V8: Add SIMD functions.

src/runtime/runtime-simd.cc

 // Copyright 2015 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/v8.h"
 
 #include "src/arguments.h"
 #include "src/base/macros.h"
 #include "src/conversions.h"
 #include "src/runtime/runtime-utils.h"
 
 // Implement Single Instruction Multiple Data (SIMD) operations as defined in
 // the SIMD.js draft spec:
 // http://littledan.github.io/simd.html
 
-#define CONVERT_SIMD_LANE_ARG_CHECKED(name, index, lanes) \
-  CONVERT_INT32_ARG_CHECKED(name, index);                 \
-  RUNTIME_ASSERT(name >= 0 && name < lanes);
-
-#define SIMD_CREATE_NUMERIC_FUNCTION(type, lane_type, lane_count) \
-  RUNTIME_FUNCTION(Runtime_Create##type) {                        \
-    static const int kLaneCount = lane_count;                     \
-    HandleScope scope(isolate);                                   \
-    DCHECK(args.length() == kLaneCount);                          \
-    lane_type lanes[kLaneCount];                                  \
-    for (int i = 0; i < kLaneCount; i++) {                        \
-      CONVERT_NUMBER_ARG_HANDLE_CHECKED(number, i);               \
-      lanes[i] = ConvertNumber<lane_type>(number->Number());      \
-    }                                                             \
-    return *isolate->factory()->New##type(lanes);                 \
-  }
-
-#define SIMD_CREATE_BOOLEAN_FUNCTION(type, lane_count) \
-  RUNTIME_FUNCTION(Runtime_Create##type) {             \
-    static const int kLaneCount = lane_count;          \
-    HandleScope scope(isolate);                        \
-    DCHECK(args.length() == kLaneCount);               \
-    bool lanes[kLaneCount];                            \
-    for (int i = 0; i < kLaneCount; i++) {             \
-      lanes[i] = args[i]->BooleanValue();              \
-    }                                                  \
-    return *isolate->factory()->New##type(lanes);      \
-  }
-
-#define SIMD_CHECK_FUNCTION(type)           \
-  RUNTIME_FUNCTION(Runtime_##type##Check) { \
-    HandleScope scope(isolate);             \
-    CONVERT_ARG_HANDLE_CHECKED(type, a, 0); \
-    return *a;                              \
-  }
-
-#define SIMD_EXTRACT_LANE_FUNCTION(type, lanes, extract_fn)    \
-  RUNTIME_FUNCTION(Runtime_##type##ExtractLane) {              \
-    HandleScope scope(isolate);                                \
-    DCHECK(args.length() == 2);                                \
-    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                    \
-    CONVERT_SIMD_LANE_ARG_CHECKED(lane, 1, lanes);             \
-    return *isolate->factory()->extract_fn(a->get_lane(lane)); \
-  }
-
-#define SIMD_REPLACE_NUMERIC_LANE_FUNCTION(type, lane_type, lane_count) \
-  RUNTIME_FUNCTION(Runtime_##type##ReplaceLane) {                       \
-    static const int kLaneCount = lane_count;                           \
-    HandleScope scope(isolate);                                         \
-    DCHECK(args.length() == 3);                                         \
-    CONVERT_ARG_HANDLE_CHECKED(type, simd, 0);                          \
-    CONVERT_SIMD_LANE_ARG_CHECKED(lane, 1, kLaneCount);                 \
-    CONVERT_NUMBER_ARG_HANDLE_CHECKED(number, 2);                       \
-    lane_type lanes[kLaneCount];                                        \
-    for (int i = 0; i < kLaneCount; i++) {                              \
-      lanes[i] = simd->get_lane(i);                                     \
-    }                                                                   \
-    lanes[lane] = ConvertNumber<lane_type>(number->Number());           \
-    Handle<type> result = isolate->factory()->New##type(lanes);         \
-    return *result;                                                     \
-  }
-
-#define SIMD_REPLACE_BOOLEAN_LANE_FUNCTION(type, lane_count)    \
-  RUNTIME_FUNCTION(Runtime_##type##ReplaceLane) {               \
-    static const int kLaneCount = lane_count;                   \
-    HandleScope scope(isolate);                                 \
-    DCHECK(args.length() == 3);                                 \
-    CONVERT_ARG_HANDLE_CHECKED(type, simd, 0);                  \
-    CONVERT_SIMD_LANE_ARG_CHECKED(lane, 1, kLaneCount);         \
-    bool lanes[kLaneCount];                                     \
-    for (int i = 0; i < kLaneCount; i++) {                      \
-      lanes[i] = simd->get_lane(i);                             \
-    }                                                           \
-    lanes[lane] = args[2]->BooleanValue();                      \
-    Handle<type> result = isolate->factory()->New##type(lanes); \
-    return *result;                                             \
-  }
-
-
 namespace v8 {
 namespace internal {
 
 namespace {
 
 // Functions to convert Numbers to SIMD component types.
 
 template <typename T>
 static T ConvertNumber(double number);
 
@@ skipping 22 matching lines @@
 }
 
 
 bool Equals(Float32x4* a, Float32x4* b) {
   for (int i = 0; i < 4; i++) {
     if (a->get_lane(i) != b->get_lane(i)) return false;
   }
   return true;
 }
 
+
+inline float RecipApprox(float a) { return 1.0f / a; }
+
+
+inline float RecipSqrtApprox(float a) { return 1.0f / std::sqrt(a); }

[Dan Ehrenberg, 2015/08/06 00:12:25]

We may have to revisit these at some point, to create consistent results for
optimized code vs fullcodegen.

[bbudge, 2015/08/06 14:07:32]

I'll add a TODO.

