[WebAssembly] Remove intrinsics for SIMD widening ops

[tlively]

Summary

Instead, pattern match extends of extract_subvectors to generate
widening operations. Since extract_subvector is not a legal node, this
is implemented via a custom combine that recognizes extract_subvector
nodes before they are legalized. The combine produces custom ISD nodes
that are later pattern matched directly, just like the intrinsic was.

Also removes the clang builtins for these operations since the
instructions can now be generated from portable code sequences.

[aheejin]

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

113

It'd be clearer to say starting indices of these don't start with 0 or [lanecount - 1] so they can't be widened using widen_low or widen_high instructions.

Question: Can we also widen these using shifts?

[tlively]

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

113

Sure, since I didn't end up testing more patterns, I can make the comment more specific.

Regarding shifts, I don't think it's possible to do widening with shifts because widening has to fundamentally change the number of lanes, which shifts can't do.

[aheejin]

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

113

What I meant was, in case of i16x8->i32x4, the current code can widen i16x8 input vector with elements in the indices 0 to 4. If those elements are instead in 1 to 5, can we first shift that to 0~4 and widen it?

[tlively]

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

113

Oh gotcha. No, unfortunately I don't think that would work. The SIMD shift instructions shift bytes within lanes but they can't shift data into a different lane. Even if we used shifts on larger lanes to try to overcome that limitation, a 64x2 shift would still not be able to shift data from the high half of the vector to the low half or vice versa, which we would need to do to implement your suggestion.

[aheejin]

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

113

Ah right... I was confused about SIMD shifts. Thanks.

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, "V16cV8sV8s", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_narrow_u_i8x16_i16x8, "V16cV8sV8s", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_narrow_s_i16x8_i32x4, "V8sV4iV4i", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_narrow_u_i16x8_i32x4, "V8sV4iV4i", "nc", "simd128")
 
-TARGET_BUILTIN(__builtin_wasm_widen_low_s_i16x8_i8x16, "V8sV16c", "nc", "simd128")
-TARGET_BUILTIN(__builtin_wasm_widen_high_s_i16x8_i8x16, "V8sV16c", "nc", "simd128")
-TARGET_BUILTIN(__builtin_wasm_widen_low_u_i16x8_i8x16, "V8sV16c", "nc", "simd128")
-TARGET_BUILTIN(__builtin_wasm_widen_high_u_i16x8_i8x16, "V8sV16c", "nc", "simd128")
-TARGET_BUILTIN(__builtin_wasm_widen_low_s_i32x4_i16x8, "V4iV8s", "nc", "simd128")
-TARGET_BUILTIN(__builtin_wasm_widen_high_s_i32x4_i16x8, "V4iV8s", "nc", "simd128")
-TARGET_BUILTIN(__builtin_wasm_widen_low_u_i32x4_i16x8, "V4iV8s", "nc", "simd128")
-TARGET_BUILTIN(__builtin_wasm_widen_high_u_i32x4_i16x8, "V4iV8s", "nc", "simd128")
-
 #undef BUILTIN
 #undef TARGET_BUILTIN

clang/lib/CodeGen/CGBuiltin.cpp

case WebAssembly::BI__builtin_wasm_narrow_u_i16x8_i32x4:
       break;
     default:
       llvm_unreachable("unexpected builtin ID");
     }
     Function *Callee =
         CGM.getIntrinsic(IntNo, {ConvertType(E->getType()), Low->getType()});
     return Builder.CreateCall(Callee, {Low, High});
   }
-  case WebAssembly::BI__builtin_wasm_widen_low_s_i16x8_i8x16:
-  case WebAssembly::BI__builtin_wasm_widen_high_s_i16x8_i8x16:
-  case WebAssembly::BI__builtin_wasm_widen_low_u_i16x8_i8x16:
-  case WebAssembly::BI__builtin_wasm_widen_high_u_i16x8_i8x16:
-  case WebAssembly::BI__builtin_wasm_widen_low_s_i32x4_i16x8:
-  case WebAssembly::BI__builtin_wasm_widen_high_s_i32x4_i16x8:
-  case WebAssembly::BI__builtin_wasm_widen_low_u_i32x4_i16x8:
-  case WebAssembly::BI__builtin_wasm_widen_high_u_i32x4_i16x8: {
-    Value *Vec = EmitScalarExpr(E->getArg(0));
-    unsigned IntNo;
-    switch (BuiltinID) {
-    case WebAssembly::BI__builtin_wasm_widen_low_s_i16x8_i8x16:
-    case WebAssembly::BI__builtin_wasm_widen_low_s_i32x4_i16x8:
-      IntNo = Intrinsic::wasm_widen_low_signed;
-      break;
-    case WebAssembly::BI__builtin_wasm_widen_high_s_i16x8_i8x16:
-    case WebAssembly::BI__builtin_wasm_widen_high_s_i32x4_i16x8:
-      IntNo = Intrinsic::wasm_widen_high_signed;
-      break;
-    case WebAssembly::BI__builtin_wasm_widen_low_u_i16x8_i8x16:
-    case WebAssembly::BI__builtin_wasm_widen_low_u_i32x4_i16x8:
-      IntNo = Intrinsic::wasm_widen_low_unsigned;
-      break;
-    case WebAssembly::BI__builtin_wasm_widen_high_u_i16x8_i8x16:
-    case WebAssembly::BI__builtin_wasm_widen_high_u_i32x4_i16x8:
-      IntNo = Intrinsic::wasm_widen_high_unsigned;
-      break;
-    default:
-      llvm_unreachable("unexpected builtin ID");
-    }
-    Function *Callee =
-        CGM.getIntrinsic(IntNo, {ConvertType(E->getType()), Vec->getType()});
-    return Builder.CreateCall(Callee, Vec);
-  }
   case WebAssembly::BI__builtin_wasm_shuffle_v8x16: {
     Value *Ops[18];
     size_t OpIdx = 0;
     Ops[OpIdx++] = EmitScalarExpr(E->getArg(0));
     Ops[OpIdx++] = EmitScalarExpr(E->getArg(1));
     while (OpIdx < 18) {
       Optional<llvm::APSInt> LaneConst =
           E->getArg(OpIdx)->getIntegerConstantExpr(getContext());

