[PATCH] powerpc: Discard ffs() function and use builtin_ffs instead

With the ffs() function as defined in arch/powerpc/include/asm/bitops.h
GCC will not optimise the code in case of constant parameter, as shown
by the small exemple below.

int ffs_test(void)
{
	return 4 << ffs(31);
}

c0012334 <ffs_test>:
c0012334:       39 20 00 01     li      r9,1
c0012338:       38 60 00 04     li      r3,4
c001233c:       7d 29 00 34     cntlzw  r9,r9
c0012340:       21 29 00 20     subfic  r9,r9,32
c0012344:       7c 63 48 30     slw     r3,r3,r9
c0012348:       4e 80 00 20     blr

With this patch, the same function will compile as follows:

c0012334 <ffs_test>:
c0012334:       38 60 00 08     li      r3,8
c0012338:       4e 80 00 20     blr

Signed-off-by: Christophe Leroy <redacted>
---
 arch/powerpc/include/asm/bitops.h | 11 +----------
 1 file changed, 1 insertion(+), 10 deletions(-)

diff --git a/arch/powerpc/include/asm/bitops.h b/arch/powerpc/include/asm/bitops.h
index 59abc62..75e3ebb 100644
--- a/arch/powerpc/include/asm/bitops.h
+++ b/arch/powerpc/include/asm/bitops.h

@@ -218,16 +218,7 @@ static __inline__ unsigned long __ffs(unsigned long x)
 	return __ilog2(x & -x);
 }
 
-/*
- * ffs: find first bit set. This is defined the same way as
- * the libc and compiler builtin ffs routines, therefore
- * differs in spirit from the above ffz (man ffs).
- */
-static __inline__ int ffs(int x)
-{
-	unsigned long i = (unsigned long)x;
-	return __ilog2(i & -i) + 1;
-}
+#include <asm-generic/bitops/builtin-ffs.h>
 
 /*
  * fls: find last (most-significant) bit set.

-- 
2.1.0

On Thu, 2016-12-05 at 15:32:22 UTC, Christophe Leroy wrote:

With the ffs() function as defined in arch/powerpc/include/asm/bitops.h
GCC will not optimise the code in case of constant parameter, as shown
by the small exemple below.

int ffs_test(void)
{
	return 4 << ffs(31);
}

c0012334 <ffs_test>:
c0012334:       39 20 00 01     li      r9,1
c0012338:       38 60 00 04     li      r3,4
c001233c:       7d 29 00 34     cntlzw  r9,r9
c0012340:       21 29 00 20     subfic  r9,r9,32
c0012344:       7c 63 48 30     slw     r3,r3,r9
c0012348:       4e 80 00 20     blr

With this patch, the same function will compile as follows:

c0012334 <ffs_test>:
c0012334:       38 60 00 08     li      r3,8
c0012338:       4e 80 00 20     blr

But what code does it generate when it's not a constant?

And which gcc version first added the builtin version?

cheers

Le 13/05/2016 à 08:16, Michael Ellerman a écrit :

On Thu, 2016-12-05 at 15:32:22 UTC, Christophe Leroy wrote:

quoted

With the ffs() function as defined in arch/powerpc/include/asm/bitops.h
GCC will not optimise the code in case of constant parameter, as shown
by the small exemple below.

int ffs_test(void)
{
	return 4 << ffs(31);
}

c0012334 <ffs_test>:
c0012334:       39 20 00 01     li      r9,1
c0012338:       38 60 00 04     li      r3,4
c001233c:       7d 29 00 34     cntlzw  r9,r9
c0012340:       21 29 00 20     subfic  r9,r9,32
c0012344:       7c 63 48 30     slw     r3,r3,r9
c0012348:       4e 80 00 20     blr

With this patch, the same function will compile as follows:

c0012334 <ffs_test>:
c0012334:       38 60 00 08     li      r3,8
c0012338:       4e 80 00 20     blr

But what code does it generate when it's not a constant?

The generated code is the same with and without the patch when not a constant:

int ffs_test2(int x)
{
    return ffs(x);
}

c001233c <ffs_test2>:
c001233c:       7d 23 00 d0     neg     r9,r3
c0012340:       7d 23 18 38     and     r3,r9,r3
c0012344:       7c 63 00 34     cntlzw  r3,r3
c0012348:       20 63 00 20     subfic  r3,r3,32
c001234c:       4e 80 00 20     blr

And which gcc version first added the builtin version?

Don't know, but __builtin_ffs() is already used in arch/powerpc/include/asm/page_32.h

Christophe

On Fri, May 13, 2016 at 04:16:57PM +1000, Michael Ellerman wrote:

On Thu, 2016-12-05 at 15:32:22 UTC, Christophe Leroy wrote:

quoted

With the ffs() function as defined in arch/powerpc/include/asm/bitops.h
GCC will not optimise the code in case of constant parameter, as shown
by the small exemple below.

int ffs_test(void)
{
	return 4 << ffs(31);
}

c0012334 <ffs_test>:
c0012334:       39 20 00 01     li      r9,1
c0012338:       38 60 00 04     li      r3,4
c001233c:       7d 29 00 34     cntlzw  r9,r9
c0012340:       21 29 00 20     subfic  r9,r9,32
c0012344:       7c 63 48 30     slw     r3,r3,r9
c0012348:       4e 80 00 20     blr

With this patch, the same function will compile as follows:

c0012334 <ffs_test>:
c0012334:       38 60 00 08     li      r3,8
c0012338:       4e 80 00 20     blr

But what code does it generate when it's not a constant?

And which gcc version first added the builtin version?

It already existed in gcc-2.95, which you do not want to use to compile
anything today but I have in a corner of a chroot environment to maintain
~1997 vintage embedded stuff, running a 2.2.12 kernel!

Hopefully this clears up your concerns :-)

    Cheers,
    Gabriel

Le 13/05/2016 à 08:53, Christophe Leroy a écrit :

Le 13/05/2016 à 08:16, Michael Ellerman a écrit :

quoted

On Thu, 2016-12-05 at 15:32:22 UTC, Christophe Leroy wrote:

quoted

With the ffs() function as defined in arch/powerpc/include/asm/bitops.h
GCC will not optimise the code in case of constant parameter, as shown
by the small exemple below.

int ffs_test(void)
{
    return 4 << ffs(31);
}

c0012334 <ffs_test>:
c0012334:       39 20 00 01     li      r9,1
c0012338:       38 60 00 04     li      r3,4
c001233c:       7d 29 00 34     cntlzw  r9,r9
c0012340:       21 29 00 20     subfic  r9,r9,32
c0012344:       7c 63 48 30     slw     r3,r3,r9
c0012348:       4e 80 00 20     blr

With this patch, the same function will compile as follows:

c0012334 <ffs_test>:
c0012334:       38 60 00 08     li      r3,8
c0012338:       4e 80 00 20     blr

But what code does it generate when it's not a constant?

The generated code is the same with and without the patch when not a constant:

int ffs_test2(int x)
{
    return ffs(x);
}

c001233c <ffs_test2>:
c001233c:       7d 23 00 d0     neg     r9,r3
c0012340:       7d 23 18 38     and     r3,r9,r3
c0012344:       7c 63 00 34     cntlzw  r3,r3
c0012348:       20 63 00 20     subfic  r3,r3,32
c001234c:       4e 80 00 20     blr

quoted

And which gcc version first added the builtin version?

Don't know, but __builtin_ffs() is already used in arch/powerpc/include/asm/page_32.h

Hi Michael,

Any change to get it into 4.9 ?

Christophe