+
+
+// Saturating addition for int16_t and int8_t.
+template <typename T>
+inline T AddSaturate(T a, T b) {
+  const T max = std::numeric_limits<T>::max();
+  const T min = std::numeric_limits<T>::min();
+  int32_t result = a + b;
+  if (result > max) return max;
+  if (result < min) return min;
+  return result;
+}
+
+
+// Saturating subtraction for int16_t and int8_t.
+template <typename T>
+inline T SubSaturate(T a, T b) {
+  const T max = std::numeric_limits<T>::max();
+  const T min = std::numeric_limits<T>::min();
+  int32_t result = a - b;
+  if (result > max) return max;
+  if (result < min) return min;
+  return result;
+}
+
+
+inline float Min(float a, float b) {
+  if (a < b) return a;
+  if (a > b) return b;
+  if (a == b) return std::signbit(a) ? a : b;
+  return std::numeric_limits<float>::quiet_NaN();
+}
+
+
+inline float Max(float a, float b) {
+  if (a > b) return a;
+  if (a < b) return b;
+  if (a == b) return std::signbit(b) ? a : b;
+  return std::numeric_limits<float>::quiet_NaN();
+}
+
+
+inline float MinNumber(float a, float b) {
+  if (std::isnan(a)) return b;
+  if (std::isnan(b)) return a;
+  return Min(a, b);
+}
+
+
+inline float MaxNumber(float a, float b) {
+  if (std::isnan(a)) return b;
+  if (std::isnan(b)) return a;
+  return Max(a, b);
+}
+
+
+inline bool CanCast(int32_t a) { return true; }
+
+
+inline bool CanCast(float a) {
+  return a > std::numeric_limits<int32_t>::min() &&
+         a < std::numeric_limits<int32_t>::max();
+}
+
 }  // namespace
 
+//-------------------------------------------------------------------
+
+// SIMD helper functions.
 
 RUNTIME_FUNCTION(Runtime_IsSimdObject) {
   HandleScope scope(isolate);
   DCHECK(args.length() == 1);
   return isolate->heap()->ToBoolean(args[0]->IsSimd128Value());
 }
 
 
 RUNTIME_FUNCTION(Runtime_SimdToObject) {
   HandleScope scope(isolate);
@@ skipping 56 matching lines @@
         result = Float32x4::cast(*a)->SameValueZero(Float32x4::cast(b));
       } else {
         result = a->BitwiseEquals(b);
       }
     }
   }
   return isolate->heap()->ToBoolean(result);
 }
 
