[mips] Support SELECT nodes for targets that don't have conditional-move instructions.

[vkalintiris]

Summary

For Mips targets that do not have conditional-move instructions, ie. targets
before MIPS32 and MIPS-IV, we have to insert a diamond control-flow
pattern in order to support SELECT nodes. In order to do that, we add
pseudo-instructions with a custom inserter that emits the necessary
control-flow that selects the correct value.

With this patch we add complete support for code generation of Mips-II targets
based on the LLVM test-suite.

[dsanders]

Great! With this, we will be able to generate code for the mips/mipsel ports of Debian.

There are quite a few comments below but most are minor things. The main one is about whether we really need the SelectFP_[TF] nodes.

lib/Target/Mips/MipsCondMov.td

269

Don't override the Predicates field, it has some surprising results since it's very easy to accidentally remove predicates. Instead override InsnPredicates, preferably using an ISA_* or INSN_* adjective. See the PredicateControl class for more information.

lib/Target/Mips/MipsISelLowering.cpp

535–536

Given that CMovFP_[TF] and SelectFP_[TF] have the same inputs and outputs, do we really need new nodes? Couldn't we use CMovFP_[TF] and use predicates to pick between the CMov instructions (added in Mips4 and Mips32) and the new pseudo's?

lib/Target/Mips/MipsInstrInfo.td

191

Nit: Based on the names of the other predicates this should either be NoCondMov or NotCondMov but, as noted above, it would be better to define and use an INSN_* or ISA_* adjective (e.g. ISA_NOT_4_32).

lib/Target/Mips/MipsSubtarget.h

199–202 ↗

(On Diff #16043)

This won't be needed once you do the above comments.

test/CodeGen/Mips/llvm-ir/select.ll

40

movn from the CMOV case has tied operands but this is not the case for seleqz/selnez. As a result, the first [[T0]] on this line isn't guaranteed to be the same register as the second [[T0]]. You should define another variable.

Likewise for the other tests below.

101–112

Please add a FIXME commenting on the redundant control flow.

115

Please check the offsets as well.

Likewise in the tests below.

135

This instruction is redundant. Could you add a FIXME comment?

143–174

I should mention that for MIPS-II, this is testing selection between two GPR's. This is because of a quirk of the O32 ABI that prevents the use of FPR's if there is a non-floating point argument as the first argument. MIPS32 is testing selection between FPR's since it's copying $5 and $6 to $f0, and $f1 first.

We should duplicate this test with the arguments re-ordered (%x, %y, %s) so that we test the FPR case for 32-bit ISA's.

Likewise for the double test below.

328

Nit: Too many blank lines

396

Nit: Too many blank lines

517

Nit: Too many blank lines

[vkalintiris]

Use CMovFP nodes for Select_Pseudo instructions and update test to address Daniel's comments.

[vkalintiris]

I replaced BC1T with BC1F and vice versa, because we were branching on the
wrong condition.

[dsanders]

LGTM. Thanks.

lib/Target/Mips/MipsCondMov.td

269

Done

lib/Target/Mips/MipsISelLowering.cpp

535–536

Done

lib/Target/Mips/MipsInstrInfo.td

191

Done

test/CodeGen/Mips/llvm-ir/select.ll

40

Done

101–112

Done

115

Done

135

Done

143–174

Done

328

Done

396

Done

517

Done

lib/Target/Mips/MipsCondMov.td

 defm : MovzPats1<GPR32, FGR64, MOVZ_I_D64, XOR>, INSN_MIPS4_32_NOT_32R6_64R6,
        FGR_64;
 defm : MovzPats1<GPR64, FGR64, MOVZ_I64_D64, XOR64>,
        INSN_MIPS4_32_NOT_32R6_64R6, FGR_64;
 defm : MovnPats<GPR32, FGR64, MOVN_I_D64, XOR>, INSN_MIPS4_32_NOT_32R6_64R6,
        FGR_64;
 defm : MovnPats<GPR64, FGR64, MOVN_I64_D64, XOR64>, INSN_MIPS4_32_NOT_32R6_64R6,
        FGR_64;
+
+// For targets that don't have conditional-move instructions
+// we have to match SELECT nodes with pseudo instructions.
+let usesCustomInserter = 1 in {
+  class Select_Pseudo<RegisterOperand RC> :
+    PseudoSE<(outs RC:$dst), (ins GPR32Opnd:$cond, RC:$T, RC:$F),
+            [(set RC:$dst, (select GPR32Opnd:$cond, RC:$T, RC:$F))]>,
+    ISA_MIPS1_NOT_4_32;
+
+  class SelectFP_Pseudo_T<RegisterOperand RC> :
+    PseudoSE<(outs RC:$dst), (ins GPR32Opnd:$cond, RC:$T, RC:$F),
+             [(set RC:$dst, (MipsCMovFP_T RC:$T, GPR32Opnd:$cond, RC:$F))]>,
+    ISA_MIPS1_NOT_4_32;
+
+  class SelectFP_Pseudo_F<RegisterOperand RC> :
+    PseudoSE<(outs RC:$dst), (ins GPR32Opnd:$cond, RC:$T, RC:$F),
+             [(set RC:$dst, (MipsCMovFP_F RC:$T, GPR32Opnd:$cond, RC:$F))]>,
+    ISA_MIPS1_NOT_4_32;
+}
+
+def PseudoSELECT_I : Select_Pseudo<GPR32Opnd>;
+def PseudoSELECT_S : Select_Pseudo<FGR32Opnd>;
+def PseudoSELECT_D32 : Select_Pseudo<AFGR64Opnd>;
+
+def PseudoSELECTFP_T_I : SelectFP_Pseudo_T<GPR32Opnd>;
+def PseudoSELECTFP_T_S : SelectFP_Pseudo_T<FGR32Opnd>;
+def PseudoSELECTFP_T_D32 : SelectFP_Pseudo_T<AFGR64Opnd>;
+
+def PseudoSELECTFP_F_I : SelectFP_Pseudo_F<GPR32Opnd>;
+def PseudoSELECTFP_F_S : SelectFP_Pseudo_F<FGR32Opnd>;
+def PseudoSELECTFP_F_D32 : SelectFP_Pseudo_F<AFGR64Opnd>;

