[WebAssembly] v128.const

[tlively]

Summary

This CL implements v128.const for each vector type. New operand types
are added to ensure the vector contents can be serialized without LEB
encoding.

[aardappel]

Generally looks good.. but what about some tests?

Should also add support for these to the assembler, or in a follow-up?

[tlively]

In D50873#1203386, @aardappel wrote:

Generally looks good.. but what about some tests?

I'm always down for more tests. What tests could I add?

Should also add support for these to the assembler, or in a follow-up?

Can you point me to the code that will need to be modified for this?

[aardappel]

Assembler is in WebAssemblyAsmParser.cpp
There's some disassembler tests in test/MC/Disassembler/wasm.txt

[aheejin]

Instruction selection part looks good. I guess other reviewers can give better feedback on the encoding/disassembling parts.

lib/Target/WebAssembly/Disassembler/WebAssemblyDisassembler.cpp

184

Are these added in order to prevent elements from being LEB-encoded one by one? SIMD immediates do not use LEB encoding?

lib/Target/WebAssembly/MCTargetDesc/WebAssemblyMCCodeEmitter.cpp

94

clang-format this switch-case block and also possibly other modified parts.

112

Fix indentation (I guess it will be fixed anyway if you run clang-format though)

[tlively]

• Add disassembler test
• Formatting
• Assembler support

[tlively]

lib/Target/WebAssembly/Disassembler/WebAssemblyDisassembler.cpp

184

That's correct. The spec says that the vector immediate is actually an array of 16 immediate bytes, i.e. not LEB-encoded. However, ISEL wants to treat v16i8, v8i16, v4i32, etc. all separately rather than converting them all to v16i8 first, so we need these new operand types in order to encode the larger lane values without LEB.

[dschuff]

lib/Target/WebAssembly/Disassembler/WebAssemblyDisassembler.cpp

95

For extra fun template points you could combine this with the FP case below for a single parseImmediate, where you could add something like
if (std::is_floating_point<T>::value) { ... } else { ... }
You can probably even do the conditional statically, although I'm not sure that would be more readable.

[tlively]

• Add disassembler test
• Formatting
• Assembler support
• Combine parseImmediate functions

lib/Target/WebAssembly/Disassembler/WebAssemblyDisassembler.cpp

static bool parseLEBImmediate(MCInst &MI, uint64_t &Size,
   if (Error)
     return false;
   Size += N;
   MI.addOperand(MCOperand::createImm(Val));
   return true;
 }
 
 template <typename T>
+bool parseImmediate(MCInst &MI, uint64_t &Size, ArrayRef<uint8_t> Bytes) {
+  if (Size + sizeof(T) > Bytes.size())
+    return false;
+  T Val;
+  memcpy(&Val, Bytes.data() + Size, sizeof(T));
+  support::endian::byte_swap<T, support::endianness::little>(Val);
+  Size += sizeof(T);
+  MI.addOperand(MCOperand::createImm(static_cast<int64_t>(Val)));
+  return true;
+}
+
+template <typename T>
 bool parseFPImmediate(MCInst &MI, uint64_t &Size, ArrayRef<uint8_t> Bytes) {
   if (Size + sizeof(T) > Bytes.size())
     return false;
   T Val;
   memcpy(&Val, Bytes.data() + Size, sizeof(T));
   support::endian::byte_swap<T, support::endianness::little>(Val);
   Size += sizeof(T);
   MI.addOperand(MCOperand::createFPImm(static_cast<double>(Val)));
case WebAssembly::OPERAND_F32IMM: {
         return MCDisassembler::Fail;
       break;
     }
     case WebAssembly::OPERAND_F64IMM: {
       if (!parseFPImmediate<double>(MI, Size, Bytes))
         return MCDisassembler::Fail;
       break;
     }
+    // Vector lane operands (not LEB encoded).
+    case WebAssembly::OPERAND_VEC_I8IMM: {
+      if (!parseImmediate<uint8_t>(MI, Size, Bytes))
+        return MCDisassembler::Fail;
+      break;
+    }
+    case WebAssembly::OPERAND_VEC_I16IMM: {
+      if (!parseImmediate<uint16_t>(MI, Size, Bytes))
+        return MCDisassembler::Fail;
+      break;
+    }
+    case WebAssembly::OPERAND_VEC_I32IMM: {
+      if (!parseImmediate<uint32_t>(MI, Size, Bytes))
+        return MCDisassembler::Fail;
+      break;
+    }
+    case WebAssembly::OPERAND_VEC_I64IMM: {
+      if (!parseImmediate<uint64_t>(MI, Size, Bytes))
+        return MCDisassembler::Fail;
+      break;
+    }
     case MCOI::OPERAND_REGISTER: {
       // These are NOT actually in the instruction stream, but MC is going to
       // expect operands to be present for them!
       // FIXME: can MC re-generate register assignments or do we have to
       // do this? Since this function decodes a single instruction, we don't
       // have the proper context for tracking an operand stack here.
       MI.addOperand(MCOperand::createReg(0));
       break;
     }
     default:
       llvm_unreachable("Unknown operand type in WebAssemblyDisassembler");
     }
   }
   return MCDisassembler::Success;
 }