 
-SIMD_CREATE_NUMERIC_FUNCTION(Float32x4, float, 4)
-SIMD_CREATE_NUMERIC_FUNCTION(Int32x4, int32_t, 4)
-SIMD_CREATE_BOOLEAN_FUNCTION(Bool32x4, 4)
-SIMD_CREATE_NUMERIC_FUNCTION(Int16x8, int16_t, 8)
-SIMD_CREATE_BOOLEAN_FUNCTION(Bool16x8, 8)
-SIMD_CREATE_NUMERIC_FUNCTION(Int8x16, int8_t, 16)
-SIMD_CREATE_BOOLEAN_FUNCTION(Bool8x16, 16)
-
-
-SIMD_CHECK_FUNCTION(Float32x4)
-SIMD_CHECK_FUNCTION(Int32x4)
-SIMD_CHECK_FUNCTION(Bool32x4)
-SIMD_CHECK_FUNCTION(Int16x8)
-SIMD_CHECK_FUNCTION(Bool16x8)
-SIMD_CHECK_FUNCTION(Int8x16)
-SIMD_CHECK_FUNCTION(Bool8x16)
-
-
-SIMD_EXTRACT_LANE_FUNCTION(Float32x4, 4, NewNumber)
-SIMD_EXTRACT_LANE_FUNCTION(Int32x4, 4, NewNumber)
-SIMD_EXTRACT_LANE_FUNCTION(Bool32x4, 4, ToBoolean)
-SIMD_EXTRACT_LANE_FUNCTION(Int16x8, 8, NewNumber)
-SIMD_EXTRACT_LANE_FUNCTION(Bool16x8, 8, ToBoolean)
-SIMD_EXTRACT_LANE_FUNCTION(Int8x16, 16, NewNumber)
-SIMD_EXTRACT_LANE_FUNCTION(Bool8x16, 16, ToBoolean)
-
+//-------------------------------------------------------------------
+
+// Utility macros.
+
+#define CONVERT_SIMD_LANE_ARG_CHECKED(name, index, lanes) \
+  CONVERT_INT32_ARG_CHECKED(name, index);                 \
+  RUNTIME_ASSERT(name >= 0 && name < lanes);
+
+#define SIMD_UNARY_OP(type, lane_type, lane_count, op, result) \
+  static const int kLaneCount = lane_count;                    \
+  DCHECK(args.length() == 1);                                  \
+  CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                      \
+  lane_type lanes[kLaneCount];                                 \
+  for (int i = 0; i < kLaneCount; i++) {                       \
+    lanes[i] = op(a->get_lane(i));                             \
+  }                                                            \
+  Handle<type> result = isolate->factory()->New##type(lanes);
+
+#define SIMD_BINARY_OP(type, lane_type, lane_count, op, result) \
+  static const int kLaneCount = lane_count;                     \
+  DCHECK(args.length() == 2);                                   \
+  CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                       \
+  CONVERT_ARG_HANDLE_CHECKED(type, b, 1);                       \
+  lane_type lanes[kLaneCount];                                  \
+  for (int i = 0; i < kLaneCount; i++) {                        \
+    lanes[i] = op(a->get_lane(i), b->get_lane(i));              \
+  }                                                             \
+  Handle<type> result = isolate->factory()->New##type(lanes);
+
+#define SIMD_RELATIONAL_OP(type, bool_type, lane_count, a, b, op, result) \
+  static const int kLaneCount = lane_count;                               \
+  DCHECK(args.length() == 2);                                             \
+  CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                                 \
+  CONVERT_ARG_HANDLE_CHECKED(type, b, 1);                                 \
+  bool lanes[kLaneCount];                                                 \
+  for (int i = 0; i < kLaneCount; i++) {                                  \
+    lanes[i] = a->get_lane(i) op b->get_lane(i);                          \
+  }                                                                       \
+  Handle<bool_type> result = isolate->factory()->New##bool_type(lanes);
+
+//-------------------------------------------------------------------
+
+// Common functions.
+
+#define GET_NUMERIC_ARG(lane_type, name, index) \
+  CONVERT_NUMBER_ARG_HANDLE_CHECKED(a, index);  \
+  name = ConvertNumber<lane_type>(a->Number());
+
+#define GET_BOOLEAN_ARG(lane_type, name, index) \
+  name = args[index]->BooleanValue();
+
+#define SIMD_ALL_TYPES(FUNCTION)                            \
+  FUNCTION(Float32x4, float, 4, NewNumber, GET_NUMERIC_ARG) \
+  FUNCTION(Int32x4, int32_t, 4, NewNumber, GET_NUMERIC_ARG) \
+  FUNCTION(Bool32x4, bool, 4, ToBoolean, GET_BOOLEAN_ARG)   \
+  FUNCTION(Int16x8, int16_t, 8, NewNumber, GET_NUMERIC_ARG) \
+  FUNCTION(Bool16x8, bool, 8, ToBoolean, GET_BOOLEAN_ARG)   \
+  FUNCTION(Int8x16, int8_t, 16, NewNumber, GET_NUMERIC_ARG) \
+  FUNCTION(Bool8x16, bool, 16, ToBoolean, GET_BOOLEAN_ARG)
+
+#define SIMD_CREATE_FUNCTION(type, lane_type, lane_count, extract, replace) \
+  RUNTIME_FUNCTION(Runtime_Create##type) {                                  \
+    static const int kLaneCount = lane_count;                               \
+    HandleScope scope(isolate);                                             \
+    DCHECK(args.length() == kLaneCount);                                    \
+    lane_type lanes[kLaneCount];                                            \
+    for (int i = 0; i < kLaneCount; i++) {                                  \
+      replace(lane_type, lanes[i], i)                                       \
+    }                                                                       \
+    return *isolate->factory()->New##type(lanes);                           \
+  }
+
+#define SIMD_EXTRACT_FUNCTION(type, lane_type, lane_count, extract, replace) \
+  RUNTIME_FUNCTION(Runtime_##type##ExtractLane) {                            \
+    HandleScope scope(isolate);                                              \
+    DCHECK(args.length() == 2);                                              \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                                  \
+    CONVERT_SIMD_LANE_ARG_CHECKED(lane, 1, lane_count);                      \
+    return *isolate->factory()->extract(a->get_lane(lane));                  \
+  }
+
+#define SIMD_REPLACE_FUNCTION(type, lane_type, lane_count, extract, replace) \
+  RUNTIME_FUNCTION(Runtime_##type##ReplaceLane) {                            \
+    static const int kLaneCount = lane_count;                                \
+    HandleScope scope(isolate);                                              \
+    DCHECK(args.length() == 3);                                              \
+    CONVERT_ARG_HANDLE_CHECKED(type, simd, 0);                               \
+    CONVERT_SIMD_LANE_ARG_CHECKED(lane, 1, kLaneCount);                      \
+    lane_type lanes[kLaneCount];                                             \
+    for (int i = 0; i < kLaneCount; i++) {                                   \
+      lanes[i] = simd->get_lane(i);                                          \
+    }                                                                        \
+    replace(lane_type, lanes[lane], 2);                                      \
+    Handle<type> result = isolate->factory()->New##type(lanes);              \
+    return *result;                                                          \
+  }
+
+#define SIMD_CHECK_FUNCTION(type, lane_type, lane_count, extract, replace) \
