[ARM] Optimised strchr and strlen

[ARM] Optimised strchr and strlen

[Dr. David Alan Gilbert]

This is a strchr and strlen optimised for ARM v6t2 or v7. It's against svn rev r15869
with my previous memchr patch.   Tested both little & big endian.

(I've checked it still applies on svn trunk, but not done a retest on that; nothing
seems to have changed around there).

Dave

2012-12-19 Dr. David Alan Gilbert <david.gilbert@linaro.org>
	* sysdeps/arm/eabi/armv6t2/strchr.S: New file
	* sysdeps/arm/eabi/armv6t2/strlen.S: New file


diff -urN ports/sysdeps/arm/eabi/armv6t2/strchr.S src/ports/sysdeps/arm/eabi/armv6t2/strchr.S
--- ports/sysdeps/arm/eabi/armv6t2/strchr.S	1970-01-01 01:00:00.000000000 +0100
+++ ports/sysdeps/arm/eabi/armv6t2/strchr.S	2011-12-16 13:43:56.704694919 +0000
@@ -0,0 +1,71 @@
+/* Copyright (C) 2011 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Code contributed by Dave Gilbert <david.gilbert@linaro.org>
+ 
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#include <sysdep.h>
+
+@ A very simple strchr routine, from benchmarks on A9 it's a bit faster than
+@ the current version in eglibc.
+@ While I have a version that does 8 bytes/loop and is a lot faster on long
+@ strings, it is slower on short strings, and short strings seem more common
+@ in strchr usage.
+@ Note: The use of cbz/cbnz means it's Thumb only
+
+@ 2011-02-07 david.gilbert@linaro.org
+@    Extracted from local git a5b438d861
+@ 2011-12-16 david.gilbert@linaro.org
+@    Copy from Cortex strings rev 65 and change license
+
+	.syntax unified
+
+	.text
+	.thumb
+
+@ ---------------------------------------------------------------------------
+
+	.thumb_func
+	.global strchr
+	.type strchr,%function
+ENTRY(strchr)
+	@ r0 = start of string
+	@ r1 = character to match
+	@ returns NULL for no match, or a pointer to the match
+	and	r1,r1, #255
+
+1:
+	ldrb	r2,[r0],#1
+	cmp	r2,r1
+	cbz	r2,10f
+	bne	1b
+
+	@ We're here if it matched
+5:
+	subs	r0,r0,#1
+	DO_RET(lr)
+
+10:
+	@ We're here if we ran off the end
+	cmp	r1, #0	@ Corner case - you can search for the nil and get a pointer to it
+	beq	5b	@ messy, if common we should branch at the start to a special loop
+	mov	r0,#0
+	DO_RET(lr)
+
+END(strchr)
+
+weak_alias (strchr, index)
+libc_hidden_builtin_def(strchr)
diff -urN ports/sysdeps/arm/eabi/armv6t2/strlen.S src/ports/sysdeps/arm/eabi/armv6t2/strlen.S
--- ports/sysdeps/arm/eabi/armv6t2/strlen.S	1970-01-01 01:00:00.000000000 +0100
+++ ports/sysdeps/arm/eabi/armv6t2/strlen.S	2011-12-16 13:43:01.991130183 +0000
@@ -0,0 +1,118 @@
+/* Copyright (C) 2011 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Code contributed by Dave Gilbert <david.gilbert@linaro.org>
