fix no compaction sb deadlock

2 files changed, 76 insertions, 17 deletions

diff --git a/src/abstract_hardware_model.cc b/src/abstract_hardware_model.cc
index 1010d00..b164c73 100644
--- a/src/abstract_hardware_model.cc
+++ b/src/abstract_hardware_model.cc
@@ -181,6 +181,16 @@ void warp_inst_t::compute_simd_sets(unsigned num_sets, unsigned set_width) {
     info.set_active_mask.reset();
     info.active_mask_in_warp.reset();
     info.num_active_threads = 0;
+    // Pipeline-register warp_inst_t objects are reused, so the per-set
+    // coissuer-tracking fields can be stale from a prior issue. The LSU
+    // gates on (!valid || source_inst == NULL); without these resets a
+    // primary's set inherits a previous coissue's source_inst != NULL
+    // and split_id, causing m_pending_writes_secondary to be incremented
+    // on a key no future decrement targets → deadlock under sb1.
+    info.split_id = (unsigned)-1;
+    info.source_inst = NULL;
+    info.source_mask.reset();
+    info.source_slot_id = (unsigned)-1;
     for (unsigned lane = 0; lane < set_width; lane++) {
       unsigned tid = s * set_width + lane;
       info.thread_ids[lane] = tid;
@@ -208,6 +218,11 @@ void warp_inst_t::compute_simd_sets_compacted(unsigned num_sets,
     info.num_active_threads = 0;
     info.valid = false;
     info.source_inst = NULL;
+    // Clear remaining coissuer-tracking state stale from prior reuse.
+    // (source_inst=NULL alone defends LSU; reset the rest for symmetry.)
+    info.split_id = (unsigned)-1;
+    info.source_mask.reset();
+    info.source_slot_id = (unsigned)-1;
     for (unsigned i = 0; i < MAX_WARP_SIZE; i++) info.thread_ids[i] = 0;
   }
 
@@ -251,6 +266,12 @@ void warp_inst_t::compute_simd_sets_xor_static(unsigned num_sets,
     info.set_active_mask.reset();
     info.active_mask_in_warp.reset();
     info.num_active_threads = 0;
+    // See compute_simd_sets — clear coissuer-tracking state stale from
+    // prior issues on this reused pipeline-register slot.
+    info.split_id = (unsigned)-1;
+    info.source_inst = NULL;
+    info.source_mask.reset();
+    info.source_slot_id = (unsigned)-1;
   }
 
   // xor_key = reverse of the low log2(warp_size) bits of m_warp_id
diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc
index e9c2e19..a86393c 100644
--- a/src/gpgpu-sim/shader.cc
+++ b/src/gpgpu-sim/shader.cc
@@ -2477,8 +2477,18 @@ void scheduler_unit::try_utilization_max_coissue(
           continue;
         }
         unsigned sec_active = sec_mask_eff.count();
-        unsigned sec_needed = (sec_active + set_width - 1) / set_width;
-        if (sec_needed > available_sets) {
+        // Use the same feasibility test as inter coissue. Under
+        // compaction != 2 partitions are fixed by thread_id, so the
+        // coissuer's set positions can overlap the composite's even
+        // when (active/set_width) <= available_sets. merge_simd_sets
+        // silently drops overlapping coissuer sets while the access
+        // was already injected, leaving m_pending_writes_secondary
+        // unincremented but later decremented at writeback (underflow
+        // on sb0/sb2, deadlock on sb1).
+        unsigned sec_needed = 0;
+        if (!check_coissue_feasibility(co_issue_composite, sec_inst,
+                                        sec_mask_eff, cand_warp_id,
+                                        available_sets, &sec_needed)) {
           m_stats->coissue_denied_by_no_sets[get_sid()]++;
           continue;
         }
@@ -2517,16 +2527,14 @@ void scheduler_unit::try_utilization_max_coissue(
        it != pool.end() && available_sets > 0; it++) {
     coissue_cand_t &c = *it;
     // Re-check feasibility: prior picks may have filled sets that this
-    // candidate overlaps (only matters for compaction_mode != 2).
+    // candidate overlaps (only matters for compaction_mode != 2). Both
+    // intra and inter must re-evaluate against the *current* composite
+    // — the cached c.sets_needed from admission can be stale once an
+    // earlier pool entry has merged its sets in.
     unsigned needed = 0;
-    if (c.is_intra) {
-      if (c.sets_needed > available_sets) continue;
-      needed = c.sets_needed;
-    } else {
-      if (!check_coissue_feasibility(co_issue_composite, c.inst, c.mask,
-                                      c.warp_id, available_sets, &needed))
-        continue;
-    }
+    if (!check_coissue_feasibility(co_issue_composite, c.inst, c.mask,
+                                    c.warp_id, available_sets, &needed))
+      continue;
 
