[csa] block cloning to optimize branch on phi

This allows to return bool values from Torque macros and branch on them
without performance penalty, reconstructing good control flow.

Drive-by cleanup: Delete EnsureDeferredCodeSingleEntryPoint(), since
it's no longer needed. Constructing a graph and then re-inferring
deferred blocks based on branch hints achieves this effect
automatically.

Bug: v8:7793
Change-Id: Idb6802372b407549e4760f290933d5b8f1e9d952
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1681132Reviewed-by: Michael Starzinger <mstarzinger@chromium.org>
Commit-Queue: Tobias Tebbi <tebbi@chromium.org>
Cr-Commit-Position: refs/heads/master@{#62979}

src/compiler/raw-machine-assembler.cc

@@ -77,7 +77,7 @@ Node* RawMachineAssembler::OptimizedAllocate(
       size);
 }
 
-Schedule* RawMachineAssembler::Export() {
+Schedule* RawMachineAssembler::ExportForTest() {
   // Compute the correct codegen order.
   DCHECK(schedule_->rpo_order()->empty());
   if (FLAG_trace_turbo_scheduler) {
@@ -106,6 +106,7 @@ Graph* RawMachineAssembler::ExportForOptimization() {
     StdoutStream{} << *schedule_;
   }
   schedule_->EnsureCFGWellFormedness();
+  OptimizeControlFlow(schedule_, graph(), common());
   Scheduler::ComputeSpecialRPO(zone(), schedule_);
   if (FLAG_trace_turbo_scheduler) {
     PrintF("--- SCHEDULE BEFORE GRAPH CREATION -------------------------\n");
@@ -117,6 +118,99 @@ Graph* RawMachineAssembler::ExportForOptimization() {
   return graph();
 }
 
