[mips][msa] Implement f16 support

[sdardis]

Summary

The MIPS MSA ASE provides instructions to convert to and from half precision
floating point. This patch teaches the MIPS backend to treat f16 as a legal
type and how to promote such values to f32 for the usual set of operations.

As a result of this, the fexup[lr].w intrinsics no longer crash LLVM during
type legalization.

[zoran.jovanovic]

lib/Target/Mips/MipsMSAInstrInfo.td

3747

Nit: Next two lines exceed 80 characters.

3752

Nit: Next two lines exceed 80 characters.

3757

Nit: Next two lines exceed 80 characters.

3762

Nit: Next two lines exceed 80 characters.

lib/Target/Mips/MipsSEISelLowering.cpp

3448

Nit: I do believe that Mips::GPR64RegClass and 64-bit variants of instructions should be used with N32 too.
Same is for code in MipsSETargetLowering::emitLD_F16_PSEUDO.

3655

Nit: Is this comment correct?
From the code below (and corresponding test cases) it seems that fexupr.d instruction should be generated after this one.

3689

if comment above is correct IsFGR64 should be replaced with IsFGR64onMips64.

[zoran.jovanovic]

LGTM otherwise.

[sdardis]

As I mention in my inline comments, I'm revising this patch. I'll highlight the differences when I repost.

lib/Target/Mips/MipsSEISelLowering.cpp

3448

Looking at this, I have found test cases where N32 requires either GPR32RegClass or GPR64RegClass depending on the context. I'll post a revised version of this patch soon. I've also found 1 or 2 places where the wrong MSA regclass is used.

3655

Yes, you're correct, I missed that when writing the comment. Probably a copy/paste thing.

3689

(For completeness' sake) The comment above is wrong, If we're expanding an f16 to FPR64, we need FEXUPRD as well. The case of Mips32/Mips64 doesn't apply here.

[sdardis]

Updated the handling of loads and stores for f16 to better match the register class that is actually used. Added test coverage for that area. We're still unable to enable the machine verifier for this, as some MSA instructions are missing their 64 bit counterparts (fill, copy_[su].[bhwd] relating to this patch).

[sdardis]

Fix usage of MSA register class for f16 to FGR64 expansion.

[sdardis]

lib/Target/Mips/MipsSEISelLowering.cpp

3448

I've updated this use the register class of the operand if it's a register, otherwise use GPR64 if it's not O32. Likewise for LD_F16.

test/CodeGen/Mips/msa/f16-llvm-ir.ll

25–63

New test coverage for testing frame indexed accesses.

[sdardis]

These changes OK?

[zoran.jovanovic]

LGTM.

[sdardis]

For cross referencing purposes, the tests included in the first diff should have been part of the last diff. They have been committed in rL287574.

lib/Target/Mips/MipsMSAInstrInfo.td

def SZ_H_PSEUDO : MSA_CBRANCH_PSEUDO_DESC_BASE<MipsVAllZero, v8i16,
                                                MSA128H, NoItinerary>;
 def SZ_W_PSEUDO : MSA_CBRANCH_PSEUDO_DESC_BASE<MipsVAllZero, v4i32,
                                                MSA128W, NoItinerary>;
 def SZ_D_PSEUDO : MSA_CBRANCH_PSEUDO_DESC_BASE<MipsVAllZero, v2i64,
                                                MSA128D, NoItinerary>;
 def SZ_V_PSEUDO : MSA_CBRANCH_PSEUDO_DESC_BASE<MipsVAnyZero, v16i8,
                                                MSA128B, NoItinerary>;
 
+// Pseudoes used to implement transparent fp16 support.
+
+let Predicates = [HasMSA] in {
+ def ST_F16 : MipsPseudo<(outs), (ins MSA128F16:$ws, mem_simm10:$addr),
+                          [(store (f16 MSA128F16:$ws), (addrimm10:$addr))]> {
+   let usesCustomInserter = 1;
+ }
+
+ def LD_F16 : MipsPseudo<(outs MSA128F16:$ws), (ins mem_simm10:$addr),
+                         [(set MSA128F16:$ws, (f16 (load addrimm10:$addr)))]> {
+   let usesCustomInserter = 1;
+ }
+
+ def MSA_FP_EXTEND_W_PSEUDO : MipsPseudo<(outs FGR32Opnd:$fd), (ins MSA128F16:$ws),
+                              [(set FGR32Opnd:$fd, (f32 (fpextend MSA128F16:$ws)))]> {
+  let usesCustomInserter = 1;
+ }
+
+ def MSA_FP_ROUND_W_PSEUDO : MipsPseudo<(outs MSA128F16:$wd), (ins FGR32Opnd:$fs),
+                              [(set MSA128F16:$wd, (f16 (fpround FGR32Opnd:$fs)))]> {
+  let usesCustomInserter = 1;
+ }
+
+ def MSA_FP_EXTEND_D_PSEUDO : MipsPseudo<(outs FGR64Opnd:$fd), (ins MSA128F16:$ws),
+                              [(set FGR64Opnd:$fd, (f64 (fpextend MSA128F16:$ws)))]> {
+  let usesCustomInserter = 1;
+ }
+
+ def MSA_FP_ROUND_D_PSEUDO : MipsPseudo<(outs MSA128F16:$wd), (ins FGR64Opnd:$fs),
+                              [(set MSA128F16:$wd, (f16 (fpround FGR64Opnd:$fs)))]> {
+  let usesCustomInserter = 1;
+ }
+
+ def : MipsPat<(MipsTruncIntFP MSA128F16:$ws),
+               (TRUNC_W_D64 (MSA_FP_EXTEND_D_PSEUDO MSA128F16:$ws))>;
+
+ def : MipsPat<(MipsFPCmp MSA128F16:$ws, MSA128F16:$wt, imm:$cond),
+               (FCMP_S32 (MSA_FP_EXTEND_W_PSEUDO MSA128F16:$ws),
+                         (MSA_FP_EXTEND_W_PSEUDO MSA128F16:$wt), imm:$cond)>,
+       ISA_MIPS1_NOT_32R6_64R6;
+}
+
 // Vector extraction with fixed index.
 //
 // Extracting 32-bit values on MSA32 should always use COPY_S_W rather than
 // COPY_U_W, even for the zero-extended case. This is because our forward
 // compatibility strategy is to consider registers to be infinitely
 // sign-extended so that a MIPS64 can execute MIPS32 code without getting
 // different register values.
 def : MSAPat<(vextract_zext_i32 (v4i32 MSA128W:$ws), immZExt2Ptr:$idx),

lib/Target/Mips/MipsRegisterInfo.td

 // FP condition code registers.
 def FCC : RegisterClass<"Mips", [i32], 32, (sequence "FCC%u", 0, 7)>,
           Unallocatable;
 
 // MIPS32r6/MIPS64r6 store FPU condition codes in normal FGR registers.
 // This class allows us to represent this in codegen patterns.
 def FGRCC : RegisterClass<"Mips", [i32], 32, (sequence "F%u", 0, 31)>;
 