     if (c.is_intra) {
       if (m_shader->m_config->gpgpu_simd_partitioning_debug) {
@@ -2798,8 +2806,18 @@ void scheduler_unit::try_utilization_max_coissue_window(
           continue;
         }
         unsigned sec_active = sec_mask_eff.count();
-        unsigned sec_needed = (sec_active + set_width - 1) / set_width;
-        if (sec_needed > available_sets) {
+        // Use the same feasibility test as inter coissue. Under
+        // compaction != 2 partitions are fixed by thread_id, so the
+        // coissuer's set positions can overlap the composite's even
+        // when (active/set_width) <= available_sets. merge_simd_sets
+        // silently drops overlapping coissuer sets while the access
+        // was already injected, leaving m_pending_writes_secondary
+        // unincremented but later decremented at writeback (underflow
+        // on sb0/sb2, deadlock on sb1).
+        unsigned sec_needed = 0;
+        if (!check_coissue_feasibility(co_issue_composite, sec_inst,
+                                        sec_mask_eff, cand_warp_id,
+                                        available_sets, &sec_needed)) {
           m_stats->coissue_denied_by_no_sets[get_sid()]++;
           continue;
         }
@@ -3120,8 +3138,18 @@ void scheduler_unit::try_utilization_max_coissue_window_centered(
           continue;
         }
         unsigned sec_active = sec_mask_eff.count();
-        unsigned sec_needed = (sec_active + set_width - 1) / set_width;
-        if (sec_needed > available_sets) {
+        // Use the same feasibility test as inter coissue. Under
+        // compaction != 2 partitions are fixed by thread_id, so the
+        // coissuer's set positions can overlap the composite's even
+        // when (active/set_width) <= available_sets. merge_simd_sets
+        // silently drops overlapping coissuer sets while the access
+        // was already injected, leaving m_pending_writes_secondary
+        // unincremented but later decremented at writeback (underflow
+        // on sb0/sb2, deadlock on sb1).
+        unsigned sec_needed = 0;
+        if (!check_coissue_feasibility(co_issue_composite, sec_inst,
+                                        sec_mask_eff, cand_warp_id,
+                                        available_sets, &sec_needed)) {
           m_stats->coissue_denied_by_no_sets[get_sid()]++;
           continue;
         }
@@ -3435,8 +3463,18 @@ void scheduler_unit::try_utilization_max_coissue_window_before(
           continue;
         }
         unsigned sec_active = sec_mask_eff.count();
-        unsigned sec_needed = (sec_active + set_width - 1) / set_width;
-        if (sec_needed > available_sets) {
+        // Use the same feasibility test as inter coissue. Under
+        // compaction != 2 partitions are fixed by thread_id, so the
+        // coissuer's set positions can overlap the composite's even
+        // when (active/set_width) <= available_sets. merge_simd_sets
+        // silently drops overlapping coissuer sets while the access
+        // was already injected, leaving m_pending_writes_secondary
+        // unincremented but later decremented at writeback (underflow
+        // on sb0/sb2, deadlock on sb1).
+        unsigned sec_needed = 0;
+        if (!check_coissue_feasibility(co_issue_composite, sec_inst,
+                                        sec_mask_eff, cand_warp_id,
+                                        available_sets, &sec_needed)) {
           m_stats->coissue_denied_by_no_sets[get_sid()]++;
           continue;
         }