From: David Abdurachmanov <david.abdurachmanov@gmail.com>
To: Christoph Muellner <christoph.muellner@vrull.eu>
Cc: libc-alpha@sourceware.org, Palmer Dabbelt <palmer@dabbelt.com>,
Darius Rad <darius@bluespec.com>,
Andrew Waterman <andrew@sifive.com>, DJ Delorie <dj@redhat.com>,
Vineet Gupta <vineetg@rivosinc.com>,
Kito Cheng <kito.cheng@sifive.com>,
Jeff Law <jeffreyalaw@gmail.com>,
Philipp Tomsich <philipp.tomsich@vrull.eu>,
Heiko Stuebner <heiko.stuebner@vrull.eu>
Subject: Re: [RFC PATCH 06/19] riscv: Adding ISA string parser for environment variables
Date: Tue, 7 Feb 2023 07:20:10 +0100 [thread overview]
Message-ID: <CAEn-LTqXy45fQmiknJx0JFZQ0CT85htjNrcqXDCmxp5W7v4n+Q@mail.gmail.com> (raw)
In-Reply-To: <20230207001618.458947-7-christoph.muellner@vrull.eu>
On Tue, Feb 7, 2023 at 1:20 AM Christoph Muellner
<christoph.muellner@vrull.eu> wrote:
>
> From: Christoph Müllner <christoph.muellner@vrull.eu>
>
> RISC-V does not have a reliable mechanism to detect hart features
> like supported ISA extensions or cache block sizes at run-time
> as of now.
>
> Not knowing the hart features limits optimization strategies of glibc
> (e.g. ifunc support requires run-time hard feature knowledge).
>
> To circumvent this limitation this patch introduces a mechanism to
> get the hart features via environment variables:
> * RISCV_RT_MARCH represents a lower-case ISA string (-march string)
> E.g. RISCV_RT_MARCH=rv64gc_zicboz
> * RISCV_RT_CBOM_BLOCKSIZE represents the cbom instruction block size
> E.g. RISCV_RT_CBOZ_BLOCKSIZE=64
> * RISCV_RT_CBOZ_BLOCKSIZE represents the cboz instruction block size
Hi,
Non-expert here, but have you considered defining RISCV glibc
tunables? It's designed for this purpose.
See: https://www.gnu.org/software/libc/manual/html_node/Tunables.html
Especially hardware capability tunables:
https://www.gnu.org/software/libc/manual/html_node/Hardware-Capability-Tunables.html
It would be nice to have tunables defined for RISC-V that could stay
with us for a long time.
Cheers,
david
> These environment variables are parsed during startup and the found
> ISA extensions are stored a struct (hart_features) for evaluation
> by dynamic dispatching code.
>
> As the parser code is executed very early, we cannot call functions
> that have direct or indirect (via getenv()) dependencies to strlen()
> and strncmp(), as these functions cannot be called before the ifunc
> support is initialized. Therefore, this patch contains its own helper
> functions for strlen(), strncmp(), and getenv().
>
> Signed-off-by: Christoph Müllner <christoph.muellner@vrull.eu>
> ---
> sysdeps/unix/sysv/linux/riscv/hart-features.c | 294 ++++++++++++++++++
> .../unix/sysv/linux/riscv/macro-for-each.h | 24 ++
> 2 files changed, 318 insertions(+)
> create mode 100644 sysdeps/unix/sysv/linux/riscv/macro-for-each.h
>
> diff --git a/sysdeps/unix/sysv/linux/riscv/hart-features.c b/sysdeps/unix/sysv/linux/riscv/hart-features.c
> index 41111eff57..6de41a26cc 100644
> --- a/sysdeps/unix/sysv/linux/riscv/hart-features.c
> +++ b/sysdeps/unix/sysv/linux/riscv/hart-features.c
> @@ -17,12 +17,17 @@
> <https://www.gnu.org/licenses/>. */
>
> #include <hart-features.h>
> +#include <macro-for-each.h>
> +#include <string_private.h>
>
> /* The code in this file is executed very early, so we cannot call
> indirect functions because ifunc support is not initialized.
