;
register int mlen = 0;
int byte_swapped = 0;
union {
char c[2];
u_short s;
} s_util;
union {
u_short s[2];
long l;
} l_util;
for (;m && len; m = m->m_next) {
if (m->m_len == 0)
continue;
w = mtod(m, u_short *);
if (mlen == -1) {
/*
* The first byte of this mbuf is the continuation
* of a word spanning between this mbuf and the
* last mbuf.
*
* s_util.c[0] is already saved when scanning previous
* mbuf.
*/
s_util.c[1] = *(char *)w;
sum += s_util.s;
w = (u_short *)((char *)w + 1);
mlen = m->m_len - 1;
len--;
} else
mlen = m->m_len;
if (len < mlen)
mlen = len;
len -= mlen;
/*
* Force to even boundary.
*/
if ((1 & (int) w) && (mlen > 0)) {
REDUCE;
sum <<= 8;
s_util.c[0] = *(u_char *)w;
w = (u_short *)((char *)w + 1);
mlen--;
byte_swapped = 1;
}
/*
* Unroll the loop to make overhead from
* branches &c small.
*/
while ((mlen -= 32) >= 0) {
sum += w[0]; sum += w[1]; sum += w[2]; sum += w[3];
sum += w[4]; sum += w[5]; sum += w[6]; sum += w[7];
sum += w[8]; sum += w[9]; sum += w[10]; sum += w[11];
sum += w[12]; sum += w[13]; sum += w[14]; sum += w[15];
w += 16;
}
mlen += 32;
while ((mlen -= 8) >= 0) {
sum += w[0]; sum += w[1]; sum += w[2]; sum += w[3];
w += 4;
}
mlen += 8;
if (mlen == 0 && byte_swapped == 0)
continue;
REDUCE;
while ((mlen -= 2) >= 0) {
sum += *w++;
}
if (byte_swapped) {
REDUCE;
sum <<= 8;
byte_swapped = 0;
if (mlen == -1) {
s_util.c[1] = *(char *)w;
sum += s_util.s;
mlen = 0;
} else
mlen = -1;
} else if (mlen == -1)
s_util.c[0] = *(char *)w;
}
if (len)
printf("cksum: out of data\n");
if (mlen == -1) {
/* The last mbuf has odd # of bytes. Follow the
standard (the odd byte may be shifted left by 8 bits
or not as determined by endian-ness of the machine) */
s_util.c[1] = 0;
sum += s_util.s;
}
REDUCE;
return (~sum & 0xffff);
}
#endif