GlobalISel: Handle llvm.roundeven

[arsenm]

Summary

I still think it's highly questionable that we have two intrinsics
with identical behavior and only vary by the name of the libcall used
if it happens to be lowered that way, but try to reduce the feature
delta between SDAG and GlobalISel for recently added intrinsics. I'm
not sure which opcode should be considered the canonical one, but
lower roundeven back to round.

[sepavloff]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

2461

I am afraid this change breaks the semantics of roundeven. Intrinsic round implements the same operation as libm function round (http://llvm.org/docs/LangRef.html#id554). In the latest draft (http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2478.pdf, 7.12.9.6p2) the function round is described:

The round functions round their argument to the nearest integer value in floating-point format,
rounding halfway cases away from zero, regardless of the current rounding direction.

In the same draft roundeven is described (7.12.9.8p2):

The roundeven functions round their argument to the nearest integer value in floating-point format,
rounding halfway cases to even (that is, to the nearest value that is an even integer), regardless of
the current rounding direction.

round and roundeven implement different rounding modes. Both functions do not depend on the current rounding mode.

LIBC variants provide roundeven but these implementations look complicated. There must be an algorithm which uses trunc and implements roundeven suitable for vector operations. Probably the algorithm that uses remainder in https://stackoverflow.com/questions/32746523/ieee-754-compliant-round-half-to-even can be used to implement roundeven similar to round in LegalizerHelper::lowerIntrinsicRound.

[paquette]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

2460

typo: operatio

[arsenm]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

2461

Oh right, this should have been G_FRINT. I can't keep all the rounding functions straight

[arsenm]

Expand to G_FRINT

[sepavloff]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

2461

Why not adding new target opcode, like G_FROUNDEVEN or G_INTRINSIC_ROUNDEVEN?

[arsenm]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

2461

That is added, and this is where it's lowered to the other equivalent operation

[sepavloff]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

2461

Indeed :)
What is the purpose of adding such replacement? In general, it is invalid, as it may be applied only in default FP mode. Using library calls is a safer solution. IIUC, if some target is required to support roundeven it must either implement custom lowering or appropriate libc must be used. Are there any reasons to allow such incomplete solution?

[arsenm]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

2461

AMDGPU has no library calls and does have the instruction. The unconstrained FP operations are defined as running in the default FP mode. It's quite undefined to execute them in another mode, these are equivalent operations.

[sepavloff]

The patch looks good to me, however I am not experienced in GlobalISel. Someone more familiar with it should look at this patch.

[arsenm]

Rebase

[paquette]

LGTM

[arsenm]

0da582d9b6cc8a96682628ff90b5e0e5d155358c

llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h

private:
 
   /// Perform generic multiplication of values held in multiple registers.
   /// Generated instructions use only types NarrowTy and i1.
   /// Destination can be same or two times size of the source.
   void multiplyRegisters(SmallVectorImpl<Register> &DstRegs,
                          ArrayRef<Register> Src1Regs,
                          ArrayRef<Register> Src2Regs, LLT NarrowTy);
 
+  void changeOpcode(MachineInstr &MI, unsigned NewOpcode);
+
 public:
   LegalizeResult fewerElementsVectorImplicitDef(MachineInstr &MI,
                                                 unsigned TypeIdx, LLT NarrowTy);
 
   /// Legalize a instruction with a vector type where each operand may have a
   /// different element type. All type indexes must have the same number of
   /// elements.
   LegalizeResult fewerElementsVectorMultiEltType(MachineInstr &MI,

llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h

public:
   LegalizeRuleSet &lowerForCartesianProduct(std::initializer_list<LLT> Types0,
                                             std::initializer_list<LLT> Types1,
                                             std::initializer_list<LLT> Types2) {
     using namespace LegalityPredicates;
     return actionForCartesianProduct(LegalizeAction::Lower, Types0, Types1,
                                      Types2);
   }
 
+  /// The instruction is emitted as a library call.
+  LegalizeRuleSet &libcall() {
+    using namespace LegalizeMutations;
+    // We have no choice but conservatively assume that predicate-less lowering
+    // properly handles all type indices by design:
+    markAllIdxsAsCovered();
+    return actionIf(LegalizeAction::Libcall, always);
+  }
+
   /// Like legalIf, but for the Libcall action.
   LegalizeRuleSet &libcallIf(LegalityPredicate Predicate) {
     // We have no choice but conservatively assume that a libcall with a
     // free-form user provided Predicate properly handles all type indices:
     markAllIdxsAsCovered();
     return actionIf(LegalizeAction::Libcall, Predicate);
   }
   LegalizeRuleSet &libcallFor(std::initializer_list<LLT> Types) {

llvm/include/llvm/Support/TargetOpcodes.def

 HANDLE_TARGET_OPCODE(G_FREEZE)
 
 /// INTRINSIC trunc intrinsic.
 HANDLE_TARGET_OPCODE(G_INTRINSIC_TRUNC)
 
 /// INTRINSIC round intrinsic.
 HANDLE_TARGET_OPCODE(G_INTRINSIC_ROUND)
 
+/// INTRINSIC roundeven intrinsic.
+HANDLE_TARGET_OPCODE(G_INTRINSIC_ROUNDEVEN)
+
 /// INTRINSIC readcyclecounter
 HANDLE_TARGET_OPCODE(G_READCYCLECOUNTER)
 
 /// Generic load (including anyext load)
 HANDLE_TARGET_OPCODE(G_LOAD)
 
 /// Generic signext load
 HANDLE_TARGET_OPCODE(G_SEXTLOAD)

llvm/include/llvm/Target/GenericOpcodes.td

 }
 
 def G_INTRINSIC_ROUND : GenericInstruction {
   let OutOperandList = (outs type0:$dst);
   let InOperandList = (ins type0:$src1);
   let hasSideEffects = 0;
 }
 
+def G_INTRINSIC_ROUNDEVEN : GenericInstruction {
+  let OutOperandList = (outs type0:$dst);
+  let InOperandList = (ins type0:$src1);
+  let hasSideEffects = 0;
+}
+
 def G_READCYCLECOUNTER : GenericInstruction {
   let OutOperandList = (outs type0:$dst);
   let InOperandList = (ins);
   let hasSideEffects = 1;
 }
 
 //------------------------------------------------------------------------------
 // Memory ops

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

switch (ID) {
     case Intrinsic::nearbyint:
       return TargetOpcode::G_FNEARBYINT;
     case Intrinsic::pow:
       return TargetOpcode::G_FPOW;
     case Intrinsic::rint:
       return TargetOpcode::G_FRINT;
     case Intrinsic::round:
       return TargetOpcode::G_INTRINSIC_ROUND;
+    case Intrinsic::roundeven:
+      return TargetOpcode::G_INTRINSIC_ROUNDEVEN;
     case Intrinsic::sin:
       return TargetOpcode::G_FSIN;
     case Intrinsic::sqrt:
       return TargetOpcode::G_FSQRT;
     case Intrinsic::trunc:
       return TargetOpcode::G_INTRINSIC_TRUNC;
     case Intrinsic::readcyclecounter:
       return TargetOpcode::G_READCYCLECOUNTER;

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

static Type *getFloatTypeForLLT(LLVMContext &Ctx, LLT Ty) {
 
   switch (Ty.getSizeInBits()) {
   case 16:
     return Type::getHalfTy(Ctx);
   case 32:
     return Type::getFloatTy(Ctx);
   case 64:
     return Type::getDoubleTy(Ctx);
+  case 80:
+    return Type::getX86_FP80Ty(Ctx);
   case 128:
     return Type::getFP128Ty(Ctx);
   default:
     return nullptr;
   }
 }
 
 LegalizerHelper::LegalizerHelper(MachineFunction &MF,
if (LCMTy.isVector()) {
     MIRBuilder.buildExtract(DstReg, Remerge, 0);
     return;
   }
 
   llvm_unreachable("unhandled case");
 }
 
 static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) {
-#define RTLIBCASE(LibcallPrefix)                                               \
+#define RTLIBCASE_INT(LibcallPrefix)                                           \
   do {                                                                         \
     switch (Size) {                                                            \
     case 32:                                                                   \
       return RTLIB::LibcallPrefix##32;                                         \
     case 64:                                                                   \
       return RTLIB::LibcallPrefix##64;                                         \
     case 128:                                                                  \
       return RTLIB::LibcallPrefix##128;                                        \
     default:                                                                   \
       llvm_unreachable("unexpected size");                                     \
     }                                                                          \
   } while (0)
 
