[PATCH] arm: Use builtins for ffs/fls

[PATCH] arm: Use builtins for ffs/fls

[Sean Anderson]

Since ARMv5, the clz instruction allows for efficient implementation of
ffs/fls with builtins. Until ARMv7 (with Thumb-2), this instruction is
only available in ARM mode. LTO makes it difficult to force specific
functions to be in ARM mode, as it is effectively a form of very
aggressive inlining. To work around this, fls/ffs are implemented in
assembly for ARMv5 and ARMv6 when compiling U-Boot in Thumb mode.
Overall, this saves around 75 bytes per call.

Signed-off-by: Sean Anderson <sean.anderson@seco.com>
---
I only tested this on ARMv8. If someone has an ARMv5 or ARMv6 board,
please test this.

 arch/arm/include/asm/bitops.h              | 27 ++++++++++++-
 arch/arm/lib/Makefile                      |  5 +++
 arch/arm/lib/bitops.S                      | 45 ++++++++++++++++++++++
 include/asm-generic/bitops/builtin-__ffs.h | 16 ++++++++
 include/asm-generic/bitops/builtin-__fls.h | 16 ++++++++
 include/asm-generic/bitops/builtin-ffs.h   | 15 ++++++++
 include/asm-generic/bitops/builtin-fls.h   | 17 ++++++++
 7 files changed, 140 insertions(+), 1 deletion(-)
 create mode 100644 arch/arm/lib/bitops.S
 create mode 100644 include/asm-generic/bitops/builtin-__ffs.h
 create mode 100644 include/asm-generic/bitops/builtin-__fls.h
 create mode 100644 include/asm-generic/bitops/builtin-ffs.h
 create mode 100644 include/asm-generic/bitops/builtin-fls.h

diff --git a/arch/arm/include/asm/bitops.h b/arch/arm/include/asm/bitops.h
index fa8548624a..8e897833bb 100644
--- a/arch/arm/include/asm/bitops.h
+++ b/arch/arm/include/asm/bitops.h
@@ -15,9 +15,34 @@
 #ifndef __ASM_ARM_BITOPS_H
 #define __ASM_ARM_BITOPS_H
 
+#if __LINUX_ARM_ARCH__ < 5
+
 #include <asm-generic/bitops/__ffs.h>
 #include <asm-generic/bitops/__fls.h>
 #include <asm-generic/bitops/fls.h>
+
+#else
+
+#define PLATFORM_FFS
+#define PLATFORM_FLS
+
+#if !IS_ENABLED(CONFIG_HAS_THUMB2) && CONFIG_IS_ENABLED(SYS_THUMB_BUILD)
+
+unsigned long __fls(unsigned long word);
+unsigned long __ffs(unsigned long word);
+int fls(unsigned int x);
+int ffs(int x);
+
+#else
+
+#include <asm-generic/bitops/builtin-__fls.h>
+#include <asm-generic/bitops/builtin-__ffs.h>
+#include <asm-generic/bitops/builtin-fls.h>
+#include <asm-generic/bitops/builtin-ffs.h>
+
+#endif
+#endif
+
 #include <asm-generic/bitops/fls64.h>
 
 #ifdef __KERNEL__
@@ -113,7 +138,7 @@ static inline int test_bit(int nr, const void * addr)
 
 static inline int __ilog2(unsigned int x)
 {
-	return generic_fls(x) - 1;
+	return fls(x) - 1;
 }
 