lib/Target/Mips/MipsISelLowering.h

private:
     MachineBasicBlock *emitAtomicBinaryPartword(MachineInstr *MI,
                     MachineBasicBlock *BB, unsigned Size, unsigned BinOpcode,
                     bool Nand = false) const;
     MachineBasicBlock *emitAtomicCmpSwap(MachineInstr *MI,
                                   MachineBasicBlock *BB, unsigned Size) const;
     MachineBasicBlock *emitAtomicCmpSwapPartword(MachineInstr *MI,
                                   MachineBasicBlock *BB, unsigned Size) const;
     MachineBasicBlock *emitSEL_D(MachineInstr *MI, MachineBasicBlock *BB) const;
+    MachineBasicBlock *emitPseudoSELECT(MachineInstr *MI,
+                                        MachineBasicBlock *BB, bool isFPCmp,
+                                        unsigned Opc) const;
   };
 
   /// Create MipsTargetLowering objects.
   const MipsTargetLowering *
   createMips16TargetLowering(const MipsTargetMachine &TM,
                              const MipsSubtarget &STI);
   const MipsTargetLowering *
   createMipsSETargetLowering(const MipsTargetMachine &TM,

lib/Target/Mips/MipsISelLowering.cpp

 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
                             SDValue False, SDLoc DL) {
   ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
   bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
   SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
 
   return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
                      True.getValueType(), True, FCC0, False, Cond);
 }
 
 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
                                     TargetLowering::DAGCombinerInfo &DCI,
                                     const MipsSubtarget &Subtarget) {
   if (DCI.isBeforeLegalizeOps())
     return SDValue();
 
   SDValue SetCC = N->getOperand(0);
 
MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
   case Mips::DDIV:
   case Mips::DDIVU:
   case Mips::DMOD:
   case Mips::DMODU:
     return insertDivByZeroTrap(
         MI, *BB, *getTargetMachine().getSubtargetImpl()->getInstrInfo(), true);
   case Mips::SEL_D:
     return emitSEL_D(MI, BB);
+
+  case Mips::PseudoSELECT_I:
+  case Mips::PseudoSELECT_S:
+  case Mips::PseudoSELECT_D32:
+    return emitPseudoSELECT(MI, BB, false, Mips::BNE);
+  case Mips::PseudoSELECTFP_F_I:
+  case Mips::PseudoSELECTFP_F_S:
+  case Mips::PseudoSELECTFP_F_D32:
+    return emitPseudoSELECT(MI, BB, true, Mips::BC1T);
+  case Mips::PseudoSELECTFP_T_I:
+  case Mips::PseudoSELECTFP_T_S:
+  case Mips::PseudoSELECTFP_T_D32:
+    return emitPseudoSELECT(MI, BB, true, Mips::BC1F);
   }
 }
 
 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
 MachineBasicBlock *
 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
                                      unsigned Size, unsigned BinOpcode,
if (State->getCallingConv() != CallingConv::Fast) {
     Size = RoundUpToAlignment(Size, RegSizeInBytes);
     for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size());
          Size -= RegSizeInBytes, ++I, ++NumRegs)
       State->AllocateReg(IntArgRegs[I], ShadowRegs[I]);
   }
 
