misched: Allow subtargets to enable misched and dependent options.

This allows me to begin enabling (or backing out) misched by default
for one subtarget at a time. To run misched we typically want to:
- Disable SelectionDAG scheduling (use the source order scheduler)
- Enable more aggressive coalescing (until we decide to always run the coalescer this way)
- Enable MachineScheduler pass itself.

Disabling PostRA sched may follow for some subtargets.

llvm-svn: 167826

llvm/include/llvm/Target/TargetSubtargetInfo.h

@@ -54,6 +54,13 @@ public:
     return 0;
   }
 
+  /// \brief True if the subtarget should run MachineScheduler after aggressive
+  /// coalescing.
+  ///
+  /// This currently replaces the SelectionDAG scheduler with the "source" order
+  /// scheduler. It does not yet disable the postRA scheduler.
+  virtual bool enableMachineScheduler() const;
+
   // enablePostRAScheduler - If the target can benefit from post-regalloc
   // scheduling and the specified optimization level meets the requirement
   // return true to enable post-register-allocation scheduling. In

llvm/lib/CodeGen/MachineScheduler.cpp

@@ -60,11 +60,11 @@ static cl::opt<unsigned> ILPWindow("ilp-window", cl::Hidden,
 
 // Experimental heuristics
 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
-  cl::desc("Enable load clustering."));
+  cl::desc("Enable load clustering."), cl::init(true));
 
 // Experimental heuristics
 static cl::opt<bool> EnableMacroFusion("misched-fusion", cl::Hidden,
-  cl::desc("Enable scheduling for macro fusion."));
+  cl::desc("Enable scheduling for macro fusion."), cl::init(true));
 
 //===----------------------------------------------------------------------===//
 // Machine Instruction Scheduling Pass and Registry

llvm/lib/CodeGen/Passes.cpp

@@ -22,6 +22,7 @@
 #include "llvm/CodeGen/RegAllocRegistry.h"
 #include "llvm/Target/TargetLowering.h"
 #include "llvm/Target/TargetOptions.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/Assembly/PrintModulePass.h"
 #include "llvm/Support/CommandLine.h"
@@ -241,7 +242,9 @@ TargetPassConfig::TargetPassConfig(TargetMachine *tm, PassManagerBase &pm)
   disablePass(&EarlyIfConverterID);
 
   // Temporarily disable experimental passes.
-  substitutePass(&MachineSchedulerID, 0);
+  const TargetSubtargetInfo &ST = TM->getSubtarget<TargetSubtargetInfo>();
+  if (!ST.enableMachineScheduler())
+    disablePass(&MachineSchedulerID);
 }
 
 /// Insert InsertedPassID pass after TargetPassID.

llvm/lib/CodeGen/RegisterCoalescer.cpp

@@ -45,6 +45,7 @@
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetOptions.h"
 #include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
 #include <algorithm>
 #include <cmath>
 using namespace llvm;
@@ -64,16 +65,16 @@ EnableJoining("join-liveintervals",
               cl::init(true));
 
 // Temporary flag to test critical edge unsplitting.
-static cl::opt<bool>
+static cl::opt<cl::boolOrDefault>
 EnableJoinSplits("join-splitedges",
-                 cl::desc("Coalesce copies on split edges (default=false)"),
-                 cl::init(false), cl::Hidden);
+  cl::desc("Coalesce copies on split edges (default=subtarget)"),
+  cl::init(cl::BOU_UNSET), cl::Hidden);
 
 // Temporary flag to test global copy optimization.
-static cl::opt<bool>
+static cl::opt<cl::boolOrDefault>
 EnableGlobalCopies("join-globalcopies",
-                   cl::desc("Coalesce copies that don't locally define an lrg"),
-                   cl::init(false));
+  cl::desc("Coalesce copies that span blocks (default=subtarget)"),
+  cl::init(cl::BOU_UNSET), cl::Hidden);
 
 static cl::opt<bool>
 VerifyCoalescing("verify-coalescing",
@@ -94,6 +95,14 @@ namespace {
     AliasAnalysis *AA;
     RegisterClassInfo RegClassInfo;
 
+    /// \brief True if the coalescer should aggressively coalesce global copies
+    /// in favor of keeping local copies.
+    bool JoinGlobalCopies;
+
+    /// \brief True if the coalescer should aggressively coalesce fall-thru
+    /// blocks exclusively containing copies.
+    bool JoinSplitEdges;
+
     /// WorkList - Copy instructions yet to be coalesced.
     SmallVector<MachineInstr*, 8> WorkList;
     SmallVector<MachineInstr*, 8> LocalWorkList;
@@ -1943,6 +1952,10 @@ namespace {
   //
   // EnableGlobalCopies assumes that the primary sort key is loop depth.
   struct MBBPriorityCompare {
+    bool JoinSplitEdges;
+
+    MBBPriorityCompare(bool joinsplits): JoinSplitEdges(joinsplits) {}
+
     bool operator()(const MBBPriorityInfo &LHS,
                     const MBBPriorityInfo &RHS) const {
       // Deeper loops first
@@ -1950,7 +1963,7 @@ namespace {
         return LHS.Depth > RHS.Depth;
 