+void RawMachineAssembler::OptimizeControlFlow(Schedule* schedule, Graph* graph,
+                                              CommonOperatorBuilder* common) {
+  for (bool changed = true; changed;) {
+    changed = false;
+    for (size_t i = 0; i < schedule->all_blocks()->size(); ++i) {
+      BasicBlock* block = (*schedule->all_blocks())[i];
+      if (block == nullptr) continue;
+
+      // Short-circuit a goto if the succeeding block is not a control-flow
+      // merge. This is not really useful on it's own since graph construction
+      // has the same effect, but combining blocks improves the pattern-match on
+      // their structure below.
+      if (block->control() == BasicBlock::kGoto) {
+        DCHECK_EQ(block->SuccessorCount(), 1);
+        BasicBlock* successor = block->SuccessorAt(0);
+        if (successor->PredecessorCount() == 1) {
+          DCHECK_EQ(successor->PredecessorAt(0), block);
+          for (Node* node : *successor) {
+            schedule->SetBlockForNode(nullptr, node);
+            schedule->AddNode(block, node);
+          }
+          block->set_control(successor->control());
+          Node* control_input = successor->control_input();
+          block->set_control_input(control_input);
+          if (control_input) {
+            schedule->SetBlockForNode(block, control_input);
+          }
+          if (successor->deferred()) block->set_deferred(true);
+          block->ClearSuccessors();
+          schedule->MoveSuccessors(successor, block);
+          schedule->ClearBlockById(successor->id());
+          changed = true;
+          --i;
+          continue;
+        }
+      }
+      // Block-cloning in the simple case where a block consists only of a phi
+      // node and a branch on that phi. This just duplicates the branch block
+      // for each predecessor, replacing the phi node with the corresponding phi
+      // input.
+      if (block->control() == BasicBlock::kBranch && block->NodeCount() == 1) {
+        Node* phi = block->NodeAt(0);
+        if (phi->opcode() != IrOpcode::kPhi) continue;
+        Node* branch = block->control_input();
+        DCHECK_EQ(branch->opcode(), IrOpcode::kBranch);
+        if (NodeProperties::GetValueInput(branch, 0) != phi) continue;
+        if (phi->UseCount() != 1) continue;
+        DCHECK_EQ(phi->op()->ValueInputCount(), block->PredecessorCount());
+
+        // Turn projection blocks into normal blocks.
+        DCHECK_EQ(block->SuccessorCount(), 2);
+        BasicBlock* true_block = block->SuccessorAt(0);
+        BasicBlock* false_block = block->SuccessorAt(1);
+        DCHECK_EQ(true_block->NodeAt(0)->opcode(), IrOpcode::kIfTrue);
+        DCHECK_EQ(false_block->NodeAt(0)->opcode(), IrOpcode::kIfFalse);
+        (*true_block->begin())->Kill();
+        true_block->RemoveNode(true_block->begin());
+        (*false_block->begin())->Kill();
+        false_block->RemoveNode(false_block->begin());
+        true_block->ClearPredecessors();
+        false_block->ClearPredecessors();
+
+        size_t arity = block->PredecessorCount();
+        for (size_t i = 0; i < arity; ++i) {
+          BasicBlock* predecessor = block->PredecessorAt(i);
+          predecessor->ClearSuccessors();
+          if (block->deferred()) predecessor->set_deferred(true);
+          Node* branch_clone = graph->CloneNode(branch);
+          int phi_input = static_cast<int>(i);
+          NodeProperties::ReplaceValueInput(
+              branch_clone, NodeProperties::GetValueInput(phi, phi_input), 0);
+          BasicBlock* new_true_block = schedule->NewBasicBlock();
+          BasicBlock* new_false_block = schedule->NewBasicBlock();
+          new_true_block->AddNode(
+              graph->NewNode(common->IfTrue(), branch_clone));
+          new_false_block->AddNode(
+              graph->NewNode(common->IfFalse(), branch_clone));
+          schedule->AddGoto(new_true_block, true_block);
+          schedule->AddGoto(new_false_block, false_block);
+          DCHECK_EQ(predecessor->control(), BasicBlock::kGoto);
+          predecessor->set_control(BasicBlock::kNone);
+          schedule->AddBranch(predecessor, branch_clone, new_true_block,
+                              new_false_block);
+        }
+        branch->Kill();
+        schedule->ClearBlockById(block->id());
+        changed = true;
+        continue;
+      }
+    }
+  }
+}
+
 void RawMachineAssembler::MakeReschedulable() {
   std::vector<Node*> block_final_control(schedule_->all_blocks_.size());
   std::vector<Node*> block_final_effect(schedule_->all_blocks_.size());

src/compiler/raw-machine-assembler.h

@@ -65,9 +65,10 @@ class V8_EXPORT_PRIVATE RawMachineAssembler {
   CallDescriptor* call_descriptor() const { return call_descriptor_; }
   PoisoningMitigationLevel poisoning_level() const { return poisoning_level_; }
 
-  // Finalizes the schedule and exports it to be used for code generation. Note
-  // that this RawMachineAssembler becomes invalid after export.
-  Schedule* Export();
+  // Only used for tests: Finalizes the schedule and exports it to be used for
+  // code generation. Note that this RawMachineAssembler becomes invalid after
+  // export.
+  Schedule* ExportForTest();
   // Finalizes the schedule and transforms it into a graph that's suitable for
   // it to be used for Turbofan optimization and re-scheduling. Note that this
   // RawMachineAssembler becomes invalid after export.
@@ -1091,6 +1092,9 @@ class V8_EXPORT_PRIVATE RawMachineAssembler {
   Schedule* schedule() { return schedule_; }
   size_t parameter_count() const { return call_descriptor_->ParameterCount(); }
 
+  static void OptimizeControlFlow(Schedule* schedule, Graph* graph,
+                                  CommonOperatorBuilder* common);
+
   Isolate* isolate_;
   Graph* graph_;
   Schedule* schedule_;

src/compiler/schedule.cc

@@ -163,6 +163,11 @@ BasicBlock* Schedule::GetBlockById(BasicBlock::Id block_id) {
   return all_blocks_[block_id.ToSize()];
 }
 