+  RUNTIME_FUNCTION(Runtime_##type##Check) {                                \
+    HandleScope scope(isolate);                                            \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                                \
+    return *a;                                                             \
+  }
+
+#define SIMD_SWIZZLE_FUNCTION(type, lane_type, lane_count, extract, replace) \
+  RUNTIME_FUNCTION(Runtime_##type##Swizzle) {                                \
+    static const int kLaneCount = lane_count;                                \
+    HandleScope scope(isolate);                                              \
+    DCHECK(args.length() == 1 + kLaneCount);                                 \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                                  \
+    lane_type lanes[kLaneCount];                                             \
+    for (int i = 0; i < kLaneCount; i++) {                                   \
+      CONVERT_SIMD_LANE_ARG_CHECKED(index, i + 1, kLaneCount);               \
+      lanes[i] = a->get_lane(index);                                         \
+    }                                                                        \
+    Handle<type> result = isolate->factory()->New##type(lanes);              \
+    return *result;                                                          \
+  }
+
+#define SIMD_SHUFFLE_FUNCTION(type, lane_type, lane_count, extract, replace) \
+  RUNTIME_FUNCTION(Runtime_##type##Shuffle) {                                \
+    static const int kLaneCount = lane_count;                                \
+    HandleScope scope(isolate);                                              \
+    DCHECK(args.length() == 2 + kLaneCount);                                 \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                                  \
+    CONVERT_ARG_HANDLE_CHECKED(type, b, 1);                                  \
+    lane_type lanes[kLaneCount];                                             \
+    for (int i = 0; i < kLaneCount; i++) {                                   \
+      CONVERT_SIMD_LANE_ARG_CHECKED(index, i + 2, kLaneCount * 2);           \
+      lanes[i] = index < kLaneCount ? a->get_lane(index)                     \
+                                    : b->get_lane(index - kLaneCount);       \
+    }                                                                        \
+    Handle<type> result = isolate->factory()->New##type(lanes);              \
+    return *result;                                                          \
+  }
+
+SIMD_ALL_TYPES(SIMD_CREATE_FUNCTION)
+SIMD_ALL_TYPES(SIMD_EXTRACT_FUNCTION)
+SIMD_ALL_TYPES(SIMD_REPLACE_FUNCTION)
+SIMD_ALL_TYPES(SIMD_CHECK_FUNCTION)
+SIMD_ALL_TYPES(SIMD_SWIZZLE_FUNCTION)
+SIMD_ALL_TYPES(SIMD_SHUFFLE_FUNCTION)
+
+//-------------------------------------------------------------------
+
+// Float-only functions.
+
+#define SIMD_ABS_FUNCTION(type, lane_type, lane_count)            \
+  RUNTIME_FUNCTION(Runtime_##type##Abs) {                         \
+    HandleScope scope(isolate);                                   \
+    SIMD_UNARY_OP(type, lane_type, lane_count, std::abs, result); \
+    return *result;                                               \
+  }
+
+#define SIMD_SQRT_FUNCTION(type, lane_type, lane_count)            \
+  RUNTIME_FUNCTION(Runtime_##type##Sqrt) {                         \
+    HandleScope scope(isolate);                                    \
+    SIMD_UNARY_OP(type, lane_type, lane_count, std::sqrt, result); \
+    return *result;                                                \
+  }
+
+#define SIMD_RECIP_APPROX_FUNCTION(type, lane_type, lane_count)      \
+  RUNTIME_FUNCTION(Runtime_##type##RecipApprox) {                    \
+    HandleScope scope(isolate);                                      \
+    SIMD_UNARY_OP(type, lane_type, lane_count, RecipApprox, result); \
+    return *result;                                                  \
+  }
+
+#define SIMD_RECIP_SQRT_APPROX_FUNCTION(type, lane_type, lane_count)     \
+  RUNTIME_FUNCTION(Runtime_##type##RecipSqrtApprox) {                    \
+    HandleScope scope(isolate);                                          \
+    SIMD_UNARY_OP(type, lane_type, lane_count, RecipSqrtApprox, result); \
+    return *result;                                                      \
+  }
+
+#define BINARY_DIV(a, b) (a) / (b)
+#define SIMD_DIV_FUNCTION(type, lane_type, lane_count)               \
+  RUNTIME_FUNCTION(Runtime_##type##Div) {                            \
+    HandleScope scope(isolate);                                      \
+    SIMD_BINARY_OP(type, lane_type, lane_count, BINARY_DIV, result); \
+    return *result;                                                  \
+  }
+
+#define SIMD_MINNUM_FUNCTION(type, lane_type, lane_count)           \
+  RUNTIME_FUNCTION(Runtime_##type##MinNum) {                        \
+    HandleScope scope(isolate);                                     \
+    SIMD_BINARY_OP(type, lane_type, lane_count, MinNumber, result); \
+    return *result;                                                 \
+  }
+
+#define SIMD_MAXNUM_FUNCTION(type, lane_type, lane_count)           \
+  RUNTIME_FUNCTION(Runtime_##type##MaxNum) {                        \
+    HandleScope scope(isolate);                                     \
+    SIMD_BINARY_OP(type, lane_type, lane_count, MaxNumber, result); \
+    return *result;                                                 \
+  }
+
+SIMD_ABS_FUNCTION(Float32x4, float, 4)
+SIMD_SQRT_FUNCTION(Float32x4, float, 4)
+SIMD_RECIP_APPROX_FUNCTION(Float32x4, float, 4)
+SIMD_RECIP_SQRT_APPROX_FUNCTION(Float32x4, float, 4)
+SIMD_DIV_FUNCTION(Float32x4, float, 4)
+SIMD_MINNUM_FUNCTION(Float32x4, float, 4)
+SIMD_MAXNUM_FUNCTION(Float32x4, float, 4)
+
+//-------------------------------------------------------------------
+
+// Int-only functions.
+
+#define SIMD_INT_TYPES(FUNCTION)            \
+  FUNCTION(Int32x4, int32_t, 0xffffffff, 4) \
+  FUNCTION(Int16x8, int16_t, 0xffff, 8)     \
+  FUNCTION(Int8x16, int8_t, 0xff, 16)
+
+#define SIMD_SHIFT_LEFT_FUNCTION(type, lane_type, mask, lane_count)    \
+  RUNTIME_FUNCTION(Runtime_##type##ShiftLeftByScalar) {                \
+    static const int kLaneCount = lane_count;                          \
+    HandleScope scope(isolate);                                        \
+    DCHECK(args.length() == 2);                                        \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                            \
+    CONVERT_INT32_ARG_CHECKED(shift, 1);                               \
+    shift = shift & 0x3f;                                              \

