[mips][msa] Fix vector insertions where the index is variable

[dsanders]

Summary

This isn't supported directly so we rotate the vector by the desired number of
elements, insert to element zero, then rotate back.

The i64 case generates rather poor code on MIPS32. There is an obvious
optimisation to be made in future (do both insert.w's inside a shared
rotate/unrotate sequence) but for now it's sufficient to select valid code
instead of aborting.

Depends on D3536

[matheusalmeida]

This looks good to me but take it with a grain of salt. The reason for this is that I'm not familiar with some methods used here (createVirtualReg...) and as you know I can't spend too much time doing these reviews.
Out of curiosity: Do you know if GCC uses the same instruction sequence as the one implemented here ?

lib/Target/Mips/MipsSEISelLowering.cpp

2964

This might be a dumb question but why MSA128WRegClass ? Would MSA128DRegClass also be OK ?

[dsanders]

Out of curiosity: Do you know if GCC uses the same instruction sequence as the one implemented here ?

GCC doesn't generate inserts/copies of any kind at the moment. It's doing insertions with a store vector + store element + load vector sequence.

lib/Target/Mips/MipsSEISelLowering.cpp

2964

That's a bug, it should be using the correct class for the type. I'll fix this.

2982

Same bug

2988

One more.

[matheusalmeida]

What about the uses of GPR32RegClass ? INSERT_D_VIDX_PSEUDO_DESC uses a GPR64RegClass.

[dsanders]

What about the uses of GPR32RegClass ? INSERT_D_VIDX_PSEUDO_DESC uses a GPR64RegClass.

I'll fix that too but I'll base it on isGP64bit() rather than just when insert.d needs emitting. I'll leave it emitting the 32-bit negation though. The slide instructions will only look at the bottom four bits anyway.

It's worth mentioning that there's no testcase for the insert.d path. This can be added when MSA64 tests are generally added.

lib/Target/Mips/MipsMSAInstrInfo.td

class MSA_INSERT_PSEUDO_BASE<SDPatternOperator OpNode, ValueType Ty,
                              RegisterOperand ROWD, RegisterOperand ROFS> :
       MSAPseudo<(outs ROWD:$wd), (ins ROWD:$wd_in, uimm6:$n, ROFS:$fs),
                 [(set ROWD:$wd, (OpNode (Ty ROWD:$wd_in), ROFS:$fs,
                                         immZExt6:$n))]> {
   bit usesCustomInserter = 1;
   string Constraints = "$wd = $wd_in";
 }
 
+class MSA_INSERT_VIDX_PSEUDO_BASE<SDPatternOperator OpNode, ValueType Ty,
+                                  RegisterOperand ROWD, RegisterOperand ROFS> :
+      MSAPseudo<(outs ROWD:$wd), (ins ROWD:$wd_in, GPR32Opnd:$n, ROFS:$fs),
+                [(set ROWD:$wd, (OpNode (Ty ROWD:$wd_in), ROFS:$fs,
+                                        GPR32Opnd:$n))]> {
+  bit usesCustomInserter = 1;
+  string Constraints = "$wd = $wd_in";
+}
+
 class MSA_INSVE_DESC_BASE<string instr_asm, SDPatternOperator OpNode,
                           RegisterOperand ROWD, RegisterOperand ROWS = ROWD,
                           InstrItinClass itin = NoItinerary> {
   dag OutOperandList = (outs ROWD:$wd);
   dag InOperandList = (ins ROWD:$wd_in, uimm6:$n, ROWS:$ws, uimmz:$n2);
   string AsmString = !strconcat(instr_asm, "\t$wd[$n], $ws[$n2]");
   list<dag> Pattern = [(set ROWD:$wd, (OpNode ROWD:$wd_in,
                                               immZExt6:$n,
class INSERT_B_DESC : MSA_INSERT_DESC_BASE<"insert.b", vinsert_v16i8,
                                            MSA128BOpnd, GPR32Opnd>;
 class INSERT_H_DESC : MSA_INSERT_DESC_BASE<"insert.h", vinsert_v8i16,
                                            MSA128HOpnd, GPR32Opnd>;
 class INSERT_W_DESC : MSA_INSERT_DESC_BASE<"insert.w", vinsert_v4i32,
                                            MSA128WOpnd, GPR32Opnd>;
 class INSERT_D_DESC : MSA_INSERT_DESC_BASE<"insert.d", vinsert_v2i64,
                                            MSA128DOpnd, GPR64Opnd>;
 