 #define ffz(x)  __ffs(~(x))
diff --git a/arch/arm/lib/Makefile b/arch/arm/lib/Makefile
index 62cf80f373..b1bcd37466 100644
--- a/arch/arm/lib/Makefile
+++ b/arch/arm/lib/Makefile
@@ -113,6 +113,11 @@ AFLAGS_REMOVE_memset.o := -mthumb -mthumb-interwork
 AFLAGS_REMOVE_memcpy.o := -mthumb -mthumb-interwork
 AFLAGS_memset.o := -DMEMSET_NO_THUMB_BUILD
 AFLAGS_memcpy.o := -DMEMCPY_NO_THUMB_BUILD
+
+# This is only necessary to force ARM mode on THUMB1 targets.
+ifneq ($(CONFIG_SYS_ARM_ARCH),4)
+obj-y   += bitops.o
+endif
 endif
 endif
 
diff --git a/arch/arm/lib/bitops.S b/arch/arm/lib/bitops.S
new file mode 100644
index 0000000000..29d1524634
--- /dev/null
+++ b/arch/arm/lib/bitops.S
@@ -0,0 +1,45 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) 2023 Sean Anderson <sean.anderson@seco.com>
+ *
+ * ARM bitops to call when using THUMB1, which doesn't have these instructions.
+ */
+#include <linux/linkage.h>
+#include <asm/assembler.h>
+
+.pushsection .text.__fls
+ENTRY(__fls)
+	clz	r0, r0
+	rsb	r0, r0, #31
+	ret	lr
+ENDPROC(__fls)
+.popsection
+
+.pushsection .text.__ffs
+ENTRY(__ffs)
+	rsb	r3, r0, #0
+	and	r0, r0, r3
+	clz	r0, r0
+	rsb	r0, r0, #31
+	ret	lr
+ENDPROC(__ffs)
+.popsection
+
+.pushsection .text.fls
+ENTRY(fls)
+	cmp	r0, #0
+	clzne	r0, r0
+	rsbne	r0, r0, #32
+	ret	lr
+ENDPROC(fls)
+.popsection
+
+.pushsection .text.ffs
+ENTRY(ffs)
+	rsb	r3, r0, #0
+	and	r0, r0, r3
+	clz	r0, r0
+	rsb	r0, r0, #32
+	ret	lr
+ENDPROC(ffs)
+.popsection
diff --git a/include/asm-generic/bitops/builtin-__ffs.h b/include/asm-generic/bitops/builtin-__ffs.h
new file mode 100644
index 0000000000..87024da44d
--- /dev/null
+++ b/include/asm-generic/bitops/builtin-__ffs.h
@@ -0,0 +1,16 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_GENERIC_BITOPS_BUILTIN___FFS_H_
+#define _ASM_GENERIC_BITOPS_BUILTIN___FFS_H_
+
+/**
+ * __ffs - find first bit in word.
+ * @word: The word to search
+ *
+ * Undefined if no bit exists, so code should check against 0 first.
+ */
+static __always_inline unsigned long __ffs(unsigned long word)
+{
+	return __builtin_ctzl(word);
+}
+
+#endif
diff --git a/include/asm-generic/bitops/builtin-__fls.h b/include/asm-generic/bitops/builtin-__fls.h
new file mode 100644
index 0000000000..43a5aa9afb
--- /dev/null
+++ b/include/asm-generic/bitops/builtin-__fls.h
@@ -0,0 +1,16 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_GENERIC_BITOPS_BUILTIN___FLS_H_
+#define _ASM_GENERIC_BITOPS_BUILTIN___FLS_H_
+
+/**
+ * __fls - find last (most-significant) set bit in a long word
+ * @word: the word to search
+ *
+ * Undefined if no set bit exists, so code should check against 0 first.
+ */
+static __always_inline unsigned long __fls(unsigned long word)
+{
+	return (sizeof(word) * 8) - 1 - __builtin_clzl(word);
+}
+
+#endif
diff --git a/include/asm-generic/bitops/builtin-ffs.h b/include/asm-generic/bitops/builtin-ffs.h
new file mode 100644
index 0000000000..7b12932904
--- /dev/null
+++ b/include/asm-generic/bitops/builtin-ffs.h
@@ -0,0 +1,15 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_GENERIC_BITOPS_BUILTIN_FFS_H_
+#define _ASM_GENERIC_BITOPS_BUILTIN_FFS_H_
+
+/**
+ * ffs - find first bit set
+ * @x: the word to search
+ *
+ * This is defined the same way as
+ * the libc and compiler builtin ffs routines, therefore
+ * differs in spirit from ffz (man ffs).
+ */
+#define ffs(x) __builtin_ffs(x)
+
+#endif
diff --git a/include/asm-generic/bitops/builtin-fls.h b/include/asm-generic/bitops/builtin-fls.h
new file mode 100644
index 0000000000..c8455cc288
--- /dev/null
+++ b/include/asm-generic/bitops/builtin-fls.h
@@ -0,0 +1,17 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_GENERIC_BITOPS_BUILTIN_FLS_H_
+#define _ASM_GENERIC_BITOPS_BUILTIN_FLS_H_
+
+/**
+ * fls - find last (most-significant) bit set
+ * @x: the word to search
+ *
+ * This is defined the same way as ffs.
+ * Note fls(0) = 0, fls(1) = 1, fls(0x80000000) = 32.
+ */
+static __always_inline int fls(unsigned int x)
+{
+	return x ? sizeof(x) * 8 - __builtin_clz(x) : 0;
+}
+
+#endif
-- 
2.40.1