-  assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
+#define RTLIBCASE(LibcallPrefix)                                               \
+  do {                                                                         \
+    switch (Size) {                                                            \
+    case 32:                                                                   \
+      return RTLIB::LibcallPrefix##32;                                         \
+    case 64:                                                                   \
+      return RTLIB::LibcallPrefix##64;                                         \
+    case 80:                                                                   \
+      return RTLIB::LibcallPrefix##80;                                         \
+    case 128:                                                                  \
+      return RTLIB::LibcallPrefix##128;                                        \
+    default:                                                                   \
+      llvm_unreachable("unexpected size");                                     \
+    }                                                                          \
+  } while (0)
 
   switch (Opcode) {
   case TargetOpcode::G_SDIV:
-    RTLIBCASE(SDIV_I);
+    RTLIBCASE_INT(SDIV_I);
   case TargetOpcode::G_UDIV:
-    RTLIBCASE(UDIV_I);
+    RTLIBCASE_INT(UDIV_I);
   case TargetOpcode::G_SREM:
-    RTLIBCASE(SREM_I);
+    RTLIBCASE_INT(SREM_I);
   case TargetOpcode::G_UREM:
-    RTLIBCASE(UREM_I);
+    RTLIBCASE_INT(UREM_I);
   case TargetOpcode::G_CTLZ_ZERO_UNDEF:
-    RTLIBCASE(CTLZ_I);
+    RTLIBCASE_INT(CTLZ_I);
   case TargetOpcode::G_FADD:
     RTLIBCASE(ADD_F);
   case TargetOpcode::G_FSUB:
     RTLIBCASE(SUB_F);
   case TargetOpcode::G_FMUL:
     RTLIBCASE(MUL_F);
   case TargetOpcode::G_FDIV:
     RTLIBCASE(DIV_F);
#define RTLIBCASE(LibcallPrefix)                                               \
   case TargetOpcode::G_FMAXNUM:
     RTLIBCASE(FMAX_F);
   case TargetOpcode::G_FSQRT:
     RTLIBCASE(SQRT_F);
   case TargetOpcode::G_FRINT:
     RTLIBCASE(RINT_F);
   case TargetOpcode::G_FNEARBYINT:
     RTLIBCASE(NEARBYINT_F);
+  case TargetOpcode::G_INTRINSIC_ROUNDEVEN:
+    RTLIBCASE(ROUNDEVEN_F);
   }
   llvm_unreachable("Unknown libcall function");
 }
 
 /// True if an instruction is in tail position in its caller. Intended for
 /// legalizing libcalls as tail calls when possible.
 static bool isLibCallInTailPosition(MachineInstr &MI) {
   MachineBasicBlock &MBB = *MI.getParent();
LegalizerHelper::libcall(MachineInstr &MI) {
   case TargetOpcode::G_FEXP:
   case TargetOpcode::G_FEXP2:
   case TargetOpcode::G_FCEIL:
   case TargetOpcode::G_FFLOOR:
   case TargetOpcode::G_FMINNUM:
   case TargetOpcode::G_FMAXNUM:
   case TargetOpcode::G_FSQRT:
   case TargetOpcode::G_FRINT:
-  case TargetOpcode::G_FNEARBYINT: {
+  case TargetOpcode::G_FNEARBYINT:
+  case TargetOpcode::G_INTRINSIC_ROUNDEVEN: {
     Type *HLTy = getFloatTypeForLLT(Ctx, LLTy);
-    if (!HLTy || (Size != 32 && Size != 64 && Size != 128)) {
-      LLVM_DEBUG(dbgs() << "No libcall available for size " << Size << ".\n");
+    if (!HLTy) {
+      LLVM_DEBUG(dbgs() << "No libcall available for type " << LLTy << ".\n");
       return UnableToLegalize;
     }
     auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy);
     if (Status != Legalized)
       return Status;
     break;
   }
   case TargetOpcode::G_FPEXT:
LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
   case TargetOpcode::G_FRINT:
   case TargetOpcode::G_FNEARBYINT:
   case TargetOpcode::G_FSQRT:
   case TargetOpcode::G_FEXP:
   case TargetOpcode::G_FEXP2:
   case TargetOpcode::G_FPOW:
   case TargetOpcode::G_INTRINSIC_TRUNC:
   case TargetOpcode::G_INTRINSIC_ROUND:
+  case TargetOpcode::G_INTRINSIC_ROUNDEVEN:
     assert(TypeIdx == 0);
     Observer.changingInstr(MI);
 
     for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I)
       widenScalarSrc(MI, WideTy, I, TargetOpcode::G_FPEXT);
 
     widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
     Observer.changedInstr(MI);
case TargetOpcode::G_XOR: {
     Observer.changedInstr(MI);
     return Legalized;
   }
   default:
     return UnableToLegalize;
   }
 }
 
+// Legalize an instruction by changing the opcode in place.
+void LegalizerHelper::changeOpcode(MachineInstr &MI, unsigned NewOpcode) {
+    Observer.changingInstr(MI);
+    MI.setDesc(MIRBuilder.getTII().get(NewOpcode));
+    Observer.changedInstr(MI);
+}
+
 LegalizerHelper::LegalizeResult
 LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
   using namespace TargetOpcode;
 
   switch(MI.getOpcode()) {
   default:
     return UnableToLegalize;
   case TargetOpcode::G_BITCAST:
case TargetOpcode::G_FSUB: {
     return Legalized;
   }
   case TargetOpcode::G_FMAD:
     return lowerFMad(MI);
   case TargetOpcode::G_FFLOOR:
     return lowerFFloor(MI);
   case TargetOpcode::G_INTRINSIC_ROUND:
     return lowerIntrinsicRound(MI);
+  case TargetOpcode::G_INTRINSIC_ROUNDEVEN: {
+    // Since round even is the assumed rounding mode for unconstrained FP
+    // operations, round and roundeven are the same operatio.
+    changeOpcode(MI, TargetOpcode::G_INTRINSIC_ROUND);
+    return Legalized;
+  }
   case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: {
     Register OldValRes = MI.getOperand(0).getReg();
     Register SuccessRes = MI.getOperand(1).getReg();
     Register Addr = MI.getOperand(2).getReg();
     Register CmpVal = MI.getOperand(3).getReg();
     Register NewVal = MI.getOperand(4).getReg();
     MIRBuilder.buildAtomicCmpXchg(OldValRes, Addr, CmpVal, NewVal,
                                   **MI.memoperands_begin());
LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx,
   case G_FLOG:
   case G_FLOG2:
   case G_FLOG10:
   case G_FNEARBYINT:
   case G_FCEIL:
   case G_FFLOOR:
   case G_FRINT:
   case G_INTRINSIC_ROUND:
+  case G_INTRINSIC_ROUNDEVEN:
   case G_INTRINSIC_TRUNC:
   case G_FCOS:
   case G_FSIN:
   case G_FSQRT:
   case G_BSWAP:
   case G_BITREVERSE:
   case G_SDIV:
   case G_UDIV:

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

if (ST.has16BitInsts())
     FPToI.legalFor({{S16, S16}});
   else
     FPToI.minScalar(1, S32);
 
   FPToI.minScalar(0, S32)
        .scalarize(0)
        .lower();
 
-  getActionDefinitionsBuilder(G_INTRINSIC_ROUND)
+  getActionDefinitionsBuilder({G_INTRINSIC_ROUND, G_INTRINSIC_ROUNDEVEN})
     .scalarize(0)
     .lower();
 
   if (ST.has16BitInsts()) {
     getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_FCEIL, G_FRINT})
       .legalFor({S16, S32, S64})
       .clampScalar(0, S16, S64)
       .scalarize(0);

llvm/lib/Target/X86/X86LegalizerInfo.cpp

X86LegalizerInfo::X86LegalizerInfo(const X86Subtarget &STI,
   setLegalizerInfoSSE2();
   setLegalizerInfoSSE41();
   setLegalizerInfoAVX();
   setLegalizerInfoAVX2();
   setLegalizerInfoAVX512();
   setLegalizerInfoAVX512DQ();
   setLegalizerInfoAVX512BW();
 
+  getActionDefinitionsBuilder(G_INTRINSIC_ROUNDEVEN)
+    .scalarize(0)
+    .minScalar(0, LLT::scalar(32))
+    .libcall();
+
   setLegalizeScalarToDifferentSizeStrategy(G_PHI, 0, widen_1);
   for (unsigned BinOp : {G_SUB, G_MUL, G_AND, G_OR, G_XOR})
     setLegalizeScalarToDifferentSizeStrategy(BinOp, 0, widen_1);
   for (unsigned MemOp : {G_LOAD, G_STORE})
     setLegalizeScalarToDifferentSizeStrategy(MemOp, 0,
        narrowToSmallerAndWidenToSmallest);
   setLegalizeScalarToDifferentSizeStrategy(
       G_PTR_ADD, 1, widenToLargerTypesUnsupportedOtherwise);

llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir

 # DEBUG-NEXT: G_INTRINSIC_TRUNC (opcode {{[0-9]+}}): 1 type index, 0 imm indices
 # DEBUG-NEXT: .. opcode {{[0-9]+}} is aliased to {{[0-9]+}}
 # DEBUG-NEXT: .. type index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: G_INTRINSIC_ROUND (opcode {{[0-9]+}}): 1 type index, 0 imm indices
 # DEBUG-NEXT: .. opcode {{[0-9]+}} is aliased to {{[0-9]+}}
 # DEBUG-NEXT: .. type index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
-
+# DEBUG-NEXT: G_INTRINSIC_ROUNDEVEN (opcode {{[0-9]+}}): 1 type index, 0 imm indices
+# DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
+# DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: G_READCYCLECOUNTER (opcode {{[0-9]+}}): 1 type index, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: no rules defined
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: no rules defined
 
