[PATCH] Workaround for ffs() on LP64 targets

[PATCH] Workaround for ffs() on LP64 targets

[Sebastian Huber]

Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
---
 newlib/libc/misc/ffs.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
index ba5700920..a09cbd3bb 100644
--- a/newlib/libc/misc/ffs.c
+++ b/newlib/libc/misc/ffs.c
@@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
 int
 ffs(int i)
 {
+#ifdef __LP64__
+	/* GCC would expand the __builtin_ffs() to ffs() in this case */
+	int bit;
+
+	if (i == 0)
+		return (0);
+	for (bit = 1; !(i & 1); bit++)
+		i = (unsigned int)i >> 1;
+	return (bit);
+#else
 
 	return (__builtin_ffs(i));
+#endif
 }
-- 
2.12.3

Re: [PATCH] Workaround for ffs() on LP64 targets

[Kito Cheng]

it's will make aarch64 gen worse code (compare to __builtin_ffs
version) since aarch64 have clz, but I don't have better idea too...

On Thu, Jul 27, 2017 at 4:06 PM, Sebastian Huber
<sebastian.huber@embedded-brains.de> wrote:
> Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
> ---
>  newlib/libc/misc/ffs.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
>
> diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
> index ba5700920..a09cbd3bb 100644
> --- a/newlib/libc/misc/ffs.c
> +++ b/newlib/libc/misc/ffs.c
> @@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
>  int
>  ffs(int i)
>  {
> +#ifdef __LP64__
> +       /* GCC would expand the __builtin_ffs() to ffs() in this case */
> +       int bit;
> +
> +       if (i == 0)
> +               return (0);
> +       for (bit = 1; !(i & 1); bit++)
> +               i = (unsigned int)i >> 1;
> +       return (bit);
> +#else
>
>         return (__builtin_ffs(i));
> +#endif
>  }
> --
> 2.12.3
>

Re: [PATCH] Workaround for ffs() on LP64 targets

[Sebastian Huber]

On 27/07/17 10:29, Kito Cheng wrote:

> it's will make aarch64 gen worse code (compare to __builtin_ffs
> version) since aarch64 have clz, but I don't have better idea too...

Maybe we should add a defined(__riscv__). A count leading zeros 
instruction is quite common.

-- 
Sebastian Huber, embedded brains GmbH

Address : Dornierstr. 4, D-82178 Puchheim, Germany
Phone   : +49 89 189 47 41-16
Fax     : +49 89 189 47 41-09
E-Mail  : sebastian.huber@embedded-brains.de
PGP     : Public key available on request.

Diese Nachricht ist keine geschäftliche Mitteilung im Sinne des EHUG.

Re: [PATCH] Workaround for ffs() on LP64 targets

[Kito Cheng]

Hi Sebastian:

You mean look like this? I am ok for this if you feel this solution is
acceptable :)

Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
---
 newlib/libc/misc/ffs.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
index ba5700920..a09cbd3bb 100644
--- a/newlib/libc/misc/ffs.c
+++ b/newlib/libc/misc/ffs.c
@@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
 int
 ffs(int i)
 {
+#if defined(__LP64__) && defined(__riscv)
+       /* GCC would expand the __builtin_ffs() to ffs() in this case */
+       int bit;
+
+       if (i == 0)
+               return (0);
+       for (bit = 1; !(i & 1); bit++)
+               i = (unsigned int)i >> 1;
+       return (bit);
+#else

        return (__builtin_ffs(i));
+#endif
 }

On Thu, Jul 27, 2017 at 4:40 PM, Sebastian Huber
<sebastian.huber@embedded-brains.de> wrote:
> On 27/07/17 10:29, Kito Cheng wrote:
>
>> it's will make aarch64 gen worse code (compare to __builtin_ffs
>> version) since aarch64 have clz, but I don't have better idea too...
>
>
> Maybe we should add a defined(__riscv__). A count leading zeros instruction
> is quite common.
>
> --
> Sebastian Huber, embedded brains GmbH
>
> Address : Dornierstr. 4, D-82178 Puchheim, Germany
> Phone   : +49 89 189 47 41-16
> Fax     : +49 89 189 47 41-09
> E-Mail  : sebastian.huber@embedded-brains.de
> PGP     : Public key available on request.
>
> Diese Nachricht ist keine geschäftliche Mitteilung im Sinne des EHUG.
>

