[commit: ghc] wip/ppc-reg-alloc: RTS: Add missing memory barrier (5c084e0)

[git at git.haskell.org]

Repository : ssh://git@git.haskell.org/ghc

On branch  : wip/ppc-reg-alloc
Link       : http://ghc.haskell.org/trac/ghc/changeset/5c084e0468be46f5ab48b2c1669a7e4d4d0f3c43/ghc

>---------------------------------------------------------------

commit 5c084e0468be46f5ab48b2c1669a7e4d4d0f3c43
Author: Peter Trommler <ptrommler at acm.org>
Date:   Sun Feb 24 17:11:00 2019 +0100

    RTS: Add missing memory barrier
    
    In the work stealing queue a load-load-barrier is required to ensure
    that a read of queue data cannot be reordered before a read of the
    bottom pointer into the queue.
    
    The added load-load-barrier ensures that the ordering of writes enforced
    at the end of `pushWSDeque` is also respected in the order of reads in
    `stealWSDeque_`. In other words, when reading `q->bottom` we want to make
    sure that we see the updates to `q->elements`.
    
    Fixes #13633


>---------------------------------------------------------------

5c084e0468be46f5ab48b2c1669a7e4d4d0f3c43
 rts/WSDeque.c                     | 9 ++++++---
 testsuite/tests/rts/testwsdeque.c | 9 ++++++---
 2 files changed, 12 insertions(+), 6 deletions(-)

diff --git a/rts/WSDeque.c b/rts/WSDeque.c
index b9393b1..60b8948 100644
--- a/rts/WSDeque.c
+++ b/rts/WSDeque.c
@@ -194,14 +194,17 @@ stealWSDeque_ (WSDeque *q)
     // concurrent popWSQueue() operation.
     if ((long)b - (long)t <= 0 ) {
         return NULL; /* already looks empty, abort */
-  }
-
+    }
+    // NB. the load of q->bottom must be ordered before the load of
+    // q->elements[t & q-> moduloSize]. See comment "KG:..." below
+    // and Ticket #13633.
+    load_load_barrier();
     /* now access array, see pushBottom() */
     stolen = q->elements[t & q->moduloSize];
 
     /* now decide whether we have won */
     if ( !(CASTOP(&(q->top),t,t+1)) ) {
-        /* lost the race, someon else has changed top in the meantime */
+        /* lost the race, someone else has changed top in the meantime */
         return NULL;
     }  /* else: OK, top has been incremented by the cas call */
 
diff --git a/testsuite/tests/rts/testwsdeque.c b/testsuite/tests/rts/testwsdeque.c
index 3f17f32..0a2a64d 100644
--- a/testsuite/tests/rts/testwsdeque.c
+++ b/testsuite/tests/rts/testwsdeque.c
@@ -50,14 +50,17 @@ myStealWSDeque_ (WSDeque *q, uint32_t n)
     // concurrent popWSQueue() operation.
     if ((long)b - (long)t <= 0 ) { 
         return NULL; /* already looks empty, abort */
-  }
-    
+    }
+    // NB. the load of q->bottom must be ordered before the load of
+    // q->elements[t & q-> moduloSize]. See comment "KG:..." below
+    // and Ticket #13633.
+    load_load_barrier();
     /* now access array, see pushBottom() */
     stolen = q->elements[t & q->moduloSize];
     
     /* now decide whether we have won */
     if ( !(CASTOP(&(q->top),t,t+1)) ) {
-        /* lost the race, someon else has changed top in the meantime */
+        /* lost the race, someone else has changed top in the meantime */
         return NULL;
     }  /* else: OK, top has been incremented by the cas call */