On Fri, 16 Mar 2007, Davide Libenzi wrote:
>
> I cannot measure any sensible difference between the two.
I'm using your previous patch (is it the same?) along with the additional
patch appended.
And yes, it's not hugely faster, but I seem to see *some* difference: this
is the real-time of ten runs of
time git log drivers/usb/ > /dev/null
Before:
0m2.673s
0m2.476s
0m2.603s
0m2.576s
0m2.625s
0m2.628s
0m2.493s
0m2.696s
0m2.525s
0m2.575s
After:
0m2.639s
0m2.519s
0m2.454s
0m2.604s
0m2.499s
0m2.497s
0m2.506s
0m2.394s
0m2.409s
0m2.562s
ie after I actually get under 2.4s once, and under 2.5s most of the time,
while before it was under 2.5s just twice, and mostly in the 2.6s..
(I did end up adding the "-g", but I trust that doesn't make things
*faster*. Generally gcc is good at not actually changing code generation
based on -g)
But yeah, not very impressive changes. We're talking *maybe* 0.1s out of
2.5, so potentially about 4% of total time but more likely about 2-3%, and
it's clearly mostly in the noise. And inflate() is still at 16%, and
inflate_fast obviously got no faster.
The nice part is that the instruction-level profile for inflate() got more
interesting. Instead of clearly peaking at the silly indirect jump, the
peak now seems to be a specific path through the thing. I've not decoded
it fully yet, but it seems to be mostly the LEN/LIT cases:
file inflate.c, line 942.
file inflate.c, line 942.
file inflate.c, line 949.
file inflate.c, line 949.
file inflate.c, line 949.
file inflate.c, line 949.
file inflate.c, line 949.
file inflate.c, line 950.
file inflate.c, line 950.
file inflate.c, line 951.
file inflate.c, line 951.
file inflate.c, line 951.
file inflate.c, line 951.
file inflate.c, line 951.
file inflate.c, line 949.
file inflate.c, line 949.
file inflate.c, line 950.
file inflate.c, line 950.
file inflate.c, line 953.
file inflate.c, line 953.
file inflate.c, line 953.
file inflate.c, line 969.
file inflate.c, line 1058.
file inflate.c, line 1059.
file inflate.c, line 1061.
file inflate.c, line 884.
file inflate.c, line 963.
file inflate.c, line 964.
(those are the line numbers *after* applying my patch for where the
hotpoints are: the same line-number showing up multiple times is just
because several hot instructions came from there and got spread out)
Linus
---
diff --git a/Makefile b/Makefile
index 2fd6e45..d8e9ff4 100644
--- a/Makefile
+++ b/Makefile
@@ -18,7 +18,7 @@
CC=cc
-CFLAGS=-O
+CFLAGS=-O -g
#CFLAGS=-O -DMAX_WBITS=14 -DMAX_MEM_LEVEL=7
#CFLAGS=-g -DDEBUG
#CFLAGS=-O3 -Wall -Wwrite-strings -Wpointer-arith -Wconversion \
diff --git a/inflate.c b/inflate.c
index 190c642..3d41d6f 100644
--- a/inflate.c
+++ b/inflate.c
@@ -568,7 +568,7 @@ int flush;
unsigned char FAR *next; /* next input */
unsigned char FAR *put; /* next output */
unsigned have, left; /* available input and output */
- unsigned long hold; /* bit buffer */
+ unsigned long hold, old_hold;/* bit buffer */
unsigned bits; /* bits in bit buffer */
unsigned in, out; /* save starting available input and output */
unsigned copy; /* number of stored or match bytes to copy */
@@ -631,8 +631,11 @@ int flush;
state->dmax = 1U << len;
Tracev((stderr, "inflate: zlib header ok\n"));
strm->adler = state->check = adler32(0L, Z_NULL, 0);
- state->mode = hold & 0x200 ? DICTID : TYPE;
+ old_hold = hold;
INITBITS();
+ if (old_hold & 0x200)
+ STATE_CHANGE(DICTID);
+ STATE_CHANGE(TYPE);
break;
#ifdef GUNZIP
CASE_DECL(FLAGS)
@@ -817,7 +820,7 @@ int flush;
state->mode = COPY;
CASE_DECL(COPY)
copy = state->length;
- if (copy) {
+ while (copy) {
if (copy > have) copy = have;
if (copy > left) copy = left;
if (copy == 0) goto inf_leave;
@@ -826,8 +829,8 @@ int flush;
next += copy;
left -= copy;
put += copy;
- state->length -= copy;
- break;
+ copy = state->length - copy;
+ state->length = copy;
}
Tracev((stderr, "inflate: stored end\n"));
STATE_CHANGE(TYPE);
-
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html