Backport More Qthreads Performance Fixes (#26328)

Backporting two more performance fixes from sandialabs/qthreads#314. [Contributed by @insertinterestingnamehere. Reviewed and merged by @jabraham17]
chapel-lang · Dec 12, 2024 · c93e20f · c93e20f
2 parents 3a79d12 + 90785f2
commit c93e20f
Show file tree

Hide file tree

Showing 3 changed files with 30 additions and 3 deletions.
diff --git a/third-party/qthread/README b/third-party/qthread/README
@@ -52,8 +52,9 @@ index 8269aa0949..bf022bc243 100644
 
 ```
 
-* Backport a performance fix from the development version of qthreads.
+* Backport a few performance fixes from the development version of qthreads.
   https://github.com/chapel-lang/chapel/pull/26138
+  https://github.com/chapel-lang/chapel/pull/26328
 
 ```
 diff --git a/include/qt_atomics.h b/include/qt_atomics.h
@@ -87,3 +88,30 @@ index 2ac887ed..2312c954 100644
  #define QTHREAD_FASTLOCK_SETUP()                                               \
    do {                                                                         \
 ```
+
+```
+--- a/src/io.c
++++ b/src/io.c
+@@ -74,7 +74,6 @@ static void qt_blocking_subsystem_internal_freemem(void) { /*{{{*/
+ static void *qt_blocking_subsystem_proxy_thread(void *Q_UNUSED(arg)) { /*{{{*/
+   while (!atomic_load_explicit(&proxy_exit, memory_order_relaxed)) {
+     if (qt_process_blocking_call()) { break; }
+-    MACHINE_FENCE;
+   }
+   atomic_fetch_sub_explicit(&io_worker_count, 1, memory_order_relaxed);
+   pthread_exit(NULL);
+```
+
+```
+--- a/include/qthread/qthread.h
++++ b/include/qthread/qthread.h
+@@ -87,7 +87,7 @@ using std::memory_order_relaxed;
+
+ #include "macros.h"
+
+-#define MACHINE_FENCE atomic_thread_fence(memory_order_acq_rel);
++#define MACHINE_FENCE atomic_thread_fence(memory_order_seq_cst);
+
+ #if QTHREAD_ASSEMBLY_ARCH == QTHREAD_AMD64
+ #define QTHREAD_SWAPS_IMPLY_ACQ_REL_FENCES
+```
diff --git a/third-party/qthread/qthread-src/include/qthread/qthread.h b/third-party/qthread/qthread-src/include/qthread/qthread.h
@@ -87,7 +87,7 @@ using std::memory_order_relaxed;
 
 #include "macros.h"
 
-#define MACHINE_FENCE atomic_thread_fence(memory_order_acq_rel);
+#define MACHINE_FENCE atomic_thread_fence(memory_order_seq_cst);
 
 #if QTHREAD_ASSEMBLY_ARCH == QTHREAD_AMD64
 #define QTHREAD_SWAPS_IMPLY_ACQ_REL_FENCES

diff --git a/third-party/qthread/qthread-src/src/io.c b/third-party/qthread/qthread-src/src/io.c
@@ -74,7 +74,6 @@ static void qt_blocking_subsystem_internal_freemem(void) { /*{{{*/
 static void *qt_blocking_subsystem_proxy_thread(void *Q_UNUSED(arg)) { /*{{{*/
   while (!atomic_load_explicit(&proxy_exit, memory_order_relaxed)) {
     if (qt_process_blocking_call()) { break; }
-    MACHINE_FENCE;
   }
   atomic_fetch_sub_explicit(&io_worker_count, 1, memory_order_relaxed);
   pthread_exit(NULL);