[WebAssembly] Custom combines for f32x4.demote_zero_f64x2

[tlively]

Summary

Replace the clang builtin function and LLVM intrinsic for
f32x4.demote_zero_f64x2 with combines from normal SDNodes. Also add missing
combines for i32x4.trunc_sat_zero_f64x2_{s,u}, which share the same pattern.

[aheejin]

clang/include/clang/Basic/BuiltinsWebAssembly.def

192–193

If these share the same pattern, do we need these builtins?

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

2367

Do we not need to check if N->getOpernad(1) is a specific type, such as v2f32 or v2i32?

2411–2412

Same here; do we not need to check if Concat.getOpernad(1) is v2f64?

[tlively]

• Check types of splats as well

clang/include/clang/Basic/BuiltinsWebAssembly.def

192–193

Hmm, good point. Maybe not. I'll investigate that in a follow up.

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

2367

I think that the concat_vectors node ensures that the two sides have the same type, but I'll explicitly check to be safe.

[tlively]

clang/include/clang/Basic/BuiltinsWebAssembly.def

192–193

Ah, the reason we can't remove these builtins is because there is no good target-independent way to represent the saturating behavior of the float-to-int conversion.

clang/include/clang/Basic/BuiltinsWebAssembly.def

 TARGET_BUILTIN(__builtin_wasm_trunc_saturate_s_i32x4_f32x4, "V4iV4f", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_trunc_saturate_u_i32x4_f32x4, "V4iV4f", "nc", "simd128")
 
 TARGET_BUILTIN(__builtin_wasm_narrow_s_i8x16_i16x8, "V16ScV8sV8s", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_narrow_u_i8x16_i16x8, "V16UcV8sV8s", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_narrow_s_i16x8_i32x4, "V8sV4iV4i", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_narrow_u_i16x8_i32x4, "V8UsV4iV4i", "nc", "simd128")
 
 TARGET_BUILTIN(__builtin_wasm_trunc_sat_zero_s_f64x2_i32x4, "V4iV2d", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_trunc_sat_zero_u_f64x2_i32x4, "V4UiV2d", "nc", "simd128")
-TARGET_BUILTIN(__builtin_wasm_demote_zero_f64x2_f32x4, "V4fV2d", "nc", "simd128")
 
 TARGET_BUILTIN(__builtin_wasm_load32_zero, "V4iiC*", "n", "simd128")
 TARGET_BUILTIN(__builtin_wasm_load64_zero, "V2LLiLLiC*", "n", "simd128")
 
 TARGET_BUILTIN(__builtin_wasm_load8_lane, "V16ScScC*V16ScIi", "n", "simd128")
 TARGET_BUILTIN(__builtin_wasm_load16_lane, "V8ssC*V8sIi", "n", "simd128")
 TARGET_BUILTIN(__builtin_wasm_load32_lane, "V4iiC*V4iIi", "n", "simd128")
 TARGET_BUILTIN(__builtin_wasm_load64_lane, "V2LLiLLiC*V2LLiIi", "n", "simd128")
 TARGET_BUILTIN(__builtin_wasm_store8_lane, "vSc*V16ScIi", "n", "simd128")
 TARGET_BUILTIN(__builtin_wasm_store16_lane, "vs*V8sIi", "n", "simd128")
 TARGET_BUILTIN(__builtin_wasm_store32_lane, "vi*V4iIi", "n", "simd128")
 TARGET_BUILTIN(__builtin_wasm_store64_lane, "vLLi*V2LLiIi", "n", "simd128")
 
