Re: [PATCH 2/2] eal/riscv: add support for zicbop extension

["Stanisław Kardach" <stanislaw.kardach@gmail.com>]

On Thu, May 30, 2024 at 7:20 PM Daniel Gregory
<daniel.gregory@bytedance.com> wrote:
>
> The zicbop extension adds instructions for prefetching data into cache.
> Use them to implement RISCV-specific versions of the rte_prefetch* and
> rte_prefetch*_write functions.
>
> - prefetch.r indicates to hardware that the cache block will be accessed
>   by a data read soon
> - prefetch.w indicates to hardware that the cache block will be accessed
>   by a data write soon
>
> These instructions are emitted by __builtin_prefetch on modern versions
> of Clang (17.0.1+) and GCC (13.1.0+). For earlier versions, we may not
> have support for assembling Zicbop instructions, so emit the word
> that encodes a 'prefetch.[rw] 0(a0)' instruction.
Is there a benefit of adding this flag instead of relying on compiler
implementation of __builtin_prefetch()? As in do you have some
requirements for older compiler support? If just using
__builtin_prefetch could simplify this code a lot and push the
detection of zicbop to the compiler implementation. The runtime
detection is another issue and would require runtime patching to keep
the code fast.
>
> This new functionality is controlled by a Meson flag that is disabled by
> default. Whilst it's a hint, like rte_pause(), and so has no effect if
> the target doesn't support the extension, it requires the address
> prefetched to be loaded into a0, which may be costly.
>
> Signed-off-by: Daniel Gregory <daniel.gregory@bytedance.com>
> Suggested-by: Punit Agrawal <punit.agrawal@bytedance.com>
> ---
>  config/riscv/meson.build             |  6 +++
>  lib/eal/riscv/include/rte_prefetch.h | 57 ++++++++++++++++++++++++++--
>  2 files changed, 59 insertions(+), 4 deletions(-)
>
> diff --git a/config/riscv/meson.build b/config/riscv/meson.build
> index 07d7d9da23..ecf9da1c39 100644
> --- a/config/riscv/meson.build
> +++ b/config/riscv/meson.build
> @@ -26,6 +26,12 @@ flags_common = [
>      # read from /proc/device-tree/cpus/timebase-frequency. This property is
>      # guaranteed on Linux, as riscv time_init() requires it.
>      ['RTE_RISCV_TIME_FREQ', 0],
> +
> +    # When true override the default implementation of the prefetching functions
> +    # (rte_prefetch*) with a version that explicitly uses the Zicbop extension.
> +    # Do not enable when using modern versions of GCC (13.1.0+) or Clang
> +    # (17.0.1+). They will emit these instructions in the default implementation
> +    ['RTE_RISCV_ZICBOP', false],
>  ]
>
>  ## SoC-specific options.
> diff --git a/lib/eal/riscv/include/rte_prefetch.h b/lib/eal/riscv/include/rte_prefetch.h
> index 748cf1b626..82cad526b3 100644
> --- a/lib/eal/riscv/include/rte_prefetch.h
> +++ b/lib/eal/riscv/include/rte_prefetch.h
> @@ -14,21 +14,42 @@ extern "C" {
>
>  #include <rte_compat.h>
>  #include <rte_common.h>
> +
> +#ifdef RTE_RISCV_ZICBOP
> +#define RTE_PREFETCH_WRITE_ARCH_DEFINED
> +#endif
> +
>  #include "generic/rte_prefetch.h"
>
> +/*
> + * Modern versions of GCC & Clang will emit prefetch instructions for
> + * __builtin_prefetch when the Zicbop extension is present.
> + * The RTE_RISCV_ZICBOP option controls whether we emit them manually for older
> + * compilers that may not have the support to assemble them.
> + */
>  static inline void rte_prefetch0(const volatile void *p)
>  {
> -       RTE_SET_USED(p);
> +#ifndef RTE_RISCV_ZICBOP
> +       /* by default __builtin_prefetch prepares for a read */
> +       __builtin_prefetch((const void *)p);
> +#else
> +       /* prefetch.r 0(a0) */
> +       register const volatile void *a0 asm("a0") = p;
> +       asm volatile (".int 0x00156013" : : "r" (a0));
> +#endif
>  }
>
> +/*
> + * The RISC-V Zicbop extension doesn't have instructions to prefetch to only a
> + * subset of cache levels, so fallback to rte_prefetch0
> + */
>  static inline void rte_prefetch1(const volatile void *p)
>  {
> -       RTE_SET_USED(p);
> +       rte_prefetch0(p);
>  }
> -
>  static inline void rte_prefetch2(const volatile void *p)
>  {
> -       RTE_SET_USED(p);
> +       rte_prefetch0(p);
>  }
>
>  static inline void rte_prefetch_non_temporal(const volatile void *p)
> @@ -44,6 +65,34 @@ rte_cldemote(const volatile void *p)
>         RTE_SET_USED(p);
>  }
>
> +#ifdef RTE_RISCV_ZICBOP
> +__rte_experimental
> +static inline void
> +rte_prefetch0_write(const void *p)
> +{
> +       /* prefetch.w 0(a0) */
> +       register const void *a0 asm("a0") = p;
> +       asm volatile (".int 0x00356013" : : "r" (a0));
> +}
> +
> +/*
> + * The RISC-V Zicbop extension doesn't have instructions to prefetch to only a
> + * subset of cache levels, so fallback to rte_prefetch0_write
> + */
> +__rte_experimental
> +static inline void
> +rte_prefetch1_write(const void *p)
> +{
> +       rte_prefetch0_write(p);
> +}
> +__rte_experimental
> +static inline void
> +rte_prefetch2_write(const void *p)
> +{
> +       rte_prefetch0_write(p);
> +}
> +#endif /* RTE_RISCV_ZICBOP */
> +
>  #ifdef __cplusplus
>  }
>  #endif
> --
> 2.39.2
>