       // Try to unsplit critical edges next.
-      if (EnableJoinSplits && LHS.IsSplit != RHS.IsSplit)
+      if (JoinSplitEdges && LHS.IsSplit != RHS.IsSplit)
         return LHS.IsSplit;
 
       // Prefer blocks that are more connected in the CFG. This takes care of
@@ -2011,7 +2024,7 @@ RegisterCoalescer::copyCoalesceInMBB(MachineBasicBlock *MBB) {
   // Collect all copy-like instructions in MBB. Don't start coalescing anything
   // yet, it might invalidate the iterator.
   const unsigned PrevSize = WorkList.size();
-  if (EnableGlobalCopies) {
+  if (JoinGlobalCopies) {
     // Coalesce copies bottom-up to coalesce local defs before local uses. They
     // are not inherently easier to resolve, but slightly preferable until we
     // have local live range splitting. In particular this is required by
@@ -2061,13 +2074,13 @@ void RegisterCoalescer::joinAllIntervals() {
     MBBs.push_back(MBBPriorityInfo(MBB, Loops->getLoopDepth(MBB),
                                    isSplitEdge(MBB)));
   }
-  std::sort(MBBs.begin(), MBBs.end(), MBBPriorityCompare());
+  std::sort(MBBs.begin(), MBBs.end(), MBBPriorityCompare(JoinSplitEdges));
 
   // Coalesce intervals in MBB priority order.
   unsigned CurrDepth = UINT_MAX;
   for (unsigned i = 0, e = MBBs.size(); i != e; ++i) {
     // Try coalescing the collected local copies for deeper loops.
-    if (EnableGlobalCopies && MBBs[i].Depth < CurrDepth)
+    if (JoinGlobalCopies && MBBs[i].Depth < CurrDepth)
       coalesceLocals();
     copyCoalesceInMBB(MBBs[i].MBB);
   }
@@ -2097,6 +2110,17 @@ bool RegisterCoalescer::runOnMachineFunction(MachineFunction &fn) {
   AA = &getAnalysis<AliasAnalysis>();
   Loops = &getAnalysis<MachineLoopInfo>();
 
+  const TargetSubtargetInfo &ST = TM->getSubtarget<TargetSubtargetInfo>();
+  if (EnableGlobalCopies == cl::BOU_UNSET)
+    JoinGlobalCopies = ST.enableMachineScheduler();
+  else
+    JoinGlobalCopies = (EnableGlobalCopies == cl::BOU_TRUE);
+
+  if (EnableJoinSplits == cl::BOU_UNSET)
+    JoinSplitEdges = ST.enableMachineScheduler();
+  else
+    JoinSplitEdges = (EnableJoinSplits == cl::BOU_TRUE);
+
   DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n"
                << "********** Function: " << MF->getName() << '\n');
 

llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

@@ -45,6 +45,7 @@
 #include "llvm/Target/TargetLowering.h"
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetOptions.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Support/Compiler.h"
 #include "llvm/Support/Debug.h"
@@ -216,8 +217,9 @@ namespace llvm {
   ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS,
                                              CodeGenOpt::Level OptLevel) {
     const TargetLowering &TLI = IS->getTargetLowering();
+    const TargetSubtargetInfo &ST = IS->TM.getSubtarget<TargetSubtargetInfo>();
 
-    if (OptLevel == CodeGenOpt::None ||
+    if (OptLevel == CodeGenOpt::None || ST.enableMachineScheduler() ||
         TLI.getSchedulingPreference() == Sched::Source)
       return createSourceListDAGScheduler(IS, OptLevel);
     if (TLI.getSchedulingPreference() == Sched::RegPressure)

llvm/lib/Target/TargetSubtargetInfo.cpp

@@ -22,6 +22,10 @@ TargetSubtargetInfo::TargetSubtargetInfo() {}
 
 TargetSubtargetInfo::~TargetSubtargetInfo() {}
 
+bool TargetSubtargetInfo::enableMachineScheduler() const {
+  return false;
+}
+
 bool TargetSubtargetInfo::enablePostRAScheduler(
           CodeGenOpt::Level OptLevel,
           AntiDepBreakMode& Mode,