 #undef BUILTIN
 #undef TARGET_BUILTIN

clang/lib/CodeGen/CGBuiltin.cpp

case WebAssembly::BI__builtin_wasm_trunc_sat_zero_u_f64x2_i32x4: {
     Function *Callee = CGM.getIntrinsic(IntNo, {TruncT, SrcT});
     Value *Trunc = Builder.CreateCall(Callee, Vec);
     Value *Splat = Builder.CreateVectorSplat(2, Builder.getInt32(0));
     Value *ConcatMask =
         llvm::ConstantVector::get({Builder.getInt32(0), Builder.getInt32(1),
                                    Builder.getInt32(2), Builder.getInt32(3)});
     return Builder.CreateShuffleVector(Trunc, Splat, ConcatMask);
   }
-  case WebAssembly::BI__builtin_wasm_demote_zero_f64x2_f32x4: {
-    Value *Vec = EmitScalarExpr(E->getArg(0));
-    Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_demote_zero);
-    return Builder.CreateCall(Callee, Vec);
-  }
   case WebAssembly::BI__builtin_wasm_load32_zero: {
     Value *Ptr = EmitScalarExpr(E->getArg(0));
     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_load32_zero);
     return Builder.CreateCall(Callee, {Ptr});
   }
   case WebAssembly::BI__builtin_wasm_load64_zero: {
     Value *Ptr = EmitScalarExpr(E->getArg(0));
     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_load64_zero);

clang/lib/Headers/wasm_simd128.h

 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_u32x4_trunc_sat_f64x2_zero(v128_t __a) {
   return (v128_t)__builtin_wasm_trunc_sat_zero_u_f64x2_i32x4((__f64x2)__a);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_f32x4_demote_f64x2_zero(v128_t __a) {
-  return (v128_t)__builtin_wasm_demote_zero_f64x2_f32x4((__f64x2)__a);
+  return (v128_t) __builtin_convertvector(
+      __builtin_shufflevector((__f64x2)__a, (__f64x2){0, 0}, 0, 1, 2, 3),
+      __f32x4);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_f64x2_promote_low_f32x4(v128_t __a) {
   return (v128_t) __builtin_convertvector(
       (__f32x2){((__f32x4)__a)[0], ((__f32x4)__a)[1]}, __f64x2);
 }
 

clang/test/CodeGen/builtins-wasm.c

 
 u32x4 trunc_sat_zero_u_f64x2_i32x4(f64x2 x) {
   return __builtin_wasm_trunc_sat_zero_u_f64x2_i32x4(x);
   // WEBASSEMBLY: %0 = tail call <2 x i32> @llvm.fptoui.sat.v2i32.v2f64(<2 x double> %x)
   // WEBASSEMBLY: %1 = shufflevector <2 x i32> %0, <2 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   // WEBASSEMBLY: ret <4 x i32> %1
 }
 
-f32x4 wasm_demote_zero_f64x2_f32x4(f64x2 x) {
-  return __builtin_wasm_demote_zero_f64x2_f32x4(x);
-  // WEBASSEMBLY: call <4 x float> @llvm.wasm.demote.zero(<2 x double> %x)
-  // WEBASSEMBLY: ret
-}
-
 i32x4 load32_zero(const int *p) {
   return __builtin_wasm_load32_zero(p);
   // WEBASSEMBLY: call <4 x i32> @llvm.wasm.load32.zero(i32* %p)
   // WEBASSEMBLY: ret
 }
 
 i64x2 load64_zero(const long long *p) {
   return __builtin_wasm_load64_zero(p);

clang/test/Headers/wasm.c

 //
 v128_t test_u32x4_trunc_sat_f64x2_zero(v128_t a) {
   return wasm_u32x4_trunc_sat_f64x2_zero(a);
 }
 
 // CHECK-LABEL: @test_f32x4_demote_f64x2_zero(
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[TMP0:%.*]] = bitcast <4 x i32> [[A:%.*]] to <2 x double>
-// CHECK-NEXT:    [[TMP1:%.*]] = tail call <4 x float> @llvm.wasm.demote.zero(<2 x double> [[TMP0]]) #[[ATTR10]]
-// CHECK-NEXT:    [[TMP2:%.*]] = bitcast <4 x float> [[TMP1]] to <4 x i32>
-// CHECK-NEXT:    ret <4 x i32> [[TMP2]]
+// CHECK-NEXT:    [[SHUFFLE_I:%.*]] = shufflevector <2 x double> [[TMP0]], <2 x double> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+// CHECK-NEXT:    [[CONV_I:%.*]] = fptrunc <4 x double> [[SHUFFLE_I]] to <4 x float>
+// CHECK-NEXT:    [[TMP1:%.*]] = bitcast <4 x float> [[CONV_I]] to <4 x i32>
+// CHECK-NEXT:    ret <4 x i32> [[TMP1]]
 //
 v128_t test_f32x4_demote_f64x2_zero(v128_t a) {
   return wasm_f32x4_demote_f64x2_zero(a);
 }
 