> Therefore this file adds a few simple helper functions to avoid
> dependencies to functions outside of this file. */
>
> +#define xstr(s) str(s)
> +#define str(s) #s
> +
> static inline void
> inhibit_loop_to_libcall
> simple_memset (void *s, int c, size_t n)
> @@ -35,9 +40,298 @@ simple_memset (void *s, int c, size_t n)
> }
> }
>
> +static inline size_t
> +inhibit_loop_to_libcall
> +simple_strlen (const char *s)
> +{
> + size_t n = 0;
> + char c = *s;
> + while (c != 0)
> + {
> + s++;
> + n++;
> + c = *s;
> + }
> + return n;
> +}
> +
> +static inline int
> +inhibit_loop_to_libcall
> +simple_strncmp (const char *s1, const char *s2, size_t n)
> +{
> + while (n != 0)
> + {
> + if (*s1 == 0 || *s1 != *s2)
> + return *((const unsigned char *)s1) - *((const unsigned char *)s2);
> + n--;
> + s1++;
> + s2++;
> + }
> + return 0;
> +}
> +
> +extern char **__environ;
> +static inline char*
> +simple_getenv (const char *name)
> +{
> + char **ep;
> + uint16_t name_start;
> +
> + if (__environ == NULL || name[0] == 0 || name[1] == 0)
> + return NULL;
> +
> + size_t len = simple_strlen (name);
> +#if _STRING_ARCH_unaligned
> + name_start = *(const uint16_t *) name;
> +#else
> + name_start = (((const unsigned char *) name)[0]
> + | (((const unsigned char *) name)[1] << 8));
> +#endif
> + len -= 2;
> + name += 2;
> +
> + for (ep = __environ; *ep != NULL; ++ep)
> + {
> +#if _STRING_ARCH_unaligned
> + uint16_t ep_start = *(uint16_t *) *ep;
> +#else
> + uint16_t ep_start = (((unsigned char *) *ep)[0]
> + | (((unsigned char *) *ep)[1] << 8));
> +#endif
> + if (name_start == ep_start && !simple_strncmp (*ep + 2, name, len)
> + && (*ep)[len + 2] == '=')
> + return &(*ep)[len + 3];
> + }
> + return NULL;
> +}
> +
> +/* Check if the given number is a power of 2.
> + Return true if so, or false otherwise. */
> +static inline int
> +is_power_of_two (unsigned long v)
> +{
> + return (v & (v - 1)) == 0;
> +}
> +
> +/* Check if the given string str starts with
> + the prefix pre. Return true if so, or false
> + otherwise. */
> +static inline int
> +starts_with (const char *str, const char *pre)
> +{
> + return simple_strncmp (pre, str, simple_strlen (pre)) == 0;
> +}
> +
> +/* Lower all characters of a string up to the
> + first NUL-character in the string. */
> +static inline void
> +strtolower (char *s)
> +{
> + char c = *s;
> + while (c != '\0')
> + {
> + if (c >= 'A' && c <= 'Z')
> + *s = c + 'a' - 'A';
> + s++;
> + c = *s;
> + }
> +}
> +
> +/* Count the number of detected extensions. */
> +static inline unsigned long
> +count_extensions (struct hart_features *hart_features)
> +{
> + unsigned long n = 0;
> +#define ISA_EXT(e) \
