Add support to model pipeline bypass / forwarding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115005 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Target/TargetSchedule.td

@@ -22,6 +22,14 @@
 //  
 class FuncUnit;
 
+//===----------------------------------------------------------------------===//
+// Pipeline bypass / forwarding - These values specifies the symbolic names of
+// pipeline bypasses which can be used to forward results of instructions
+// that are forwarded to uses.
+class Bypass;
+
+def NoBypass : Bypass;
+
 class ReservationKind<bits<1> val> {
   int Value = val;
 }
@@ -81,22 +89,26 @@ def NoItinerary : InstrItinClass;
 // instruction itinerary class (name) to its itinerary data.
 //
 class InstrItinData<InstrItinClass Class, list<InstrStage> stages,
-                    list<int> operandcycles = []> {
+                    list<int> operandcycles = [],
+                    list<Bypass> bypasses = []> {
   InstrItinClass TheClass = Class;
   list<InstrStage> Stages = stages;
   list<int> OperandCycles = operandcycles;
+  list<Bypass> Bypasses = bypasses;
 }
 
 //===----------------------------------------------------------------------===//
 // Processor itineraries - These values represent the set of all itinerary
 // classes for a given chip set.
 //
-class ProcessorItineraries<list<FuncUnit> fu, list<InstrItinData> iid> {
+class ProcessorItineraries<list<FuncUnit> fu, list<Bypass> bp,
+                           list<InstrItinData> iid> {
   list<FuncUnit> FU = fu;
+  list<Bypass> BP = bp;
   list<InstrItinData> IID = iid;
 }
 
 // NoItineraries - A marker that can be used by processors without schedule
 // info.
-def NoItineraries : ProcessorItineraries<[], []>;
+def NoItineraries : ProcessorItineraries<[], [], []>;
 

lib/Target/ARM/ARMSchedule.td

@@ -156,7 +156,7 @@ def IIC_VTBX4      : InstrItinClass;
 //===----------------------------------------------------------------------===//
 // Processor instruction itineraries.
 
-def GenericItineraries : ProcessorItineraries<[], []>;
+def GenericItineraries : ProcessorItineraries<[], [], []>;
 
 include "ARMScheduleV6.td"
 include "ARMScheduleA8.td"

lib/Target/ARM/ARMScheduleA8.td

@@ -25,7 +25,8 @@ def A8_NLSPipe : FuncUnit; // NEON LS pipe
 // Dual issue pipeline represented by A8_Pipe0 | A8_Pipe1
 //
 def CortexA8Itineraries : ProcessorItineraries<
-  [A8_Issue, A8_Pipe0, A8_Pipe1, A8_LdSt0, A8_LdSt1, A8_NPipe, A8_NLSPipe], [
+  [A8_Issue, A8_Pipe0, A8_Pipe1, A8_LdSt0, A8_LdSt1, A8_NPipe, A8_NLSPipe],
+  [], [
   // Two fully-pipelined integer ALU pipelines
   //
   // No operand cycles

lib/Target/ARM/ARMScheduleA9.td

@@ -26,7 +26,7 @@ def A9_DRegsN  : FuncUnit; // FP register set, NEON side
 // Dual issue pipeline represented by A9_Pipe0 | A9_Pipe1
 //
 def CortexA9Itineraries : ProcessorItineraries<
-  [A9_NPipe, A9_DRegsN, A9_DRegsVFP, A9_LSPipe, A9_Pipe0, A9_Pipe1], [
+  [A9_NPipe, A9_DRegsN, A9_DRegsVFP, A9_LSPipe, A9_Pipe0, A9_Pipe1], [], [
   // Two fully-pipelined integer ALU pipelines
   // FIXME: There are no operand latencies for these instructions at all!
   //

lib/Target/ARM/ARMScheduleV6.td

@@ -19,7 +19,7 @@ def V6_Pipe : FuncUnit; // pipeline
 // Scheduling information derived from "ARM1176JZF-S Technical Reference Manual"
 //
 def ARMV6Itineraries : ProcessorItineraries<
-  [V6_Pipe], [
+  [V6_Pipe], [], [
   //
   // No operand cycles
   InstrItinData<IIC_iALUx    , [InstrStage<1, [V6_Pipe]>]>,

lib/Target/Alpha/AlphaSchedule.td

@@ -54,7 +54,7 @@ def s_pseudo : InstrItinClass;
 //modified some
 
 def Alpha21264Itineraries : ProcessorItineraries<
-  [L0, L1, FST0, FST1, U0, U1, FA, FM], [
+  [L0, L1, FST0, FST1, U0, U1, FA, FM], [], [
   InstrItinData<s_ild    , [InstrStage<3, [L0, L1]>]>,
   InstrItinData<s_fld    , [InstrStage<4, [L0, L1]>]>,
   InstrItinData<s_ist    , [InstrStage<0, [L0, L1]>]>,

lib/Target/CellSPU/SPUSchedule.td

@@ -36,7 +36,7 @@ def RotateShift  : InstrItinClass;              // EVEN_UNIT
 def ImmLoad      : InstrItinClass;              // EVEN_UNIT
 
