[WebAssembly] SIMD comparisons

[tlively]

Summary

Match the ordering semantics of non-vector comparisons. For
floating point comparisons that do not correspond to instructions, the
tests check that some vector comparison instruction was emitted but do
not care about the full implementation.

[aheejin]

lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

149 ↗

(On Diff #164331)

When do you use simdop and when do you use baseInst? Do you use it interchangeably? If so we should use the same variable name I guess. (In previous patches too.)

169 ↗

(On Diff #164331)

I can see why you added this class, but that some instructions use this and the others bypasses this and directly use SIMDConditionInt and SIMDConditionFP directly looks a bit confusing to me. Even if we delete this and manually add _u instructions below, it's gonna be only (?) +8 lines, so I guess it'd be easier to read that way..?

test/CodeGen/WebAssembly/comparisons_simd.ll

1 ↗

(On Diff #164331)

How about renaming this to simd-comparisons.ll to match other file name conventions you've added so far?

[tlively]

• Remove unnecessary layer of multiclasses

[tlively]

lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

149 ↗

(On Diff #164331)

I try to use baseInst when the multiclass will add constants to it to create new opcodes and simdop when it will be used as-is. I will go through and make sure this usage is consistent. I would also be fine with changing them all to simdop if you think that would make more sense.

test/CodeGen/WebAssembly/comparisons_simd.ll

1 ↗

(On Diff #164331)

I named it this way to be consistent with comparisons_f32.ll, comparisons_i32.ll and friends. Those files should probably be renamed for consistency in a separate CL as well. What do you think of comparisons-i32.ll, comparisons-simd.ll, etc?

[aheejin]

lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

149 ↗

(On Diff #164331)

Oh, I see. The usage makes sense I think.

test/CodeGen/WebAssembly/comparisons_simd.ll

1 ↗

(On Diff #164331)

Hmm, yeah, they look like the only ones with underscores in file names. I think renaming them also is a good idea. But other SIMD-related test files all start with simd-, like simd-arith.ll or simd-conversions.ll, so I guess putting simd first makes more sense. Wait, there's an exception, there are exceptions that start with offset-. :facepalm: But anyway.

[tlively]

• Rename test file for consistency

[aheejin]

lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

309 ↗

(On Diff #164499)

Could you elaborate what this comment means and why we need this only for floats?

[tlively]

• Add comment explaining don't care float comparison lowering

[aheejin]

lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

309 ↗

(On Diff #164499)

I don't really understand what "Lower don't float comparisons" means... Did you mean "Don't lower float comparisons?"

[tlively]

• Fix editing mistake

[aheejin]

lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

348 ↗

(On Diff #164780)

1. Shouldn't "NE" be "ONE" here, according to the comments?
2. What kind of source pattern does this rule cover? I test your patch with commenting this part out, and all you tests succeed anyway.

[tlively]

• Remove unnecessary float compare patterns

[tlively]

• Unroll newly shortened and complicated loop

[tlively]

lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

348 ↗

(On Diff #164780)

1. No, these aren't the ordered or unordered comparison modes, these are the "don't care" comparison modes, which are allowed to be either ordered or unordered. These patterns say that these "don't care" comparisons should be lowered to whichever comparison type wasm natively supports.

2. I tried commenting this section out and the tests failed, but I was able to cut the set of comparisons here down to just eq and ne. I'm not sure why it would have passed for you. The "don't care" float comparisons are not expressible in LLVM IR, but are emitted as isel dag nodes in the target-independent automatic expansion of fcmp ueq and fcmp one, respectively.