+ 
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#include <sysdep.h>
+
+@ This strlen routine is optimised on a Cortex-A9 and should work on
+@ all ARMv7 processors.   This routine is reasonably fast for short
+@ strings, but is probably slower than a simple implementation if all
+@ your strings are very short
+@ Note: The use of cbz/cbnz means it's Thumb only
+
+@ 2011-02-08 david.gilbert@linaro.org
+@    Extracted from local git 6848613a
+@ 2011-12-16 david.gilbert@linaro.org
+@    Copy from Cortex strings rev 65 and change license
+@    Add cfi magic, switch to ldrd
+
+
+@ this lets us check a flag in a 00/ff byte easily in either endianness
+#ifdef __ARMEB__
+#define CHARTSTMASK(c) 1<<(31-(c*8))
+#else
+#define CHARTSTMASK(c) 1<<(c*8)
+#endif
+
+@-----------------------------------------------------------------------------
+	.syntax unified
+
+	.text
+	.thumb
+
+	.thumb_func
+	.global strlen
+	.type strlen,%function
+ENTRY(strlen)
+	@ r0 = string
+	@ returns count of bytes in string not including terminator
+	mov	r1, r0
+	push	{ r4,r6 }
+	cfi_adjust_cfa_offset (8)
+	cfi_rel_offset (r4, 0)
+	cfi_rel_offset (r6, 4)
+
+	cfi_remember_state
+
+	mvns	r6, #0		@ all F
+	movs	r4, #0
+	tst	r0, #7
+	beq	2f
+
+1:
+	ldrb	r2, [r1], #1
+	tst	r1, #7		@ Hit alignment yet?
+	cbz	r2, 10f		@ Exit if we found the 0
+	bne	1b
+
+	@ So we're now aligned
+2:
+	ldrd	r2,r3,[r1],#8
+	uadd8	r2, r2, r6	@ Parallel add 0xff - sets the GE bits for anything that wasn't 0
+	sel	r2, r4, r6	@ bytes are 00 for none-00 bytes, or ff for 00 bytes - NOTE INVERSION
+	uadd8	r3, r3, r6	@ Parallel add 0xff - sets the GE bits for anything that wasn't 0
+	sel	r3, r2, r6	@ bytes are 00 for none-00 bytes, or ff for 00 bytes - NOTE INVERSION
+	cmp	r3, #0
+	beq	2b
+
+strlenendtmp:
+	@ One (or more) of the bytes we loaded was 0 - but which one?
+	@ r2 has the mask corresponding to the first loaded word
+	@ r3 has a combined mask of the two words - but if r2 was all-non 0 
+	@ then it's just the 2nd words
+	cmp	r2, #0
+	itte	eq
+	moveq	r2, r3		@ the end is in the 2nd word
+	subeq	r1,r1,#3
+	subne	r1,r1,#7
+
+	@ r1 currently points to the 2nd byte of the word containing the 0
+	tst	r2, # CHARTSTMASK(0)	@ 1st character
+	bne	10f
+	adds	r1,r1,#1
+	tst	r2, # CHARTSTMASK(1)	@ 2nd character
+	ittt	eq
+	addeq	r1,r1,#1
+	tsteq	r2, # (3<<15)		@ 2nd & 3rd character
+	@ If not the 3rd must be the last one
+	addeq	r1,r1,#1
+
+10:
+	@ r0 is still at the beginning, r1 is pointing 1 byte after terminator
+	sub	r0, r1, r0
+	subs	r0, r0, #1
+	pop	{ r4, r6 }
+
+	cfi_adjust_cfa_offset (-8)
+	cfi_restore (r4)
+	cfi_restore (r6)
+
+	DO_RET(lr)
+
+END(strlen)
+libc_hidden_builtin_def (strlen)