 /* Note: The itinerary for the Cell SPU is somewhat contrived... */
-def SPUItineraries : ProcessorItineraries<[ODD_UNIT, EVEN_UNIT], [
+def SPUItineraries : ProcessorItineraries<[ODD_UNIT, EVEN_UNIT], [], [
   InstrItinData<LoadStore   , [InstrStage<6,  [ODD_UNIT]>]>,
   InstrItinData<BranchHints , [InstrStage<6,  [ODD_UNIT]>]>,
   InstrItinData<BranchResolv, [InstrStage<4,  [ODD_UNIT]>]>,

lib/Target/MBlaze/MBlazeSchedule.td

@@ -41,7 +41,7 @@ def IIPseudo           : InstrItinClass;
 // MBlaze Generic instruction itineraries.
 //===----------------------------------------------------------------------===//
 def MBlazeGenericItineraries : ProcessorItineraries<
-  [ALU, IMULDIV], [
+  [ALU, IMULDIV], [], [
   InstrItinData<IIAlu              , [InstrStage<1,  [ALU]>]>,
   InstrItinData<IILoad             , [InstrStage<3,  [ALU]>]>,
   InstrItinData<IIStore            , [InstrStage<1,  [ALU]>]>,

lib/Target/Mips/MipsSchedule.td

@@ -40,7 +40,7 @@ def IIPseudo           : InstrItinClass;
 //===----------------------------------------------------------------------===//
 // Mips Generic instruction itineraries.
 //===----------------------------------------------------------------------===//
-def MipsGenericItineraries : ProcessorItineraries<[ALU, IMULDIV], [
+def MipsGenericItineraries : ProcessorItineraries<[ALU, IMULDIV], [], [
   InstrItinData<IIAlu              , [InstrStage<1,  [ALU]>]>,
   InstrItinData<IILoad             , [InstrStage<3,  [ALU]>]>,
   InstrItinData<IIStore            , [InstrStage<1,  [ALU]>]>,

lib/Target/PowerPC/PPCScheduleG3.td

@@ -13,7 +13,7 @@
 
 
 def G3Itineraries : ProcessorItineraries<
-  [IU1, IU2, FPU1, BPU, SRU, SLU], [
+  [IU1, IU2, FPU1, BPU, SRU, SLU], [], [
   InstrItinData<IntGeneral  , [InstrStage<1, [IU1, IU2]>]>,
   InstrItinData<IntCompare  , [InstrStage<1, [IU1, IU2]>]>,
   InstrItinData<IntDivW     , [InstrStage<19, [IU1]>]>,

lib/Target/PowerPC/PPCScheduleG4.td

@@ -12,7 +12,7 @@
 //===----------------------------------------------------------------------===//
 
 def G4Itineraries : ProcessorItineraries<
-  [IU1, IU2, SLU, SRU, BPU, FPU1, VIU1, VIU2, VPU, VFPU], [
+  [IU1, IU2, SLU, SRU, BPU, FPU1, VIU1, VIU2, VPU, VFPU], [], [
   InstrItinData<IntGeneral  , [InstrStage<1, [IU1, IU2]>]>,
   InstrItinData<IntCompare  , [InstrStage<1, [IU1, IU2]>]>,
   InstrItinData<IntDivW     , [InstrStage<19, [IU1]>]>,

lib/Target/PowerPC/PPCScheduleG4Plus.td

@@ -15,7 +15,7 @@ def IU3    : FuncUnit; // integer unit 3 (7450 simple)
 def IU4    : FuncUnit; // integer unit 4 (7450 simple)
 
 def G4PlusItineraries : ProcessorItineraries<
-  [IU1, IU2, IU3, IU4, BPU, SLU, FPU1, VFPU, VIU1, VIU2, VPU], [
+  [IU1, IU2, IU3, IU4, BPU, SLU, FPU1, VFPU, VIU1, VIU2, VPU], [], [
   InstrItinData<IntGeneral  , [InstrStage<1, [IU1, IU2, IU3, IU4]>]>,
   InstrItinData<IntCompare  , [InstrStage<1, [IU1, IU2, IU3, IU4]>]>,
   InstrItinData<IntDivW     , [InstrStage<23, [IU2]>]>,

lib/Target/PowerPC/PPCScheduleG5.td

@@ -12,7 +12,7 @@
 //===----------------------------------------------------------------------===//
 
 def G5Itineraries : ProcessorItineraries<
-  [IU1, IU2, SLU, BPU, FPU1, FPU2, VFPU, VIU1, VIU2, VPU], [
+  [IU1, IU2, SLU, BPU, FPU1, FPU2, VFPU, VIU1, VIU2, VPU], [], [
   InstrItinData<IntGeneral  , [InstrStage<2, [IU1, IU2]>]>,
   InstrItinData<IntCompare  , [InstrStage<3, [IU1, IU2]>]>,
   InstrItinData<IntDivD     , [InstrStage<68, [IU1]>]>,

utils/TableGen/SubtargetEmitter.cpp

@@ -262,6 +262,24 @@ void SubtargetEmitter::FormItineraryOperandCycleString(Record *ItinData,
   }
 }
 