+void Schedule::ClearBlockById(BasicBlock::Id block_id) {
+  DCHECK(block_id.ToSize() < all_blocks_.size());
+  all_blocks_[block_id.ToSize()] = nullptr;
+}
+
 bool Schedule::SameBasicBlock(Node* a, Node* b) const {
   BasicBlock* block = this->block(a);
   return block != nullptr && block == this->block(b);
@@ -321,9 +326,6 @@ void Schedule::EnsureCFGWellFormedness() {
       if (block != end_) {
         EnsureSplitEdgeForm(block);
       }
-      if (block->deferred()) {
-        EnsureDeferredCodeSingleEntryPoint(block);
-      }
     }
   }
 
@@ -356,6 +358,7 @@ void Schedule::EliminateRedundantPhiNodes() {
           }
           if (!inputs_equal) continue;
           node->ReplaceUses(first_input);
+          node->Kill();
           block->RemoveNode(block->begin() + node_pos);
           --node_pos;
           reached_fixed_point = false;
@@ -376,43 +379,6 @@ void Schedule::EnsureSplitEdgeForm(BasicBlock* block) {
 #endif
 }
 
-void Schedule::EnsureDeferredCodeSingleEntryPoint(BasicBlock* block) {
-  // If a deferred block has multiple predecessors, they have to
-  // all be deferred. Otherwise, we can run into a situation where a range
-  // that spills only in deferred blocks inserts its spill in the block, but
-  // other ranges need moves inserted by ResolveControlFlow in the predecessors,
-  // which may clobber the register of this range.
-  // To ensure that, when a deferred block has multiple predecessors, and some
-  // are not deferred, we add a non-deferred block to collect all such edges.
-
-  DCHECK(block->deferred() && block->PredecessorCount() > 1);
-  bool all_deferred = true;
-  for (auto current_pred = block->predecessors().begin();
-       current_pred != block->predecessors().end(); ++current_pred) {
-    BasicBlock* pred = *current_pred;
-    if (!pred->deferred()) {
-      all_deferred = false;
-      break;
-    }
-  }
-
-  if (all_deferred) return;
-  BasicBlock* merger = NewBasicBlock();
-  merger->set_control(BasicBlock::kGoto);
-  merger->successors().push_back(block);
-  for (auto current_pred = block->predecessors().begin();
-       current_pred != block->predecessors().end(); ++current_pred) {
-    BasicBlock* pred = *current_pred;
-    merger->predecessors().push_back(pred);
-    pred->successors().clear();
-    pred->successors().push_back(merger);
-  }
-  merger->set_deferred(false);
-  block->predecessors().clear();
-  block->predecessors().push_back(merger);
-  MovePhis(block, merger);
-}
-
 void Schedule::MovePhis(BasicBlock* from, BasicBlock* to) {
   for (size_t i = 0; i < from->NodeCount();) {
     Node* node = from->NodeAt(i);
@@ -481,6 +447,7 @@ void Schedule::SetBlockForNode(BasicBlock* block, Node* node) {
 std::ostream& operator<<(std::ostream& os, const Schedule& s) {
   for (BasicBlock* block :
        ((s.RpoBlockCount() == 0) ? *s.all_blocks() : *s.rpo_order())) {
+    if (block == nullptr) continue;
     if (block->rpo_number() == -1) {
       os << "--- BLOCK id:" << block->id().ToInt();
     } else {

src/compiler/schedule.h

@@ -200,6 +200,7 @@ class V8_EXPORT_PRIVATE Schedule final : public NON_EXPORTED_BASE(ZoneObject) {
 
   bool IsScheduled(Node* node);
   BasicBlock* GetBlockById(BasicBlock::Id block_id);
+  void ClearBlockById(BasicBlock::Id block_id);
 
