This commit adds the improvements that were made as part of the apr-m…

…ay2021 ACO paper.

* Replaced some cost fn checks w/ calls to needsSLIL()

* Changed DCF(dual cost fn) ACO to use better cost comparisons

* simplified logic and fixed 2 pass bug

* added compilation checks around expensive ACO debugging statements

* Added stalling list scheduler

* Previously was impossible to select stalling_list

* Fixed up ACO 2 cost fn alogrithm. Added code for assessing ACO schedule quality

* Changed update rule to not update in the second pass if rp constraints are violated, added profiling print stmts

* compile time optimizations

* switched variables to use pheromone_t

* added back ability to use no heuristic

* Clang formatted and removed irrelevant comments & #if directive

* changes in response to Justin's PR comments

* Dealing with a problem that I noticed which would affect cost functions other than APRP

example/llvm7-CPU2006-cfg/sched.ini

@@ -147,6 +147,7 @@ LATENCY_PRECISION LLVM
 # The scheduler used to find an initial feasible schedule.
 # LIST: List scheduler
 # SEQ: Sequential list scheduler
+# STALLING_LIST: Schedules stalls until instruction with top heuristic score becomes ready
 HEUR_SCHED_TYPE LIST
 
 # What circumstances the ACO dual cost algorithm should be applied
@@ -161,7 +162,7 @@ ACO_DUAL_COST_FN_ENABLE OFF
 # The option NONE disables the constraint cost function for the selected pass
 # NOTE: If the constraint cost function is SLIL then the cost function must also be SLIL
 # NOTE: The value for ACO2P is used in the second pass
-ACO_DUAL_COST_FN TARGET
+ACO_DUAL_COST_FN SLIL
 ACO2P_DUAL_COST_FN NONE
 
 #use 3-tournament
@@ -199,8 +200,10 @@ ACO_DBG_REGIONS NONE
 
 ACO_DBG_REGIONS_OUT_PATH /home/user/path_to_graph_output_directory/
 
-# The importance of the heuristic in ACO. ACO uses (1/heuristic)^importance, so
-# importance of 0 means don't use the heuristic.
+# Previously the heuristic was raised to the power of the heuristic importance,
+# but this has proved to not be useful, and added an expensive pow operation.
+# Now a heuristic importance of 0 disables the heuristic. Any other value leaves the
+# heuristic enabled.
 ACO_HEURISTIC_IMPORTANCE 1
 ACO2P_HEURISTIC_IMPORTANCE 1
 

example/optsched-cfg/sched.ini

@@ -147,6 +147,7 @@ LATENCY_PRECISION LLVM
 # The scheduler used to find an initial feasible schedule.
 # LIST: List scheduler
 # SEQ: Sequential list scheduler
+# STALLING_LIST: Schedules stalls until instruction with top heuristic score becomes ready
 HEUR_SCHED_TYPE LIST
 
 # What circumstances the ACO dual cost algorithm should be applied
@@ -161,7 +162,7 @@ ACO_DUAL_COST_FN_ENABLE OFF
 # The option NONE disables the constraint cost function for the selected pass
 # NOTE: If the constraint cost function is SLIL then the cost function must also be SLIL
 # NOTE: The value for ACO2P is used in the second pass
-ACO_DUAL_COST_FN TARGET
+ACO_DUAL_COST_FN SLIL
 ACO2P_DUAL_COST_FN NONE
 
 #use 3-tournament
@@ -199,8 +200,10 @@ ACO_DBG_REGIONS NONE
 
 ACO_DBG_REGIONS_OUT_PATH /home/user/path_to_graph_output_directory/
 
-# The importance of the heuristic in ACO. ACO uses (1/heuristic)^importance, so
-# importance of 0 means don't use the heuristic.
+# Previously the heuristic was raised to the power of the heuristic importance,
+# but this has proved to not be useful, and added an expensive pow operation.
+# Now a heuristic importance of 0 disables the heuristic. Any other value leaves the
+# heuristic enabled.
 ACO_HEURISTIC_IMPORTANCE 1
 ACO2P_HEURISTIC_IMPORTANCE 1
 

include/opt-sched/Scheduler/aco.h

@@ -17,7 +17,6 @@ Last Update:  Jan. 2020
 #include <map>
 #include <memory>
 #include <utility>
-
 namespace llvm {
 namespace opt_sched {
 
@@ -32,7 +31,7 @@ enum class DCF_OPT {
 
 struct Choice {
   SchedInstruction *inst;
-  double heuristic;  // range 1 to 2
+  pheromone_t heuristic; // range 1 to 2
   InstCount readyOn; // number of cycles until this instruction becomes ready
 };
 
