Since gcc does not provide lock free 16 byte access we must do so ourselves.

2 files changed, 42 insertions, 10 deletions

diff --git a/vespamalloc/src/vespamalloc/malloc/allocchunk.cpp b/vespamalloc/src/vespamalloc/malloc/allocchunk.cpp
index 41165244e5c..fb95018e94e 100644
--- a/vespamalloc/src/vespamalloc/malloc/allocchunk.cpp
+++ b/vespamalloc/src/vespamalloc/malloc/allocchunk.cpp
@@ -7,7 +7,7 @@ namespace vespamalloc {
 void AFListBase::linkInList(AtomicHeadPtr & head, AFListBase * list)
 {
     AFListBase * tail;
-    for (tail = list; tail->_next != NULL ;tail = tail->_next) { }
+    for (tail = list; tail->_next != nullptr ;tail = tail->_next) { }
     linkIn(head, list, tail);
 }
 
@@ -16,7 +16,7 @@ void AFListBase::linkIn(AtomicHeadPtr & head, AFListBase * csl, AFListBase * tai
     HeadPtr oldHead = head.load(std::memory_order_relaxed);
     HeadPtr newHead(csl, oldHead._tag + 1);
     tail->_next = static_cast<AFListBase *>(oldHead._ptr);
-    while ( ! head.compare_exchange_weak(oldHead, newHead, std::memory_order_release, std::memory_order_relaxed) ) {
+    while ( __builtin_expect(! head.compare_exchange_weak(oldHead, newHead, std::memory_order_release, std::memory_order_relaxed), false) ) {
         newHead._tag =  oldHead._tag + 1;
         tail->_next = static_cast<AFListBase *>(oldHead._ptr);
     }
@@ -26,19 +26,19 @@ AFListBase * AFListBase::linkOut(AtomicHeadPtr & head)
 {
     HeadPtr oldHead = head.load(std::memory_order_relaxed);
     AFListBase *csl = static_cast<AFListBase *>(oldHead._ptr);
-    if (csl == NULL) {
-        return NULL;
+    if (csl == nullptr) {
+        return nullptr;
     }
     HeadPtr newHead(csl->_next, oldHead._tag + 1);
-    while ( ! head.compare_exchange_weak(oldHead, newHead, std::memory_order_acquire, std::memory_order_relaxed) ) {
+    while ( __builtin_expect(! head.compare_exchange_weak(oldHead, newHead, std::memory_order_acquire, std::memory_order_relaxed), false) ) {
         csl = static_cast<AFListBase *>(oldHead._ptr);
-        if (csl == NULL) {
-            return NULL;
+        if (csl == nullptr) {
+            return nullptr;
         }
         newHead._ptr = csl->_next;
         newHead._tag = oldHead._tag + 1;
     }
-    csl->_next = NULL;
+    csl->_next = nullptr;
     return csl;
 }
 
diff --git a/vespamalloc/src/vespamalloc/malloc/allocchunk.h b/vespamalloc/src/vespamalloc/malloc/allocchunk.h
index 8b6a2ce3fc8..64f45067e79 100644
--- a/vespamalloc/src/vespamalloc/malloc/allocchunk.h
+++ b/vespamalloc/src/vespamalloc/malloc/allocchunk.h
@@ -20,16 +20,48 @@ struct TaggedPtr {
     TaggedPtr() noexcept : _ptr(nullptr), _tag(0) { }
     TaggedPtr(void *h, size_t t) noexcept : _ptr(h), _tag(t) {}
 
+#if defined(__x86_64__)
+#define VESPA_USE_ATOMIC_TAGGEDPTR
+    TaggedPtr load(std::memory_order = std::memory_order_seq_cst) {
+        return *this;
+    }
+    void store(TaggedPtr ptr) {
+        *this = ptr;
+    }
+    bool
+    compare_exchange_weak(TaggedPtr & oldPtr, TaggedPtr newPtr, std::memory_order, std::memory_order) {
+        char result;
+        __asm__ volatile (
+        "lock ;"
+        "cmpxchg16b %6;"
+        "setz %1;"
+        : "+m" (*this),
+          "=q" (result),
+          "+a" (oldPtr._ptr),
+          "+d" (oldPtr._tag)
+        : "b" (newPtr._ptr),
+          "c" (newPtr._tag)
+        : "cc", "memory"
+        );
+        return result;
+    }
+#endif
+
     void *_ptr;
     size_t _tag;
-};
+} __attribute__ ((aligned (16)));
 
 class AFListBase
 {
 public:
     using HeadPtr = TaggedPtr;
+#ifdef VESPA_USE_ATOMIC_TAGGEDPTR
+    using AtomicHeadPtr = HeadPtr;
+#else
     using AtomicHeadPtr = std::atomic<HeadPtr>;
-    AFListBase() : _next(NULL) { }
+#endif
+
+    AFListBase() : _next(nullptr) { }
     void setNext(AFListBase * csl)           { _next = csl; }
     static void init();
     static void linkInList(AtomicHeadPtr & head, AFListBase * list);