   State->addInRegsParamInfo(FirstReg, FirstReg + NumRegs);
 }
+
+MachineBasicBlock *
+MipsTargetLowering::emitPseudoSELECT(MachineInstr *MI, MachineBasicBlock *BB,
+                                     bool isFPCmp, unsigned Opc) const {
+  assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) &&
+         "Subtarget already supports SELECT nodes with the use of"
+         "conditional-move instructions.");
+
+  const TargetInstrInfo *TII =
+      getTargetMachine().getSubtargetImpl()->getInstrInfo();
+  DebugLoc DL = MI->getDebugLoc();
+
+  // To "insert" a SELECT instruction, we actually have to insert the
+  // diamond control-flow pattern.  The incoming instruction knows the
+  // destination vreg to set, the condition code register to branch on, the
+  // true/false values to select between, and a branch opcode to use.
+  const BasicBlock *LLVM_BB = BB->getBasicBlock();
+  MachineFunction::iterator It = BB;
+  ++It;
+
+  //  thisMBB:
+  //  ...
+  //   TrueVal = ...
+  //   setcc r1, r2, r3
+  //   bNE   r1, r0, copy1MBB
+  //   fallthrough --> copy0MBB
+  MachineBasicBlock *thisMBB  = BB;
+  MachineFunction *F = BB->getParent();
+  MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
+  MachineBasicBlock *sinkMBB  = F->CreateMachineBasicBlock(LLVM_BB);
+  F->insert(It, copy0MBB);
+  F->insert(It, sinkMBB);
+
+  // Transfer the remainder of BB and its successor edges to sinkMBB.
+  sinkMBB->splice(sinkMBB->begin(), BB,
+                  std::next(MachineBasicBlock::iterator(MI)), BB->end());
+  sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
+
+  // Next, add the true and fallthrough blocks as its successors.
+  BB->addSuccessor(copy0MBB);
+  BB->addSuccessor(sinkMBB);
+
+  if (isFPCmp) {
+    // bc1[tf] cc, sinkMBB
+    BuildMI(BB, DL, TII->get(Opc))
+      .addReg(MI->getOperand(1).getReg())
+      .addMBB(sinkMBB);
+  } else {
+    // bne rs, $0, sinkMBB
+    BuildMI(BB, DL, TII->get(Opc))
+      .addReg(MI->getOperand(1).getReg())
+      .addReg(Mips::ZERO)
+      .addMBB(sinkMBB);
+  }
+
+  //  copy0MBB:
+  //   %FalseValue = ...
+  //   # fallthrough to sinkMBB
+  BB = copy0MBB;
+
+  // Update machine-CFG edges
+  BB->addSuccessor(sinkMBB);
+
+  //  sinkMBB:
+  //   %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
+  //  ...
+  BB = sinkMBB;
+
+  BuildMI(*BB, BB->begin(), DL,
+          TII->get(Mips::PHI), MI->getOperand(0).getReg())
+    .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
+    .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB);
+
+  MI->eraseFromParent();   // The pseudo instruction is gone now.
+
+  return BB;
+}

lib/Target/Mips/MipsInstrInfo.td

 def HasMips3_32  :    Predicate<"Subtarget->hasMips3_32()">,
                       AssemblerPredicate<"FeatureMips3_32">;
 def HasMips3_32r2 :   Predicate<"Subtarget->hasMips3_32r2()">,
                       AssemblerPredicate<"FeatureMips3_32r2">;
 def HasMips3     :    Predicate<"Subtarget->hasMips3()">,
                       AssemblerPredicate<"FeatureMips3">;
 def HasMips4_32  :    Predicate<"Subtarget->hasMips4_32()">,
                       AssemblerPredicate<"FeatureMips4_32">;
+def NotMips4_32  :    Predicate<"!Subtarget->hasMips4_32()">,
+                      AssemblerPredicate<"FeatureMips4_32">;
 def HasMips4_32r2 :   Predicate<"Subtarget->hasMips4_32r2()">,
                       AssemblerPredicate<"FeatureMips4_32r2">;
 def HasMips5_32r2 :   Predicate<"Subtarget->hasMips5_32r2()">,
                       AssemblerPredicate<"FeatureMips5_32r2">;
 def HasMips32    :    Predicate<"Subtarget->hasMips32()">,
                       AssemblerPredicate<"FeatureMips32">;
 def HasMips32r2  :    Predicate<"Subtarget->hasMips32r2()">,
                       AssemblerPredicate<"FeatureMips32r2">;

 def NotMips64r6  :    Predicate<"!Subtarget->hasMips64r6()">,
                       AssemblerPredicate<"!FeatureMips64r6">;
 def IsN64       :     Predicate<"Subtarget->isABI_N64()">,
                       AssemblerPredicate<"FeatureN64">;
 def InMips16Mode :    Predicate<"Subtarget->inMips16Mode()">,
                       AssemblerPredicate<"FeatureMips16">;
 def HasCnMips    :    Predicate<"Subtarget->hasCnMips()">,
                       AssemblerPredicate<"FeatureCnMips">;
 def RelocStatic :     Predicate<"TM.getRelocationModel() == Reloc::Static">,
                       AssemblerPredicate<"FeatureMips32">;
 def RelocPIC    :     Predicate<"TM.getRelocationModel() == Reloc::PIC_">,
                       AssemblerPredicate<"FeatureMips32">;
 def NoNaNsFPMath :    Predicate<"TM.Options.NoNaNsFPMath">;
 def HasStdEnc :       Predicate<"Subtarget->hasStandardEncoding()">,
                       AssemblerPredicate<"!FeatureMips16">;
 def NotDSP :          Predicate<"!Subtarget->hasDSP()">;
 def InMicroMips    :  Predicate<"Subtarget->inMicroMipsMode()">,

 // They are mutually exclusive.
 //===----------------------------------------------------------------------===//
 
 // FIXME: I'd prefer to use additive predicates to build the instruction sets
 //        but we are short on assembler feature bits at the moment. Using a
 //        subtractive predicate will hopefully keep us under the 32 predicate
 //        limit long enough to develop an alternative way to handle P1||P2
 //        predicates.
+class ISA_MIPS1_NOT_4_32 {
+  list<Predicate> InsnPredicates = [NotMips4_32];
+}
 class ISA_MIPS1_NOT_32R6_64R6 {
   list<Predicate> InsnPredicates = [NotMips32r6, NotMips64r6];
 }
 class ISA_MIPS2    { list<Predicate> InsnPredicates = [HasMips2]; }
 class ISA_MIPS2_NOT_32R6_64R6 {
   list<Predicate> InsnPredicates = [HasMips2, NotMips32r6, NotMips64r6];
 }
 class ISA_MIPS3    { list<Predicate> InsnPredicates = [HasMips3]; }

