fix atomic counters; add more dbg prints

1 files changed, 9 insertions, 0 deletions

diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index fb72f25..2a88e5a 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -1157,6 +1157,15 @@ void ptx_instruction::pre_decode() {
       break;
   }
 
+  // Pre-tag atomics on the static template so co-issue eligibility checks
+  // (mem_coissue_disallowed in shader.cc) can detect them at candidate-scan
+  // time. Without this, m_isatomic only gets set during per-instance
+  // functional execution (add_callback path), which is too late — the
+  // intra co-issue scan reads the SHARED template's flag and would
+  // mistakenly admit an atomic into a composite, leaving the warp's
+  // n_atomic counter unbalanced and the warp stuck in waiting().
+  if (m_opcode == ATOM_OP) m_isatomic = true;
+
   set_opcode_and_latency();
   set_bar_type();
   // Get register operands