Aren/NetBSD
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

rewrite and hand-tune copy_gap16 loops:

Browse Source 
* hoist code to deal with  fragments of a 16-byte chunk outside
    main copy loop.
  * over 98% of dynamic calls are 2-byte-algnied but not 4-byte-aligned
   (due to 14-byte Ether headers).  kernel bcopy is poorly tune for this.
   Replace bcopy() with  inline code tuned to  minimize accesses
   to DMA buffers, which aren't uncached on mips.

Tested on 5000/240 (3MAXPLUS) and 3000/700 (sandpiper) (mjacob@feral.com)
This commit is contained in:
jonathan 1997-08-26 01:27:12 +00:00
parent 9ffe59d049
commit 6cd95f2a55
1 changed files with 124 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

130
sys/dev/tc/if_le_ioasic.c
View File

@ -1,4 +1,4 @@
/* $NetBSD: if_le_ioasic.c,v 1.7 1997/07/22 03:44:30 jonathan Exp $ */
/* $NetBSD: if_le_ioasic.c,v 1.8 1997/08/26 01:27:12 jonathan Exp $ */
/*
* Copyright (c) 1996 Carnegie-Mellon University.
@ -204,19 +204,79 @@ le_ioasic_copytobuf_gap16(sc, fromv, boff, len)
volatile caddr_t buf = sc->sc_mem;
register caddr_t from = fromv;
register caddr_t bptr;
register int xfer;
bptr = buf + ((boff << 1) & ~0x1f);
boff &= 0xf;
/*
* Dispose of boff so destination of subsequent copies is
* 16-byte aligned.
*/
if (boff) {
register int xfer;
xfer = min(len, 16 - boff);
while (len > 0) {
bcopy(from, bptr + boff, xfer);
from += xfer;
bptr += 32;
boff = 0;
len -= xfer;
xfer = min(len, 16);
}
/* Destination of copies is now 16-byte aligned. */
if (len >= 16)
switch ((u_long)from & (sizeof(u_int32_t) -1)) {
case 2:
/* Ethernet headers make this the dominant case. */
do {
register u_int32_t *dst = (u_int32_t*)bptr;
register u_int16_t t0;
register u_int32_t t1, t2, t3, t4;
/* read from odd-16-bit-aligned, cached src */
t0 = *(u_int16_t*)from;
t1 = *(u_int32_t*)(from+2);
t2 = *(u_int32_t*)(from+6);
t3 = *(u_int32_t*)(from+10);
t4 = *(u_int16_t*)(from+14);
/* DMA buffer is uncached on mips */
dst[0] = t0 | (t1 << 16);
dst[1] = (t1 >> 16) | (t2 << 16);
dst[2] = (t2 >> 16) | (t3 << 16);
dst[3] = (t3 >> 16) | (t4 << 16);
from += 16;
bptr += 32;
len -= 16;
} while (len >= 16);
break;
case 0:
do {
register u_int32_t *src = (u_int32_t*)from;
register u_int32_t *dst = (u_int32_t*)bptr;
register u_int32_t t0, t1, t2, t3;
t0 = src[0]; t1 = src[1]; t2 = src[2]; t3 = src[3];
dst[0] = t0; dst[1] = t1; dst[2] = t2; dst[3] = t3;
from += 16;
bptr += 32;
len -= 16;
} while (len >= 16);
break;
default:
/* Does odd-aligned case ever happen? */
do {
bcopy(from, bptr, 16);
from += 16;
bptr += 32;
len -= 16;
} while (len >= 16);
break;
}
if (len)
bcopy(from, bptr, len);
}
void
@ -228,19 +288,71 @@ le_ioasic_copyfrombuf_gap16(sc, tov, boff, len)
volatile caddr_t buf = sc->sc_mem;
register caddr_t to = tov;
register caddr_t bptr;
register int xfer;
bptr = buf + ((boff << 1) & ~0x1f);
boff &= 0xf;
/* Dispose of boff. source of copy is subsequently 16-byte aligned. */
if (boff) {
register int xfer;
xfer = min(len, 16 - boff);
while (len > 0) {
bcopy(bptr+boff, to, xfer);
to += xfer;
bptr += 32;
boff = 0;
len -= xfer;
xfer = min(len, 16);
}
if (len >= 16)
switch ((u_long)to & (sizeof(u_int32_t) -1)) {
case 2:
/*
* to is aligned to an odd 16-bit boundary. Ethernet headers
* make this the dominant case (98% or more).
*/
do {
register u_int32_t *src = (u_int32_t*)bptr;
register u_int32_t t0, t1, t2, t3;
/* read from uncached aligned DMA buf */
t0 = src[0]; t1 = src[1]; t2 = src[2]; t3 = src[3];
/* write to odd-16-bit-word aligned dst */
*(u_int16_t *) (to+0) = (u_short) t0;
*(u_int32_t *) (to+2) = (t0 >> 16) | (t1 << 16);
*(u_int32_t *) (to+6) = (t1 >> 16) | (t2 << 16);
*(u_int32_t *) (to+10) = (t2 >> 16) | (t3 << 16);
*(u_int16_t *) (to+14) = (t3 >> 16);
bptr += 32;
to += 16;
len -= 16;
} while (len > 16);
break;
case 0:
/* 32-bit aligned aligned copy. Rare. */
do {
register u_int32_t *src = (u_int32_t*)bptr;
register u_int32_t *dst = (u_int32_t*)to;
register u_int32_t t0, t1, t2, t3;
t0 = src[0]; t1 = src[1]; t2 = src[2]; t3 = src[3];
dst[0] = t0; dst[1] = t1; dst[2] = t2; dst[3] = t3;
to += 16;
bptr += 32;
len -= 16;
} while (len > 16);
break;
/* XXX Does odd-byte-aligned case ever happen? */
default:
do {
bcopy(bptr, to, 16);
to += 16;
bptr += 32;
len -= 16;
} while (len > 16);
break;
}
if (len)
bcopy(bptr, to, len);
}
void