clang/lib/Headers/wasm_simd128.h

 typedef unsigned int __u32x4
     __attribute__((__vector_size__(16), __aligned__(16)));
 typedef long long __i64x2 __attribute__((__vector_size__(16), __aligned__(16)));
 typedef unsigned long long __u64x2
     __attribute__((__vector_size__(16), __aligned__(16)));
 typedef float __f32x4 __attribute__((__vector_size__(16), __aligned__(16)));
 typedef double __f64x2 __attribute__((__vector_size__(16), __aligned__(16)));
 
+typedef signed char __i8x8 __attribute__((__vector_size__(8), __aligned__(8)));
+typedef unsigned char __u8x8
+    __attribute__((__vector_size__(8), __aligned__(8)));
+typedef short __i16x4 __attribute__((__vector_size__(8), __aligned__(8)));
+typedef unsigned short __u16x4
+    __attribute__((__vector_size__(8), __aligned__(8)));
+
 #define __DEFAULT_FN_ATTRS                                                     \
   __attribute__((__always_inline__, __nodebug__, __target__("simd128"),        \
                  __min_vector_width__(128)))
 
 #define __REQUIRE_CONSTANT(e)                                                  \
   _Static_assert(__builtin_constant_p(e), "Expected constant")
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS wasm_v128_load(const void *__mem) {

 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_u16x8_narrow_i32x4(v128_t __a, v128_t __b) {
   return (v128_t)__builtin_wasm_narrow_u_i16x8_i32x4((__i32x4)__a,
                                                      (__i32x4)__b);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_i16x8_widen_low_i8x16(v128_t __a) {
-  return (v128_t)__builtin_wasm_widen_low_s_i16x8_i8x16((__i8x16)__a);
+  return (v128_t) __builtin_convertvector(
+      (__i8x8){((__i8x16)__a)[0], ((__i8x16)__a)[1], ((__i8x16)__a)[2],
+               ((__i8x16)__a)[3], ((__i8x16)__a)[4], ((__i8x16)__a)[5],
+               ((__i8x16)__a)[6], ((__i8x16)__a)[7]},
+      __i16x8);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_i16x8_widen_high_i8x16(v128_t __a) {
-  return (v128_t)__builtin_wasm_widen_high_s_i16x8_i8x16((__i8x16)__a);
+  return (v128_t) __builtin_convertvector(
+      (__i8x8){((__i8x16)__a)[8], ((__i8x16)__a)[9], ((__i8x16)__a)[10],
+               ((__i8x16)__a)[11], ((__i8x16)__a)[12], ((__i8x16)__a)[13],
+               ((__i8x16)__a)[14], ((__i8x16)__a)[15]},
+      __i16x8);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_i16x8_widen_low_u8x16(v128_t __a) {
-  return (v128_t)__builtin_wasm_widen_low_u_i16x8_i8x16((__i8x16)__a);
+  return (v128_t) __builtin_convertvector(
+      (__u8x8){((__u8x16)__a)[0], ((__u8x16)__a)[1], ((__u8x16)__a)[2],
+               ((__u8x16)__a)[3], ((__u8x16)__a)[4], ((__u8x16)__a)[5],
+               ((__u8x16)__a)[6], ((__u8x16)__a)[7]},
+      __u16x8);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_i16x8_widen_high_u8x16(v128_t __a) {
-  return (v128_t)__builtin_wasm_widen_high_u_i16x8_i8x16((__i8x16)__a);
+  return (v128_t) __builtin_convertvector(
+      (__u8x8){((__u8x16)__a)[8], ((__u8x16)__a)[9], ((__u8x16)__a)[10],
+               ((__u8x16)__a)[11], ((__u8x16)__a)[12], ((__u8x16)__a)[13],
+               ((__u8x16)__a)[14], ((__u8x16)__a)[15]},
+      __u16x8);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_i32x4_widen_low_i16x8(v128_t __a) {
-  return (v128_t)__builtin_wasm_widen_low_s_i32x4_i16x8((__i16x8)__a);
+  return (v128_t) __builtin_convertvector(
+      (__i16x4){((__i16x8)__a)[0], ((__i16x8)__a)[1], ((__i16x8)__a)[2],
+                ((__i16x8)__a)[3]},
+      __i32x4);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_i32x4_widen_high_i16x8(v128_t __a) {
-  return (v128_t)__builtin_wasm_widen_high_s_i32x4_i16x8((__i16x8)__a);
+  return (v128_t) __builtin_convertvector(
+      (__i16x4){((__i16x8)__a)[4], ((__i16x8)__a)[5], ((__i16x8)__a)[6],
+                ((__i16x8)__a)[7]},
+      __i32x4);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_i32x4_widen_low_u16x8(v128_t __a) {
-  return (v128_t)__builtin_wasm_widen_low_u_i32x4_i16x8((__i16x8)__a);
+  return (v128_t) __builtin_convertvector(
+      (__u16x4){((__u16x8)__a)[0], ((__u16x8)__a)[1], ((__u16x8)__a)[2],
+                ((__u16x8)__a)[3]},
+      __u32x4);
 }
 
