RE: [PATCH 12/13] AArch64: Cleanup fenv implementation

["Wilco Dijkstra" <wdijkstr at arm dot com>]

ping

> -----Original Message-----
> From: Wilco Dijkstra [mailto:wdijkstr@arm.com]
> Sent: 23 October 2014 18:36
> To: 'libc-alpha@sourceware.org'
> Subject: [PATCH 12/13] AArch64: Cleanup fenv implementation
> 
> Improve fesetenv performance by avoiding unnecessary FPSR/FPCR reads/writes.
> It uses the same logic as the ARM version. The common case removes 1 FPSR
> and 1 FPCR read. For FE_DFL_ENV and FE_NOMASK_ENV a FPCR read is avoided in
> case the FPCR does not change.
> 
> ChangeLog:
> 2014-10-23  Wilco Dijkstra  <wdijkstr@arm.com>
> 
> 	* sysdeps/aarch64/fpu/fesetenv.c (fesetenv):
> 	Optimize to reduce FPCR/FPSR accesses.
> 
> ---
>  sysdeps/aarch64/fpu/fesetenv.c | 40 +++++++++++++++++++++++-----------------
>  1 file changed, 23 insertions(+), 17 deletions(-)
> 
> diff --git a/sysdeps/aarch64/fpu/fesetenv.c b/sysdeps/aarch64/fpu/fesetenv.c
> index c19680d..f71014d 100644
> --- a/sysdeps/aarch64/fpu/fesetenv.c
> +++ b/sysdeps/aarch64/fpu/fesetenv.c
> @@ -29,8 +29,20 @@ fesetenv (const fenv_t *envp)
>    fpu_fpsr_t fpsr_new;
> 
>    _FPU_GETCW (fpcr);
> -  _FPU_GETFPSR (fpsr);
> 
> +  if ((envp != FE_DFL_ENV) && (envp != FE_NOMASK_ENV))
> +    {
> +      /* The new FPCR/FPSR are valid, so don't merge the reserved flags.  */
> +      fpcr_new = envp->__fpcr;
> +
> +      if (fpcr != fpcr_new)
> +	_FPU_SETCW (fpcr_new);
> +
> +      _FPU_SETFPSR (envp->__fpsr);
> +      return 0;
> +    }
> +
> +  _FPU_GETFPSR (fpsr);
>    fpcr_new = fpcr & _FPU_RESERVED;
>    fpsr_new = fpsr & _FPU_FPSR_RESERVED;
> 
> @@ -39,31 +51,25 @@ fesetenv (const fenv_t *envp)
>        fpcr_new |= _FPU_DEFAULT;
>        fpsr_new |= _FPU_FPSR_DEFAULT;
>      }
> -  else if (envp == FE_NOMASK_ENV)
> +  else
>      {
>        fpcr_new |= _FPU_FPCR_IEEE;
>        fpsr_new |= _FPU_FPSR_IEEE;
>      }
> -  else
> -    {
> -      fpcr_new |= envp->__fpcr & ~_FPU_RESERVED;
> -      fpsr_new |= envp->__fpsr & ~_FPU_FPSR_RESERVED;
> -    }
> 
> -  if (fpsr != fpsr_new)
> -    _FPU_SETFPSR (fpsr_new);
> +  _FPU_SETFPSR (fpsr_new);
> 
>    if (fpcr != fpcr_new)
> -    _FPU_SETCW (fpcr_new);
> +    {
> +      _FPU_SETCW (fpcr_new);
> 
> -  /* Trapping exceptions are optional in AArch64 the relevant enable
> -     bits in FPCR are RES0 hence the absence of support can be
> -     detected by reading back the FPCR and comparing with the required
> -     value.  */
> +      /* Trapping exceptions are optional in AArch64; the relevant enable
> +	 bits in FPCR are RES0 hence the absence of support can be detected
> +	 by reading back the FPCR and comparing with the required value.  */
> +      _FPU_GETCW (updated_fpcr);
> 
> -  _FPU_GETCW (updated_fpcr);
> -  if ((updated_fpcr & fpcr_new) != fpcr_new)
> -    return 1;
> +      return fpcr_new & ~updated_fpcr;
> +    }
> 
>    return 0;
>  }
> --
> 1.9.1