Skip to contentSkip to conversation

support: user more portable atomic wrappers

PATCH In review Vineet Gupta 2 participants 4 messages Sep 9, 2020 – Sep 10, 2020 Synopsys ARC architecture

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 support: user more portable atomic wrappers

This came up in a nascent arc64 port, lacking gcc atomics for now

---
 support/support_record_failure.c | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)
support/support_record_failure.c +8 -7
index f766c0623683..65e576c6e901 100644
@@ -25,6 +25,7 @@
 #include <stdlib.h>
 #include <sys/mman.h>
 #include <unistd.h>
+#include <atomic.h>
 
 /* This structure keeps track of test failures.  The counter is
    incremented on each failure.  The failed member is set to true if a
@@ -66,8 +67,8 @@ support_record_failure (void)
     }
   /* Relaxed MO is sufficient because we are only interested in the
      values themselves, in isolation.  */
-  __atomic_store_n (&state->failed, 1, __ATOMIC_RELEASE);
-  __atomic_add_fetch (&state->counter, 1, __ATOMIC_RELEASE);
+  atomic_store_release (&state->failed, 1);
+  atomic_fetch_add_release (&state->counter, 1);
 }
 
 int
@@ -84,10 +85,10 @@ support_report_failure (int status)
      assumes that exiting from the main thread happens before the
      error reporting via support_record_failure, which requires some
      form of external synchronization.  */
-  bool failed = __atomic_load_n (&state->failed, __ATOMIC_RELAXED);
+  bool failed = atomic_load_relaxed (&state->failed);
   if (failed)
     printf ("error: %u test failures\n",
-            __atomic_load_n (&state->counter, __ATOMIC_RELAXED));
+            atomic_load_relaxed (&state->counter));
 
   if ((status == 0 || status == EXIT_UNSUPPORTED) && failed)
     /* If we have a recorded failure, it overrides a non-failure
@@ -101,8 +102,8 @@ support_record_failure_reset (void)
 {
   /* Only used for testing the test framework, with external
      synchronization, but use release MO for consistency.  */
-  __atomic_store_n (&state->failed, 0, __ATOMIC_RELAXED);
-  __atomic_add_fetch (&state->counter, 0, __ATOMIC_RELAXED);
+  atomic_store_relaxed (&state->failed, 0);
+  atomic_fetch_add_release (&state->counter, 0);
 }
 
 int
@@ -110,5 +111,5 @@ support_record_failure_is_failed (void)
 {
   /* Relaxed MO is sufficient because we need (blocking) external
      synchronization for reliable test error reporting anyway.  */
-  return __atomic_load_n (&state->failed, __ATOMIC_RELAXED);
+  return atomic_load_relaxed (&state->failed);
 }
Signature · 7 lines
2.25.1


_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
#1 lore raw

* Vineet Gupta via Libc-alpha:

Quoted 1 line

This came up in a nascent arc64 port, lacking gcc atomics for now

Can you please change the GCC port to provide atomics instead?

It does not make sense to maintain those atomics in many projects separately.

Thanks,
Florian

Signature · 9 lines
Red Hat GmbH, https://de.redhat.com/ , Registered seat: Grasbrunn,
Commercial register: Amtsgericht Muenchen, HRB 153243,
Managing Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neill


_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
#2 lore raw
Quoted 5 lines · Florian Weimer via Libc-alpha

* Vineet Gupta via Libc-alpha:

Quoted 1 line

This came up in a nascent arc64 port, lacking gcc atomics for now

Can you please change the GCC port to provide atomics instead?

Sure, they are added now.

Quoted 2 lines

It does not make sense to maintain those atomics in many projects separately.

I agree that gcc atomics should be baseline. I would still propose to carry this patch as it makes code less verbose if nothing else and the wrappers are part of glibc already.
_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc

#3 lore raw

* Vineet Gupta:

Quoted 3 lines

I agree that gcc atomics should be baseline. I would still propose to carry this patch as it makes code less verbose if nothing else and the wrappers are part of glibc already.

The advantage of the GCC built-ins is that they have easy-to-find documentation. Our wrapper macros are in the same file as the legacy atomics that no longer should be used. They are also not easy to use correctly (e.g., there is no support for access to single bytes, but the code will still compile on some architectures).

The GCC built-ins have other problems, of course. For example, one might accidentally introduce a dependency on libatomic (leading to a link failure later). But I think these issues are less severe.

Thanks,
Florian

Signature · 9 lines
Red Hat GmbH, https://de.redhat.com/ , Registered seat: Grasbrunn,
Commercial register: Amtsgericht Muenchen, HRB 153243,
Managing Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neill


_______________________________________________
linux-snps-arc mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-snps-arc