Merge pull request #551 from ValeevGroup/evaleev/feature/lazy-deleter-skip-sync-in-do-cleanup

DistArray::lazy_deleter: skip lazy_sync when invoked from fence's do_cleanup

Author: evaleev

external/versions.cmake

@@ -12,8 +12,8 @@ set(TA_INSTALL_EIGEN_URL_HASH SHA256=b4c198460eba6f28d34894e3a5710998818515104d6
 set(TA_INSTALL_EIGEN_PREVIOUS_URL_HASH MD5=b9e98a200d2455f06db9c661c5610496)
 
 set(TA_TRACKED_MADNESS_URL https://github.com/m-a-d-n-e-s-s/madness.git CACHE STRING "GIT_REPOSITORY for cloning MADNESS source")
-set(TA_TRACKED_MADNESS_TAG f7aa1401e CACHE STRING "GIT_TAG (branch or hash) for cloning MADNESS")
-set(TA_TRACKED_MADNESS_PREVIOUS_TAG 7d8aaf9d51981e4accf4d84742270d1473f8ca2e)
+set(TA_TRACKED_MADNESS_TAG 666765ca6 CACHE STRING "GIT_TAG (branch or hash) for cloning MADNESS")
+set(TA_TRACKED_MADNESS_PREVIOUS_TAG f7aa1401e)
 set(TA_TRACKED_MADNESS_VERSION 0.10.1)
 set(TA_TRACKED_MADNESS_PREVIOUS_VERSION 0.10.1)
 

src/TiledArray/array_impl.h

@@ -480,6 +480,24 @@ class ArrayImpl : public TensorImpl<Policy>,
         // wait for all DelayedSet's to vanish
         world.await([&]() { return (pimpl->num_live_ds() == 0); }, true);
 
+        // Fast path when invoked from inside the fence's deferred-cleanup
+        // phase: the global-termination protocol has already established
+        // global quiescence (no in-flight AM, all ranks at the same point),
+        // and symmetric collective use of `defer_deleter_to_next_fence()`
+        // guarantees every rank has this same pimpl in its deferred list
+        // and so reaches this same delete in lockstep. The cross-rank
+        // lazy_sync handshake below is therefore redundant; it would also
+        // schedule a lazy_sync_children task on this world's taskq that the
+        // fence cannot drain (do_cleanup runs after the drain loop) and
+        // that would later be run by some unrelated fence -- against freed
+        // state if this world is destroyed before then (e.g. einsum's
+        // per-Hadamard sub-Worlds).
+        if (world.gop.is_in_do_cleanup()) {
+          delete pimpl;
+          cleanup_counter_--;
+          return;
+        }
+
         try {
           world.gop.lazy_sync(id, [pimpl]() {
             delete pimpl;