MIsched: add an ILP window property to machine model.

This was an experimental option, but needs to be defined
per-target. e.g. PPC A2 needs to aggressively hide latency.

I converted some in-order scheduling tests to A2. Hal is working on
more test cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@171946 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/CodeGen/TargetSchedule.h

@@ -84,6 +84,9 @@ public:
   /// \brief Maximum number of micro-ops that may be scheduled per cycle.
   unsigned getIssueWidth() const { return SchedModel.IssueWidth; }
 
+  /// \brief Number of cycles the OOO processor is expected to hide.
+  unsigned getILPWindow() const { return SchedModel.ILPWindow; }
+
   /// \brief Return the number of issue slots required for this MI.
   unsigned getNumMicroOps(const MachineInstr *MI,
                           const MCSchedClassDesc *SC = 0) const;

include/llvm/MC/MCSchedule.h

@@ -155,7 +155,7 @@ public:
   //      Optional InstrItinerary OperandCycles provides expected latency.
   //      TODO: can't yet specify both min and expected latency per operand.
   int MinLatency;
-  static const unsigned DefaultMinLatency = -1;
+  static const int DefaultMinLatency = -1;
 
   // LoadLatency is the expected latency of load instructions.
   //
@@ -172,6 +172,16 @@ public:
   unsigned HighLatency;
   static const unsigned DefaultHighLatency = 10;
 
+  // ILPWindow is the number of cycles that the scheduler effectively ignores
+  // before attempting to hide latency. This should be zero for in-order cpus to
+  // always hide expected latency. For out-of-order cpus, it may be tweaked as
+  // desired to roughly approximate instruction buffers. The actual threshold is
+  // not very important for an OOO processor, as long as it isn't too high. A
+  // nonzero value helps avoid rescheduling to hide latency when its is fairly
+  // obviously useless and makes register pressure heuristics more effective.
+  unsigned ILPWindow;
+  static const unsigned DefaultILPWindow = 0;
+
   // MispredictPenalty is the typical number of extra cycles the processor
   // takes to recover from a branch misprediction.
   unsigned MispredictPenalty;
@@ -196,6 +206,7 @@ public:
                   MinLatency(DefaultMinLatency),
                   LoadLatency(DefaultLoadLatency),
                   HighLatency(DefaultHighLatency),
+                  ILPWindow(DefaultILPWindow),
                   MispredictPenalty(DefaultMispredictPenalty),
                   ProcID(0), ProcResourceTable(0), SchedClassTable(0),
                   NumProcResourceKinds(0), NumSchedClasses(0),
@@ -205,12 +216,12 @@ public:
   }
 
   // Table-gen driven ctor.
-  MCSchedModel(unsigned iw, int ml, unsigned ll, unsigned hl, unsigned mp,
-               unsigned pi, const MCProcResourceDesc *pr,
+  MCSchedModel(unsigned iw, int ml, unsigned ll, unsigned hl, unsigned ilp,
+               unsigned mp, unsigned pi, const MCProcResourceDesc *pr,
                const MCSchedClassDesc *sc, unsigned npr, unsigned nsc,
                const InstrItinerary *ii):
     IssueWidth(iw), MinLatency(ml), LoadLatency(ll), HighLatency(hl),
-    MispredictPenalty(mp), ProcID(pi), ProcResourceTable(pr),
+    ILPWindow(ilp), MispredictPenalty(mp), ProcID(pi), ProcResourceTable(pr),
     SchedClassTable(sc), NumProcResourceKinds(npr), NumSchedClasses(nsc),
     InstrItineraries(ii) {}
 

include/llvm/Target/TargetSchedule.td

@@ -76,6 +76,7 @@ class SchedMachineModel {
   int IssueWidth = -1; // Max micro-ops that may be scheduled per cycle.
   int MinLatency = -1; // Determines which instrucions are allowed in a group.
                        // (-1) inorder (0) ooo, (1): inorder +var latencies.
+  int ILPWindow = -1;  // Cycles of latency likely hidden by hardware buffers.
   int LoadLatency = -1; // Cycles for loads to access the cache.
   int HighLatency = -1; // Approximation of cycles for "high latency" ops.
   int MispredictPenalty = -1; // Extra cycles for a mispredicted branch.

lib/CodeGen/MachineScheduler.cpp

@@ -48,15 +48,6 @@ static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
 static bool ViewMISchedDAGs = false;
 #endif // NDEBUG
 
-// Threshold to very roughly model an out-of-order processor's instruction
-// buffers. If the actual value of this threshold matters much in practice, then
-// it can be specified by the machine model. For now, it's an experimental
-// tuning knob to determine when and if it matters.
-static cl::opt<unsigned> ILPWindow("ilp-window", cl::Hidden,
-  cl::desc("Allow expected latency to exceed the critical path by N cycles "
-           "before attempting to balance ILP"),
-  cl::init(10U));
-
 // Experimental heuristics
 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
   cl::desc("Enable load clustering."), cl::init(true));
@@ -1297,7 +1288,8 @@ void ConvergingScheduler::SchedBoundary::setLatencyPolicy(CandPolicy &Policy) {
     if (L > RemLatency)
       RemLatency = L;
   }
-  if (RemLatency + ExpectedLatency >= Rem->CriticalPath + ILPWindow
+  unsigned CriticalPathLimit = Rem->CriticalPath + SchedModel->getILPWindow();
+  if (RemLatency + ExpectedLatency >= CriticalPathLimit
       && RemLatency > Rem->getMaxRemainingCount(SchedModel)) {
     Policy.ReduceLatency = true;
     DEBUG(dbgs() << "Increase ILP: " << Available.getName() << '\n');