test/CodeGen/Mips/llvm-ir/select.ll

+; RUN: llc < %s -march=mips -mcpu=mips2 | FileCheck %s \
+; RUN:    -check-prefix=ALL -check-prefix=M2
+; RUN: llc < %s -march=mips -mcpu=mips32 | FileCheck %s \
+; RUN:    -check-prefix=ALL -check-prefix=CMOV \
+; RUN:    -check-prefix=CMOV-32 -check-prefix=CMOV-32R1
+; RUN: llc < %s -march=mips -mcpu=mips32r2 | FileCheck %s \
+; RUN:    -check-prefix=ALL -check-prefix=CMOV \
+; RUN:    -check-prefix=CMOV-32 -check-prefix=CMOV-32R2
+; RUN: llc < %s -march=mips -mcpu=mips32r6 | FileCheck %s \
+; RUN:    -check-prefix=ALL -check-prefix=SEL -check-prefix=SEL-32
+; RUN: llc < %s -march=mips64 -mcpu=mips4 | FileCheck %s \
+; RUN:    -check-prefix=ALL -check-prefix=CMOV -check-prefix=CMOV-64
+; RUN: llc < %s -march=mips64 -mcpu=mips64 | FileCheck %s \
+; RUN:    -check-prefix=ALL -check-prefix=CMOV -check-prefix=CMOV-64
+; RUN: llc < %s -march=mips64 -mcpu=mips64r2 | FileCheck %s \
+; RUN:    -check-prefix=ALL -check-prefix=CMOV -check-prefix=CMOV-64
+; RUN: llc < %s -march=mips64 -mcpu=mips64r6 | FileCheck %s \
+; RUN:    -check-prefix=ALL -check-prefix=SEL -check-prefix=SEL-64
+
+define signext i1 @tst_select_i1_i1(i1 signext %s,
+                                    i1 signext %x, i1 signext %y) {
+entry:
+  ; ALL-LABEL: tst_select_i1_i1:
+
+  ; M2:     andi    $[[T0:[0-9]+]], $4, 1
+  ; M2:     bnez    $[[T0]], $[[BB0:BB[0-9_]+]]
+  ; M2:     nop
+  ; M2:     move    $5, $6
+  ; M2:     $[[BB0]]:
+  ; M2:     jr      $ra
+  ; M2:     move    $2, $5
+
+  ; CMOV:   andi    $[[T0:[0-9]+]], $4, 1
+  ; CMOV:   movn    $6, $5, $[[T0]]
+  ; CMOV:   move    $2, $6
+
+  ; SEL:    andi    $[[T0:[0-9]+]], $4, 1
+  ; SEL:    seleqz  $[[T1:[0-9]+]], $6, $[[T0]]
+  ; SEL:    selnez  $[[T2:[0-9]+]], $5, $[[T0]]
+  ; SEL:    or      $2, $[[T2]], $[[T1]]
+  %r = select i1 %s, i1 %x, i1 %y
+  ret i1 %r
+}
+
+define signext i8 @tst_select_i1_i8(i1 signext %s,
+                                    i8 signext %x, i8 signext %y) {
+entry:
+  ; ALL-LABEL: tst_select_i1_i8:
+
+  ; M2:     andi    $[[T0:[0-9]+]], $4, 1
+  ; M2:     bnez    $[[T0]], $[[BB0:BB[0-9_]+]]
+  ; M2:     nop
+  ; M2:     move    $5, $6
+  ; M2:     $[[BB0]]:
+  ; M2:     jr      $ra
+  ; M2:     move    $2, $5
+
+  ; CMOV:   andi    $[[T0:[0-9]+]], $4, 1
+  ; CMOV:   movn    $6, $5, $[[T0]]
+  ; CMOV:   move    $2, $6
+
+  ; SEL:    andi    $[[T0:[0-9]+]], $4, 1
+  ; SEL:    seleqz  $[[T1:[0-9]+]], $6, $[[T0]]
+  ; SEL:    selnez  $[[T2:[0-9]+]], $5, $[[T0]]
+  ; SEL:    or      $2, $[[T2]], $[[T1]]
+  %r = select i1 %s, i8 %x, i8 %y
+  ret i8 %r
+}
+
+define signext i32 @tst_select_i1_i32(i1 signext %s,
+                                      i32 signext %x, i32 signext %y) {
+entry:
+  ; ALL-LABEL: tst_select_i1_i32:
+
+  ; M2:     andi    $[[T0:[0-9]+]], $4, 1
+  ; M2:     bnez    $[[T0]], $[[BB0:BB[0-9_]+]]
+  ; M2:     nop
+  ; M2:     move    $5, $6
+  ; M2:     $[[BB0]]:
+  ; M2:     jr      $ra
+  ; M2:     move    $2, $5
+
+  ; CMOV:   andi    $[[T0:[0-9]+]], $4, 1
+  ; CMOV:   movn    $6, $5, $[[T0]]
+  ; CMOV:   move    $2, $6
+
+  ; SEL:    andi    $[[T0:[0-9]+]], $4, 1
+  ; SEL:    seleqz  $[[T1:[0-9]+]], $6, $[[T0]]
+  ; SEL:    selnez  $[[T2:[0-9]+]], $5, $[[T0]]
+  ; SEL:    or      $2, $[[T2]], $[[T1]]
+  %r = select i1 %s, i32 %x, i32 %y
+  ret i32 %r
+}
+
+define signext i64 @tst_select_i1_i64(i1 signext %s,
+                                      i64 signext %x, i64 signext %y) {
+entry:
+  ; ALL-LABEL: tst_select_i1_i64:
+
+  ; M2:     andi    $[[T0:[0-9]+]], $4, 1
+  ; M2:     bnez    $[[T0]], $[[BB0:BB[0-9_]+]]
+  ; M2:     nop
+  ; M2:     lw      $[[T1:[0-9]+]], 16($sp)
+  ; M2:     $[[BB0]]:
+  ; FIXME: This branch is redundant
+  ; M2:     bnez    $[[T0]], $[[BB1:BB[0-9_]+]]
+  ; M2:     nop
+  ; M2:     lw      $[[T2:[0-9]+]], 20($sp)
+  ; M2:     $[[BB1]]:
+  ; M2:     move    $2, $[[T1]]
+  ; M2:     jr      $ra
+  ; M2:     move    $3, $[[T2]]
+
+  ; CMOV-32:    andi    $[[T0:[0-9]+]], $4, 1
+  ; CMOV-32:    lw      $2, 16($sp)