@@ -51,7 +50,7 @@ class ACOScheduler : public ConstrainedScheduler {
 private:
   pheromone_t &Pheromone(SchedInstruction *from, SchedInstruction *to);
   pheromone_t &Pheromone(InstCount from, InstCount to);
-  double Score(SchedInstruction *from, Choice choice);
+  pheromone_t Score(SchedInstruction *from, Choice choice);
   bool shouldReplaceSchedule(InstSchedule *OldSched, InstSchedule *NewSched,
                              bool IsGlobal);
   DCF_OPT ParseDCFOpt(const std::string &opt);
@@ -78,7 +77,7 @@ class ACOScheduler : public ConstrainedScheduler {
   Choice SelectInstruction(const llvm::ArrayRef<Choice> &ready,
                            SchedInstruction *lastInst);
   void UpdatePheromone(InstSchedule *schedule);
-  std::unique_ptr<InstSchedule> FindOneSchedule();
+  std::unique_ptr<InstSchedule> FindOneSchedule(InstCount TargetRPCost);
   llvm::SmallVector<pheromone_t, 0> pheromone_;
   pheromone_t initialValue_;
   bool use_fixed_bias;
@@ -90,6 +89,8 @@ class ACOScheduler : public ConstrainedScheduler {
   double local_decay;
   double decay_factor;
   int ants_per_iteration;
+  int ants_per_iteration1p;
+  int ants_per_iteration2p;
   int noImprovementMax;
   bool print_aco_trace;
   std::unique_ptr<InstSchedule> InitialSchedule;
@@ -99,6 +100,7 @@ class ACOScheduler : public ConstrainedScheduler {
   pheromone_t ScRelMax;
   DCF_OPT DCFOption;
   SPILL_COST_FUNCTION DCFCostFn;
+  int localCmp = 0, localCmpRej = 0, globalCmp = 0, globalCmpRej = 0;
 };
 
 } // namespace opt_sched

include/opt-sched/Scheduler/bb_spill.h

@@ -80,6 +80,7 @@ class BBWithSpill : public SchedRegion {
   InstCount totSpillCost_;
   InstCount slilSpillCost_;
   bool trackLiveRangeLngths_;
+  bool NeedsComputeSLIL;
 
   // Virtual Functions:
   // Given a schedule, compute the cost function value
@@ -127,6 +128,8 @@ class BBWithSpill : public SchedRegion {
   // cost of the schedule using Scf whenever the spill cost updates
   void addRecordedCost(SPILL_COST_FUNCTION Scf);
   void storeExtraCost(InstSchedule *sched, SPILL_COST_FUNCTION Scf);
+  InstCount getUnnormalizedIncrementalRPCost() const;
+
   void CmputAndSetCostLwrBound();
   int cmputSpillCostLwrBound();
 
@@ -149,6 +152,7 @@ class BBWithSpill : public SchedRegion {
                     InstCount slotNum, EnumTreeNode *trgtNode);
   void SetSttcLwrBounds(EnumTreeNode *node);
   bool ChkInstLglty(SchedInstruction *inst);
+  bool needsSLIL() const;
   void InitForSchdulng();
 
 protected:

include/opt-sched/Scheduler/data_dep.h

@@ -633,7 +633,7 @@ class InstSchedule {
   InstCount NormSpillCost;
 
   // Stores the spill cost of other spill cost functions
-  InstCount storedSC[MAX_SCHED_PRIRTS];
+  InstCount storedSC[MAX_SCF_TYPES];
 
   // An array of peak reg pressures for all reg types in the schedule
   InstCount *peakRegPressures_;

include/opt-sched/Scheduler/gen_sched.h

@@ -33,7 +33,9 @@ enum SchedulerType {
   // List scheduler.
   SCHED_LIST,
   // Sequential list scheduler.
-  SCHED_SEQ
+  SCHED_SEQ,
+  // Stalling list scheduler
+  SCHED_STALLING_LIST,
 };
 
 // Forward declarations used to reduce the number of #includes.

include/opt-sched/Scheduler/list_sched.h

@@ -52,6 +52,20 @@ class SequentialListScheduler : public ListScheduler {
   bool ChkInstLglty_(SchedInstruction *inst) const override;
 };
 
+// A list scheduler that schedules the instruction with the top heuristic value
+// Unalike ListScheduler this class considers instructions that are ready
+// in terms of data dependencies, but not in terms of latencies.
+// If the instruction with the top heuristic is not ready in terms of latency
+// Then stalls will be inserted until it is ready
+class StallSchedulingListScheduler : public ListScheduler {
+public:
+  StallSchedulingListScheduler(DataDepGraph *dataDepGraph,
+                               MachineModel *machMdl, InstCount schedUprBound,
+                               SchedPriorities prirts);
+
+  SchedInstruction *PickInst() const;
+};
+
 } // namespace opt_sched
 } // namespace llvm
 

include/opt-sched/Scheduler/sched_basic_data.h

@@ -83,6 +83,8 @@ enum SPILL_COST_FUNCTION {
   SCF_TARGET
 };
 
+#define MAX_SCF_TYPES 10
+
 // The type of instruction signatures, used by the enumerator's history table to
 // keep track of partial schedules.
 typedef UDT_HASHKEY InstSignature;

include/opt-sched/Scheduler/sched_region.h

@@ -73,6 +73,8 @@ class SchedRegion {
   // Get the number of simulated spills code added for this block.
   inline int GetSimSpills() { return totalSimSpills_; }
 
+  // Gets the un-normalized incremental RP cost for the region(used by ACO)
+  virtual InstCount getUnnormalizedIncrementalRPCost() const = 0;
   // Get schedLength for best-so-far sched
   inline InstCount getBestSchedLength() { return bestSchedLngth_; }
 
@@ -312,6 +314,8 @@ class SchedRegion {
 
   virtual bool EnableEnum_() = 0;
 
+  virtual bool needsSLIL() const = 0;
+
   // Prepares the region for being scheduled.
   virtual void SetupForSchdulng_() = 0;