[mips][msa] Implemented extract_vector_elt for v4f32 or v2f64

For v4f32 and v2f64, EXTRACT_VECTOR_ELT is matched by a pseudo-insn which may
be expanded to subregister copies and/or instructions as appropriate.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@191514 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 104 insertions, 4 deletions

diff --git a/lib/Target/Mips/MipsMSAInstrInfo.td b/lib/Target/Mips/MipsMSAInstrInfo.td
index 7ed2f84..13b5775 100644
--- a/lib/Target/Mips/MipsMSAInstrInfo.td
+++ b/lib/Target/Mips/MipsMSAInstrInfo.td
@@ -1076,6 +1076,13 @@ class MSA_COPY_DESC_BASE<string instr_asm, SDPatternOperator OpNode,
   InstrItinClass Itinerary = itin;
 }
 
+class MSA_COPY_PSEUDO_BASE<SDPatternOperator OpNode, ValueType VecTy,
+                           RegisterClass RCD, RegisterClass RCWS> :
+      MipsPseudo<(outs RCD:$wd), (ins RCWS:$ws, uimm4:$n),
+                 [(set RCD:$wd, (OpNode (VecTy RCWS:$ws), immZExt4:$n))]> {
+  bit usesCustomInserter = 1;
+}
+
 class MSA_I5_DESC_BASE<string instr_asm, SDPatternOperator OpNode,
                        SplatComplexPattern SplatImm, RegisterClass RCWD,
                        RegisterClass RCWS = RCWD,
@@ -1581,6 +1588,11 @@ class COPY_U_H_DESC : MSA_COPY_DESC_BASE<"copy_u.h", vextract_zext_i16, v8i16,
 class COPY_U_W_DESC : MSA_COPY_DESC_BASE<"copy_u.w", vextract_zext_i32, v4i32,
                                          GPR32, MSA128W>;
 
+class COPY_FW_PSEUDO_DESC : MSA_COPY_PSEUDO_BASE<vector_extract, v4f32, FGR32,
+                                                 MSA128W>;
+class COPY_FD_PSEUDO_DESC : MSA_COPY_PSEUDO_BASE<vector_extract, v2f64, FGR64,
+                                                 MSA128D>;
+
 class CTCMSA_DESC {
   dag OutOperandList = (outs);
   dag InOperandList = (ins MSACtrl:$cd, GPR32:$rs);
@@ -2579,6 +2591,9 @@ def COPY_U_B : COPY_U_B_ENC, COPY_U_B_DESC;
 def COPY_U_H : COPY_U_H_ENC, COPY_U_H_DESC;
 def COPY_U_W : COPY_U_W_ENC, COPY_U_W_DESC;
 
+def COPY_FW_PSEUDO : COPY_FW_PSEUDO_DESC;
+def COPY_FD_PSEUDO : COPY_FD_PSEUDO_DESC;
+
 def CTCMSA : CTCMSA_ENC, CTCMSA_DESC;
 
 def DIV_S_B : DIV_S_B_ENC, DIV_S_B_DESC;
diff --git a/lib/Target/Mips/MipsSEISelLowering.cpp b/lib/Target/Mips/MipsSEISelLowering.cpp
index 84db5ce..99c7019 100644
--- a/lib/Target/Mips/MipsSEISelLowering.cpp
+++ b/lib/Target/Mips/MipsSEISelLowering.cpp
@@ -827,6 +827,10 @@ MipsSETargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
     return emitMSACBranchPseudo(MI, BB, Mips::BZ_D);
   case Mips::SZ_V_PSEUDO:
     return emitMSACBranchPseudo(MI, BB, Mips::BZ_V);
+  case Mips::COPY_FW_PSEUDO:
+    return emitCOPY_FW(MI, BB);
+  case Mips::COPY_FD_PSEUDO:
+    return emitCOPY_FD(MI, BB);
   }
 }
 