+  ; CMOV-32:    movn    $2, $6, $[[T0]]
+  ; CMOV-32:    lw      $3, 20($sp)
+  ; CMOV-32:    movn    $3, $7, $[[T0]]
+
+  ; SEL-32:     andi    $[[T0:[0-9]+]], $4, 1
+  ; SEL-32:     selnez  $[[T1:[0-9]+]], $6, $[[T0]]
+  ; SEL-32:     lw      $[[T2:[0-9]+]], 16($sp)
+  ; SEL-32:     seleqz  $[[T3:[0-9]+]], $[[T2]], $[[T0]]
+  ; SEL-32:     or      $2, $[[T1]], $[[T3]]
+  ; SEL-32:     selnez  $[[T4:[0-9]+]], $7, $[[T0]]
+  ; SEL-32:     lw      $[[T5:[0-9]+]], 20($sp)
+  ; SEL-32:     seleqz  $[[T6:[0-9]+]], $[[T5]], $[[T0]]
+  ; SEL-32:     or      $3, $[[T4]], $[[T6]]
+
+  ; CMOV-64:    andi    $[[T0:[0-9]+]], $4, 1
+  ; CMOV-64:    movn    $6, $5, $[[T0]]
+  ; CMOV-64:    move    $2, $6
+
+  ; SEL-64:     andi    $[[T0:[0-9]+]], $4, 1
+  ; FIXME: This shift is redundant
+  ; SEL-64:     sll     $[[T0]], $[[T0]], 0
+  ; SEL-64:     seleqz  $[[T1:[0-9]+]], $6, $[[T0]]
+  ; SEL-64:     selnez  $[[T0]], $5, $[[T0]]
+  ; SEL-64:     or      $2, $[[T0]], $[[T1]]
+  %r = select i1 %s, i64 %x, i64 %y
+  ret i64 %r
+}
+
+define float @tst_select_i1_float(i1 signext %s, float %x, float %y) {
+entry:
+  ; ALL-LABEL: tst_select_i1_float:
+
+  ; M2:         andi    $[[T0:[0-9]+]], $4, 1
+  ; M2:         bnez    $[[T0]], $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         jr      $ra
+  ; M2:         mtc1    $6, $f0
+  ; M2:         $[[BB0]]:
+  ; M2:         jr      $ra
+  ; M2:         mtc1    $5, $f0
+
+  ; CMOV-32:    mtc1    $6, $f0
+  ; CMOV-32:    mtc1    $5, $f1
+  ; CMOV-32:    andi    $[[T0:[0-9]+]], $4, 1
+  ; CMOV-32:    movn.s  $f0, $f1, $[[T0]]
+
+  ; SEL-32:     mtc1    $5, $[[F0:f[0-9]+]]
+  ; SEL-32:     mtc1    $6, $[[F1:f[0-9]+]]
+  ; SEL-32:     mtc1    $4, $f0
+  ; SEL-32:     sel.s   $f0, $[[F1]], $[[F0]]
+
+  ; CMOV-64:    andi    $[[T0:[0-9]+]], $4, 1
+  ; CMOV-64:    movn.s  $f14, $f13, $[[T0]]
+  ; CMOV-64:    mov.s   $f0, $f14
+
+  ; SEL-64:     mtc1    $4, $f0
+  ; SEL-64:     sel.s   $f0, $f14, $f13
+  %r = select i1 %s, float %x, float %y
+  ret float %r
+}
+
+define float @tst_select_i1_float_reordered(float %x, float %y,
+                                            i1 signext %s) {
+entry:
+  ; ALL-LABEL: tst_select_i1_float_reordered:
+
+  ; M2:         andi    $[[T0:[0-9]+]], $6, 1
+  ; M2:         bnez    $[[T0]], $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.s   $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr      $ra
+  ; M2:         mov.s   $f0, $f12
+
+  ; CMOV-32:    andi    $[[T0:[0-9]+]], $6, 1
+  ; CMOV-32:    movn.s  $f14, $f12, $[[T0]]
+  ; CMOV-32:    mov.s   $f0, $f14
+
+  ; SEL-32:     mtc1    $6, $f0
+  ; SEL-32:     sel.s   $f0, $f14, $f12
+
+  ; CMOV-64:    andi    $[[T0:[0-9]+]], $6, 1
+  ; CMOV-64:    movn.s  $f13, $f12, $[[T0]]
+  ; CMOV-64:    mov.s   $f0, $f13
+
+  ; SEL-64:     mtc1    $6, $f0
+  ; SEL-64:     sel.s   $f0, $f13, $f12
+  %r = select i1 %s, float %x, float %y
+  ret float %r
+}
+
+define double @tst_select_i1_double(i1 signext %s, double %x, double %y) {
+entry:
+  ; ALL-LABEL: tst_select_i1_double:
+
+  ; M2:         andi    $[[T0:[0-9]+]], $4, 1
+  ; M2:         bnez    $[[T0]], $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         ldc1    $f0, 16($sp)
+  ; M2:         jr      $ra
+  ; M2:         nop
+  ; M2:         $[[BB0]]:
+  ; M2:         mtc1    $7, $f0
+  ; M2:         jr      $ra
+  ; M2:         mtc1    $6, $f1
+
+  ; CMOV-32:    mtc1    $7, $[[F0:f[0-9]+]]
+  ; CMOV-32R1:  mtc1    $6, $f{{[0-9]+}}
+  ; CMOV-32R2   mthc1   $6, $[[F0]]
+  ; CMOV-32:    andi    $[[T0:[0-9]+]], $4, 1
+  ; CMOV-32:    ldc1    $f0, 16($sp)
+  ; CMOV-32:    movn.d  $f0, $[[F0]], $[[T0]]
+
+  ; SEL-32:     mtc1    $7, $[[F0:f[0-9]+]]
+  ; SEL-32:     mthc1   $6, $[[F0]]
+  ; SEL-32:     ldc1    $[[F1:f[0-9]+]], 16($sp)
+  ; SEL-32:     mtc1    $4, $f0
+  ; SEL-32:     sel.d   $f0, $[[F1]], $[[F0]]
+
+  ; CMOV-64:    andi    $[[T0:[0-9]+]], $4, 1
+  ; CMOV-64:    movn.d  $f14, $f13, $[[T0]]
+  ; CMOV-64:    mov.d   $f0, $f14
+
+  ; SEL-64:     mtc1    $4, $f0
+  ; SEL-64:     sel.d   $f0, $f14, $f13
