Add function that translates subregister lane masks to other subregs.

This works like the composeSubRegisterIndices() function but transforms
a subregister lane mask instead of a subregister index.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223874 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Target/TargetRegisterInfo.h

@@ -510,6 +510,15 @@ public:
     return composeSubRegIndicesImpl(a, b);
   }
 
+  /// Transforms a LaneMask computed for one subregister to the lanemask that
+  /// would have been computed when composing the subsubregisters with IdxA
+  /// first. @sa composeSubRegIndices()
+  unsigned composeSubRegIndexLaneMask(unsigned IdxA, unsigned LaneMask) const {
+    if (!IdxA)
+      return LaneMask;
+    return composeSubRegIndexLaneMaskImpl(IdxA, LaneMask);
+  }
+
   /// Debugging helper: dump register in human readable form to dbgs() stream.
   static void dumpReg(unsigned Reg, unsigned SubRegIndex = 0,
                       const TargetRegisterInfo* TRI = nullptr);
@@ -520,6 +529,12 @@ protected:
     llvm_unreachable("Target has no sub-registers");
   }
 
+  /// Overridden by TableGen in targets that have sub-registers.
+  virtual unsigned
+  composeSubRegIndexLaneMaskImpl(unsigned, unsigned) const {
+    llvm_unreachable("Target has no sub-registers");
+  }
+
 public:
   /// getCommonSuperRegClass - Find a common super-register class if it exists.
   ///

utils/TableGen/CodeGenRegisters.cpp

@@ -1192,6 +1192,70 @@ void CodeGenRegBank::computeSubRegLaneMasks() {
     }
   }
 
+  // Compute transformation sequences for composeSubRegIndexLaneMask. The idea
+  // here is that for each possible target subregister we look at the leafs
+  // in the subregister graph that compose for this target and create
+  // transformation sequences for the lanemasks. Each step in the sequence
+  // consists of a bitmask and a bitrotate operation. As the rotation amounts
+  // are usually the same for many subregisters we can easily combine the steps
+  // by combining the masks.
+  for (const auto &Idx : SubRegIndices) {
+    const auto &Composites = Idx.getComposites();
+    auto &LaneTransforms = Idx.CompositionLaneMaskTransform;
+    // Go through all leaf subregisters and find the ones that compose with Idx.
+    // These make out all possible valid bits in the lane mask we want to
+    // transform. Looking only at the leafs ensure that only a single bit in
+    // the mask is set.
+    unsigned NextBit = 0;
+    for (auto &Idx2 : SubRegIndices) {
+      // Skip non-leaf subregisters.
+      if (!Idx2.getComposites().empty())
+        continue;
+      // Replicate the behaviour from the lane mask generation loop above.
+      unsigned SrcBit = NextBit;
+      unsigned SrcMask = 1u << SrcBit;
+      if (NextBit < 31)
+        ++NextBit;
+      assert(Idx2.LaneMask == SrcMask);
+
+      // Get the composed subregister if there is any.
+      auto C = Composites.find(&Idx2);
+      if (C == Composites.end())
+        continue;
+      const CodeGenSubRegIndex *Composite = C->second;
+      // The Composed subreg should be a leaf subreg too
+      assert(Composite->getComposites().empty());
+
+      // Create Mask+Rotate operation and merge with existing ops if possible.
+      unsigned DstBit = Log2_32(Composite->LaneMask);
+      int Shift = DstBit - SrcBit;
+      uint8_t RotateLeft = Shift >= 0 ? (uint8_t)Shift : 32+Shift;
+      for (auto &I : LaneTransforms) {
+        if (I.RotateLeft == RotateLeft) {
+          I.Mask |= SrcMask;
+          SrcMask = 0;
+        }
+      }
+      if (SrcMask != 0) {
+        MaskRolPair MaskRol = { SrcMask, RotateLeft };
+        LaneTransforms.push_back(MaskRol);
+      }
+    }
+    // Optimize if the transformation consists of one step only: Set mask to
+    // 0xffffffff (including some irrelevant invalid bits) so that it should
+    // merge with more entries later while compressing the table.
+    if (LaneTransforms.size() == 1)
+      LaneTransforms[0].Mask = ~0u;
+
+    // Further compression optimization: For invalid compositions resulting
+    // in a sequence with 0 entries we can just pick any other. Choose
+    // Mask 0xffffffff with Rotation 0.
+    if (LaneTransforms.size() == 0) {
+      MaskRolPair P = { ~0u, 0 };
+      LaneTransforms.push_back(P);
+    }
+  }
+
   // FIXME: What if ad-hoc aliasing introduces overlaps that aren't represented
   // by the sub-register graph? This doesn't occur in any known targets.
 