[aheejin]

lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

342 ↗

(On Diff #164789)

LGTM with one nit: wrap comment around 80 col

llvm/trunk/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

WebAssemblyTargetLowering::WebAssemblyTargetLowering(
 
   // Take the default expansion for va_arg, va_copy, and va_end. There is no
   // default action for va_start, so we do that custom.
   setOperationAction(ISD::VASTART, MVT::Other, Custom);
   setOperationAction(ISD::VAARG, MVT::Other, Expand);
   setOperationAction(ISD::VACOPY, MVT::Other, Expand);
   setOperationAction(ISD::VAEND, MVT::Other, Expand);
 
-  for (auto T : {MVT::f32, MVT::f64}) {
+  for (auto T : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
     // Don't expand the floating-point types to constant pools.
     setOperationAction(ISD::ConstantFP, T, Legal);
     // Expand floating-point comparisons.
     for (auto CC : {ISD::SETO, ISD::SETUO, ISD::SETUEQ, ISD::SETONE,
                     ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE})
       setCondCodeAction(CC, T, Expand);
     // Expand floating-point library function operators.
     for (auto Op :

llvm/trunk/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

defm NOT_#vec_t : SIMD_I<(outs V128:$dst), (ins V128:$vec),
                              (vec_t V128:$dst),
                              (vec_t (xor
                                (vec_t V128:$vec),
                                (vec_t (splat_pat (lane_t -1)))
                              ))
                            )],
                            "v128.not\t$dst, $vec", "v128.not", 62>;
 }
+multiclass SIMDCondition<ValueType vec_t, ValueType out_t, string vec,
+                             string name, CondCode cond, bits<32> simdop> {
+  defm _#vec_t :
+    SIMD_I<(outs V128:$dst), (ins V128:$lhs, V128: $rhs), (outs), (ins),
+           [(set (out_t V128:$dst),
+             (setcc (vec_t V128:$lhs), (vec_t V128:$rhs), cond))],
+           vec#"."#name#"\t$dst, $lhs, $rhs", vec#"."#name, simdop>;
+}
+multiclass SIMDConditionInt<string name, CondCode cond, bits<32> baseInst,
+                            int step = 1> {
+  defm "" : SIMDCondition<v16i8, v16i8, "i8x16", name, cond, baseInst>;
+  defm "" : SIMDCondition<v8i16, v8i16, "i16x8", name, cond,
+                              !add(baseInst, step)>;
+  defm "" : SIMDCondition<v4i32, v4i32, "i32x4", name, cond,
+                              !add(!add(baseInst, step), step)>;
+}
+multiclass SIMDConditionFP<string name, CondCode cond, bits<32> baseInst> {
+  defm "" : SIMDCondition<v4f32, v4i32, "f32x4", name, cond, baseInst>;
+  defm "" : SIMDCondition<v2f64, v2i64, "f64x2", name, cond,
+                              !add(baseInst, 1)>;
+}
 
 let Defs = [ARGUMENTS] in {
 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,

 defm XOR : SIMDBitwise<xor, "xor", 61>;
 } // isCommutable = 1
 
 defm "" : SIMDNot<v16i8, splat16, i32>;
 defm "" : SIMDNot<v8i16, splat8, i32>;
 defm "" : SIMDNot<v4i32, splat4, i32>;
 defm "" : SIMDNot<v2i64, splat2, i64>;
 
+let isCommutable = 1 in {
+defm EQ : SIMDConditionInt<"eq", SETEQ, 72>;
+defm EQ : SIMDConditionFP<"eq", SETOEQ, 75>;
+defm NE : SIMDConditionInt<"ne", SETNE, 77>;
+defm NE : SIMDConditionFP<"ne", SETUNE, 80>;
+} // isCommutable = 1
+
+defm LT_S : SIMDConditionInt<"lt_s", SETLT, 82, 2>;
+defm LT_U : SIMDConditionInt<"lt_u", SETULT, 83, 2>;
+defm LT : SIMDConditionFP<"lt", SETOLT, 88>;
+defm LE_S : SIMDConditionInt<"le_s", SETLE, 90, 2>;
+defm LE_U : SIMDConditionInt<"le_u", SETULE, 91, 2>;
+defm LE : SIMDConditionFP<"le", SETOLE, 96>;
+defm GT_S : SIMDConditionInt<"gt_s", SETGT, 98, 2>;
+defm GT_U : SIMDConditionInt<"gt_u", SETUGT, 99, 2>;
+defm GT : SIMDConditionFP<"gt", SETOGT, 104>;
+defm GE_S : SIMDConditionInt<"ge_s", SETGE, 106, 2>;
+defm GE_U : SIMDConditionInt<"ge_u", SETUGE, 107, 2>;
+defm GE : SIMDConditionFP<"ge", SETOGE, 112>;
+
 } // Defs = [ARGUMENTS]
 