+def MSA128F16 : RegisterClass<"Mips", [f16], 128, (sequence "W%u", 0, 31)>;
+
 def MSA128B: RegisterClass<"Mips", [v16i8], 128,
                            (sequence "W%u", 0, 31)>;
 def MSA128H: RegisterClass<"Mips", [v8i16, v8f16], 128,
                            (sequence "W%u", 0, 31)>;
 def MSA128W: RegisterClass<"Mips", [v4i32, v4f32], 128,
                            (sequence "W%u", 0, 31)>;
 def MSA128D: RegisterClass<"Mips", [v2i64, v2f64], 128,
                            (sequence "W%u", 0, 31)>;

 def COP2Opnd : RegisterOperand<COP2> {
   let ParserMatchClass = COP2AsmOperand;
 }
 
 def COP3Opnd : RegisterOperand<COP3> {
   let ParserMatchClass = COP3AsmOperand;
 }
 
+def MSA128F16Opnd : RegisterOperand<MSA128F16> {
+  let ParserMatchClass = MSA128AsmOperand;
+}
+
 def MSA128BOpnd : RegisterOperand<MSA128B> {
   let ParserMatchClass = MSA128AsmOperand;
 }
 
 def MSA128HOpnd : RegisterOperand<MSA128H> {
   let ParserMatchClass = MSA128AsmOperand;
 }
 

lib/Target/Mips/MipsSEISelLowering.h

private:
     MachineBasicBlock *emitFILL_FD(MachineInstr &MI,
                                    MachineBasicBlock *BB) const;
     /// \brief Emit the FEXP2_W_1 pseudo instructions.
     MachineBasicBlock *emitFEXP2_W_1(MachineInstr &MI,
                                      MachineBasicBlock *BB) const;
     /// \brief Emit the FEXP2_D_1 pseudo instructions.
     MachineBasicBlock *emitFEXP2_D_1(MachineInstr &MI,
                                      MachineBasicBlock *BB) const;
+    /// \brief Emit the FILL_FW pseudo instruction
+    MachineBasicBlock *emitLD_F16_PSEUDO(MachineInstr &MI,
+                                   MachineBasicBlock *BB) const;
+    /// \brief Emit the FILL_FD pseudo instruction
+    MachineBasicBlock *emitST_F16_PSEUDO(MachineInstr &MI,
+                                   MachineBasicBlock *BB) const;
+    /// \brief Emit the FEXP2_W_1 pseudo instructions.
+    MachineBasicBlock *emitFPEXTEND_PSEUDO(MachineInstr &MI,
+                                           MachineBasicBlock *BB,
+                                           bool IsFGR64) const;
+    /// \brief Emit the FEXP2_D_1 pseudo instructions.
+    MachineBasicBlock *emitFPROUND_PSEUDO(MachineInstr &MI,
+                                          MachineBasicBlock *BBi,
+                                          bool IsFGR64) const;
   };
 }
 
 #endif

lib/Target/Mips/MipsSEISelLowering.cpp

if (Subtarget.hasMSA()) {
     addMSAIntType(MVT::v16i8, &Mips::MSA128BRegClass);
     addMSAIntType(MVT::v8i16, &Mips::MSA128HRegClass);
     addMSAIntType(MVT::v4i32, &Mips::MSA128WRegClass);
     addMSAIntType(MVT::v2i64, &Mips::MSA128DRegClass);
     addMSAFloatType(MVT::v8f16, &Mips::MSA128HRegClass);
     addMSAFloatType(MVT::v4f32, &Mips::MSA128WRegClass);
     addMSAFloatType(MVT::v2f64, &Mips::MSA128DRegClass);
 
+    // f16 is a storage-only type, always promote it to f32.
+    addRegisterClass(MVT::f16, &Mips::MSA128HRegClass);
+    setOperationAction(ISD::SETCC, MVT::f16, Promote);
+    setOperationAction(ISD::BR_CC, MVT::f16, Promote);
+    setOperationAction(ISD::SELECT_CC, MVT::f16, Promote);
+    setOperationAction(ISD::SELECT, MVT::f16, Promote);
+    setOperationAction(ISD::FADD, MVT::f16, Promote);
+    setOperationAction(ISD::FSUB, MVT::f16, Promote);
+    setOperationAction(ISD::FMUL, MVT::f16, Promote);
+    setOperationAction(ISD::FDIV, MVT::f16, Promote);
+    setOperationAction(ISD::FREM, MVT::f16, Promote);
+    setOperationAction(ISD::FMA, MVT::f16, Promote);
+    setOperationAction(ISD::FNEG, MVT::f16, Promote);
+    setOperationAction(ISD::FABS, MVT::f16, Promote);
+    setOperationAction(ISD::FCEIL, MVT::f16, Promote);
+    setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
+    setOperationAction(ISD::FCOS, MVT::f16, Promote);
+    setOperationAction(ISD::FP_EXTEND, MVT::f16, Promote);
+    setOperationAction(ISD::FFLOOR, MVT::f16, Promote);
+    setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote);
+    setOperationAction(ISD::FPOW, MVT::f16, Promote);
+    setOperationAction(ISD::FPOWI, MVT::f16, Promote);
+    setOperationAction(ISD::FRINT, MVT::f16, Promote);
+    setOperationAction(ISD::FSIN, MVT::f16, Promote);
+    setOperationAction(ISD::FSINCOS, MVT::f16, Promote);
+    setOperationAction(ISD::FSQRT, MVT::f16, Promote);
+    setOperationAction(ISD::FEXP, MVT::f16, Promote);
+    setOperationAction(ISD::FEXP2, MVT::f16, Promote);
+    setOperationAction(ISD::FLOG, MVT::f16, Promote);
+    setOperationAction(ISD::FLOG2, MVT::f16, Promote);
+    setOperationAction(ISD::FLOG10, MVT::f16, Promote);
+    setOperationAction(ISD::FROUND, MVT::f16, Promote);
+    setOperationAction(ISD::FTRUNC, MVT::f16, Promote);
+    setOperationAction(ISD::FMINNUM, MVT::f16, Promote);
+    setOperationAction(ISD::FMAXNUM, MVT::f16, Promote);
+    setOperationAction(ISD::FMINNAN, MVT::f16, Promote);
+    setOperationAction(ISD::FMAXNAN, MVT::f16, Promote);
+
     setTargetDAGCombine(ISD::AND);
     setTargetDAGCombine(ISD::OR);
     setTargetDAGCombine(ISD::SRA);
     setTargetDAGCombine(ISD::VSELECT);
     setTargetDAGCombine(ISD::XOR);
   }
 
   if (!Subtarget.useSoftFloat()) {
MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
   case Mips::FILL_FW_PSEUDO:
     return emitFILL_FW(MI, BB);
   case Mips::FILL_FD_PSEUDO:
     return emitFILL_FD(MI, BB);
   case Mips::FEXP2_W_1_PSEUDO:
     return emitFEXP2_W_1(MI, BB);
   case Mips::FEXP2_D_1_PSEUDO:
     return emitFEXP2_D_1(MI, BB);
+  case Mips::ST_F16:
+    return emitST_F16_PSEUDO(MI, BB);
+  case Mips::LD_F16:
+    return emitLD_F16_PSEUDO(MI, BB);
+  case Mips::MSA_FP_EXTEND_W_PSEUDO:
+    return emitFPEXTEND_PSEUDO(MI, BB, false);
+  case Mips::MSA_FP_ROUND_W_PSEUDO:
+    return emitFPROUND_PSEUDO(MI, BB, false);
+  case Mips::MSA_FP_EXTEND_D_PSEUDO:
+    return emitFPEXTEND_PSEUDO(MI, BB, true);
+  case Mips::MSA_FP_ROUND_D_PSEUDO:
+    return emitFPROUND_PSEUDO(MI, BB, true);
   }
 }
 
 bool MipsSETargetLowering::isEligibleForTailCallOptimization(
     const CCState &CCInfo, unsigned NextStackOffset,
     const MipsFunctionInfo &FI) const {
   if (!UseMipsTailCalls)
     return false;
BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_SUBREG), Wt2)
       .addReg(Fs)
       .addImm(Mips::sub_64);
   BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wd).addReg(Wt2).addImm(0);
 