[Dan Ehrenberg, 2015/08/06 00:12:25]

Looks like you're implementing wrap-around behavior here (like JS <<). However, 
SIMD operations max out at 0 for shifting too far left; see
http://littledan.github.io/simd.html#simd-shift-left-by-scalar . This is done to
match what the processors apparently do. Analogous behavior exists on the right
shifts.

[bbudge, 2015/08/06 14:07:32]

Done.

+    lane_type lanes[kLaneCount];                                       \
+    for (int i = 0; i < kLaneCount; i++) {                             \
+      int64_t shifted = static_cast<int64_t>(a->get_lane(i)) << shift; \
+      lanes[i] = static_cast<lane_type>(shifted);                      \
+    }                                                                  \
+    Handle<type> result = isolate->factory()->New##type(lanes);        \
+    return *result;                                                    \
+  }
+
+#define SIMD_SHIFT_RIGHT_LOGICAL_FUNCTION(type, lane_type, mask, lane_count) \
+  RUNTIME_FUNCTION(Runtime_##type##ShiftRightLogicalByScalar) {              \
+    static const int kLaneCount = lane_count;                                \
+    HandleScope scope(isolate);                                              \
+    DCHECK(args.length() == 2);                                              \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                                  \
+    CONVERT_INT32_ARG_CHECKED(shift, 1);                                     \
+    shift = shift & 0x3f;                                                    \
+    lane_type lanes[kLaneCount];                                             \
+    for (int i = 0; i < kLaneCount; i++) {                                   \
+      int64_t shifted =                                                      \
+          (static_cast<int64_t>(a->get_lane(i)) & mask) >> shift;            \
+      lanes[i] = static_cast<lane_type>(shifted);                            \
+    }                                                                        \
+    Handle<type> result = isolate->factory()->New##type(lanes);              \
+    return *result;                                                          \
+  }
+
+#define SIMD_SHIFT_RIGHT_ARITHMETIC_FUNCTION(type, lane_type, mask,    \
+                                             lane_count)               \
+  RUNTIME_FUNCTION(Runtime_##type##ShiftRightArithmeticByScalar) {     \
+    static const int kLaneCount = lane_count;                          \
+    HandleScope scope(isolate);                                        \
+    DCHECK(args.length() == 2);                                        \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);                            \
+    CONVERT_INT32_ARG_CHECKED(shift, 1);                               \
+    shift = shift & 0x3f;                                              \
+    lane_type lanes[kLaneCount];                                       \
+    for (int i = 0; i < kLaneCount; i++) {                             \
+      int64_t shifted = static_cast<int64_t>(a->get_lane(i)) >> shift; \
+      lanes[i] = static_cast<lane_type>(shifted);                      \
+    }                                                                  \
+    Handle<type> result = isolate->factory()->New##type(lanes);        \
+    return *result;                                                    \
+  }
+
+SIMD_INT_TYPES(SIMD_SHIFT_LEFT_FUNCTION)
+SIMD_INT_TYPES(SIMD_SHIFT_RIGHT_LOGICAL_FUNCTION)
+SIMD_INT_TYPES(SIMD_SHIFT_RIGHT_ARITHMETIC_FUNCTION)
+
+//-------------------------------------------------------------------
+
+// Bool-only functions.
+
+#define SIMD_BOOL_TYPES(FUNCTION) \
+  FUNCTION(Bool32x4, 4)           \
+  FUNCTION(Bool16x8, 8)           \
+  FUNCTION(Bool8x16, 16)
+
+#define SIMD_ANY_FUNCTION(type, lane_count)    \
+  RUNTIME_FUNCTION(Runtime_##type##AnyTrue) {  \
+    HandleScope scope(isolate);                \
+    DCHECK(args.length() == 1);                \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);    \
+    bool result = false;                       \
+    for (int i = 0; i < lane_count; i++) {     \
+      if (a->get_lane(i)) {                    \
+        result = true;                         \
+        break;                                 \
+      }                                        \
+    }                                          \
+    return isolate->heap()->ToBoolean(result); \
+  }
+
+#define SIMD_ALL_FUNCTION(type, lane_count)    \
+  RUNTIME_FUNCTION(Runtime_##type##AllTrue) {  \
+    HandleScope scope(isolate);                \
+    DCHECK(args.length() == 1);                \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 0);    \
+    bool result = true;                        \
+    for (int i = 0; i < lane_count; i++) {     \
+      if (!a->get_lane(i)) {                   \
+        result = false;                        \
+        break;                                 \
+      }                                        \
+    }                                          \
+    return isolate->heap()->ToBoolean(result); \
+  }
+
+SIMD_BOOL_TYPES(SIMD_ANY_FUNCTION)
+SIMD_BOOL_TYPES(SIMD_ALL_FUNCTION)
+
+//-------------------------------------------------------------------
+
+// Small Int-only functions.
+
+#define SIMD_SMALL_INT_TYPES(FUNCTION) \
+  FUNCTION(Int16x8, int16_t, 8)        \
+  FUNCTION(Int8x16, int8_t, 16)
+
+#define SIMD_ADD_SATURATE_FUNCTION(type, lane_type, lane_count)       \
+  RUNTIME_FUNCTION(Runtime_##type##AddSaturate) {                     \
+    HandleScope scope(isolate);                                       \
+    SIMD_BINARY_OP(type, lane_type, lane_count, AddSaturate, result); \
+    return *result;                                                   \
+  }
+
+#define BINARY_SUB(a, b) (a) - (b)
+#define SIMD_SUB_SATURATE_FUNCTION(type, lane_type, lane_count)       \
+  RUNTIME_FUNCTION(Runtime_##type##SubSaturate) {                     \
+    HandleScope scope(isolate);                                       \
+    SIMD_BINARY_OP(type, lane_type, lane_count, SubSaturate, result); \
+    return *result;                                                   \
+  }
+
+SIMD_SMALL_INT_TYPES(SIMD_ADD_SATURATE_FUNCTION)
+SIMD_SMALL_INT_TYPES(SIMD_SUB_SATURATE_FUNCTION)
+
+//-------------------------------------------------------------------
+
+// Numeric functions.
+
+#define SIMD_NUMERIC_TYPES(FUNCTION) \
+  FUNCTION(Float32x4, float, 4)      \
+  FUNCTION(Int32x4, int32_t, 4)      \
+  FUNCTION(Int16x8, int16_t, 8)      \
+  FUNCTION(Int8x16, int8_t, 16)
+
+#define SIMD_NEG_FUNCTION(type, lane_type, lane_count)     \
+  RUNTIME_FUNCTION(Runtime_##type##Neg) {                  \
+    HandleScope scope(isolate);                            \
+    SIMD_UNARY_OP(type, lane_type, lane_count, -, result); \
+    return *result;                                        \
+  }
+
+#define BINARY_ADD(a, b) (a) + (b)

