[RISCV] Add 16 bit GPR sub-register for Zhinx. (#107446)

This patches adds a 16 bit register class for use with Zhinx
instructions. This makes them more similar to Zfh instructions and
allows us to only spill 16 bits.

I've added CodeGenOnly instructions for load/store using GPRF16 as that
gave better results than insert_subreg/extract_subreg. I'm using FSGNJ
for GPRF16 copy with Zhinx as that gave better results. Zhinxmin will
use ADDI+subreg operations.

Function arguments use this new GPRF16 register class for f16 arguments
with Zhinxmin. Eliminating the need to use RISCVISD::FMV* nodes.

I plan to extend this idea to Zfinx next.

llvm/lib/Target/RISCV/AsmParser/RISCVAsmParser.cpp

@@ -480,7 +480,13 @@ struct RISCVOperand final : public MCParsedAsmOperand {
            RISCVMCRegisterClasses[RISCV::GPRRegClassID].contains(Reg.RegNum);
   }
 
+  bool isGPRF16() const {
+    return Kind == KindTy::Register &&
+           RISCVMCRegisterClasses[RISCV::GPRF16RegClassID].contains(Reg.RegNum);
+  }
+
   bool isGPRAsFPR() const { return isGPR() && Reg.IsGPRAsFPR; }
+  bool isGPRAsFPR16() const { return isGPRF16() && Reg.IsGPRAsFPR; }
   bool isGPRPairAsFPR() const { return isGPRPair() && Reg.IsGPRAsFPR; }
 
   bool isGPRPair() const {
@@ -1342,6 +1348,10 @@ unsigned RISCVAsmParser::validateTargetOperandClass(MCParsedAsmOperand &AsmOp,
     Op.Reg.RegNum = convertFPR64ToFPR16(Reg);
     return Match_Success;
   }
+  if (Kind == MCK_GPRAsFPR16 && Op.isGPRAsFPR()) {
+    Op.Reg.RegNum = Reg - RISCV::X0 + RISCV::X0_H;
+    return Match_Success;
+  }
 
   // There are some GPRF64AsFPR instructions that have no RV32 equivalent. We
   // reject them at parsing thinking we should match as GPRPairAsFPR for RV32.

llvm/lib/Target/RISCV/Disassembler/RISCVDisassembler.cpp

@@ -81,6 +81,19 @@ static DecodeStatus DecodeGPRRegisterClass(MCInst &Inst, uint32_t RegNo,
   return MCDisassembler::Success;
 }
 
+static DecodeStatus DecodeGPRF16RegisterClass(MCInst &Inst, uint32_t RegNo,
+                                              uint64_t Address,
+                                              const MCDisassembler *Decoder) {
+  bool IsRVE = Decoder->getSubtargetInfo().hasFeature(RISCV::FeatureStdExtE);
+
+  if (RegNo >= 32 || (IsRVE && RegNo >= 16))
+    return MCDisassembler::Fail;
+
+  MCRegister Reg = RISCV::X0_H + RegNo;
+  Inst.addOperand(MCOperand::createReg(Reg));
+  return MCDisassembler::Success;
+}
+
 static DecodeStatus DecodeGPRX1X5RegisterClass(MCInst &Inst, uint32_t RegNo,
                                                uint64_t Address,
                                                const MCDisassembler *Decoder) {

llvm/lib/Target/RISCV/RISCVCallingConv.cpp

@@ -139,6 +139,23 @@ ArrayRef<MCPhysReg> RISCV::getArgGPRs(const RISCVABI::ABI ABI) {
   return ArrayRef(ArgIGPRs);
 }
 