 # DEBUG-NEXT: G_LOAD (opcode {{[0-9]+}}): 2 type indices, 0 imm indices
 # DEBUG-NEXT: .. type index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: G_SEXTLOAD (opcode {{[0-9]+}}): 2 type indices, 0 imm indices

llvm/test/CodeGen/AMDGPU/GlobalISel/roundeven.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=fiji < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s
+
+define float @v_roundeven_f32(float %x) {
+; GFX6-LABEL: v_roundeven_f32:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_trunc_f32_e32 v1, v0
+; GFX6-NEXT:    v_sub_f32_e32 v2, v0, v1
+; GFX6-NEXT:    v_and_b32_e32 v0, 0x80000000, v0
+; GFX6-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v2|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX6-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_f32:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_trunc_f32_e32 v1, v0
+; GFX7-NEXT:    v_sub_f32_e32 v2, v0, v1
+; GFX7-NEXT:    v_and_b32_e32 v0, 0x80000000, v0
+; GFX7-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v2|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX7-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_f32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f32_e32 v1, v0
+; GFX8-NEXT:    v_sub_f32_e32 v2, v0, v1
+; GFX8-NEXT:    v_and_b32_e32 v0, 0x80000000, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v2|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX8-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_f32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f32_e32 v1, v0
+; GFX9-NEXT:    v_sub_f32_e32 v2, v0, v1
+; GFX9-NEXT:    v_bfrev_b32_e32 v3, 1
+; GFX9-NEXT:    v_and_or_b32 v0, v0, v3, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v2|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX9-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %roundeven = call float @llvm.roundeven.f32(float %x)
+  ret float %roundeven
+}
+
+define <2 x float> @v_roundeven_v2f32(<2 x float> %x) {
+; GFX6-LABEL: v_roundeven_v2f32:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_trunc_f32_e32 v2, v0
+; GFX6-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX6-NEXT:    v_sub_f32_e32 v3, v0, v2
+; GFX6-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX6-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX6-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX6-NEXT:    v_trunc_f32_e32 v2, v1
+; GFX6-NEXT:    v_sub_f32_e32 v3, v1, v2
+; GFX6-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX6-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX6-NEXT:    v_add_f32_e32 v1, v2, v1
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_v2f32:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_trunc_f32_e32 v2, v0
+; GFX7-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX7-NEXT:    v_sub_f32_e32 v3, v0, v2
+; GFX7-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX7-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX7-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX7-NEXT:    v_trunc_f32_e32 v2, v1
+; GFX7-NEXT:    v_sub_f32_e32 v3, v1, v2
+; GFX7-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX7-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX7-NEXT:    v_add_f32_e32 v1, v2, v1
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_v2f32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f32_e32 v2, v0
+; GFX8-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX8-NEXT:    v_sub_f32_e32 v3, v0, v2
+; GFX8-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX8-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX8-NEXT:    v_trunc_f32_e32 v2, v1
+; GFX8-NEXT:    v_sub_f32_e32 v3, v1, v2
+; GFX8-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX8-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX8-NEXT:    v_add_f32_e32 v1, v2, v1
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_v2f32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f32_e32 v2, v0
+; GFX9-NEXT:    v_sub_f32_e32 v3, v0, v2
+; GFX9-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX9-NEXT:    v_and_or_b32 v0, v0, s4, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX9-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX9-NEXT:    v_trunc_f32_e32 v2, v1
+; GFX9-NEXT:    v_sub_f32_e32 v3, v1, v2
+; GFX9-NEXT:    v_and_or_b32 v1, v1, s4, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v3|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX9-NEXT:    v_add_f32_e32 v1, v2, v1
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %roundeven = call <2 x float> @llvm.roundeven.v2f32(<2 x float> %x)
+  ret <2 x float> %roundeven
+}
+
+define <3 x float> @v_roundeven_v3f32(<3 x float> %x) {
+; GFX6-LABEL: v_roundeven_v3f32:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_trunc_f32_e32 v3, v0
+; GFX6-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX6-NEXT:    v_sub_f32_e32 v4, v0, v3
+; GFX6-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX6-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX6-NEXT:    v_add_f32_e32 v0, v3, v0
+; GFX6-NEXT:    v_trunc_f32_e32 v3, v1
+; GFX6-NEXT:    v_sub_f32_e32 v4, v1, v3
+; GFX6-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX6-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX6-NEXT:    v_add_f32_e32 v1, v3, v1
+; GFX6-NEXT:    v_trunc_f32_e32 v3, v2
+; GFX6-NEXT:    v_sub_f32_e32 v4, v2, v3
+; GFX6-NEXT:    v_and_b32_e32 v2, s4, v2
+; GFX6-NEXT:    v_or_b32_e32 v2, 1.0, v2
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX6-NEXT:    v_add_f32_e32 v2, v3, v2
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_v3f32:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_trunc_f32_e32 v3, v0
+; GFX7-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX7-NEXT:    v_sub_f32_e32 v4, v0, v3
+; GFX7-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX7-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX7-NEXT:    v_add_f32_e32 v0, v3, v0
+; GFX7-NEXT:    v_trunc_f32_e32 v3, v1
+; GFX7-NEXT:    v_sub_f32_e32 v4, v1, v3
+; GFX7-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX7-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX7-NEXT:    v_add_f32_e32 v1, v3, v1
+; GFX7-NEXT:    v_trunc_f32_e32 v3, v2
+; GFX7-NEXT:    v_sub_f32_e32 v4, v2, v3
+; GFX7-NEXT:    v_and_b32_e32 v2, s4, v2
+; GFX7-NEXT:    v_or_b32_e32 v2, 1.0, v2
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX7-NEXT:    v_add_f32_e32 v2, v3, v2
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_v3f32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f32_e32 v3, v0
+; GFX8-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX8-NEXT:    v_sub_f32_e32 v4, v0, v3
+; GFX8-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX8-NEXT:    v_add_f32_e32 v0, v3, v0
+; GFX8-NEXT:    v_trunc_f32_e32 v3, v1
+; GFX8-NEXT:    v_sub_f32_e32 v4, v1, v3
+; GFX8-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX8-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX8-NEXT:    v_add_f32_e32 v1, v3, v1
+; GFX8-NEXT:    v_trunc_f32_e32 v3, v2
+; GFX8-NEXT:    v_sub_f32_e32 v4, v2, v3
+; GFX8-NEXT:    v_and_b32_e32 v2, s4, v2
+; GFX8-NEXT:    v_or_b32_e32 v2, 1.0, v2
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX8-NEXT:    v_add_f32_e32 v2, v3, v2
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_v3f32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f32_e32 v3, v0
+; GFX9-NEXT:    v_sub_f32_e32 v4, v0, v3
+; GFX9-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX9-NEXT:    v_and_or_b32 v0, v0, s4, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX9-NEXT:    v_add_f32_e32 v0, v3, v0
+; GFX9-NEXT:    v_trunc_f32_e32 v3, v1
+; GFX9-NEXT:    v_sub_f32_e32 v4, v1, v3
+; GFX9-NEXT:    v_and_or_b32 v1, v1, s4, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX9-NEXT:    v_add_f32_e32 v1, v3, v1
+; GFX9-NEXT:    v_trunc_f32_e32 v3, v2
+; GFX9-NEXT:    v_sub_f32_e32 v4, v2, v3
+; GFX9-NEXT:    v_and_or_b32 v2, v2, s4, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v4|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX9-NEXT:    v_add_f32_e32 v2, v3, v2
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %roundeven = call <3 x float> @llvm.roundeven.v3f32(<3 x float> %x)
+  ret <3 x float> %roundeven
+}
+
+define <4 x float> @v_roundeven_v4f32(<4 x float> %x) {
+; GFX6-LABEL: v_roundeven_v4f32:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_trunc_f32_e32 v4, v0
+; GFX6-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX6-NEXT:    v_sub_f32_e32 v5, v0, v4
+; GFX6-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX6-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX6-NEXT:    v_add_f32_e32 v0, v4, v0
+; GFX6-NEXT:    v_trunc_f32_e32 v4, v1
+; GFX6-NEXT:    v_sub_f32_e32 v5, v1, v4
+; GFX6-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX6-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX6-NEXT:    v_add_f32_e32 v1, v4, v1
+; GFX6-NEXT:    v_trunc_f32_e32 v4, v2
+; GFX6-NEXT:    v_sub_f32_e32 v5, v2, v4
+; GFX6-NEXT:    v_and_b32_e32 v2, s4, v2
+; GFX6-NEXT:    v_or_b32_e32 v2, 1.0, v2
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX6-NEXT:    v_add_f32_e32 v2, v4, v2
+; GFX6-NEXT:    v_trunc_f32_e32 v4, v3
+; GFX6-NEXT:    v_sub_f32_e32 v5, v3, v4
+; GFX6-NEXT:    v_and_b32_e32 v3, 0x80000000, v3
+; GFX6-NEXT:    v_or_b32_e32 v3, 1.0, v3
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v3, 0, v3, vcc
+; GFX6-NEXT:    v_add_f32_e32 v3, v4, v3
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_v4f32:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_trunc_f32_e32 v4, v0
+; GFX7-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX7-NEXT:    v_sub_f32_e32 v5, v0, v4
+; GFX7-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX7-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX7-NEXT:    v_add_f32_e32 v0, v4, v0
+; GFX7-NEXT:    v_trunc_f32_e32 v4, v1
+; GFX7-NEXT:    v_sub_f32_e32 v5, v1, v4
+; GFX7-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX7-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX7-NEXT:    v_add_f32_e32 v1, v4, v1
+; GFX7-NEXT:    v_trunc_f32_e32 v4, v2
+; GFX7-NEXT:    v_sub_f32_e32 v5, v2, v4
+; GFX7-NEXT:    v_and_b32_e32 v2, s4, v2
+; GFX7-NEXT:    v_or_b32_e32 v2, 1.0, v2
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX7-NEXT:    v_add_f32_e32 v2, v4, v2