+  %r = select i1 %s, double %x, double %y
+  ret double %r
+}
+
+define double @tst_select_i1_double_reordered(double %x, double %y,
+                                              i1 signext %s) {
+entry:
+  ; ALL-LABEL: tst_select_i1_double_reordered:
+
+  ; M2:         lw      $[[T0:[0-9]+]], 16($sp)
+  ; M2:         andi    $[[T1:[0-9]+]], $[[T0]], 1
+  ; M2:         bnez    $[[T1]], $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.d   $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr      $ra
+  ; M2:         mov.d   $f0, $f12
+
+  ; CMOV-32:    lw      $[[T0:[0-9]+]], 16($sp)
+  ; CMOV-32:    andi    $[[T1:[0-9]+]], $[[T0]], 1
+  ; CMOV-32:    movn.d  $f14, $f12, $[[T1]]
+  ; CMOV-32:    mov.d   $f0, $f14
+
+  ; SEL-32:     lw      $[[T0:[0-9]+]], 16($sp)
+  ; SEL-32:     mtc1    $[[T0]], $f0
+  ; SEL-32:     sel.d   $f0, $f14, $f12
+
+  ; CMOV-64:    andi    $[[T0:[0-9]+]], $6, 1
+  ; CMOV-64:    movn.d  $f13, $f12, $[[T0]]
+  ; CMOV-64:    mov.d   $f0, $f13
+
+  ; SEL-64:     mtc1    $6, $f0
+  ; SEL-64:     sel.d   $f0, $f13, $f12
+  %r = select i1 %s, double %x, double %y
+  ret double %r
+}
+
+define float @tst_select_fcmp_olt_float(float %x, float %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_olt_float:
+
+  ; M2:         c.olt.s   $f12, $f14
+  ; M2:         bc1f      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.s     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.s     $f0, $f12
+
+  ; CMOV-32:    c.olt.s   $f12, $f14
+  ; CMOV-32:    movt.s    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.s     $f0, $f14
+
+  ; SEL-32:     cmp.lt.s  $f0, $f12, $f14
+  ; SEL-32:     sel.s     $f0, $f14, $f12
+
+  ; CMOV-64:    c.olt.s   $f12, $f13
+  ; CMOV-64:    movt.s    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.s     $f0, $f13
+
+  ; SEL-64:     cmp.lt.s  $f0, $f12, $f13
+  ; SEL-64:     sel.s     $f0, $f13, $f12
+  %s = fcmp olt float %x, %y
+  %r = select i1 %s, float %x, float %y
+  ret float %r
+}
+
+define float @tst_select_fcmp_ole_float(float %x, float %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_ole_float:
+
+  ; M2:         c.ole.s   $f12, $f14
+  ; M2:         bc1f      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.s     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.s     $f0, $f12
+
+  ; CMOV-32:    c.ole.s   $f12, $f14
+  ; CMOV-32:    movt.s    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.s     $f0, $f14
+
+  ; SEL-32:     cmp.le.s  $f0, $f12, $f14
+  ; SEL-32:     sel.s     $f0, $f14, $f12
+
+  ; CMOV-64:    c.ole.s   $f12, $f13
+  ; CMOV-64:    movt.s    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.s     $f0, $f13
+
+  ; SEL-64:     cmp.le.s  $f0, $f12, $f13
+  ; SEL-64:     sel.s     $f0, $f13, $f12
+  %s = fcmp ole float %x, %y
+  %r = select i1 %s, float %x, float %y
+  ret float %r
+}
+
+define float @tst_select_fcmp_ogt_float(float %x, float %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_ogt_float:
+
+  ; M2:         c.ule.s   $f12, $f14
+  ; M2:         bc1t      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.s     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.s     $f0, $f12
+
+  ; CMOV-32:    c.ule.s   $f12, $f14
+  ; CMOV-32:    movf.s    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.s     $f0, $f14
+
+  ; SEL-32:     cmp.lt.s  $f0, $f14, $f12
+  ; SEL-32:     sel.s     $f0, $f14, $f12
+
+  ; CMOV-64:    c.ule.s   $f12, $f13
+  ; CMOV-64:    movf.s    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.s     $f0, $f13
+
+  ; SEL-64:     cmp.lt.s  $f0, $f13, $f12
+  ; SEL-64:     sel.s     $f0, $f13, $f12
+  %s = fcmp ogt float %x, %y
+  %r = select i1 %s, float %x, float %y
+  ret float %r
+}
+
+define float @tst_select_fcmp_oge_float(float %x, float %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_oge_float:
+
+  ; M2:         c.ult.s   $f12, $f14
+  ; M2:         bc1t      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.s     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.s     $f0, $f12