+void SubtargetEmitter::FormItineraryBypassString(const std::string &Name,
+                                                 Record *ItinData,
+                                                 std::string &ItinString,
+                                                 unsigned NOperandCycles) {
+  const std::vector<Record*> &BypassList =
+    ItinData->getValueAsListOfDefs("Bypasses");
+  unsigned N = BypassList.size();
+  for (unsigned i = 0; i < N;) {
+    ItinString += Name + "Bypass::" + BypassList[i]->getName();
+    if (++i < N) ItinString += ", ";
+  }
+
+  for (; N < NOperandCycles;) {
+    ItinString += " 0";
+    if (++N < NOperandCycles) ItinString += ", ";
+  }
+}
+
 //
 // EmitStageAndOperandCycleData - Generate unique itinerary stages and
 // operand cycle tables.  Record itineraries for processors.
@@ -296,6 +314,16 @@ void SubtargetEmitter::EmitStageAndOperandCycleData(raw_ostream &OS,
          << " = 1 << " << j << ";\n";
 
     OS << "}\n";
+
+    std::vector<Record*> BPs = Proc->getValueAsListOfDefs("BP");
+    OS << "\n// Pipeline bypasses for itineraries \"" << Name << "\"\n"
+       << "namespace " << Name << "Bypass {\n";
+
+    for (unsigned j = 0, BPN = BPs.size(); j < BPN; ++j)
+      OS << "  const unsigned " << BPs[j]->getName()
+         << " = 1 << " << j << ";\n";
+
+    OS << "}\n";
   }
 
   // Begin stages table
@@ -305,6 +333,10 @@ void SubtargetEmitter::EmitStageAndOperandCycleData(raw_ostream &OS,
   // Begin operand cycle table
   std::string OperandCycleTable = "static const unsigned OperandCycles[] = {\n";
   OperandCycleTable += "  0, // No itinerary\n";
+
+  // Begin pipeline bypass table
+  std::string BypassTable = "static const unsigned Bypasses[] = {\n";
+  BypassTable += "  0, // No itinerary\n";
         
   unsigned StageCount = 1, OperandCycleCount = 1;
   unsigned ItinStageEnum = 1, ItinOperandCycleEnum = 1;
@@ -342,6 +374,10 @@ void SubtargetEmitter::EmitStageAndOperandCycleData(raw_ostream &OS,
       FormItineraryOperandCycleString(ItinData, ItinOperandCycleString,
                                       NOperandCycles);
 
+      std::string ItinBypassString;
+      FormItineraryBypassString(Name, ItinData, ItinBypassString,
+                                NOperandCycles);
+
       // Check to see if stage already exists and create if it doesn't
       unsigned FindStage = 0;
       if (NStages > 0) {
@@ -367,6 +403,11 @@ void SubtargetEmitter::EmitStageAndOperandCycleData(raw_ostream &OS,
           // Record Itin class number.
           ItinOperandCycleMap[ItinOperandCycleString] = 
             FindOperandCycle = OperandCycleCount;
+
+          // Emit as bypass, // index
+          BypassTable += ItinBypassString + ", // " + 
+            itostr(ItinOperandCycleEnum) + "\n";
+
           OperandCycleCount += NOperandCycles;
           ItinOperandCycleEnum++;
         }
@@ -389,7 +430,7 @@ void SubtargetEmitter::EmitStageAndOperandCycleData(raw_ostream &OS,
     // Add process itinerary to list
     ProcList.push_back(ItinList);
   }
-  
+
   // Closing stage
   StageTable += "  { 0, 0, 0, llvm::InstrStage::Required } // End itinerary\n";
   StageTable += "};\n";
@@ -398,9 +439,13 @@ void SubtargetEmitter::EmitStageAndOperandCycleData(raw_ostream &OS,
   OperandCycleTable += "  0 // End itinerary\n";
   OperandCycleTable += "};\n";
 
+  BypassTable += "  0 // End itinerary\n";
+  BypassTable += "};\n";
+
   // Emit tables.
   OS << StageTable;
   OS << OperandCycleTable;
+  OS << BypassTable;
   
   // Emit size of tables
   OS<<"\nenum {\n";

utils/TableGen/SubtargetEmitter.h

@@ -40,6 +40,9 @@ class SubtargetEmitter : public TableGenBackend {
                                 unsigned &NStages);
   void FormItineraryOperandCycleString(Record *ItinData, std::string &ItinString,
                                        unsigned &NOperandCycles);
+  void FormItineraryBypassString(const std::string &Names,
+                                 Record *ItinData,
+                                 std::string &ItinString, unsigned NOperandCycles);
   void EmitStageAndOperandCycleData(raw_ostream &OS, unsigned NItinClasses,
                      std::map<std::string, unsigned> &ItinClassesMap,
                      std::vector<Record*> &ItinClassList,