Re: [PATCH] Workaround for ffs() on LP64 targets

[Eric Blake]

[-- Attachment #1.1: Type: text/plain, Size: 1144 bytes --]

On 07/27/2017 03:06 AM, Sebastian Huber wrote:
> Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
> ---
>  newlib/libc/misc/ffs.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
> index ba5700920..a09cbd3bb 100644
> --- a/newlib/libc/misc/ffs.c
> +++ b/newlib/libc/misc/ffs.c
> @@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
>  int
>  ffs(int i)
>  {
> +#ifdef __LP64__
> +	/* GCC would expand the __builtin_ffs() to ffs() in this case */
> +	int bit;
> +
> +	if (i == 0)
> +		return (0);
> +	for (bit = 1; !(i & 1); bit++)
> +		i = (unsigned int)i >> 1;
> +	return (bit);

If we're going to open-code it to work around the compiler creating an
infloop recursion to ffs(), at least code a straight-line version
without branches, rather than the painfully slow bit-by-bit loop.
There's plenty of examples on the web of writing ffs() by using
bit-twiddling without branching.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 619 bytes --]

Re: [PATCH] Workaround for ffs() on LP64 targets

[Sebastian Huber]

On 27/07/17 13:13, Eric Blake wrote:

> On 07/27/2017 03:06 AM, Sebastian Huber wrote:
>> Signed-off-by: Sebastian Huber<sebastian.huber@embedded-brains.de>
>> ---
>>   newlib/libc/misc/ffs.c | 11 +++++++++++
>>   1 file changed, 11 insertions(+)
>>
>> diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
>> index ba5700920..a09cbd3bb 100644
>> --- a/newlib/libc/misc/ffs.c
>> +++ b/newlib/libc/misc/ffs.c
>> @@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
>>   int
>>   ffs(int i)
>>   {
>> +#ifdef __LP64__
>> +	/* GCC would expand the __builtin_ffs() to ffs() in this case */
>> +	int bit;
>> +
>> +	if (i == 0)
>> +		return (0);
>> +	for (bit = 1; !(i & 1); bit++)
>> +		i = (unsigned int)i >> 1;
>> +	return (bit);
> If we're going to open-code it to work around the compiler creating an
> infloop recursion to ffs(), at least code a straight-line version
> without branches, rather than the painfully slow bit-by-bit loop.
> There's plenty of examples on the web of writing ffs() by using
> bit-twiddling without branching.

This is roughly the same implementation we had before. I do not intend 
to optimize this.

-- 
Sebastian Huber, embedded brains GmbH

Address : Dornierstr. 4, D-82178 Puchheim, Germany
Phone   : +49 89 189 47 41-16
Fax     : +49 89 189 47 41-09
E-Mail  : sebastian.huber@embedded-brains.de
PGP     : Public key available on request.

Diese Nachricht ist keine geschäftliche Mitteilung im Sinne des EHUG.

Re: [PATCH] Workaround for ffs() on LP64 targets

[Corinna Vinschen]

[-- Attachment #1: Type: text/plain, Size: 1619 bytes --]