[Dan Ehrenberg, 2015/08/06 00:12:25]

We'll have to revisit ops like this for proper denormal behavior on floats on
ARMv7, to be consistent with optimized code. Fine for now though.

+#define SIMD_ADD_FUNCTION(type, lane_type, lane_count)               \
+  RUNTIME_FUNCTION(Runtime_##type##Add) {                            \
+    HandleScope scope(isolate);                                      \
+    SIMD_BINARY_OP(type, lane_type, lane_count, BINARY_ADD, result); \
+    return *result;                                                  \
+  }
+
+#define BINARY_SUB(a, b) (a) - (b)
+#define SIMD_SUB_FUNCTION(type, lane_type, lane_count)               \
+  RUNTIME_FUNCTION(Runtime_##type##Sub) {                            \
+    HandleScope scope(isolate);                                      \
+    SIMD_BINARY_OP(type, lane_type, lane_count, BINARY_SUB, result); \
+    return *result;                                                  \
+  }
+
+#define BINARY_MUL(a, b) (a) * (b)
+#define SIMD_MUL_FUNCTION(type, lane_type, lane_count)               \
+  RUNTIME_FUNCTION(Runtime_##type##Mul) {                            \
+    HandleScope scope(isolate);                                      \
+    SIMD_BINARY_OP(type, lane_type, lane_count, BINARY_MUL, result); \
+    return *result;                                                  \
+  }
+
+#define SIMD_MIN_FUNCTION(type, lane_type, lane_count)        \
+  RUNTIME_FUNCTION(Runtime_##type##Min) {                     \
+    HandleScope scope(isolate);                               \
+    SIMD_BINARY_OP(type, lane_type, lane_count, Min, result); \
+    return *result;                                           \
+  }
+
+#define SIMD_MAX_FUNCTION(type, lane_type, lane_count)        \
+  RUNTIME_FUNCTION(Runtime_##type##Max) {                     \
+    HandleScope scope(isolate);                               \
+    SIMD_BINARY_OP(type, lane_type, lane_count, Max, result); \
+    return *result;                                           \
+  }
+
+SIMD_NUMERIC_TYPES(SIMD_NEG_FUNCTION)
+SIMD_NUMERIC_TYPES(SIMD_ADD_FUNCTION)
+SIMD_NUMERIC_TYPES(SIMD_SUB_FUNCTION)
+SIMD_NUMERIC_TYPES(SIMD_MUL_FUNCTION)
+SIMD_NUMERIC_TYPES(SIMD_MIN_FUNCTION)
+SIMD_NUMERIC_TYPES(SIMD_MAX_FUNCTION)
+
+//-------------------------------------------------------------------
+
+// Relational functions.
+
+#define SIMD_RELATIONAL_TYPES(FUNCTION) \
+  FUNCTION(Float32x4, Bool32x4, 4)      \
+  FUNCTION(Int32x4, Bool32x4, 4)        \
+  FUNCTION(Int16x8, Bool16x8, 8)        \
+  FUNCTION(Int8x16, Bool8x16, 16)
+
+#define SIMD_EQUALITY_TYPES(FUNCTION) \
+  SIMD_RELATIONAL_TYPES(FUNCTION)     \
+  FUNCTION(Bool32x4, Bool32x4, 4)     \
+  FUNCTION(Bool16x8, Bool16x8, 8)     \
+  FUNCTION(Bool8x16, Bool8x16, 16)
+
+#define SIMD_EQUAL_FUNCTION(type, bool_type, lane_count)               \
+  RUNTIME_FUNCTION(Runtime_##type##Equal) {                            \
+    HandleScope scope(isolate);                                        \
+    SIMD_RELATIONAL_OP(type, bool_type, lane_count, a, b, ==, result); \
+    return *result;                                                    \
+  }
+
+#define SIMD_NOT_EQUAL_FUNCTION(type, bool_type, lane_count)           \
+  RUNTIME_FUNCTION(Runtime_##type##NotEqual) {                         \
+    HandleScope scope(isolate);                                        \
+    SIMD_RELATIONAL_OP(type, bool_type, lane_count, a, b, !=, result); \
+    return *result;                                                    \
+  }
+
+SIMD_EQUALITY_TYPES(SIMD_EQUAL_FUNCTION)
+SIMD_EQUALITY_TYPES(SIMD_NOT_EQUAL_FUNCTION)
+
+#define SIMD_LESS_THAN_FUNCTION(type, bool_type, lane_count)          \
+  RUNTIME_FUNCTION(Runtime_##type##LessThan) {                        \
+    HandleScope scope(isolate);                                       \
+    SIMD_RELATIONAL_OP(type, bool_type, lane_count, a, b, <, result); \
+    return *result;                                                   \
+  }
+
+#define SIMD_LESS_THAN_OR_EQUAL_FUNCTION(type, bool_type, lane_count)  \
+  RUNTIME_FUNCTION(Runtime_##type##LessThanOrEqual) {                  \
+    HandleScope scope(isolate);                                        \
+    SIMD_RELATIONAL_OP(type, bool_type, lane_count, a, b, <=, result); \
+    return *result;                                                    \
+  }
+
+#define SIMD_GREATER_THAN_FUNCTION(type, bool_type, lane_count)       \
+  RUNTIME_FUNCTION(Runtime_##type##GreaterThan) {                     \
+    HandleScope scope(isolate);                                       \
+    SIMD_RELATIONAL_OP(type, bool_type, lane_count, a, b, >, result); \
+    return *result;                                                   \
+  }
+
+#define SIMD_GREATER_THAN_OR_EQUAL_FUNCTION(type, bool_type, lane_count) \
+  RUNTIME_FUNCTION(Runtime_##type##GreaterThanOrEqual) {                 \
+    HandleScope scope(isolate);                                          \
+    SIMD_RELATIONAL_OP(type, bool_type, lane_count, a, b, >=, result);   \
+    return *result;                                                      \
+  }
+
+SIMD_RELATIONAL_TYPES(SIMD_LESS_THAN_FUNCTION)
+SIMD_RELATIONAL_TYPES(SIMD_LESS_THAN_OR_EQUAL_FUNCTION)
+SIMD_RELATIONAL_TYPES(SIMD_GREATER_THAN_FUNCTION)
+SIMD_RELATIONAL_TYPES(SIMD_GREATER_THAN_OR_EQUAL_FUNCTION)
+
+//-------------------------------------------------------------------
+
+// Logical functions.
+
+#define SIMD_LOGICAL_TYPES(FUNCTION)  \
+  FUNCTION(Int32x4, int32_t, 4, _INT) \
+  FUNCTION(Int16x8, int16_t, 8, _INT) \
+  FUNCTION(Int8x16, int8_t, 16, _INT) \
+  FUNCTION(Bool32x4, bool, 4, _BOOL)  \
+  FUNCTION(Bool16x8, bool, 8, _BOOL)  \
+  FUNCTION(Bool8x16, bool, 16, _BOOL)
+
+#define BINARY_AND_INT(a, b) (a) & (b)
+#define BINARY_AND_BOOL(a, b) (a) && (b)
+#define SIMD_AND_FUNCTION(type, lane_type, lane_count, op)               \
+  RUNTIME_FUNCTION(Runtime_##type##And) {                                \
+    HandleScope scope(isolate);                                          \
+    SIMD_BINARY_OP(type, lane_type, lane_count, BINARY_AND##op, result); \