Re: [ARM] Optimised strchr and strlen

[David Gilbert]

On 19 December 2011 17:30, Hector Oron <hector.oron@gmail.com> wrote:
> Hello,
>
> 2011/12/19 Dr. David Alan Gilbert <david.gilbert@linaro.org>:
>
>> 2012-12-19 Dr. David Alan Gilbert <david.gilbert@linaro.org>
>>        * sysdeps/arm/eabi/armv6t2/strchr.S: New file
>>        * sysdeps/arm/eabi/armv6t2/strlen.S: New file
>
> As a minor issue, I think the date on your ChangeLog entry is bogus by
> a year. :-)

Oops yes - well spotted!


Dave

Re: [ARM] Optimised strchr and strlen

[Richard Henderson]

On 12/19/2011 09:21 AM, Dr. David Alan Gilbert wrote:
> +	@ r1 currently points to the 2nd byte of the word containing the 0
> +	tst	r2, # CHARTSTMASK(0)	@ 1st character
> +	bne	10f
> +	adds	r1,r1,#1
> +	tst	r2, # CHARTSTMASK(1)	@ 2nd character
> +	ittt	eq
> +	addeq	r1,r1,#1
> +	tsteq	r2, # (3<<15)		@ 2nd & 3rd character
> +	@ If not the 3rd must be the last one
> +	addeq	r1,r1,#1

No need to search -- clz can do that for you.

#ifdef __ARMEL__
	rbit	r2, r2
#endif
	clz	r2, r2
	lsrs	r2, r2, #3
	adds	r1, r1, r2


r~

Re: [ARM] Optimised strchr and strlen

[Joseph S. Myers]

On Mon, 19 Dec 2011, Dr. David Alan Gilbert wrote:

> +@ A very simple strchr routine, from benchmarks on A9 it's a bit faster than
> +@ the current version in eglibc.
> +@ While I have a version that does 8 bytes/loop and is a lot faster on long
> +@ strings, it is slower on short strings, and short strings seem more common
> +@ in strchr usage.

That sounds like a possible case for a hybrid function, with an unrolled 
initial part testing some number of characters to cover short strings (it 
might be possible to get things aligned at the same time) and the more 
complicated version for longer strings.  What's the actual size 
distribution you see in strchr use?

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: [ARM] Optimised strchr and strlen

[David Gilbert]

On 20 December 2011 23:07, Joseph S. Myers <joseph@codesourcery.com> wrote:
> On Mon, 19 Dec 2011, Dr. David Alan Gilbert wrote:
>
>> +@ A very simple strchr routine, from benchmarks on A9 it's a bit faster than
>> +@ the current version in eglibc.
>> +@ While I have a version that does 8 bytes/loop and is a lot faster on long
>> +@ strings, it is slower on short strings, and short strings seem more common
>> +@ in strchr usage.
>
> That sounds like a possible case for a hybrid function, with an unrolled
> initial part testing some number of characters to cover short strings (it
> might be possible to get things aligned at the same time) and the more
> complicated version for longer strings.

Of course the difficulty with strchr (compared with memchr) is that you have
no length parameter to hint at how much is left, and thus whether it's
worth making that switch; so it has to be a heuristic and it's going to cost
you something in the small case.

>  What's the actual size distribution you see in strchr use?

It varies heavily by program.  I only found one of the SPEC benchmarks using
it to a measurable amount (i.e. showed up in profile), and that was
mostly in 24byte strings
with the match at random positions within the string.

From some embedded benchmarks I found that almost all the uses of strchr
were calls with less than 8 byte strings (mostly unexpected fallout
from within libc rather
than the meat of the benchmark).
One of the things people commonly seem to do with strchr is see whether a
character is in a set, and run that strchr along a string - e.g. something like

char *mystr=....
char tmp;

for(tmp=*mystr;tmp;mystr++) {
   tmp=*(mystr++);
   if (strchr(" \t\n\r", tmp)!=NULL) { do something }
}

With the string being searched depressingly small.   There are some places where
you might hope the compiler could be smart and do something with that,
although often the string being searched is actually a variable (think
$IFS in shell).
(I fancy something like a strchr_short with the compiler calling that
when it knows
the string is less than some length - but how do you define that
length between the
compiler and library?)

The other thing I found in some examples was splitting env strings; so
searching for
the '=' in NAME=VALUE, the NAME parts tend not to be particularly long.

The worst case tends to be where you're using strchr on a long string and you
don't actually find the match.

A ltrace of gcc's cc1 shows a few of the cases - e.g.

