[RISCV] Optimize the spill/reload of segment registers

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

@@ -382,7 +382,7 @@ void RISCVInstrInfo::copyPhysRegVector(
     MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
     const DebugLoc &DL, MCRegister DstReg, MCRegister SrcReg, bool KillSrc,
     const TargetRegisterClass *RegClass) const {
-  const TargetRegisterInfo *TRI = STI.getRegisterInfo();
+  const RISCVRegisterInfo *TRI = STI.getRegisterInfo();
   RISCVVType::VLMUL LMul = RISCVRI::getLMul(RegClass->TSFlags);
   unsigned NF = RISCVRI::getNF(RegClass->TSFlags);
 
@@ -444,13 +444,7 @@ void RISCVInstrInfo::copyPhysRegVector(
     return {RISCVVType::LMUL_1, RISCV::VRRegClass, RISCV::VMV1R_V,
             RISCV::PseudoVMV_V_V_M1, RISCV::PseudoVMV_V_I_M1};
   };
-  auto FindRegWithEncoding = [TRI](const TargetRegisterClass &RegClass,
-                                   uint16_t Encoding) {
-    MCRegister Reg = RISCV::V0 + Encoding;
-    if (RISCVRI::getLMul(RegClass.TSFlags) == RISCVVType::LMUL_1)
-      return Reg;
-    return TRI->getMatchingSuperReg(Reg, RISCV::sub_vrm1_0, &RegClass);
-  };
+
   while (I != NumRegs) {
     // For non-segment copying, we only do this once as the registers are always
     // aligned.
@@ -470,9 +464,9 @@ void RISCVInstrInfo::copyPhysRegVector(
 
     // Emit actual copying.
     // For reversed copying, the encoding should be decreased.
-    MCRegister ActualSrcReg = FindRegWithEncoding(
+    MCRegister ActualSrcReg = TRI->findVRegWithEncoding(
         RegClass, ReversedCopy ? (SrcEncoding - NumCopied + 1) : SrcEncoding);
-    MCRegister ActualDstReg = FindRegWithEncoding(
+    MCRegister ActualDstReg = TRI->findVRegWithEncoding(
         RegClass, ReversedCopy ? (DstEncoding - NumCopied + 1) : DstEncoding);
 
     auto MIB = BuildMI(MBB, MBBI, DL, get(Opc), ActualDstReg);

llvm/lib/Target/RISCV/RISCVRegisterInfo.cpp

@@ -389,9 +389,25 @@ void RISCVRegisterInfo::adjustReg(MachineBasicBlock &MBB,
       .setMIFlag(Flag);
 }
 