   MI.eraseFromParent(); // The pseudo instruction is gone now.
   return BB;
 }
 
+// Emit the ST_F16_PSEDUO instruction to store a f16 value from an MSA
+// register.
+//
+// STF16 MSA128F16:$wd, mem_simm10:$addr
+// =>
+//  copy_u.h $rtemp,$wd[0]
+//  sh $rtemp, $addr
+//
+// Safety: We can't use st.h & co as they would over write the memory after
+// the destination. It would require half floats be allocated 16 bytes(!) of
+// space.
+MachineBasicBlock *
+MipsSETargetLowering::emitST_F16_PSEUDO(MachineInstr &MI,
+                                       MachineBasicBlock *BB) const {
+
+  const TargetInstrInfo *TII = Subtarget.getInstrInfo();
+  MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
+  DebugLoc DL = MI.getDebugLoc();
+  unsigned Ws = MI.getOperand(0).getReg();
+  unsigned Rt = MI.getOperand(1).getReg();
+  const MachineMemOperand &MMO = **MI.memoperands_begin();
+  unsigned Imm = MMO.getOffset();
+
+  const bool IsN64 = Subtarget.isABI_N64();
+  const TargetRegisterClass *RC =
+      IsN64 ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
+  unsigned Rs = RegInfo.createVirtualRegister(RC);
+
+  BuildMI(*BB, MI, DL, TII->get(Mips::COPY_U_H), Rs).addReg(Ws).addImm(0);
+  BuildMI(*BB, MI, DL, TII->get(IsN64 ? Mips::SH64 : Mips::SH))
+      .addReg(Rs)
+      .addReg(Rt)
+      .addImm(Imm)
+      .addMemOperand(BB->getParent()->getMachineMemOperand(
+          &MMO, MMO.getOffset(), MMO.getSize()));
+
+  MI.eraseFromParent();
+  return BB;
+}
+
+// Emit the LD_F16_PSEDUO instruction to load a f16 value into an MSA register.
+//
+// LD_F16 MSA128F16:$wd, mem_simm10:$addr
+// =>
+//  lh $rtemp, $addr
+//  fill.h $wd, $rtemp
+//
+// Safety: We can't use ld.h & co as they over-read from the source.
+// Additionally, if the address is not modulo 16, 2 cases can occur:
+//  a) Segmentation fault as the load instruction reads from a memory page
+//     memory it's not supposed to.
+//  b) The load crosses an implementation specific boundary, requiring OS
+//     intervention.
+//
+MachineBasicBlock *
+MipsSETargetLowering::emitLD_F16_PSEUDO(MachineInstr &MI,
+                                       MachineBasicBlock *BB) const {
+
+  const TargetInstrInfo *TII = Subtarget.getInstrInfo();
+  MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
+  DebugLoc DL = MI.getDebugLoc();
+  unsigned Wd = MI.getOperand(0).getReg();
+  unsigned Rs = MI.getOperand(1).getReg();
+  unsigned Imm = MI.getOperand(2).getImm();
+  const MachineMemOperand &MMO = **MI.memoperands_begin();
+
+  const bool IsN64 = Subtarget.isABI_N64();
+  const TargetRegisterClass *RC =
+      IsN64 ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
+
+  unsigned Rt = RegInfo.createVirtualRegister(RC);
+
+  BuildMI(*BB, MI, DL, TII->get(IsN64 ? Mips::LH64 : Mips::LH), Rt)
+      .addReg(Rs)
+      .addImm(Imm)
+      .addMemOperand(BB->getParent()->getMachineMemOperand(
+          &MMO, MMO.getOffset(), MMO.getSize()));
+
+  BuildMI(*BB, MI, DL, TII->get(Mips::FILL_H), Wd).addReg(Rt);
+
+  MI.eraseFromParent();
+  return BB;
+}
+
+// Emit the FPROUND_PSEUDO instruction.
+//
+// Round an FGR64Opnd, FGR32Opnd to an f16.
+//
+// Safety: Cycle the operand through the GPRs so the result always ends up
+//         the correct MSA register.
+//
+// FIXME: This copying is strictly unnecessary. If we could tie FGR32Opnd:$Fs
+//        / FGR64Opnd:$Fs and MSA128F16:$Wd to the same physical register
+//        (which they can be, as the MSA registers are defined to alias the
+//        FPU's 64 bit and 32 bit registers) the result can be accessed using
+//        the correct register class. That requires operands be tie-able across
+//        register classes which have a sub/super register class relationship.
+//
+// For FPG32Opnd:
+//
+// FPROUND MSA128F16:$wd, FGR32Opnd:$fs
+// =>
+//  mfc1 $rtemp, $fs
+//  fill.w $rtemp, $wtemp
+//  fexdo.w $wd, $wtemp, $wtemp
+//
+// For FPG64Opnd on mips32r2+:
+//
+// FPROUND MSA128F16:$wd, FGR64Opnd:$fs
+// =>
+//  mfc1 $rtemp, $fs
+//  fill.w $rtemp, $wtemp
+//  mfhc1 $rtemp2, $fs
+//  insert.w $wtemp[1], $rtemp2
+//  insert.w $wtemp[3], $rtemp2
+//  fexdo.w $wtemp2, $wtemp, $wtemp
+//  fexdo.h $wd, $temp2, $temp2
+//
+// For FGR64Opnd on mips64r2+:
+//
+// FPROUND MSA128F16:$wd, FGR64Opnd:$fs
+// =>
+//  dmfc1 $rtemp, $fs
+//  fill.d $rtemp, $wtemp
+//  fexdo.w $wtemp2, $wtemp, $wtemp
+//  fexdo.h $wd, $wtemp2, $wtemp2
+//
+// Safety note: As $wtemp is UNDEF, we may provoke a spurious exception if the
+//              undef bits are "just right" and the exception enable bits are
+//              set. By using fill.w to replicate $fs into all elements over
+//              insert.w for one element, we avoid that potiential case. If
+//              fexdo.[hw] causes an exception in, the exception is valid and it
+//              occurs for all elements.
+//
+MachineBasicBlock *
+MipsSETargetLowering::emitFPROUND_PSEUDO(MachineInstr &MI,
+                                         MachineBasicBlock *BB,
+                                         bool IsFGR64) const {
+
+  // Strictly speaking, we need MIPS32R5 to support MSA. We'll be generous
+  // here. It's technically doable to support MIPS32 here, but the ISA forbids
+  // it.
+  assert(Subtarget.hasMSA() && Subtarget.hasMips32r2());
+
+  bool IsFGR64onMips64 = Subtarget.hasMips64() && IsFGR64;
+
+  const TargetInstrInfo *TII = Subtarget.getInstrInfo();