+class INSERT_B_VIDX_PSEUDO_DESC :
+    MSA_INSERT_VIDX_PSEUDO_BASE<vector_insert, v16i8, MSA128BOpnd, GPR32Opnd>;
+class INSERT_H_VIDX_PSEUDO_DESC :
+    MSA_INSERT_VIDX_PSEUDO_BASE<vector_insert, v8i16, MSA128HOpnd, GPR32Opnd>;
+class INSERT_W_VIDX_PSEUDO_DESC :
+    MSA_INSERT_VIDX_PSEUDO_BASE<vector_insert, v4i32, MSA128WOpnd, GPR32Opnd>;
+class INSERT_D_VIDX_PSEUDO_DESC :
+    MSA_INSERT_VIDX_PSEUDO_BASE<vector_insert, v2i64, MSA128DOpnd, GPR64Opnd>;
+
 class INSERT_FW_PSEUDO_DESC : MSA_INSERT_PSEUDO_BASE<vector_insert, v4f32,
                                                      MSA128WOpnd, FGR32Opnd>;
 class INSERT_FD_PSEUDO_DESC : MSA_INSERT_PSEUDO_BASE<vector_insert, v2f64,
                                                      MSA128DOpnd, FGR64Opnd>;
 
+class INSERT_FW_VIDX_PSEUDO_DESC :
+    MSA_INSERT_VIDX_PSEUDO_BASE<vector_insert, v4f32, MSA128WOpnd, FGR32Opnd>;
+class INSERT_FD_VIDX_PSEUDO_DESC :
+    MSA_INSERT_VIDX_PSEUDO_BASE<vector_insert, v2f64, MSA128DOpnd, FGR64Opnd>;
+
 class INSVE_B_DESC : MSA_INSVE_DESC_BASE<"insve.b", insve_v16i8,
                                          MSA128BOpnd>;
 class INSVE_H_DESC : MSA_INSVE_DESC_BASE<"insve.h", insve_v8i16,
                                          MSA128HOpnd>;
 class INSVE_W_DESC : MSA_INSVE_DESC_BASE<"insve.w", insve_v4i32,
                                          MSA128WOpnd>;
 class INSVE_D_DESC : MSA_INSVE_DESC_BASE<"insve.d", insve_v2i64,
                                          MSA128DOpnd>;
let DecoderMethod = "DecodeINSVE_DF" in {
   def INSVE_H : INSVE_H_ENC, INSVE_H_DESC;
   def INSVE_W : INSVE_W_ENC, INSVE_W_DESC;
   def INSVE_D : INSVE_D_ENC, INSVE_D_DESC;
 }
 
 def INSERT_FW_PSEUDO : INSERT_FW_PSEUDO_DESC;
 def INSERT_FD_PSEUDO : INSERT_FD_PSEUDO_DESC;
 
+def INSERT_B_VIDX_PSEUDO : INSERT_B_VIDX_PSEUDO_DESC;
+def INSERT_H_VIDX_PSEUDO : INSERT_H_VIDX_PSEUDO_DESC;
+def INSERT_W_VIDX_PSEUDO : INSERT_W_VIDX_PSEUDO_DESC;
+def INSERT_D_VIDX_PSEUDO : INSERT_D_VIDX_PSEUDO_DESC;
+def INSERT_FW_VIDX_PSEUDO : INSERT_FW_VIDX_PSEUDO_DESC;
+def INSERT_FD_VIDX_PSEUDO : INSERT_FD_VIDX_PSEUDO_DESC;
+
 def LD_B: LD_B_ENC, LD_B_DESC;
 def LD_H: LD_H_ENC, LD_H_DESC;
 def LD_W: LD_W_ENC, LD_W_DESC;
 def LD_D: LD_D_ENC, LD_D_DESC;
 
