From e5c2a4e4166a660a853406bc8ebb3cce244c4fc0 Mon Sep 17 00:00:00 2001
From: Nick Gasson <nick@nickg.me.uk>
Date: Sat, 5 Nov 2022 14:30:39 +0000
Subject: [PATCH] Use relaxed atomic operations for some work queue functions

---
 src/thread.c      | 9 ++++++---
 src/thread.h      | 3 +++
 test/workqbench.c | 2 +-
 3 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/src/thread.c b/src/thread.c
index 97cf614b..351882aa 100644
--- a/src/thread.c
+++ b/src/thread.c
@@ -513,16 +513,16 @@ void __scoped_unlock(nvc_lock_t **plock)
 
 static void push_bot(threadq_t *tq, const task_t *tasks, size_t count)
 {
-   const abp_idx_t bot = atomic_load(&tq->bot);
+   const abp_idx_t bot = relaxed_load(&tq->bot);
    assert(bot + count <= THREADQ_SIZE);
 
    memcpy(tq->deque + bot, tasks, count * sizeof(task_t));
-   atomic_store(&tq->bot, bot + count);
+   store_release(&tq->bot, bot + count);
 }
 
 static bool pop_bot(threadq_t *tq, task_t *task)
 {
-   const abp_idx_t old_bot = atomic_load(&tq->bot);
+   const abp_idx_t old_bot = relaxed_load(&tq->bot);
    if (old_bot == 0)
       return false;
 
@@ -547,6 +547,7 @@ static bool pop_bot(threadq_t *tq, task_t *task)
    return false;
 }
 
+__attribute__((no_sanitize("thread")))
 static bool pop_top(threadq_t *tq, task_t *task)
 {
    const abp_age_t old_age = { .bits = atomic_load(&tq->age.bits) };
@@ -555,6 +556,8 @@ static bool pop_top(threadq_t *tq, task_t *task)
    if (bot <= old_age.top)
       return false;
 
+   // This triggers a TSAN false-positive: we will never read from *task
+   // if the CAS fails below, so it's safe to ignore
    *task = tq->deque[old_age.top];
 
    const abp_age_t new_age = {
diff --git a/src/thread.h b/src/thread.h
index 994e7e8c..c3158f04 100644
--- a/src/thread.h
+++ b/src/thread.h
@@ -39,6 +39,9 @@
 #define relaxed_load(p) __atomic_load_n((p), __ATOMIC_RELAXED)
 #define relaxed_store(p, v) __atomic_store_n((p), (v), __ATOMIC_RELAXED)
 
+#define store_release(p, v) __atomic_store_n((p), (v), __ATOMIC_RELEASE)
+#define load_acquire(p) __atomic_load_n((p), __ATOMIC_ACQUIRE)
+
 typedef struct _nvc_thread nvc_thread_t;
 
 void thread_init(void);
diff --git a/test/workqbench.c b/test/workqbench.c
index 6a6d3d0d..d4e8ab38 100644
--- a/test/workqbench.c
+++ b/test/workqbench.c
@@ -35,7 +35,7 @@ static void is_prime_cb(void *context, void *arg)
 {
    int n = (intptr_t)arg;
 
-   for (int i = 2; i < n / 2; i++) {
+   for (int i = 2; i <= n / 2; i++) {
       if (n % i == 0) {
          results[n] = 1;
          break;
-- 
2.39.2