Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

[Peter Stephenson]

On Mon, 19 Jun 2017 17:16:01 +0100
Stephane Chazelas <stephane.chazelas@gmail.com> wrote:
> That defeats a benefit of stdio saving read() systems calls by
> reading in chunk if we end up doing one system call per byte
> anyway.

Yes, if it's line buffered we should ideally only be unblocking
interrupts once around reading the line, which will dominate the
timing.

How about something like this?  As far as I can tell, fgets is designed
from the ground up as Gets Done Properly, so if you have it on your
system it will work correctly.  I can't think of a case where this
wouldn't do the right thing --- fgets will read at most one line and if
it does we were going to get the big STDIO overhead at that point
anyway.

pws

diff --git a/Src/input.c b/Src/input.c
index 92abaec..03d6476 100644
--- a/Src/input.c
+++ b/Src/input.c
@@ -147,6 +147,53 @@ shingetline(void)
     for (;;) {
 	winch_unblock();
 	dont_queue_signals();
+#ifdef HAVE_FGETS
+	do {
+	    errno = 0;
+	    p = fgets(buf, BUFSIZ, bshin);
+	} while (!p && errno == EINTR);
+	winch_block();
+	if (!p) {
+	    restore_queue_signals(q);
+	    return NULL;
+	}
+	c = 0;
+	while (*p) {
+	    if (imeta(STOUC(*p++)))
+		c++;
+	}
+	if (p > buf) {
+	    /* p is pointing to '\0' */
+	    ptrdiff_t nread = p - buf;
+	    queue_signals();
+	    line = zrealloc(line, ll + c + nread + 1);
+	    if (c) {
+		char *dest = line + ll;
+		char *src = buf;
+		while (src < p) {
+		    if (imeta(STOUC(*src))) {
+			*dest++ = Meta;
+			*dest++ = STOUC(*src++) ^ 32;
+		    } else
+			*dest++ = *src++;
+		}
+		*dest++ = '\0';
+	    } else {
+		/* copy null */
+		memcpy(line + ll, buf, nread + 1);
+	    }
+	    unqueue_signals();
+	    /* fgets stops at first newline but stores '\0' after */
+	    if (p[-1] == '\n') {
+		restore_queue_signals(q);
+		return line;
+	    }
+	    ll += nread;
+	} else {
+	    restore_queue_signals(q);
+	    return NULL;
+	}
+#else
 	do {
 	    errno = 0;
 	    c = fgetc(bshin);
@@ -178,6 +225,7 @@ shingetline(void)
 	    p = buf;
 	    unqueue_signals();
 	}
+#endif
     }
 }
 
diff --git a/configure.ac b/configure.ac
index 88da89e..2f19a73 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1325,7 +1325,8 @@ AC_CHECK_FUNCS(strftime strptime mktime timelocal \
 	       cygwin_conv_path \
 	       nanosleep \
 	       srand_deterministic \
-	       setutxent getutxent endutxent getutent)
+	       setutxent getutxent endutxent getutent \
+	       fgets)
 AC_FUNC_STRCOLL
 
 AH_TEMPLATE([REALPATH_ACCEPTS_NULL],

Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

[Bart Schaefer]

On Jun 19,  8:28pm, Peter Stephenson wrote:
}
} How about something like this?

If that's really OK, then I think there's a simpler patch:

diff --git a/Src/input.c b/Src/input.c
index 92abaec..73e8b26 100644
--- a/Src/input.c
+++ b/Src/input.c
@@ -144,9 +144,9 @@ shingetline(void)
     int q = queue_signal_level();
 
     p = buf;
+    winch_unblock();
+    dont_queue_signals();
     for (;;) {
-	winch_unblock();
-	dont_queue_signals();
 	do {
 	    errno = 0;
 	    c = fgetc(bshin);
@@ -176,7 +176,8 @@ shingetline(void)
 	    ll += p - buf;
 	    line[ll] = '\0';
 	    p = buf;
-	    unqueue_signals();
+	    winch_unblock();
+	    dont_queue_signals();
 	}
     }
 }

-- 
Barton E. Schaefer

Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

[Bart Schaefer]

Parsing 10,000 lines, each a colon-command followed by misc. text, and
repeated 10 times.

HEAD from git:

Src/zsh -fs  0.31s user 1.24s system 94% cpu 1.634 total
Src/zsh -fs  0.45s user 1.31s system 98% cpu 1.795 total
Src/zsh -fs  0.39s user 1.41s system 95% cpu 1.877 total
Src/zsh -fs  0.31s user 1.16s system 88% cpu 1.652 total
Src/zsh -fs  0.43s user 1.33s system 94% cpu 1.854 total
Src/zsh -fs  0.43s user 1.22s system 92% cpu 1.784 total
Src/zsh -fs  0.37s user 1.32s system 91% cpu 1.848 total
Src/zsh -fs  0.31s user 1.48s system 91% cpu 1.948 total
Src/zsh -fs  0.42s user 1.37s system 92% cpu 1.926 total
Src/zsh -fs  0.30s user 1.52s system 92% cpu 1.964 total

Peter's patch:

Src/zsh -fs  0.11s user 0.59s system 73% cpu 0.953 total
Src/zsh -fs  0.10s user 0.54s system 68% cpu 0.940 total
Src/zsh -fs  0.17s user 0.56s system 71% cpu 1.017 total
Src/zsh -fs  0.19s user 0.58s system 79% cpu 0.969 total
Src/zsh -fs  0.13s user 0.65s system 74% cpu 1.040 total
Src/zsh -fs  0.05s user 0.65s system 74% cpu 0.945 total
Src/zsh -fs  0.19s user 0.64s system 82% cpu 1.009 total
Src/zsh -fs  0.18s user 0.59s system 81% cpu 0.944 total
Src/zsh -fs  0.34s user 0.47s system 80% cpu 1.002 total
Src/zsh -fs  0.35s user 0.52s system 89% cpu 0.971 total