lib/Target/WebAssembly/MCTargetDesc/WebAssemblyMCCodeEmitter.cpp

if (MO.isReg()) {
       /* nothing to encode */
     } else if (MO.isImm()) {
       if (i < Desc.getNumOperands()) {
         assert(Desc.TSFlags == 0 &&
                "WebAssembly non-variable_ops don't use TSFlags");
         const MCOperandInfo &Info = Desc.OpInfo[i];
         LLVM_DEBUG(dbgs() << "Encoding immediate: type="
                           << int(Info.OperandType) << "\n");
-        if (Info.OperandType == WebAssembly::OPERAND_I32IMM) {
+        switch (Info.OperandType) {
+          case WebAssembly::OPERAND_I32IMM:
             encodeSLEB128(int32_t(MO.getImm()), OS);
-        } else if (Info.OperandType == WebAssembly::OPERAND_OFFSET32) {
+            break;
+          case WebAssembly::OPERAND_OFFSET32:
             encodeULEB128(uint32_t(MO.getImm()), OS);
-        } else if (Info.OperandType == WebAssembly::OPERAND_I64IMM) {
+            break;
+          case WebAssembly::OPERAND_I64IMM:
             encodeSLEB128(int64_t(MO.getImm()), OS);
-        } else if (Info.OperandType == WebAssembly::OPERAND_GLOBAL) {
-          llvm_unreachable("wasm globals should only be accessed symbolicly");
-        } else if (Info.OperandType == WebAssembly::OPERAND_SIGNATURE) {
+            break;
+          case WebAssembly::OPERAND_SIGNATURE:
             OS << uint8_t(MO.getImm());
-        } else {
+            break;
+          case WebAssembly::OPERAND_VEC_I8IMM:
+            support::endian::write<uint8_t>(OS, MO.getImm(), support::little);
+            break;
+          case WebAssembly::OPERAND_VEC_I16IMM:
+            support::endian::write<uint16_t>(OS, MO.getImm(), support::little);
+            break;
+               case WebAssembly::OPERAND_VEC_I32IMM:
+            support::endian::write<uint32_t>(OS, MO.getImm(), support::little);
+            break;
+          case WebAssembly::OPERAND_VEC_I64IMM:
+            support::endian::write<uint64_t>(OS, MO.getImm(), support::little);
+            break;
+          case WebAssembly::OPERAND_GLOBAL:
+            llvm_unreachable("wasm globals should only be accessed symbolicly");
+          default:
             encodeULEB128(uint64_t(MO.getImm()), OS);
         }
       } else {
         assert(Desc.TSFlags == (WebAssemblyII::VariableOpIsImmediate |
                                 WebAssemblyII::VariableOpImmediateIsLabel));
         encodeULEB128(uint64_t(MO.getImm()), OS);
       }
     } else if (MO.isFPImm()) {
       assert(i < Desc.getNumOperands() &&