+; GFX7-NEXT:    v_trunc_f32_e32 v4, v3
+; GFX7-NEXT:    v_sub_f32_e32 v5, v3, v4
+; GFX7-NEXT:    v_and_b32_e32 v3, 0x80000000, v3
+; GFX7-NEXT:    v_or_b32_e32 v3, 1.0, v3
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v3, 0, v3, vcc
+; GFX7-NEXT:    v_add_f32_e32 v3, v4, v3
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_v4f32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f32_e32 v4, v0
+; GFX8-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX8-NEXT:    v_sub_f32_e32 v5, v0, v4
+; GFX8-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, 1.0, v0
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX8-NEXT:    v_add_f32_e32 v0, v4, v0
+; GFX8-NEXT:    v_trunc_f32_e32 v4, v1
+; GFX8-NEXT:    v_sub_f32_e32 v5, v1, v4
+; GFX8-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX8-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX8-NEXT:    v_add_f32_e32 v1, v4, v1
+; GFX8-NEXT:    v_trunc_f32_e32 v4, v2
+; GFX8-NEXT:    v_sub_f32_e32 v5, v2, v4
+; GFX8-NEXT:    v_and_b32_e32 v2, s4, v2
+; GFX8-NEXT:    v_or_b32_e32 v2, 1.0, v2
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX8-NEXT:    v_add_f32_e32 v2, v4, v2
+; GFX8-NEXT:    v_trunc_f32_e32 v4, v3
+; GFX8-NEXT:    v_sub_f32_e32 v5, v3, v4
+; GFX8-NEXT:    v_and_b32_e32 v3, 0x80000000, v3
+; GFX8-NEXT:    v_or_b32_e32 v3, 1.0, v3
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, 0, v3, vcc
+; GFX8-NEXT:    v_add_f32_e32 v3, v4, v3
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_v4f32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f32_e32 v4, v0
+; GFX9-NEXT:    v_sub_f32_e32 v5, v0, v4
+; GFX9-NEXT:    s_mov_b32 s4, 0x80000000
+; GFX9-NEXT:    v_and_or_b32 v0, v0, s4, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX9-NEXT:    v_add_f32_e32 v0, v4, v0
+; GFX9-NEXT:    v_trunc_f32_e32 v4, v1
+; GFX9-NEXT:    v_sub_f32_e32 v5, v1, v4
+; GFX9-NEXT:    v_and_or_b32 v1, v1, s4, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX9-NEXT:    v_add_f32_e32 v1, v4, v1
+; GFX9-NEXT:    v_trunc_f32_e32 v4, v2
+; GFX9-NEXT:    v_sub_f32_e32 v5, v2, v4
+; GFX9-NEXT:    v_and_or_b32 v2, v2, s4, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX9-NEXT:    v_add_f32_e32 v2, v4, v2
+; GFX9-NEXT:    v_trunc_f32_e32 v4, v3
+; GFX9-NEXT:    v_sub_f32_e32 v5, v3, v4
+; GFX9-NEXT:    v_mov_b32_e32 v6, 0x80000000
+; GFX9-NEXT:    v_and_or_b32 v3, v3, v6, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v5|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, 0, v3, vcc
+; GFX9-NEXT:    v_add_f32_e32 v3, v4, v3
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %roundeven = call <4 x float> @llvm.roundeven.v4f32(<4 x float> %x)
+  ret <4 x float> %roundeven
+}
+
+define half @v_roundeven_f16(half %x) {
+; GFX6-LABEL: v_roundeven_f16:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v0
+; GFX6-NEXT:    v_and_b32_e32 v0, 0x8000, v0
+; GFX6-NEXT:    v_or_b32_e32 v0, 0x3c00, v0
+; GFX6-NEXT:    v_trunc_f32_e32 v2, v1
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v3, -v2
+; GFX6-NEXT:    v_add_f32_e32 v1, v1, v3
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, 0.5
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v1, |v1|
+; GFX6-NEXT:    v_cmp_ge_f32_e32 vcc, v1, v3
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_f16:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v0
+; GFX7-NEXT:    v_and_b32_e32 v0, 0x8000, v0
+; GFX7-NEXT:    v_or_b32_e32 v0, 0x3c00, v0
+; GFX7-NEXT:    v_trunc_f32_e32 v2, v1
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v3, -v2
+; GFX7-NEXT:    v_add_f32_e32 v1, v1, v3
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, 0.5
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v1, |v1|
+; GFX7-NEXT:    v_cmp_ge_f32_e32 vcc, v1, v3
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_f16:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f16_e32 v1, v0
+; GFX8-NEXT:    v_add_f16_e64 v2, v0, -v1
+; GFX8-NEXT:    v_and_b32_e32 v0, 0x8000, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, 0x3c00, v0
+; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v2|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX8-NEXT:    v_add_f16_e32 v0, v1, v0
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_f16:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f16_e32 v1, v0
+; GFX9-NEXT:    v_add_f16_e64 v2, v0, -v1
+; GFX9-NEXT:    v_and_b32_e32 v0, 0x8000, v0
+; GFX9-NEXT:    v_or_b32_e32 v0, 0x3c00, v0
+; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v2|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX9-NEXT:    v_add_f16_e32 v0, v1, v0
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %roundeven = call half @llvm.roundeven.f16(half %x)
+  ret half %roundeven
+}
+
+define <2 x half> @v_roundeven_v2f16(<2 x half> %x) {
+; GFX6-LABEL: v_roundeven_v2f16:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v0
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v1
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v7, 0.5
+; GFX6-NEXT:    s_mov_b32 s4, 0x8000
+; GFX6-NEXT:    v_trunc_f32_e32 v3, v2
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT:    v_trunc_f32_e32 v6, v4
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v5, -v3
+; GFX6-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX6-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX6-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX6-NEXT:    v_add_f32_e32 v2, v2, v5
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v5, -v6
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT:    v_add_f32_e32 v4, v4, v5
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v2, |v2|
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT:    v_cmp_ge_f32_e32 vcc, v2, v7
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v2, |v4|
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT:    v_cmp_ge_f32_e32 vcc, v2, v7
+; GFX6-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v6
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT:    v_add_f32_e32 v0, v3, v0
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT:    v_add_f32_e32 v1, v2, v1
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_v2f16:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v0
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v1
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v7, 0.5
+; GFX7-NEXT:    s_mov_b32 s4, 0x8000
+; GFX7-NEXT:    v_trunc_f32_e32 v3, v2
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT:    v_trunc_f32_e32 v6, v4
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v5, -v3
+; GFX7-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX7-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX7-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX7-NEXT:    v_add_f32_e32 v2, v2, v5
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v5, -v6
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT:    v_add_f32_e32 v4, v4, v5
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v2, |v2|
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT:    v_cmp_ge_f32_e32 vcc, v2, v7
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v2, |v4|
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT:    v_cmp_ge_f32_e32 vcc, v2, v7
+; GFX7-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v6
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT:    v_add_f32_e32 v0, v3, v0
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT:    v_add_f32_e32 v1, v2, v1
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_v2f16:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f16_e32 v2, v0
+; GFX8-NEXT:    s_mov_b32 s4, 0x8000
+; GFX8-NEXT:    v_add_f16_e64 v3, v0, -v2
+; GFX8-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
+; GFX8-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX8-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX8-NEXT:    v_add_f16_e32 v0, v2, v0
+; GFX8-NEXT:    v_trunc_f16_e32 v2, v1
+; GFX8-NEXT:    v_add_f16_e64 v3, v1, -v2
+; GFX8-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX8-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX8-NEXT:    v_add_f16_e32 v1, v2, v1
+; GFX8-NEXT:    v_mov_b32_e32 v2, 16
+; GFX8-NEXT:    v_lshlrev_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_v2f16:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f16_e32 v2, v0
+; GFX9-NEXT:    s_mov_b32 s4, 0x8000
+; GFX9-NEXT:    v_add_f16_e64 v3, v0, -v2
+; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
+; GFX9-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX9-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX9-NEXT:    v_add_f16_e32 v0, v2, v0
+; GFX9-NEXT:    v_trunc_f16_e32 v2, v1
+; GFX9-NEXT:    v_add_f16_e64 v3, v1, -v2
+; GFX9-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX9-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX9-NEXT:    v_add_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0xffff
+; GFX9-NEXT:    v_and_or_b32 v0, v0, v2, v1
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %roundeven = call <2 x half> @llvm.roundeven.v2f16(<2 x half> %x)
+  ret <2 x half> %roundeven
+}
+
+define <2 x half> @v_roundeven_v2f16_fneg(<2 x half> %x) {
+; GFX6-LABEL: v_roundeven_v2f16_fneg:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX6-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX6-NEXT:    v_or_b32_e32 v0, v1, v0
+; GFX6-NEXT:    v_xor_b32_e32 v0, 0x80008000, v0
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v0
+; GFX6-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v3
+; GFX6-NEXT:    s_mov_b32 s4, 0x8000
+; GFX6-NEXT:    v_trunc_f32_e32 v2, v1
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT:    v_trunc_f32_e32 v6, v4
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v5, -v2
+; GFX6-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v7, -v6
+; GFX6-NEXT:    v_and_b32_e32 v3, s4, v3
+; GFX6-NEXT:    v_add_f32_e32 v1, v1, v5
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT:    v_add_f32_e32 v4, v4, v7
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v5, 0.5
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v1, |v1|
+; GFX6-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX6-NEXT:    v_or_b32_e32 v3, s5, v3
+; GFX6-NEXT:    v_cmp_ge_f32_e32 vcc, v1, v5
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v2