 def LDI_B : LDI_B_ENC, LDI_B_DESC;
 def LDI_H : LDI_H_ENC, LDI_H_DESC;
 def LDI_W : LDI_W_ENC, LDI_W_DESC;

lib/Target/Mips/MipsSEISelLowering.h

private:
     MachineBasicBlock *emitCOPY_FD(MachineInstr *MI,
                                    MachineBasicBlock *BB) const;
     /// \brief Emit the INSERT_FW pseudo instruction
     MachineBasicBlock *emitINSERT_FW(MachineInstr *MI,
                                      MachineBasicBlock *BB) const;
     /// \brief Emit the INSERT_FD pseudo instruction
     MachineBasicBlock *emitINSERT_FD(MachineInstr *MI,
                                      MachineBasicBlock *BB) const;
+    /// \brief Emit the INSERT_([BHWD]|F[WD])_VIDX pseudo instruction
+    MachineBasicBlock *emitINSERT_DF_VIDX(MachineInstr *MI,
+                                          MachineBasicBlock *BB,
+                                          unsigned EltSizeInBytes,
+                                          bool IsFP) const;
     /// \brief Emit the FILL_FW pseudo instruction
     MachineBasicBlock *emitFILL_FW(MachineInstr *MI,
                                    MachineBasicBlock *BB) const;
     /// \brief Emit the FILL_FD pseudo instruction
     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;
   };
 }
 
 #endif // MipsSEISELLOWERING_H

lib/Target/Mips/MipsSEISelLowering.cpp

MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
   case Mips::COPY_FW_PSEUDO:
     return emitCOPY_FW(MI, BB);
   case Mips::COPY_FD_PSEUDO:
     return emitCOPY_FD(MI, BB);
   case Mips::INSERT_FW_PSEUDO:
     return emitINSERT_FW(MI, BB);
   case Mips::INSERT_FD_PSEUDO:
     return emitINSERT_FD(MI, BB);
+  case Mips::INSERT_B_VIDX_PSEUDO:
+    return emitINSERT_DF_VIDX(MI, BB, 1, false);
+  case Mips::INSERT_H_VIDX_PSEUDO:
+    return emitINSERT_DF_VIDX(MI, BB, 2, false);
+  case Mips::INSERT_W_VIDX_PSEUDO:
+    return emitINSERT_DF_VIDX(MI, BB, 4, false);
+  case Mips::INSERT_D_VIDX_PSEUDO:
+    return emitINSERT_DF_VIDX(MI, BB, 8, false);
+  case Mips::INSERT_FW_VIDX_PSEUDO:
+    return emitINSERT_DF_VIDX(MI, BB, 4, true);
+  case Mips::INSERT_FD_VIDX_PSEUDO:
+    return emitINSERT_DF_VIDX(MI, BB, 8, true);
   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);
BuildMI(*BB, MI, DL, TII->get(Mips::INSVE_D), Wd)
       .addImm(Lane)
       .addReg(Wt)
       .addImm(0);
 