lib/Target/ARM/ARMScheduleA9.td

@@ -1887,6 +1887,9 @@ def CortexA9Model : SchedMachineModel {
   let LoadLatency = 2; // Optimistic load latency assuming bypass.
                        // This is overriden by OperandCycles if the
                        // Itineraries are queried instead.
+  let ILPWindow = 10; // Don't reschedule small blocks to hide
+                      // latency. Minimum latency requirements are already
+                      // modeled strictly by reserving resources.
   let MispredictPenalty = 8; // Based on estimate of pipeline depth.
 
   let Itineraries = CortexA9Itineraries;

lib/Target/X86/X86Schedule.td

@@ -470,12 +470,17 @@ def IIC_NOP : InstrItinClass;
 // latencies. Since these latencies are not used for pipeline hazards,
 // they do not need to be exact.
 //
+// ILPWindow=10 is an arbitrary threshold that approximates cycles of
+// latency hidden by instruction buffers. The actual value is not very
+// important but should be zero for inorder and nonzero for OOO processors.
+//
 // The GenericModel contains no instruciton itineraries.
 def GenericModel : SchedMachineModel {
   let IssueWidth = 4;
   let MinLatency = 0;
   let LoadLatency = 4;
   let HighLatency = 10;
+  let ILPWindow = 10;
 }
 
 include "X86ScheduleAtom.td"

lib/Target/X86/X86ScheduleAtom.td

@@ -525,6 +525,7 @@ def AtomModel : SchedMachineModel {
                        // OperandCycles may be used for expected latency.
   let LoadLatency = 3; // Expected cycles, may be overriden by OperandCycles.
   let HighLatency = 30;// Expected, may be overriden by OperandCycles.
+  let ILPWindow = 0; // Always try to hide expected latency.
 
   let Itineraries = AtomItineraries;
 }

test/CodeGen/PowerPC/misched-inorder-latency.ll

@@ -1,15 +1,15 @@
-; RUN: llc < %s -enable-misched -march=thumb -mcpu=swift \
-; RUN:          -pre-RA-sched=source -scheditins=false -ilp-window=0 \
+; RUN: llc < %s -enable-misched -pre-RA-sched=source -scheditins=false \
 ; RUN:          -disable-ifcvt-triangle-false -disable-post-ra | FileCheck %s
 ;
-; For these tests, we set -ilp-window=0 to simulate in order processor.
+target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f128:128:128-v128:128:128-n32:64"
+target triple = "powerpc64-bgq-linux"
 
-; %val1 is a 3-cycle load live out of %entry. It should be hoisted
+; %val1 is a load live out of %entry. It should be hoisted
 ; above the add.
-; CHECK: @testload
+; CHECK: testload:
 ; CHECK: %entry
-; CHECK: ldr
-; CHECK: adds
+; CHECK: lwz
+; CHECK: addi
 ; CHECK: bne
 ; CHECK: %true
 define i32 @testload(i32 *%ptr, i32 %sumin) {
@@ -34,15 +34,22 @@ end:
 ; The prefetch gets a default latency of 3 cycles and should be hoisted
 ; above the add.
 ;
-; CHECK: @testprefetch
+; CHECK: testprefetch:
 ; CHECK: %entry
-; CHECK: pld
-; CHECK: adds
-; CHECK: bx
+; CHECK: dcbt
+; CHECK: addi
+; CHECK: blr
 define i32 @testprefetch(i8 *%ptr, i32 %i) {
 entry:
-  %tmp = add i32 %i, 1
+  %val1 = add i32 %i, 1
   tail call void @llvm.prefetch( i8* %ptr, i32 0, i32 3, i32 1 )
-  ret i32 %tmp
+  %p = icmp eq i32 %i, 0
+  br i1 %p, label %true, label %end
+true:
+  %val2 = add i32 %val1, 1
+  br label %end
+end:
+  %valmerge = phi i32 [ %val1, %entry], [ %val2, %true ]
+  ret i32 %valmerge
 }
 declare void @llvm.prefetch(i8*, i32, i32, i32) nounwind

utils/TableGen/SubtargetEmitter.cpp

@@ -1108,6 +1108,7 @@ void SubtargetEmitter::EmitProcessorModels(raw_ostream &OS) {
     EmitProcessorProp(OS, PI->ModelDef, "MinLatency", ',');
     EmitProcessorProp(OS, PI->ModelDef, "LoadLatency", ',');
     EmitProcessorProp(OS, PI->ModelDef, "HighLatency", ',');
+    EmitProcessorProp(OS, PI->ModelDef, "ILPWindow", ',');
     EmitProcessorProp(OS, PI->ModelDef, "MispredictPenalty", ',');
     OS << "  " << PI->Index << ", // Processor ID\n";
     if (PI->hasInstrSchedModel())