+
+  ; CMOV-32:    c.ult.s   $f12, $f14
+  ; CMOV-32:    movf.s    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.s     $f0, $f14
+
+  ; SEL-32:     cmp.le.s  $f0, $f14, $f12
+  ; SEL-32:     sel.s     $f0, $f14, $f12
+
+  ; CMOV-64:    c.ult.s   $f12, $f13
+  ; CMOV-64:    movf.s    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.s     $f0, $f13
+
+  ; SEL-64:     cmp.le.s  $f0, $f13, $f12
+  ; SEL-64:     sel.s     $f0, $f13, $f12
+  %s = fcmp oge float %x, %y
+  %r = select i1 %s, float %x, float %y
+  ret float %r
+}
+
+define float @tst_select_fcmp_oeq_float(float %x, float %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_oeq_float:
+
+  ; M2:         c.eq.s    $f12, $f14
+  ; M2:         bc1f      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.s     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.s     $f0, $f12
+
+  ; CMOV-32:    c.eq.s    $f12, $f14
+  ; CMOV-32:    movt.s    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.s     $f0, $f14
+
+  ; SEL-32:     cmp.eq.s  $f0, $f12, $f14
+  ; SEL-32:     sel.s     $f0, $f14, $f12
+
+  ; CMOV-64:    c.eq.s    $f12, $f13
+  ; CMOV-64:    movt.s    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.s     $f0, $f13
+
+  ; SEL-64:     cmp.eq.s  $f0, $f12, $f13
+  ; SEL-64:     sel.s     $f0, $f13, $f12
+  %s = fcmp oeq float %x, %y
+  %r = select i1 %s, float %x, float %y
+  ret float %r
+}
+
+define float @tst_select_fcmp_one_float(float %x, float %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_one_float:
+
+  ; M2:         c.ueq.s   $f12, $f14
+  ; M2:         bc1t      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.s     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.s     $f0, $f12
+
+  ; CMOV-32:    c.ueq.s   $f12, $f14
+  ; CMOV-32:    movf.s    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.s     $f0, $f14
+
+  ; SEL-32:     cmp.ueq.s $f0, $f12, $f14
+  ; SEL-32:     mfc1      $[[T0:[0-9]+]], $f0
+  ; SEL-32:     not       $[[T0]], $[[T0]]
+  ; SEL-32:     mtc1      $[[T0:[0-9]+]], $f0
+  ; SEL-32:     sel.s     $f0, $f14, $f12
+
+  ; CMOV-64:    c.ueq.s   $f12, $f13
+  ; CMOV-64:    movf.s    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.s     $f0, $f13
+
+  ; SEL-64:     cmp.ueq.s $f0, $f12, $f13
+  ; SEL-64:     mfc1      $[[T0:[0-9]+]], $f0
+  ; SEL-64:     not       $[[T0]], $[[T0]]
+  ; SEL-64:     mtc1      $[[T0:[0-9]+]], $f0
+  ; SEL-64:     sel.s     $f0, $f13, $f12
+
+  %s = fcmp one float %x, %y
+  %r = select i1 %s, float %x, float %y
+  ret float %r
+}
+
+define double @tst_select_fcmp_olt_double(double %x, double %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_olt_double:
+
+  ; M2:         c.olt.d   $f12, $f14
+  ; M2:         bc1f      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.d     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.d     $f0, $f12
+
+  ; CMOV-32:    c.olt.d   $f12, $f14
+  ; CMOV-32:    movt.d    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.d     $f0, $f14
+
+  ; SEL-32:     cmp.lt.d  $f0, $f12, $f14
+  ; SEL-32:     sel.d     $f0, $f14, $f12
+
+  ; CMOV-64:    c.olt.d   $f12, $f13
+  ; CMOV-64:    movt.d    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.d     $f0, $f13
+
+  ; SEL-64:     cmp.lt.d  $f0, $f12, $f13
+  ; SEL-64:     sel.d     $f0, $f13, $f12
+  %s = fcmp olt double %x, %y
+  %r = select i1 %s, double %x, double %y
+  ret double %r
+}
+
+define double @tst_select_fcmp_ole_double(double %x, double %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_ole_double:
+
+  ; M2:         c.ole.d   $f12, $f14
+  ; M2:         bc1f      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.d     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.d     $f0, $f12
+
+  ; CMOV-32:    c.ole.d   $f12, $f14
+  ; CMOV-32:    movt.d    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.d     $f0, $f14
+
+  ; SEL-32:     cmp.le.d  $f0, $f12, $f14
+  ; SEL-32:     sel.d     $f0, $f14, $f12
+
+  ; CMOV-64:    c.ole.d   $f12, $f13