 // Def load and store patterns from WebAssemblyInstrMemory.td for vector types
 foreach vec_t = [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64] in {
 
 def : LoadPatNoOffset<vec_t, load, !cast<NI>("LOAD_"#vec_t)>;
 def : LoadPatImmOff<vec_t, load, regPlusImm, !cast<NI>("LOAD_"#vec_t)>;
 def : LoadPatImmOff<vec_t, load, or_is_add, !cast<NI>("LOAD_"#vec_t)>;

 def : StorePatGlobalAddr<vec_t, store, !cast<NI>("STORE_"#vec_t)>;
 def : StorePatExternalSym<vec_t, store, !cast<NI>("STORE_"#vec_t)>;
 def : StorePatOffsetOnly<vec_t, store, !cast<NI>("STORE_"#vec_t)>;
 def : StorePatGlobalAddrOffOnly<vec_t, store, !cast<NI>("STORE_"#vec_t)>;
 def : StorePatExternSymOffOnly<vec_t, store, !cast<NI>("STORE_"#vec_t)>;
 
 }
 
+// Lower float comparisons that don't care about NaN to standard
+// WebAssembly float comparisons. These instructions are generated in
+// the target-independent expansion of unordered comparisons and
+// ordered ne.
+def : Pat<(v4i32 (seteq (v4f32 V128:$lhs), (v4f32 V128:$rhs))),
+          (v4i32 (EQ_v4f32 (v4f32 V128:$lhs), (v4f32 V128:$rhs)))>;
+def : Pat<(v4i32 (setne (v4f32 V128:$lhs), (v4f32 V128:$rhs))),
+          (v4i32 (NE_v4f32 (v4f32 V128:$lhs), (v4f32 V128:$rhs)))>;
+def : Pat<(v2i64 (seteq (v2f64 V128:$lhs), (v2f64 V128:$rhs))),
+          (v2i64 (EQ_v2f64 (v2f64 V128:$lhs), (v2f64 V128:$rhs)))>;
+def : Pat<(v2i64 (setne (v2f64 V128:$lhs), (v2f64 V128:$rhs))),
+          (v2i64 (NE_v2f64 (v2f64 V128:$lhs), (v2f64 V128:$rhs)))>;
+
 // follow convention of making implicit expansions unsigned
 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))>;
 
 // Bitcasts are nops
 // Matching bitcast t1 to t1 causes strange errors, so avoid repeating types