Mostly short identifiers:

strchr("nothrow", ' ')                           = NULL
strchr("final", ' ')                             = NULL
strchr("dfinish", ' ')                           = NULL


This is a case that can get long - finding '.' in filenames; this can
be one where
you get longer strings without  a match.

strchr("/usr/local/include", '.')                = NULL
strchr("/usr/local/include", '.')                = NULL
strchr("/usr/lib/arm-linux-gnueabi/gcc/arm-linux-gnueabi/4.5.2/include",
'.') = ".5.2/include"

I suspect something like parsing a format string:
strchr("-+ #0", 's')                             = NULL

Splitting assignments:
strchr("__UINT16_C(c)=c", '=')                   = "=c"
strchr("__UINT_LEAST32_MAX__=4294967295U", '=')  = "=4294967295U"


Here's a profile graph of different strlen's on an ARM:
https://wiki.linaro.org/WorkingGroups/ToolChain/Benchmarks/InitialStrchr?action=AttachFile&do=get&target=panda-01-strchr-git44154ec-strchr-abs.png

That 'simple' one is showing the benefit at the short lengths,
the 'smarter' one I have is doing 8 bytes/loop and is nice on the long
strings - but as you can see worse at the short ones.

Dave

Re: [ARM] Optimised strchr and strlen

[David Gilbert]

On 20 December 2011 20:29, Richard Henderson <rth@twiddle.net> wrote:
> On 12/19/2011 09:21 AM, Dr. David Alan Gilbert wrote:
>> +     @ r1 currently points to the 2nd byte of the word containing the 0
>> +     tst     r2, # CHARTSTMASK(0)    @ 1st character
>> +     bne     10f
>> +     adds    r1,r1,#1
>> +     tst     r2, # CHARTSTMASK(1)    @ 2nd character
>> +     ittt    eq
>> +     addeq   r1,r1,#1
>> +     tsteq   r2, # (3<<15)           @ 2nd & 3rd character
>> +     @ If not the 3rd must be the last one
>> +     addeq   r1,r1,#1
>
> No need to search -- clz can do that for you.
>
> #ifdef __ARMEL__
>        rbit    r2, r2
> #endif
>        clz     r2, r2
>        lsrs    r2, r2, #3
>        adds    r1, r1, r2

Hi Richard,
  Thanks for the suggestion - I'd seen that form in some code from another
library and decided because of that not to roll it into my code; however
since there have now been 3 separate people who've suggested it to me
independently I don't see why not!

Thanks,

Dave

Re: [ARM] Optimised strchr and strlen

[Richard Henderson]

On 12/21/2011 02:55 AM, David Gilbert wrote:
> That 'simple' one is showing the benefit at the short lengths,
> the 'smarter' one I have is doing 8 bytes/loop and is nice on the long
> strings - but as you can see worse at the short ones.

Having not seen your "smarter" strchr, it's hard to suggest anything
concrete.  I'd have thought that there's enough slack in load delay
that one or two arithmetic operations could be done without penalty...

Something like performing a simple compare loop looking for "alignment plus":

...
	bic	r3, r0, #7
	and	r1, r1, #255
	adds	r3, r3, #32
1:
	ldrb	r2, [r0],#1
	cmp	r2, r1
	cbz	r2, .Lfound_zero
	it	ne
	cmpne	r0, r3
	bne	1b
	cmp	r2, r1
	beq	.Lfound
	@ Here, r0 is aligned.  Do something word-based.
...

or even just

	and	r3, r0, #7
	and	r1, r1, #255
	rsb	r3, r3, #32
1:
	ldrb	r2, [r0],#1
	cmp	r2, r1
	beq	.Lfound
	subs	r3, r3, #1
	cbz	r2, .Lfound_zero
	bne	1b
	@ Here, r0 is aligned.  Do something word-based.


r~

Re: [ARM] Optimised strchr and strlen

[David Gilbert]

On 21 December 2011 21:20, Richard Henderson <rth@twiddle.net> wrote:
> On 12/21/2011 02:55 AM, David Gilbert wrote:
>> That 'simple' one is showing the benefit at the short lengths,
>> the 'smarter' one I have is doing 8 bytes/loop and is nice on the long
>> strings - but as you can see worse at the short ones.
>
> Having not seen your "smarter" strchr, it's hard to suggest anything
> concrete.  I'd have thought that there's enough slack in load delay
> that one or two arithmetic operations could be done without penalty...

