[Turbofan] Add native ARM support for basic SIMD 32x4 operations.

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

Index: test/cctest/wasm/test-run-wasm-simd.cc
diff --git a/test/cctest/wasm/test-run-wasm-simd.cc b/test/cctest/wasm/test-run-wasm-simd.cc
index 2905473b17d048031ff224856eb73c82bb94ecd2..83ff149c9fb85d7be52c7deeb7c3e8a237a08895 100644
--- a/test/cctest/wasm/test-run-wasm-simd.cc
+++ b/test/cctest/wasm/test-run-wasm-simd.cc
@@ -13,6 +13,30 @@ using namespace v8::internal;
 using namespace v8::internal::compiler;
 using namespace v8::internal::wasm;
 
+namespace {
+
+template <typename T>
+T Add(T a, T b) {
+  return a + b;
+}
+
+template <typename T>
+T Sub(T a, T b) {
+  return a - b;
+}
+
+template <typename T>
+int32_t Equal(T a, T b) {
+  return a == b ? 0xFFFFFFFF : 0;
+}
+
+template <typename T>
+int32_t NotEqual(T a, T b) {
+  return a != b ? 0xFFFFFFFF : 0;
+}
+
+}  // namespace
+
 // TODO(gdeepti): These are tests using sample values to verify functional
 // correctness of opcodes, add more tests for a range of values and macroize
 // tests.
@@ -32,6 +56,157 @@ using namespace v8::internal::wasm;
 #define WASM_SIMD_CHECK_SPLAT4(TYPE, value, LANE_TYPE, lv) \
   WASM_SIMD_CHECK4(TYPE, value, LANE_TYPE, lv, lv, lv, lv)
 
+#define WASM_SIMD_CHECK_F32_LANE(TYPE, value, lane_value, lane_index)       \
+  WASM_IF(                                                                  \
+      WASM_I32_NE(WASM_I32_REINTERPRET_F32(WASM_GET_LOCAL(lane_value)),     \
+                  WASM_I32_REINTERPRET_F32(WASM_SIMD_##TYPE##_EXTRACT_LANE( \
+                      lane_index, WASM_GET_LOCAL(value)))),                 \
+      WASM_RETURN1(WASM_ZERO))
+
+#define WASM_SIMD_CHECK4_F32(TYPE, value, lv0, lv1, lv2, lv3) \
+  WASM_SIMD_CHECK_F32_LANE(TYPE, value, lv0, 0)               \
+  , WASM_SIMD_CHECK_F32_LANE(TYPE, value, lv1, 1),            \
+      WASM_SIMD_CHECK_F32_LANE(TYPE, value, lv2, 2),          \
+      WASM_SIMD_CHECK_F32_LANE(TYPE, value, lv3, 3)
+
+#define WASM_SIMD_CHECK_SPLAT4_F32(TYPE, value, lv) \
+  WASM_SIMD_CHECK4_F32(TYPE, value, lv, lv, lv, lv)
+
+#if V8_TARGET_ARCH_ARM
+WASM_EXEC_TEST(F32x4Splat) {
+  FLAG_wasm_simd_prototype = true;
+
+  WasmRunner<int32_t, float> r(kExecuteCompiled);
+  byte lane_val = 0;
+  byte simd = r.AllocateLocal(kAstS128);
+  BUILD(r, WASM_BLOCK(WASM_SET_LOCAL(simd, WASM_SIMD_F32x4_SPLAT(
+                                               WASM_GET_LOCAL(lane_val))),
+                      WASM_SIMD_CHECK_SPLAT4_F32(F32x4, simd, lane_val),
+                      WASM_RETURN1(WASM_ONE)));
+
+  FOR_FLOAT32_INPUTS(i) { CHECK_EQ(1, r.Call(*i)); }
+}
+
+WASM_EXEC_TEST(F32x4ReplaceLane) {
+  FLAG_wasm_simd_prototype = true;
+  WasmRunner<int32_t, float, float> r(kExecuteCompiled);
+  byte old_val = 0;
+  byte new_val = 1;
+  byte simd = r.AllocateLocal(kAstS128);
+  BUILD(r, WASM_BLOCK(
+               WASM_SET_LOCAL(simd,
+                              WASM_SIMD_F32x4_SPLAT(WASM_GET_LOCAL(old_val))),
+               WASM_SET_LOCAL(
+                   simd, WASM_SIMD_F32x4_REPLACE_LANE(0, WASM_GET_LOCAL(simd),
+                                                      WASM_GET_LOCAL(new_val))),
+               WASM_SIMD_CHECK4(F32x4, simd, F32, new_val, old_val, old_val,
+                                old_val),
+               WASM_SET_LOCAL(
+                   simd, WASM_SIMD_F32x4_REPLACE_LANE(1, WASM_GET_LOCAL(simd),
+                                                      WASM_GET_LOCAL(new_val))),
+               WASM_SIMD_CHECK4(F32x4, simd, F32, new_val, new_val, old_val,
+                                old_val),
+               WASM_SET_LOCAL(
+                   simd, WASM_SIMD_F32x4_REPLACE_LANE(2, WASM_GET_LOCAL(simd),
+                                                      WASM_GET_LOCAL(new_val))),
+               WASM_SIMD_CHECK4(F32x4, simd, F32, new_val, new_val, new_val,
+                                old_val),
+               WASM_SET_LOCAL(
+                   simd, WASM_SIMD_F32x4_REPLACE_LANE(3, WASM_GET_LOCAL(simd),
+                                                      WASM_GET_LOCAL(new_val))),
+               WASM_SIMD_CHECK_SPLAT4(F32x4, simd, F32, new_val),
+               WASM_RETURN1(WASM_ONE)));
+
+  CHECK_EQ(1, r.Call(1, 2));

