[turboshaft][wasm] Implement some wasm requirements

- Add Turboshaft to the wasm pipeline (behind a flag).
- Add a few operators.
- Implement SimplifyLoopsPhase, which ensures each loop has at most
  two inputs.
- Remove the unneeded effect argument from
  {FlagContinuation::FromTrap}.

Bug: v8:12783
Change-Id: I03a3f8cf3af40fc75bf57cfbad973b754b13dd8c
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3899126
Commit-Queue: Manos Koukoutos <manoskouk@chromium.org>
Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Cr-Commit-Position: refs/heads/main@{#83365}

BUILD.bazel

@@ -2881,6 +2881,8 @@ filegroup(
         "src/compiler/turboshaft/representations.cc",
         "src/compiler/turboshaft/representations.h",
         "src/compiler/turboshaft/sidetable.h",
+        "src/compiler/turboshaft/simplify-tf-loops.cc",
+        "src/compiler/turboshaft/simplify-tf-loops.h",
         "src/compiler/turboshaft/utils.h",
         "src/compiler/turboshaft/value-numbering-assembler.h",
         "src/compiler/type-cache.cc",

BUILD.gn

@@ -2935,6 +2935,7 @@ v8_header_set("v8_internal_headers") {
     "src/compiler/turboshaft/recreate-schedule.h",
     "src/compiler/turboshaft/representations.h",
     "src/compiler/turboshaft/sidetable.h",
+    "src/compiler/turboshaft/simplify-tf-loops.h",
     "src/compiler/turboshaft/utils.h",
     "src/compiler/turboshaft/value-numbering-assembler.h",
     "src/compiler/type-cache.h",
@@ -4231,6 +4232,7 @@ v8_source_set("v8_turboshaft") {
     "src/compiler/turboshaft/optimization-phase.cc",
     "src/compiler/turboshaft/recreate-schedule.cc",
     "src/compiler/turboshaft/representations.cc",
+    "src/compiler/turboshaft/simplify-tf-loops.cc",
   ]
 
   public_deps = [

src/compiler/backend/instruction-selector.cc

@@ -14,12 +14,9 @@
 #include "src/compiler/backend/instruction-selector-impl.h"
 #include "src/compiler/compiler-source-position-table.h"
 #include "src/compiler/js-heap-broker.h"
-#include "src/compiler/node-matchers.h"
 #include "src/compiler/node-properties.h"
-#include "src/compiler/pipeline.h"
 #include "src/compiler/schedule.h"
 #include "src/compiler/state-values-utils.h"
-#include "src/deoptimizer/deoptimizer.h"
 
 #if V8_ENABLE_WEBASSEMBLY
 #include "src/wasm/simd-shuffle.h"
@@ -3137,14 +3134,12 @@ void InstructionSelector::VisitSelect(Node* node) {
 }
 
 void InstructionSelector::VisitTrapIf(Node* node, TrapId trap_id) {
-  FlagsContinuation cont =
-      FlagsContinuation::ForTrap(kNotEqual, trap_id, node->InputAt(1));
+  FlagsContinuation cont = FlagsContinuation::ForTrap(kNotEqual, trap_id);
   VisitWordCompareZero(node, node->InputAt(0), &cont);
 }
 
 void InstructionSelector::VisitTrapUnless(Node* node, TrapId trap_id) {
-  FlagsContinuation cont =
-      FlagsContinuation::ForTrap(kEqual, trap_id, node->InputAt(1));
+  FlagsContinuation cont = FlagsContinuation::ForTrap(kEqual, trap_id);
   VisitWordCompareZero(node, node->InputAt(0), &cont);
 }
 

src/compiler/backend/instruction-selector.h

@@ -8,7 +8,6 @@
 #include <map>
 
 #include "src/codegen/cpu-features.h"
-#include "src/common/globals.h"
 #include "src/compiler/backend/instruction-scheduler.h"
 #include "src/compiler/backend/instruction.h"
 #include "src/compiler/common-operator.h"
@@ -78,9 +77,8 @@ class FlagsContinuation final {
   }
 