 static __inline__ v128_t __DEFAULT_FN_ATTRS
 wasm_i32x4_widen_high_u16x8(v128_t __a) {
-  return (v128_t)__builtin_wasm_widen_high_u_i32x4_i16x8((__i16x8)__a);
+  return (v128_t) __builtin_convertvector(
+      (__u16x4){((__u16x8)__a)[4], ((__u16x8)__a)[5], ((__u16x8)__a)[6],
+                ((__u16x8)__a)[7]},
+      __u32x4);
 }
 
 // Undefine helper macros
 #undef __DEFAULT_FN_ATTRS
 
 #endif // __WASM_SIMD128_H

clang/test/CodeGen/builtins-wasm.c

 
 i16x8 narrow_u_i16x8_i32x4(i32x4 low, i32x4 high) {
   return __builtin_wasm_narrow_u_i16x8_i32x4(low, high);
   // WEBASSEMBLY: call <8 x i16> @llvm.wasm.narrow.unsigned.v8i16.v4i32(
   // WEBASSEMBLY-SAME: <4 x i32> %low, <4 x i32> %high)
   // WEBASSEMBLY: ret
 }
 
-i16x8 widen_low_s_i16x8_i8x16(i8x16 v) {
-  return __builtin_wasm_widen_low_s_i16x8_i8x16(v);
-  // WEBASSEMBLY: call <8 x i16> @llvm.wasm.widen.low.signed.v8i16.v16i8(<16 x i8> %v)
-  // WEBASSEMBLY: ret
-}
-
-i16x8 widen_high_s_i16x8_i8x16(i8x16 v) {
-  return __builtin_wasm_widen_high_s_i16x8_i8x16(v);
-  // WEBASSEMBLY: call <8 x i16> @llvm.wasm.widen.high.signed.v8i16.v16i8(<16 x i8> %v)
-  // WEBASSEMBLY: ret
-}
-
-i16x8 widen_low_u_i16x8_i8x16(i8x16 v) {
-  return __builtin_wasm_widen_low_u_i16x8_i8x16(v);
-  // WEBASSEMBLY: call <8 x i16> @llvm.wasm.widen.low.unsigned.v8i16.v16i8(<16 x i8> %v)
-  // WEBASSEMBLY: ret
-}
-
-i16x8 widen_high_u_i16x8_i8x16(i8x16 v) {
-  return __builtin_wasm_widen_high_u_i16x8_i8x16(v);
-  // WEBASSEMBLY: call <8 x i16> @llvm.wasm.widen.high.unsigned.v8i16.v16i8(<16 x i8> %v)
-  // WEBASSEMBLY: ret
-}
-
-i32x4 widen_low_s_i32x4_i16x8(i16x8 v) {
-  return __builtin_wasm_widen_low_s_i32x4_i16x8(v);
-  // WEBASSEMBLY: call <4 x i32> @llvm.wasm.widen.low.signed.v4i32.v8i16(<8 x i16> %v)
-  // WEBASSEMBLY: ret
-}
-
-i32x4 widen_high_s_i32x4_i16x8(i16x8 v) {
-  return __builtin_wasm_widen_high_s_i32x4_i16x8(v);
-  // WEBASSEMBLY: call <4 x i32> @llvm.wasm.widen.high.signed.v4i32.v8i16(<8 x i16> %v)
-  // WEBASSEMBLY: ret
-}
-
-i32x4 widen_low_u_i32x4_i16x8(i16x8 v) {
-  return __builtin_wasm_widen_low_u_i32x4_i16x8(v);
-  // WEBASSEMBLY: call <4 x i32> @llvm.wasm.widen.low.unsigned.v4i32.v8i16(<8 x i16> %v)
-  // WEBASSEMBLY: ret
-}
-
-i32x4 widen_high_u_i32x4_i16x8(i16x8 v) {
-  return __builtin_wasm_widen_high_u_i32x4_i16x8(v);
-  // WEBASSEMBLY: call <4 x i32> @llvm.wasm.widen.high.unsigned.v4i32.v8i16(<8 x i16> %v)
-  // WEBASSEMBLY: ret
-}
-
 i8x16 swizzle_v8x16(i8x16 x, i8x16 y) {
   return __builtin_wasm_swizzle_v8x16(x, y);
   // WEBASSEMBLY: call <16 x i8> @llvm.wasm.swizzle(<16 x i8> %x, <16 x i8> %y)
 }
 
