Reland "[maglev] Deopt on overflow in >>>"

This is a reland of commit 24e60017d4

The reland changes %ClearFunctionFeedback to clear *all* feedback
slot kinds including binary/compare/for-in slots. In the tests we
thus no longer have to resort to tricks to restore the function to
it's initial state, instead simply call %ClearFunctionFeedback.

Original change's description:
> [maglev] Deopt on overflow in >>>
>
> Re-enable the int32 fast path for ShiftRightLogical, but account for
> Maglev's missing signed/unsigned representation tracking by a)
> removing rhs==0 as the identity value (a shift by 0 is still a
> signed-unsigned conversion) and b) deoptimizing if the result cannot
> be converted to a non-negative smi.
>
> Note this is not a deopt loop, since a non-smi result will change the
> feedback to kSignedSmallInputs (from kSignedSmall).
>
> To fix this properly, we should track signed/unsigned representations
> and convert the result to a heap number if it doesn't fit within smi
> range.
>
> Bug: v8:7700
> Change-Id: Ifd538d227a6f1290eb7f008d9bfad586ff91ea0f
> Fixed: v8:13251
> Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3876366
> Reviewed-by: Leszek Swirski <leszeks@chromium.org>
> Commit-Queue: Jakob Linke <jgruber@chromium.org>
> Cr-Commit-Position: refs/heads/main@{#83025}

Bug: v8:7700
Change-Id: I2f607a0fb863b80e8589c9c1e86ee31fbac48c25
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3879491
Auto-Submit: Jakob Linke <jgruber@chromium.org>
Commit-Queue: Jakob Linke <jgruber@chromium.org>
Reviewed-by: Leszek Swirski <leszeks@chromium.org>
Cr-Commit-Position: refs/heads/main@{#83057}

src/maglev/maglev-graph-builder.cc

@@ -176,8 +176,7 @@ bool BinaryOperationHasInt32FastPath() {
     case Operation::kBitwiseXor:
     case Operation::kShiftLeft:
     case Operation::kShiftRight:
-      // TODO(v8:13251): Implement with overflow protection.
-      //    case Operation::kShiftRightLogical:
+    case Operation::kShiftRightLogical:
     case Operation::kEqual:
     case Operation::kStrictEqual:
     case Operation::kLessThan:
@@ -205,21 +204,21 @@ bool BinaryOperationHasFloat64FastPath() {
 }  // namespace
 
 // MAP_OPERATION_TO_NODES are tuples with the following format:
-// (Operation name,
-//  Int32 operation node,
-//  Unit of int32 operation (e.g, 0 for add/sub and 1 for mul/div))
-#define MAP_OPERATION_TO_INT32_NODE(V)                      \
-  V(Add, Int32AddWithOverflow, 0)                           \
-  V(Subtract, Int32SubtractWithOverflow, 0)                 \
-  V(Multiply, Int32MultiplyWithOverflow, 1)                 \
-  V(Divide, Int32DivideWithOverflow, 1)                     \
-  V(BitwiseAnd, Int32BitwiseAnd, ~0)                        \
-  V(BitwiseOr, Int32BitwiseOr, 0)                           \
-  V(BitwiseXor, Int32BitwiseXor, 0)                         \
-  V(ShiftLeft, Int32ShiftLeft, 0)                           \
-  V(ShiftRight, Int32ShiftRight, 0)                         \
-  /* TODO(v8:13251): Implement with overflow protection. */ \
-  V(ShiftRightLogical, Int32ShiftRightLogical, 0)
+// - Operation name,
+// - Int32 operation node,
+// - Identity of int32 operation (e.g, 0 for add/sub and 1 for mul/div), if it
+//   exists, or otherwise {}.
+#define MAP_OPERATION_TO_INT32_NODE(V)      \
+  V(Add, Int32AddWithOverflow, 0)           \
+  V(Subtract, Int32SubtractWithOverflow, 0) \
+  V(Multiply, Int32MultiplyWithOverflow, 1) \
+  V(Divide, Int32DivideWithOverflow, 1)     \
+  V(BitwiseAnd, Int32BitwiseAnd, ~0)        \
+  V(BitwiseOr, Int32BitwiseOr, 0)           \
+  V(BitwiseXor, Int32BitwiseXor, 0)         \
+  V(ShiftLeft, Int32ShiftLeft, 0)           \
+  V(ShiftRight, Int32ShiftRight, 0)         \
+  V(ShiftRightLogical, Int32ShiftRightLogical, {})
 
