__llsc_futex_atomic_set() is implmented using
LLSC_FUTEX_ATOMIC_OP() macro with "mov  %w3, %w5".
But this instruction isn't required to implement fux_atomic_set()
so make a small optimisation by implementing __llsc_futex_atomic_set()
as separate function.

This will make usage of LLSC_FUTEX_ATOMIC_OP() macro more simple.

Signed-off-by: Yeoreum Yun <yeoreum.yun@arm.com>
---
 arch/arm64/include/asm/futex.h | 43 ++++++++++++++++++++++++++++------
 1 file changed, 36 insertions(+), 7 deletions(-)

diff --git a/arch/arm64/include/asm/futex.h b/arch/arm64/include/asm/futex.h
index ab7003cb4724..22a6301a9f3d 100644
--- a/arch/arm64/include/asm/futex.h
+++ b/arch/arm64/include/asm/futex.h
@@ -13,7 +13,7 @@

 #define LLSC_MAX_LOOPS	128 /* What's the largest number you can think of? */

-#define LLSC_FUTEX_ATOMIC_OP(op, insn)					\
+#define LLSC_FUTEX_ATOMIC_OP(op, asm_op)				\
 static __always_inline int						\
 __llsc_futex_atomic_##op(int oparg, u32 __user *uaddr, int *oval)	\
 {									\
@@ -24,7 +24,7 @@ __llsc_futex_atomic_##op(int oparg, u32 __user *uaddr, int *oval)	\
 	asm volatile("// __llsc_futex_atomic_" #op "\n"			\
 "	prfm	pstl1strm, %2\n"					\
 "1:	ldxr	%w1, %2\n"						\
-	insn "\n"							\
+"	" #asm_op "	%w3, %w1, %w5\n"				\
 "2:	stlxr	%w0, %w3, %2\n"						\
 "	cbz	%w0, 3f\n"						\
 "	sub	%w4, %w4, %w0\n"					\
@@ -46,11 +46,40 @@ __llsc_futex_atomic_##op(int oparg, u32 __user *uaddr, int *oval)	\
 	return ret;							\
 }

-LLSC_FUTEX_ATOMIC_OP(add, "add	%w3, %w1, %w5")
-LLSC_FUTEX_ATOMIC_OP(or, "orr	%w3, %w1, %w5")
-LLSC_FUTEX_ATOMIC_OP(and, "and	%w3, %w1, %w5")
-LLSC_FUTEX_ATOMIC_OP(eor, "eor	%w3, %w1, %w5")
-LLSC_FUTEX_ATOMIC_OP(set, "mov	%w3, %w5")
+LLSC_FUTEX_ATOMIC_OP(add, add)
+LLSC_FUTEX_ATOMIC_OP(or, orr)
+LLSC_FUTEX_ATOMIC_OP(and, and)
+LLSC_FUTEX_ATOMIC_OP(eor, eor)
+
+static __always_inline int
+__llsc_futex_atomic_set(int oparg, u32 __user *uaddr, int *oval)
+{
+	unsigned int loops = LLSC_MAX_LOOPS;
+	int ret, oldval;
+
+	uaccess_enable_privileged();
+	asm volatile("//__llsc_futex_xchg\n"
+"	prfm	pstl1strm, %2\n"
+"1:	ldxr	%w1, %2\n"
+"2:	stlxr	%w0, %w4, %2\n"
+"	cbz	%w3, 3f\n"
+"	sub	%w3, %w3, %w0\n"
+"	cbnz	%w3, 1b\n"
+"	mov	%w0, %w5\n"
+"3:\n"
+"	dmb	ish\n"
+	_ASM_EXTABLE_UACCESS_ERR(1b, 3b, %w0)
+	_ASM_EXTABLE_UACCESS_ERR(2b, 3b, %w0)
+	: "=&r" (ret), "=&r" (oldval), "+Q" (*uaddr), "+r" (loops)
+	: "r" (oparg), "Ir" (-EAGAIN)
+	: "memory");
+	uaccess_disable_privileged();
+
+	if (!ret)
+		*oval = oldval;
+
+	return ret;
+}

 static __always_inline int
 __llsc_futex_cmpxchg(u32 __user *uaddr, u32 oldval, u32 newval, u32 *oval)
--
LEVI:{C3F47F37-75D8-414A-A8BA-3980EC8A46D7}