+static ArrayRef<MCPhysReg> getArgGPR16s(const RISCVABI::ABI ABI) {
+  // The GPRs used for passing arguments in the ILP32* and LP64* ABIs, except
+  // the ILP32E ABI.
+  static const MCPhysReg ArgIGPRs[] = {RISCV::X10_H, RISCV::X11_H, RISCV::X12_H,
+                                       RISCV::X13_H, RISCV::X14_H, RISCV::X15_H,
+                                       RISCV::X16_H, RISCV::X17_H};
+  // The GPRs used for passing arguments in the ILP32E/LP64E ABI.
+  static const MCPhysReg ArgEGPRs[] = {RISCV::X10_H, RISCV::X11_H,
+                                       RISCV::X12_H, RISCV::X13_H,
+                                       RISCV::X14_H, RISCV::X15_H};
+
+  if (ABI == RISCVABI::ABI_ILP32E || ABI == RISCVABI::ABI_LP64E)
+    return ArrayRef(ArgEGPRs);
+
+  return ArrayRef(ArgIGPRs);
+}
+
 static ArrayRef<MCPhysReg> getFastCCArgGPRs(const RISCVABI::ABI ABI) {
   // The GPRs used for passing arguments in the FastCC, X5 and X6 might be used
   // for save-restore libcall, so we don't use them.
@@ -157,6 +174,26 @@ static ArrayRef<MCPhysReg> getFastCCArgGPRs(const RISCVABI::ABI ABI) {
   return ArrayRef(FastCCIGPRs);
 }
 
+static ArrayRef<MCPhysReg> getFastCCArgGPRF16s(const RISCVABI::ABI ABI) {
+  // The GPRs used for passing arguments in the FastCC, X5 and X6 might be used
+  // for save-restore libcall, so we don't use them.
+  // Don't use X7 for fastcc, since Zicfilp uses X7 as the label register.
+  static const MCPhysReg FastCCIGPRs[] = {
+      RISCV::X10_H, RISCV::X11_H, RISCV::X12_H, RISCV::X13_H,
+      RISCV::X14_H, RISCV::X15_H, RISCV::X16_H, RISCV::X17_H,
+      RISCV::X28_H, RISCV::X29_H, RISCV::X30_H, RISCV::X31_H};
+
+  // The GPRs used for passing arguments in the FastCC when using ILP32E/LP64E.
+  static const MCPhysReg FastCCEGPRs[] = {RISCV::X10_H, RISCV::X11_H,
+                                          RISCV::X12_H, RISCV::X13_H,
+                                          RISCV::X14_H, RISCV::X15_H};
+
+  if (ABI == RISCVABI::ABI_ILP32E || ABI == RISCVABI::ABI_LP64E)
+    return ArrayRef(FastCCEGPRs);
+
+  return ArrayRef(FastCCIGPRs);
+}
+
 // Pass a 2*XLEN argument that has been split into two XLEN values through
 // registers or the stack as necessary.
 static bool CC_RISCVAssign2XLen(unsigned XLen, CCState &State, CCValAssign VA1,
@@ -320,6 +357,13 @@ bool llvm::CC_RISCV(unsigned ValNo, MVT ValVT, MVT LocVT,
     }
   }
 
+  if ((ValVT == MVT::f16 && Subtarget.hasStdExtZhinxmin())) {
+    if (MCRegister Reg = State.AllocateReg(getArgGPR16s(ABI))) {
+      State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
+      return false;
+    }
+  }
+
   ArrayRef<MCPhysReg> ArgGPRs = RISCV::getArgGPRs(ABI);
 
   // Zfinx/Zdinx use GPR without a bitcast when possible.
@@ -564,9 +608,16 @@ bool llvm::CC_RISCV_FastCC(unsigned ValNo, MVT ValVT, MVT LocVT,
 
   MVT XLenVT = Subtarget.getXLenVT();
 
