Re: [PATCH] riscv: Turn has_fpu into a static key if FPU

[Palmer Dabbelt]

On Tue, 11 May 2021 11:01:52 PDT (-0700), jszhang3@xxxxxxxxxxxxxxxx wrote:
> From: Jisheng Zhang <jszhang@xxxxxxxxxx>
>
> The has_fpu check sits at hot code path: switch_to(). Currently, has_fpu
> is a bool variable if FPU, switch_to() checks it each time, we can
> optimize out this check by turning the has_fpu into a static key.
>
> Before the patch:
> 	...
> 	ccc7c783		lbu     a5,-820(a5) # ffffffff80cab098 <has_fpu>
> 	e3b5			bnez    a5,ffffffff8031a438 <__schedule+0x232>
> 	...
>
> After the patch:
> If FPU isn't supported:
> 	...
> 	00000013		nop
> 	...
>
> If FPU is supported:
> 	...
> 	0540006f                j       ffffffff8031a418 <__schedule+0x21a>
> 	...
>
> Signed-off-by: Jisheng Zhang <jszhang@xxxxxxxxxx>
> ---
>  arch/riscv/include/asm/switch_to.h | 11 ++++++++---
>  arch/riscv/kernel/cpufeature.c     |  4 ++--
>  arch/riscv/kernel/process.c        |  2 +-
>  arch/riscv/kernel/signal.c         |  4 ++--
>  4 files changed, 13 insertions(+), 8 deletions(-)
>
> diff --git a/arch/riscv/include/asm/switch_to.h b/arch/riscv/include/asm/switch_to.h
> index 407bcc96a710..ea25ccb7785e 100644
> --- a/arch/riscv/include/asm/switch_to.h
> +++ b/arch/riscv/include/asm/switch_to.h
> @@ -6,6 +6,7 @@
>  #ifndef _ASM_RISCV_SWITCH_TO_H
>  #define _ASM_RISCV_SWITCH_TO_H
>
> +#include <linux/jump_label.h>
>  #include <linux/sched/task_stack.h>
>  #include <asm/processor.h>
>  #include <asm/ptrace.h>
> @@ -55,9 +56,13 @@ static inline void __switch_to_aux(struct task_struct *prev,
>  	fstate_restore(next, task_pt_regs(next));
>  }
>
> -extern bool has_fpu;
> +extern struct static_key_false cpu_hwcap_fpu;
> +static __always_inline bool has_fpu(void)
> +{
> +	return static_branch_unlikely(&cpu_hwcap_fpu);
> +}
>  #else
> -#define has_fpu false
> +static __always_inline bool has_fpu(void) { return false; }
>  #define fstate_save(task, regs) do { } while (0)
>  #define fstate_restore(task, regs) do { } while (0)
>  #define __switch_to_aux(__prev, __next) do { } while (0)
> @@ -70,7 +75,7 @@ extern struct task_struct *__switch_to(struct task_struct *,
>  do {							\
>  	struct task_struct *__prev = (prev);		\
>  	struct task_struct *__next = (next);		\
> -	if (has_fpu)					\
> +	if (has_fpu())					\
>  		__switch_to_aux(__prev, __next);	\
>  	((last) = __switch_to(__prev, __next));		\
>  } while (0)
> diff --git a/arch/riscv/kernel/cpufeature.c b/arch/riscv/kernel/cpufeature.c
> index ac202f44a670..a2848dc36927 100644
> --- a/arch/riscv/kernel/cpufeature.c
> +++ b/arch/riscv/kernel/cpufeature.c
> @@ -19,7 +19,7 @@ unsigned long elf_hwcap __read_mostly;
>  static DECLARE_BITMAP(riscv_isa, RISCV_ISA_EXT_MAX) __read_mostly;
>
>  #ifdef CONFIG_FPU
> -bool has_fpu __read_mostly;
> +__ro_after_init DEFINE_STATIC_KEY_FALSE(cpu_hwcap_fpu);
>  #endif
>
>  /**
> @@ -146,6 +146,6 @@ void riscv_fill_hwcap(void)
>
>  #ifdef CONFIG_FPU
>  	if (elf_hwcap & (COMPAT_HWCAP_ISA_F | COMPAT_HWCAP_ISA_D))
> -		has_fpu = true;
> +		static_branch_enable(&cpu_hwcap_fpu);
>  #endif
>  }
> diff --git a/arch/riscv/kernel/process.c b/arch/riscv/kernel/process.c
> index f9cd57c9c67d..03ac3aa611f5 100644
> --- a/arch/riscv/kernel/process.c
> +++ b/arch/riscv/kernel/process.c
> @@ -87,7 +87,7 @@ void start_thread(struct pt_regs *regs, unsigned long pc,
>  	unsigned long sp)
>  {
>  	regs->status = SR_PIE;
> -	if (has_fpu) {
> +	if (has_fpu()) {
>  		regs->status |= SR_FS_INITIAL;
>  		/*
>  		 * Restore the initial value to the FP register
> diff --git a/arch/riscv/kernel/signal.c b/arch/riscv/kernel/signal.c
> index 65942b3748b4..c2d5ecbe5526 100644
> --- a/arch/riscv/kernel/signal.c
> +++ b/arch/riscv/kernel/signal.c
> @@ -90,7 +90,7 @@ static long restore_sigcontext(struct pt_regs *regs,
>  	/* sc_regs is structured the same as the start of pt_regs */
>  	err = __copy_from_user(regs, &sc->sc_regs, sizeof(sc->sc_regs));
>  	/* Restore the floating-point state. */
> -	if (has_fpu)
> +	if (has_fpu())
>  		err |= restore_fp_state(regs, &sc->sc_fpregs);
>  	return err;
>  }
> @@ -143,7 +143,7 @@ static long setup_sigcontext(struct rt_sigframe __user *frame,
>  	/* sc_regs is structured the same as the start of pt_regs */
>  	err = __copy_to_user(&sc->sc_regs, regs, sizeof(sc->sc_regs));
>  	/* Save the floating-point state. */
> -	if (has_fpu)
> +	if (has_fpu())
>  		err |= save_fp_state(regs, &sc->sc_fpregs);
>  	return err;
>  }

Thanks, this is on for-next.