impl option w/ only 2 SBs; stall so SBs are cleared before div or reconv

1 files changed, 21 insertions, 1 deletions

diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index 6331d6b..09d59f9 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -1166,6 +1166,11 @@ struct simd_set_info {
   // source_inst that owns this set. Used at writeback to release the
   // matching (reg, mask) entry. Zero/empty when source_inst is NULL.
   active_mask_t source_mask;
+  // Mode 2 (slot-pinned scoreboard): which ibuffer half (0 or 1) the
+  // source_inst issued from in its own warp. Used at writeback to release
+  // the matching slot's scoreboard. (unsigned)-1 when source_inst is NULL
+  // (i.e., the set carries primary's own work).
+  unsigned source_slot_id;
 
   // Mixed-space MEM co-issue (v2): per-set shared state.
   // Populated only when source_inst's space is shared/sstarr; otherwise
@@ -1177,7 +1182,8 @@ struct simd_set_info {
 
   simd_set_info()
       : set_id(0), warp_id(0), split_id((unsigned)-1),
-        num_active_threads(0), valid(false), source_inst(NULL), cycles(0) {
+        num_active_threads(0), valid(false), source_inst(NULL),
+        source_slot_id((unsigned)-1), cycles(0) {
     set_active_mask.reset();
     active_mask_in_warp.reset();
     source_mask.reset();
@@ -1337,6 +1343,16 @@ class warp_inst_t : public inst_t {
   // Mode 1: stamp the source_inst's full active_mask onto each valid set
   // for use at writeback (releaseRegisterMask).
   void set_source_mask_on_sets(const active_mask_t &mask);
+  // Mode 2: stamp the source_inst's owning ibuffer half (0 or 1) onto
+  // each valid set for use at writeback (releaseRegisterSlot).
+  void set_source_slot_on_sets(unsigned slot_id);
+
+  // Mode 2: which ibuffer half (0 or 1) this inst was issued from. Set
+  // by the issue path; read by the scoreboard reservation/release
+  // dispatcher to pick the correct per-slot scoreboard. (unsigned)-1
+  // means "not from a slot" (modes 0/1 don't use this).
+  unsigned get_ibuffer_half_id() const { return m_ibuffer_half_id; }
+  void set_ibuffer_half_id(unsigned h) { m_ibuffer_half_id = h; }
   static bool simd_sets_overlap(const std::vector<simd_set_info> &a,
                                 const std::vector<simd_set_info> &b);
 
@@ -1575,6 +1591,10 @@ class warp_inst_t : public inst_t {
   int m_dbg_split_id = -1;
   const void *m_dbg_source_inst = NULL;
 
+  // Mode 2 (slot-pinned scoreboard): which ibuffer half (0 or 1) this
+  // inst was issued from. (unsigned)-1 = unset / N/A in modes 0,1.
+  unsigned m_ibuffer_half_id = (unsigned)-1;
+
   // Jin: cdp support
  public:
   int m_is_cdp;