+    return *result;                                                      \
+  }
+
+#define BINARY_OR_INT(a, b) (a) | (b)
+#define BINARY_OR_BOOL(a, b) (a) || (b)
+#define SIMD_OR_FUNCTION(type, lane_type, lane_count, op)               \
+  RUNTIME_FUNCTION(Runtime_##type##Or) {                                \
+    HandleScope scope(isolate);                                         \
+    SIMD_BINARY_OP(type, lane_type, lane_count, BINARY_OR##op, result); \
+    return *result;                                                     \
+  }
+
+#define BINARY_XOR_INT(a, b) (a) ^ (b)
+#define BINARY_XOR_BOOL(a, b) (a) != (b)
+#define SIMD_XOR_FUNCTION(type, lane_type, lane_count, op)               \
+  RUNTIME_FUNCTION(Runtime_##type##Xor) {                                \
+    HandleScope scope(isolate);                                          \
+    SIMD_BINARY_OP(type, lane_type, lane_count, BINARY_XOR##op, result); \
+    return *result;                                                      \
+  }
+
+#define UNARY_NOT_INT ~
+#define UNARY_NOT_BOOL !
+#define SIMD_NOT_FUNCTION(type, lane_type, lane_count, op)             \
+  RUNTIME_FUNCTION(Runtime_##type##Not) {                              \
+    HandleScope scope(isolate);                                        \
+    SIMD_UNARY_OP(type, lane_type, lane_count, UNARY_NOT##op, result); \
+    return *result;                                                    \
+  }
+
+SIMD_LOGICAL_TYPES(SIMD_AND_FUNCTION)
+SIMD_LOGICAL_TYPES(SIMD_OR_FUNCTION)
+SIMD_LOGICAL_TYPES(SIMD_XOR_FUNCTION)
+SIMD_LOGICAL_TYPES(SIMD_NOT_FUNCTION)
+
+//-------------------------------------------------------------------
+
+// Select functions.
+
+#define SIMD_SELECT_TYPES(FUNCTION)       \
+  FUNCTION(Float32x4, float, Bool32x4, 4) \
+  FUNCTION(Int32x4, int32_t, Bool32x4, 4) \
+  FUNCTION(Int16x8, int16_t, Bool16x8, 8) \
+  FUNCTION(Int8x16, int8_t, Bool8x16, 16) \
+  FUNCTION(Bool32x4, bool, Bool32x4, 4)   \
+  FUNCTION(Bool16x8, bool, Bool16x8, 8)   \
+  FUNCTION(Bool8x16, bool, Bool8x16, 16)
+
+#define SIMD_SELECT_FUNCTION(type, lane_type, bool_type, lane_count)  \
+  RUNTIME_FUNCTION(Runtime_##type##Select) {                          \
+    static const int kLaneCount = lane_count;                         \
+    HandleScope scope(isolate);                                       \
+    DCHECK(args.length() == 3);                                       \
+    CONVERT_ARG_HANDLE_CHECKED(bool_type, mask, 0);                   \
+    CONVERT_ARG_HANDLE_CHECKED(type, a, 1);                           \
+    CONVERT_ARG_HANDLE_CHECKED(type, b, 2);                           \
+    lane_type lanes[kLaneCount];                                      \
+    for (int i = 0; i < kLaneCount; i++) {                            \
+      lanes[i] = mask->get_lane(i) ? a->get_lane(i) : b->get_lane(i); \
+    }                                                                 \
+    Handle<type> result = isolate->factory()->New##type(lanes);       \
+    return *result;                                                   \
+  }
+
+SIMD_SELECT_TYPES(SIMD_SELECT_FUNCTION)
+
+//-------------------------------------------------------------------
+
+// Casting functions.
+
+#define SIMD_FROM_TYPES(FUNCTION)                 \
+  FUNCTION(Float32x4, float, 4, Int32x4, int32_t) \
+  FUNCTION(Int32x4, int32_t, 4, Float32x4, float)
+
+#define SIMD_FROM_FUNCTION(type, lane_type, lane_count, from_type, from_ctype) \
+  RUNTIME_FUNCTION(Runtime_##type##From##from_type) {                          \
+    static const int kLaneCount = lane_count;                                  \
+    HandleScope scope(isolate);                                                \
+    DCHECK(args.length() == 1);                                                \
+    CONVERT_ARG_HANDLE_CHECKED(from_type, a, 0);                               \
+    lane_type lanes[kLaneCount];                                               \
+    for (int i = 0; i < kLaneCount; i++) {                                     \
+      from_ctype a_value = a->get_lane(i);                                     \
+      RUNTIME_ASSERT(CanCast(a_value));                                        \
+      lanes[i] = static_cast<lane_type>(a_value);                              \
+    }                                                                          \
+    Handle<type> result = isolate->factory()->New##type(lanes);                \
+    return *result;                                                            \
+  }
+
+SIMD_FROM_TYPES(SIMD_FROM_FUNCTION)
+
+#define SIMD_FROM_BITS_TYPES(FUNCTION)     \
+  FUNCTION(Float32x4, float, 4, Int32x4)   \
+  FUNCTION(Float32x4, float, 4, Int16x8)   \
+  FUNCTION(Float32x4, float, 4, Int8x16)   \
+  FUNCTION(Int32x4, int32_t, 4, Float32x4) \
+  FUNCTION(Int32x4, int32_t, 4, Int16x8)   \
+  FUNCTION(Int32x4, int32_t, 4, Int8x16)   \
+  FUNCTION(Int16x8, int16_t, 8, Float32x4) \
+  FUNCTION(Int16x8, int16_t, 8, Int32x4)   \
+  FUNCTION(Int16x8, int16_t, 8, Int8x16)   \
+  FUNCTION(Int8x16, int8_t, 16, Float32x4) \
+  FUNCTION(Int8x16, int8_t, 16, Int32x4)   \
+  FUNCTION(Int8x16, int8_t, 16, Int16x8)
+
+#define SIMD_FROM_BITS_FUNCTION(type, lane_type, lane_count, from_type) \
+  RUNTIME_FUNCTION(Runtime_##type##From##from_type##Bits) {             \
+    static const int kLaneCount = lane_count;                           \
+    HandleScope scope(isolate);                                         \
+    DCHECK(args.length() == 1);                                         \
+    CONVERT_ARG_HANDLE_CHECKED(from_type, a, 0);                        \
+    lane_type lanes[kLaneCount];                                        \
+    a->CopyBits(lanes);                                                 \
+    Handle<type> result = isolate->factory()->New##type(lanes);         \
+    return *result;                                                     \
+  }
+
+SIMD_FROM_BITS_TYPES(SIMD_FROM_BITS_FUNCTION)
+
+//-------------------------------------------------------------------
+
+// Unsigned extract functions.
+// TODO(bbudge): remove when spec changes to include unsigned int types.
 