+  // Check if there is an available GPRF16 before hitting the stack.
+  if ((LocVT == MVT::f16 && Subtarget.hasStdExtZhinxmin())) {
+    if (MCRegister Reg = State.AllocateReg(getFastCCArgGPRF16s(ABI))) {
+      State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
+      return false;
+    }
+  }
+
   // Check if there is an available GPR before hitting the stack.
-  if ((LocVT == MVT::f16 && Subtarget.hasStdExtZhinxmin()) ||
-      (LocVT == MVT::f32 && Subtarget.hasStdExtZfinx()) ||
+  if ((LocVT == MVT::f32 && Subtarget.hasStdExtZfinx()) ||
       (LocVT == MVT::f64 && Subtarget.is64Bit() &&
        Subtarget.hasStdExtZdinx())) {
     if (MCRegister Reg = State.AllocateReg(getFastCCArgGPRs(ABI))) {

llvm/lib/Target/RISCV/RISCVDeadRegisterDefinitions.cpp

@@ -93,14 +93,19 @@ bool RISCVDeadRegisterDefinitions::runOnMachineFunction(MachineFunction &MF) {
           continue;
         LLVM_DEBUG(dbgs() << "    Dead def operand #" << I << " in:\n      ";
                    MI.print(dbgs()));
+        Register X0Reg;
         const TargetRegisterClass *RC = TII->getRegClass(Desc, I, TRI, MF);
-        if (!(RC && RC->contains(RISCV::X0))) {
+        if (RC && RC->contains(RISCV::X0)) {
+          X0Reg = RISCV::X0;
+        } else if (RC && RC->contains(RISCV::X0_H)) {
+          X0Reg = RISCV::X0_H;
+        } else {
           LLVM_DEBUG(dbgs() << "    Ignoring, register is not a GPR.\n");
           continue;
         }
         assert(LIS.hasInterval(Reg));
         LIS.removeInterval(Reg);
-        MO.setReg(RISCV::X0);
+        MO.setReg(X0Reg);
         LLVM_DEBUG(dbgs() << "    Replacing with zero register. New:\n      ";
                    MI.print(dbgs()));
         ++NumDeadDefsReplaced;

llvm/lib/Target/RISCV/RISCVExpandPseudoInsts.cpp

@@ -48,6 +48,8 @@ class RISCVExpandPseudo : public MachineFunctionPass {
                   MachineBasicBlock::iterator &NextMBBI);
   bool expandVMSET_VMCLR(MachineBasicBlock &MBB,
                          MachineBasicBlock::iterator MBBI, unsigned Opcode);
+  bool expandMV_FPR16INX(MachineBasicBlock &MBB,
+                         MachineBasicBlock::iterator MBBI);
   bool expandRV32ZdinxStore(MachineBasicBlock &MBB,
                             MachineBasicBlock::iterator MBBI);
   bool expandRV32ZdinxLoad(MachineBasicBlock &MBB,
@@ -104,6 +106,8 @@ bool RISCVExpandPseudo::expandMI(MachineBasicBlock &MBB,
   // expanded instructions for each pseudo is correct in the Size field of the
   // tablegen definition for the pseudo.
   switch (MBBI->getOpcode()) {
+  case RISCV::PseudoMV_FPR16INX:
+    return expandMV_FPR16INX(MBB, MBBI);
   case RISCV::PseudoRV32ZdinxSD:
     return expandRV32ZdinxStore(MBB, MBBI);
   case RISCV::PseudoRV32ZdinxLD:
@@ -266,6 +270,23 @@ bool RISCVExpandPseudo::expandVMSET_VMCLR(MachineBasicBlock &MBB,
   return true;
 }
 