-// Split a VSPILLx_Mx pseudo into multiple whole register stores separated by
-// LMUL*VLENB bytes.
-void RISCVRegisterInfo::lowerVSPILL(MachineBasicBlock::iterator II) const {
+static std::tuple<RISCVVType::VLMUL, const TargetRegisterClass &, unsigned>
+getSpillReloadInfo(unsigned NumRemaining, uint16_t RegEncoding, bool IsSpill) {
+  if (NumRemaining >= 8 && RegEncoding % 8 == 0)
+    return {RISCVVType::LMUL_8, RISCV::VRM8RegClass,
+            IsSpill ? RISCV::VS8R_V : RISCV::VL8RE8_V};
+  if (NumRemaining >= 4 && RegEncoding % 4 == 0)
+    return {RISCVVType::LMUL_4, RISCV::VRM4RegClass,
+            IsSpill ? RISCV::VS4R_V : RISCV::VL4RE8_V};
+  if (NumRemaining >= 2 && RegEncoding % 2 == 0)
+    return {RISCVVType::LMUL_2, RISCV::VRM2RegClass,
+            IsSpill ? RISCV::VS2R_V : RISCV::VL2RE8_V};
+  return {RISCVVType::LMUL_1, RISCV::VRRegClass,
+          IsSpill ? RISCV::VS1R_V : RISCV::VL1RE8_V};
+}
+
+// Split a VSPILLx_Mx/VSPILLx_Mx pseudo into multiple whole register stores
+// separated by LMUL*VLENB bytes.
+void RISCVRegisterInfo::lowerSegmentSpillReload(MachineBasicBlock::iterator II,
+                                                bool IsSpill) const {
   DebugLoc DL = II->getDebugLoc();
   MachineBasicBlock &MBB = *II->getParent();
   MachineFunction &MF = *MBB.getParent();
@@ -403,148 +419,75 @@ void RISCVRegisterInfo::lowerVSPILL(MachineBasicBlock::iterator II) const {
   auto ZvlssegInfo = RISCV::isRVVSpillForZvlsseg(II->getOpcode());
   unsigned NF = ZvlssegInfo->first;
   unsigned LMUL = ZvlssegInfo->second;
-  assert(NF * LMUL <= 8 && "Invalid NF/LMUL combinations.");
-  unsigned Opcode, SubRegIdx;
-  switch (LMUL) {
-  default:
-    llvm_unreachable("LMUL must be 1, 2, or 4.");
-  case 1:
-    Opcode = RISCV::VS1R_V;
-    SubRegIdx = RISCV::sub_vrm1_0;
-    break;
-  case 2:
-    Opcode = RISCV::VS2R_V;
-    SubRegIdx = RISCV::sub_vrm2_0;
-    break;
-  case 4:
-    Opcode = RISCV::VS4R_V;
-    SubRegIdx = RISCV::sub_vrm4_0;
-    break;
-  }
-  static_assert(RISCV::sub_vrm1_7 == RISCV::sub_vrm1_0 + 7,
-                "Unexpected subreg numbering");
-  static_assert(RISCV::sub_vrm2_3 == RISCV::sub_vrm2_0 + 3,
-                "Unexpected subreg numbering");
-  static_assert(RISCV::sub_vrm4_1 == RISCV::sub_vrm4_0 + 1,
-                "Unexpected subreg numbering");
-
-  Register VL = MRI.createVirtualRegister(&RISCV::GPRRegClass);
-  // Optimize for constant VLEN.
-  if (auto VLEN = STI.getRealVLen()) {
-    const int64_t VLENB = *VLEN / 8;
-    int64_t Offset = VLENB * LMUL;
-    STI.getInstrInfo()->movImm(MBB, II, DL, VL, Offset);
-  } else {
-    BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENB), VL);
-    uint32_t ShiftAmount = Log2_32(LMUL);
-    if (ShiftAmount != 0)
-      BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), VL)
-          .addReg(VL)
-          .addImm(ShiftAmount);
-  }
+  unsigned NumRegs = NF * LMUL;
+  assert(NumRegs <= 8 && "Invalid NF/LMUL combinations.");
 
-  Register SrcReg = II->getOperand(0).getReg();
+  Register Reg = II->getOperand(0).getReg();
+  uint16_t RegEncoding = TRI->getEncodingValue(Reg);
   Register Base = II->getOperand(1).getReg();
   bool IsBaseKill = II->getOperand(1).isKill();
   Register NewBase = MRI.createVirtualRegister(&RISCV::GPRRegClass);
 