 #define MAP_COMPARE_OPERATION_TO_INT32_NODE(V) \
   V(Equal, Int32Equal)                         \
@@ -238,11 +237,11 @@ bool BinaryOperationHasFloat64FastPath() {
   V(Divide, Float64Divide)
 
 template <Operation kOperation>
-static int Int32Unit() {
+static constexpr base::Optional<int> Int32Identity() {
   switch (kOperation) {
-#define CASE(op, OpNode, unit) \
-  case Operation::k##op:       \
-    return unit;
+#define CASE(op, _, identity) \
+  case Operation::k##op:      \
+    return identity;
     MAP_OPERATION_TO_INT32_NODE(CASE)
 #undef CASE
     default:
@@ -254,8 +253,8 @@ template <Operation kOperation>
 ValueNode* MaglevGraphBuilder::AddNewInt32BinaryOperationNode(
     std::initializer_list<ValueNode*> inputs) {
   switch (kOperation) {
-#define CASE(op, OpNode, unit) \
-  case Operation::k##op:       \
+#define CASE(op, OpNode, _) \
+  case Operation::k##op:    \
     return AddNewNode<OpNode>(inputs);
     MAP_OPERATION_TO_INT32_NODE(CASE)
 #undef CASE
@@ -328,7 +327,7 @@ void MaglevGraphBuilder::BuildInt32BinarySmiOperationNode() {
   // TODO(v8:7700): Do constant folding.
   ValueNode* left = GetAccumulatorInt32();
   int32_t constant = iterator_.GetImmediateOperand(0);
-  if (constant == Int32Unit<kOperation>()) {
+  if (base::Optional<int>(constant) == Int32Identity<kOperation>()) {
     // If the constant is the unit of the operation, it already has the right
     // value, so we can just return.
     return;

src/maglev/maglev-ir.cc

@@ -2056,6 +2056,13 @@ void Int32ShiftRightLogical::GenerateCode(MaglevAssembler* masm,
   Register left = ToRegister(left_input());
   DCHECK_EQ(rcx, ToRegister(right_input()));
   __ shrl_cl(left);
+  // The result of >>> is unsigned, but Maglev doesn't yet track
+  // signed/unsigned representations. Instead, deopt if the resulting smi would
+  // be negative.
+  // TODO(jgruber): Properly track signed/unsigned representations and
+  // allocated a heap number if the result is outside smi range.
+  __ testl(left, Immediate((1 << 31) | (1 << 30)));
+  EmitEagerDeoptIf(not_equal, masm, DeoptimizeReason::kOverflow, this);
 }
 
 namespace {

src/objects/feedback-vector.cc

@@ -479,7 +479,7 @@ void FeedbackVector::EvictOptimizedCodeMarkedForDeoptimization(
   }
 }
 
-bool FeedbackVector::ClearSlots(Isolate* isolate) {
+bool FeedbackVector::ClearSlots(Isolate* isolate, ClearBehavior behavior) {
   if (!shared_function_info().HasFeedbackMetadata()) return false;
   MaybeObject uninitialized_sentinel = MaybeObject::FromObject(
       FeedbackVector::RawUninitializedSentinel(isolate));
@@ -492,7 +492,7 @@ bool FeedbackVector::ClearSlots(Isolate* isolate) {
     MaybeObject obj = Get(slot);
     if (obj != uninitialized_sentinel) {
       FeedbackNexus nexus(*this, slot);
-      feedback_updated |= nexus.Clear();
+      feedback_updated |= nexus.Clear(behavior);
     }
   }
   return feedback_updated;
@@ -609,23 +609,37 @@ void FeedbackNexus::ConfigureUninitialized() {
   }
 }
 
-bool FeedbackNexus::Clear() {
+bool FeedbackNexus::Clear(ClearBehavior behavior) {
   bool feedback_updated = false;
 
   switch (kind()) {
     case FeedbackSlotKind::kTypeProfile:
-      // We don't clear these kinds ever.
+      if (V8_LIKELY(behavior == ClearBehavior::kDefault)) {
+        // We don't clear these kinds ever.
+      } else if (!IsCleared()) {
+        DCHECK_EQ(behavior, ClearBehavior::kClearAll);
+        SetFeedback(UninitializedSentinel(), SKIP_WRITE_BARRIER);
+        feedback_updated = true;
+      }
       break;
 