> + if (hart_features->have_##e == 1) \
> + n++;
> +#define ISA_EXT_GROUP(g, ...) \
> + if (hart_features->have_##g == 1) \
> + n++;
> +#include "isa-extensions.def"
> + return n;
> +}
> +
> +/* Check if the given charater is not '0'-'9'. */
> +static inline int
> +notanumber (const char c)
> +{
> + return (c < '0' || c > '9');
> +}
> +
> +/* Parse RISCV_RT_MARCH and store found extensions. */
> +static inline void
> +parse_rt_march (struct hart_features *hart_features)
> +{
> + const char* s = simple_getenv ("RISCV_RT_MARCH");
> + if (s == NULL)
> + goto end;
> +
> + hart_features->rt_march = s;
> +
> + /* "RISC-V ISA strings begin with either RV32I, RV32E, RV64I, or RV128I
> + indicating the supported address space size in bits for the base
> + integer ISA." */
> + if (starts_with (s, "rv32") && notanumber (*(s+4)))
> + {
> + hart_features->xlen = 32;
> + s += 4;
> + }
> + else if (starts_with (s, "rv64") && notanumber (*(s+4)))
> + {
> + hart_features->xlen = 64;
> + s += 4;
> + }
> + else if (starts_with (s, "rv128") && notanumber (*(s+5)))
> + {
> + hart_features->xlen = 128;
> + s += 5;
> + }
> + else
> + {
> + goto fail;
> + }
> +
> + /* Parse the extensions. */
> + const char *s_old = s;
> + while (*s != '\0')
> + {
> +#define ISA_EXT(e) \
> + else if (starts_with (s, xstr (e))) \
> + { \
> + hart_features->have_##e = 1; \
> + s += simple_strlen (xstr (e)); \
> + }
> +#define ISA_EXT_GROUP(g, ...) \
> + ISA_EXT (g)
> + if (0);
> +#include "isa-extensions.def"
> +
> + /* Consume optional version information. */
> + while (*s >= '0' && *s <= '9')
> + s++;
> + while (*s == 'p')
> + s++;
> + while (*s >= '0' && *s <= '9')
> + s++;
> +
> + /* Consume optional '_'. */
> + if (*s == '_')
> + s++;
> +
> + /* If we got stuck, bail out. */
> + if (s == s_old)
> + goto fail;
> + }
> +
> + /* Propagate subsets (until we reach a fixpoint). */
> + unsigned long n = count_extensions (hart_features);
> + while (1)
> + {
> + /* Forward-propagation. E.g.:
> + if (hart_features->have_g == 1)
> + {
> + hart_features->have_i = 1;
> + ...
> + hart_features->have_zifencei = 1;
> + } */
> +#define ISA_EXT_GROUP_HEAD(y) \
> + if (hart_features->have_##y) \
> + {
> +#define ISA_EXT_GROUP_SUBSET(s) \
> + hart_features->have_##s = 1;
> +#define ISA_EXT_GROUP_TAIL(z) \
> + }
> +#define ISA_EXT_GROUP(x, ...) \
> + ISA_EXT_GROUP_HEAD (x) \
> + FOR_EACH (ISA_EXT_GROUP_SUBSET, __VA_ARGS__) \
> + ISA_EXT_GROUP_TAIL (x)
> +#include "isa-extensions.def"
> +#undef ISA_EXT_GROUP_HEAD
> +#undef ISA_EXT_GROUP_SUBSET
> +#undef ISA_EXT_GROUP_TAIL
> +
> + /* Backward-propagation. E.g.:
> + if (1
> + && hart_features->have_i == 1
> + ...