+  DebugLoc DL = MI.getDebugLoc();
+  unsigned Wd = MI.getOperand(0).getReg();
+  unsigned Fs = MI.getOperand(1).getReg();
+
+  MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
+  unsigned Wtemp = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+  const TargetRegisterClass *GPRRC =
+      IsFGR64onMips64 ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
+  unsigned MFC1Opc = IsFGR64onMips64 ? Mips::DMFC1 : Mips::MFC1;
+  unsigned FILLOpc = IsFGR64onMips64 ? Mips::FILL_D : Mips::FILL_W;
+
+  // Perform the register class copy as mentioned above.
+  unsigned Rtemp = RegInfo.createVirtualRegister(GPRRC);
+  BuildMI(*BB, MI, DL, TII->get(MFC1Opc), Rtemp).addReg(Fs);
+  BuildMI(*BB, MI, DL, TII->get(FILLOpc), Wtemp).addReg(Rtemp);
+  unsigned WPHI = Wtemp;
+
+  if (!Subtarget.hasMips64() && IsFGR64) {
+    unsigned Rtemp2 = RegInfo.createVirtualRegister(GPRRC);
+    BuildMI(*BB, MI, DL, TII->get(Mips::MFHC1_D64), Rtemp2).addReg(Fs);
+    unsigned Wtemp2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+    unsigned Wtemp3 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+    BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_W), Wtemp2)
+        .addReg(Wtemp)
+        .addReg(Rtemp2)
+        .addImm(1);
+    BuildMI(*BB, MI, DL, TII->get(Mips::INSERT_W), Wtemp3)
+        .addReg(Wtemp2)
+        .addReg(Rtemp2)
+        .addImm(3);
+    WPHI = Wtemp3;
+  }
+
+  if (IsFGR64) {
+    unsigned Wtemp2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+    BuildMI(*BB, MI, DL, TII->get(Mips::FEXDO_W), Wtemp2)
+        .addReg(WPHI)
+        .addReg(WPHI);
+    WPHI = Wtemp2;
+  }
+
+  BuildMI(*BB, MI, DL, TII->get(Mips::FEXDO_H), Wd).addReg(WPHI).addReg(WPHI);
+
+  MI.eraseFromParent();
+  return BB;
+}
+
+// Emit the FPEXTEND_PSEUDO instruction.
+//
+// Expand an f16 to either a FGR32Opnd or FGR64Opnd.
+//
+// Safety: Cycle the result through the GPRs so the result always ends up
+//         the correct floating point register.
+//
+// FIXME: This copying is strictly unnecessary. If we could tie FGR32Opnd:$Fd
+//        / FGR64Opnd:$Fd and MSA128F16:$Ws to the same physical register
+//        (which they can be, as the MSA registers are defined to alias the
+//        FPU's 64 bit and 32 bit registers) the result can be accessed using
+//        the correct register class. That requires operands be tie-able across
+//        register classes which have a sub/super register class relationship. I
+//        haven't checked.
+//
+// For FGR32Opnd:
+//
+// FPEXTEND FGR32Opnd:$fd, MSA128F16:$ws
+// =>
+//  fexupr.w $wtemp, $ws
+//  copy_s.w $rtemp, $ws[0]
+//  mtc1 $rtemp, $fd
+//
+// For FGR64Opnd on Mips64:
+//
+// FPEXTEND FGR64Opnd:$fd, MSA128F16:$ws
+// =>
+//  fexupr.w $wtemp, $ws
+//  fexupr.d $wtemp2, $wtemp
+//  copy_s.d $rtemp, $wtemp2s[0]
+//  dmtc1 $rtemp, $fd
+//
+// For FGR64Opnd on Mips32:
+//
+// FPEXTEND FGR64Opnd:$fd, MSA128F16:$ws
+// =>
+//  fexupr.w $wtemp, $ws
+//  copy_s.w $rtemp, $wtemp[0]
+//  mtc1 $rtemp, $ftemp
+//  copy_s.w $rtemp2, $wtemp[1]
+//  $fd = mthc1 $rtemp2, $ftemp
+//
+MachineBasicBlock *
+MipsSETargetLowering::emitFPEXTEND_PSEUDO(MachineInstr &MI,
+                                          MachineBasicBlock *BB,
+                                          bool IsFGR64) const {
+
+  // Strictly speaking, we need MIPS32R5 to support MSA. We'll be generous
+  // here. It's technically doable to support MIPS32 here, but the ISA forbids
+  // it.
+  assert(Subtarget.hasMSA() && Subtarget.hasMips32r2());
+
+  bool IsFGR64onMips64 = Subtarget.hasMips64() && IsFGR64;
+  bool IsFGR64onMips32 = !Subtarget.hasMips64() && IsFGR64;
+
+  const TargetInstrInfo *TII = Subtarget.getInstrInfo();
+  DebugLoc DL = MI.getDebugLoc();
+  unsigned Fd = MI.getOperand(0).getReg();
+  unsigned Ws = MI.getOperand(1).getReg();
+
+  MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
+  const TargetRegisterClass *GPRRC =
+      IsFGR64onMips64 ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
+  unsigned MTC1Opc = IsFGR64onMips64 ? Mips::DMTC1 : Mips::MTC1;
+  unsigned COPYOpc = IsFGR64onMips64 ? Mips::COPY_S_D : Mips::COPY_S_W;
+
+  unsigned Wtemp = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+  unsigned WPHI = Wtemp;
+
+  BuildMI(*BB, MI, DL, TII->get(Mips::FEXUPR_W), Wtemp).addReg(Ws);
+  if (IsFGR64) {
+    WPHI = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+    BuildMI(*BB, MI, DL, TII->get(Mips::FEXUPR_D), WPHI).addReg(Wtemp);
+  }
+
+  // Perform the safety regclass copy mentioned above.
+  unsigned Rtemp = RegInfo.createVirtualRegister(GPRRC);
+  unsigned FPRPHI = IsFGR64onMips32
+                        ? RegInfo.createVirtualRegister(&Mips::FGR64RegClass)
+                        : Fd;
+  BuildMI(*BB, MI, DL, TII->get(COPYOpc), Rtemp).addReg(WPHI).addImm(0);
+  BuildMI(*BB, MI, DL, TII->get(MTC1Opc), FPRPHI).addReg(Rtemp);
+
+  if (IsFGR64onMips32) {
+    unsigned Rtemp2 = RegInfo.createVirtualRegister(GPRRC);
+    BuildMI(*BB, MI, DL, TII->get(Mips::COPY_S_W), Rtemp2)
+        .addReg(WPHI)
+        .addImm(1);
+    BuildMI(*BB, MI, DL, TII->get(Mips::MTHC1_D64), Fd)
+        .addReg(FPRPHI)
+        .addReg(Rtemp2);
+  }
+
+  MI.eraseFromParent();
+  return BB;
+}
+
 // Emit the FEXP2_W_1 pseudo instructions.
 //
 // fexp2_w_1_pseudo $wd, $wt
 // =>
 // ldi.w $ws, 1
 // fexp2.w $wd, $ws, $wt
 MachineBasicBlock *
 MipsSETargetLowering::emitFEXP2_W_1(MachineInstr &MI,