     case FeedbackSlotKind::kCompareOp:
     case FeedbackSlotKind::kForIn:
     case FeedbackSlotKind::kBinaryOp:
-      // We don't clear these, either.
+      if (V8_LIKELY(behavior == ClearBehavior::kDefault)) {
+        // We don't clear these, either.
+      } else if (!IsCleared()) {
+        DCHECK_EQ(behavior, ClearBehavior::kClearAll);
+        SetFeedback(Smi::zero(), SKIP_WRITE_BARRIER);
+        feedback_updated = true;
+      }
       break;
 
     case FeedbackSlotKind::kLiteral:
-      SetFeedback(Smi::zero(), SKIP_WRITE_BARRIER);
-      feedback_updated = true;
+      if (!IsCleared()) {
+        SetFeedback(Smi::zero(), SKIP_WRITE_BARRIER);
+        feedback_updated = true;
+      }
       break;
 
     case FeedbackSlotKind::kSetNamedSloppy:

src/objects/feedback-vector.h

@@ -28,6 +28,12 @@ class IsCompiledScope;
 
 enum class UpdateFeedbackMode { kOptionalFeedback, kGuaranteedFeedback };
 
+// Which feedback slots to clear in Clear().
+enum class ClearBehavior {
+  kDefault,
+  kClearAll,  // .. also ForIn, CompareOp, BinaryOp.
+};
+
 enum class FeedbackSlotKind : uint8_t {
   // This kind means that the slot points to the middle of other slot
   // which occupies more than one feedback vector element.
@@ -357,7 +363,14 @@ class FeedbackVector
   void FeedbackSlotPrint(std::ostream& os, FeedbackSlot slot);
 
   // Clears the vector slots. Return true if feedback has changed.
-  bool ClearSlots(Isolate* isolate);
+  bool ClearSlots(Isolate* isolate) {
+    return ClearSlots(isolate, ClearBehavior::kDefault);
+  }
+  // As above, but clears *all* slots - even those that we usually keep (e.g.:
+  // BinaryOp feedback).
+  bool ClearAllSlotsForTesting(Isolate* isolate) {
+    return ClearSlots(isolate, ClearBehavior::kClearAll);
+  }
 
   // The object that indicates an uninitialized cache.
   static inline Handle<Symbol> UninitializedSentinel(Isolate* isolate);
@@ -384,6 +397,8 @@ class FeedbackVector
   TQ_OBJECT_CONSTRUCTORS(FeedbackVector)
 
  private:
+  bool ClearSlots(Isolate* isolate, ClearBehavior behavior);
+
   static void AddToVectorsForProfilingTools(Isolate* isolate,
                                             Handle<FeedbackVector> vector);
 
@@ -811,7 +826,7 @@ class V8_EXPORT_PRIVATE FeedbackNexus final {
   }
 
   // Clear() returns true if the state of the underlying vector was changed.
-  bool Clear();
+  bool Clear(ClearBehavior behavior);
   void ConfigureUninitialized();
   // ConfigureMegamorphic() returns true if the state of the underlying vector
   // was changed. Extra feedback is cleared if the 0 parameter version is used.

src/objects/js-function.cc

@@ -1364,14 +1364,14 @@ void JSFunction::CalculateInstanceSizeHelper(InstanceType instance_type,
            static_cast<unsigned>(JSObject::kMaxInstanceSize));
 }
 
-void JSFunction::ClearTypeFeedbackInfo() {
+void JSFunction::ClearAllTypeFeedbackInfoForTesting() {
   ResetIfCodeFlushed();
   if (has_feedback_vector()) {
     FeedbackVector vector = feedback_vector();
     Isolate* isolate = GetIsolate();
-    if (vector.ClearSlots(isolate)) {
+    if (vector.ClearAllSlotsForTesting(isolate)) {
       IC::OnFeedbackChanged(isolate, vector, FeedbackSlot::Invalid(),
-                            "ClearTypeFeedbackInfo");
+                            "ClearAllTypeFeedbackInfoForTesting");
     }
   }
 }

src/objects/js-function.h

@@ -243,8 +243,9 @@ class JSFunction : public TorqueGeneratedJSFunction<
                                      IsCompiledScope* compiled_scope,
                                      bool reset_budget_for_feedback_allocation);
 
-  // Unconditionally clear the type feedback vector.
-  void ClearTypeFeedbackInfo();
+  // Unconditionally clear the type feedback vector, even those that we usually
+  // keep (e.g.: BinaryOp feedback).
+  void ClearAllTypeFeedbackInfoForTesting();
 