> + && hart_features->have_zifencei == 1
> + )
> + hart_features->have_g = 1; */
> +#define ISA_EXT_GROUP_HEAD(y) \
> + if (1
> +#define ISA_EXT_GROUP_SUBSET(s) \
> + && hart_features->have_##s == 1
> +#define ISA_EXT_GROUP_TAIL(z) \
> + ) \
> + hart_features->have_##z = 1;
> +#define ISA_EXT_GROUP(x, ...) \
> + ISA_EXT_GROUP_HEAD (x) \
> + FOR_EACH (ISA_EXT_GROUP_SUBSET, __VA_ARGS__) \
> + ISA_EXT_GROUP_TAIL (x)
> +#include "isa-extensions.def"
> +#undef ISA_EXT_GROUP_HEAD
> +#undef ISA_EXT_GROUP_SUBSET
> +#undef ISA_EXT_GROUP_TAIL
> +
> + unsigned long n2 = count_extensions (hart_features);
> + /* Stop if fix-point reached. */
> + if (n == n2)
> + break;
> + n = n2;
> + }
> +
> +end:
> + return;
> +
> +fail:
> + hart_features->rt_march = NULL;
> +}
> +
> +/* Parse RISCV_RT_CBOM_BLOCKSIZE and store value. */
> +static inline void
> +parse_rt_cbom_blocksize (struct hart_features *hart_features)
> +{
> + hart_features->rt_cbom_blocksize = NULL;
> + hart_features->cbom_blocksize = 0;
> +
> + const char *s = simple_getenv ("RISCV_RT_CBOM_BLOCKSIZE");
> + if (s == NULL)
> + return;
> +
> + uint64_t v = _dl_strtoul (s, NULL);
> + if (!is_power_of_two (v))
> + return;
> +
> + hart_features->rt_cbom_blocksize = s;
> + hart_features->cbom_blocksize = v;
> +}
> +
> +/* Parse RISCV_RT_CBOZ_BLOCKSIZE and store value. */
> +static inline void
> +parse_rt_cboz_blocksize (struct hart_features *hart_features)
> +{
> + hart_features->rt_cboz_blocksize = NULL;
> + hart_features->cboz_blocksize = 0;
> +
> + const char *s = simple_getenv ("RISCV_RT_CBOZ_BLOCKSIZE");
> + if (s == NULL)
> + return;
> +
> + uint64_t v = _dl_strtoul (s, NULL);
> + if (!is_power_of_two (v))
> + return;
> +
> + hart_features->rt_cboz_blocksize = s;
> + hart_features->cboz_blocksize = v;
> +}
> +
> /* Discover hart features and store them. */
> static inline void
> init_hart_features (struct hart_features *hart_features)
> {
> simple_memset (hart_features, 0, sizeof (*hart_features));
> + parse_rt_march (hart_features);
> + parse_rt_cbom_blocksize (hart_features);
> + parse_rt_cboz_blocksize (hart_features);
> }
> diff --git a/sysdeps/unix/sysv/linux/riscv/macro-for-each.h b/sysdeps/unix/sysv/linux/riscv/macro-for-each.h
> new file mode 100644
> index 0000000000..524bef3c0a
> --- /dev/null
> +++ b/sysdeps/unix/sysv/linux/riscv/macro-for-each.h
> @@ -0,0 +1,24 @@
> +/* Recursive macros implementation by David Mazières
> + https://www.scs.stanford.edu/~dm/blog/va-opt.html */
> +
> +#ifndef _MACRO_FOR_EACH_H
> +#define _MACRO_FOR_EACH_H
> +
> +#define EXPAND1(...) __VA_ARGS__
> +#define EXPAND2(...) EXPAND1 (EXPAND1 (EXPAND1 (EXPAND1 (__VA_ARGS__))))
> +#define EXPAND3(...) EXPAND2 (EXPAND2 (EXPAND2 (EXPAND2 (__VA_ARGS__))))
> +#define EXPAND4(...) EXPAND3 (EXPAND3 (EXPAND3 (EXPAND3 (__VA_ARGS__))))
> +#define EXPAND(...) EXPAND4 (EXPAND4 (EXPAND4 (EXPAND4 (__VA_ARGS__))))
> +
> +#define FOR_EACH(macro, ...) \
> + __VA_OPT__ (EXPAND (FOR_EACH_HELPER (macro, __VA_ARGS__)))
> +
> +#define PARENS ()
> +
> +#define FOR_EACH_HELPER(macro, a1, ...) \
> + macro (a1) \
> + __VA_OPT__ (FOR_EACH_AGAIN PARENS (macro, __VA_ARGS__))
> +
> +#define FOR_EACH_AGAIN() FOR_EACH_HELPER
> +
> +#endif /* _MACRO_FOR_EACH_H */
> --
> 2.39.1
>
next prev parent reply other threads:[~2023-02-07 6:20 UTC|newest]