utils/TableGen/CodeGenRegisters.h

@@ -34,6 +34,20 @@
 namespace llvm {
   class CodeGenRegBank;
 
+  /// Used to encode a step in a register lane mask transformation.
+  /// Mask the bits specified in Mask, then rotate them Rol bits to the left
+  /// assuming a wraparound at 32bits.
+  struct MaskRolPair {
+    unsigned Mask;
+    uint8_t RotateLeft;
+    bool operator==(const MaskRolPair Other) {
+      return Mask == Other.Mask && RotateLeft == Other.RotateLeft;
+    }
+    bool operator!=(const MaskRolPair Other) {
+      return Mask != Other.Mask || RotateLeft != Other.RotateLeft;
+    }
+  };
+
   /// CodeGenSubRegIndex - Represents a sub-register index.
   class CodeGenSubRegIndex {
     Record *const TheDef;
@@ -45,6 +59,7 @@ namespace llvm {
     uint16_t Offset;
     const unsigned EnumValue;
     mutable unsigned LaneMask;
+    mutable SmallVector<MaskRolPair,1> CompositionLaneMaskTransform;
 
     // Are all super-registers containing this SubRegIndex covered by their
     // sub-registers?

utils/TableGen/RegisterInfoEmitter.cpp

@@ -62,6 +62,8 @@ private:
                            const std::string &ClassName);
   void emitComposeSubRegIndices(raw_ostream &OS, CodeGenRegBank &RegBank,
                                 const std::string &ClassName);
+  void emitComposeSubRegIndexLaneMask(raw_ostream &OS, CodeGenRegBank &RegBank,
+                                      const std::string &ClassName);
 };
 } // End anonymous namespace
 
@@ -687,6 +689,95 @@ RegisterInfoEmitter::emitComposeSubRegIndices(raw_ostream &OS,
   OS << "}\n\n";
 }
 