   // Creates a new flags continuation for a wasm trap.
-  static FlagsContinuation ForTrap(FlagsCondition condition, TrapId trap_id,
-                                   Node* result) {
-    return FlagsContinuation(condition, trap_id, result);
+  static FlagsContinuation ForTrap(FlagsCondition condition, TrapId trap_id) {
+    return FlagsContinuation(condition, trap_id);
   }
 
   static FlagsContinuation ForSelect(FlagsCondition condition, Node* result,
@@ -218,13 +216,8 @@ class FlagsContinuation final {
     DCHECK_NOT_NULL(result);
   }
 
-  FlagsContinuation(FlagsCondition condition, TrapId trap_id, Node* result)
-      : mode_(kFlags_trap),
-        condition_(condition),
-        frame_state_or_result_(result),
-        trap_id_(trap_id) {
-    DCHECK_NOT_NULL(result);
-  }
+  FlagsContinuation(FlagsCondition condition, TrapId trap_id)
+      : mode_(kFlags_trap), condition_(condition), trap_id_(trap_id) {}
 
   FlagsContinuation(FlagsCondition condition, Node* result, Node* true_value,
                     Node* false_value)

src/compiler/pipeline.cc

@@ -85,6 +85,7 @@
 #include "src/compiler/turboshaft/graph.h"
 #include "src/compiler/turboshaft/optimization-phase.h"
 #include "src/compiler/turboshaft/recreate-schedule.h"
+#include "src/compiler/turboshaft/simplify-tf-loops.h"
 #include "src/compiler/turboshaft/value-numbering-assembler.h"
 #include "src/compiler/type-narrowing-reducer.h"
 #include "src/compiler/typed-optimization.h"
@@ -2104,6 +2105,19 @@ struct WasmGCOptimizationPhase {
   }
 };
 
+struct SimplifyLoopsPhase {
+  DECL_PIPELINE_PHASE_CONSTANTS(SimplifyLoops)
+
+  void Run(PipelineData* data, Zone* temp_zone) {
+    GraphReducer graph_reducer(
+        temp_zone, data->graph(), &data->info()->tick_counter(), data->broker(),
+        data->jsgraph()->Dead(), data->observe_node_manager());
+    SimplifyTFLoops simplify_loops(&graph_reducer, data->mcgraph());
+    AddReducer(data, &graph_reducer, &simplify_loops);
+    graph_reducer.ReduceGraph();
+  }
+};
+
 struct WasmGCLoweringPhase {
   DECL_PIPELINE_PHASE_CONSTANTS(WasmGCLowering)
 
@@ -3384,10 +3398,12 @@ void Pipeline::GenerateCodeForWasmFunction(
     pipeline.Run<MachineOperatorOptimizationPhase>();
     pipeline.RunPrintAndVerify(MachineOperatorOptimizationPhase::phase_name(),
                                true);
+    if (!v8_flags.turboshaft_wasm) {
       pipeline.Run<DecompressionOptimizationPhase>();
       pipeline.RunPrintAndVerify(DecompressionOptimizationPhase::phase_name(),
                                  true);
     }
+  }
 
   if (v8_flags.wasm_opt) {
     pipeline.Run<BranchConditionDuplicationPhase>();
@@ -3395,6 +3411,11 @@ void Pipeline::GenerateCodeForWasmFunction(
                                true);
   }
 
+  if (v8_flags.turboshaft_wasm) {
+    pipeline.Run<SimplifyLoopsPhase>();
+    pipeline.RunPrintAndVerify(SimplifyLoopsPhase::phase_name(), true);
+  }
+
   if (v8_flags.turbo_splitting && !is_asm_js) {
     data.info()->set_splitting();
   }
@@ -3407,6 +3428,30 @@ void Pipeline::GenerateCodeForWasmFunction(
   pipeline.ComputeScheduledGraph();
 
