[wasm]implement simd lowering for simple F32x4 and I32x4 unops

BUG=v8:4124
R=bradnelson@chromium.org,bbudge@chromium.org,gdeepti@chromium.org,mtrofin@chromium.org,titzer@chromium.org

Review-Url: https://codereview.chromium.org/2726453003
Cr-Commit-Position: refs/heads/master@{#43509}

src/compiler/simd-scalar-lowering.cc

@@ -75,6 +75,8 @@ void SimdScalarLowering::LowerGraph() {
   V(Int32x4Splat)                 \
   V(Int32x4ExtractLane)           \
   V(Int32x4ReplaceLane)           \
+  V(Int32x4Neg)                   \
+  V(Simd128Not)                   \
   V(Int32x4Add)                   \
   V(Int32x4Sub)                   \
   V(Int32x4Mul)                   \
@@ -86,12 +88,16 @@ void SimdScalarLowering::LowerGraph() {
   V(Float32x4Splat)                 \
   V(Float32x4ExtractLane)           \
   V(Float32x4ReplaceLane)           \
+  V(Float32x4Abs)                   \
+  V(Float32x4Neg)                   \
   V(Float32x4Add)                   \
   V(Float32x4Sub)                   \
   V(Float32x4Mul)                   \
   V(Float32x4Div)                   \
   V(Float32x4Min)                   \
-  V(Float32x4Max)
+  V(Float32x4Max)                   \
+  V(Float32x4FromInt32x4)           \
+  V(Float32x4FromUint32x4)
 
 void SimdScalarLowering::SetLoweredType(Node* node, Node* output) {
   switch (node->opcode()) {
@@ -108,8 +114,14 @@ void SimdScalarLowering::SetLoweredType(Node* node, Node* output) {
         break;
       }
 #undef CASE_STMT
-    default:
-      replacements_[node->id()].type = replacements_[output->id()].type;
+    default: {
+      if (output->opcode() == IrOpcode::kFloat32x4FromInt32x4 ||
+          output->opcode() == IrOpcode::kFloat32x4FromUint32x4) {
+        replacements_[node->id()].type = SimdType::kInt32;
+      } else {
+        replacements_[node->id()].type = replacements_[output->id()].type;
+      }
+    }
   }
 }
 
@@ -229,11 +241,11 @@ void SimdScalarLowering::LowerStoreOp(MachineRepresentation rep, Node* node,
   }
 }
 
-void SimdScalarLowering::LowerBinaryOp(Node* node, SimdType rep_type,
+void SimdScalarLowering::LowerBinaryOp(Node* node, SimdType input_rep_type,
                                        const Operator* op) {
   DCHECK(node->InputCount() == 2);
-  Node** rep_left = GetReplacementsWithType(node->InputAt(0), rep_type);
-  Node** rep_right = GetReplacementsWithType(node->InputAt(1), rep_type);
+  Node** rep_left = GetReplacementsWithType(node->InputAt(0), input_rep_type);
+  Node** rep_right = GetReplacementsWithType(node->InputAt(1), input_rep_type);
   Node* rep_node[kMaxLanes];
   for (int i = 0; i < kMaxLanes; ++i) {
     rep_node[i] = graph()->NewNode(op, rep_left[i], rep_right[i]);
@@ -241,6 +253,17 @@ void SimdScalarLowering::LowerBinaryOp(Node* node, SimdType rep_type,
   ReplaceNode(node, rep_node);
 }
 