+bool RISCVExpandPseudo::expandMV_FPR16INX(MachineBasicBlock &MBB,
+                                          MachineBasicBlock::iterator MBBI) {
+  DebugLoc DL = MBBI->getDebugLoc();
+  const TargetRegisterInfo *TRI = STI->getRegisterInfo();
+  Register DstReg = TRI->getMatchingSuperReg(
+      MBBI->getOperand(0).getReg(), RISCV::sub_16, &RISCV::GPRRegClass);
+  Register SrcReg = TRI->getMatchingSuperReg(
+      MBBI->getOperand(1).getReg(), RISCV::sub_16, &RISCV::GPRRegClass);
+
+  BuildMI(MBB, MBBI, DL, TII->get(RISCV::ADDI), DstReg)
+      .addReg(SrcReg, getKillRegState(MBBI->getOperand(1).isKill()))
+      .addImm(0);
+
+  MBBI->eraseFromParent(); // The pseudo instruction is gone now.
+  return true;
+}
+
 // This function expands the PseudoRV32ZdinxSD for storing a double-precision
 // floating-point value into memory by generating an equivalent instruction
 // sequence for RV32.

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

@@ -928,7 +928,10 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
     }
 
     SDNode *Res;
-    if (Opc == RISCV::FCVT_D_W_IN32X || Opc == RISCV::FCVT_D_W)
+    if (VT.SimpleTy == MVT::f16 && Opc == RISCV::COPY) {
+      Res =
+          CurDAG->getTargetExtractSubreg(RISCV::sub_16, DL, VT, Imm).getNode();
+    } else if (Opc == RISCV::FCVT_D_W_IN32X || Opc == RISCV::FCVT_D_W)
       Res = CurDAG->getMachineNode(
           Opc, DL, VT, Imm,
           CurDAG->getTargetConstant(RISCVFPRndMode::RNE, DL, XLenVT));

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

@@ -104,6 +104,7 @@ Register RISCVInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
     MemBytes = 1;
     break;
   case RISCV::LH:
+  case RISCV::LH_INX:
   case RISCV::LHU:
   case RISCV::FLH:
     MemBytes = 2;
@@ -144,6 +145,7 @@ Register RISCVInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
     MemBytes = 1;
     break;
   case RISCV::SH:
+  case RISCV::SH_INX:
   case RISCV::FSH:
     MemBytes = 2;
     break;
@@ -462,6 +464,13 @@ void RISCVInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
     return;
   }
 
+  if (RISCV::GPRF16RegClass.contains(DstReg, SrcReg)) {
+    BuildMI(MBB, MBBI, DL, get(RISCV::PseudoMV_FPR16INX), DstReg)
+        .addReg(SrcReg,
+                getKillRegState(KillSrc) | getRenamableRegState(RenamableSrc));
+    return;
+  }
+
   if (RISCV::GPRPairRegClass.contains(DstReg, SrcReg)) {
     // Emit an ADDI for both parts of GPRPair.
     BuildMI(MBB, MBBI, DL, get(RISCV::ADDI),
@@ -583,6 +592,9 @@ void RISCVInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
     Opcode = TRI->getRegSizeInBits(RISCV::GPRRegClass) == 32 ?
              RISCV::SW : RISCV::SD;
     IsScalableVector = false;
+  } else if (RISCV::GPRF16RegClass.hasSubClassEq(RC)) {
+    Opcode = RISCV::SH_INX;
+    IsScalableVector = false;
   } else if (RISCV::GPRPairRegClass.hasSubClassEq(RC)) {
     Opcode = RISCV::PseudoRV32ZdinxSD;
     IsScalableVector = false;
@@ -666,6 +678,9 @@ void RISCVInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
     Opcode = TRI->getRegSizeInBits(RISCV::GPRRegClass) == 32 ?
              RISCV::LW : RISCV::LD;
     IsScalableVector = false;
+  } else if (RISCV::GPRF16RegClass.hasSubClassEq(RC)) {
+    Opcode = RISCV::LH_INX;
+    IsScalableVector = false;
   } else if (RISCV::GPRPairRegClass.hasSubClassEq(RC)) {
     Opcode = RISCV::PseudoRV32ZdinxLD;
     IsScalableVector = false;
@@ -1538,6 +1553,9 @@ unsigned RISCVInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
   }
 