 i8x16 shuffle(i8x16 x, i8x16 y) {
   return __builtin_wasm_shuffle_v8x16(x, y, 0, 1, 2, 3, 4, 5, 6, 7,
                                       8, 9, 10, 11, 12, 13, 14, 15);
   // WEBASSEMBLY: call <16 x i8> @llvm.wasm.shuffle(<16 x i8> %x, <16 x i8> %y,
   // WEBASSEMBLY-SAME: i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7,
   // WEBASSEMBLY-SAME: i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14,
   // WEBASSEMBLY-SAME: i32 15
   // WEBASSEMBLY-NEXT: ret
 }

llvm/include/llvm/IR/IntrinsicsWebAssembly.td

 def int_wasm_narrow_signed :
   Intrinsic<[llvm_anyvector_ty],
             [llvm_anyvector_ty, LLVMMatchType<1>],
             [IntrNoMem, IntrSpeculatable]>;
 def int_wasm_narrow_unsigned :
   Intrinsic<[llvm_anyvector_ty],
             [llvm_anyvector_ty, LLVMMatchType<1>],
             [IntrNoMem, IntrSpeculatable]>;
-def int_wasm_widen_low_signed :
-  Intrinsic<[llvm_anyvector_ty],
-            [llvm_anyvector_ty],
-            [IntrNoMem, IntrSpeculatable]>;
-def int_wasm_widen_high_signed :
-  Intrinsic<[llvm_anyvector_ty],
-            [llvm_anyvector_ty],
-            [IntrNoMem, IntrSpeculatable]>;
-def int_wasm_widen_low_unsigned :
-  Intrinsic<[llvm_anyvector_ty],
-            [llvm_anyvector_ty],
-            [IntrNoMem, IntrSpeculatable]>;
-def int_wasm_widen_high_unsigned :
-  Intrinsic<[llvm_anyvector_ty],
-            [llvm_anyvector_ty],
-            [IntrNoMem, IntrSpeculatable]>;
 
 // TODO: Replace these intrinsics with normal ISel patterns
 def int_wasm_pmin :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMMatchType<0>, LLVMMatchType<0>],
             [IntrNoMem, IntrSpeculatable]>;
 def int_wasm_pmax :
   Intrinsic<[llvm_anyvector_ty],

llvm/lib/Target/WebAssembly/WebAssemblyISD.def

 HANDLE_NODETYPE(WrapperPIC)
 HANDLE_NODETYPE(BR_IF)
 HANDLE_NODETYPE(BR_TABLE)
 HANDLE_NODETYPE(SHUFFLE)
 HANDLE_NODETYPE(SWIZZLE)
 HANDLE_NODETYPE(VEC_SHL)
 HANDLE_NODETYPE(VEC_SHR_S)
 HANDLE_NODETYPE(VEC_SHR_U)
+HANDLE_NODETYPE(WIDEN_LOW_S)
+HANDLE_NODETYPE(WIDEN_LOW_U)
+HANDLE_NODETYPE(WIDEN_HIGH_S)
+HANDLE_NODETYPE(WIDEN_HIGH_U)
 HANDLE_NODETYPE(THROW)
 HANDLE_NODETYPE(MEMORY_COPY)
 HANDLE_NODETYPE(MEMORY_FILL)
 
 // Memory intrinsics
 HANDLE_MEM_NODETYPE(LOAD_SPLAT)

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

if (Subtarget->hasSIMD128())
         setOperationAction(Op, T, Expand);
   }
 
   // SIMD-specific configuration
   if (Subtarget->hasSIMD128()) {
     // Hoist bitcasts out of shuffles
     setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
 