test/CodeGen/Mips/msa/f16-llvm-ir.ll

+; RUN: llc -relocation-model=pic -march=mipsel -mcpu=mips32r5 \
+; RUN:     -mattr=+fp64,+msa < %s | FileCheck %s \
+; RUN:     --check-prefixes=ALL,MIPS32,MIPSR5,MIPS32-O32,MIPS32R5-O32
+; RUN: llc -relocation-model=pic -march=mips64el -mcpu=mips64r5 \
+; RUN:     -mattr=+fp64,+msa -target-abi n32 < %s | FileCheck %s \
+; RUN:     --check-prefixes=ALL,MIPS64,MIPSR5,MIPS64-N32,MIPS64R5-N32
+; RUN: llc -relocation-model=pic -march=mips64el -mcpu=mips64r5 \
+; RUN:     -mattr=+fp64,+msa -target-abi n64 < %s | FileCheck %s \
+; RUN:     --check-prefixes=ALL,MIPS64,MIPSR5,MIPS64-N64,MIPS64R5-N64
+
+; RUN: llc -relocation-model=pic -march=mipsel -mcpu=mips32r6 \
+; RUN:     -mattr=+fp64,+msa < %s | FileCheck %s \
+; RUN:     --check-prefixes=ALL,MIPS32,MIPSR6,MIPSR6-O32
+; RUN: llc -relocation-model=pic -march=mips64el -mcpu=mips64r6 \
+; RUN:     -mattr=+fp64,+msa -target-abi n32 < %s | FileCheck %s \
+; RUN:     --check-prefixes=ALL,MIPS64,MIPSR6,MIPS64-N32,MIPSR6-N32
+; RUN: llc -relocation-model=pic -march=mips64el -mcpu=mips64r6 \
+; RUN:     -mattr=+fp64,+msa -target-abi n64 < %s | FileCheck %s \
+; RUN:     --check-prefixes=ALL,MIPS64,MIPSR6,MIPS64-N64,MIPSR6-N64
+
+
+@g = external global i16, align 2
+@h = external global half, align 2
+
+; Check that fext f16 to double has a fexupr.w, fexupr.d sequence.
+; Check that ftrunc double to f16 has fexdo.w, fexdo.h sequence.
+; Check that MIPS64R5+ uses 64-bit floating point <-> 64-bit GPR transfers.
+
+; We don't need to check if pre-MIPSR5 expansions occur, the MSA ASE requires
+; MIPSR5. Additionally, fp64 mode / FR=1 is required to use MSA.
+
+define void @fadd_f64() {
+entry:
+; ALL-LABEL: fadd_f64:
+  %0 = load half, half * @h, align 2
+  %1 = fpext half %0 to double
+; ALL:    lh $[[R0:[0-9]+]]
+; ALL:    fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:    fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:    fexupr.d $w[[W2:[0-9]+]], $w[[W1]]
+; MIPS32: copy_s.w $[[R1:[0-9]+]], $w[[W2]][0]
+; MIPS32: mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32: copy_s.w $[[R2:[0-9]+]], $w[[W2]][1]
+; MIPS32: mthc1 $[[R2]], $f[[F0]]
+; MIPS64: copy_s.d $[[R2:[0-9]+]], $w[[W2]][0]
+; MIPS64: dmtc1 $[[R2]], $f[[F0:[0-9]+]]
+
+  %2 = load half, half * @h, align 2
+  %3 = fpext half %2 to double
+  %add = fadd double %1, %3
+
+; ALL: add.d $f[[F1:[0-9]+]], $f[[F0]], $f[[F0]]
+
+  %4 = fptrunc double %add to half
+
+; MIPS32: mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; MIPS32: fill.w $w[[W2:[0-9]+]], $[[R2]]
+; MIPS32: mfhc1 $[[R3:[0-9]+]], $f[[F1]]
+; MIPS32: insert.w $w[[W2]][1], $[[R3]]
+; MIPS32: insert.w $w[[W2]][3], $[[R3]]
+
+; MIPS64: dmfc1 $[[R2:[0-9]+]], $f[[F1]]
+; MIPS64: fill.d $w[[W2:[0-9]+]], $[[R2]]
+
+; ALL:    fexdo.w $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:    fexdo.h $w[[W4:[0-9]+]], $w[[W3]], $w[[W3]]
+; ALL:    copy_u.h $[[R3:[0-9]+]], $w[[W4]][0]
+; ALL:    sh $[[R3]]
+   store half %4, half * @h, align 2
+  ret void
+}
+
+define i32 @ffptoui() {
+entry:
+; ALL-LABEL: ffptoui:
+  %0 = load half, half * @h, align 2
+  %1 = fptoui half %0 to i32
+
+; MIPS32:       lwc1 $f[[FC:[0-9]+]], %lo($CPI1_0)
+; MIPS64-N32:   lwc1 $f[[FC:[0-9]+]], %got_ofst(.LCPI1_0)
+; MIPS64-N64:   lwc1 $f[[FC:[0-9]+]], %got_ofst(.LCPI1_0)
+
+; ALL:          lh $[[R0:[0-9]+]]
+; ALL:          fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:          fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:          copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL:          mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPSR6:       cmp.lt.s  $f[[F1:[0-9]+]], $f[[F0]], $f[[FC]]
+; ALL:          sub.s $f[[F2:[0-9]+]], $f[[F0]], $f[[FC]]
+; ALL:          mfc1 $[[R2:[0-9]]], $f[[F2]]
+; ALL:          fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:          fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:          fexupr.w $w[[W4:[0-9]+]], $w[[W3]]
+; ALL:          fexupr.d $w[[W5:[0-9]+]], $w[[W4]]
+
+; MIPS32:       copy_s.w $[[R3:[0-9]+]], $w[[W5]][0]
+; MIPS32:       mtc1 $[[R3]], $f[[F3:[0-9]+]]
+; MIPS32:       copy_s.w $[[R4:[0-9]+]], $w[[W5]][1]
+; MIPS32:       mthc1 $[[R3]], $f[[F3]]
+
+; MIPS64:       copy_s.d $[[R2:[0-9]+]], $w[[W2]][0]
+; MIPS64:       dmtc1 $[[R2]], $f[[F3:[0-9]+]]
+
+; ALL:          trunc.w.d $f[[F4:[0-9]+]], $f[[F3]]
+; ALL:          mfc1 $[[R4:[0-9]+]], $f[[F4]]
+; ALL:          fexupr.d $w[[W6:[0-9]+]], $w[[W1]]
+
+; MIPS32:       copy_s.w $[[R5:[0-9]+]], $w[[W6]][0]
+; MIPS32:       mtc1 $[[R5]], $f[[F5:[0-9]+]]
+; MIPS32:       copy_s.w $[[R6:[0-9]+]], $w[[W6]][1]
+; MIPS32:       mthc1 $[[R6]], $f[[F5]]
+
+; MIPS64:       copy_s.d $[[R2:[0-9]+]], $w[[W2]][0]
+; MIPS64:       dmtc1 $[[R2]], $f[[F5:[0-9]+]]
+
+; ALL:          trunc.w.d $f[[F6:[0-9]]], $f[[F5]]
+; ALL:          mfc1 $[[R7:[0-9]]], $f[[F6]]
+
+; MIPS32R5-O32: lw $[[R13:[0-9]+]], %got($CPI1_1)
+; MIPS32R5-O32: addiu $[[R14:[0-9]+]], $[[R13]], %lo($CPI1_1)
+
+; MIPS64R5-N32: lw $[[R13:[0-9]+]], %got_page(.LCPI1_1)
+; MIPS64R5-N32: addiu $[[R14:[0-9]+]], $[[R13]], %got_ofst(.LCPI1_1)
+
+; MIPS64R5-N64: ld $[[R13:[0-9]+]], %got_page(.LCPI1_1)
+; MIPS64R5-N64: daddiu $[[R14:[0-9]+]], $[[R13]], %got_ofst(.LCPI1_1)
+
+; ALL:          lui $[[R8:[0-9]+]], 32768
+; ALL:          xor $[[R9:[0-9]+]], $[[R4]], $[[R8]]
+
+; MIPSR5:       lh $[[R15:[0-9]+]], 0($[[R14]])
+; MIPSR5:       fill.h $w[[W7:[0-9]+]], $[[R15]]
+; MIPSR5:       fexupr.w $w[[W8:[0-9]+]], $w[[W7]]
+; MIPSR5:       copy_s.w $[[R16:[0-9]+]], $w[[W8]][0]
+; MIPSR5:       mtc1 $[[R16]], $f[[F7:[0-9]+]]
+; MIPSR5:       c.olt.s $f[[F0]], $f[[F7]]
+; MIPSR5:       movt $[[R9]], $[[R7]], $fcc0
+
+; MIPSR6:       mfc1 $[[R10:[0-9]+]], $f[[F1]]
+; MIPSR6:       seleqz $[[R11:[0-9]]], $[[R9]], $[[R10]]
+; MIPSR6:       selnez $[[R12:[0-9]]], $[[R7]], $[[R10]]
+; MIPSR6:       or $2, $[[R12]], $[[R11]]
+
+  ret i32 %1
+}
+
+define i32 @ffptosi() {
+entry:
+; ALL-LABEL: ffptosi:
+  %0 = load half, half * @h, align 2
+  %1 = fptosi half %0 to i32
+  ret i32 %1
+
+; ALL:    lh $[[R0:[0-9]+]]
+; ALL:    fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:    fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:    fexupr.d $w[[W2:[0-9]+]], $w[[W1]]
+
+; MIPS32: copy_s.w $[[R2:[0-9]+]], $w[[W2]][0]
+; MIPS32: mtc1 $[[R2]], $f[[F0:[0-9]+]]
+; MIPS32: copy_s.w $[[R3:[0-9]+]], $w[[W2]][1]
+; MIPS32: mthc1 $[[R3]], $f[[F0]]
+
+; MIPS64: copy_s.d $[[R2:[0-9]+]], $w[[W2]][0]
+; MIPS64: dmtc1 $[[R2]], $f[[F0:[0-9]+]]
+
+; ALL:    trunc.w.d $f[[F1:[0-9]+]], $f[[F0]]
+; ALL:    mfc1 $2, $f[[F1]]
+}
+
+define void @uitofp(i32 %a) {
+entry:
+; ALL-LABEL: uitofp:
+
+; MIPS32-O32: ldc1 $f[[F0:[0-9]+]], %lo($CPI3_0)
+; MIPS32-O32: ldc1 $f[[F1:[0-9]+]], 0($sp)
+
+; MIPS64-N32: ldc1 $f[[F0:[0-9]+]], %got_ofst(.LCPI3_0)
+; MIPS64-N32: ldc1 $f[[F1:[0-9]+]], 8($sp)
+
+; MIPS64-N64: ldc1 $f[[F0:[0-9]+]], %got_ofst(.LCPI3_0)
+; MIPS64-N64: ldc1 $f[[F1:[0-9]+]], 8($sp)
+
+; MIPSR5:     sub.d $f[[F2:[0-9]+]], $f[[F1]], $f[[F0]]
+; MIPSR6-O32: sub.d $f[[F2:[0-9]+]], $f[[F0]], $f[[F1]]
+; MIPSR6-N32: sub.d $f[[F2:[0-9]+]], $f[[F1]], $f[[F0]]
+; MIPSR6-N64: sub.d $f[[F2:[0-9]+]], $f[[F1]], $f[[F0]]
+
+; MIPS32:     mfc1 $[[R0:[0-9]+]], $f[[F2]]
+; MIPS32:     fill.w $w[[W0:[0-9]+]], $[[R0]]
+; MIPS32:     mfhc1 $[[R1:[0-9]+]], $f[[F2]]
+; MIPS32:     insert.w $w[[W0]][1], $[[R1]]
+; MIPS32:     insert.w $w[[W0]][3], $[[R1]]
+
+; MIPS64-N64: ld $[[R3:[0-9]+]], %got_disp(h)
+; MIPS64-N32: lw $[[R3:[0-9]+]], %got_disp(h)
+; MIPS64:     dmfc1 $[[R1:[0-9]+]], $f[[F2]]
+; MIPS64:     fill.d $w[[W0:[0-9]+]], $[[R1]]
+
+; ALL:        fexdo.w $w[[W1:[0-9]+]], $w[[W0]], $w[[W0]]
+; ALL:        fexdo.h $w[[W2:[0-9]+]], $w[[W1]], $w[[W1]]
+
+; MIPS32:     lw $[[R3:[0-9]+]], %got(h)
+
+; ALL:        copy_u.h $[[R2:[0-9]+]], $w[[W2]]
+; ALL:        sh $[[R2]], 0($[[R3]])
+  %0 = uitofp i32 %a to half
+  store half %0, half * @h, align 2
+  ret void
+}
+
+
+; Check that f16 is expanded to f32 and relevant transfer ops occur.
+; We don't check f16 -> f64 expansion occurs, as we expand f16 to f32.
+
+define void @fadd() {
+entry:
+; ALL-LABEL: fadd:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL: lh $[[R0:[0-9]+]]
+; ALL: fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL: fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL: copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL: mtc1 $[[R1]], $f[[F0:[0-9]+]]
+
+  %2 = load i16, i16* @g, align 2
+  %3 = call float @llvm.convert.from.fp16.f32(i16 %2)
+  %add = fadd float %1, %3
+
+; ALL: add.s $f[[F1:[0-9]+]], $f[[F0]], $f[[F0]]
+
+ %4 = call i16 @llvm.convert.to.fp16.f32(float %add)
+
