fix double sb from fifo rotations during intra coissue

1 files changed, 47 insertions, 2 deletions

diff --git a/src/gpgpu-sim/scoreboard.h b/src/gpgpu-sim/scoreboard.h
index da34af5..8f736d7 100644
--- a/src/gpgpu-sim/scoreboard.h
+++ b/src/gpgpu-sim/scoreboard.h
@@ -136,8 +136,15 @@ class Scoreboard {
   std::vector<std::set<unsigned> > sec_reg_table;
 
   // Mode 2: per-warp, per-slot register sets.
-  // slot_reg_table[wid][slot] = set of reserved register numbers.
-  std::vector<std::array<std::set<unsigned>, NUM_SLOTS> > slot_reg_table;
+  // slot_reg_table[wid][slot][reg] = refcount of pending writes to reg
+  // from this slot. Idempotent insert (set semantics) is *wrong* — two
+  // consecutive issues from the same slot writing the same reg would
+  // collapse to a single entry, and the first retire would erase it
+  // before the second retires, losing the second's pending state.
+  // Ref-counted map fixes that: each reserve ++, each release --, entry
+  // erased when count hits 0.
+  std::vector<std::array<std::map<unsigned, unsigned>, NUM_SLOTS> >
+      slot_reg_table;
 
   // Mode 1: per-warp list of mask-aware reservations.
   struct mask_resv {
@@ -180,6 +187,44 @@ class Scoreboard {
       m_primary_resv_cycle;
   std::map<std::pair<unsigned, unsigned>, unsigned long long>
       m_sec_resv_cycle;
+
+ public:
+  // Mode 2 (slot-pinned) cross-slot hazard diagnostics. Increments when an
+  // issue from one slot would have collided with the *opposite* slot's
+  // pending writes — i.e. a hazard the slot-pinned design currently misses.
+  // Updated from shader.cc at issue-eligibility check sites.
+  // The unmasked variant counts any reg-name overlap (including
+  // lane-disjoint cases that aren't real RAWs). The "_real" variant uses
+  // a mask-aware shadow table to count only true RAWs.
+  unsigned long long m_xslot_primary_issue_would_stall      = 0;
+  unsigned long long m_xslot_primary_issue_real_would_stall = 0;
+  unsigned long long m_xslot_primary_issue_checks           = 0;
+  unsigned long long m_xslot_inter_coissue_would_stall      = 0;
+  unsigned long long m_xslot_inter_coissue_real_would_stall = 0;
+  unsigned long long m_xslot_inter_coissue_checks           = 0;
+  unsigned long long m_xslot_intra_coissue_would_stall      = 0;
+  unsigned long long m_xslot_intra_coissue_checks           = 0;
+
+  // Mask-aware shadow table maintained alongside the slot tables (mode 2
+  // only) for diagnostic comparison vs sb=1 ground truth. Each reservation
+  // pushes (reg, inst's active mask, refcount); each release pops the
+  // matching (reg, mask) entry. checkCollisionShadow performs the
+  // sb=1-style mask intersection across ALL entries for the warp.
+  void reserveShadowMask(unsigned wid, unsigned reg,
+                         const active_mask_t &mask);
+  void releaseShadowMask(unsigned wid, unsigned reg,
+                         const active_mask_t &mask);
+  bool checkCollisionShadow(unsigned wid, const class inst_t *inst,
+                            const active_mask_t &mask) const;
+  // Dump shadow entries that intersect (mask) for any of inst's regs.
+  // Used by diag tracing to attribute real_would_stall events.
+  void dumpShadowOverlap(unsigned wid, const class inst_t *inst,
+                         const active_mask_t &mask, FILE *out,
+                         const char *site, unsigned long long cyc) const;
+
+ private:
+  std::vector<std::vector<mask_resv> > m_xslot_shadow_mask_table;
+ public:
 };
 
 #endif /* SCOREBOARD_H_ */