+void SimdScalarLowering::LowerUnaryOp(Node* node, SimdType input_rep_type,
+                                      const Operator* op) {
+  DCHECK(node->InputCount() == 1);
+  Node** rep = GetReplacementsWithType(node->InputAt(0), input_rep_type);
+  Node* rep_node[kMaxLanes];
+  for (int i = 0; i < kMaxLanes; ++i) {
+    rep_node[i] = graph()->NewNode(op, rep[i]);
+  }
+  ReplaceNode(node, rep_node);
+}
+
 void SimdScalarLowering::LowerNode(Node* node) {
   SimdType rep_type = ReplacementType(node);
   switch (node->opcode()) {
@@ -399,6 +422,28 @@ void SimdScalarLowering::LowerNode(Node* node) {
       I32X4_BINOP_CASE(kSimd128Or, Word32Or)
       I32X4_BINOP_CASE(kSimd128Xor, Word32Xor)
 #undef I32X4_BINOP_CASE
+    case IrOpcode::kInt32x4Neg: {
+      DCHECK(node->InputCount() == 1);
+      Node** rep = GetReplacementsWithType(node->InputAt(0), rep_type);
+      Node* rep_node[kMaxLanes];
+      Node* zero = graph()->NewNode(common()->Int32Constant(0));
+      for (int i = 0; i < kMaxLanes; ++i) {
+        rep_node[i] = graph()->NewNode(machine()->Int32Sub(), zero, rep[i]);
+      }
+      ReplaceNode(node, rep_node);
+      break;
+    }
+    case IrOpcode::kSimd128Not: {
+      DCHECK(node->InputCount() == 1);
+      Node** rep = GetReplacementsWithType(node->InputAt(0), rep_type);
+      Node* rep_node[kMaxLanes];
+      Node* mask = graph()->NewNode(common()->Int32Constant(0xffffffff));
+      for (int i = 0; i < kMaxLanes; ++i) {
+        rep_node[i] = graph()->NewNode(machine()->Word32Xor(), rep[i], mask);
+      }
+      ReplaceNode(node, rep_node);
+      break;
+    }
 #define F32X4_BINOP_CASE(name)                                 \
   case IrOpcode::kFloat32x4##name: {                           \
     LowerBinaryOp(node, rep_type, machine()->Float32##name()); \
@@ -411,6 +456,23 @@ void SimdScalarLowering::LowerNode(Node* node) {
       F32X4_BINOP_CASE(Min)
       F32X4_BINOP_CASE(Max)
 #undef F32X4_BINOP_CASE
+#define F32X4_UNOP_CASE(name)                                 \
+  case IrOpcode::kFloat32x4##name: {                          \
+    LowerUnaryOp(node, rep_type, machine()->Float32##name()); \
+    break;                                                    \
+  }
+      F32X4_UNOP_CASE(Abs)
+      F32X4_UNOP_CASE(Neg)
+      F32X4_UNOP_CASE(Sqrt)
+#undef F32x4_UNOP_CASE
+    case IrOpcode::kFloat32x4FromInt32x4: {
+      LowerUnaryOp(node, SimdType::kInt32, machine()->RoundInt32ToFloat32());
+      break;
+    }
+    case IrOpcode::kFloat32x4FromUint32x4: {
+      LowerUnaryOp(node, SimdType::kInt32, machine()->RoundUint32ToFloat32());
+      break;
+    }
     case IrOpcode::kInt32x4Splat:
     case IrOpcode::kFloat32x4Splat: {
       Node* rep_node[kMaxLanes];

src/compiler/simd-scalar-lowering.h

@@ -59,7 +59,8 @@ class SimdScalarLowering {
                    const Operator* load_op);
   void LowerStoreOp(MachineRepresentation rep, Node* node,
                     const Operator* store_op, SimdType rep_type);
-  void LowerBinaryOp(Node* node, SimdType rep_type, const Operator* op);
+  void LowerBinaryOp(Node* node, SimdType input_rep_type, const Operator* op);
+  void LowerUnaryOp(Node* node, SimdType input_rep_type, const Operator* op);
 
   struct NodeState {
     Node* node;

src/compiler/wasm-compiler.cc

@@ -3383,6 +3383,8 @@ Node* WasmGraphBuilder::SimdOp(wasm::WasmOpcode opcode,
       return graph()->NewNode(jsgraph()->machine()->Float32x4Abs(), inputs[0]);
     case wasm::kExprF32x4Neg:
       return graph()->NewNode(jsgraph()->machine()->Float32x4Neg(), inputs[0]);
+    case wasm::kExprF32x4Sqrt:
+      return graph()->NewNode(jsgraph()->machine()->Float32x4Sqrt(), inputs[0]);
     case wasm::kExprF32x4Add:
       return graph()->NewNode(jsgraph()->machine()->Float32x4Add(), inputs[0],
                               inputs[1]);