   MI->eraseFromParent(); // The pseudo instruction is gone now.
   return BB;
 }
 
+// Emit the INSERT_([BHWD]|F[WD])_VIDX pseudo instruction.
+//
+// For integer:
+// (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $rs)
+// =>
+// (SLL $lanetmp1, $lane, <log2size)
+// (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
+// (INSERT_[BHWD], $wdtmp2, $wdtmp1, 0, $rs)
+// (NEG $lanetmp2, $lanetmp1)
+// (SLD_B $wd, $wdtmp2, $wdtmp2,  $lanetmp2)
+//
+// For floating point:
+// (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $fs)
+// =>
+// (SUBREG_TO_REG $wt, $fs, <subreg>)
+// (SLL $lanetmp1, $lane, <log2size)
+// (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
+// (INSVE_[WD], $wdtmp2, 0, $wdtmp1, 0)
+// (NEG $lanetmp2, $lanetmp1)
+// (SLD_B $wd, $wdtmp2, $wdtmp2,  $lanetmp2)
+MachineBasicBlock *
+MipsSETargetLowering::emitINSERT_DF_VIDX(MachineInstr *MI,
+                                         MachineBasicBlock *BB,
+                                         unsigned EltSizeInBytes,
+                                         bool IsFP) const {
+  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+  MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
+  DebugLoc DL = MI->getDebugLoc();
+  unsigned Wd = MI->getOperand(0).getReg();
+  unsigned SrcVecReg = MI->getOperand(1).getReg();
+  unsigned LaneReg = MI->getOperand(2).getReg();
+  unsigned SrcValReg = MI->getOperand(3).getReg();
+
+  unsigned EltLog2Size;
+  unsigned InsertOp = 0;
+  unsigned InsveOp = 0;
+  switch (EltSizeInBytes) {
+  default:
+    llvm_unreachable("Unexpected size");
+  case 1:
+    EltLog2Size = 0;
+    InsertOp = Mips::INSERT_B;
+    InsveOp = Mips::INSVE_B;
+    break;
+  case 2:
+    EltLog2Size = 1;
+    InsertOp = Mips::INSERT_H;
+    InsveOp = Mips::INSVE_H;
+    break;
+  case 4:
+    EltLog2Size = 2;
+    InsertOp = Mips::INSERT_W;
+    InsveOp = Mips::INSVE_W;
+    break;
+  case 8:
+    EltLog2Size = 3;
+    InsertOp = Mips::INSERT_D;
+    InsveOp = Mips::INSVE_D;
+    break;
+  }
+
+  if (IsFP) {
+    unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+    BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
+        .addImm(0)
+        .addReg(SrcValReg)
+        .addImm(EltSizeInBytes == 8 ? Mips::sub_64 : Mips::sub_lo);
+    SrcValReg = Wt;
+  }
+
+  // Convert the lane index into a byte index
+  if (EltSizeInBytes != 1) {
+    unsigned LaneTmp1 = RegInfo.createVirtualRegister(&Mips::GPR32RegClass);
+    BuildMI(*BB, MI, DL, TII->get(Mips::SLL), LaneTmp1)
+        .addReg(LaneReg)
+        .addImm(EltLog2Size);
+    LaneReg = LaneTmp1;
+  }
+
+  // Rotate bytes around so that the desired lane is element zero
+  unsigned WdTmp1 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+  BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), WdTmp1)
+      .addReg(SrcVecReg)
+      .addReg(SrcVecReg)
+      .addReg(LaneReg);
+
+  unsigned WdTmp2 = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+  if (IsFP) {
+    // Use insve.df to insert to element zero
+    BuildMI(*BB, MI, DL, TII->get(InsveOp), WdTmp2)
+        .addReg(WdTmp1)
+        .addImm(0)
+        .addReg(SrcValReg)
+        .addImm(0);
+  } else {
+    // Use insert.df to insert to element zero
+    BuildMI(*BB, MI, DL, TII->get(InsertOp), WdTmp2)
+        .addReg(WdTmp1)
+        .addReg(SrcValReg)
+        .addImm(0);
+  }
+
+  // Rotate elements the rest of the way for a full rotation.
+  // sld.df inteprets $rt modulo the number of columns so we only need to negate
+  // the lane index to do this.
+  unsigned LaneTmp2 = RegInfo.createVirtualRegister(&Mips::GPR32RegClass);
+  BuildMI(*BB, MI, DL, TII->get(Mips::SUB), LaneTmp2)
+      .addReg(Mips::ZERO)
+      .addReg(LaneReg);
+  BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), Wd)
+      .addReg(WdTmp2)
+      .addReg(WdTmp2)
+      .addReg(LaneTmp2);
+
+  MI->eraseFromParent(); // The pseudo instruction is gone now.
+  return BB;
+}
+
 // Emit the FILL_FW pseudo instruction.
 //
 // fill_fw_pseudo $wd, $fs
 // =>
 // implicit_def $wt1
 // insert_subreg $wt2:subreg_lo, $wt1, $fs
 // splati.w $wd, $wt2[0]
 MachineBasicBlock *

test/CodeGen/Mips/msa/basic_operations.ll

define void @insert_v2i64(i64 %a) nounwind {
 
   store <2 x i64> %2, <2 x i64>* @v2i64
   ; MIPS32-AE-DAG: st.w [[R1]]
 