Sure; it's pretty much the same trick as my strlen routine.

> Something like performing a simple compare loop looking for "alignment plus":

<snip>

> or even just
>
>        and     r3, r0, #7
>        and     r1, r1, #255
>        rsb     r3, r3, #32
> 1:
>        ldrb    r2, [r0],#1
>        cmp     r2, r1
>        beq     .Lfound
>        subs    r3, r3, #1
>        cbz     r2, .Lfound_zero
>        bne     1b
>        @ Here, r0 is aligned.  Do something word-based.

OK, so I gave that a go - and the results are:

https://wiki.linaro.org/WorkingGroups/ToolChain/Benchmarks/InitialStrchr?action=AttachFile&do=get&target=strchr-fiddle20111223-rth-strchr-abs.png

The line on that graph above labelled 'fiddle' is the one where I
tried that code; I can't explain the
really grim dip (at 16 bytes) - I checked with a debugger that it is
going around the loop
upto 32 times.  However, even if we ignored the dip at 16, it's still
not a nice result - it's worse
than the simple routine I posted up to 64 bytes.  One guess is that
loop with 3 branches in might
be too much for the branch predictor (I remember there are
restrictions on how close together they can go.

I might be able to make that a bit better for the larger cases by
using the clz trick you suggested for the strlen
for the end-of-fastloop case; and I'll give that a go; but it's not
going to help at the bottom.

(OT: Does anyone know how to get R to use a log2 X graph rather than log 10?)

Here are the raw numbers that maybe easier than the graphs (Note the
strings tested are aligned):

smarter_strchr_armv7fiddle: ,6553600, loops of ,64, bytes=400.000000
MB, transferred in ,1395904541.000000 ns, giving, 286.552546 MB/s
smarter_strchr_armv7fiddle: ,6553600, loops of ,62, bytes=387.500000
MB, transferred in ,1258880615.000000 ns, giving, 307.813144 MB/s
smarter_strchr_armv7fiddle: ,13107200, loops of ,32, bytes=400.000000
MB, transferred in ,1813507080.000000 ns, giving, 220.567101 MB/s
smarter_strchr_armv7fiddle: ,26214400, loops of ,16, bytes=400.000000
MB, transferred in ,2922332764.000000 ns, giving, 136.876951 MB/s
smarter_strchr_armv7fiddle: ,52428800, loops of ,8, bytes=400.000000
MB, transferred in ,1904724122.000000 ns, giving, 210.004166 MB/s
smarter_strchr_armv7fiddle: ,104857600, loops of ,4, bytes=400.000000
MB, transferred in ,2191650391.000000 ns, giving, 182.510861 MB/s
smarter_strchr_armv7fiddle: ,209715200, loops of ,2, bytes=400.000000
MB, transferred in ,3548339844.000000 ns, giving, 112.728774 MB/s

smarter_strchr_armv7: ,6553600, loops of ,64, bytes=400.000000 MB,
transferred in ,726409912.000000 ns, giving, 550.653279 MB/s
smarter_strchr_armv7: ,6553600, loops of ,62, bytes=387.500000 MB,
transferred in ,743652344.000000 ns, giving, 521.076822 MB/s
smarter_strchr_armv7: ,13107200, loops of ,32, bytes=400.000000 MB,
transferred in ,874023437.000000 ns, giving, 457.653632 MB/s
smarter_strchr_armv7: ,26214400, loops of ,16, bytes=400.000000 MB,
transferred in ,1160278321.000000 ns, giving, 344.744871 MB/s
smarter_strchr_armv7: ,52428800, loops of ,8, bytes=400.000000 MB,
transferred in ,1617584229.000000 ns, giving, 247.282332 MB/s
smarter_strchr_armv7: ,104857600, loops of ,4, bytes=400.000000 MB,
transferred in ,3235260010.000000 ns, giving, 123.637667 MB/s
smarter_strchr_armv7: ,209715200, loops of ,2, bytes=400.000000 MB,
transferred in ,7044738769.000000 ns, giving, 56.779962 MB/s