On Sat, Aug 16, 2025 at 04:19:28PM +0100, Yeoreum Yun wrote:
> __llsc_futex_atomic_set() is implmented using
> LLSC_FUTEX_ATOMIC_OP() macro with "mov  %w3, %w5".
> But this instruction isn't required to implement fux_atomic_set()
> so make a small optimisation by implementing __llsc_futex_atomic_set()
> as separate function.
> 
> This will make usage of LLSC_FUTEX_ATOMIC_OP() macro more simple.
> 
> Signed-off-by: Yeoreum Yun <yeoreum.yun@arm.com>
> ---
>  arch/arm64/include/asm/futex.h | 43 ++++++++++++++++++++++++++++------
>  1 file changed, 36 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/futex.h b/arch/arm64/include/asm/futex.h
> index ab7003cb4724..22a6301a9f3d 100644
> --- a/arch/arm64/include/asm/futex.h
> +++ b/arch/arm64/include/asm/futex.h
> @@ -13,7 +13,7 @@
> 
>  #define LLSC_MAX_LOOPS	128 /* What's the largest number you can think of? */
> 
> -#define LLSC_FUTEX_ATOMIC_OP(op, insn)					\
> +#define LLSC_FUTEX_ATOMIC_OP(op, asm_op)				\
>  static __always_inline int						\
>  __llsc_futex_atomic_##op(int oparg, u32 __user *uaddr, int *oval)	\
>  {									\
> @@ -24,7 +24,7 @@ __llsc_futex_atomic_##op(int oparg, u32 __user *uaddr, int *oval)	\
>  	asm volatile("// __llsc_futex_atomic_" #op "\n"			\
>  "	prfm	pstl1strm, %2\n"					\
>  "1:	ldxr	%w1, %2\n"						\
> -	insn "\n"							\
> +"	" #asm_op "	%w3, %w1, %w5\n"				\
>  "2:	stlxr	%w0, %w3, %2\n"						\
>  "	cbz	%w0, 3f\n"						\
>  "	sub	%w4, %w4, %w0\n"					\
> @@ -46,11 +46,40 @@ __llsc_futex_atomic_##op(int oparg, u32 __user *uaddr, int *oval)	\
>  	return ret;							\
>  }
> 
> -LLSC_FUTEX_ATOMIC_OP(add, "add	%w3, %w1, %w5")
> -LLSC_FUTEX_ATOMIC_OP(or, "orr	%w3, %w1, %w5")
> -LLSC_FUTEX_ATOMIC_OP(and, "and	%w3, %w1, %w5")
> -LLSC_FUTEX_ATOMIC_OP(eor, "eor	%w3, %w1, %w5")
> -LLSC_FUTEX_ATOMIC_OP(set, "mov	%w3, %w5")
> +LLSC_FUTEX_ATOMIC_OP(add, add)
> +LLSC_FUTEX_ATOMIC_OP(or, orr)
> +LLSC_FUTEX_ATOMIC_OP(and, and)
> +LLSC_FUTEX_ATOMIC_OP(eor, eor)
> +
> +static __always_inline int
> +__llsc_futex_atomic_set(int oparg, u32 __user *uaddr, int *oval)
> +{
> +	unsigned int loops = LLSC_MAX_LOOPS;
> +	int ret, oldval;
> +
> +	uaccess_enable_privileged();
> +	asm volatile("//__llsc_futex_xchg\n"
> +"	prfm	pstl1strm, %2\n"
> +"1:	ldxr	%w1, %2\n"
> +"2:	stlxr	%w0, %w4, %2\n"
> +"	cbz	%w3, 3f\n"
> +"	sub	%w3, %w3, %w0\n"
> +"	cbnz	%w3, 1b\n"
> +"	mov	%w0, %w5\n"
> +"3:\n"
> +"	dmb	ish\n"
> +	_ASM_EXTABLE_UACCESS_ERR(1b, 3b, %w0)
> +	_ASM_EXTABLE_UACCESS_ERR(2b, 3b, %w0)
> +	: "=&r" (ret), "=&r" (oldval), "+Q" (*uaddr), "+r" (loops)
> +	: "r" (oparg), "Ir" (-EAGAIN)
> +	: "memory");
> +	uaccess_disable_privileged();
> +
> +	if (!ret)
> +		*oval = oldval;

Hmm, I'm really not sure this is worthwhile. I doubt the "optimisation"
actually does anything and adding a whole new block of asm just for the
SET case isn't much of an improvement on the maintainability side, either.

Will

Hi Will,

[...]