+    // Combine extends of extract_subvectors into widening ops
+    setTargetDAGCombine(ISD::SIGN_EXTEND);
+    setTargetDAGCombine(ISD::ZERO_EXTEND);
+
     // 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
     for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
       setOperationAction(ISD::ABS, T, Legal);
performVECTOR_SHUFFLECombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
   if (!SrcType.is128BitVector() ||
       SrcType.getVectorNumElements() != DstType.getVectorNumElements())
     return SDValue();
   SDValue NewShuffle = DAG.getVectorShuffle(
       SrcType, SDLoc(N), CastOp, DAG.getUNDEF(SrcType), Shuffle->getMask());
   return DAG.getBitcast(DstType, NewShuffle);
 }
 
+static SDValue performVectorWidenCombine(SDNode *N,
+                                         TargetLowering::DAGCombinerInfo &DCI) {
+  auto &DAG = DCI.DAG;
+  assert(N->getOpcode() == ISD::SIGN_EXTEND ||
+         N->getOpcode() == ISD::ZERO_EXTEND);
+
+  // Combine ({s,z}ext (extract_subvector src, i)) into a widening operation if
+  // possible before the extract_subvector can be expanded.
+  auto Extract = N->getOperand(0);
+  if (Extract.getOpcode() != ISD::EXTRACT_SUBVECTOR)
+    return SDValue();
+  auto Source = Extract.getOperand(0);
+  auto *IndexNode = dyn_cast<ConstantSDNode>(Extract.getOperand(1));
+  if (IndexNode == nullptr)
+    return SDValue();
+  auto Index = IndexNode->getZExtValue();
+
+  // Only v8i8 and v4i16 extracts can be widened, and only if the extracted
+  // subvector is the low or high half of its source.
+  EVT ResVT = N->getValueType(0);
+  if (ResVT == MVT::v8i16) {
+    if (Extract.getValueType() != MVT::v8i8 ||
+        Source.getValueType() != MVT::v16i8 || (Index != 0 && Index != 8))
+      return SDValue();
+  } else if (ResVT == MVT::v4i32) {
+    if (Extract.getValueType() != MVT::v4i16 ||
+        Source.getValueType() != MVT::v8i16 || (Index != 0 && Index != 4))
+      return SDValue();
+  } else {
+    return SDValue();
+  }
+
+  bool IsSext = N->getOpcode() == ISD::SIGN_EXTEND;
+  bool IsLow = Index == 0;
+
+  unsigned Op = IsSext ? (IsLow ? WebAssemblyISD::WIDEN_LOW_S
+                                : WebAssemblyISD::WIDEN_HIGH_S)
+                       : (IsLow ? WebAssemblyISD::WIDEN_LOW_U
+                                : WebAssemblyISD::WIDEN_HIGH_U);
+
+  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 performVectorWidenCombine(N, DCI);
   }
 }

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

 
 // Lower llvm.wasm.trunc.saturate.* to saturating instructions
 def : Pat<(v4i32 (int_wasm_trunc_saturate_signed (v4f32 V128:$src))),
           (fp_to_sint_v4i32_v4f32 (v4f32 V128:$src))>;
 def : Pat<(v4i32 (int_wasm_trunc_saturate_unsigned (v4f32 V128:$src))),
           (fp_to_uint_v4i32_v4f32 (v4f32 V128:$src))>;
 
 // Widening operations
+def widen_t : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>]>;
+def widen_low_s : SDNode<"WebAssemblyISD::WIDEN_LOW_S", widen_t>;
+def widen_high_s : SDNode<"WebAssemblyISD::WIDEN_HIGH_S", widen_t>;
+def widen_low_u : SDNode<"WebAssemblyISD::WIDEN_LOW_U", widen_t>;
+def widen_high_u : SDNode<"WebAssemblyISD::WIDEN_HIGH_U", widen_t>;
+
 multiclass SIMDWiden<ValueType vec_t, string vec, ValueType arg_t, string arg,
                      bits<32> baseInst> {
-  defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_low_signed,
+  defm "" : SIMDConvert<vec_t, arg_t, widen_low_s,
                         vec#".widen_low_"#arg#"_s", baseInst>;
-  defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_high_signed,
+  defm "" : SIMDConvert<vec_t, arg_t, widen_high_s,
                         vec#".widen_high_"#arg#"_s", !add(baseInst, 1)>;
-  defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_low_unsigned,
+  defm "" : SIMDConvert<vec_t, arg_t, widen_low_u,
                         vec#".widen_low_"#arg#"_u", !add(baseInst, 2)>;
-  defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_high_unsigned,
+  defm "" : SIMDConvert<vec_t, arg_t, widen_high_u,
                         vec#".widen_high_"#arg#"_u", !add(baseInst, 3)>;
 }
 
 defm "" : SIMDWiden<v8i16, "i16x8", v16i8, "i8x16", 135>;
 defm "" : SIMDWiden<v4i32, "i32x4", v8i16, "i16x8", 167>;
 