 // CHECK-LABEL: @test_f64x2_promote_low_f32x4(
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[TMP0:%.*]] = bitcast <4 x i32> [[A:%.*]] to <4 x float>

llvm/include/llvm/IR/IntrinsicsWebAssembly.td

def int_wasm_extadd_pairwise_signed :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMSubdivide2VectorType<0>],
             [IntrNoMem, IntrSpeculatable]>;
 def int_wasm_extadd_pairwise_unsigned :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMSubdivide2VectorType<0>],
             [IntrNoMem, IntrSpeculatable]>;
 
-// TODO: Remove this if possible.
-def int_wasm_demote_zero :
-  Intrinsic<[llvm_v4f32_ty], [llvm_v2f64_ty],
-            [IntrNoMem, IntrSpeculatable]>;
-
 //===----------------------------------------------------------------------===//
 // Thread-local storage intrinsics
 //===----------------------------------------------------------------------===//
 
 def int_wasm_tls_size :
   Intrinsic<[llvm_anyint_ty],
             [],
             [IntrNoMem, IntrSpeculatable]>;

llvm/lib/Target/WebAssembly/WebAssemblyISD.def

 HANDLE_NODETYPE(EXTEND_LOW_U)
 HANDLE_NODETYPE(EXTEND_HIGH_S)
 HANDLE_NODETYPE(EXTEND_HIGH_U)
 HANDLE_NODETYPE(CONVERT_LOW_S)
 HANDLE_NODETYPE(CONVERT_LOW_U)
 HANDLE_NODETYPE(PROMOTE_LOW)
 HANDLE_NODETYPE(TRUNC_SAT_ZERO_S)
 HANDLE_NODETYPE(TRUNC_SAT_ZERO_U)
+HANDLE_NODETYPE(DEMOTE_ZERO)
 HANDLE_NODETYPE(THROW)
 HANDLE_NODETYPE(CATCH)
 HANDLE_NODETYPE(MEMORY_COPY)
 HANDLE_NODETYPE(MEMORY_FILL)
 
 // Memory intrinsics
 HANDLE_MEM_NODETYPE(LOAD_SPLAT)
 HANDLE_MEM_NODETYPE(GLOBAL_GET)
 HANDLE_MEM_NODETYPE(GLOBAL_SET)

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

if (Subtarget->hasSIMD128()) {
 
     // Combine int_to_fp or fp_extend of extract_vectors and vice versa into
     // conversions ops
     setTargetDAGCombine(ISD::SINT_TO_FP);
     setTargetDAGCombine(ISD::UINT_TO_FP);
     setTargetDAGCombine(ISD::FP_EXTEND);
     setTargetDAGCombine(ISD::EXTRACT_SUBVECTOR);
 
-    // Combine concat of {s,u}int_to_fp_sat to i32x4.trunc_sat_f64x2_zero_{s,u}
+    // Combine fp_to_{s,u}int_sat or fp_round of concat_vectors or vice versa
+    // into conversion ops
+    setTargetDAGCombine(ISD::FP_TO_SINT_SAT);
+    setTargetDAGCombine(ISD::FP_TO_UINT_SAT);
+    setTargetDAGCombine(ISD::FP_ROUND);
     setTargetDAGCombine(ISD::CONCAT_VECTORS);
 
     // Support saturating add for i8x16 and i16x8
     for (auto Op : {ISD::SADDSAT, ISD::UADDSAT})
       for (auto T : {MVT::v16i8, MVT::v8i16})
         setOperationAction(Op, T, Legal);
 