   // Resets function to clear compiled data after bytecode has been flushed.
   inline bool NeedsResetDueToFlushedBytecode();

src/runtime/runtime-test.cc

@@ -952,7 +952,7 @@ RUNTIME_FUNCTION(Runtime_ClearFunctionFeedback) {
   HandleScope scope(isolate);
   DCHECK_EQ(1, args.length());
   Handle<JSFunction> function = args.at<JSFunction>(0);
-  function->ClearTypeFeedbackInfo();
+  function->ClearAllTypeFeedbackInfoForTesting();
   return ReadOnlyRoots(isolate).undefined_value();
 }
 

test/mjsunit/maglev/shift-right-logical-smi.js

@@ -0,0 +1,48 @@
+// Copyright 2022 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.
+
+// Flags: --allow-natives-syntax --maglev
+
+function gen_shrl_smi(y) {
+  return function shrl_smi(x) { return x >>> y; };
+}
+
+function shrl_test(lhs, rhs, expected_result) {
+  const shrl = gen_shrl_smi(rhs);
+
+  // Warmup.
+  %PrepareFunctionForOptimization(shrl);
+  %ClearFunctionFeedback(shrl);
+  shrl(1);
+  %OptimizeMaglevOnNextCall(shrl);
+
+  assertEquals(expected_result, shrl(lhs));
+  assertTrue(isMaglevved(shrl));
+
+  %DeoptimizeFunction(shrl);
+  assertEquals(expected_result, shrl(lhs));
+}
+
+
+function shrl_test_expect_deopt(lhs, rhs, expected_result) {
+  const shrl = gen_shrl_smi(rhs);
+
+  // Warmup.
+  %PrepareFunctionForOptimization(shrl);
+  %ClearFunctionFeedback(shrl);
+  shrl(1);
+  %OptimizeMaglevOnNextCall(shrl);
+
+  assertEquals(expected_result, shrl(lhs));
+  assertFalse(isMaglevved(shrl));
+}
+
+shrl_test(8, 2, 2);
+shrl_test_expect_deopt(-1, 1, 2147483647);
+shrl_test(-8, 2, 1073741822);
+shrl_test_expect_deopt(-8, 0, 4294967288);
+shrl_test_expect_deopt(-892396978, 0, 3402570318);
+shrl_test(8, 10, 0);
+shrl_test(8, 33, 4);
+shrl_test_expect_deopt(0xFFFFFFFF, 0x3FFFFFFF, 1);

test/mjsunit/maglev/shift-right-logical.js

@@ -0,0 +1,44 @@
+// Copyright 2022 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.
+
+// Flags: --allow-natives-syntax --maglev
+
+function shrl(x, y) {
+  return x >>> y;
+}
+
+function shrl_test(lhs, rhs, expected_result) {
+  // Warmup.
+  %PrepareFunctionForOptimization(shrl);
+  %ClearFunctionFeedback(shrl);
+  shrl(1, 1);
+  %OptimizeMaglevOnNextCall(shrl);
+
+  assertEquals(expected_result, shrl(lhs, rhs));
+  assertTrue(isMaglevved(shrl));
+
+  %DeoptimizeFunction(shrl);
+  assertEquals(expected_result, shrl(lhs, rhs));
+}
+
+
+function shrl_test_expect_deopt(lhs, rhs, expected_result) {
+  // Warmup.
+  %PrepareFunctionForOptimization(shrl);
+  %ClearFunctionFeedback(shrl);
+  shrl(1, 1);
+  %OptimizeMaglevOnNextCall(shrl);
+
+  assertEquals(expected_result, shrl(lhs, rhs));
+  assertFalse(isMaglevved(shrl));
+}
+
+shrl_test(8, 2, 2);
+shrl_test_expect_deopt(-1, 1, 2147483647);
+shrl_test(-8, 2, 1073741822);
+shrl_test_expect_deopt(-8, 0, 4294967288);
+shrl_test_expect_deopt(-892396978, 0, 3402570318);
+shrl_test(8, 10, 0);
+shrl_test(8, 33, 4);
+shrl_test_expect_deopt(0xFFFFFFFF, 0x3FFFFFFF, 1);

test/mjsunit/maglev/shift-right-smi.js

@@ -4,55 +4,42 @@
 
 // Flags: --allow-natives-syntax --maglev
 
-// Checks Smi shift_right operation and deopt while untagging.
-(function() {
-  function shift_right(x, y) {
-    return x >> y;
-  }
+function gen_sarl_smi(y) {
+  return function sarl_smi(x) { return x >> y; };
+}
 