 RUNTIME_FUNCTION(Runtime_Int16x8UnsignedExtractLane) {
   HandleScope scope(isolate);
   DCHECK(args.length() == 2);
   CONVERT_ARG_HANDLE_CHECKED(Int16x8, a, 0);
   CONVERT_SIMD_LANE_ARG_CHECKED(lane, 1, 8);
   return *isolate->factory()->NewNumber(bit_cast<uint16_t>(a->get_lane(lane)));
 }
 
 
 RUNTIME_FUNCTION(Runtime_Int8x16UnsignedExtractLane) {
   HandleScope scope(isolate);
   DCHECK(args.length() == 2);
   CONVERT_ARG_HANDLE_CHECKED(Int8x16, a, 0);
   CONVERT_SIMD_LANE_ARG_CHECKED(lane, 1, 16);
   return *isolate->factory()->NewNumber(bit_cast<uint8_t>(a->get_lane(lane)));
 }
-
-
-SIMD_REPLACE_NUMERIC_LANE_FUNCTION(Float32x4, float, 4)
-SIMD_REPLACE_NUMERIC_LANE_FUNCTION(Int32x4, int32_t, 4)
-SIMD_REPLACE_BOOLEAN_LANE_FUNCTION(Bool32x4, 4)
-SIMD_REPLACE_NUMERIC_LANE_FUNCTION(Int16x8, int16_t, 8)
-SIMD_REPLACE_BOOLEAN_LANE_FUNCTION(Bool16x8, 8)
-SIMD_REPLACE_NUMERIC_LANE_FUNCTION(Int8x16, int8_t, 16)
-SIMD_REPLACE_BOOLEAN_LANE_FUNCTION(Bool8x16, 16)
 }  // namespace internal
 }  // namespace v8