     // Support integer abs
performVectorConvertLowCombine(SDNode *N,
   if (IndexNode == nullptr || IndexNode->getZExtValue() != 0)
     return SDValue();
 
   unsigned Op = GetWasmConversionOp(ConversionOp);
   return DAG.getNode(Op, SDLoc(N), ResVT, Source);
 }
 
 static SDValue
-performVectorTruncSatLowCombine(SDNode *N,
-                                TargetLowering::DAGCombinerInfo &DCI) {
+performVectorTruncZeroCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
   auto &DAG = DCI.DAG;
-  assert(N->getOpcode() == ISD::CONCAT_VECTORS);
 
+  auto GetWasmConversionOp = [](unsigned Op) {
+    switch (Op) {
+    case ISD::FP_TO_SINT_SAT:
+      return WebAssemblyISD::TRUNC_SAT_ZERO_S;
+    case ISD::FP_TO_UINT_SAT:
+      return WebAssemblyISD::TRUNC_SAT_ZERO_U;
+    case ISD::FP_ROUND:
+      return WebAssemblyISD::DEMOTE_ZERO;
+    }
+    llvm_unreachable("unexpected op");
+  };
+
+  auto IsZeroSplat = [](SDValue SplatVal) {
+    auto *Splat = dyn_cast<BuildVectorSDNode>(SplatVal.getNode());
+    APInt SplatValue, SplatUndef;
+    unsigned SplatBitSize;
+    bool HasAnyUndefs;
+    return Splat &&
+           Splat->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
+                                  HasAnyUndefs) &&
+           SplatValue == 0;
+  };
+
+  if (N->getOpcode() == ISD::CONCAT_VECTORS) {
     // Combine this:
     //
     //   (concat_vectors (v2i32 (fp_to_{s,u}int_sat $x, 32)), (v2i32 (splat 0)))
     //
     // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x).
-  EVT ResVT = N->getValueType(0);
-  if (ResVT != MVT::v4i32)
+    //
+    // Or this:
+    //
+    //   (concat_vectors (v2f32 (fp_round (v2f64 $x))), (v2f32 (splat 0)))
+    //
+    // into (f32x4.demote_zero_f64x2 $x).
+    EVT ResVT;
+    EVT ExpectedConversionType;
+    auto Conversion = N->getOperand(0);
+    auto ConversionOp = Conversion.getOpcode();
+    switch (ConversionOp) {
+    case ISD::FP_TO_SINT_SAT:
+    case ISD::FP_TO_UINT_SAT:
+      ResVT = MVT::v4i32;
+      ExpectedConversionType = MVT::v2i32;
+      break;
+    case ISD::FP_ROUND:
+      ResVT = MVT::v4f32;
+      ExpectedConversionType = MVT::v2f32;
+      break;
+    default:
       return SDValue();
+    }
 
-  auto FPToInt = N->getOperand(0);
-  auto FPToIntOp = FPToInt.getOpcode();
-  if (FPToIntOp != ISD::FP_TO_SINT_SAT && FPToIntOp != ISD::FP_TO_UINT_SAT)
+    if (N->getValueType(0) != ResVT)
       return SDValue();
-  if (cast<VTSDNode>(FPToInt.getOperand(1))->getVT() != MVT::i32)
+
+    if (Conversion.getValueType() != ExpectedConversionType)
       return SDValue();
 
-  auto Source = FPToInt.getOperand(0);
+    auto Source = Conversion.getOperand(0);
     if (Source.getValueType() != MVT::v2f64)
       return SDValue();
 
