[PBQP] Tweak spill costs and coalescing benefits

This patch improves how the different costs (register, interference, spill
and coalescing) relates together. The assumption is now that:
 - coalescing (or any other "side effect" of reg alloc) is negative, and
   instead of being derived from a spill cost, they use the block
   frequency info.
 - spill costs are in the [MinSpillCost:+inf( range
 - register or interference costs are in [0.0:MinSpillCost( or +inf

The current MinSpillCost is set to 10.0, which is a random value high
enough that the current constraint builders do not need to worry about
when settings costs. It would however be worth adding a normalization
step for register and interference costs as the last step in the
constraint builder chain to ensure they are not greater than SpillMinCost
(unless this has some sense for some architectures). This would work well
with the current builder pipeline, where all costs are tweaked relatively
to each others, but could grow above MinSpillCost if the pipeline is
deep enough.

The current heuristic is tuned to depend rather on the number of uses of
a live interval rather than a density of uses, as used by the greedy
allocator. This heuristic provides a few percent improvement on a number
of benchmarks (eembc, spec, ...) and will definitely need to change once
spill placement is implemented: the current spill placement is really
ineficient, so making the cost proportionnal to the number of use is a
clear win.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@221292 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/CodeGen/CalcSpillWeights.h

@@ -30,8 +30,10 @@ namespace llvm {
   /// @param UseDefFreq Expected number of executed use and def instructions
   ///                   per function call. Derived from block frequencies.
   /// @param Size       Size of live interval as returnexd by getSize()
+  /// @param NumInstr   Number of instructions using this live interval
   ///
-  static inline float normalizeSpillWeight(float UseDefFreq, unsigned Size) {
+  static inline float normalizeSpillWeight(float UseDefFreq, unsigned Size,
+                                           unsigned NumInstr) {
     // The constant 25 instructions is added to avoid depending too much on
     // accidental SlotIndex gaps for small intervals. The effect is that small
     // intervals have a spill weight that is mostly proportional to the number
@@ -44,7 +46,7 @@ namespace llvm {
   /// spill weight and allocation hint.
   class VirtRegAuxInfo {
   public:
-    typedef float (*NormalizingFn)(float, unsigned);
+    typedef float (*NormalizingFn)(float, unsigned, unsigned);
 
   private:
     MachineFunction &MF;

lib/CodeGen/CalcSpillWeights.cpp

@@ -100,6 +100,7 @@ VirtRegAuxInfo::calculateSpillWeightAndHint(LiveInterval &li) {
   MachineLoop *loop = nullptr;
   bool isExiting = false;
   float totalWeight = 0;
+  unsigned numInstr = 0; // Number of instructions using li
   SmallPtrSet<MachineInstr*, 8> visited;
 
   // Find the best physreg hint and the best virtreg hint.
@@ -116,6 +117,7 @@ VirtRegAuxInfo::calculateSpillWeightAndHint(LiveInterval &li) {
        I = mri.reg_instr_begin(li.reg), E = mri.reg_instr_end();
        I != E; ) {
     MachineInstr *mi = &*(I++);
+    numInstr++;
     if (mi->isIdentityCopy() || mi->isImplicitDef() || mi->isDebugValue())
       continue;
     if (!visited.insert(mi))
@@ -189,5 +191,5 @@ VirtRegAuxInfo::calculateSpillWeightAndHint(LiveInterval &li) {
   if (isRematerializable(li, LIS, *MF.getSubtarget().getInstrInfo()))
     totalWeight *= 0.5F;
 
-  li.weight = normalize(totalWeight, li.getSize());
+  li.weight = normalize(totalWeight, li.getSize(), numInstr);
 }

lib/CodeGen/RegAllocGreedy.cpp

@@ -1789,9 +1789,11 @@ unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
         // instructions.
         //
         // Try to guess the size of the new interval.
-        const float EstWeight = normalizeSpillWeight(blockFreq * (NewGaps + 1),
+        const float EstWeight = normalizeSpillWeight(
-                                 Uses[SplitBefore].distance(Uses[SplitAfter]) +
+            blockFreq * (NewGaps + 1),
-                                 (LiveBefore + LiveAfter)*SlotIndex::InstrDist);
+            Uses[SplitBefore].distance(Uses[SplitAfter]) +
+                (LiveBefore + LiveAfter) * SlotIndex::InstrDist,
+            1);
         // Would this split be possible to allocate?
         // Never allocate all gaps, we wouldn't be making progress.
         DEBUG(dbgs() << " w=" << EstWeight);

lib/CodeGen/RegAllocPBQP.cpp

@@ -150,11 +150,17 @@ public:
   void apply(PBQPRAGraph &G) override {
     LiveIntervals &LIS = G.getMetadata().LIS;
 
+    // A minimum spill costs, so that register constraints can can be set
+    // without normalization in the [0.0:MinSpillCost( interval.
+    const PBQP::PBQPNum MinSpillCost = 10.0;
+
     for (auto NId : G.nodeIds()) {
       PBQP::PBQPNum SpillCost =
         LIS.getInterval(G.getNodeMetadata(NId).getVReg()).weight;
       if (SpillCost == 0.0)
         SpillCost = std::numeric_limits<PBQP::PBQPNum>::min();
+      else
+        SpillCost += MinSpillCost;
       PBQPRAGraph::RawVector NodeCosts(G.getNodeCosts(NId));
       NodeCosts[PBQP::RegAlloc::getSpillOptionIdx()] = SpillCost;
       G.setNodeCosts(NId, std::move(NodeCosts));
@@ -350,11 +356,8 @@ public:
         unsigned DstReg = CP.getDstReg();
         unsigned SrcReg = CP.getSrcReg();
 
-        const float CopyFactor = 0.5; // Cost of copy relative to load. Current
+        const float Scale = 1.0f / MBFI.getEntryFreq();
-                                      // value plucked randomly out of the air.
+        PBQP::PBQPNum CBenefit = MBFI.getBlockFreq(&MBB).getFrequency() * Scale;
-
-        PBQP::PBQPNum CBenefit =
-          CopyFactor * LiveIntervals::getSpillWeight(false, true, &MBFI, &MI);
 
         if (CP.isPhys()) {
           if (!MF.getRegInfo().isAllocatable(DstReg))
@@ -607,12 +610,20 @@ void RegAllocPBQP::finalizeAlloc(MachineFunction &MF,
   }
 }
 
+static inline float normalizePBQPSpillWeight(float UseDefFreq, unsigned Size,
+                                         unsigned NumInstr) {
+  // All intervals have a spill weight that is mostly proportional to the number
+  // of uses, with uses in loops having a bigger weight.
+  return NumInstr * normalizeSpillWeight(UseDefFreq, Size, 1);
+}
+
 bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
   LiveIntervals &LIS = getAnalysis<LiveIntervals>();
   MachineBlockFrequencyInfo &MBFI =
     getAnalysis<MachineBlockFrequencyInfo>();
 
-  calculateSpillWeightsAndHints(LIS, MF, getAnalysis<MachineLoopInfo>(), MBFI);
+  calculateSpillWeightsAndHints(LIS, MF, getAnalysis<MachineLoopInfo>(), MBFI,
+                                normalizePBQPSpillWeight);
 
   VirtRegMap &VRM = getAnalysis<VirtRegMap>();