Re: [PATCH] arm: Use builtins for ffs/fls

[Tom Rini]

[-- Attachment #1: Type: text/plain, Size: 676 bytes --]

On Mon, Jul 31, 2023 at 05:27:33PM -0400, Sean Anderson wrote:

> Since ARMv5, the clz instruction allows for efficient implementation of
> ffs/fls with builtins. Until ARMv7 (with Thumb-2), this instruction is
> only available in ARM mode. LTO makes it difficult to force specific
> functions to be in ARM mode, as it is effectively a form of very
> aggressive inlining. To work around this, fls/ffs are implemented in
> assembly for ARMv5 and ARMv6 when compiling U-Boot in Thumb mode.
> Overall, this saves around 75 bytes per call.
> 
> Signed-off-by: Sean Anderson <sean.anderson@seco.com>

This looks like it's synced from the kernel, what tag?

-- 
Tom

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 659 bytes --]

Re: [PATCH] arm: Use builtins for ffs/fls

[Sean Anderson]

On 7/31/23 17:36, Tom Rini wrote:
> On Mon, Jul 31, 2023 at 05:27:33PM -0400, Sean Anderson wrote:
> 
>> Since ARMv5, the clz instruction allows for efficient implementation of
>> ffs/fls with builtins. Until ARMv7 (with Thumb-2), this instruction is
>> only available in ARM mode. LTO makes it difficult to force specific
>> functions to be in ARM mode, as it is effectively a form of very
>> aggressive inlining. To work around this, fls/ffs are implemented in
>> assembly for ARMv5 and ARMv6 when compiling U-Boot in Thumb mode.
>> Overall, this saves around 75 bytes per call.
>>
>> Signed-off-by: Sean Anderson <sean.anderson@seco.com>
> 
> This looks like it's synced from the kernel, what tag?
> 

The builtins are synced from v5.15 and they haven't changed since.

--Sean

Re: [PATCH] arm: Use builtins for ffs/fls

[Sean Anderson]

On 7/31/23 17:27, Sean Anderson wrote:
> Since ARMv5, the clz instruction allows for efficient implementation of
> ffs/fls with builtins. Until ARMv7 (with Thumb-2), this instruction is
> only available in ARM mode. LTO makes it difficult to force specific
> functions to be in ARM mode, as it is effectively a form of very
> aggressive inlining. To work around this, fls/ffs are implemented in
> assembly for ARMv5 and ARMv6 when compiling U-Boot in Thumb mode.
> Overall, this saves around 75 bytes per call.
> 
> Signed-off-by: Sean Anderson <sean.anderson@seco.com>
> ---
> I only tested this on ARMv8. If someone has an ARMv5 or ARMv6 board,
> please test this.
>

+CC Linus since he may care about integrator.