   size_t BasicBlockCount() const { return all_blocks_.size(); }
   size_t RpoBlockCount() const { return rpo_order_.size(); }
@@ -280,8 +281,6 @@ class V8_EXPORT_PRIVATE Schedule final : public NON_EXPORTED_BASE(ZoneObject) {
   void EliminateRedundantPhiNodes();
   // Ensure split-edge form for a hand-assembled schedule.
   void EnsureSplitEdgeForm(BasicBlock* block);
-  // Ensure entry into a deferred block happens from a single hot block.
-  void EnsureDeferredCodeSingleEntryPoint(BasicBlock* block);
   // Move Phi operands to newly created merger blocks
   void MovePhis(BasicBlock* from, BasicBlock* to);
   // Copy deferred block markers down as far as possible

test/cctest/compiler/codegen-tester.h

@@ -79,7 +79,7 @@ class RawMachineAssemblerTester : public HandleAndZoneScope,
  protected:
   Address Generate() override {
     if (code_.is_null()) {
-      Schedule* schedule = this->Export();
+      Schedule* schedule = this->ExportForTest();
       auto call_descriptor = this->call_descriptor();
       Graph* graph = this->graph();
       OptimizedCompilationInfo info(ArrayVector("testing"), main_zone(), kind_);

test/cctest/compiler/test-multiple-return.cc

@@ -167,11 +167,11 @@ void TestReturnMultipleValues(MachineType type) {
 
       OptimizedCompilationInfo info(ArrayVector("testing"), handles.main_zone(),
                                     Code::WASM_FUNCTION);
-      Handle<Code> code =
-          Pipeline::GenerateCodeForTesting(
-              &info, handles.main_isolate(), desc, m.graph(),
-              AssemblerOptions::Default(handles.main_isolate()), m.Export())
-              .ToHandleChecked();
+      Handle<Code> code = Pipeline::GenerateCodeForTesting(
+                              &info, handles.main_isolate(), desc, m.graph(),
+                              AssemblerOptions::Default(handles.main_isolate()),
+                              m.ExportForTest())
+                              .ToHandleChecked();
 #ifdef ENABLE_DISASSEMBLER
       if (FLAG_print_code) {
         StdoutStream os;
@@ -272,11 +272,11 @@ void ReturnLastValue(MachineType type) {
 
     OptimizedCompilationInfo info(ArrayVector("testing"), handles.main_zone(),
                                   Code::WASM_FUNCTION);
-    Handle<Code> code =
-        Pipeline::GenerateCodeForTesting(
-            &info, handles.main_isolate(), desc, m.graph(),
-            AssemblerOptions::Default(handles.main_isolate()), m.Export())
-            .ToHandleChecked();
+    Handle<Code> code = Pipeline::GenerateCodeForTesting(
+                            &info, handles.main_isolate(), desc, m.graph(),
+                            AssemblerOptions::Default(handles.main_isolate()),
+                            m.ExportForTest())
+                            .ToHandleChecked();
 
     std::shared_ptr<wasm::NativeModule> module = AllocateNativeModule(
         handles.main_isolate(), code->raw_instruction_size());
@@ -334,11 +334,11 @@ void ReturnSumOfReturns(MachineType type) {
 
     OptimizedCompilationInfo info(ArrayVector("testing"), handles.main_zone(),
                                   Code::WASM_FUNCTION);
-    Handle<Code> code =
-        Pipeline::GenerateCodeForTesting(
-            &info, handles.main_isolate(), desc, m.graph(),
-            AssemblerOptions::Default(handles.main_isolate()), m.Export())
-            .ToHandleChecked();
+    Handle<Code> code = Pipeline::GenerateCodeForTesting(
+                            &info, handles.main_isolate(), desc, m.graph(),
+                            AssemblerOptions::Default(handles.main_isolate()),
+                            m.ExportForTest())
+                            .ToHandleChecked();
 
     std::shared_ptr<wasm::NativeModule> module = AllocateNativeModule(
         handles.main_isolate(), code->raw_instruction_size());

test/cctest/compiler/test-run-native-calls.cc

@@ -439,7 +439,7 @@ class Computer {
       Graph graph(&zone);
       RawMachineAssembler raw(isolate, &graph, desc);
       build(desc, raw);
-      inner = CompileGraph("Compute", desc, &graph, raw.Export());
+      inner = CompileGraph("Compute", desc, &graph, raw.ExportForTest());
     }
 