Thread overview: 42+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-02-07 0:15 [RFC PATCH 00/19] riscv: ifunc support with optimized mem*/str*/cpu_relax routines Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 01/19] Inhibit early libcalls before ifunc support is ready Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 02/19] riscv: LEAF: Use C_LABEL() to construct the asm name for a C symbol Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 03/19] riscv: Add ENTRY_ALIGN() macro Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 04/19] riscv: Add hart feature run-time detection framework Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 05/19] riscv: Introduction of ISA extensions Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 06/19] riscv: Adding ISA string parser for environment variables Christoph Muellner
2023-02-07 6:20 ` David Abdurachmanov [this message]
2023-02-07 0:16 ` [RFC PATCH 07/19] riscv: hart-features: Add fast_unaligned property Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 08/19] riscv: Add (empty) ifunc framework Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 09/19] riscv: Add ifunc support for memset Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 10/19] riscv: Add accelerated memset routines for RV64 Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 11/19] riscv: Add ifunc support for memcpy/memmove Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 12/19] riscv: Add accelerated memcpy/memmove routines for RV64 Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 13/19] riscv: Add ifunc support for strlen Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 14/19] riscv: Add accelerated strlen routine Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 15/19] riscv: Add ifunc support for strcmp Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 16/19] riscv: Add accelerated strcmp routines Christoph Muellner
2023-02-07 11:57 ` Xi Ruoyao
2023-02-07 14:15 ` Christoph Müllner
2023-03-31 5:06 ` Jeff Law
2023-03-31 12:31 ` Adhemerval Zanella Netto
2023-03-31 14:30 ` Jeff Law
2023-03-31 14:48 ` Adhemerval Zanella Netto
2023-03-31 17:19 ` Palmer Dabbelt
2023-03-31 14:32 ` Jeff Law
2023-02-07 0:16 ` [RFC PATCH 17/19] riscv: Add ifunc support for strncmp Christoph Muellner
2023-02-07 0:16 ` [RFC PATCH 18/19] riscv: Add an optimized strncmp routine Christoph Muellner
2023-02-07 1:19 ` Noah Goldstein
2023-02-08 15:13 ` Philipp Tomsich
2023-02-08 17:55 ` Palmer Dabbelt
2023-02-08 19:48 ` Adhemerval Zanella Netto
2023-02-08 18:04 ` Noah Goldstein
2023-02-07 0:16 ` [RFC PATCH 19/19] riscv: Add __riscv_cpu_relax() to allow yielding in busy loops Christoph Muellner
2023-02-07 0:23 ` Andrew Waterman
2023-02-07 0:29 ` Christoph Müllner
2023-02-07 2:59 ` [RFC PATCH 00/19] riscv: ifunc support with optimized mem*/str*/cpu_relax routines Kito Cheng
2023-02-07 16:40 ` Adhemerval Zanella Netto
2023-02-07 17:16 ` DJ Delorie
2023-02-07 19:32 ` Philipp Tomsich
2023-02-07 21:14 ` DJ Delorie
2023-02-08 11:26 ` Christoph Müllner
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAEn-LTqXy45fQmiknJx0JFZQ0CT85htjNrcqXDCmxp5W7v4n+Q@mail.gmail.com \
--to=david.abdurachmanov@gmail.com \
--cc=andrew@sifive.com \
--cc=christoph.muellner@vrull.eu \
--cc=darius@bluespec.com \
--cc=dj@redhat.com \
--cc=heiko.stuebner@vrull.eu \
--cc=jeffreyalaw@gmail.com \
--cc=kito.cheng@sifive.com \
--cc=libc-alpha@sourceware.org \
--cc=palmer@dabbelt.com \
--cc=philipp.tomsich@vrull.eu \
--cc=vineetg@rivosinc.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).