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v2, |v4|
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT:    v_cmp_ge_f32_e32 vcc, v2, v5
+; GFX6-NEXT:    v_cndmask_b32_e32 v2, 0, v3, vcc
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v6
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT:    v_add_f32_e32 v1, v3, v2
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_v2f16_fneg:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX7-NEXT:    v_or_b32_e32 v0, v1, v0
+; GFX7-NEXT:    v_xor_b32_e32 v0, 0x80008000, v0
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v0
+; GFX7-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v3
+; GFX7-NEXT:    s_mov_b32 s4, 0x8000
+; GFX7-NEXT:    v_trunc_f32_e32 v2, v1
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT:    v_trunc_f32_e32 v6, v4
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v5, -v2
+; GFX7-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v7, -v6
+; GFX7-NEXT:    v_and_b32_e32 v3, s4, v3
+; GFX7-NEXT:    v_add_f32_e32 v1, v1, v5
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT:    v_add_f32_e32 v4, v4, v7
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v5, 0.5
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v1, |v1|
+; GFX7-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX7-NEXT:    v_or_b32_e32 v3, s5, v3
+; GFX7-NEXT:    v_cmp_ge_f32_e32 vcc, v1, v5
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v2
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v2, |v4|
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT:    v_cmp_ge_f32_e32 vcc, v2, v5
+; GFX7-NEXT:    v_cndmask_b32_e32 v2, 0, v3, vcc
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v6
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT:    v_add_f32_e32 v0, v1, v0
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT:    v_add_f32_e32 v1, v3, v2
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_v2f16_fneg:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_xor_b32_e32 v0, 0x80008000, v0
+; GFX8-NEXT:    v_trunc_f16_e32 v2, v0
+; GFX8-NEXT:    s_mov_b32 s4, 0x8000
+; GFX8-NEXT:    v_add_f16_e64 v3, v0, -v2
+; GFX8-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
+; GFX8-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX8-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX8-NEXT:    v_add_f16_e32 v0, v2, v0
+; GFX8-NEXT:    v_trunc_f16_e32 v2, v1
+; GFX8-NEXT:    v_add_f16_e64 v3, v1, -v2
+; GFX8-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX8-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX8-NEXT:    v_add_f16_e32 v1, v2, v1
+; GFX8-NEXT:    v_mov_b32_e32 v2, 16
+; GFX8-NEXT:    v_lshlrev_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_v2f16_fneg:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_xor_b32_e32 v0, 0x80008000, v0
+; GFX9-NEXT:    v_trunc_f16_e32 v2, v0
+; GFX9-NEXT:    s_mov_b32 s4, 0x8000
+; GFX9-NEXT:    v_add_f16_e64 v3, v0, -v2
+; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
+; GFX9-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX9-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX9-NEXT:    v_add_f16_e32 v0, v2, v0
+; GFX9-NEXT:    v_trunc_f16_e32 v2, v1
+; GFX9-NEXT:    v_add_f16_e64 v3, v1, -v2
+; GFX9-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX9-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v3|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX9-NEXT:    v_add_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0xffff
+; GFX9-NEXT:    v_and_or_b32 v0, v0, v2, v1
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %x.fneg = fneg <2 x half> %x
+  %roundeven = call <2 x half> @llvm.roundeven.v2f16(<2 x half> %x.fneg)
+  ret <2 x half> %roundeven
+}
+
+define <4 x half> @v_roundeven_v4f16(<4 x half> %x) {
+; GFX6-LABEL: v_roundeven_v4f16:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v0
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v8, v1
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v7, 0.5
+; GFX6-NEXT:    s_mov_b32 s4, 0x8000
+; GFX6-NEXT:    v_trunc_f32_e32 v5, v4
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX6-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX6-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v6, -v5
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; GFX6-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX6-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX6-NEXT:    v_add_f32_e32 v4, v4, v6
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; GFX6-NEXT:    v_trunc_f32_e32 v6, v8
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v11, 0.5
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v4, |v4|
+; GFX6-NEXT:    v_mov_b32_e32 v10, 0x3c00
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v9, -v6
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; GFX6-NEXT:    v_cmp_ge_f32_e32 vcc, v4, v7
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX6-NEXT:    v_add_f32_e32 v8, v8, v9
+; GFX6-NEXT:    v_mov_b32_e32 v4, 0x8000
+; GFX6-NEXT:    v_add_f32_e32 v0, v5, v0
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v8
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v8, v2
+; GFX6-NEXT:    v_and_b32_e32 v2, v2, v4
+; GFX6-NEXT:    v_or_b32_e32 v2, v10, v2
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v5, |v5|
+; GFX6-NEXT:    v_trunc_f32_e32 v9, v8
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v9, v9
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; GFX6-NEXT:    v_cmp_ge_f32_e32 vcc, v5, v7
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v7, v3
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v5, -v9
+; GFX6-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; GFX6-NEXT:    v_and_b32_e32 v3, v3, v4
+; GFX6-NEXT:    v_add_f32_e32 v5, v8, v5
+; GFX6-NEXT:    v_trunc_f32_e32 v8, v7
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v8, v8
+; GFX6-NEXT:    v_add_f32_e32 v1, v6, v1
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v5
+; GFX6-NEXT:    v_or_b32_e32 v3, v10, v3
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v6, -v8
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v8
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v5, |v5|
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX6-NEXT:    v_add_f32_e32 v6, v7, v6
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; GFX6-NEXT:    v_cmp_ge_f32_e32 vcc, v5, v11
+; GFX6-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v5, v9
+; GFX6-NEXT:    v_cvt_f32_f16_e64 v6, |v6|
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; GFX6-NEXT:    v_cmp_ge_f32_e32 vcc, v6, v11
+; GFX6-NEXT:    v_cndmask_b32_e32 v3, 0, v3, vcc
+; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; GFX6-NEXT:    v_add_f32_e32 v2, v5, v2
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; GFX6-NEXT:    v_add_f32_e32 v3, v4, v3
+; GFX6-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_v4f16:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v0
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v8, v1
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v7, 0.5
+; GFX7-NEXT:    s_mov_b32 s4, 0x8000
+; GFX7-NEXT:    v_trunc_f32_e32 v5, v4
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX7-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX7-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v6, -v5
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v5, v5
+; GFX7-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX7-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX7-NEXT:    v_add_f32_e32 v4, v4, v6
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v4, v4
+; GFX7-NEXT:    v_trunc_f32_e32 v6, v8
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v11, 0.5
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v4, |v4|
+; GFX7-NEXT:    v_mov_b32_e32 v10, 0x3c00
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v9, -v6
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v6, v6
+; GFX7-NEXT:    v_cmp_ge_f32_e32 vcc, v4, v7
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX7-NEXT:    v_add_f32_e32 v8, v8, v9
+; GFX7-NEXT:    v_mov_b32_e32 v4, 0x8000
+; GFX7-NEXT:    v_add_f32_e32 v0, v5, v0
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v8
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v8, v2
+; GFX7-NEXT:    v_and_b32_e32 v2, v2, v4
+; GFX7-NEXT:    v_or_b32_e32 v2, v10, v2
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v5, |v5|
+; GFX7-NEXT:    v_trunc_f32_e32 v9, v8
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v9, v9
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v0, v0
+; GFX7-NEXT:    v_cmp_ge_f32_e32 vcc, v5, v7
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v7, v3
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v5, -v9
+; GFX7-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; GFX7-NEXT:    v_and_b32_e32 v3, v3, v4
+; GFX7-NEXT:    v_add_f32_e32 v5, v8, v5
+; GFX7-NEXT:    v_trunc_f32_e32 v8, v7
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v8, v8
+; GFX7-NEXT:    v_add_f32_e32 v1, v6, v1
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v5
+; GFX7-NEXT:    v_or_b32_e32 v3, v10, v3
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v6, -v8
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v8
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v5, |v5|
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; GFX7-NEXT:    v_add_f32_e32 v6, v7, v6
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v6, v6
+; GFX7-NEXT:    v_cmp_ge_f32_e32 vcc, v5, v11
+; GFX7-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v5, v9
+; GFX7-NEXT:    v_cvt_f32_f16_e64 v6, |v6|
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; GFX7-NEXT:    v_cmp_ge_f32_e32 vcc, v6, v11
+; GFX7-NEXT:    v_cndmask_b32_e32 v3, 0, v3, vcc
+; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; GFX7-NEXT:    v_add_f32_e32 v2, v5, v2
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; GFX7-NEXT:    v_add_f32_e32 v3, v4, v3
+; GFX7-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_v4f16:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f16_e32 v4, v0
+; GFX8-NEXT:    s_mov_b32 s4, 0x8000
+; GFX8-NEXT:    v_add_f16_e64 v5, v0, -v4
+; GFX8-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