     CSignatureOf<int32_t> csig;
@@ -466,8 +466,8 @@ class Computer {
         Node* store = io.StoreOutput(raw, call);
         USE(store);
         raw.Return(raw.Int32Constant(seed));
-        wrapper =
-            CompileGraph("Compute-wrapper-const", cdesc, &graph, raw.Export());
+        wrapper = CompileGraph("Compute-wrapper-const", cdesc, &graph,
+                               raw.ExportForTest());
       }
 
       CodeRunner<int32_t> runnable(isolate, wrapper, &csig);
@@ -501,7 +501,8 @@ class Computer {
         Node* store = io.StoreOutput(raw, call);
         USE(store);
         raw.Return(raw.Int32Constant(seed));
-        wrapper = CompileGraph("Compute-wrapper", cdesc, &graph, raw.Export());
+        wrapper =
+            CompileGraph("Compute-wrapper", cdesc, &graph, raw.ExportForTest());
       }
 
       CodeRunner<int32_t> runnable(isolate, wrapper, &csig);
@@ -576,7 +577,7 @@ static void CopyTwentyInt32(CallDescriptor* desc) {
                 kNoWriteBarrier);
     }
     raw.Return(raw.Int32Constant(42));
-    inner = CompileGraph("CopyTwentyInt32", desc, &graph, raw.Export());
+    inner = CompileGraph("CopyTwentyInt32", desc, &graph, raw.ExportForTest());
   }
 
   CSignatureOf<int32_t> csig;
@@ -599,8 +600,8 @@ static void CopyTwentyInt32(CallDescriptor* desc) {
 
     Node* call = raw.CallN(desc, input_count, inputs);
     raw.Return(call);
-    wrapper =
-        CompileGraph("CopyTwentyInt32-wrapper", cdesc, &graph, raw.Export());
+    wrapper = CompileGraph("CopyTwentyInt32-wrapper", cdesc, &graph,
+                           raw.ExportForTest());
   }
 
   CodeRunner<int32_t> runnable(isolate, wrapper, &csig);
@@ -962,7 +963,8 @@ static void Build_Select_With_Call(
     Graph graph(&zone);
     RawMachineAssembler raw(isolate, &graph, desc);
     raw.Return(raw.Parameter(which));
-    inner = CompileGraph("Select-indirection", desc, &graph, raw.Export());
+    inner =
+        CompileGraph("Select-indirection", desc, &graph, raw.ExportForTest());
     CHECK(!inner.is_null());
     CHECK(inner->IsCode());
   }
@@ -1058,7 +1060,7 @@ void MixedParamTest(int start) {
       Graph graph(&zone);
       RawMachineAssembler raw(isolate, &graph, desc);
       raw.Return(raw.Parameter(which));
-      select = CompileGraph("Compute", desc, &graph, raw.Export());
+      select = CompileGraph("Compute", desc, &graph, raw.ExportForTest());
     }
 
     {
@@ -1117,7 +1119,7 @@ void MixedParamTest(int start) {
         expected_ret = static_cast<int32_t>(constant);
         raw.Return(raw.Int32Constant(expected_ret));
         wrapper = CompileGraph("Select-mixed-wrapper-const", cdesc, &graph,
-                               raw.Export());
+                               raw.ExportForTest());
       }
 
       CodeRunner<int32_t> runnable(isolate, wrapper, &csig);
@@ -1176,7 +1178,7 @@ void TestStackSlot(MachineType slot_type, T expected) {
   g.Store(slot_type.representation(), g.Parameter(11), g.Parameter(10),
           WriteBarrierKind::kNoWriteBarrier);
   g.Return(g.Parameter(9));
-  inner = CompileGraph("Compute", desc, &graph, g.Export());
+  inner = CompileGraph("Compute", desc, &graph, g.ExportForTest());
 