+; ALL: mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL: fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL: fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL: copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+; ALL: sh $[[R3]]
+   store i16 %4, i16* @g, align 2
+  ret void
+}
+
+; Function Attrs: nounwind readnone
+declare float @llvm.convert.from.fp16.f32(i16)
+
+; Function Attrs: nounwind readnone
+declare i16 @llvm.convert.to.fp16.f32(float)
+
+; Function Attrs: nounwind
+define void @fsub() {
+entry:
+; ALL-LABEL: fsub:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL: lh $[[R0:[0-9]+]]
+; ALL: fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL: fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL: copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL: mtc1 $[[R1]], $f[[F0:[0-9]+]]
+
+  %2 = load i16, i16* @g, align 2
+  %3 = call float @llvm.convert.from.fp16.f32(i16 %2)
+  %sub = fsub float %1, %3
+
+; ALL: sub.s $f[[F1:[0-9]+]], $f[[F0]], $f[[F0]]
+
+  %4 = call i16 @llvm.convert.to.fp16.f32(float %sub)
+
+; ALL: mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL: fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL: fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL: copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %4, i16* @g, align 2
+; ALL: sh $[[R3]]
+  ret void
+}
+
+define void @fmult() {
+entry:
+; ALL-LABEL: fmult:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL: lh $[[R0:[0-9]+]]
+; ALL: fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL: fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL: copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL: mtc1 $[[R1]], $f[[F0:[0-9]+]]
+
+  %2 = load i16, i16* @g, align 2
+  %3 = call float @llvm.convert.from.fp16.f32(i16 %2)
+  %mul = fmul float %1, %3
+
+; ALL: mul.s $f[[F1:[0-9]+]], $f[[F0]], $f[[F0]]
+
+  %4 = call i16 @llvm.convert.to.fp16.f32(float %mul)
+
+; ALL: mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL: fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL: fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL: copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %4, i16* @g, align 2
+
+; ALL: sh $[[R3]]
+  ret void
+}
+
+define void @fdiv() {
+entry:
+; ALL-LABEL: fdiv:
+
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL: lh $[[R0:[0-9]+]]
+; ALL: fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL: fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL: copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL: mtc1 $[[R1]], $f[[F0:[0-9]+]]
+
+  %2 = load i16, i16* @g, align 2
+  %3 = call float @llvm.convert.from.fp16.f32(i16 %2)
+  %div = fdiv float %1, %3
+
+; ALL: div.s $f[[F1:[0-9]+]], $f[[F0]], $f[[F0]]
+
+  %4 = call i16 @llvm.convert.to.fp16.f32(float %div)
+
+; ALL: mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL: fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL: fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL: copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+  store i16 %4, i16* @g, align 2
+; ALL: sh $[[R3]]
+  ret void
+}
+
+define void @frem() {
+entry:
+; ALL-LABEL: frem:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:        lh $[[R0:[0-9]+]]
+; ALL:        fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:        fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:        copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+
+  %2 = load i16, i16* @g, align 2
+  %3 = call float @llvm.convert.from.fp16.f32(i16 %2)
+  %rem = frem float %1, %3
+
+; MIPS32:     lw $25, %call16(fmodf)($gp)
+; MIPS64-N32: lw $25, %call16(fmodf)($gp)
+; MIPS64-N64: ld $25, %call16(fmodf)($gp)
+; ALL:        jalr $25
+
+  %4 = call i16 @llvm.convert.to.fp16.f32(float %rem)
+
+; ALL:        mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:        fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:        fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:        copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %4, i16* @g, align 2
+; ALL:        sh $[[R3]]
+
+  ret void
+}
+
+@i1 = external global i16, align 1
+
+define void @fcmp() {
+entry:
+; ALL-LABEL: fcmp:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+; ALL:        lh $[[R0:[0-9]+]]
+; ALL:        fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:        fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:        copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+
+  %2 = load i16, i16* @g, align 2
+  %3 = call float @llvm.convert.from.fp16.f32(i16 %2)
+  %fcmp = fcmp oeq float %1, %3
+
+; MIPSR5: addiu $[[R2:[0-9]+]], $zero, 1
+; MIPSR5: c.un.s $f[[F0]], $f[[F0]]
+; MIPSR5: movt $[[R2]], $zero, $fcc0
+; MIPSR6: cmp.un.s $f[[F1:[0-9]+]], $f[[F0]], $f[[F0]]
+; MIPSR6: mfc1 $[[R3:[0-9]]], $f[[F1]]
+; MIPSR6: not $[[R4:[0-9]+]], $[[R3]]
+; MIPSR6: andi $[[R2:[0-9]+]], $[[R4]], 1
+
+  %4 = zext i1 %fcmp to i16
+  store i16 %4, i16* @i1, align 2
+; ALL:        sh $[[R2]]
+
+  ret void
+}
+
+declare float @llvm.powi.f32(float, i32)
+
+define void @fpowi() {
+entry:
+; ALL-LABEL: fpowi:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL: lh $[[R0:[0-9]+]]
+; ALL: fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL: fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL: copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL: mtc1 $[[R1]], $f[[F0:[0-9]+]]
+
+  %powi = call float @llvm.powi.f32(float %1, i32 2)
+
+; ALL: mul.s $f[[F1:[0-9]+]], $f[[F0]], $f[[F0]]
+
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %powi)
+
+; ALL: mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL: fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL: fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL: copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL: sh $[[R3]]
+  ret void
+}
+
+define void @fpowi_var(i32 %var) {
+entry:
+; ALL-LABEL: fpowi_var:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+
+  %powi = call float @llvm.powi.f32(float %1, i32 %var)
+
+; ALL-DAG: mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(__powisf2)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(__powisf2)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(__powisf2)($gp)
+; ALL-DAG:        jalr $25
+
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %powi)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+  ret void
+}
+
+declare float @llvm.pow.f32(float %Val, float %power)
+
+define void @fpow(float %var) {
+entry:
+; ALL-LABEL: fpow:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+
+  %powi = call float @llvm.pow.f32(float %1, float %var)
+
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(powf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(powf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(powf)($gp)
+; ALL-DAG:        jalr $25
+
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %powi)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+  ret void
+}
+
+declare float @llvm.log2.f32(float %Val)
+
+define void @flog2() {
+entry:
+; ALL-LABEL: flog2:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(log2f)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(log2f)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(log2f)($gp)
+; ALL-DAG:        jalr $25
+
+  %log2 = call float @llvm.log2.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %log2)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.log10.f32(float %Val)
+
+define void @flog10() {
+entry:
+; ALL-LABEL: flog10:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(log10f)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(log10f)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(log10f)($gp)
+; ALL-DAG:        jalr $25
+
+  %log10 = call float @llvm.log10.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %log10)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.sqrt.f32(float %Val)
+
+define void @fsqrt() {
+entry:
+; ALL-LABEL: fsqrt:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL: lh $[[R0:[0-9]+]]
+; ALL: fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL: fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL: copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL: mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; ALL: sqrt.s $f[[F1:[0-9]+]], $f[[F0]]
+
+  %sqrt = call float @llvm.sqrt.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %sqrt)
+
+; ALL: mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL: fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL: fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL: copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL: sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.sin.f32(float %Val)
+
+define void @fsin() {