+; GFX8-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX8-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v5|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX8-NEXT:    v_add_f16_e32 v0, v4, v0
+; GFX8-NEXT:    v_trunc_f16_e32 v4, v2
+; GFX8-NEXT:    v_add_f16_e64 v5, v2, -v4
+; GFX8-NEXT:    v_and_b32_e32 v2, s4, v2
+; GFX8-NEXT:    v_or_b32_e32 v2, s5, v2
+; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v5|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX8-NEXT:    v_add_f16_e32 v2, v4, v2
+; GFX8-NEXT:    v_trunc_f16_e32 v4, v1
+; GFX8-NEXT:    v_add_f16_e64 v5, v1, -v4
+; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
+; GFX8-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX8-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v5|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX8-NEXT:    v_add_f16_e32 v1, v4, v1
+; GFX8-NEXT:    v_trunc_f16_e32 v4, v3
+; GFX8-NEXT:    v_add_f16_e64 v5, v3, -v4
+; GFX8-NEXT:    v_and_b32_e32 v3, 0x8000, v3
+; GFX8-NEXT:    v_or_b32_e32 v3, 0x3c00, v3
+; GFX8-NEXT:    v_cmp_ge_f16_e64 vcc, |v5|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v3, 0, v3, vcc
+; GFX8-NEXT:    v_add_f16_e32 v3, v4, v3
+; GFX8-NEXT:    v_mov_b32_e32 v4, 16
+; GFX8-NEXT:    v_lshlrev_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    v_lshlrev_b32_sdwa v2, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
+; GFX8-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_v4f16:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f16_e32 v4, v0
+; GFX9-NEXT:    s_mov_b32 s4, 0x8000
+; GFX9-NEXT:    v_add_f16_e64 v5, v0, -v4
+; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
+; GFX9-NEXT:    s_movk_i32 s5, 0x3c00
+; GFX9-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v5|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc
+; GFX9-NEXT:    v_add_f16_e32 v0, v4, v0
+; GFX9-NEXT:    v_trunc_f16_e32 v4, v2
+; GFX9-NEXT:    v_add_f16_e64 v5, v2, -v4
+; GFX9-NEXT:    v_and_b32_e32 v2, s4, v2
+; GFX9-NEXT:    v_or_b32_e32 v2, s5, v2
+; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v5|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, 0, v2, vcc
+; GFX9-NEXT:    v_add_f16_sdwa v2, v4, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX9-NEXT:    v_trunc_f16_e32 v4, v1
+; GFX9-NEXT:    v_add_f16_e64 v5, v1, -v4
+; GFX9-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
+; GFX9-NEXT:    v_and_b32_e32 v1, s4, v1
+; GFX9-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v5|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX9-NEXT:    v_add_f16_e32 v1, v4, v1
+; GFX9-NEXT:    v_trunc_f16_e32 v4, v3
+; GFX9-NEXT:    v_add_f16_e64 v5, v3, -v4
+; GFX9-NEXT:    v_and_b32_e32 v3, 0x8000, v3
+; GFX9-NEXT:    v_or_b32_e32 v3, 0x3c00, v3
+; GFX9-NEXT:    v_cmp_ge_f16_e64 vcc, |v5|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, 0, v3, vcc
+; GFX9-NEXT:    v_add_f16_sdwa v3, v4, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; GFX9-NEXT:    v_mov_b32_e32 v4, 0xffff
+; GFX9-NEXT:    v_and_or_b32 v0, v0, v4, v2
+; GFX9-NEXT:    v_and_or_b32 v1, v1, v4, v3
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %roundeven = call <4 x half> @llvm.roundeven.v4f16(<4 x half> %x)
+  ret <4 x half> %roundeven
+}
+
+
+define float @v_roundeven_f32_fabs(float %x) {
+; GFX6-LABEL: v_roundeven_f32_fabs:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_and_b32_e32 v1, 0x7fffffff, v0
+; GFX6-NEXT:    v_trunc_f32_e64 v2, |v0|
+; GFX6-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX6-NEXT:    v_sub_f32_e64 v0, |v0|, v2
+; GFX6-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX6-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_f32_fabs:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_and_b32_e32 v1, 0x7fffffff, v0
+; GFX7-NEXT:    v_trunc_f32_e64 v2, |v0|
+; GFX7-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX7-NEXT:    v_sub_f32_e64 v0, |v0|, v2
+; GFX7-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX7-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_f32_fabs:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_and_b32_e32 v1, 0x7fffffff, v0
+; GFX8-NEXT:    v_trunc_f32_e64 v2, |v0|
+; GFX8-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX8-NEXT:    v_sub_f32_e64 v0, |v0|, v2
+; GFX8-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX8-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_f32_fabs:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f32_e64 v2, |v0|
+; GFX9-NEXT:    v_and_b32_e32 v1, 0x7fffffff, v0
+; GFX9-NEXT:    v_sub_f32_e64 v0, |v0|, v2
+; GFX9-NEXT:    v_bfrev_b32_e32 v3, 1
+; GFX9-NEXT:    v_and_or_b32 v1, v1, v3, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX9-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %fabs.x = call float @llvm.fabs.f32(float %x)
+  %roundeven = call float @llvm.roundeven.f32(float %fabs.x)
+  ret float %roundeven
+}
+
+define amdgpu_ps float @s_roundeven_f32(float inreg %x) {
+; GFX6-LABEL: s_roundeven_f32:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    v_trunc_f32_e32 v0, s0
+; GFX6-NEXT:    v_sub_f32_e32 v1, s0, v0
+; GFX6-NEXT:    s_and_b32 s0, s0, 0x80000000
+; GFX6-NEXT:    s_or_b32 s0, 1.0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v1, 0, v2, vcc
+; GFX6-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX6-NEXT:    ; return to shader part epilog
+;
+; GFX7-LABEL: s_roundeven_f32:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    v_trunc_f32_e32 v0, s0
+; GFX7-NEXT:    v_sub_f32_e32 v1, s0, v0
+; GFX7-NEXT:    s_and_b32 s0, s0, 0x80000000
+; GFX7-NEXT:    s_or_b32 s0, 1.0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s0
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v1, 0, v2, vcc
+; GFX7-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX7-NEXT:    ; return to shader part epilog
+;
+; GFX8-LABEL: s_roundeven_f32:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    v_trunc_f32_e32 v0, s0
+; GFX8-NEXT:    v_sub_f32_e32 v1, s0, v0
+; GFX8-NEXT:    s_and_b32 s0, s0, 0x80000000
+; GFX8-NEXT:    s_or_b32 s0, 1.0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v2, s0
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, 0, v2, vcc
+; GFX8-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX8-NEXT:    ; return to shader part epilog
+;
+; GFX9-LABEL: s_roundeven_f32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    v_trunc_f32_e32 v0, s0
+; GFX9-NEXT:    v_sub_f32_e32 v1, s0, v0
+; GFX9-NEXT:    s_and_b32 s0, s0, 0x80000000
+; GFX9-NEXT:    s_or_b32 s0, 1.0, s0
+; GFX9-NEXT:    v_mov_b32_e32 v2, s0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v1|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, 0, v2, vcc
+; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
+; GFX9-NEXT:    ; return to shader part epilog
+  %roundeven = call float @llvm.roundeven.f32(float %x)
+  ret float %roundeven
+}
+
+define float @v_roundeven_f32_fneg(float %x) {
+; GFX6-LABEL: v_roundeven_f32_fneg:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_xor_b32_e32 v1, 0x80000000, v0
+; GFX6-NEXT:    v_trunc_f32_e64 v2, -v0
+; GFX6-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX6-NEXT:    v_sub_f32_e64 v0, -v0, v2
+; GFX6-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX6-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, 0.5
+; GFX6-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX6-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_f32_fneg:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_xor_b32_e32 v1, 0x80000000, v0
+; GFX7-NEXT:    v_trunc_f32_e64 v2, -v0
+; GFX7-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX7-NEXT:    v_sub_f32_e64 v0, -v0, v2
+; GFX7-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX7-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, 0.5
+; GFX7-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX7-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_f32_fneg:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_xor_b32_e32 v1, 0x80000000, v0
+; GFX8-NEXT:    v_trunc_f32_e64 v2, -v0
+; GFX8-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX8-NEXT:    v_sub_f32_e64 v0, -v0, v2
+; GFX8-NEXT:    v_or_b32_e32 v1, 1.0, v1
+; GFX8-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX8-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_f32_fneg:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f32_e64 v2, -v0
+; GFX9-NEXT:    v_xor_b32_e32 v1, 0x80000000, v0
+; GFX9-NEXT:    v_sub_f32_e64 v0, -v0, v2
+; GFX9-NEXT:    v_bfrev_b32_e32 v3, 1
+; GFX9-NEXT:    v_and_or_b32 v1, v1, v3, 1.0
+; GFX9-NEXT:    v_cmp_ge_f32_e64 vcc, |v0|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, 0, v1, vcc
+; GFX9-NEXT:    v_add_f32_e32 v0, v2, v0
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %neg.x = fneg float %x
+  %roundeven = call float @llvm.roundeven.f32(float %neg.x)
+  ret float %roundeven
+}
+
+define double @v_roundeven_f64(double %x) {
+; GFX6-LABEL: v_roundeven_f64:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_bfe_u32 v2, v1, 20, 11
+; GFX6-NEXT:    v_subrev_i32_e32 v4, vcc, 0x3ff, v2
+; GFX6-NEXT:    s_mov_b32 s4, -1
+; GFX6-NEXT:    s_mov_b32 s5, 0xfffff
+; GFX6-NEXT:    v_ashr_i64 v[2:3], s[4:5], v4
+; GFX6-NEXT:    v_and_b32_e32 v5, 0x80000000, v1
+; GFX6-NEXT:    v_xor_b32_e32 v2, -1, v2
+; GFX6-NEXT:    v_xor_b32_e32 v3, -1, v3
+; GFX6-NEXT:    v_and_b32_e32 v2, v0, v2
+; GFX6-NEXT:    v_and_b32_e32 v3, v1, v3
+; GFX6-NEXT:    v_cmp_gt_i32_e32 vcc, 0, v4
+; GFX6-NEXT:    v_mov_b32_e32 v6, 0
+; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc
+; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v5, vcc
+; GFX6-NEXT:    v_cmp_lt_i32_e32 vcc, 51, v4
+; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
+; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v1, vcc
+; GFX6-NEXT:    v_add_f64 v[4:5], v[0:1], -v[2:3]
+; GFX6-NEXT:    v_and_b32_e32 v0, 0x80000000, v1
+; GFX6-NEXT:    v_cmp_ge_f64_e64 vcc, |v[4:5]|, 0.5
+; GFX6-NEXT:    v_or_b32_e32 v1, 0x3ff00000, v0
+; GFX6-NEXT:    v_mov_b32_e32 v0, 0
+; GFX6-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX6-NEXT:    v_add_f64 v[0:1], v[2:3], v[0:1]
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_f64:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_trunc_f64_e32 v[2:3], v[0:1]
+; GFX7-NEXT:    v_add_f64 v[4:5], v[0:1], -v[2:3]