-  %PrepareFunctionForOptimization(shift_right);
-  assertEquals(2, shift_right(8, 2));
-  assertEquals(-2, shift_right(-8, 2));
-  assertEquals(-8, shift_right(-8, 0));
-  assertEquals(0, shift_right(8, 10));
-  assertEquals(4, shift_right(8, 33));
+function sarl_test(lhs, rhs, expected_result) {
+  const sarl = gen_sarl_smi(rhs);
 
-  %OptimizeMaglevOnNextCall(shift_right);
-  assertEquals(2, shift_right(8, 2));
-  assertTrue(isMaglevved(shift_right));
+  // Warmup.
+  %PrepareFunctionForOptimization(sarl);
+  %ClearFunctionFeedback(sarl);
+  sarl(1);
+  %OptimizeMaglevOnNextCall(sarl);
 
-  assertEquals(-2, shift_right(-8, 2));
-  assertTrue(isMaglevved(shift_right));
+  assertEquals(expected_result, sarl(lhs));
+  assertTrue(isMaglevved(sarl));
 
-  assertEquals(-8, shift_right(-8, 0));
-  assertTrue(isMaglevved(shift_right));
+  %DeoptimizeFunction(sarl);
+  assertEquals(expected_result, sarl(lhs));
+}
 
-  assertEquals(0, shift_right(8, 10));
-  assertTrue(isMaglevved(shift_right));
+function sarl_test_expect_deopt(lhs, rhs, expected_result) {
+  const sarl = gen_sarl_smi(rhs);
 
-  // Shifts are mod 32
-  assertEquals(4, shift_right(8, 33));
-  assertTrue(isMaglevved(shift_right));
+  // Warmup.
+  %PrepareFunctionForOptimization(sarl);
+  %ClearFunctionFeedback(sarl);
+  sarl(1);
+  %OptimizeMaglevOnNextCall(sarl);
 
-  // // We should deopt here in SmiUntag.
-  // assertEquals(0x40000000, shift_right(1, 0x3FFFFFFF));
-  // assertFalse(isMaglevved(shift_right));
-})();
+  assertEquals(expected_result, sarl(lhs));
+  assertFalse(isMaglevved(sarl));
+}
 
-// // Checks when we deopt due to tagging.
-// (function() {
-//   function shift_right(x, y) {
-//     return x + y;
-//   }
-
-//   %PrepareFunctionForOptimization(shift_right);
-//   assertEquals(3, shift_right(1, 2));
-
-//   %OptimizeMaglevOnNextCall(shift_right);
-//   assertEquals(3, shift_right(1, 2));
-//   assertTrue(isMaglevved(shift_right));
-
-//   // We should deopt here in SmiTag.
-//   assertEquals(3.2, shift_right(1.2, 2));
-//   assertFalse(isMaglevved(shift_right));
-// })();
+sarl_test(8, 2, 2);
+sarl_test(-8, 2, -2);
+sarl_test(-8, 0, -8);
+sarl_test(8, 10, 0);
+sarl_test(8, 33, 4);
+sarl_test_expect_deopt(0xFFFFFFFF, 0x3FFFFFFF, -1);

test/mjsunit/maglev/shift-right.js

@@ -0,0 +1,41 @@
+// Copyright 2022 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.
+
+// Flags: --allow-natives-syntax --maglev
+
+function sarl(x, y) {
+  return x >> y;
+}
+
+function sarl_test(lhs, rhs, expected_result) {
+  // Warmup.
+  %PrepareFunctionForOptimization(sarl);
+  %ClearFunctionFeedback(sarl);
+  sarl(1, 1);
+  %OptimizeMaglevOnNextCall(sarl);
+
+  assertEquals(expected_result, sarl(lhs, rhs));
+  assertTrue(isMaglevved(sarl));
+
+  %DeoptimizeFunction(sarl);
+  assertEquals(expected_result, sarl(lhs, rhs));
+}
+
+function sarl_test_expect_deopt(lhs, rhs, expected_result) {
+  // Warmup.
+  %PrepareFunctionForOptimization(sarl);
+  %ClearFunctionFeedback(sarl);
+  sarl(1, 1);
+  %OptimizeMaglevOnNextCall(sarl);
+
+  assertEquals(expected_result, sarl(lhs, rhs));
+  assertFalse(isMaglevved(sarl));
+}
+
+sarl_test(8, 2, 2);
+sarl_test(-8, 2, -2);
+sarl_test(-8, 0, -8);
+sarl_test(8, 10, 0);
+sarl_test(8, 33, 4);
+sarl_test_expect_deopt(0xFFFFFFFF, 0x3FFFFFFF, -1);