-  auto *Splat = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
-  APInt SplatValue, SplatUndef;
-  unsigned SplatBitSize;
-  bool HasAnyUndefs;
-  if (!Splat || !Splat->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
-                                        HasAnyUndefs))
+    if (!IsZeroSplat(N->getOperand(1)) ||
+        N->getOperand(1).getValueType() != ExpectedConversionType)
       return SDValue();
-  if (SplatValue != 0)
+
+    unsigned Op = GetWasmConversionOp(ConversionOp);
+    return DAG.getNode(Op, SDLoc(N), ResVT, Source);
+  }
+
+  // Combine this:
+  //
+  //   (fp_to_{s,u}int_sat (concat_vectors $x, (v2f64 (splat 0))), 32)
+  //
+  // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x).
+  //
+  // Or this:
+  //
+  //   (v4f32 (fp_round (concat_vectors $x, (v2f64 (splat 0)))))
+  //
+  // into (f32x4.demote_zero_f64x2 $x).
+  EVT ResVT;
+  auto ConversionOp = N->getOpcode();
+  switch (ConversionOp) {
+  case ISD::FP_TO_SINT_SAT:
+  case ISD::FP_TO_UINT_SAT:
+    ResVT = MVT::v4i32;
+    break;
+  case ISD::FP_ROUND:
+    ResVT = MVT::v4f32;
+    break;
+  default:
+    llvm_unreachable("unexpected op");
+  }
+
+  if (N->getValueType(0) != ResVT)
+    return SDValue();
+
+  auto Concat = N->getOperand(0);
+  if (Concat.getValueType() != MVT::v4f64)
+    return SDValue();
+
+  auto Source = Concat.getOperand(0);
+  if (Source.getValueType() != MVT::v2f64)
     return SDValue();
 
-  unsigned Op = FPToIntOp == ISD::FP_TO_SINT_SAT
-                    ? WebAssemblyISD::TRUNC_SAT_ZERO_S
-                    : WebAssemblyISD::TRUNC_SAT_ZERO_U;
+  if (!IsZeroSplat(Concat.getOperand(1)) ||
+      Concat.getOperand(1).getValueType() != MVT::v2f64)
+    return SDValue();
 
+  unsigned Op = GetWasmConversionOp(ConversionOp);
   return DAG.getNode(Op, SDLoc(N), ResVT, Source);
 }
 
 SDValue
 WebAssemblyTargetLowering::PerformDAGCombine(SDNode *N,
                                              DAGCombinerInfo &DCI) const {
   switch (N->getOpcode()) {
   default:
     return SDValue();
   case ISD::VECTOR_SHUFFLE:
     return performVECTOR_SHUFFLECombine(N, DCI);
   case ISD::SIGN_EXTEND:
   case ISD::ZERO_EXTEND:
     return performVectorExtendCombine(N, DCI);
   case ISD::SINT_TO_FP:
   case ISD::UINT_TO_FP:
   case ISD::FP_EXTEND:
   case ISD::EXTRACT_SUBVECTOR:
     return performVectorConvertLowCombine(N, DCI);
+  case ISD::FP_TO_SINT_SAT:
+  case ISD::FP_TO_UINT_SAT:
+  case ISD::FP_ROUND:
   case ISD::CONCAT_VECTORS:
-    return performVectorTruncSatLowCombine(N, DCI);
+    return performVectorTruncZeroCombine(N, DCI);
   }
 }

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

 defm "" : SIMDConvert<I16x8, I8x16, int_wasm_extadd_pairwise_unsigned,
                       "extadd_pairwise_i8x16_u", 0x7d>;
 defm "" : SIMDConvert<I32x4, I16x8, int_wasm_extadd_pairwise_signed,
                       "extadd_pairwise_i16x8_s", 0x7e>;
 defm "" : SIMDConvert<I32x4, I16x8, int_wasm_extadd_pairwise_unsigned,
                       "extadd_pairwise_i16x8_u", 0x7f>;
 
 // f64x2 <-> f32x4 conversions
-defm "" : SIMDConvert<F32x4, F64x2, int_wasm_demote_zero,
+def demote_t : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>]>;
+def demote_zero : SDNode<"WebAssemblyISD::DEMOTE_ZERO", demote_t>;
+defm "" : SIMDConvert<F32x4, F64x2, demote_zero,
                       "demote_zero_f64x2", 0x5e>;
 
 def promote_t : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>]>;
 def promote_low : SDNode<"WebAssemblyISD::PROMOTE_LOW", promote_t>;
 defm "" : SIMDConvert<F64x2, F32x4, promote_low, "promote_low_f32x4", 0x5f>;
 