   ret void
   ; MIPS32-AE: .size insert_v2i64
 }
 
+define void @insert_v16i8_vidx(i32 %a) nounwind {
+  ; MIPS32-AE: insert_v16i8_vidx:
+
+  %1 = load <16 x i8>* @v16i8
+  ; MIPS32-AE-DAG: ld.b [[R1:\$w[0-9]+]],
+
+  %2 = load i32* @i32
+  ; MIPS32-AE-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)(
+  ; MIPS32-AE-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]])
+
+  %a2 = trunc i32 %a to i8
+  %a3 = sext i8 %a2 to i32
+  %a4 = trunc i32 %a3 to i8
+  ; MIPS32-AE-NOT: andi
+  ; MIPS32-AE-NOT: sra
+
+  %3 = insertelement <16 x i8> %1, i8 %a4, i32 %2
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[IDX]]]
+  ; MIPS32-AE-DAG: insert.b [[R1]][0], $4
+  ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[IDX]]
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]]
+
+  store <16 x i8> %3, <16 x i8>* @v16i8
+  ; MIPS32-AE-DAG: st.b [[R1]]
+
+  ret void
+  ; MIPS32-AE: .size insert_v16i8_vidx
+}
+
+define void @insert_v8i16_vidx(i32 %a) nounwind {
+  ; MIPS32-AE: insert_v8i16_vidx:
+
+  %1 = load <8 x i16>* @v8i16
+  ; MIPS32-AE-DAG: ld.h [[R1:\$w[0-9]+]],
+
+  %2 = load i32* @i32
+  ; MIPS32-AE-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)(
+  ; MIPS32-AE-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]])
+
+  %a2 = trunc i32 %a to i16
+  %a3 = sext i16 %a2 to i32
+  %a4 = trunc i32 %a3 to i16
+  ; MIPS32-AE-NOT: andi
+  ; MIPS32-AE-NOT: sra
+
+  %3 = insertelement <8 x i16> %1, i16 %a4, i32 %2
+  ; MIPS32-AE-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 1
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]]
+  ; MIPS32-AE-DAG: insert.h [[R1]][0], $4
+  ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]]
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]]
+
+  store <8 x i16> %3, <8 x i16>* @v8i16
+  ; MIPS32-AE-DAG: st.h [[R1]]
+
+  ret void
+  ; MIPS32-AE: .size insert_v8i16_vidx
+}
+
+define void @insert_v4i32_vidx(i32 %a) nounwind {
+  ; MIPS32-AE: insert_v4i32_vidx:
+
+  %1 = load <4 x i32>* @v4i32
+  ; MIPS32-AE-DAG: ld.w [[R1:\$w[0-9]+]],
+
+  %2 = load i32* @i32
+  ; MIPS32-AE-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)(
+  ; MIPS32-AE-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]])
+
+  ; MIPS32-AE-NOT: andi
+  ; MIPS32-AE-NOT: sra
+
+  %3 = insertelement <4 x i32> %1, i32 %a, i32 %2
+  ; MIPS32-AE-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 2
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]]
+  ; MIPS32-AE-DAG: insert.w [[R1]][0], $4
+  ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]]
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]]
+
+  store <4 x i32> %3, <4 x i32>* @v4i32
+  ; MIPS32-AE-DAG: st.w [[R1]]
+
+  ret void
+  ; MIPS32-AE: .size insert_v4i32_vidx
+}
+
+define void @insert_v2i64_vidx(i64 %a) nounwind {
+  ; MIPS32-AE: insert_v2i64_vidx:
+
+  %1 = load <2 x i64>* @v2i64
+  ; MIPS32-AE-DAG: ld.w [[R1:\$w[0-9]+]],
+
+  %2 = load i32* @i32
+  ; MIPS32-AE-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)(
+  ; MIPS32-AE-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]])
+
+  ; MIPS32-AE-NOT: andi
+  ; MIPS32-AE-NOT: sra
+
+  %3 = insertelement <2 x i64> %1, i64 %a, i32 %2
+  ; TODO: This code could be a lot better but it works. The legalizer splits
+  ; 64-bit inserts into two 32-bit inserts because there is no i64 type on
+  ; MIPS32. The obvious optimisation is to perform both insert.w's at once while
+  ; the vector is rotated.
+  ; MIPS32-AE-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 2
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]]
+  ; MIPS32-AE-DAG: insert.w [[R1]][0], $4
+  ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]]
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]]
+  ; MIPS32-AE-DAG: addiu [[IDX2:\$[0-9]+]], [[IDX]], 1
+  ; MIPS32-AE-DAG: sll [[BIDX:\$[0-9]+]], [[IDX2]], 2
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]]
+  ; MIPS32-AE-DAG: insert.w [[R1]][0], $5
+  ; MIPS32-AE-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]]
+  ; MIPS32-AE-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]]
+
+  store <2 x i64> %3, <2 x i64>* @v2i64
+  ; MIPS32-AE-DAG: st.w [[R1]]
+
+  ret void
+  ; MIPS32-AE: .size insert_v2i64_vidx
+}
+
 define void @truncstore() nounwind {
   ; MIPS32-AE-LABEL: truncstore:
 