lib/Target/WebAssembly/MCTargetDesc/WebAssemblyMCTargetDesc.h

enum OperandType {
   /// 32-bit integer immediates.
   OPERAND_I32IMM,
   /// 64-bit integer immediates.
   OPERAND_I64IMM,
   /// 32-bit floating-point immediates.
   OPERAND_F32IMM,
   /// 64-bit floating-point immediates.
   OPERAND_F64IMM,
+  /// 8-bit vector lane immediate
+  OPERAND_VEC_I8IMM,
+  /// 16-bit vector lane immediate
+  OPERAND_VEC_I16IMM,
+  /// 32-bit vector lane immediate
+  OPERAND_VEC_I32IMM,
+  /// 64-bit vector lane immediate
+  OPERAND_VEC_I64IMM,
   /// 32-bit unsigned function indices.
   OPERAND_FUNCTION32,
   /// 32-bit unsigned memory offsets.
   OPERAND_OFFSET32,
   /// p2align immediate for load and store address alignment.
   OPERAND_P2ALIGN,
   /// signature immediate for block/loop.
   OPERAND_SIGNATURE,

lib/Target/WebAssembly/WebAssemblyInstrInfo.td

 def i64imm_op : Operand<i64>;
 
 let OperandType = "OPERAND_F32IMM" in
 def f32imm_op : Operand<f32>;
 
 let OperandType = "OPERAND_F64IMM" in
 def f64imm_op : Operand<f64>;
 
+let OperandType = "OPERAND_VEC_I8IMM" in
+def vec_i8imm_op : Operand<i32>;
+
+let OperandType = "OPERAND_VEC_I16IMM" in
+def vec_i16imm_op : Operand<i32>;
+
+let OperandType = "OPERAND_VEC_I32IMM" in
+def vec_i32imm_op : Operand<i32>;
+
+let OperandType = "OPERAND_VEC_I64IMM" in
+def vec_i64imm_op : Operand<i64>;
+
 let OperandType = "OPERAND_FUNCTION32" in
 def function32_op : Operand<i32>;
 
 let OperandType = "OPERAND_OFFSET32" in
 def offset32_op : Operand<i32>;
 
 let OperandType = "OPERAND_P2ALIGN" in {
 def P2Align : Operand<i32> {

lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

 // WebAssemblyInstrSIMD.td - WebAssembly SIMD codegen support -*- tablegen -*-//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 ///
 /// \file
 /// WebAssembly SIMD operand code-gen constructs.
 ///
 //===----------------------------------------------------------------------===//
 
-// immediate argument types
-def ImmByte : ImmLeaf<i32, [{ return 0 <= Imm && Imm < 256; }]>;
+// constrained immediate argument types
+foreach SIZE = [8, 16] in
+def ImmI#SIZE : ImmLeaf<i32, "return (Imm & ((1UL << "#SIZE#") - 1)) == Imm;">;
 foreach SIZE = [2, 4, 8, 16, 32] in
 def LaneIdx#SIZE : ImmLeaf<i32, "return 0 <= Imm && Imm < "#SIZE#";">;
 