[gdeepti, 2016/12/16 23:35:24]

Is there a reason we do not use floating point values here? 

I guess this is a general question for WASM_SIMD_CHECK_F32_LANE using
WASM_I32_NE and WASM_I32_REINTERPRET_F32 instead of WASM_F32_NE. 

[bbudge, 2016/12/17 01:51:05]

Changed to be more obviously floating point.

Comparing floating point values is a pain, so I'm avoiding Nan values in these
tests. I'll add a TODO to figure out how to generate WASM code to compare FP
values.

+}
+
+// Tests both signed and unsigned conversion.
+WASM_EXEC_TEST(F32x4FromInt32x4) {
+  FLAG_wasm_simd_prototype = true;
+  WasmRunner<int32_t, int32_t, float, float> r(kExecuteCompiled);
+  byte a = 0;
+  byte expected_signed = 1;
+  byte expected_unsigned = 2;
+  byte simd0 = r.AllocateLocal(kAstS128);
+  byte simd1 = r.AllocateLocal(kAstS128);
+  byte simd2 = r.AllocateLocal(kAstS128);
+  BUILD(r, WASM_BLOCK(
+               WASM_SET_LOCAL(simd0, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(a))),
+               WASM_SET_LOCAL(
+                   simd1, WASM_SIMD_F32x4_FROM_INT32x4(WASM_GET_LOCAL(simd0))),
+               WASM_SIMD_CHECK_SPLAT4_F32(F32x4, simd1, expected_signed),
+               WASM_SET_LOCAL(
+                   simd2, WASM_SIMD_F32x4_FROM_UINT32x4(WASM_GET_LOCAL(simd0))),
+               WASM_SIMD_CHECK_SPLAT4_F32(F32x4, simd2, expected_unsigned),
+               WASM_RETURN1(WASM_ONE)));
+
+  FOR_INT32_INPUTS(i) {
+    CHECK_EQ(1, r.Call(*i, static_cast<float>(*i),
+                       static_cast<float>(static_cast<uint32_t>(*i))));
+  }
+}
+
+WASM_EXEC_TEST(S32x4Select) {
+  FLAG_wasm_simd_prototype = true;
+  WasmRunner<int32_t, int32_t, int32_t> r(kExecuteCompiled);
+  byte val1 = 0;
+  byte val2 = 1;
+  byte mask = r.AllocateLocal(kAstS128);
+  byte src1 = r.AllocateLocal(kAstS128);
+  byte src2 = r.AllocateLocal(kAstS128);
+  BUILD(r,
+        WASM_BLOCK(
+            WASM_SET_LOCAL(mask, WASM_SIMD_I32x4_SPLAT(WASM_ZERO)),
+            WASM_SET_LOCAL(src1, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(val1))),
+            WASM_SET_LOCAL(src2, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(val2))),
+            WASM_SET_LOCAL(mask, WASM_SIMD_I32x4_REPLACE_LANE(
+                                     1, WASM_GET_LOCAL(mask), WASM_I32V(-1))),
+            WASM_SET_LOCAL(mask, WASM_SIMD_I32x4_REPLACE_LANE(
+                                     2, WASM_GET_LOCAL(mask), WASM_I32V(-1))),
+            WASM_SET_LOCAL(mask, WASM_SIMD_32x4_SELECT(WASM_GET_LOCAL(mask),
+                                                       WASM_GET_LOCAL(src1),
+                                                       WASM_GET_LOCAL(src2))),
+            WASM_SIMD_CHECK_LANE(I32x4, mask, I32, val2, 0),
+            WASM_SIMD_CHECK_LANE(I32x4, mask, I32, val1, 1),
+            WASM_SIMD_CHECK_LANE(I32x4, mask, I32, val1, 2),
+            WASM_SIMD_CHECK_LANE(I32x4, mask, I32, val2, 3),
+            WASM_RETURN1(WASM_ONE)));
+
+  CHECK_EQ(1, r.Call(0x1234, 0x5678));
+}
+
+#define WASM_SIMD_F32x4_BINOP_TEST(Name, OP, op)                               \
+  WASM_EXEC_TEST(F32x4##Name) {                                                \
+    FLAG_wasm_simd_prototype = true;                                           \
+    WasmRunner<int32_t, float, float, float> r(kExecuteCompiled);              \
+    byte a = 0;                                                                \
+    byte b = 1;                                                                \
+    byte expected = 2;                                                         \
+    byte simd0 = r.AllocateLocal(kAstS128);                                    \
+    byte simd1 = r.AllocateLocal(kAstS128);                                    \
+    BUILD(r,                                                                   \
+          WASM_BLOCK(                                                          \
+              WASM_SET_LOCAL(simd0, WASM_SIMD_F32x4_SPLAT(WASM_GET_LOCAL(a))), \
+              WASM_SET_LOCAL(simd1, WASM_SIMD_F32x4_SPLAT(WASM_GET_LOCAL(b))), \
+              WASM_SET_LOCAL(simd1,                                            \
+                             WASM_SIMD_F32x4_##OP(WASM_GET_LOCAL(simd0),       \
+                                                  WASM_GET_LOCAL(simd1))),     \
+              WASM_SIMD_CHECK_SPLAT4_F32(F32x4, simd1, expected),              \
+              WASM_RETURN1(WASM_ONE)));                                        \
+                                                                               \
+    FOR_FLOAT32_INPUTS(i) {                                                    \
+      if (std::isnan(*i)) continue;                                            \
+      FOR_FLOAT32_INPUTS(j) {                                                  \
+        if (std::isnan(*j)) continue;                                          \
+        CHECK_EQ(1, r.Call(*i, *j, *i op *j));                                 \
+      }                                                                        \
+    }                                                                          \
+  }
+
+WASM_SIMD_F32x4_BINOP_TEST(Add, ADD, +);
+WASM_SIMD_F32x4_BINOP_TEST(Sub, SUB, -);
+#endif  // V8_TARGET_ARCH_ARM
+
 WASM_EXEC_TEST(I32x4Splat) {
   FLAG_wasm_simd_prototype = true;
 