+void
+RegisterInfoEmitter::emitComposeSubRegIndexLaneMask(raw_ostream &OS,
+                                                    CodeGenRegBank &RegBank,
+                                                    const std::string &ClName) {
+  // See the comments in computeSubRegLaneMasks() for our goal here.
+  const auto &SubRegIndices = RegBank.getSubRegIndices();
+
+  // Create a list of Mask+Rotate operations, with equivalent entries merged.
+  SmallVector<unsigned, 4> SubReg2SequenceIndexMap;
+  SmallVector<SmallVector<MaskRolPair, 1>, 4> Sequences;
+  for (const auto &Idx : SubRegIndices) {
+    const SmallVector<MaskRolPair, 1> &IdxSequence
+      = Idx.CompositionLaneMaskTransform;
+
+    unsigned Found = ~0u;
+    unsigned SIdx = 0;
+    unsigned NextSIdx;
+    for (size_t s = 0, se = Sequences.size(); s != se; ++s, SIdx = NextSIdx) {
+      SmallVectorImpl<MaskRolPair> &Sequence = Sequences[s];
+      NextSIdx = SIdx + Sequence.size() + 1;
+      if (Sequence.size() != IdxSequence.size())
+        continue;
+      bool Identical = true;
+      for (size_t o = 0, oe = Sequence.size(); o != oe; ++o) {
+        if (Sequence[o] != IdxSequence[o]) {
+          Identical = false;
+          break;
+        }
+      }
+      if (Identical) {
+        Found = SIdx;
+        break;
+      }
+    }
+    if (Found == ~0u) {
+      Sequences.push_back(IdxSequence);
+      Found = SIdx;
+    }
+    SubReg2SequenceIndexMap.push_back(Found);
+  }
+
+  OS << "unsigned " << ClName
+     << "::composeSubRegIndexLaneMaskImpl(unsigned IdxA, unsigned LaneMask)"
+        " const {\n";
+
+  OS << "  struct MaskRolOp {\n"
+        "    unsigned Mask;\n"
+        "    uint8_t  RotateLeft;\n"
+        "  };\n"
+        "  static const MaskRolOp Seqs[] = {\n";
+  unsigned Idx = 0;
+  for (size_t s = 0, se = Sequences.size(); s != se; ++s) {
+    OS << "    ";
+    const SmallVectorImpl<MaskRolPair> &Sequence = Sequences[s];
+    for (size_t p = 0, pe = Sequence.size(); p != pe; ++p) {
+      const MaskRolPair &P = Sequence[p];
+      OS << format("{ 0x%08X, %2u }, ", P.Mask, P.RotateLeft);
+    }
+    OS << "{ 0, 0 }";
+    if (s+1 != se)
+      OS << ", ";
+    OS << "  // Sequence " << Idx << "\n";
+    Idx += Sequence.size() + 1;
+  }
+  OS << "  };\n"
+        "  static const MaskRolOp *CompositeSequences[] = {\n";
+  for (size_t i = 0, e = SubRegIndices.size(); i != e; ++i) {
+    OS << "    ";
+    unsigned Idx = SubReg2SequenceIndexMap[i];
+    OS << format("&Seqs[%u]", Idx);
+    if (i+1 != e)
+      OS << ",";
+    OS << " // to " << SubRegIndices[i].getName() << "\n";
+  }
+  OS << "  };\n\n";
+
+  OS << "  --IdxA; assert(IdxA < " << SubRegIndices.size()
+     << " && \"Subregister index out of bounds\");\n"
+        "  unsigned Result = 0;\n"
+        "  for (const MaskRolOp *Ops = CompositeSequences[IdxA]; Ops->Mask != 0; ++Ops)"
+        " {\n"
+        "    unsigned Masked = LaneMask & Ops->Mask;\n"
+        "    Result |= (Masked << Ops->RotateLeft) & 0xFFFFFFFF;\n"
+        "    Result |= (Masked >> ((32 - Ops->RotateLeft) & 0x1F));\n"
+        "  }\n"
+        "  return Result;\n"
+        "}\n";
+}
+
 //
 // runMCDesc - Print out MC register descriptions.
 //
@@ -966,6 +1057,8 @@ RegisterInfoEmitter::runTargetHeader(raw_ostream &OS, CodeGenTarget &Target,
      << "     { return false; }\n";
   if (!RegBank.getSubRegIndices().empty()) {
     OS << "  unsigned composeSubRegIndicesImpl"
+       << "(unsigned, unsigned) const override;\n"
+       << "  unsigned composeSubRegIndexLaneMaskImpl"
        << "(unsigned, unsigned) const override;\n"
        << "  const TargetRegisterClass *getSubClassWithSubReg"
        << "(const TargetRegisterClass*, unsigned) const override;\n";
@@ -1214,8 +1307,10 @@ RegisterInfoEmitter::runTargetDesc(raw_ostream &OS, CodeGenTarget &Target,
   auto SubRegIndicesSize =
       std::distance(SubRegIndices.begin(), SubRegIndices.end());
 
-  if (!SubRegIndices.empty())
+  if (!SubRegIndices.empty()) {
     emitComposeSubRegIndices(OS, RegBank, ClassName);
+    emitComposeSubRegIndexLaneMask(OS, RegBank, ClassName);
+  }
 
   // Emit getSubClassWithSubReg.
   if (!SubRegIndices.empty()) {