   // Create function f with a stack slot which calls the inner function g.
   BufferedRawMachineAssemblerTester<T> f(slot_type);

test/cctest/torque/test-torque.cc

@@ -585,6 +585,23 @@ TEST(TestGenericStruct2) {
   ft.Call();
 }
 
+TEST(TestBranchOnBoolOptimization) {
+  CcTest::InitializeVM();
+  Isolate* isolate(CcTest::i_isolate());
+  i::HandleScope scope(isolate);
+  Handle<Context> context =
+      Utils::OpenHandle(*v8::Isolate::GetCurrent()->GetCurrentContext());
+  CodeAssemblerTester asm_tester(isolate, 1);
+  TestTorqueAssembler m(asm_tester.state());
+  {
+    m.TestBranchOnBoolOptimization(
+        m.UncheckedCast<Context>(m.HeapConstant(context)),
+        m.UncheckedCast<Smi>(m.Parameter(0)));
+    m.Return(m.UndefinedConstant());
+  }
+  asm_tester.GenerateCode();
+}
+
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8

test/fuzzer/multi-return.cc

@@ -238,10 +238,11 @@ extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) {
   callee.Return(static_cast<int>(desc->ReturnCount()), returns.get());
 
   OptimizedCompilationInfo info(ArrayVector("testing"), &zone, Code::STUB);
-  Handle<Code> code = Pipeline::GenerateCodeForTesting(
-                          &info, i_isolate, desc, callee.graph(),
-                          AssemblerOptions::Default(i_isolate), callee.Export())
-                          .ToHandleChecked();
+  Handle<Code> code =
+      Pipeline::GenerateCodeForTesting(&info, i_isolate, desc, callee.graph(),
+                                       AssemblerOptions::Default(i_isolate),
+                                       callee.ExportForTest())
+          .ToHandleChecked();
 
   std::shared_ptr<wasm::NativeModule> module =
       AllocateNativeModule(i_isolate, code->raw_instruction_size());
@@ -286,7 +287,7 @@ extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) {
   Handle<Code> wrapper_code =
       Pipeline::GenerateCodeForTesting(
           &wrapper_info, i_isolate, wrapper_desc, caller.graph(),
-          AssemblerOptions::Default(i_isolate), caller.Export())
+          AssemblerOptions::Default(i_isolate), caller.ExportForTest())
           .ToHandleChecked();
 
   auto fn = GeneratedCode<int32_t>::FromCode(*wrapper_code);

test/torque/test-torque.tq

@@ -1008,4 +1008,26 @@ namespace test {
     return smiAndIntptr;
   }
 
+  macro BranchAndWriteResult(x: Smi, box: SmiBox): bool {
+    if (x > 5 || x < 0) {
+      box.value = 1;
+      return true;
+    } else {
+      box.value = 2;
+      return false;
+    }
+  }
+
+  @export
+  macro TestBranchOnBoolOptimization(implicit context: Context)(input: Smi) {
+    const box = NewSmiBox(1);
+    // If the two branches get combined into one, we should be able to determine
+    // the value of {box} statically.
+    if (BranchAndWriteResult(input, box)) {
+      StaticAssert(box.value == 1);
+    } else {
+      StaticAssert(box.value == 2);
+    }
+  }
+
 }

test/unittests/compiler/backend/instruction-selector-unittest.cc

@@ -25,7 +25,7 @@ InstructionSelectorTest::Stream InstructionSelectorTest::StreamBuilder::Build(
     InstructionSelector::Features features,
     InstructionSelectorTest::StreamBuilderMode mode,
     InstructionSelector::SourcePositionMode source_position_mode) {
-  Schedule* schedule = Export();
+  Schedule* schedule = ExportForTest();
   if (FLAG_trace_turbo) {
     StdoutStream{} << "=== Schedule before instruction selection ==="
                    << std::endl