   auto *OldMMO = *(II->memoperands_begin());
   LocationSize OldLoc = OldMMO->getSize();
   assert(OldLoc.isPrecise() && OldLoc.getValue().isKnownMultipleOf(NF));
-  TypeSize NewSize = OldLoc.getValue().divideCoefficientBy(NF);
-  auto *NewMMO = MF.getMachineMemOperand(OldMMO, OldMMO->getOffset(), NewSize);
-  for (unsigned I = 0; I < NF; ++I) {
-    // Adding implicit-use of super register to describe we are using part of
-    // super register, that prevents machine verifier complaining when part of
-    // subreg is undef, see comment in MachineVerifier::checkLiveness for more
-    // detail.
-    BuildMI(MBB, II, DL, TII->get(Opcode))
-        .addReg(TRI->getSubReg(SrcReg, SubRegIdx + I))
-        .addReg(Base, getKillRegState(I == NF - 1))
-        .addMemOperand(NewMMO)
-        .addReg(SrcReg, RegState::Implicit);
-    if (I != NF - 1)
+  TypeSize VRegSize = OldLoc.getValue().divideCoefficientBy(NumRegs);
+
+  Register VLENB = 0;
+  unsigned PreHandledNum = 0;
+  unsigned I = 0;
+  while (I != NumRegs) {
+    auto [LMulHandled, RegClass, Opcode] =
+        getSpillReloadInfo(NumRegs - I, RegEncoding, IsSpill);
+    auto [RegNumHandled, _] = RISCVVType::decodeVLMUL(LMulHandled);
+    bool IsLast = I + RegNumHandled == NumRegs;
+    if (PreHandledNum) {
+      Register Step;
+      // Optimize for constant VLEN.
+      if (auto VLEN = STI.getRealVLen()) {
+        int64_t Offset = *VLEN / 8 * PreHandledNum;
+        Step = MRI.createVirtualRegister(&RISCV::GPRRegClass);
+        STI.getInstrInfo()->movImm(MBB, II, DL, Step, Offset);
+      } else {
+        if (!VLENB) {
+          VLENB = MRI.createVirtualRegister(&RISCV::GPRRegClass);
+          BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENB), VLENB);
+        }
+        uint32_t ShiftAmount = Log2_32(PreHandledNum);
+        if (ShiftAmount == 0)
+          Step = VLENB;
+        else {
+          Step = MRI.createVirtualRegister(&RISCV::GPRRegClass);
+          BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), Step)
+              .addReg(VLENB, getKillRegState(IsLast))
+              .addImm(ShiftAmount);
+        }
+      }
+
       BuildMI(MBB, II, DL, TII->get(RISCV::ADD), NewBase)
           .addReg(Base, getKillRegState(I != 0 || IsBaseKill))
-          .addReg(VL, getKillRegState(I == NF - 2));
-    Base = NewBase;
-  }
-  II->eraseFromParent();
-}
+          .addReg(Step, getKillRegState(Step != VLENB || IsLast));
+      Base = NewBase;
+    }
 
-// Split a VSPILLx_Mx pseudo into multiple whole register loads separated by
-// LMUL*VLENB bytes.
-void RISCVRegisterInfo::lowerVRELOAD(MachineBasicBlock::iterator II) const {
-  DebugLoc DL = II->getDebugLoc();
-  MachineBasicBlock &MBB = *II->getParent();
-  MachineFunction &MF = *MBB.getParent();
-  MachineRegisterInfo &MRI = MF.getRegInfo();
-  const RISCVSubtarget &STI = MF.getSubtarget<RISCVSubtarget>();
-  const TargetInstrInfo *TII = STI.getInstrInfo();
-  const TargetRegisterInfo *TRI = STI.getRegisterInfo();
+    MCRegister ActualReg = findVRegWithEncoding(RegClass, RegEncoding);
+    MachineInstrBuilder MIB =
+        BuildMI(MBB, II, DL, TII->get(Opcode))
+            .addReg(ActualReg, getDefRegState(!IsSpill))
+            .addReg(Base, getKillRegState(IsLast))
+            .addMemOperand(MF.getMachineMemOperand(OldMMO, OldMMO->getOffset(),
+                                                   VRegSize * RegNumHandled));
 
-  auto ZvlssegInfo = RISCV::isRVVSpillForZvlsseg(II->getOpcode());
-  unsigned NF = ZvlssegInfo->first;
-  unsigned LMUL = ZvlssegInfo->second;
-  assert(NF * LMUL <= 8 && "Invalid NF/LMUL combinations.");
-  unsigned Opcode, SubRegIdx;
-  switch (LMUL) {
-  default:
-    llvm_unreachable("LMUL must be 1, 2, or 4.");
-  case 1:
-    Opcode = RISCV::VL1RE8_V;
-    SubRegIdx = RISCV::sub_vrm1_0;
-    break;
-  case 2:
-    Opcode = RISCV::VL2RE8_V;
-    SubRegIdx = RISCV::sub_vrm2_0;
-    break;
-  case 4:
-    Opcode = RISCV::VL4RE8_V;
-    SubRegIdx = RISCV::sub_vrm4_0;
-    break;
-  }
-  static_assert(RISCV::sub_vrm1_7 == RISCV::sub_vrm1_0 + 7,
-                "Unexpected subreg numbering");
-  static_assert(RISCV::sub_vrm2_3 == RISCV::sub_vrm2_0 + 3,
-                "Unexpected subreg numbering");
-  static_assert(RISCV::sub_vrm4_1 == RISCV::sub_vrm4_0 + 1,
-                "Unexpected subreg numbering");
-
-  Register VL = MRI.createVirtualRegister(&RISCV::GPRRegClass);
-  // Optimize for constant VLEN.
-  if (auto VLEN = STI.getRealVLen()) {
-    const int64_t VLENB = *VLEN / 8;
-    int64_t Offset = VLENB * LMUL;
-    STI.getInstrInfo()->movImm(MBB, II, DL, VL, Offset);
-  } else {
-    BuildMI(MBB, II, DL, TII->get(RISCV::PseudoReadVLENB), VL);
-    uint32_t ShiftAmount = Log2_32(LMUL);
-    if (ShiftAmount != 0)
-      BuildMI(MBB, II, DL, TII->get(RISCV::SLLI), VL)
-          .addReg(VL)
-          .addImm(ShiftAmount);
-  }
+    // Adding implicit-use of super register to describe we are using part of
+    // super register, that prevents machine verifier complaining when part of
+    // subreg is undef, see comment in MachineVerifier::checkLiveness for more
+    // detail.
+    if (IsSpill)
+      MIB.addReg(Reg, RegState::Implicit);
 