llvm/trunk/test/CodeGen/WebAssembly/simd-comparisons.ll

+; RUN: llc < %s -disable-wasm-fallthrough-return-opt -wasm-keep-registers -wasm-disable-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 -wasm-keep-registers -wasm-disable-explicit-locals -mattr=+simd128,+sign-ext --show-mc-encoding | FileCheck %s --check-prefixes CHECK,SIMD128-VM
+; RUN: llc < %s -disable-wasm-fallthrough-return-opt -wasm-keep-registers -wasm-disable-explicit-locals -mattr=-simd128,+sign-ext --show-mc-encoding | FileCheck %s --check-prefixes CHECK,NO-SIMD128
+
+; Test SIMD comparison operators
+
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32-unknown-unknown"
+
+; CHECK-LABEL: compare_eq_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.eq $push0=, $0, $1 # encoding: [0xfd,0x48]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_eq_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp eq <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_ne_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.ne $push0=, $0, $1 # encoding: [0xfd,0x4d]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_ne_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp ne <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_slt_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.lt_s $push0=, $0, $1 # encoding: [0xfd,0x52]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_slt_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp slt <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_ult_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.lt_u $push0=, $0, $1 # encoding: [0xfd,0x53]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_ult_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp ult <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_sle_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.le_s $push0=, $0, $1 # encoding: [0xfd,0x5a]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_sle_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp sle <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_ule_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.le_u $push0=, $0, $1 # encoding: [0xfd,0x5b]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_ule_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp ule <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_sgt_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.gt_s $push0=, $0, $1 # encoding: [0xfd,0x62]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_sgt_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp sgt <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_ugt_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.gt_u $push0=, $0, $1 # encoding: [0xfd,0x63]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_ugt_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp ugt <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_sge_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.ge_s $push0=, $0, $1 # encoding: [0xfd,0x6a]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_sge_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp sge <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_uge_v16i8:
+; NO-SIMD128-NOT: i8x16
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i8x16.ge_u $push0=, $0, $1 # encoding: [0xfd,0x6b]{{$}}
+; SIMD128: return $pop0 #
+define <16 x i1> @compare_uge_v16i8 (<16 x i8> %x, <16 x i8> %y) {
+  %res = icmp uge <16 x i8> %x, %y
+  ret <16 x i1> %res
+}
+
+; CHECK-LABEL: compare_eq_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.eq $push0=, $0, $1 # encoding: [0xfd,0x49]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_eq_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp eq <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_ne_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.ne $push0=, $0, $1 # encoding: [0xfd,0x4e]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_ne_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp ne <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_slt_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.lt_s $push0=, $0, $1 # encoding: [0xfd,0x54]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_slt_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp slt <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_ult_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.lt_u $push0=, $0, $1 # encoding: [0xfd,0x55]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_ult_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp ult <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_sle_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.le_s $push0=, $0, $1 # encoding: [0xfd,0x5c]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_sle_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp sle <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_ule_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.le_u $push0=, $0, $1 # encoding: [0xfd,0x5d]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_ule_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp ule <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_sgt_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.gt_s $push0=, $0, $1 # encoding: [0xfd,0x64]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_sgt_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp sgt <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_ugt_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.gt_u $push0=, $0, $1 # encoding: [0xfd,0x65]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_ugt_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp ugt <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_sge_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.ge_s $push0=, $0, $1 # encoding: [0xfd,0x6c]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_sge_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp sge <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_uge_v8i16:
+; NO-SIMD128-NOT: i16x8
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i16x8.ge_u $push0=, $0, $1 # encoding: [0xfd,0x6d]{{$}}
+; SIMD128: return $pop0 #
+define <8 x i1> @compare_uge_v8i16 (<8 x i16> %x, <8 x i16> %y) {
+  %res = icmp uge <8 x i16> %x, %y
+  ret <8 x i1> %res
+}
+
+; CHECK-LABEL: compare_eq_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.eq $push0=, $0, $1 # encoding: [0xfd,0x4a]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_eq_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp eq <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ne_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.ne $push0=, $0, $1 # encoding: [0xfd,0x4f]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_ne_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp ne <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_slt_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.lt_s $push0=, $0, $1 # encoding: [0xfd,0x56]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_slt_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp slt <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ult_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.lt_u $push0=, $0, $1 # encoding: [0xfd,0x57]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_ult_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp ult <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_sle_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.le_s $push0=, $0, $1 # encoding: [0xfd,0x5e]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_sle_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp sle <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ule_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.le_u $push0=, $0, $1 # encoding: [0xfd,0x5f]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_ule_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp ule <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_sgt_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.gt_s $push0=, $0, $1 # encoding: [0xfd,0x66]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_sgt_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp sgt <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ugt_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.gt_u $push0=, $0, $1 # encoding: [0xfd,0x67]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_ugt_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp ugt <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_sge_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.ge_s $push0=, $0, $1 # encoding: [0xfd,0x6e]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_sge_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp sge <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_uge_v4i32:
+; NO-SIMD128-NOT: i32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: i32x4.ge_u $push0=, $0, $1 # encoding: [0xfd,0x6f]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_uge_v4i32 (<4 x i32> %x, <4 x i32> %y) {
+  %res = icmp uge <4 x i32> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_oeq_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.eq $push0=, $0, $1 # encoding: [0xfd,0x4b]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_oeq_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp oeq <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ogt_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.gt $push0=, $0, $1 # encoding: [0xfd,0x68]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_ogt_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp ogt <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_oge_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.ge $push0=, $0, $1 # encoding: [0xfd,0x70]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_oge_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp oge <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_olt_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.lt $push0=, $0, $1 # encoding: [0xfd,0x58]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_olt_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp olt <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ole_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.le $push0=, $0, $1 # encoding: [0xfd,0x60]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_ole_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp ole <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_one_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.ne
+define <4 x i1> @compare_one_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp one <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ord_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.eq
+define <4 x i1> @compare_ord_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp ord <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ueq_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.eq
+define <4 x i1> @compare_ueq_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp ueq <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ugt_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.le
+define <4 x i1> @compare_ugt_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp ugt <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_uge_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.lt
+define <4 x i1> @compare_uge_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp uge <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ult_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.ge
+define <4 x i1> @compare_ult_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp ult <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_ule_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.gt
+define <4 x i1> @compare_ule_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp ule <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_une_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.ne $push0=, $0, $1 # encoding: [0xfd,0x50]{{$}}
+; SIMD128: return $pop0 #
+define <4 x i1> @compare_une_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp une <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_uno_v4f32:
+; NO-SIMD128-NOT: f32x4
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f32x4.ne
+define <4 x i1> @compare_uno_v4f32 (<4 x float> %x, <4 x float> %y) {
+  %res = fcmp uno <4 x float> %x, %y
+  ret <4 x i1> %res
+}
+
+; CHECK-LABEL: compare_oeq_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.eq $push0=, $0, $1 # encoding: [0xfd,0x4c]{{$}}
+; SIMD128: return $pop0 #
+define <2 x i1> @compare_oeq_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp oeq <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_ogt_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.gt $push0=, $0, $1 # encoding: [0xfd,0x69]{{$}}
+; SIMD128: return $pop0 #
+define <2 x i1> @compare_ogt_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp ogt <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_oge_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.ge $push0=, $0, $1 # encoding: [0xfd,0x71]{{$}}
+; SIMD128: return $pop0 #
+define <2 x i1> @compare_oge_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp oge <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_olt_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.lt $push0=, $0, $1 # encoding: [0xfd,0x59]{{$}}
+; SIMD128: return $pop0 #
+define <2 x i1> @compare_olt_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp olt <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_ole_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.le $push0=, $0, $1 # encoding: [0xfd,0x61]{{$}}
+; SIMD128: return $pop0 #
+define <2 x i1> @compare_ole_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp ole <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_one_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.ne
+define <2 x i1> @compare_one_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp one <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_ord_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.eq
+define <2 x i1> @compare_ord_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp ord <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_ueq_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.eq
+define <2 x i1> @compare_ueq_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp ueq <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_ugt_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.le
+define <2 x i1> @compare_ugt_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp ugt <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_uge_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.lt
+define <2 x i1> @compare_uge_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp uge <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_ult_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.ge
+define <2 x i1> @compare_ult_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp ult <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_ule_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.gt
+define <2 x i1> @compare_ule_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp ule <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_une_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.ne $push0=, $0, $1 # encoding: [0xfd,0x51]{{$}}
+; SIMD128: return $pop0 #
+define <2 x i1> @compare_une_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp une <2 x double> %x, %y
+  ret <2 x i1> %res
+}
+
+; CHECK-LABEL: compare_uno_v2f64:
+; NO-SIMD128-NOT: f64x2
+; SIMD128-VM-NOT: f64x2
+; SIMD128: .param v128, v128{{$}}
+; SIMD128: .result v128{{$}}
+; SIMD128: f64x2.ne
+define <2 x i1> @compare_uno_v2f64 (<2 x double> %x, <2 x double> %y) {
+  %res = fcmp uno <2 x double> %x, %y
+  ret <2 x i1> %res
+}