@@ -89,46 +264,104 @@ WASM_EXEC_TEST(I32x4ReplaceLane) {
   CHECK_EQ(1, r.Call(1, 2));
 }
 
-WASM_EXEC_TEST(I32x4Add) {
-  FLAG_wasm_simd_prototype = true;
-  WasmRunner<int32_t, int32_t, int32_t, int32_t> r(kExecuteCompiled);
-  byte a = 0;
-  byte b = 1;
-  byte expected = 2;
-  byte simd0 = r.AllocateLocal(kAstS128);
-  byte simd1 = r.AllocateLocal(kAstS128);
-  BUILD(r,
-        WASM_BLOCK(
-            WASM_SET_LOCAL(simd0, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(a))),
-            WASM_SET_LOCAL(simd1, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(b))),
-            WASM_SET_LOCAL(simd1, WASM_SIMD_I32x4_ADD(WASM_GET_LOCAL(simd0),
-                                                      WASM_GET_LOCAL(simd1))),
-            WASM_SIMD_CHECK_SPLAT4(I32x4, simd1, I32, expected),
-            WASM_RETURN1(WASM_ONE)));
+#if V8_TARGET_ARCH_ARM
 
-  FOR_INT32_INPUTS(i) {
-    FOR_INT32_INPUTS(j) { CHECK_EQ(1, r.Call(*i, *j, *i + *j)); }
+// Determines if conversion from float to int will be valid.
+bool CanRoundToZeroAndConvert(double val, bool unsigned_integer) {
+  const double max_uint = static_cast<double>(0xffffffffu);
+  const double max_int = static_cast<double>(kMaxInt);
+  const double min_int = static_cast<double>(kMinInt);
+
+  // Check for NaN.
+  if (val != val) {
+    return false;
+  }
+
+  // Round to zero and check for overflow. This code works because 32 bit
+  // integers can be exactly represented by ieee-754 64bit floating-point
+  // values.
+  return unsigned_integer ? (val < (max_uint + 1.0)) && (val > -1)
+                          : (val < (max_int + 1.0)) && (val > (min_int - 1.0));
+}
+
+int ConvertInvalidValue(double val, bool unsigned_integer) {
+  if (val != val) {
+    return 0;
+  } else {
+    if (unsigned_integer) {
+      return (val < 0) ? 0 : 0xffffffffu;
+    } else {
+      return (val < 0) ? kMinInt : kMaxInt;
+    }
   }
 }
 