-  Register DestReg = II->getOperand(0).getReg();
-  Register Base = II->getOperand(1).getReg();
-  bool IsBaseKill = II->getOperand(1).isKill();
-  Register NewBase = MRI.createVirtualRegister(&RISCV::GPRRegClass);
-  auto *OldMMO = *(II->memoperands_begin());
-  LocationSize OldLoc = OldMMO->getSize();
-  assert(OldLoc.isPrecise() && OldLoc.getValue().isKnownMultipleOf(NF));
-  TypeSize NewSize = OldLoc.getValue().divideCoefficientBy(NF);
-  auto *NewMMO = MF.getMachineMemOperand(OldMMO, OldMMO->getOffset(), NewSize);
-  for (unsigned I = 0; I < NF; ++I) {
-    BuildMI(MBB, II, DL, TII->get(Opcode),
-            TRI->getSubReg(DestReg, SubRegIdx + I))
-        .addReg(Base, getKillRegState(I == NF - 1))
-        .addMemOperand(NewMMO);
-    if (I != NF - 1)
-      BuildMI(MBB, II, DL, TII->get(RISCV::ADD), NewBase)
-          .addReg(Base, getKillRegState(I != 0 || IsBaseKill))
-          .addReg(VL, getKillRegState(I == NF - 2));
-    Base = NewBase;
+    PreHandledNum = RegNumHandled;
+    RegEncoding += RegNumHandled;
+    I += RegNumHandled;
   }
   II->eraseFromParent();
 }
@@ -635,9 +578,7 @@ bool RISCVRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
   }
 
   // Handle spill/fill of synthetic register classes for segment operations to
-  // ensure correctness in the edge case one gets spilled. There are many
-  // possible optimizations here, but given the extreme rarity of such spills,
-  // we prefer simplicity of implementation for now.
+  // ensure correctness in the edge case one gets spilled.
   switch (MI.getOpcode()) {
   case RISCV::PseudoVSPILL2_M1:
   case RISCV::PseudoVSPILL2_M2:
@@ -650,7 +591,7 @@ bool RISCVRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
   case RISCV::PseudoVSPILL6_M1:
   case RISCV::PseudoVSPILL7_M1:
   case RISCV::PseudoVSPILL8_M1:
-    lowerVSPILL(II);
+    lowerSegmentSpillReload(II, /*IsSpill=*/true);
     return true;
   case RISCV::PseudoVRELOAD2_M1:
   case RISCV::PseudoVRELOAD2_M2:
@@ -663,7 +604,7 @@ bool RISCVRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
   case RISCV::PseudoVRELOAD6_M1:
   case RISCV::PseudoVRELOAD7_M1:
   case RISCV::PseudoVRELOAD8_M1:
-    lowerVRELOAD(II);
+    lowerSegmentSpillReload(II, /*IsSpill=*/false);
     return true;
   }
 