   store volatile <4 x i8> <i8 -1, i8 -1, i8 -1, i8 -1>, <4 x i8>*@v4i8
   ; TODO: What code should be emitted?
 
   ret void
   ; MIPS32-AE: .size truncstore
 }

test/CodeGen/Mips/msa/basic_operations_float.ll

define void @insert_v2f64(double %a) nounwind {
   ; MIPS32-DAG: insve.d [[R1]][1], $w12[0]
 
   store <2 x double> %2, <2 x double>* @v2f64
   ; MIPS32-DAG: st.d [[R1]]
 
   ret void
   ; MIPS32: .size insert_v2f64
 }
+
+define void @insert_v4f32_vidx(float %a) nounwind {
+  ; MIPS32-LABEL: insert_v4f32_vidx:
+
+  %1 = load <4 x float>* @v4f32
+  ; MIPS32-DAG: lw [[PTR_V:\$[0-9]+]], %got(v4f32)(
+  ; MIPS32-DAG: ld.w [[R1:\$w[0-9]+]], 0([[PTR_V]])
+
+  %2 = load i32* @i32
+  ; MIPS32-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)(
+  ; MIPS32-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]])
+
+  %3 = insertelement <4 x float> %1, float %a, i32 %2
+  ; float argument passed in $f12
+  ; MIPS32-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 2
+  ; MIPS32-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]]
+  ; MIPS32-DAG: insve.w [[R1]][0], $w12[0]
+  ; MIPS32-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]]
+  ; MIPS32-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]]
+
+  store <4 x float> %3, <4 x float>* @v4f32
+  ; MIPS32-DAG: st.w [[R1]]
+
+  ret void
+  ; MIPS32: .size insert_v4f32_vidx
+}
+
+define void @insert_v2f64_vidx(double %a) nounwind {
+  ; MIPS32-LABEL: insert_v2f64_vidx:
+
+  %1 = load <2 x double>* @v2f64
+  ; MIPS32-DAG: lw [[PTR_V:\$[0-9]+]], %got(v2f64)(
+  ; MIPS32-DAG: ld.d [[R1:\$w[0-9]+]], 0([[PTR_V]])
+
+  %2 = load i32* @i32
+  ; MIPS32-DAG: lw [[PTR_I:\$[0-9]+]], %got(i32)(
+  ; MIPS32-DAG: lw [[IDX:\$[0-9]+]], 0([[PTR_I]])
+
+  %3 = insertelement <2 x double> %1, double %a, i32 %2
+  ; double argument passed in $f12
+  ; MIPS32-DAG: sll [[BIDX:\$[0-9]+]], [[IDX]], 3
+  ; MIPS32-DAG: sld.b [[R1]], [[R1]]{{\[}}[[BIDX]]]
+  ; MIPS32-DAG: insve.d [[R1]][0], $w12[0]
+  ; MIPS32-DAG: neg [[NIDX:\$[0-9]+]], [[BIDX]]
+  ; MIPS32-DAG: sld.b [[R1]], [[R1]]{{\[}}[[NIDX]]]
+
+  store <2 x double> %3, <2 x double>* @v2f64
+  ; MIPS32-DAG: st.d [[R1]]
+
+  ret void
+  ; MIPS32: .size insert_v2f64_vidx
+}