   switch (Opcode) {
+  case RISCV::PseudoMV_FPR16INX:
+    // MV is always compressible to either c.mv or c.li rd, 0.
+    return STI.hasStdExtCOrZca() ? 2 : 4;
   case TargetOpcode::STACKMAP:
     // The upper bound for a stackmap intrinsic is the full length of its shadow
     return StackMapOpers(&MI).getNumPatchBytes();
@@ -2593,6 +2611,7 @@ bool RISCVInstrInfo::canFoldIntoAddrMode(const MachineInstr &MemI, Register Reg,
   case RISCV::LB:
   case RISCV::LBU:
   case RISCV::LH:
+  case RISCV::LH_INX:
   case RISCV::LHU:
   case RISCV::LW:
   case RISCV::LWU:
@@ -2602,6 +2621,7 @@ bool RISCVInstrInfo::canFoldIntoAddrMode(const MachineInstr &MemI, Register Reg,
   case RISCV::FLD:
   case RISCV::SB:
   case RISCV::SH:
+  case RISCV::SH_INX:
   case RISCV::SW:
   case RISCV::SD:
   case RISCV::FSH:
@@ -2665,9 +2685,11 @@ bool RISCVInstrInfo::getMemOperandsWithOffsetWidth(
   case RISCV::LBU:
   case RISCV::SB:
   case RISCV::LH:
+  case RISCV::LH_INX:
   case RISCV::LHU:
   case RISCV::FLH:
   case RISCV::SH:
+  case RISCV::SH_INX:
   case RISCV::FSH:
   case RISCV::LW:
   case RISCV::LWU:

llvm/lib/Target/RISCV/RISCVInstrInfo.td

@@ -520,8 +520,8 @@ class BranchCC_rri<bits<3> funct3, string opcodestr>
 }
 
 let hasSideEffects = 0, mayLoad = 1, mayStore = 0 in {
-class Load_ri<bits<3> funct3, string opcodestr>
-    : RVInstI<funct3, OPC_LOAD, (outs GPR:$rd), (ins GPRMem:$rs1, simm12:$imm12),
+class Load_ri<bits<3> funct3, string opcodestr, DAGOperand rty = GPR>
+    : RVInstI<funct3, OPC_LOAD, (outs rty:$rd), (ins GPRMem:$rs1, simm12:$imm12),
               opcodestr, "$rd, ${imm12}(${rs1})">;
 
 class HLoad_r<bits<7> funct7, bits<5> funct5, string opcodestr>
@@ -535,9 +535,9 @@ class HLoad_r<bits<7> funct7, bits<5> funct5, string opcodestr>
 // reflecting the order these fields are specified in the instruction
 // encoding.
 let hasSideEffects = 0, mayLoad = 0, mayStore = 1 in {
-class Store_rri<bits<3> funct3, string opcodestr>
+class Store_rri<bits<3> funct3, string opcodestr, DAGOperand rty = GPR>
     : RVInstS<funct3, OPC_STORE, (outs),
-              (ins GPR:$rs2, GPRMem:$rs1, simm12:$imm12),
+              (ins rty:$rs2, GPRMem:$rs1, simm12:$imm12),
               opcodestr, "$rs2, ${imm12}(${rs1})">;
 
 class HStore_rr<bits<7> funct7, string opcodestr>
@@ -549,8 +549,8 @@ class HStore_rr<bits<7> funct7, string opcodestr>
 }
 
 let hasSideEffects = 0, mayLoad = 0, mayStore = 0 in
-class ALU_ri<bits<3> funct3, string opcodestr>
-    : RVInstI<funct3, OPC_OP_IMM, (outs GPR:$rd), (ins GPR:$rs1, simm12:$imm12),
+class ALU_ri<bits<3> funct3, string opcodestr, DAGOperand rty = GPR>
+    : RVInstI<funct3, OPC_OP_IMM, (outs rty:$rd), (ins rty:$rs1, simm12:$imm12),
               opcodestr, "$rd, $rs1, $imm12">,
       Sched<[WriteIALU, ReadIALU]>;
 