simple_strchr: ,6553600, loops of ,64, bytes=400.000000 MB,
transferred in ,1298034668.000000 ns, giving, 308.158179 MB/s
simple_strchr: ,6553600, loops of ,62, bytes=387.500000 MB,
transferred in ,1265472413.000000 ns, giving, 306.209757 MB/s
simple_strchr: ,13107200, loops of ,32, bytes=400.000000 MB,
transferred in ,1447998047.000000 ns, giving, 276.243467 MB/s
simple_strchr: ,26214400, loops of ,16, bytes=400.000000 MB,
transferred in ,1245819092.000000 ns, giving, 321.073904 MB/s
simple_strchr: ,52428800, loops of ,8, bytes=400.000000 MB,
transferred in ,1435119629.000000 ns, giving, 278.722409 MB/s
simple_strchr: ,104857600, loops of ,4, bytes=400.000000 MB,
transferred in ,1878479004.000000 ns, giving, 212.938233 MB/s
simple_strchr: ,209715200, loops of ,2, bytes=400.000000 MB,
transferred in ,3339569092.000000 ns, giving, 119.775932 MB/s

Dave

Re: [ARM] Optimised strchr and strlen

[Richard Henderson]

[-- Attachment #1: Type: text/plain, Size: 553 bytes --]

On 12/23/2011 12:31 PM, David Gilbert wrote:
> Sure; it's pretty much the same trick as my strlen routine.
...
> OK, so I gave that a go - and the results are:

I can't help but wonder if just the one branch in the first loop is best.

Also, it appears one can use uqadd8 and do the aligned two words in parallel
rather than having everything serialize on the GT flags and SEL.

I've run this through glibc's test-strchr, but havn't gotten around to
benchmarking it at all.  Since you've already got that set up, perhaps
you could give it a whirl.


r~

[-- Attachment #2: strchr.S --]
[-- Type: text/plain, Size: 3303 bytes --]

@ We occasionally want to use the S form simply to achieve a smaller
@ instruction form in Thumb mode.  Never set the flags in ARM mode.
#ifdef __thumb2__
# define s(insn)	insn##s
#else
# define s(insn)	insn
#endif

	.syntax unified
	.text

#ifdef __thumb2__
	.thumb
	.thumb_func
	.arch	armv6t2
#else
	.arch	armv7-a
#endif

	.global strchr
	.type	strchr, %function

strchr:
	@ r0 = start of string
	@ r1 = character to match
	@ returns NULL for no match, or a pointer to the match

	@ To cater to long strings, we want to search through a few
	@ characters until we reach an aligned pointer.  To cater to
	@ small strings, we don't want to start doing word operations
	@ immediately.  The compromise is a maximum of 32 bytes less
	@ whatever is required to end with an aligned pointer.
	@ r3 = number of characters to search in alignment loop
	and	r3, r0, #7
	uxtb	r1, r1
	rsb	r3, r3, #32

	@ Loop until we find ...
1:	ldrb	r2, [r0], #1
	subs	r3, r3, #1		@ ... the aligment point
	it	ne
	cmpne	r2, r1			@ ... or the character
	it	ne
	cmpne	r2, #0			@ ... or EOS
	bne	1b

	@ Disambiguate the exit possibilites above
	cmp	r2, r1			@ Found the character
	itt	eq
	subeq	r0, r0, #1
	bxeq	lr

	cmp	r2, #0			@ Found EOS
	itt	eq
	moveq	r0, #0
	bxeq	lr

	@ So now we're aligned.  Now we actually need a stack frame.
	push	{ r4, r5, r6, r7 }
	orr	r1, r1, r1, lsl #8	@ Replicate C to all bytes
	movw	ip, #0xfefe
	orr	r1, r1, r1, lsl #16
	movt	ip, #0xfefe