+entry:
+; ALL-LABEL: fsin:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(sinf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(sinf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(sinf)($gp)
+; ALL-DAG:        jalr $25
+
+  %sin = call float @llvm.sin.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %sin)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.cos.f32(float %Val)
+
+define void @fcos() {
+entry:
+; ALL-LABEL: fcos:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(cosf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(cosf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(cosf)($gp)
+; ALL-DAG:        jalr $25
+
+  %cos = call float @llvm.cos.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %cos)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.exp.f32(float %Val)
+
+define void @fexp() {
+entry:
+; ALL-LABEL: fexp:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(expf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(expf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(expf)($gp)
+; ALL-DAG:        jalr $25
+
+  %exp = call float @llvm.exp.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %exp)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.exp2.f32(float %Val)
+
+define void @fexp2() {
+entry:
+; ALL-LABEL: fexp2:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(exp2f)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(exp2f)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(exp2f)($gp)
+; ALL-DAG:        jalr $25
+
+  %exp2 = call float @llvm.exp2.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %exp2)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.fma.f32(float, float, float)
+
+define void @ffma(float %b, float %c) {
+entry:
+; ALL-LABEL: ffma:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(fmaf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(fmaf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(fmaf)($gp)
+; ALL-DAG:        jalr $25
+
+  %fma = call float @llvm.fma.f32(float %1, float %b, float %c)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %fma)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+; FIXME: For MIPSR6, this should produced the maddf.s instruction. MIPSR5 cannot
+;        fuse the operation such that the intermediate result is not rounded.
+
+declare float @llvm.fmuladd.f32(float, float, float)
+
+define void @ffmuladd(float %b, float %c) {
+entry:
+; ALL-LABEL: ffmuladd:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL:            mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-O32:     madd.s $f[[F1:[0-9]]], $f14, $f[[F0]], $f12
+; MIPS32-N32:     madd.s $f[[F1:[0-9]]], $f13, $f[[F0]], $f12
+; MIPS32-N64:     madd.s $f[[F1:[0-9]]], $f13, $f[[F0]], $f12
+; MIPSR6:         mul.s $f[[F2:[0-9]+]], $f[[F0]], $f12
+; MIPSR6-O32:     add.s $f[[F1:[0-9]+]], $f[[F2]], $f14
+; MIPSR6-N32:     add.s $f[[F1:[0-9]+]], $f[[F2]], $f13
+; MIPSR6-N64:     add.s $f[[F1:[0-9]+]], $f[[F2]], $f13
+
+  %fmuladd = call float @llvm.fmuladd.f32(float %1, float %b, float %c)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %fmuladd)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.fabs.f32(float %Val)
+
+define void @ffabs() {
+entry:
+; ALL-LABEL: ffabs:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL:            mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; ALL:            abs.s $f[[F1:[0-9]+]], $f[[F0]]
+
+  %fabs = call float @llvm.fabs.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %fabs)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+
+; ALL:            sh $[[R3]]
+  ret void
+}
+
+declare float @llvm.minnum.f32(float %Val, float %b)
+
+define void @fminnum(float %b) {
+entry:
+; ALL-LABEL: fminnum:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(fminf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(fminf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(fminf)($gp)
+; ALL-DAG:        jalr $25
+
+  %minnum = call float @llvm.minnum.f32(float %1, float %b)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %minnum)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.maxnum.f32(float %Val, float %b)
+
+define void @fmaxnum(float %b) {
+entry:
+; ALL-LABEL: fmaxnum:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(fmaxf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(fmaxf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(fmaxf)($gp)
+; ALL-DAG:        jalr $25
+
+  %maxnum = call float @llvm.maxnum.f32(float %1, float %b)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %maxnum)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:             sh $[[R3]]
+
+  ret void
+}
+
+; This expansion of fcopysign could be done without converting f16 to float.
+
+declare float @llvm.copysign.f32(float %Val, float %b)
+
+define void @fcopysign(float %b) {
+entry:
+; ALL-LABEL: fcopysign:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+
+  %copysign = call float @llvm.copysign.f32(float %1, float %b)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %copysign)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f12
+; ALL:            ext $[[R3:[0-9]+]], $3, 31, 1
+; ALL:            ins $[[R1]], $[[R3]], 31, 1
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R1]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.floor.f32(float %Val)
+
+define void @ffloor() {
+entry:
+; ALL-LABEL: ffloor:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(floorf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(floorf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(floorf)($gp)
+; ALL-DAG:        jalr $25
+
+  %floor = call float @llvm.floor.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %floor)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.ceil.f32(float %Val)
+
+define void @fceil() {
+entry:
+; ALL-LABEL: fceil:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(ceilf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(ceilf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(ceilf)($gp)
+; ALL-DAG:        jalr $25
+
+  %ceil = call float @llvm.ceil.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %ceil)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.trunc.f32(float %Val)
+
+define void @ftrunc() {
+entry:
+; ALL-LABEL: ftrunc:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(truncf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(truncf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(truncf)($gp)
+; ALL-DAG:        jalr $25
+
+  %trunc = call float @llvm.trunc.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %trunc)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.rint.f32(float %Val)
+
+define void @frint() {
+entry:
+; ALL-LABEL: frint:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(rintf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(rintf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(rintf)($gp)
+; ALL-DAG:        jalr $25
+  %rint = call float @llvm.rint.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %rint)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+  store i16 %2, i16* @g, align 2
+
+; ALL:            sh $[[R3]]
+  ret void
+}
+
+declare float @llvm.nearbyint.f32(float %Val)
+
+define void @fnearbyint() {
+entry:
+; ALL-LABEL: fnearbyint:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(nearbyintf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(nearbyintf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(nearbyintf)($gp)
+; ALL-DAG:        jalr $25
+
+  %nearbyint = call float @llvm.nearbyint.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %nearbyint)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}
+
+declare float @llvm.round.f32(float %Val)
+
+define void @fround() {
+entry:
+; ALL-LABEL: fround:
+  %0 = load i16, i16* @g, align 2
+  %1 = call float @llvm.convert.from.fp16.f32(i16 %0)
+
+; ALL:            lh $[[R0:[0-9]+]]
+; ALL:            fill.h $w[[W0:[0-9]+]], $[[R0]]
+; ALL:            fexupr.w $w[[W1:[0-9]+]], $w[[W0]]
+; ALL:            copy_s.w $[[R1:[0-9]+]], $w[[W1]][0]
+; ALL-DAG:        mtc1 $[[R1]], $f[[F0:[0-9]+]]
+; MIPS32-DAG:     lw $25, %call16(roundf)($gp)
+; MIPS64-N32-DAG: lw $25, %call16(roundf)($gp)
+; MIPS64-N64-DAG: ld $25, %call16(roundf)($gp)
+; ALL-DAG:        jalr $25
+
+  %round = call float @llvm.round.f32(float %1)
+  %2 = call i16 @llvm.convert.to.fp16.f32(float %round)
+
+; ALL:            mfc1 $[[R2:[0-9]+]], $f[[F1]]
+; ALL:            fill.w $w[[W2:[0-9]+]], $[[R2]]
+; ALL:            fexdo.h $w[[W3:[0-9]+]], $w[[W2]], $w[[W2]]
+; ALL:            copy_u.h $[[R3:[0-9]+]], $w[[W3]][0]
+
+  store i16 %2, i16* @g, align 2
+; ALL:            sh $[[R3]]
+
+  ret void
+}