-WASM_EXEC_TEST(I32x4Sub) {
+int32_t ConvertToInt(double val, bool unsigned_integer) {
+  int32_t result =
+      unsigned_integer ? static_cast<uint32_t>(val) : static_cast<int32_t>(val);
+
+  if (!CanRoundToZeroAndConvert(val, unsigned_integer)) {
+    result = ConvertInvalidValue(val, unsigned_integer);
+  }
+  return result;
+}
+
+// Tests both signed and unsigned conversion.
+WASM_EXEC_TEST(I32x4FromFloat32x4) {
   FLAG_wasm_simd_prototype = true;
-  WasmRunner<int32_t, int32_t, int32_t, int32_t> r(kExecuteCompiled);
+  WasmRunner<int32_t, float, int32_t, int32_t> r(kExecuteCompiled);
   byte a = 0;
-  byte b = 1;
-  byte expected = 2;
+  byte expected_signed = 1;
+  byte expected_unsigned = 2;
   byte simd0 = r.AllocateLocal(kAstS128);
   byte simd1 = r.AllocateLocal(kAstS128);
+  byte simd2 = r.AllocateLocal(kAstS128);
   BUILD(r,
         WASM_BLOCK(
-            WASM_SET_LOCAL(simd0, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(a))),
-            WASM_SET_LOCAL(simd1, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(b))),
-            WASM_SET_LOCAL(simd1, WASM_SIMD_I32x4_SUB(WASM_GET_LOCAL(simd0),
-                                                      WASM_GET_LOCAL(simd1))),
-            WASM_SIMD_CHECK_SPLAT4(I32x4, simd1, I32, expected),
+            WASM_SET_LOCAL(simd0, WASM_SIMD_F32x4_SPLAT(WASM_GET_LOCAL(a))),
+            WASM_SET_LOCAL(
+                simd1, WASM_SIMD_I32x4_FROM_FLOAT32x4(WASM_GET_LOCAL(simd0))),
+            WASM_SIMD_CHECK_SPLAT4(I32x4, simd1, I32, expected_signed),
+            WASM_SET_LOCAL(
+                simd2, WASM_SIMD_UI32x4_FROM_FLOAT32x4(WASM_GET_LOCAL(simd0))),
+            WASM_SIMD_CHECK_SPLAT4(I32x4, simd2, I32, expected_unsigned),
             WASM_RETURN1(WASM_ONE)));
 
-  FOR_INT32_INPUTS(i) {
-    FOR_INT32_INPUTS(j) { CHECK_EQ(1, r.Call(*i, *j, *i - *j)); }
+  FOR_FLOAT32_INPUTS(i) {
+    int32_t signed_value = ConvertToInt(*i, false);
+    int32_t unsigned_value = ConvertToInt(*i, true);
+    CHECK_EQ(1, r.Call(*i, signed_value, unsigned_value));
   }
 }
+#endif  // V8_TARGET_ARCH_ARM
+
+#define WASM_SIMD_I32x4_BINOP_TEST(Name, OP, op)                               \
+  WASM_EXEC_TEST(I32x4##Name) {                                                \
+    FLAG_wasm_simd_prototype = true;                                           \
+    WasmRunner<int32_t, int32_t, int32_t, int32_t> r(kExecuteCompiled);        \
+    byte a = 0;                                                                \
+    byte b = 1;                                                                \
+    byte expected = 2;                                                         \
+    byte simd0 = r.AllocateLocal(kAstS128);                                    \
+    byte simd1 = r.AllocateLocal(kAstS128);                                    \
+    BUILD(r,                                                                   \
+          WASM_BLOCK(                                                          \
+              WASM_SET_LOCAL(simd0, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(a))), \
+              WASM_SET_LOCAL(simd1, WASM_SIMD_I32x4_SPLAT(WASM_GET_LOCAL(b))), \
+              WASM_SET_LOCAL(simd1,                                            \
+                             WASM_SIMD_I32x4_##OP(WASM_GET_LOCAL(simd0),       \
+                                                  WASM_GET_LOCAL(simd1))),     \
+              WASM_SIMD_CHECK_SPLAT4(I32x4, simd1, I32, expected),             \
+              WASM_RETURN1(WASM_ONE)));                                        \
+                                                                               \
+    FOR_INT32_INPUTS(i) {                                                      \
+      FOR_INT32_INPUTS(j) { CHECK_EQ(1, r.Call(*i, *j, op(*i, *j))); }         \
+    }                                                                          \
+  }
+
+WASM_SIMD_I32x4_BINOP_TEST(Add, ADD, Add);
+WASM_SIMD_I32x4_BINOP_TEST(Sub, SUB, Sub);
+#if V8_TARGET_ARCH_ARM
+WASM_SIMD_I32x4_BINOP_TEST(Equal, EQUAL, Equal);
+WASM_SIMD_I32x4_BINOP_TEST(NotEqual, NOT_EQUAL, NotEqual);
+#endif  // V8_TARGET_ARCH_ARM