--Sean

Re: [PATCH] arm: Use builtins for ffs/fls

[Tom Rini]

[-- Attachment #1: Type: text/plain, Size: 748 bytes --]

On Mon, Jul 31, 2023 at 05:27:33PM -0400, Sean Anderson wrote:

> Since ARMv5, the clz instruction allows for efficient implementation of
> ffs/fls with builtins. Until ARMv7 (with Thumb-2), this instruction is
> only available in ARM mode. LTO makes it difficult to force specific
> functions to be in ARM mode, as it is effectively a form of very
> aggressive inlining. To work around this, fls/ffs are implemented in
> assembly for ARMv5 and ARMv6 when compiling U-Boot in Thumb mode.
> Overall, this saves around 75 bytes per call.
> 
> Signed-off-by: Sean Anderson <sean.anderson@seco.com>

I'll add some wording about when this was synced from the kernel when
applying.

Reviewed-by: Tom Rini <trini@konsulko.com>

-- 
Tom

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 659 bytes --]

Re: [PATCH] arm: Use builtins for ffs/fls

[Tom Rini]

[-- Attachment #1: Type: text/plain, Size: 758 bytes --]

On Mon, Jul 31, 2023 at 05:27:33PM -0400, Sean Anderson wrote:

> Since ARMv5, the clz instruction allows for efficient implementation of
> ffs/fls with builtins. Until ARMv7 (with Thumb-2), this instruction is
> only available in ARM mode. LTO makes it difficult to force specific
> functions to be in ARM mode, as it is effectively a form of very
> aggressive inlining. To work around this, fls/ffs are implemented in
> assembly for ARMv5 and ARMv6 when compiling U-Boot in Thumb mode.
> Overall, this saves around 75 bytes per call.
> 
> This code is synced with v5.15 of the Linux kernel.
> 
> Signed-off-by: Sean Anderson <sean.anderson@seco.com>
> Reviewed-by: Tom Rini <trini@konsulko.com>

Applied to u-boot/next, thanks!

-- 
Tom

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 659 bytes --]

Re: [PATCH] arm: Use builtins for ffs/fls

[Tony Dinh]

Hi Sean,

On Mon, Jul 31, 2023 at 2:28 PM Sean Anderson <sean.anderson@seco.com> wrote:
>
> Since ARMv5, the clz instruction allows for efficient implementation of
> ffs/fls with builtins. Until ARMv7 (with Thumb-2), this instruction is
> only available in ARM mode. LTO makes it difficult to force specific
> functions to be in ARM mode, as it is effectively a form of very
> aggressive inlining. To work around this, fls/ffs are implemented in
> assembly for ARMv5 and ARMv6 when compiling U-Boot in Thumb mode.
> Overall, this saves around 75 bytes per call.
>
> Signed-off-by: Sean Anderson <sean.anderson@seco.com>
> ---
> I only tested this on ARMv8. If someone has an ARMv5 or ARMv6 board,
> please test this.

Thanks for the patch! I've tested it with the Pogo V4 board (Kirkwood
88F6192, ARMv5) with LTO and SYS_THUMB_BUILD.