+; GFX7-NEXT:    v_and_b32_e32 v0, 0x80000000, v1
+; GFX7-NEXT:    v_cmp_ge_f64_e64 vcc, |v[4:5]|, 0.5
+; GFX7-NEXT:    v_or_b32_e32 v1, 0x3ff00000, v0
+; GFX7-NEXT:    v_mov_b32_e32 v0, 0
+; GFX7-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX7-NEXT:    v_add_f64 v[0:1], v[2:3], v[0:1]
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_f64:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f64_e32 v[2:3], v[0:1]
+; GFX8-NEXT:    v_add_f64 v[4:5], v[0:1], -v[2:3]
+; GFX8-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX8-NEXT:    v_or_b32_e32 v1, 0x3ff00000, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, 0
+; GFX8-NEXT:    v_cmp_ge_f64_e64 vcc, |v[4:5]|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX8-NEXT:    v_add_f64 v[0:1], v[2:3], v[0:1]
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_f64:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f64_e32 v[2:3], v[0:1]
+; GFX9-NEXT:    v_add_f64 v[4:5], v[0:1], -v[2:3]
+; GFX9-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX9-NEXT:    v_or_b32_e32 v1, 0x3ff00000, v1
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_cmp_ge_f64_e64 vcc, |v[4:5]|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX9-NEXT:    v_add_f64 v[0:1], v[2:3], v[0:1]
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %roundeven = call double @llvm.roundeven.f64(double %x)
+  ret double %roundeven
+}
+
+define double @v_roundeven_f64_fneg(double %x) {
+; GFX6-LABEL: v_roundeven_f64_fneg:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_xor_b32_e32 v4, 0x80000000, v1
+; GFX6-NEXT:    v_bfe_u32 v2, v4, 20, 11
+; GFX6-NEXT:    v_subrev_i32_e32 v5, vcc, 0x3ff, v2
+; GFX6-NEXT:    s_mov_b32 s4, -1
+; GFX6-NEXT:    s_mov_b32 s5, 0xfffff
+; GFX6-NEXT:    v_ashr_i64 v[2:3], s[4:5], v5
+; GFX6-NEXT:    v_and_b32_e32 v6, 0x80000000, v4
+; GFX6-NEXT:    v_xor_b32_e32 v2, -1, v2
+; GFX6-NEXT:    v_xor_b32_e32 v3, -1, v3
+; GFX6-NEXT:    v_and_b32_e32 v2, v0, v2
+; GFX6-NEXT:    v_and_b32_e32 v3, v4, v3
+; GFX6-NEXT:    v_cmp_gt_i32_e32 vcc, 0, v5
+; GFX6-NEXT:    v_mov_b32_e32 v7, 0
+; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v7, vcc
+; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v6, vcc
+; GFX6-NEXT:    v_cmp_lt_i32_e32 vcc, 51, v5
+; GFX6-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
+; GFX6-NEXT:    v_cndmask_b32_e32 v2, v2, v0, vcc
+; GFX6-NEXT:    v_add_f64 v[0:1], -v[0:1], -v[2:3]
+; GFX6-NEXT:    v_and_b32_e32 v4, 0x80000000, v4
+; GFX6-NEXT:    v_cmp_ge_f64_e64 vcc, |v[0:1]|, 0.5
+; GFX6-NEXT:    v_or_b32_e32 v5, 0x3ff00000, v4
+; GFX6-NEXT:    v_mov_b32_e32 v4, 0
+; GFX6-NEXT:    v_cndmask_b32_e32 v5, 0, v5, vcc
+; GFX6-NEXT:    v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_f64_fneg:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_trunc_f64_e64 v[2:3], -v[0:1]
+; GFX7-NEXT:    v_xor_b32_e32 v4, 0x80000000, v1
+; GFX7-NEXT:    v_add_f64 v[0:1], -v[0:1], -v[2:3]
+; GFX7-NEXT:    v_and_b32_e32 v4, 0x80000000, v4
+; GFX7-NEXT:    v_cmp_ge_f64_e64 vcc, |v[0:1]|, 0.5
+; GFX7-NEXT:    v_or_b32_e32 v5, 0x3ff00000, v4
+; GFX7-NEXT:    v_mov_b32_e32 v4, 0
+; GFX7-NEXT:    v_cndmask_b32_e32 v5, 0, v5, vcc
+; GFX7-NEXT:    v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_f64_fneg:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f64_e64 v[2:3], -v[0:1]
+; GFX8-NEXT:    v_add_f64 v[4:5], -v[0:1], -v[2:3]
+; GFX8-NEXT:    v_xor_b32_e32 v1, 0x80000000, v1
+; GFX8-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX8-NEXT:    v_or_b32_e32 v1, 0x3ff00000, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, 0
+; GFX8-NEXT:    v_cmp_ge_f64_e64 vcc, |v[4:5]|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX8-NEXT:    v_add_f64 v[0:1], v[2:3], v[0:1]
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_f64_fneg:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f64_e64 v[2:3], -v[0:1]
+; GFX9-NEXT:    v_add_f64 v[4:5], -v[0:1], -v[2:3]
+; GFX9-NEXT:    v_xor_b32_e32 v1, 0x80000000, v1
+; GFX9-NEXT:    v_and_b32_e32 v1, 0x80000000, v1
+; GFX9-NEXT:    v_or_b32_e32 v1, 0x3ff00000, v1
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    v_cmp_ge_f64_e64 vcc, |v[4:5]|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v1, 0, v1, vcc
+; GFX9-NEXT:    v_add_f64 v[0:1], v[2:3], v[0:1]
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %neg.x = fneg double %x
+  %roundeven = call double @llvm.roundeven.f64(double %neg.x)
+  ret double %roundeven
+}
+
+define <2 x double> @v_roundeven_v2f64(<2 x double> %x) {
+; GFX6-LABEL: v_roundeven_v2f64:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_bfe_u32 v4, v1, 20, 11
+; GFX6-NEXT:    s_movk_i32 s4, 0x3ff
+; GFX6-NEXT:    v_subrev_i32_e32 v6, vcc, s4, v4
+; GFX6-NEXT:    s_mov_b32 s4, -1
+; GFX6-NEXT:    s_mov_b32 s5, 0xfffff
+; GFX6-NEXT:    v_ashr_i64 v[4:5], s[4:5], v6
+; GFX6-NEXT:    v_and_b32_e32 v7, 0x80000000, v1
+; GFX6-NEXT:    v_xor_b32_e32 v4, -1, v4
+; GFX6-NEXT:    v_xor_b32_e32 v5, -1, v5
+; GFX6-NEXT:    v_and_b32_e32 v4, v0, v4
+; GFX6-NEXT:    v_and_b32_e32 v5, v1, v5
+; GFX6-NEXT:    v_cmp_gt_i32_e32 vcc, 0, v6
+; GFX6-NEXT:    v_mov_b32_e32 v11, 0
+; GFX6-NEXT:    v_cndmask_b32_e32 v4, v4, v11, vcc
+; GFX6-NEXT:    v_cndmask_b32_e32 v5, v5, v7, vcc
+; GFX6-NEXT:    v_cmp_lt_i32_e32 vcc, 51, v6
+; GFX6-NEXT:    v_cndmask_b32_e32 v4, v4, v0, vcc
+; GFX6-NEXT:    v_cndmask_b32_e32 v5, v5, v1, vcc
+; GFX6-NEXT:    v_add_f64 v[6:7], v[0:1], -v[4:5]
+; GFX6-NEXT:    s_brev_b32 s6, 1
+; GFX6-NEXT:    v_cmp_ge_f64_e64 vcc, |v[6:7]|, 0.5
+; GFX6-NEXT:    v_and_b32_e32 v0, s6, v1
+; GFX6-NEXT:    s_mov_b32 s7, 0x3ff00000
+; GFX6-NEXT:    v_or_b32_e32 v0, s7, v0
+; GFX6-NEXT:    v_mov_b32_e32 v8, 0
+; GFX6-NEXT:    v_cndmask_b32_e32 v9, 0, v0, vcc
+; GFX6-NEXT:    v_add_f64 v[0:1], v[4:5], v[8:9]
+; GFX6-NEXT:    v_mov_b32_e32 v10, 0x3ff
+; GFX6-NEXT:    v_bfe_u32 v4, v3, 20, 11
+; GFX6-NEXT:    v_sub_i32_e32 v6, vcc, v4, v10
+; GFX6-NEXT:    v_ashr_i64 v[4:5], s[4:5], v6
+; GFX6-NEXT:    v_and_b32_e32 v7, 0x80000000, v3
+; GFX6-NEXT:    v_xor_b32_e32 v4, -1, v4
+; GFX6-NEXT:    v_xor_b32_e32 v5, -1, v5
+; GFX6-NEXT:    v_and_b32_e32 v4, v2, v4
+; GFX6-NEXT:    v_cmp_gt_i32_e32 vcc, 0, v6
+; GFX6-NEXT:    v_and_b32_e32 v5, v3, v5
+; GFX6-NEXT:    v_cndmask_b32_e32 v4, v4, v11, vcc
+; GFX6-NEXT:    v_cndmask_b32_e32 v5, v5, v7, vcc
+; GFX6-NEXT:    v_cmp_lt_i32_e32 vcc, 51, v6
+; GFX6-NEXT:    v_cndmask_b32_e32 v4, v4, v2, vcc
+; GFX6-NEXT:    v_cndmask_b32_e32 v5, v5, v3, vcc
+; GFX6-NEXT:    v_add_f64 v[6:7], v[2:3], -v[4:5]
+; GFX6-NEXT:    v_and_b32_e32 v2, s6, v3
+; GFX6-NEXT:    v_cmp_ge_f64_e64 vcc, |v[6:7]|, 0.5
+; GFX6-NEXT:    v_or_b32_e32 v2, s7, v2
+; GFX6-NEXT:    v_cndmask_b32_e32 v9, 0, v2, vcc
+; GFX6-NEXT:    v_add_f64 v[2:3], v[4:5], v[8:9]
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX7-LABEL: v_roundeven_v2f64:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_trunc_f64_e32 v[4:5], v[0:1]
+; GFX7-NEXT:    s_brev_b32 s4, 1
+; GFX7-NEXT:    v_add_f64 v[6:7], v[0:1], -v[4:5]
+; GFX7-NEXT:    v_and_b32_e32 v0, s4, v1
+; GFX7-NEXT:    s_mov_b32 s5, 0x3ff00000
+; GFX7-NEXT:    v_cmp_ge_f64_e64 vcc, |v[6:7]|, 0.5
+; GFX7-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX7-NEXT:    v_trunc_f64_e32 v[6:7], v[2:3]
+; GFX7-NEXT:    v_mov_b32_e32 v8, 0
+; GFX7-NEXT:    v_cndmask_b32_e32 v9, 0, v0, vcc
+; GFX7-NEXT:    v_add_f64 v[0:1], v[4:5], v[8:9]
+; GFX7-NEXT:    v_add_f64 v[4:5], v[2:3], -v[6:7]
+; GFX7-NEXT:    v_and_b32_e32 v2, s4, v3
+; GFX7-NEXT:    v_cmp_ge_f64_e64 vcc, |v[4:5]|, 0.5
+; GFX7-NEXT:    v_or_b32_e32 v2, s5, v2
+; GFX7-NEXT:    v_cndmask_b32_e32 v9, 0, v2, vcc
+; GFX7-NEXT:    v_add_f64 v[2:3], v[6:7], v[8:9]
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_roundeven_v2f64:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_trunc_f64_e32 v[4:5], v[0:1]
+; GFX8-NEXT:    v_trunc_f64_e32 v[6:7], v[2:3]
+; GFX8-NEXT:    s_brev_b32 s4, 1
+; GFX8-NEXT:    s_mov_b32 s5, 0x3ff00000
+; GFX8-NEXT:    v_mov_b32_e32 v10, 0
+; GFX8-NEXT:    v_add_f64 v[8:9], v[0:1], -v[4:5]
+; GFX8-NEXT:    v_add_f64 v[12:13], v[2:3], -v[6:7]
+; GFX8-NEXT:    v_and_b32_e32 v0, s4, v1
+; GFX8-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX8-NEXT:    v_and_b32_e32 v1, s4, v3
+; GFX8-NEXT:    v_or_b32_e32 v2, s5, v1
+; GFX8-NEXT:    v_cmp_ge_f64_e64 vcc, |v[8:9]|, 0.5
+; GFX8-NEXT:    v_cndmask_b32_e32 v11, 0, v0, vcc
+; GFX8-NEXT:    v_cmp_ge_f64_e64 vcc, |v[12:13]|, 0.5
+; GFX8-NEXT:    v_add_f64 v[0:1], v[4:5], v[10:11]
+; GFX8-NEXT:    v_cndmask_b32_e32 v11, 0, v2, vcc
+; GFX8-NEXT:    v_add_f64 v[2:3], v[6:7], v[10:11]
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_roundeven_v2f64:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_trunc_f64_e32 v[4:5], v[0:1]
+; GFX9-NEXT:    v_trunc_f64_e32 v[6:7], v[2:3]
+; GFX9-NEXT:    s_brev_b32 s4, 1
+; GFX9-NEXT:    s_mov_b32 s5, 0x3ff00000
+; GFX9-NEXT:    v_mov_b32_e32 v10, 0
+; GFX9-NEXT:    v_add_f64 v[8:9], v[0:1], -v[4:5]
+; GFX9-NEXT:    v_add_f64 v[12:13], v[2:3], -v[6:7]
+; GFX9-NEXT:    v_and_b32_e32 v0, s4, v1
+; GFX9-NEXT:    v_or_b32_e32 v0, s5, v0
+; GFX9-NEXT:    v_and_b32_e32 v1, s4, v3
+; GFX9-NEXT:    v_or_b32_e32 v2, s5, v1
+; GFX9-NEXT:    v_cmp_ge_f64_e64 vcc, |v[8:9]|, 0.5
+; GFX9-NEXT:    v_cndmask_b32_e32 v11, 0, v0, vcc
+; GFX9-NEXT:    v_cmp_ge_f64_e64 vcc, |v[12:13]|, 0.5
+; GFX9-NEXT:    v_add_f64 v[0:1], v[4:5], v[10:11]
+; GFX9-NEXT:    v_cndmask_b32_e32 v11, 0, v2, vcc
+; GFX9-NEXT:    v_add_f64 v[2:3], v[6:7], v[10:11]
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %roundeven = call <2 x double> @llvm.roundeven.v2f64(<2 x double> %x)
+  ret <2 x double> %roundeven
+}
+
+declare half @llvm.roundeven.f16(half) #0
+declare <2 x half> @llvm.roundeven.v2f16(<2 x half>) #0
+declare <4 x half> @llvm.roundeven.v4f16(<4 x half>) #0
+
+declare float @llvm.roundeven.f32(float) #0
+declare <2 x float> @llvm.roundeven.v2f32(<2 x float>) #0
+declare <3 x float> @llvm.roundeven.v3f32(<3 x float>) #0
+declare <4 x float> @llvm.roundeven.v4f32(<4 x float>) #0
+
+declare double @llvm.roundeven.f64(double) #0
+declare <2 x double> @llvm.roundeven.v2f64(<2 x double>) #0
+
+declare half @llvm.fabs.f16(half) #0
+declare float @llvm.fabs.f32(float) #0
+
+attributes #0 = { nounwind readnone speculatable willreturn }

