From: Dinar Temirbulatov <dtemirbulatov@gmail.com>
To: "Joseph S. Myers" <joseph@codesourcery.com>
Cc: libc-ports@sourceware.org, adeb@nvidia.com,
Maxim Kuvyrkov <maxim@kugelworks.com>
Subject: Re: [Patch] ARM define atomic_exchange_acq/atomic_exchange_rel to __atomic_exchange_n
Date: Thu, 05 Sep 2013 12:56:00 -0000 [thread overview]
Message-ID: <CAMnfPmOyoRNNu62uAvTPDfa42XCwBdrW0n5jAW0Ab+nPS8u86Q@mail.gmail.com> (raw)
In-Reply-To: <Pine.LNX.4.64.1309021457010.17654@digraph.polyomino.org.uk>
[-- Attachment #1: Type: text/plain, Size: 2450 bytes --]
Hi,
Here is updated version of change. Ok to commit?
thanks, Dinar.
On Mon, Sep 2, 2013 at 7:08 PM, Joseph S. Myers <joseph@codesourcery.com> wrote:
> On Mon, 2 Sep 2013, Dinar Temirbulatov wrote:
>
>> Hi,
>> Another version of change, I added
>> atomic_compare_and_exchange_val_acq/atomic_compare_and_exchange_val_rel
>> and atomic_compare_and_exchange_bool_acq/atomic_compare_and_exchange_boot_rel
>> defenitions and also for gcc-4.7 and higher in the case of unsupported
>> atomic compare and swap operation, it uses the kernel helper inlines.
>> Tested on arm a9 with no new regressions. Ok to commit?
>> Oh, sorry. I missed to attach the change. Here it is.
>
> For subsequent patch revisions, please note there should be an extra space
> after "#" for preprocessor directives inside #if conditionals, one per
> level of #if nesting (other than toplevel multiple-inclusion guards) -
> which means that if conditioning existing code, you need to adjust
> directives inside that code.
>
> This patch appears to have too much duplication. For example, you
> duplicate the definition of __arm_assisted_compare_and_exchange_val_32_acq
> - but that should not need any extra conditionals at all (beyond the
> existing #ifndef __arm_assisted_compare_and_exchange_val_32_acq), there's
> no reason ever not to define it. Similarly, you duplicate
> __arch_compare_and_exchange_val_64_acq, but with proper #if structure
> there should only need to be one copy of the version that uses
> __arm_link_error.
>
> What I think you should aim for is that each definition, or small group of
> definitions, uses conditionals in the form
>
> #if __GNUC_PREREQ (4, 7) && defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_4
> /* Version using __atomic_*. */
> #elif defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_4
> /* Version using __sync_*. */
> #else
> /* Version using __arm_assisted_*. */
> #endif
>
> with cases omitted if not useful for that particular macro (this may
> include some macros not being defined at all in some cases). So don't
> insert any global conditionals affecting all the existing definitions at
> all - look at each block of conditionals and add a third case as needed,
> along with any new macros (again with conditionals in that form) that are
> appropriate.
>
> Where you use abort () in some definitions, use __arm_link_error ()
> instead, like for the existing definitions.
>
> --
> Joseph S. Myers
> joseph@codesourcery.com
[-- Attachment #2: arm_atomic7.patch --]
[-- Type: application/octet-stream, Size: 5558 bytes --]
--- glibc-orig/ports/sysdeps/arm/bits/atomic.h 2013-08-19 21:46:44.592731210 +0400
+++ glibc/ports/sysdeps/arm/bits/atomic.h 2013-09-05 15:53:52.752801514 +0400
@@ -35,9 +35,6 @@ typedef uintmax_t uatomic_max_t;
void __arm_link_error (void);
-/* Use the atomic builtins provided by GCC in case the backend provides
- a pattern to do this efficiently. */
-
#ifdef __GCC_HAVE_SYNC_COMPARE_AND_SWAP_4
# define atomic_full_barrier() __sync_synchronize ()
#else
@@ -51,26 +48,107 @@ void __arm_link_error (void);
# define __arm_assisted_full_barrier() __arm_link_error()
#endif
-/* Atomic compare and exchange. */
+/* Use the atomic builtins provided by GCC in case the backend provides
+ a pattern to do this efficiently. */
+#if __GNUC_PREREQ (4, 7) && defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_4
-#ifdef __GCC_HAVE_SYNC_COMPARE_AND_SWAP_4
-# define __arch_compare_and_exchange_val_32_acq(mem, newval, oldval) \
+# define atomic_exchange_acq(mem, value) \
+ __atomic_val_bysize (__arch_exchange, int, mem, value, __ATOMIC_ACQUIRE)
+
+# define atomic_exchange_rel(mem, value) \
+ __atomic_val_bysize (__arch_exchange, int, mem, value, __ATOMIC_RELEASE)
+
+/* Atomic exchange (without compare). */
+
+# define __arch_exchange_8_int(mem, newval, model) \
+ (__arm_link_error (), (typeof(*mem)) 0)
+
+# define __arch_exchange_16_int(mem, newval, model) \
+ (__arm_link_error (), (typeof(*mem)) 0)
+
+# define __arch_exchange_32_int(mem, newval, model) \
+ __atomic_exchange_n (mem, newval, model)
+
+# define __arch_exchange_64_int(mem, newval, model) \
+ (__arm_link_error (), (typeof(*mem)) 0)
+
+/* Compare and exchange with "acquire" semantics, ie barrier after. */
+
+# define atomic_compare_and_exchange_bool_acq(mem, new, old) \
+ __atomic_bool_bysize (__arch_compare_and_exchange_bool, int, \
+ mem, new, old, __ATOMIC_ACQUIRE)
+
+# define atomic_compare_and_exchange_val_acq(mem, new, old) \
+ __atomic_val_bysize (__arch_compare_and_exchange_val, int, \
+ mem, new, old, __ATOMIC_ACQUIRE)
+
+/* Compare and exchange with "release" semantics, ie barrier before. */
+
+# define atomic_compare_and_exchange_bool_rel(mem, new, old) \
+ __atomic_bool_bysize (__arch_compare_and_exchange_bool, int, \
+ mem, new, old, __ATOMIC_RELEASE)
+
+# define atomic_compare_and_exchange_val_rel(mem, new, old) \
+ __atomic_val_bysize (__arch_compare_and_exchange_val, int, \
+ mem, new, old, __ATOMIC_RELEASE)
+
+/* Compare and exchange.