+// const vectors
+multiclass ConstVec<ValueType vec_t, dag ops, dag pat, string args> {
+  defm CONST_V128_#vec_t : SIMD_I<(outs V128:$dst), ops, (outs), ops,
+                                  [(set V128:$dst, (vec_t pat))],
+                                  "v128.const\t$dst, "#args,
+                                  "v128.const\t"#args, 0>;
+}
+defm "" : ConstVec<v16i8,
+                   (ins vec_i8imm_op:$i0, vec_i8imm_op:$i1,
+                        vec_i8imm_op:$i2, vec_i8imm_op:$i3,
+                        vec_i8imm_op:$i4, vec_i8imm_op:$i5,
+                        vec_i8imm_op:$i6, vec_i8imm_op:$i7,
+                        vec_i8imm_op:$i8, vec_i8imm_op:$i9,
+                        vec_i8imm_op:$iA, vec_i8imm_op:$iB,
+                        vec_i8imm_op:$iC, vec_i8imm_op:$iD,
+                        vec_i8imm_op:$iE, vec_i8imm_op:$iF),
+                   (build_vector ImmI8:$i0, ImmI8:$i1, ImmI8:$i2, ImmI8:$i3,
+                                 ImmI8:$i4, ImmI8:$i5, ImmI8:$i6, ImmI8:$i7,
+                                 ImmI8:$i8, ImmI8:$i9, ImmI8:$iA, ImmI8:$iB,
+                                 ImmI8:$iC, ImmI8:$iD, ImmI8:$iE, ImmI8:$iF),
+                   !strconcat("$i0, $i1, $i2, $i3, $i4, $i5, $i6, $i7, ",
+                              "$i8, $i9, $iA, $iB, $iC, $iD, $iE, $iF")>;
+defm "" : ConstVec<v8i16,
+                   (ins vec_i16imm_op:$i0, vec_i16imm_op:$i1,
+                        vec_i16imm_op:$i2, vec_i16imm_op:$i3,
+                        vec_i16imm_op:$i4, vec_i16imm_op:$i5,
+                        vec_i16imm_op:$i6, vec_i16imm_op:$i7),
+                   (build_vector
+                     ImmI16:$i0, ImmI16:$i1, ImmI16:$i2, ImmI16:$i3,
+                     ImmI16:$i4, ImmI16:$i5, ImmI16:$i6, ImmI16:$i7),
+                   "$i0, $i1, $i2, $i3, $i4, $i5, $i6, $i7">;
+defm "" : ConstVec<v4i32,
+                   (ins vec_i32imm_op:$i0, vec_i32imm_op:$i1,
+                        vec_i32imm_op:$i2, vec_i32imm_op:$i3),
+                   (build_vector (i32 imm:$i0), (i32 imm:$i1),
+                                 (i32 imm:$i2), (i32 imm:$i3)),
+                   "$i0, $i1, $i2, $i3">;
+defm "" : ConstVec<v2i64,
+                  (ins vec_i64imm_op:$i0, vec_i64imm_op:$i1),
+                  (build_vector (i64 imm:$i0), (i64 imm:$i1)),
+                  "$i0, $i1">;
+defm "" : ConstVec<v4f32,
+                   (ins f32imm_op:$i0, f32imm_op:$i1,
+                        f32imm_op:$i2, f32imm_op:$i3),
+                   (build_vector (f32 fpimm:$i0), (f32 fpimm:$i1),
+                                 (f32 fpimm:$i2), (f32 fpimm:$i3)),
+                   "$i0, $i1, $i2, $i3">;
+defm "" : ConstVec<v2f64,
+                  (ins f64imm_op:$i0, f64imm_op:$i1),
+                  (build_vector (f64 fpimm:$i0), (f64 fpimm:$i1)),
+                  "$i0, $i1">;
+
 // lane extraction
 multiclass ExtractLane<ValueType vec_t, string vec, ImmLeaf imm_t,
                        WebAssemblyRegClass reg_t, bits<32> simdop,
                        string suffix = "", SDNode extract = vector_extract> {
   defm EXTRACT_LANE_#vec_t#suffix :
-      SIMD_I<(outs reg_t:$dst), (ins V128:$vec, I32:$idx),
-             (outs), (ins I32:$idx),
+      SIMD_I<(outs reg_t:$dst), (ins V128:$vec, i32imm_op:$idx),
+             (outs), (ins i32imm_op:$idx),
              [(set reg_t:$dst, (extract (vec_t V128:$vec), (i32 imm_t:$idx)))],
              vec#".extract_lane"#suffix#"\t$dst, $vec, $idx",
              vec#".extract_lane"#suffix#"\t$idx", simdop>;
 }
 multiclass ExtractPat<ValueType lane_t, int mask> {
   def _s : PatFrag<(ops node:$vec, node:$idx),
                    (i32 (sext_inreg
                      (i32 (vector_extract
def : Pat<(i32 (vector_extract (v16i8 V128:$vec), (i32 LaneIdx16:$idx))),
           (EXTRACT_LANE_v16i8_u V128:$vec, (i32 LaneIdx16:$idx))>;
 def : Pat<(i32 (vector_extract (v8i16 V128:$vec), (i32 LaneIdx8:$idx))),
           (EXTRACT_LANE_v8i16_u V128:$vec, (i32 LaneIdx8:$idx))>;
 