> >  arch/arm64/include/asm/futex.h | 43 ++++++++++++++++++++++++++++------
> >  1 file changed, 36 insertions(+), 7 deletions(-)
> >
> > diff --git a/arch/arm64/include/asm/futex.h b/arch/arm64/include/asm/futex.h
> > index ab7003cb4724..22a6301a9f3d 100644
> > --- a/arch/arm64/include/asm/futex.h
> > +++ b/arch/arm64/include/asm/futex.h
> > @@ -13,7 +13,7 @@
> >
> >  #define LLSC_MAX_LOOPS	128 /* What's the largest number you can think of? */
> >
> > -#define LLSC_FUTEX_ATOMIC_OP(op, insn)					\
> > +#define LLSC_FUTEX_ATOMIC_OP(op, asm_op)				\
> >  static __always_inline int						\
> >  __llsc_futex_atomic_##op(int oparg, u32 __user *uaddr, int *oval)	\
> >  {									\
> > @@ -24,7 +24,7 @@ __llsc_futex_atomic_##op(int oparg, u32 __user *uaddr, int *oval)	\
> >  	asm volatile("// __llsc_futex_atomic_" #op "\n"			\
> >  "	prfm	pstl1strm, %2\n"					\
> >  "1:	ldxr	%w1, %2\n"						\
> > -	insn "\n"							\
> > +"	" #asm_op "	%w3, %w1, %w5\n"				\
> >  "2:	stlxr	%w0, %w3, %2\n"						\
> >  "	cbz	%w0, 3f\n"						\
> >  "	sub	%w4, %w4, %w0\n"					\
> > @@ -46,11 +46,40 @@ __llsc_futex_atomic_##op(int oparg, u32 __user *uaddr, int *oval)	\
> >  	return ret;							\
> >  }
> >
> > -LLSC_FUTEX_ATOMIC_OP(add, "add	%w3, %w1, %w5")
> > -LLSC_FUTEX_ATOMIC_OP(or, "orr	%w3, %w1, %w5")
> > -LLSC_FUTEX_ATOMIC_OP(and, "and	%w3, %w1, %w5")
> > -LLSC_FUTEX_ATOMIC_OP(eor, "eor	%w3, %w1, %w5")
> > -LLSC_FUTEX_ATOMIC_OP(set, "mov	%w3, %w5")
> > +LLSC_FUTEX_ATOMIC_OP(add, add)
> > +LLSC_FUTEX_ATOMIC_OP(or, orr)
> > +LLSC_FUTEX_ATOMIC_OP(and, and)
> > +LLSC_FUTEX_ATOMIC_OP(eor, eor)
> > +
> > +static __always_inline int
> > +__llsc_futex_atomic_set(int oparg, u32 __user *uaddr, int *oval)
> > +{
> > +	unsigned int loops = LLSC_MAX_LOOPS;
> > +	int ret, oldval;
> > +
> > +	uaccess_enable_privileged();
> > +	asm volatile("//__llsc_futex_xchg\n"
> > +"	prfm	pstl1strm, %2\n"
> > +"1:	ldxr	%w1, %2\n"
> > +"2:	stlxr	%w0, %w4, %2\n"
> > +"	cbz	%w3, 3f\n"
> > +"	sub	%w3, %w3, %w0\n"
> > +"	cbnz	%w3, 1b\n"
> > +"	mov	%w0, %w5\n"
> > +"3:\n"
> > +"	dmb	ish\n"
> > +	_ASM_EXTABLE_UACCESS_ERR(1b, 3b, %w0)
> > +	_ASM_EXTABLE_UACCESS_ERR(2b, 3b, %w0)
> > +	: "=&r" (ret), "=&r" (oldval), "+Q" (*uaddr), "+r" (loops)
> > +	: "r" (oparg), "Ir" (-EAGAIN)
> > +	: "memory");
> > +	uaccess_disable_privileged();
> > +
> > +	if (!ret)
> > +		*oval = oldval;
>
> Hmm, I'm really not sure this is worthwhile. I doubt the "optimisation"
> actually does anything and adding a whole new block of asm just for the
> SET case isn't much of an improvement on the maintainability side, either.

TBH, I had the same question, but I thought this code seems to modify
freqenetly, I decide even a small optimisation -- reduce one instruction
only.

But I don't have strong opinion for this patch.
If it's not good for maintainability perspective,
This patch can be dropped.

Thanks!

--
Sincerely,
Yeoreum Yun

On Thu, Sep 11, 2025 at 05:19:11PM +0100, Yeoreum Yun wrote:
> > > +static __always_inline int
> > > +__llsc_futex_atomic_set(int oparg, u32 __user *uaddr, int *oval)
[...]
> > Hmm, I'm really not sure this is worthwhile. I doubt the "optimisation"
> > actually does anything and adding a whole new block of asm just for the
> > SET case isn't much of an improvement on the maintainability side, either.
> 
> TBH, I had the same question, but I thought this code seems to modify
> freqenetly, I decide even a small optimisation -- reduce one instruction
> only.
> 
> But I don't have strong opinion for this patch.
> If it's not good for maintainability perspective,
> This patch can be dropped.

I'd drop it for now unless you can show some performance benefits
(unlikely).

-- 
Catalin

On Fri, Sep 12, 2025 at 05:36:20PM +0100, Catalin Marinas wrote:
> On Thu, Sep 11, 2025 at 05:19:11PM +0100, Yeoreum Yun wrote:
> > > > +static __always_inline int
> > > > +__llsc_futex_atomic_set(int oparg, u32 __user *uaddr, int *oval)
> [...]
> > > Hmm, I'm really not sure this is worthwhile. I doubt the "optimisation"
> > > actually does anything and adding a whole new block of asm just for the
> > > SET case isn't much of an improvement on the maintainability side, either.
> >
> > TBH, I had the same question, but I thought this code seems to modify
> > freqenetly, I decide even a small optimisation -- reduce one instruction
> > only.
> >
> > But I don't have strong opinion for this patch.
> > If it's not good for maintainability perspective,
> > This patch can be dropped.
>
> I'd drop it for now unless you can show some performance benefits
> (unlikely).

Yes. not much of improvement. So I'll drop this patch.
Thanks.

--
Sincerely,
Yeoreum Yun