+ For all "bool" routines, we return FALSE if exchange succesful. */
+
+# define __arch_compare_and_exchange_bool_8_int(mem, newval, oldval, model) \
+ ({__arm_link_error (); oldval; })
+
+# define __arch_compare_and_exchange_bool_16_int(mem, newval, oldval, model) \
+ ({__arm_link_error (); oldval; })
+
+# define __arch_compare_and_exchange_bool_32_int(mem, newval, oldval, model) \
+ ({ \
+ typeof (*mem) __oldval = (oldval); \
+ !__atomic_compare_exchange_n (mem, (void *) &__oldval, newval, 0, \
+ model, __ATOMIC_RELAXED); \
+ })
+
+# define __arch_compare_and_exchange_bool_64_int(mem, newval, oldval, model) \
+ ({__arm_link_error (); oldval; })
+
+# define __arch_compare_and_exchange_val_8_int(mem, newval, oldval, model) \
+ ({__arm_link_error (); oldval; })
+
+# define __arch_compare_and_exchange_val_16_int(mem, newval, oldval, model) \
+ ({__arm_link_error (); oldval; })
+
+# define __arch_compare_and_exchange_val_32_int(mem, newval, oldval, model) \
+ ({ \
+ typeof (*mem) __oldval = (oldval); \
+ __atomic_compare_exchange_n (mem, (void *) &__oldval, newval, 0, \
+ model, __ATOMIC_RELAXED); \
+ __oldval; \
+ })
+
+# define __arch_compare_and_exchange_val_64_int(mem, newval, oldval, model) \
+ ({__arm_link_error (); oldval; })
+
+#elif defined __GCC_HAVE_SYNC_COMPARE_AND_SWAP_4
+/* Atomic compare and exchange. */
+# define __arch_compare_and_exchange_val_32_acq(mem, newval, oldval) \$
__sync_val_compare_and_swap ((mem), (oldval), (newval))
#else
# define __arch_compare_and_exchange_val_32_acq(mem, newval, oldval) \
__arm_assisted_compare_and_exchange_val_32_acq ((mem), (newval), (oldval))
#endif
+#if !__GNUC_PREREQ (4, 7) || !defined (__GCC_HAVE_SYNC_COMPARE_AND_SWAP_4)
/* We don't support atomic operations on any non-word types.
So make them link errors. */
-#define __arch_compare_and_exchange_val_8_acq(mem, newval, oldval) \
+# define __arch_compare_and_exchange_val_8_acq(mem, newval, oldval) \
({ __arm_link_error (); oldval; })
-#define __arch_compare_and_exchange_val_16_acq(mem, newval, oldval) \
+# define __arch_compare_and_exchange_val_16_acq(mem, newval, oldval) \
({ __arm_link_error (); oldval; })
-#define __arch_compare_and_exchange_val_64_acq(mem, newval, oldval) \
+# define __arch_compare_and_exchange_val_64_acq(mem, newval, oldval) \
({ __arm_link_error (); oldval; })
+#endif
/* An OS-specific bits/atomic.h file will define this macro if
the OS can provide something. If not, we'll fail to build
next prev parent reply other threads:[~2013-09-05 12:56 UTC|newest]
Thread overview: 21+ messages / expand[flat|nested] mbox.gz Atom feed top
2013-09-01 20:43 Dinar Temirbulatov
2013-09-02 15:08 ` Joseph S. Myers
2013-09-03 17:31 ` Abhishek Deb
2013-09-05 12:56 ` Dinar Temirbulatov [this message]
2013-09-05 15:48 ` Joseph S. Myers
2013-09-05 17:09 ` Abhishek Deb
2013-09-11 19:25 ` Dinar Temirbulatov
2013-09-11 19:27 ` Dinar Temirbulatov
2013-09-11 20:40 ` Abhishek Deb
2013-09-13 17:14 ` Joseph S. Myers
2013-09-14 3:29 ` Dinar Temirbulatov
2013-09-18 17:16 ` Joseph S. Myers
2013-09-19 6:53 ` Maxim Kuvyrkov
2013-09-19 16:29 ` Joseph S. Myers
-- strict thread matches above, loose matches on Subject: below --
2013-09-01 20:42 Dinar Temirbulatov
2013-08-13 17:11 Dinar Temirbulatov
2013-08-15 21:46 ` Abhishek Deb
2013-08-16 22:57 ` Dinar Temirbulatov
2013-08-17 2:42 ` Abhishek Deb
2013-08-19 19:21 ` Dinar Temirbulatov
2013-08-18 20:29 ` Joseph S. Myers
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=CAMnfPmOyoRNNu62uAvTPDfa42XCwBdrW0n5jAW0Ab+nPS8u86Q@mail.gmail.com \
--to=dtemirbulatov@gmail.com \
--cc=adeb@nvidia.com \
--cc=joseph@codesourcery.com \
--cc=libc-ports@sourceware.org \
--cc=maxim@kugelworks.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).