My "simpler" patch:

Src/zsh -fs  0.17s user 0.79s system 85% cpu 1.119 total
Src/zsh -fs  0.28s user 0.77s system 90% cpu 1.160 total
Src/zsh -fs  0.19s user 0.82s system 88% cpu 1.146 total
Src/zsh -fs  0.20s user 0.70s system 82% cpu 1.087 total
Src/zsh -fs  0.20s user 0.65s system 84% cpu 1.011 total
Src/zsh -fs  0.31s user 0.57s system 86% cpu 1.021 total
Src/zsh -fs  0.22s user 0.72s system 81% cpu 1.148 total
Src/zsh -fs  0.21s user 0.60s system 81% cpu 0.999 total
Src/zsh -fs  0.19s user 0.63s system 78% cpu 1.046 total
Src/zsh -fs  0.23s user 0.81s system 83% cpu 1.242 total

The additional speedup in PWS's patch is probably due to fewer stdio
function calls, and skip of the check for buffer overflow on every
character read.

I don't think the HAVE_FGETS + configure change is needed, there are
already at least four places where we assume fgets() is available.

Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

[Peter Stephenson]

On Mon, 19 Jun 2017 13:38:12 -0700
Bart Schaefer <schaefer@brasslantern.com> wrote:
> I don't think the HAVE_FGETS + configure change is needed, there are
> already at least four places where we assume fgets() is available.

It's been there since 2001 and I added one of them myself in 2011.

pws

Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

[Peter Stephenson]

On Mon, 19 Jun 2017 20:28:35 +0100
Peter Stephenson <p.w.stephenson@ntlworld.com> wrote:
> How about something like this?  As far as I can tell, fgets is designed
> from the ground up as Gets Done Properly, so if you have it on your
> system it will work correctly.  I can't think of a case where this
> wouldn't do the right thing --- fgets will read at most one line and if
> it does we were going to get the big STDIO overhead at that point
> anyway.

I know what the problem with it is --- we can't tell the difference
betwen a terminating '\0' and a '\0' that's part of the input stream.

pws

Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

[Bart Schaefer]

On Jun 19,  9:52pm, Peter Stephenson wrote:
}
} I know what the problem with it is --- we can't tell the difference
} betwen a terminating '\0' and a '\0' that's part of the input stream.

I'll add a comment to that effect and commit my patch from w/41322.

fgets() portability (Was: Why sourcing a file is not faster than doing a loop with eval, zle -N)

[Stephane Chazelas]

2017-06-19 21:44:14 +0100, Peter Stephenson:
> On Mon, 19 Jun 2017 13:38:12 -0700
> Bart Schaefer <schaefer@brasslantern.com> wrote:
> > I don't think the HAVE_FGETS + configure change is needed, there are
> > already at least four places where we assume fgets() is available.
> 
> It's been there since 2001 and I added one of them myself in 2011.
[...]

Out of curiosity, are there really systems that have fgetc() and not
fgets()? AFAICT, both were already there in the "original"
stdio implementation in Unix v7 in the 70s.

-- 
Stephane

Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

[Sebastian Gniazdowski]

On 20 czerwca 2017 at 01:01:26, Bart Schaefer (schaefer@brasslantern.com) wrote:
> On Jun 19, 9:52pm, Peter Stephenson wrote:
> }
> } I know what the problem with it is --- we can't tell the difference
> } betwen a terminating '\0' and a '\0' that's part of the input stream.
> 
> I'll add a comment to that effect and commit my patch from w/41322.

I have just noted a huge speed up of Zsh startup and ran bisect. I do:

repeat 10 { time /usr/local/bin/zsh -i -c exit }

and before 41322, speed is 400 ms, 350 ms when everything is compiled. With this patch, it is 220/200 ms. I thought this effect will be only for sourcing not-compiled files.

Basically this means that all the frameworks like Oh-My-Zsh can grow in size a "little" more ;)

With my hasher optimization that's 205/185 ms, but 41322 is just paradigm-shift.

-- 
Sebastian Gniazdowski
psprint /at/ zdharma.org

Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

[Sebastian Gniazdowski]

On 19 czerwca 2017 at 22:52:37, Peter Stephenson (p.w.stephenson@ntlworld.com) wrote:
> On Mon, 19 Jun 2017 20:28:35 +0100
> Peter Stephenson wrote:
> > How about something like this? As far as I can tell, fgets is designed
> > from the ground up as Gets Done Properly, so if you have it on your
> > system it will work correctly. I can't think of a case where this
> > wouldn't do the right thing --- fgets will read at most one line and if
> > it does we were going to get the big STDIO overhead at that point
> > anyway.
> 
> I know what the problem with it is --- we can't tell the difference
> betwen a terminating '\0' and a '\0' that's part of the input stream.

We could do a pair of calls, first fgets(), then fgetc(), and this way solve this problem?

-- 
Sebastian Gniazdowski
psprint /at/ zdharma.org

Re: Why sourcing a file is not faster than doing a loop with eval, zle -N

[Sebastian Gniazdowski]

On 23 grudnia 2017 at 16:01:38, Sebastian Gniazdowski (psprint@zdharma.org) wrote:
> We could do a pair of calls, first fgets(), then fgetc(), and this way solve this problem? 

Sorry rather no, seems more complex than that.

-- 
Sebastian Gniazdowski
psprint /at/ zdharma.org