On Jul 27 13:24, Sebastian Huber wrote:
> On 27/07/17 13:13, Eric Blake wrote:
> 
> > On 07/27/2017 03:06 AM, Sebastian Huber wrote:
> > > Signed-off-by: Sebastian Huber<sebastian.huber@embedded-brains.de>
> > > ---
> > >   newlib/libc/misc/ffs.c | 11 +++++++++++
> > >   1 file changed, 11 insertions(+)
> > > 
> > > diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
> > > index ba5700920..a09cbd3bb 100644
> > > --- a/newlib/libc/misc/ffs.c
> > > +++ b/newlib/libc/misc/ffs.c
> > > @@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
> > >   int
> > >   ffs(int i)
> > >   {
> > > +#ifdef __LP64__
> > > +	/* GCC would expand the __builtin_ffs() to ffs() in this case */
> > > +	int bit;
> > > +
> > > +	if (i == 0)
> > > +		return (0);
> > > +	for (bit = 1; !(i & 1); bit++)
> > > +		i = (unsigned int)i >> 1;
> > > +	return (bit);
> > If we're going to open-code it to work around the compiler creating an
> > infloop recursion to ffs(), at least code a straight-line version
> > without branches, rather than the painfully slow bit-by-bit loop.
> > There's plenty of examples on the web of writing ffs() by using
> > bit-twiddling without branching.
> 
> This is roughly the same implementation we had before. I do not intend to
> optimize this.

Still, __LP64__ is unacceptable.  Cygwin would be affected by this as
well and would have to revert to its former own ffs implementation.

Reverting to a C-based implementation should only be performed on a
case-by-case basis.


Thanks,
Corinna

-- 
Corinna Vinschen
Cygwin Maintainer
Red Hat

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]

Re: [PATCH] Workaround for ffs() on LP64 targets

[Sebastian Huber]

On 27/07/17 14:27, Corinna Vinschen wrote:

> On Jul 27 13:24, Sebastian Huber wrote:
>> On 27/07/17 13:13, Eric Blake wrote:
>>
>>> On 07/27/2017 03:06 AM, Sebastian Huber wrote:
>>>> Signed-off-by: Sebastian Huber<sebastian.huber@embedded-brains.de>
>>>> ---
>>>>    newlib/libc/misc/ffs.c | 11 +++++++++++
>>>>    1 file changed, 11 insertions(+)
>>>>
>>>> diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
>>>> index ba5700920..a09cbd3bb 100644
>>>> --- a/newlib/libc/misc/ffs.c
>>>> +++ b/newlib/libc/misc/ffs.c
>>>> @@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
>>>>    int
>>>>    ffs(int i)
>>>>    {
>>>> +#ifdef __LP64__
>>>> +	/* GCC would expand the __builtin_ffs() to ffs() in this case */
>>>> +	int bit;
>>>> +
>>>> +	if (i == 0)
>>>> +		return (0);
>>>> +	for (bit = 1; !(i & 1); bit++)
>>>> +		i = (unsigned int)i >> 1;
>>>> +	return (bit);
>>> If we're going to open-code it to work around the compiler creating an
>>> infloop recursion to ffs(), at least code a straight-line version
>>> without branches, rather than the painfully slow bit-by-bit loop.
>>> There's plenty of examples on the web of writing ffs() by using
>>> bit-twiddling without branching.
>> This is roughly the same implementation we had before. I do not intend to
>> optimize this.
> Still, __LP64__ is unacceptable.  Cygwin would be affected by this as
> well and would have to revert to its former own ffs implementation.
>
> Reverting to a C-based implementation should only be performed on a
> case-by-case basis.

Yes, so maybe something like this

#if defined(__LP64__) && defined(__riscv)

or a target-specific ffs.c file similar to memcpy.c, etc.

-- 
Sebastian Huber, embedded brains GmbH

Address : Dornierstr. 4, D-82178 Puchheim, Germany
Phone   : +49 89 189 47 41-16
Fax     : +49 89 189 47 41-09
E-Mail  : sebastian.huber@embedded-brains.de
PGP     : Public key available on request.

Diese Nachricht ist keine geschäftliche Mitteilung im Sinne des EHUG.