	@ Loop searching for EOS or C, 8 bytes at a time.
2:	ldrd	r2, r3, [r0], #8
	@ Adding (unsigned saturating) 0xfe means result of 0xfe for any byte
	@ that was originally zero and 0xff otherwise.  Therefore we consider
	@ the lsb of each byte the "found" bit, with 0 for a match.
	uqadd8	r4, r2, ip		@ Find EOS
	uqadd8	r5, r3, ip
	eor	r6, r2, r1		@ Convert C bytes to 0
	eor	r7, r3, r1
	uqadd8	r6, r6, ip		@ Find C
	uqadd8	r7, r7, ip
	s(and)	r4, r4, r6		@ Combine found for EOS and C
	s(and)	r5, r5, r7
	and	r6, r4, r5		@ Combine the two words
	mvns	r6, r6			@ Test for any found bit true
	beq	2b

	@ Invert the sense of the found bits.  After this we have 1 in
	@ any byte that contains a match, and 0 otherwise.
	s(mvn)	r5, r5
	mvns	r4, r4

	@ Found something.  Disambiguate between first and second words.
	@ Adjust r0 to point to the word containing the match.
	@ Adjust r2 to the contents of the word containing the match.
	@ Adjust r4 to the found bits for the word containing the match.
	iteee	ne
	subne	r0, r0, #8
	subeq	r0, r0, #4
	moveq	r4, r5
	moveq	r2, r3

	@ Find the bit-offset of the match within the word.
#ifdef __ARMEL__
	@ For little-endian, we only need to reverse the bits so that
	@ count-leading-zeros becomes in effect count-trailing-zeros.
	rbit	r4, r4
	clz	r3, r4
#else
	@ For big-endian, we're matching 0x01 (not 0x80), and so the
	@ bit offset is 7 too high.  Also, we byte-swap the word so
	@ that we can shift down to extract the found byte.
	clz	r3, r4
	rev	r2, r2
	s(sub)	r3, r3, #7
#endif
	s(lsr)	r2, r2, r3		@ Shift down found byte
	add	r0, r0, r3, lsr #3	@ Adjust the pointer to the found byte
	uxtb	r2, r2			@ Extract found byte
	uxtb	r1, r1			@ Undo replication of C

	pop	{ r4, r5, r6, r7 }

	@ Disambiguate between EOS and C.
	cmp	r2, r1
	it	ne
	movne	r0, #0			@ Found EOS, return NULL
	bx	lr

	.size	strchr, . - strchr

Re: [ARM] Optimised strchr and strlen

[David Gilbert]

On 24 December 2011 21:01, Richard Henderson <rth@twiddle.net> wrote:
> On 12/23/2011 12:31 PM, David Gilbert wrote:
>> Sure; it's pretty much the same trick as my strlen routine.
> ...
>> OK, so I gave that a go - and the results are:
>
> I can't help but wonder if just the one branch in the first loop is best.

Yes.

> Also, it appears one can use uqadd8 and do the aligned two words in parallel
> rather than having everything serialize on the GT flags and SEL.
>
> I've run this through glibc's test-strchr, but havn't gotten around to
> benchmarking it at all.  Since you've already got that set up, perhaps
> you could give it a whirl.

Here we go - you're code is the green line; rth_strchr - your uqadd8
trick is very nice;
the peak speed is a nice bit higher than my version using a set of uadd8's and
sel (you get 1 instruction less in the main loop).

https://wiki.linaro.org/WorkingGroups/ToolChain/Benchmarks/InitialStrchr?action=AttachFile&do=view&target=strchr-withrth-strchr-abs.png

The simple routine is still easily winning below 32 bytes though, and
there is still that odd notch at 16.

(I think your uqadd8 trick would be a nice improvement on my strlen
and memchr routines).

Dave