 // Narrowing operations
 multiclass SIMDNarrow<ValueType vec_t, string vec, ValueType arg_t, string arg,

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

 declare <8 x i16> @llvm.wasm.narrow.unsigned.v8i16.v4i32(<4 x i32>, <4 x i32>)
 define <8 x i16> @narrow_unsigned_v8i16(<4 x i32> %low, <4 x i32> %high) {
   %a = call <8 x i16> @llvm.wasm.narrow.unsigned.v8i16.v4i32(
     <4 x i32> %low, <4 x i32> %high
   )
   ret <8 x i16> %a
 }
 
-; CHECK-LABEL: widen_low_signed_v8i16:
-; SIMD128-NEXT: .functype widen_low_signed_v8i16 (v128) -> (v128){{$}}
-; SIMD128-NEXT: i16x8.widen_low_i8x16_s $push[[R:[0-9]+]]=, $0{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
-declare <8 x i16> @llvm.wasm.widen.low.signed.v8i16.v16i8(<16 x i8>)
-define <8 x i16> @widen_low_signed_v8i16(<16 x i8> %v) {
-  %a = call <8 x i16> @llvm.wasm.widen.low.signed.v8i16.v16i8(<16 x i8> %v)
-  ret <8 x i16> %a
-}
-
-; CHECK-LABEL: widen_high_signed_v8i16:
-; SIMD128-NEXT: .functype widen_high_signed_v8i16 (v128) -> (v128){{$}}
-; SIMD128-NEXT: i16x8.widen_high_i8x16_s $push[[R:[0-9]+]]=, $0{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
-declare <8 x i16> @llvm.wasm.widen.high.signed.v8i16.v16i8(<16 x i8>)
-define <8 x i16> @widen_high_signed_v8i16(<16 x i8> %v) {
-  %a = call <8 x i16> @llvm.wasm.widen.high.signed.v8i16.v16i8(<16 x i8> %v)
-  ret <8 x i16> %a
-}
-
-; CHECK-LABEL: widen_low_unsigned_v8i16:
-; SIMD128-NEXT: .functype widen_low_unsigned_v8i16 (v128) -> (v128){{$}}
-; SIMD128-NEXT: i16x8.widen_low_i8x16_u $push[[R:[0-9]+]]=, $0{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
-declare <8 x i16> @llvm.wasm.widen.low.unsigned.v8i16.v16i8(<16 x i8>)
-define <8 x i16> @widen_low_unsigned_v8i16(<16 x i8> %v) {
-  %a = call <8 x i16> @llvm.wasm.widen.low.unsigned.v8i16.v16i8(<16 x i8> %v)
-  ret <8 x i16> %a
-}
-
-; CHECK-LABEL: widen_high_unsigned_v8i16:
-; SIMD128-NEXT: .functype widen_high_unsigned_v8i16 (v128) -> (v128){{$}}
-; SIMD128-NEXT: i16x8.widen_high_i8x16_u $push[[R:[0-9]+]]=, $0{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
-declare <8 x i16> @llvm.wasm.widen.high.unsigned.v8i16.v16i8(<16 x i8>)
-define <8 x i16> @widen_high_unsigned_v8i16(<16 x i8> %v) {
-  %a = call <8 x i16> @llvm.wasm.widen.high.unsigned.v8i16.v16i8(<16 x i8> %v)
-  ret <8 x i16> %a
-}
-
 ; ==============================================================================
 ; 4 x i32
 ; ==============================================================================
 ; CHECK-LABEL: dot:
 ; SIMD128-NEXT: .functype dot (v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: i32x4.dot_i16x8_s $push[[R:[0-9]+]]=, $0, $1{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 declare <4 x i32> @llvm.wasm.dot(<8 x i16>, <8 x i16>)

 ; SIMD128-NEXT: i32x4.trunc_sat_f32x4_u $push[[R:[0-9]+]]=, $0
 ; SIMD128-NEXT: return $pop[[R]]
 declare <4 x i32> @llvm.wasm.trunc.saturate.unsigned.v4i32.v4f32(<4 x float>)
 define <4 x i32> @trunc_sat_u_v4i32(<4 x float> %x) {
   %a = call <4 x i32> @llvm.wasm.trunc.saturate.unsigned.v4i32.v4f32(<4 x float> %x)
   ret <4 x i32> %a
 }
 