Re: [PATCH] Workaround for ffs() on LP64 targets

[Corinna Vinschen]

[-- Attachment #1: Type: text/plain, Size: 2130 bytes --]

On Jul 27 14:33, Sebastian Huber wrote:
> On 27/07/17 14:27, Corinna Vinschen wrote:
> 
> > On Jul 27 13:24, Sebastian Huber wrote:
> > > On 27/07/17 13:13, Eric Blake wrote:
> > > 
> > > > On 07/27/2017 03:06 AM, Sebastian Huber wrote:
> > > > > Signed-off-by: Sebastian Huber<sebastian.huber@embedded-brains.de>
> > > > > ---
> > > > >    newlib/libc/misc/ffs.c | 11 +++++++++++
> > > > >    1 file changed, 11 insertions(+)
> > > > > 
> > > > > diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
> > > > > index ba5700920..a09cbd3bb 100644
> > > > > --- a/newlib/libc/misc/ffs.c
> > > > > +++ b/newlib/libc/misc/ffs.c
> > > > > @@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
> > > > >    int
> > > > >    ffs(int i)
> > > > >    {
> > > > > +#ifdef __LP64__
> > > > > +	/* GCC would expand the __builtin_ffs() to ffs() in this case */
> > > > > +	int bit;
> > > > > +
> > > > > +	if (i == 0)
> > > > > +		return (0);
> > > > > +	for (bit = 1; !(i & 1); bit++)
> > > > > +		i = (unsigned int)i >> 1;
> > > > > +	return (bit);
> > > > If we're going to open-code it to work around the compiler creating an
> > > > infloop recursion to ffs(), at least code a straight-line version
> > > > without branches, rather than the painfully slow bit-by-bit loop.
> > > > There's plenty of examples on the web of writing ffs() by using
> > > > bit-twiddling without branching.
> > > This is roughly the same implementation we had before. I do not intend to
> > > optimize this.
> > Still, __LP64__ is unacceptable.  Cygwin would be affected by this as
> > well and would have to revert to its former own ffs implementation.
> > 
> > Reverting to a C-based implementation should only be performed on a
> > case-by-case basis.
> 
> Yes, so maybe something like this
> 
> #if defined(__LP64__) && defined(__riscv)
> 
> or a target-specific ffs.c file similar to memcpy.c, etc.

I'm inclined to favor a target-specific file.  This would also allow
to implement the replacement in assembler easily.


Corinna

-- 
Corinna Vinschen
Cygwin Maintainer
Red Hat

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 819 bytes --]

Re: [PATCH] Workaround for ffs() on LP64 targets

[Sebastian Huber]

On 27/07/17 14:48, Corinna Vinschen wrote:

> On Jul 27 14:33, Sebastian Huber wrote:
>> On 27/07/17 14:27, Corinna Vinschen wrote:
>>
>>> On Jul 27 13:24, Sebastian Huber wrote:
>>>> On 27/07/17 13:13, Eric Blake wrote:
>>>>
>>>>> On 07/27/2017 03:06 AM, Sebastian Huber wrote:
>>>>>> Signed-off-by: Sebastian Huber<sebastian.huber@embedded-brains.de>
>>>>>> ---
>>>>>>     newlib/libc/misc/ffs.c | 11 +++++++++++
>>>>>>     1 file changed, 11 insertions(+)
>>>>>>
>>>>>> diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
>>>>>> index ba5700920..a09cbd3bb 100644
>>>>>> --- a/newlib/libc/misc/ffs.c
>>>>>> +++ b/newlib/libc/misc/ffs.c
>>>>>> @@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
>>>>>>     int
>>>>>>     ffs(int i)
>>>>>>     {
>>>>>> +#ifdef __LP64__
>>>>>> +	/* GCC would expand the __builtin_ffs() to ffs() in this case */
>>>>>> +	int bit;
>>>>>> +
>>>>>> +	if (i == 0)
>>>>>> +		return (0);
>>>>>> +	for (bit = 1; !(i & 1); bit++)
>>>>>> +		i = (unsigned int)i >> 1;
>>>>>> +	return (bit);
>>>>> If we're going to open-code it to work around the compiler creating an
>>>>> infloop recursion to ffs(), at least code a straight-line version
>>>>> without branches, rather than the painfully slow bit-by-bit loop.
>>>>> There's plenty of examples on the web of writing ffs() by using
>>>>> bit-twiddling without branching.
>>>> This is roughly the same implementation we had before. I do not intend to
>>>> optimize this.
>>> Still, __LP64__ is unacceptable.  Cygwin would be affected by this as
>>> well and would have to revert to its former own ffs implementation.
>>>
>>> Reverting to a C-based implementation should only be performed on a
>>> case-by-case basis.
>> Yes, so maybe something like this
>>
>> #if defined(__LP64__) && defined(__riscv)
>>
>> or a target-specific ffs.c file similar to memcpy.c, etc.
> I'm inclined to favor a target-specific file.  This would also allow
> to implement the replacement in assembler easily.