 // lane replacement
 multiclass ReplaceLane<ValueType vec_t, string vec, WebAssemblyRegClass reg_t,
                        ValueType lane_t, ImmLeaf imm_t, bits<32> simdop> {
   defm REPLACE_LANE_#vec_t :
-      SIMD_I<(outs V128:$dst), (ins V128:$vec, I32:$idx, reg_t:$x),
-             (outs), (ins I32:$idx),
+      SIMD_I<(outs V128:$dst), (ins V128:$vec, i32imm_op:$idx, reg_t:$x),
+             (outs), (ins i32imm_op:$idx),
              [(set V128:$dst, (vector_insert
                (vec_t V128:$vec), (lane_t reg_t:$x), (i32 imm_t:$idx)))],
              vec#".replace_lane\t$dst, $vec, $idx, $x",
              vec#".replace_lane\t$idx", simdop>;
 }
 let Defs = [ARGUMENTS] in {
 defm "" : ReplaceLane<v16i8, "i8x16", I32, i32, LaneIdx16, 17>;
 defm "" : ReplaceLane<v8i16, "i16x8", I32, i32, LaneIdx8, 18>;

test/CodeGen/WebAssembly/simd.ll

 ; RUN: llc < %s -disable-wasm-fallthrough-return-opt -disable-wasm-explicit-locals -wasm-enable-unimplemented-simd -mattr=+simd128,+sign-ext --show-mc-encoding | FileCheck %s --check-prefixes CHECK,SIMD128
 ; RUN: llc < %s -disable-wasm-fallthrough-return-opt -disable-wasm-explicit-locals -mattr=+simd128,+sign-ext --show-mc-encoding | FileCheck %s --check-prefixes CHECK,SIMD128-VM
 ; RUN: llc < %s -disable-wasm-fallthrough-return-opt -disable-wasm-explicit-locals -mattr=-simd128,+sign-ext --show-mc-encoding | FileCheck %s --check-prefixes CHECK,NO-SIMD128
 
 ; Test that basic SIMD128 vector manipulation operations assemble as expected.
 
 target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
 target triple = "wasm32-unknown-unknown"
 
 ; ==============================================================================
 ; 16 x i8
 ; ==============================================================================
+; CHECK-LABEL: const_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .result v128{{$}}
+; SIMD128: v128.const $push0=,
+; SIMD128-SAME: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
+; SIMD128-SAME: # encoding: [0xfd,0x00,
+; SIMD128-SAME: 0x00,0x01,0x02,0x03,0x04,0x05,0x06,0x07,
+; SIMD128-SAME: 0x08,0x09,0x0a,0x0b,0x0c,0x0d,0x0e,0x0f]{{$}}
+define <16 x i8> @const_v16i8() {
+  ret <16 x i8> <i8 00, i8 01, i8 02, i8 03, i8 04, i8 05, i8 06, i8 07,
+                 i8 08, i8 09, i8 10, i8 11, i8 12, i8 13, i8 14, i8 15>
+}
+
 ; CHECK-LABEL: splat_v16i8:
 ; NO-SIMD128-NOT: i8x16
 ; SIMD128: .param i32{{$}}
 ; SIMD128: .result v128{{$}}
 ; SIMD128: i8x16.splat $push0=, $0 # encoding: [0xfd,0x03]{{$}}
 ; SIMD128: return $pop0 # encoding: [0x0f]{{$}}
 define <16 x i8> @splat_v16i8(i8 %x) {
   %v = insertelement <16 x i8> undef, i8 %x, i32 0

 define <16 x i8> @replace_v16i8(<16 x i8> %v, i8 %x) {
   %res = insertelement <16 x i8> %v, i8 %x, i32 11
   ret <16 x i8> %res
 }
 