test/cctest/wasm/test-run-wasm-simd.cc

@@ -212,6 +212,11 @@ T Not(T a) {
   return ~a;
 }
 
+template <typename T>
+T Sqrt(T a) {
+  return std::sqrt(a);
+}
+
 }  // namespace
 
 #if !V8_TARGET_ARCH_ARM && !V8_TARGET_ARCH_X64
@@ -379,9 +384,7 @@ WASM_EXEC_COMPILED_TEST(F32x4ReplaceLane) {
 
   CHECK_EQ(1, r.Call(3.14159f, -1.5f));
 }
-#endif  // V8_TARGET_ARCH_ARM || SIMD_LOWERING_TARGET
 
-#if V8_TARGET_ARCH_ARM
 // Tests both signed and unsigned conversion.
 WASM_EXEC_COMPILED_TEST(F32x4FromInt32x4) {
   FLAG_wasm_simd_prototype = true;
@@ -417,13 +420,18 @@ void RunF32x4UnOpTest(WasmOpcode simd_op, FloatUnOp expected_op) {
 
   FOR_FLOAT32_INPUTS(i) {
     if (std::isnan(*i)) continue;
+    if (std::isnan(expected_op(*i))) continue;
     CHECK_EQ(1, r.Call(*i, expected_op(*i)));
   }
 }
 
 WASM_EXEC_COMPILED_TEST(F32x4Abs) { RunF32x4UnOpTest(kExprF32x4Abs, std::abs); }
 WASM_EXEC_COMPILED_TEST(F32x4Neg) { RunF32x4UnOpTest(kExprF32x4Neg, Negate); }
-#endif  // V8_TARGET_ARCH_ARM
+#endif  // V8_TARGET_ARCH_ARM || SIMD_LOWERING_TARGET
+
+#if SIMD_LOWERING_TARGET
+WASM_EXEC_COMPILED_TEST(F32x4Sqrt) { RunF32x4UnOpTest(kExprF32x4Sqrt, Sqrt); }
+#endif  // SIMD_LOWERING_TARGET
 
 #if V8_TARGET_ARCH_ARM || SIMD_LOWERING_TARGET
 void RunF32x4BinOpTest(WasmOpcode simd_op, FloatBinOp expected_op,
@@ -808,7 +816,9 @@ WASM_EXEC_COMPILED_TEST(I32x4FromFloat32x4) {
     CHECK_EQ(1, r.Call(*i, signed_value, unsigned_value));
   }
 }
+#endif  // V8_TARGET_ARCH_ARM
 
+#if V8_TARGET_ARCH_ARM || SIMD_LOWERING_TARGET
 void RunI32x4UnOpTest(WasmOpcode simd_op, Int32UnOp expected_op) {
   FLAG_wasm_simd_prototype = true;
   WasmRunner<int32_t, int32_t, int32_t> r(kExecuteCompiled);
@@ -825,7 +835,7 @@ void RunI32x4UnOpTest(WasmOpcode simd_op, Int32UnOp expected_op) {
 WASM_EXEC_COMPILED_TEST(I32x4Neg) { RunI32x4UnOpTest(kExprI32x4Neg, Negate); }
 
 WASM_EXEC_COMPILED_TEST(S128Not) { RunI32x4UnOpTest(kExprS128Not, Not); }
-#endif  // V8_TARGET_ARCH_ARM
+#endif  // V8_TARGET_ARCH_ARM || SIMD_LOWERING_TARGET
 
 void RunI32x4BinOpTest(WasmOpcode simd_op, Int32BinOp expected_op) {
   FLAG_wasm_simd_prototype = true;