-; CHECK-LABEL: widen_low_signed_v4i32:
-; SIMD128-NEXT: .functype widen_low_signed_v4i32 (v128) -> (v128){{$}}
-; SIMD128-NEXT: i32x4.widen_low_i16x8_s $push[[R:[0-9]+]]=, $0{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
-declare <4 x i32> @llvm.wasm.widen.low.signed.v4i32.v8i16(<8 x i16>)
-define <4 x i32> @widen_low_signed_v4i32(<8 x i16> %v) {
-  %a = call <4 x i32> @llvm.wasm.widen.low.signed.v4i32.v8i16(<8 x i16> %v)
-  ret <4 x i32> %a
-}
-
-; CHECK-LABEL: widen_high_signed_v4i32:
-; SIMD128-NEXT: .functype widen_high_signed_v4i32 (v128) -> (v128){{$}}
-; SIMD128-NEXT: i32x4.widen_high_i16x8_s $push[[R:[0-9]+]]=, $0{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
-declare <4 x i32> @llvm.wasm.widen.high.signed.v4i32.v8i16(<8 x i16>)
-define <4 x i32> @widen_high_signed_v4i32(<8 x i16> %v) {
-  %a = call <4 x i32> @llvm.wasm.widen.high.signed.v4i32.v8i16(<8 x i16> %v)
-  ret <4 x i32> %a
-}
-
-; CHECK-LABEL: widen_low_unsigned_v4i32:
-; SIMD128-NEXT: .functype widen_low_unsigned_v4i32 (v128) -> (v128){{$}}
-; SIMD128-NEXT: i32x4.widen_low_i16x8_u $push[[R:[0-9]+]]=, $0{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
-declare <4 x i32> @llvm.wasm.widen.low.unsigned.v4i32.v8i16(<8 x i16>)
-define <4 x i32> @widen_low_unsigned_v4i32(<8 x i16> %v) {
-  %a = call <4 x i32> @llvm.wasm.widen.low.unsigned.v4i32.v8i16(<8 x i16> %v)
-  ret <4 x i32> %a
-}
-
-; CHECK-LABEL: widen_high_unsigned_v4i32:
-; SIMD128-NEXT: .functype widen_high_unsigned_v4i32 (v128) -> (v128){{$}}
-; SIMD128-NEXT: i32x4.widen_high_i16x8_u $push[[R:[0-9]+]]=, $0{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
-declare <4 x i32> @llvm.wasm.widen.high.unsigned.v4i32.v8i16(<8 x i16>)
-define <4 x i32> @widen_high_unsigned_v4i32(<8 x i16> %v) {
-  %a = call <4 x i32> @llvm.wasm.widen.high.unsigned.v4i32.v8i16(<8 x i16> %v)
-  ret <4 x i32> %a
-}
-
 ; ==============================================================================
 ; 2 x i64
 ; ==============================================================================
 ; CHECK-LABEL: any_v2i64:
 ; SIMD128-NEXT: .functype any_v2i64 (v128) -> (i32){{$}}
 ; SIMD128-NEXT: i64x2.any_true $push[[R:[0-9]+]]=, $0{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 declare i32 @llvm.wasm.anytrue.v2i64(<2 x i64>)