 ; ==============================================================================
 ; 8 x i16
 ; ==============================================================================
+; CHECK-LABEL: const_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .result v128{{$}}
+; SIMD128: v128.const $push0=, 256, 770, 1284, 1798, 2312, 2826, 3340, 3854
+; SIMD128-SAME: # encoding: [0xfd,0x00,
+; SIMD128-SAME: 0x00,0x01,0x02,0x03,0x04,0x05,0x06,0x07,
+; SIMD128-SAME: 0x08,0x09,0x0a,0x0b,0x0c,0x0d,0x0e,0x0f]{{$}}
+define <8 x i16> @const_v8i16() {
+  ret <8 x i16> <i16 256, i16 770, i16 1284, i16 1798,
+                 i16 2312, i16 2826, i16 3340, i16 3854>
+}
+
 ; CHECK-LABEL: splat_v8i16:
 ; NO-SIMD128-NOT: i16x8
 ; SIMD128: .param i32{{$}}
 ; SIMD128: .result v128{{$}}
 ; SIMD128: i16x8.splat $push0=, $0 # encoding: [0xfd,0x04]{{$}}
 ; SIMD128: return $pop0 # encoding: [0x0f]{{$}}
 define <8 x i16> @splat_v8i16(i16 %x) {
   %v = insertelement <8 x i16> undef, i16 %x, i32 0

 define <8 x i16> @replace_v8i16(<8 x i16> %v, i16 %x) {
   %res = insertelement <8 x i16> %v, i16 %x, i32 7
   ret <8 x i16> %res
 }
 
 ; ==============================================================================
 ; 4 x i32
 ; ==============================================================================
+; CHECK-LABEL: const_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .result v128{{$}}
+; SIMD128: v128.const $push0=, 50462976, 117835012, 185207048, 252579084
+; SIMD128-SAME: # encoding: [0xfd,0x00,
+; SIMD128-SAME: 0x00,0x01,0x02,0x03,0x04,0x05,0x06,0x07,
+; SIMD128-SAME: 0x08,0x09,0x0a,0x0b,0x0c,0x0d,0x0e,0x0f]{{$}}
+define <4 x i32> @const_v4i32() {
+  ret <4 x i32> <i32 50462976, i32 117835012, i32 185207048, i32 252579084>
+}
+
 ; CHECK-LABEL: splat_v4i32:
 ; NO-SIMD128-NOT: i32x4
 ; SIMD128: .param i32{{$}}
 ; SIMD128: .result v128{{$}}
 ; SIMD128: i32x4.splat $push0=, $0 # encoding: [0xfd,0x05]{{$}}
 ; SIMD128: return $pop0 # encoding: [0x0f]{{$}}
 define <4 x i32> @splat_v4i32(i32 %x) {
   %v = insertelement <4 x i32> undef, i32 %x, i32 0

 define <4 x i32> @replace_v4i32(<4 x i32> %v, i32 %x) {
   %res = insertelement <4 x i32> %v, i32 %x, i32 2
   ret <4 x i32> %res
 }
 