test/CodeGen/Mips/msa/fexuprl.ll

+; RUN: llc < %s -march=mips -mcpu=mips32r5 -mattr=+fp64,+msa | FileCheck %s
+
+; Test that fexup[rl].w don't crash LLVM during type legalization.
+
+@g = local_unnamed_addr global <8 x half> <half 0xH5BF8, half 0xH5BF8, half 0xH5BF8, half 0xH5BF8, half 0xH73C0, half 0xH73C0, half 0xH73C0, half 0xH73C0>, align 16
+@i = local_unnamed_addr global <4 x float> zeroinitializer, align 16
+@j = local_unnamed_addr global <4 x float> zeroinitializer, align 16
+
+define i32 @test() local_unnamed_addr {
+entry:
+  %0 = load <8 x half>, <8 x half>* @g, align 16
+  %1 = tail call <4 x float> @llvm.mips.fexupl.w(<8 x half> %0)
+  store <4 x float> %1, <4 x float>* @i, align 16
+; CHECK: ld.h $w[[W0:[0-9]+]], 0(${{[0-9]+}})
+; CHECK: fexupl.w $w[[W1:[0-9]+]], $w[[W0]]
+; CHECK: st.w $w[[W1]], 0(${{[0-9]+}})
+
+  %2 = tail call <4 x float> @llvm.mips.fexupr.w(<8 x half> %0)
+  store <4 x float> %2, <4 x float>* @j, align 16
+
+; CHECK: fexupr.w $w[[W2:[0-9]+]], $w[[W0]]
+; CHECK: st.w $w[[W2]], 0(${{[0-9]+}})
+
+  ret i32 0
+}
+
+declare <4 x float> @llvm.mips.fexupl.w(<8 x half>)
+declare <4 x float> @llvm.mips.fexupr.w(<8 x half>)