Also nice size reduction:
Before: 502K
After: 499K

So for ARMv5,
Tested-by: Tony Dinh <mibodhi@gmail.com>

All the best,
Tony

>
>  arch/arm/include/asm/bitops.h              | 27 ++++++++++++-
>  arch/arm/lib/Makefile                      |  5 +++
>  arch/arm/lib/bitops.S                      | 45 ++++++++++++++++++++++
>  include/asm-generic/bitops/builtin-__ffs.h | 16 ++++++++
>  include/asm-generic/bitops/builtin-__fls.h | 16 ++++++++
>  include/asm-generic/bitops/builtin-ffs.h   | 15 ++++++++
>  include/asm-generic/bitops/builtin-fls.h   | 17 ++++++++
>  7 files changed, 140 insertions(+), 1 deletion(-)
>  create mode 100644 arch/arm/lib/bitops.S
>  create mode 100644 include/asm-generic/bitops/builtin-__ffs.h
>  create mode 100644 include/asm-generic/bitops/builtin-__fls.h
>  create mode 100644 include/asm-generic/bitops/builtin-ffs.h
>  create mode 100644 include/asm-generic/bitops/builtin-fls.h
>
> diff --git a/arch/arm/include/asm/bitops.h b/arch/arm/include/asm/bitops.h
> index fa8548624a..8e897833bb 100644
> --- a/arch/arm/include/asm/bitops.h
> +++ b/arch/arm/include/asm/bitops.h
> @@ -15,9 +15,34 @@
>  #ifndef __ASM_ARM_BITOPS_H
>  #define __ASM_ARM_BITOPS_H
>
> +#if __LINUX_ARM_ARCH__ < 5
> +
>  #include <asm-generic/bitops/__ffs.h>
>  #include <asm-generic/bitops/__fls.h>
>  #include <asm-generic/bitops/fls.h>
> +
> +#else
> +
> +#define PLATFORM_FFS
> +#define PLATFORM_FLS
> +
> +#if !IS_ENABLED(CONFIG_HAS_THUMB2) && CONFIG_IS_ENABLED(SYS_THUMB_BUILD)
> +
> +unsigned long __fls(unsigned long word);
> +unsigned long __ffs(unsigned long word);
> +int fls(unsigned int x);
> +int ffs(int x);
> +
> +#else
> +
> +#include <asm-generic/bitops/builtin-__fls.h>
> +#include <asm-generic/bitops/builtin-__ffs.h>
> +#include <asm-generic/bitops/builtin-fls.h>
> +#include <asm-generic/bitops/builtin-ffs.h>
> +
> +#endif
> +#endif
> +
>  #include <asm-generic/bitops/fls64.h>
>
>  #ifdef __KERNEL__
> @@ -113,7 +138,7 @@ static inline int test_bit(int nr, const void * addr)
>
>  static inline int __ilog2(unsigned int x)
>  {
> -       return generic_fls(x) - 1;
> +       return fls(x) - 1;
>  }
>
>  #define ffz(x)  __ffs(~(x))
> diff --git a/arch/arm/lib/Makefile b/arch/arm/lib/Makefile
> index 62cf80f373..b1bcd37466 100644
> --- a/arch/arm/lib/Makefile
> +++ b/arch/arm/lib/Makefile
> @@ -113,6 +113,11 @@ AFLAGS_REMOVE_memset.o := -mthumb -mthumb-interwork
>  AFLAGS_REMOVE_memcpy.o := -mthumb -mthumb-interwork
>  AFLAGS_memset.o := -DMEMSET_NO_THUMB_BUILD
>  AFLAGS_memcpy.o := -DMEMCPY_NO_THUMB_BUILD
> +
> +# This is only necessary to force ARM mode on THUMB1 targets.
> +ifneq ($(CONFIG_SYS_ARM_ARCH),4)
> +obj-y   += bitops.o
> +endif
>  endif
>  endif
>
> diff --git a/arch/arm/lib/bitops.S b/arch/arm/lib/bitops.S
> new file mode 100644
> index 0000000000..29d1524634
> --- /dev/null
> +++ b/arch/arm/lib/bitops.S
> @@ -0,0 +1,45 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +/*
> + * Copyright (C) 2023 Sean Anderson <sean.anderson@seco.com>
> + *
> + * ARM bitops to call when using THUMB1, which doesn't have these instructions.
> + */
> +#include <linux/linkage.h>
> +#include <asm/assembler.h>
> +
> +.pushsection .text.__fls
> +ENTRY(__fls)
> +       clz     r0, r0
> +       rsb     r0, r0, #31
> +       ret     lr
> +ENDPROC(__fls)
> +.popsection
> +
> +.pushsection .text.__ffs