 ; ==============================================================================
 ; 2 x i64
 ; ==============================================================================
+; CHECK-LABEL: const_v2i64:
+; NO-SIMD128-NOT: i64x2
+; SIMD128-VM-NOT: i64x2
+; SIMD128: .result v128{{$}}
+; SIMD128: v128.const $push0=, 506097522914230528, 1084818905618843912
+; SIMD128-SAME: # encoding: [0xfd,0x00,
+; SIMD128-SAME: 0x00,0x01,0x02,0x03,0x04,0x05,0x06,0x07,
+; SIMD128-SAME: 0x08,0x09,0x0a,0x0b,0x0c,0x0d,0x0e,0x0f]{{$}}
+define <2 x i64> @const_v2i64() {
+  ret <2 x i64> <i64 506097522914230528, i64 1084818905618843912>
+}
+
 ; CHECK-LABEL: splat_v2i64:
 ; NO-SIMD128-NOT: i64x2
 ; SIMD128-VM-NOT: i64x2
 ; SIMD128: .param i64{{$}}
 ; SIMD128: .result v128{{$}}
 ; SIMD128: i64x2.splat $push0=, $0 # encoding: [0xfd,0x06]{{$}}
 ; SIMD128: return $pop0 # encoding: [0x0f]{{$}}
 define <2 x i64> @splat_v2i64(i64 %x) {

 define <2 x i64> @replace_v2i64(<2 x i64> %v, i64 %x) {
   %res = insertelement <2 x i64> %v, i64 %x, i32 0
   ret <2 x i64> %res
 }
 
 ; ==============================================================================
 ; 4 x f32
 ; ==============================================================================
+; CHECK-LABEL: const_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .result v128{{$}}
+; SIMD128: v128.const $push0=,
+; SIMD128-SAME: 0x1.0402p-121, 0x1.0c0a08p-113, 0x1.14121p-105, 0x1.1c1a18p-97
+; SIMD128-SAME: # encoding: [0xfd,0x00,
+; SIMD128-SAME: 0x00,0x01,0x02,0x03,0x04,0x05,0x06,0x07,
+; SIMD128-SAME: 0x08,0x09,0x0a,0x0b,0x0c,0x0d,0x0e,0x0f]{{$}}
+define <4 x float> @const_v4f32() {
+  ret <4 x float> <float 0x3860402000000000, float 0x38e0c0a080000000,
+                   float 0x3961412100000000, float 0x39e1c1a180000000>
+}
+
 ; CHECK-LABEL: splat_v4f32:
 ; NO-SIMD128-NOT: f32x4
 ; SIMD128: .param f32{{$}}
 ; SIMD128: .result v128{{$}}
 ; SIMD128: f32x4.splat $push0=, $0 # encoding: [0xfd,0x07]{{$}}
 ; SIMD128: return $pop0 # encoding: [0x0f]{{$}}
 define <4 x float> @splat_v4f32(float %x) {
   %v = insertelement <4 x float> undef, float %x, i32 0

 define <4 x float> @replace_v4f32(<4 x float> %v, float %x) {
   %res = insertelement <4 x float> %v, float %x, i32 2
   ret <4 x float> %res
 }
 
 ; ==============================================================================
 ; 2 x f64
 ; ==============================================================================
+; CHECK-LABEL: const_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128: .result v128{{$}}
+; SIMD128: v128.const $push0=, 0x1.60504030201p-911, 0x1.e0d0c0b0a0908p-783
+; SIMD128-SAME: # encoding: [0xfd,0x00,
+; SIMD128-SAME: 0x00,0x01,0x02,0x03,0x04,0x05,0x06,0x07,
+; SIMD128-SAME: 0x08,0x09,0x0a,0x0b,0x0c,0x0d,0x0e,0x0f]{{$}}
+define <2 x double> @const_v2f64() {
+  ret <2 x double> <double 0x0706050403020100, double 0x0F0E0D0C0B0A0908>
+}
+
 ; CHECK-LABEL: splat_v2f64:
 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-VM-NOT: f64x2
 ; SIMD128: .param f64{{$}}
 ; SIMD128: .result v128{{$}}
 ; SIMD128: f64x2.splat $push0=, $0 # encoding: [0xfd,0x08]{{$}}
 ; SIMD128: return $pop0 # encoding: [0x0f]{{$}}
 define <2 x double> @splat_v2f64(double %x) {