llvm/lib/Target/RISCV/RISCVInstrInfoZc.td

@@ -112,8 +112,9 @@ class CLoadB_ri<bits<6> funct6, string OpcodeStr>
 }
 
 let hasSideEffects = 0, mayLoad = 1, mayStore = 0 in
-class CLoadH_ri<bits<6> funct6, bit funct1, string OpcodeStr>
-    : RVInst16CLH<funct6, funct1, 0b00, (outs GPRC:$rd),
+class CLoadH_ri<bits<6> funct6, bit funct1, string OpcodeStr,
+                DAGOperand rty = GPRC>
+    : RVInst16CLH<funct6, funct1, 0b00, (outs rty:$rd),
                   (ins GPRCMem:$rs1, uimm2_lsb0:$imm),
                   OpcodeStr, "$rd, ${imm}(${rs1})"> {
   bits<2> imm;
@@ -132,9 +133,10 @@ class CStoreB_rri<bits<6> funct6, string OpcodeStr>
 }
 
 let hasSideEffects = 0, mayLoad = 0, mayStore = 1 in
-class CStoreH_rri<bits<6> funct6, bit funct1, string OpcodeStr>
+class CStoreH_rri<bits<6> funct6, bit funct1, string OpcodeStr,
+                  DAGOperand rty = GPRC>
     : RVInst16CSH<funct6, funct1, 0b00, (outs),
-                  (ins GPRC:$rs2, GPRCMem:$rs1, uimm2_lsb0:$imm),
+                  (ins rty:$rs2, GPRCMem:$rs1, uimm2_lsb0:$imm),
                   OpcodeStr, "$rs2, ${imm}(${rs1})"> {
   bits<2> imm;
 
@@ -202,7 +204,15 @@ def C_SB : CStoreB_rri<0b100010, "c.sb">,
            Sched<[WriteSTB, ReadStoreData, ReadMemBase]>;
 def C_SH : CStoreH_rri<0b100011, 0b0, "c.sh">,
            Sched<[WriteSTH, ReadStoreData, ReadMemBase]>;
+
+// Compressed versions of Zhinx load/store.
+let isCodeGenOnly = 1 in {
+def C_LH_INX : CLoadH_ri<0b100001, 0b1, "c.lh", GPRF16C>,
+               Sched<[WriteLDH, ReadMemBase]>;
+def C_SH_INX : CStoreH_rri<0b100011, 0b0, "c.sh", GPRF16C>,
+               Sched<[WriteSTH, ReadStoreData, ReadMemBase]>;
 }
+} // Predicates = [HasStdExtZcb]
 
 // Zcmp
 let DecoderNamespace = "RVZcmp", Predicates = [HasStdExtZcmp],
@@ -318,6 +328,13 @@ def : CompressPat<(SB GPRC:$rs2, GPRCMem:$rs1, uimm2:$imm),
                   (C_SB GPRC:$rs2, GPRCMem:$rs1, uimm2:$imm)>;
 def : CompressPat<(SH GPRC:$rs2, GPRCMem:$rs1, uimm2_lsb0:$imm),
                   (C_SH GPRC:$rs2, GPRCMem:$rs1, uimm2_lsb0:$imm)>;
+
+let isCompressOnly = true in {
+def : CompressPat<(LH_INX GPRF16C:$rd, GPRCMem:$rs1, uimm2_lsb0:$imm),
+                  (C_LH_INX GPRF16C:$rd, GPRCMem:$rs1, uimm2_lsb0:$imm)>;
+def : CompressPat<(SH_INX GPRF16C:$rs2, GPRCMem:$rs1, uimm2_lsb0:$imm),
+                  (C_SH_INX GPRF16C:$rs2, GPRCMem:$rs1, uimm2_lsb0:$imm)>;
+}
 }// Predicates = [HasStdExtZcb]