 //===----------------------------------------------------------------------===//
 // Saturating Rounding Q-Format Multiplication
 //===----------------------------------------------------------------------===//
 
 defm Q15MULR_SAT_S :
   SIMDBinary<I16x8, int_wasm_q15mulr_sat_signed, "q15mulr_sat_s", 0x82>;

llvm/test/CodeGen/WebAssembly/simd-conversions.ll

 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-NOT: i64x2.trunc_sat_f64x2_u
 ; SIMD128-NEXT: .functype trunc_sat_u_v2i64 (v128) -> (v128){{$}}
 define <2 x i64> @trunc_sat_u_v2i64(<2 x double> %x) {
   %a = fptoui <2 x double> %x to <2 x i64>
   ret <2 x i64> %a
 }
 
+; CHECK-LABEL: demote_zero_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128-NEXT: .functype demote_zero_v4f32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f32x4.demote_zero_f64x2 $push[[R:[0-9]+]]=, $0
+; SIMD128-NEXT: return $pop[[R]]
+define <4 x float> @demote_zero_v4f32(<2 x double> %x) {
+  %v = shufflevector <2 x double> %x, <2 x double> zeroinitializer,
+         <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %a = fptrunc <4 x double> %v to <4 x float>
+  ret <4 x float> %a
+}
+
+; CHECK-LABEL: demote_zero_v4f32_2:
+; NO-SIMD128-NOT: f32x4
+; SIMD128-NEXT: .functype demote_zero_v4f32_2 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f32x4.demote_zero_f64x2 $push[[R:[0-9]+]]=, $0
+; SIMD128-NEXT: return $pop[[R]]
+define <4 x float> @demote_zero_v4f32_2(<2 x double> %x) {
+  %v = fptrunc <2 x double> %x to <2 x float>
+  %a = shufflevector <2 x float> %v, <2 x float> zeroinitializer,
+         <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  ret <4 x float> %a
+}
+
 ; CHECK-LABEL: convert_low_s_v2f64:
 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-NEXT: .functype convert_low_s_v2f64 (v128) -> (v128){{$}}
 ; SIMD128-NEXT: f64x2.convert_low_i32x4_s $push[[R:[0-9]+]]=, $0
 ; SIMD128-NEXT: return $pop[[R]]
 define <2 x double> @convert_low_s_v2f64(<4 x i32> %x) {
   %v = shufflevector <4 x i32> %x, <4 x i32> undef, <2 x i32> <i32 0, i32 1>
   %a = sitofp <2 x i32> %v to <2 x double>

llvm/test/CodeGen/WebAssembly/simd-intrinsics.ll

-; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+simd128 | FileCheck %s
+; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+simd128 | FileCheck %s --check-prefixes=CHECK,SLOW
 ; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+simd128 -fast-isel | FileCheck %s
 
 ; Test that SIMD128 intrinsics lower as expected. These intrinsics are
 ; only expected to lower successfully if the simd128 attribute is
 ; enabled and legal types are used.
 
 target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
 target triple = "wasm32-unknown-unknown"

 declare <2 x i32> @llvm.fptosi.sat.v2i32.v2f64(<2 x double>)
 define <4 x i32> @trunc_sat_zero_s_v4i32(<2 x double> %x) {
   %v = call <2 x i32> @llvm.fptosi.sat.v2i32.v2f64(<2 x double> %x)
   %a = shufflevector <2 x i32> %v, <2 x i32> <i32 0, i32 0>,
            <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   ret <4 x i32> %a
 }
 