Ok, good. So, this is something for the new RISC-V port.

-- 
Sebastian Huber, embedded brains GmbH

Address : Dornierstr. 4, D-82178 Puchheim, Germany
Phone   : +49 89 189 47 41-16
Fax     : +49 89 189 47 41-09
E-Mail  : sebastian.huber@embedded-brains.de
PGP     : Public key available on request.

Diese Nachricht ist keine geschäftliche Mitteilung im Sinne des EHUG.

Re: [PATCH] Workaround for ffs() on LP64 targets

[Brian Inglis]

On 2017-07-27 05:24, Sebastian Huber wrote:
> On 27/07/17 13:13, Eric Blake wrote:
>> On 07/27/2017 03:06 AM, Sebastian Huber wrote:
>>> Signed-off-by: Sebastian Huber
>>> ---
>>>   newlib/libc/misc/ffs.c | 11 +++++++++++
>>>   1 file changed, 11 insertions(+)
>>>
>>> diff --git a/newlib/libc/misc/ffs.c b/newlib/libc/misc/ffs.c
>>> index ba5700920..a09cbd3bb 100644
>>> --- a/newlib/libc/misc/ffs.c
>>> +++ b/newlib/libc/misc/ffs.c
>>> @@ -31,6 +31,17 @@ No supporting OS subroutines are required.  */
>>>   int
>>>   ffs(int i)
>>>   {
>>> +#ifdef __LP64__
>>> +    /* GCC would expand the __builtin_ffs() to ffs() in this case */
>>> +    int bit;
>>> +
>>> +    if (i == 0)
>>> +        return (0);
>>> +    for (bit = 1; !(i & 1); bit++)
>>> +        i = (unsigned int)i >> 1;
>>> +    return (bit);
>> If we're going to open-code it to work around the compiler creating an
>> infloop recursion to ffs(), at least code a straight-line version
>> without branches, rather than the painfully slow bit-by-bit loop.
>> There's plenty of examples on the web of writing ffs() by using
>> bit-twiddling without branching.

Definitive twiddling reference is now Hacker's Delight 2nd ed, Henry S. Warren,
Jr., 2013, Pearson/InformIT/AW; available in ebook formats:
https://github.com/jyfc/ebook/blob/master/02_algorithm/Hacker's%20Delight%202nd%20Edition.pdf

https://www.safaribooksonline.com/library/view/hackers-delight-second/9780133084993/

https://en.wikipedia.org/wiki/Hacker%27s_Delight
http://www.hackersdelight.org/

> This is roughly the same implementation we had before. I do not intend to 
> optimize this.

Programmers using these functions expect the usage cost to be trivial and fairly
constant ~ O(log2(bits)) not O(bits); if not, they may implement their own!

Try this one, seems decently short; adjust for different word sizes; with gcc
-O3 on x86-64 compiles to 32 instructions branch free: YMMV

int
ffsll( long long in )
{
/* find first set == 1 + count trailing zeros */
	int index = 64;


	if (!in)			return 0;

	in &= -in;			/* clear all but lsb set */
/*
 * for ctz remove above test and add next line
 *	if (in)				--index;
 */
	if (in & 0x00000000FFFFFFFF)	index -= 32;
	if (in & 0x0000FFFF0000FFFF)	index -= 16;
	if (in & 0x00FF00FF00FF00FF)	index -= 8;
	if (in & 0x0F0F0F0F0F0F0F0F)	index -= 4;
	if (in & 0x3333333333333333)	index -= 2;
	if (in & 0x5555555555555555)	index -= 1;

	return index;
}

-- 
Take care. Thanks, Brian Inglis, Calgary, Alberta, Canada