llvm/test/CodeGen/X86/GlobalISel/roundeven.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -mtriple=x86_64-linux-gnu < %s | FileCheck %s
+
+; FIXME: Calling convention lowering fails
+; define half @roundeven_f16(half %x) {
+;   %roundeven = call half @llvm.roundeven.f16(half %x)
+;   ret half %roundeven
+; }
+
+define float @roundeven_f32(float %x) {
+; CHECK-LABEL: roundeven_f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pushq %rax
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    callq roundevenf
+; CHECK-NEXT:    popq %rax
+; CHECK-NEXT:    .cfi_def_cfa_offset 8
+; CHECK-NEXT:    retq
+  %roundeven = call float @llvm.roundeven.f32(float %x)
+  ret float %roundeven
+}
+
+define double @roundeven_f64(double %x) {
+; CHECK-LABEL: roundeven_f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pushq %rax
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    callq roundeven
+; CHECK-NEXT:    popq %rax
+; CHECK-NEXT:    .cfi_def_cfa_offset 8
+; CHECK-NEXT:    retq
+  %roundeven = call double @llvm.roundeven.f64(double %x)
+  ret double %roundeven
+}
+
+; FIXME: Insert fails
+; define x86_fp80 @roundeven_fp80(x86_fp80 %x) {
+;   %roundeven = call x86_fp80 @llvm.roundeven.f80(x86_fp80 %x)
+;   ret x86_fp80 %roundeven
+; }
+
+define fp128 @roundeven_f128(fp128 %x) {
+; CHECK-LABEL: roundeven_f128:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pushq %rax
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    callq roundevenl
+; CHECK-NEXT:    popq %rax
+; CHECK-NEXT:    .cfi_def_cfa_offset 8
+; CHECK-NEXT:    retq
+  %roundeven = call fp128 @llvm.roundeven.f128(fp128 %x)
+  ret fp128 %roundeven
+}
+
+; FIXME: Fails on build_vector
+; define <4 x float> @roundeven_v4f32(<4 x float> %x) {
+;   %roundeven = call <4 x float> @llvm.roundeven.v4f32(<4 x float> %x)
+;   ret <4 x float> %roundeven
+; }
+
+declare half @llvm.roundeven.f16(half) #0
+declare float @llvm.roundeven.f32(float) #0
+declare <4 x float> @llvm.roundeven.v4f32(<4 x float>) #0
+declare double @llvm.roundeven.f64(double) #0
+declare x86_fp80 @llvm.roundeven.f80(x86_fp80) #0
+declare fp128 @llvm.roundeven.f128(fp128) #0
+
+attributes #0 = { nounwind readnone speculatable willreturn }