+; CHECK-LABEL: trunc_sat_zero_s_v4i32_2:
+; CHECK-NEXT: .functype trunc_sat_zero_s_v4i32_2 (v128) -> (v128){{$}}
+; SLOW-NEXT: i32x4.trunc_sat_zero_f64x2_s $push[[R:[0-9]+]]=, $0{{$}}
+; SLOW-NEXT: return $pop[[R]]{{$}}
+declare <4 x i32> @llvm.fptosi.sat.v4i32.v4f64(<4 x double>)
+define <4 x i32> @trunc_sat_zero_s_v4i32_2(<2 x double> %x) {
+  %v = shufflevector <2 x double> %x, <2 x double> zeroinitializer,
+           <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %a = call <4 x i32> @llvm.fptosi.sat.v4i32.v4f64(<4 x double> %v)
+  ret <4 x i32> %a
+}
+
 ; CHECK-LABEL: trunc_sat_zero_u_v4i32:
 ; CHECK-NEXT: .functype trunc_sat_zero_u_v4i32 (v128) -> (v128){{$}}
 ; CHECK-NEXT: i32x4.trunc_sat_zero_f64x2_u $push[[R:[0-9]+]]=, $0{{$}}
 ; CHECK-NEXT: return $pop[[R]]{{$}}
 declare <2 x i32> @llvm.fptoui.sat.v2i32.v2f64(<2 x double>)
 define <4 x i32> @trunc_sat_zero_u_v4i32(<2 x double> %x) {
   %v = call <2 x i32> @llvm.fptoui.sat.v2i32.v2f64(<2 x double> %x)
   %a = shufflevector <2 x i32> %v, <2 x i32> <i32 0, i32 0>,
            <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   ret <4 x i32> %a
 }
 
+; CHECK-LABEL: trunc_sat_zero_u_v4i32_2:
+; CHECK-NEXT: .functype trunc_sat_zero_u_v4i32_2 (v128) -> (v128){{$}}
+; SLOW-NEXT: i32x4.trunc_sat_zero_f64x2_u $push[[R:[0-9]+]]=, $0{{$}}
+; SLOW-NEXT: return $pop[[R]]{{$}}
+declare <4 x i32> @llvm.fptoui.sat.v4i32.v4f64(<4 x double>)
+define <4 x i32> @trunc_sat_zero_u_v4i32_2(<2 x double> %x) {
+  %v = shufflevector <2 x double> %x, <2 x double> zeroinitializer,
+           <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %a = call <4 x i32> @llvm.fptoui.sat.v4i32.v4f64(<4 x double> %v)
+  ret <4 x i32> %a
+}
+
 ; ==============================================================================
 ; 2 x i64
 ; ==============================================================================
 ; CHECK-LABEL: extmul_low_s_v2i64:
 ; CHECK-NEXT: .functype extmul_low_s_v2i64 (v128, v128) -> (v128){{$}}
 ; CHECK-NEXT: i64x2.extmul_low_i32x4_s $push[[R:[0-9]+]]=, $0, $1{{$}}
 ; CHECK-NEXT: return $pop[[R]]{{$}}
 declare <2 x i64> @llvm.wasm.extmul.low.signed.v2i64(<4 x i32>, <4 x i32>)

 ; CHECK-NEXT: f32x4.nearest $push[[R:[0-9]+]]=, $0{{$}}
 ; CHECK-NEXT: return $pop[[R]]{{$}}
 declare <4 x float> @llvm.nearbyint.v4f32(<4 x float>)
 define <4 x float> @nearest_v4f32(<4 x float> %a) {
   %v = call <4 x float> @llvm.nearbyint.v4f32(<4 x float> %a)
   ret <4 x float> %v
 }
 
-; CHECK-LABEL: demote_zero_v4f32:
-; CHECK-NEXT: .functype demote_zero_v4f32 (v128) -> (v128){{$}}
-; CHECK-NEXT: f32x4.demote_zero_f64x2 $push[[R:[0-9]+]]=, $0{{$}}
-; CHECK-NEXT: return $pop[[R]]{{$}}
-declare <4 x float> @llvm.wasm.demote.zero(<2 x double>)
-define <4 x float> @demote_zero_v4f32(<2 x double> %a) {
-  %v = call <4 x float> @llvm.wasm.demote.zero(<2 x double> %a)
-  ret <4 x float> %v
-}
-
 ; ==============================================================================
 ; 2 x f64
 ; ==============================================================================
 ; CHECK-LABEL: bitselect_v2f64:
 ; CHECK-NEXT: .functype bitselect_v2f64 (v128, v128, v128) -> (v128){{$}}
 ; CHECK-NEXT: v128.bitselect $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
 ; CHECK-NEXT: return $pop[[R]]{{$}}
 declare <2 x double> @llvm.wasm.bitselect.v2f64(<2 x double>, <2 x double>, <2 x double>)