+  ; CMOV-64:    movt.d    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.d     $f0, $f13
+
+  ; SEL-64:     cmp.le.d  $f0, $f12, $f13
+  ; SEL-64:     sel.d     $f0, $f13, $f12
+  %s = fcmp ole double %x, %y
+  %r = select i1 %s, double %x, double %y
+  ret double %r
+}
+
+define double @tst_select_fcmp_ogt_double(double %x, double %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_ogt_double:
+
+  ; M2:         c.ule.d   $f12, $f14
+  ; M2:         bc1t      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.d     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.d     $f0, $f12
+
+  ; CMOV-32:    c.ule.d   $f12, $f14
+  ; CMOV-32:    movf.d    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.d     $f0, $f14
+
+  ; SEL-32:     cmp.lt.d  $f0, $f14, $f12
+  ; SEL-32:     sel.d     $f0, $f14, $f12
+
+  ; CMOV-64:    c.ule.d   $f12, $f13
+  ; CMOV-64:    movf.d    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.d     $f0, $f13
+
+  ; SEL-64:     cmp.lt.d  $f0, $f13, $f12
+  ; SEL-64:     sel.d     $f0, $f13, $f12
+  %s = fcmp ogt double %x, %y
+  %r = select i1 %s, double %x, double %y
+  ret double %r
+}
+
+define double @tst_select_fcmp_oge_double(double %x, double %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_oge_double:
+
+  ; M2:         c.ult.d   $f12, $f14
+  ; M2:         bc1t      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.d     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.d     $f0, $f12
+
+  ; CMOV-32:    c.ult.d   $f12, $f14
+  ; CMOV-32:    movf.d    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.d     $f0, $f14
+
+  ; SEL-32:     cmp.le.d  $f0, $f14, $f12
+  ; SEL-32:     sel.d     $f0, $f14, $f12
+
+  ; CMOV-64:    c.ult.d   $f12, $f13
+  ; CMOV-64:    movf.d    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.d     $f0, $f13
+
+  ; SEL-64:     cmp.le.d  $f0, $f13, $f12
+  ; SEL-64:     sel.d     $f0, $f13, $f12
+  %s = fcmp oge double %x, %y
+  %r = select i1 %s, double %x, double %y
+  ret double %r
+}
+
+define double @tst_select_fcmp_oeq_double(double %x, double %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_oeq_double:
+
+  ; M2:         c.eq.d    $f12, $f14
+  ; M2:         bc1f      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.d     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.d     $f0, $f12
+
+  ; CMOV-32:    c.eq.d    $f12, $f14
+  ; CMOV-32:    movt.d    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.d     $f0, $f14
+
+  ; SEL-32:     cmp.eq.d  $f0, $f12, $f14
+  ; SEL-32:     sel.d     $f0, $f14, $f12
+
+  ; CMOV-64:    c.eq.d    $f12, $f13
+  ; CMOV-64:    movt.d    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.d     $f0, $f13
+
+  ; SEL-64:     cmp.eq.d  $f0, $f12, $f13
+  ; SEL-64:     sel.d     $f0, $f13, $f12
+  %s = fcmp oeq double %x, %y
+  %r = select i1 %s, double %x, double %y
+  ret double %r
+}
+
+define double @tst_select_fcmp_one_double(double %x, double %y) {
+entry:
+  ; ALL-LABEL: tst_select_fcmp_one_double:
+
+  ; M2:         c.ueq.d   $f12, $f14
+  ; M2:         bc1t      $[[BB0:BB[0-9_]+]]
+  ; M2:         nop
+  ; M2:         mov.d     $f12, $f14
+  ; M2:         $[[BB0]]:
+  ; M2:         jr        $ra
+  ; M2:         mov.d     $f0, $f12
+
+  ; CMOV-32:    c.ueq.d   $f12, $f14
+  ; CMOV-32:    movf.d    $f14, $f12, $fcc0
+  ; CMOV-32:    mov.d     $f0, $f14
+
+  ; SEL-32:     cmp.ueq.d $f0, $f12, $f14
+  ; SEL-32:     mfc1      $[[T0:[0-9]+]], $f0
+  ; SEL-32:     not       $[[T0]], $[[T0]]
+  ; SEL-32:     mtc1      $[[T0:[0-9]+]], $f0
+  ; SEL-32:     sel.d     $f0, $f14, $f12
+
+  ; CMOV-64:    c.ueq.d   $f12, $f13
+  ; CMOV-64:    movf.d    $f13, $f12, $fcc0
+  ; CMOV-64:    mov.d     $f0, $f13
+
+  ; SEL-64:     cmp.ueq.d $f0, $f12, $f13
+  ; SEL-64:     mfc1      $[[T0:[0-9]+]], $f0
+  ; SEL-64:     not       $[[T0]], $[[T0]]
+  ; SEL-64:     mtc1      $[[T0:[0-9]+]], $f0
+  ; SEL-64:     sel.d     $f0, $f13, $f12
+  %s = fcmp one double %x, %y
+  %r = select i1 %s, double %x, double %y
+  ret double %r
+}