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

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mattr=+simd128 | FileCheck %s
+
+;; Test that SIMD widening operations can be successfully selected
+
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32-unknown-unknown"
+
+define <8 x i16> @widen_low_i8x16_s(<16 x i8> %v) {
+; CHECK-LABEL: widen_low_i8x16_s:
+; CHECK:         .functype widen_low_i8x16_s (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i16x8.widen_low_i8x16_s
+; CHECK-NEXT:    # fallthrough-return
+  %low = shufflevector <16 x i8> %v, <16 x i8> undef,
+           <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %widened = sext <8 x i8> %low to <8 x i16>
+  ret <8 x i16> %widened
+}
+
+define <8 x i16> @widen_low_i8x16_u(<16 x i8> %v) {
+; CHECK-LABEL: widen_low_i8x16_u:
+; CHECK:         .functype widen_low_i8x16_u (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i16x8.widen_low_i8x16_u
+; CHECK-NEXT:    # fallthrough-return
+  %low = shufflevector <16 x i8> %v, <16 x i8> undef,
+           <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %widened = zext <8 x i8> %low to <8 x i16>
+  ret <8 x i16> %widened
+}
+
+define <8 x i16> @widen_high_i8x16_s(<16 x i8> %v) {
+; CHECK-LABEL: widen_high_i8x16_s:
+; CHECK:         .functype widen_high_i8x16_s (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i16x8.widen_high_i8x16_s
+; CHECK-NEXT:    # fallthrough-return
+  %low = shufflevector <16 x i8> %v, <16 x i8> undef,
+           <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %widened = sext <8 x i8> %low to <8 x i16>
+  ret <8 x i16> %widened
+}
+
+define <8 x i16> @widen_high_i8x16_u(<16 x i8> %v) {
+; CHECK-LABEL: widen_high_i8x16_u:
+; CHECK:         .functype widen_high_i8x16_u (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i16x8.widen_high_i8x16_u
+; CHECK-NEXT:    # fallthrough-return
+  %low = shufflevector <16 x i8> %v, <16 x i8> undef,
+           <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %widened = zext <8 x i8> %low to <8 x i16>
+  ret <8 x i16> %widened
+}
+
+define <4 x i32> @widen_low_i16x8_s(<8 x i16> %v) {
+; CHECK-LABEL: widen_low_i16x8_s:
+; CHECK:         .functype widen_low_i16x8_s (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i32x4.widen_low_i16x8_s
+; CHECK-NEXT:    # fallthrough-return
+  %low = shufflevector <8 x i16> %v, <8 x i16> undef,
+           <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %widened = sext <4 x i16> %low to <4 x i32>
+  ret <4 x i32> %widened
+}
+
+define <4 x i32> @widen_low_i16x8_u(<8 x i16> %v) {
+; CHECK-LABEL: widen_low_i16x8_u:
+; CHECK:         .functype widen_low_i16x8_u (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i32x4.widen_low_i16x8_u
+; CHECK-NEXT:    # fallthrough-return
+  %low = shufflevector <8 x i16> %v, <8 x i16> undef,
+           <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %widened = zext <4 x i16> %low to <4 x i32>
+  ret <4 x i32> %widened
+}
+
+define <4 x i32> @widen_high_i16x8_s(<8 x i16> %v) {
+; CHECK-LABEL: widen_high_i16x8_s:
+; CHECK:         .functype widen_high_i16x8_s (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i32x4.widen_high_i16x8_s
+; CHECK-NEXT:    # fallthrough-return
+  %low = shufflevector <8 x i16> %v, <8 x i16> undef,
+           <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %widened = sext <4 x i16> %low to <4 x i32>
+  ret <4 x i32> %widened
+}
+
+define <4 x i32> @widen_high_i16x8_u(<8 x i16> %v) {
+; CHECK-LABEL: widen_high_i16x8_u:
+; CHECK:         .functype widen_high_i16x8_u (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i32x4.widen_high_i16x8_u
+; CHECK-NEXT:    # fallthrough-return
+  %low = shufflevector <8 x i16> %v, <8 x i16> undef,
+           <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %widened = zext <4 x i16> %low to <4 x i32>
+  ret <4 x i32> %widened
+}
+
+;; Also test that similar patterns are still expanded correctly
+
+define <8 x i16> @widen_lowish_i8x16_s(<16 x i8> %v) {
+; CHECK-LABEL: widen_lowish_i8x16_s:
+; CHECK:         .functype widen_lowish_i8x16_s (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.extract_lane_u 1
+; CHECK-NEXT:    i16x8.splat
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.extract_lane_u 2
+; CHECK-NEXT:    i16x8.replace_lane 1
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.extract_lane_u 3
+; CHECK-NEXT:    i16x8.replace_lane 2
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.extract_lane_u 4
+; CHECK-NEXT:    i16x8.replace_lane 3
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.extract_lane_u 5
+; CHECK-NEXT:    i16x8.replace_lane 4
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.extract_lane_u 6
+; CHECK-NEXT:    i16x8.replace_lane 5
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.extract_lane_u 7
+; CHECK-NEXT:    i16x8.replace_lane 6
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i8x16.extract_lane_u 8
+; CHECK-NEXT:    i16x8.replace_lane 7
+; CHECK-NEXT:    i32.const 8
+; CHECK-NEXT:    i16x8.shl
+; CHECK-NEXT:    i32.const 8
+; CHECK-NEXT:    i16x8.shr_s
+; CHECK-NEXT:    # fallthrough-return
+  %lowish = shufflevector <16 x i8> %v, <16 x i8> undef,
+           <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>
+  %widened = sext <8 x i8> %lowish to <8 x i16>
+  ret <8 x i16> %widened
+}
+
+define <4 x i32> @widen_lowish_i16x8_s(<8 x i16> %v) {
+; CHECK-LABEL: widen_lowish_i16x8_s:
+; CHECK:         .functype widen_lowish_i16x8_s (v128) -> (v128)
+; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i16x8.extract_lane_u 1
+; CHECK-NEXT:    i32x4.splat
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i16x8.extract_lane_u 2
+; CHECK-NEXT:    i32x4.replace_lane 1
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i16x8.extract_lane_u 3
+; CHECK-NEXT:    i32x4.replace_lane 2
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i16x8.extract_lane_u 4
+; CHECK-NEXT:    i32x4.replace_lane 3
+; CHECK-NEXT:    i32.const 16
+; CHECK-NEXT:    i32x4.shl
+; CHECK-NEXT:    i32.const 16
+; CHECK-NEXT:    i32x4.shr_s
+; CHECK-NEXT:    # fallthrough-return
+  %lowish = shufflevector <8 x i16> %v, <8 x i16> undef,
+           <4 x i32> <i32 1, i32 2, i32 3, i32 4>
+  %widened = sext <4 x i16> %lowish to <4 x i32>
+  ret <4 x i32> %widened
+}