> +ENTRY(__ffs)
> +       rsb     r3, r0, #0
> +       and     r0, r0, r3
> +       clz     r0, r0
> +       rsb     r0, r0, #31
> +       ret     lr
> +ENDPROC(__ffs)
> +.popsection
> +
> +.pushsection .text.fls
> +ENTRY(fls)
> +       cmp     r0, #0
> +       clzne   r0, r0
> +       rsbne   r0, r0, #32
> +       ret     lr
> +ENDPROC(fls)
> +.popsection
> +
> +.pushsection .text.ffs
> +ENTRY(ffs)
> +       rsb     r3, r0, #0
> +       and     r0, r0, r3
> +       clz     r0, r0
> +       rsb     r0, r0, #32
> +       ret     lr
> +ENDPROC(ffs)
> +.popsection
> diff --git a/include/asm-generic/bitops/builtin-__ffs.h b/include/asm-generic/bitops/builtin-__ffs.h
> new file mode 100644
> index 0000000000..87024da44d
> --- /dev/null
> +++ b/include/asm-generic/bitops/builtin-__ffs.h
> @@ -0,0 +1,16 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef _ASM_GENERIC_BITOPS_BUILTIN___FFS_H_
> +#define _ASM_GENERIC_BITOPS_BUILTIN___FFS_H_
> +
> +/**
> + * __ffs - find first bit in word.
> + * @word: The word to search
> + *
> + * Undefined if no bit exists, so code should check against 0 first.
> + */
> +static __always_inline unsigned long __ffs(unsigned long word)
> +{
> +       return __builtin_ctzl(word);
> +}
> +
> +#endif
> diff --git a/include/asm-generic/bitops/builtin-__fls.h b/include/asm-generic/bitops/builtin-__fls.h
> new file mode 100644
> index 0000000000..43a5aa9afb
> --- /dev/null
> +++ b/include/asm-generic/bitops/builtin-__fls.h
> @@ -0,0 +1,16 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef _ASM_GENERIC_BITOPS_BUILTIN___FLS_H_
> +#define _ASM_GENERIC_BITOPS_BUILTIN___FLS_H_
> +
> +/**
> + * __fls - find last (most-significant) set bit in a long word
> + * @word: the word to search
> + *
> + * Undefined if no set bit exists, so code should check against 0 first.
> + */
> +static __always_inline unsigned long __fls(unsigned long word)
> +{
> +       return (sizeof(word) * 8) - 1 - __builtin_clzl(word);
> +}
> +
> +#endif
> diff --git a/include/asm-generic/bitops/builtin-ffs.h b/include/asm-generic/bitops/builtin-ffs.h
> new file mode 100644
> index 0000000000..7b12932904
> --- /dev/null
> +++ b/include/asm-generic/bitops/builtin-ffs.h
> @@ -0,0 +1,15 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef _ASM_GENERIC_BITOPS_BUILTIN_FFS_H_
> +#define _ASM_GENERIC_BITOPS_BUILTIN_FFS_H_
> +
> +/**
> + * ffs - find first bit set
> + * @x: the word to search
> + *
> + * This is defined the same way as
> + * the libc and compiler builtin ffs routines, therefore
> + * differs in spirit from ffz (man ffs).
> + */
> +#define ffs(x) __builtin_ffs(x)
> +
> +#endif
> diff --git a/include/asm-generic/bitops/builtin-fls.h b/include/asm-generic/bitops/builtin-fls.h
> new file mode 100644
> index 0000000000..c8455cc288
> --- /dev/null
> +++ b/include/asm-generic/bitops/builtin-fls.h
> @@ -0,0 +1,17 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef _ASM_GENERIC_BITOPS_BUILTIN_FLS_H_
> +#define _ASM_GENERIC_BITOPS_BUILTIN_FLS_H_
> +
> +/**
> + * fls - find last (most-significant) bit set
> + * @x: the word to search
> + *
> + * This is defined the same way as ffs.
> + * Note fls(0) = 0, fls(1) = 1, fls(0x80000000) = 32.
> + */
> +static __always_inline int fls(unsigned int x)
> +{
> +       return x ? sizeof(x) * 8 - __builtin_clz(x) : 0;
> +}
> +
> +#endif
> --
> 2.40.1
>