@@ -1052,3 +993,12 @@ bool RISCVRegisterInfo::getRegAllocationHints(
 
   return BaseImplRetVal;
 }
+
+Register
+RISCVRegisterInfo::findVRegWithEncoding(const TargetRegisterClass &RegClass,
+                                        uint16_t Encoding) const {
+  MCRegister Reg = RISCV::V0 + Encoding;
+  if (RISCVRI::getLMul(RegClass.TSFlags) == RISCVVType::LMUL_1)
+    return Reg;
+  return getMatchingSuperReg(Reg, RISCV::sub_vrm1_0, &RegClass);
+}

llvm/lib/Target/RISCV/RISCVRegisterInfo.h

@@ -107,8 +107,8 @@ struct RISCVRegisterInfo : public RISCVGenRegisterInfo {
   int64_t getFrameIndexInstrOffset(const MachineInstr *MI,
                                    int Idx) const override;
 
-  void lowerVSPILL(MachineBasicBlock::iterator II) const;
-  void lowerVRELOAD(MachineBasicBlock::iterator II) const;
+  void lowerSegmentSpillReload(MachineBasicBlock::iterator II,
+                               bool IsSpill) const;
 
   Register getFrameRegister(const MachineFunction &MF) const override;
 
@@ -144,6 +144,9 @@ struct RISCVRegisterInfo : public RISCVGenRegisterInfo {
                              const MachineFunction &MF, const VirtRegMap *VRM,
                              const LiveRegMatrix *Matrix) const override;
 
+  Register findVRegWithEncoding(const TargetRegisterClass &RegClass,
+                                uint16_t Encoding) const;
+
   static bool isVRRegClass(const TargetRegisterClass *RC) {
     return RISCVRI::isVRegClass(RC->TSFlags) &&
            RISCVRI::getNF(RC->TSFlags) == 1;

llvm/test/CodeGen/RISCV/early-clobber-tied-def-subreg-liveness.ll

@@ -40,15 +40,7 @@ define void @_Z3foov() {
 ; CHECK-NEXT:    addi a0, a0, %lo(.L__const._Z3foov.var_45)
 ; CHECK-NEXT:    vle16.v v12, (a0)
 ; CHECK-NEXT:    addi a0, sp, 16
-; CHECK-NEXT:    csrr a1, vlenb
-; CHECK-NEXT:    slli a1, a1, 1
-; CHECK-NEXT:    vs2r.v v8, (a0) # vscale x 16-byte Folded Spill
-; CHECK-NEXT:    add a0, a0, a1
-; CHECK-NEXT:    vs2r.v v10, (a0) # vscale x 16-byte Folded Spill
-; CHECK-NEXT:    add a0, a0, a1
-; CHECK-NEXT:    vs2r.v v12, (a0) # vscale x 16-byte Folded Spill
-; CHECK-NEXT:    add a0, a0, a1
-; CHECK-NEXT:    vs2r.v v14, (a0) # vscale x 16-byte Folded Spill
+; CHECK-NEXT:    vs8r.v v8, (a0) # vscale x 64-byte Folded Spill
 ; CHECK-NEXT:    lui a0, %hi(.L__const._Z3foov.var_40)
 ; CHECK-NEXT:    addi a0, a0, %lo(.L__const._Z3foov.var_40)
 ; CHECK-NEXT:    #APP
@@ -59,15 +51,7 @@ define void @_Z3foov() {
 ; CHECK-NEXT:    addi a0, a0, 928
 ; CHECK-NEXT:    vmsbc.vx v0, v8, a0
 ; CHECK-NEXT:    addi a0, sp, 16
-; CHECK-NEXT:    csrr a1, vlenb
-; CHECK-NEXT:    slli a1, a1, 1
-; CHECK-NEXT:    vl2r.v v8, (a0) # vscale x 16-byte Folded Reload
-; CHECK-NEXT:    add a0, a0, a1
-; CHECK-NEXT:    vl2r.v v10, (a0) # vscale x 16-byte Folded Reload
-; CHECK-NEXT:    add a0, a0, a1
-; CHECK-NEXT:    vl2r.v v12, (a0) # vscale x 16-byte Folded Reload
-; CHECK-NEXT:    add a0, a0, a1
-; CHECK-NEXT:    vl2r.v v14, (a0) # vscale x 16-byte Folded Reload
+; CHECK-NEXT:    vl8r.v v8, (a0) # vscale x 64-byte Folded Reload
 ; CHECK-NEXT:    csrr a0, vlenb
 ; CHECK-NEXT:    slli a0, a0, 3
 ; CHECK-NEXT:    add a0, sp, a0