@@ -1662,10 +1666,19 @@ lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
   SDLoc DL(Op);
   EVT ResTy = Op->getValueType(0);
   SDValue Op0 = Op->getOperand(0);
-  SDValue Op1 = Op->getOperand(1);
-  EVT EltTy = Op0->getValueType(0).getVectorElementType();
-  return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
-                     DAG.getValueType(EltTy));
+  EVT VecTy = Op0->getValueType(0);
+
+  if (!VecTy.is128BitVector())
+    return SDValue();
+
+  if (ResTy.isInteger()) {
+    SDValue Op1 = Op->getOperand(1);
+    EVT EltTy = VecTy.getVectorElementType();
+    return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
+                       DAG.getValueType(EltTy));
+  }
+
+  return Op;
 }
 
 static bool isConstantOrUndef(const SDValue Op) {
@@ -2236,3 +2249,69 @@ emitMSACBranchPseudo(MachineInstr *MI, MachineBasicBlock *BB,
   MI->eraseFromParent();   // The pseudo instruction is gone now.
   return Sink;
 }
+
+// Emit the COPY_FW pseudo instruction.
+//
+// copy_fw_pseudo $fd, $ws, n
+// =>
+// copy_u_w $rt, $ws, $n
+// mtc1     $rt, $fd
+//
+// When n is zero, the equivalent operation can be performed with (potentially)
+// zero instructions due to register overlaps. This optimization is never valid
+// for lane 1 because it would require FR=0 mode which isn't supported by MSA.
+MachineBasicBlock * MipsSETargetLowering::
+emitCOPY_FW(MachineInstr *MI, MachineBasicBlock *BB) const{
+  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+  MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
+  DebugLoc DL = MI->getDebugLoc();
+  unsigned Fd = MI->getOperand(0).getReg();
+  unsigned Ws = MI->getOperand(1).getReg();
+  unsigned Lane = MI->getOperand(2).getImm();
+
+  if (Lane == 0)
+    BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_lo);
+  else {
+    unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128WRegClass);
+
+    BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_W), Wt).addReg(Ws).addImm(1);
+    BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_lo);
+  }
+
+  MI->eraseFromParent();   // The pseudo instruction is gone now.
+  return BB;
+}
+
+// Emit the COPY_FD pseudo instruction.
+//
+// copy_fd_pseudo $fd, $ws, n
+// =>
+// splati.d $wt, $ws, $n
+// copy $fd, $wt:sub_64
+//
+// When n is zero, the equivalent operation can be performed with (potentially)
+// zero instructions due to register overlaps. This optimization is always
+// valid because FR=1 mode which is the only supported mode in MSA.
+MachineBasicBlock * MipsSETargetLowering::
+emitCOPY_FD(MachineInstr *MI, MachineBasicBlock *BB) const{
+  assert(Subtarget->isFP64bit());
+
+  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+  MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
+  unsigned Fd  = MI->getOperand(0).getReg();
+  unsigned Ws  = MI->getOperand(1).getReg();
+  unsigned Lane = MI->getOperand(2).getImm() * 2;
+  DebugLoc DL = MI->getDebugLoc();
+
+  if (Lane == 0)
+    BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Ws, 0, Mips::sub_64);
+  else {
+    unsigned Wt = RegInfo.createVirtualRegister(&Mips::MSA128DRegClass);
+
+    BuildMI(*BB, MI, DL, TII->get(Mips::SPLATI_D), Wt).addReg(Ws).addImm(1);
+    BuildMI(*BB, MI, DL, TII->get(Mips::COPY), Fd).addReg(Wt, 0, Mips::sub_64);
+  }
+
+  MI->eraseFromParent();   // The pseudo instruction is gone now.
+  return BB;
+}
diff --git a/lib/Target/Mips/MipsSEISelLowering.h b/lib/Target/Mips/MipsSEISelLowering.h
index 9b69fb5..03f1cc5 100644
--- a/lib/Target/Mips/MipsSEISelLowering.h
+++ b/lib/Target/Mips/MipsSEISelLowering.h
@@ -84,6 +84,12 @@ namespace llvm {
     MachineBasicBlock *emitMSACBranchPseudo(MachineInstr *MI,
                                             MachineBasicBlock *BB,
                                             unsigned BranchOp) const;
+    /// \brief Emit the COPY_FW pseudo instruction
+    MachineBasicBlock *emitCOPY_FW(MachineInstr *MI,
+                                   MachineBasicBlock *BB) const;
+    /// \brief Emit the COPY_FD pseudo instruction
+    MachineBasicBlock *emitCOPY_FD(MachineInstr *MI,
+                                   MachineBasicBlock *BB) const;
   };
 }