   Linkage linkage(call_descriptor);
+
+  if (v8_flags.turboshaft_wasm) {
+    if (base::Optional<BailoutReason> bailout =
+            pipeline.Run<BuildTurboshaftPhase>()) {
+      pipeline.info()->AbortOptimization(*bailout);
+      data.EndPhaseKind();
+      info->SetWasmCompilationResult({});
+      return;
+    }
+    pipeline.Run<PrintTurboshaftGraphPhase>(BuildTurboshaftPhase::phase_name());
+
+    pipeline.Run<OptimizeTurboshaftPhase>();
+    pipeline.Run<PrintTurboshaftGraphPhase>(
+        OptimizeTurboshaftPhase::phase_name());
+
+    pipeline.Run<DecompressionOptimizationPhase>();
+    pipeline.Run<PrintTurboshaftGraphPhase>(
+        DecompressionOptimizationPhase::phase_name());
+
+    pipeline.Run<TurboshaftRecreateSchedulePhase>(&linkage);
+    TraceSchedule(data.info(), &data, data.schedule(),
+                  TurboshaftRecreateSchedulePhase::phase_name());
+  }
+
   if (!pipeline.SelectInstructions(&linkage)) return;
   pipeline.AssembleCode(&linkage);
 

src/compiler/turboshaft/assembler.h

@@ -11,16 +11,13 @@
 #include <memory>
 #include <type_traits>
 
-#include "src/base/iterator.h"
 #include "src/base/logging.h"
 #include "src/base/macros.h"
 #include "src/base/small-vector.h"
 #include "src/base/template-utils.h"
-#include "src/codegen/machine-type.h"
-#include "src/codegen/source-position.h"
+#include "src/codegen/reloc-info.h"
 #include "src/compiler/turboshaft/graph.h"
 #include "src/compiler/turboshaft/operations.h"
-#include "src/zone/zone-containers.h"
 
 namespace v8::internal::compiler::turboshaft {
 
@@ -359,9 +356,29 @@ class AssemblerInterface : public Superclass {
                         WordRepresentation::Word32())
   DECL_SINGLE_REP_UNARY(Word64CountLeadingZeros, WordUnary, CountLeadingZeros,
                         WordRepresentation::Word64())
+  DECL_MULTI_REP_UNARY(WordCountTrailingZeros, WordUnary, WordRepresentation,
+                       CountTrailingZeros)
+  DECL_SINGLE_REP_UNARY(Word32CountTrailingZeros, WordUnary, CountTrailingZeros,
+                        WordRepresentation::Word32())
+  DECL_SINGLE_REP_UNARY(Word64CountTrailingZeros, WordUnary, CountTrailingZeros,
+                        WordRepresentation::Word64())
+  DECL_MULTI_REP_UNARY(WordPopCount, WordUnary, WordRepresentation, PopCount)
+  DECL_SINGLE_REP_UNARY(Word32PopCount, WordUnary, PopCount,
+                        WordRepresentation::Word32())
+  DECL_SINGLE_REP_UNARY(Word64PopCount, WordUnary, PopCount,
+                        WordRepresentation::Word64())
 #undef DECL_SINGLE_REP_UNARY
 #undef DECL_MULTI_REP_UNARY
 
+  OpIndex Word32Select(OpIndex condition, OpIndex left, OpIndex right) {
+    return subclass().Select(condition, left, right,
+                             WordRepresentation::Word32());
+  }
+  OpIndex Word64Select(OpIndex condition, OpIndex left, OpIndex right) {
+    return subclass().Select(condition, left, right,
+                             WordRepresentation::Word64());
+  }
+
   OpIndex Word32Constant(uint32_t value) {
     return subclass().Constant(ConstantOp::Kind::kWord32, uint64_t{value});
   }
@@ -413,6 +430,13 @@ class AssemblerInterface : public Superclass {
   OpIndex ExternalConstant(ExternalReference value) {
     return subclass().Constant(ConstantOp::Kind::kExternal, value);
   }
+  OpIndex RelocatableConstant(int64_t value, RelocInfo::Mode mode) {
+    DCHECK_EQ(mode, any_of(RelocInfo::WASM_CALL, RelocInfo::WASM_STUB_CALL));
+    return subclass().Constant(mode == RelocInfo::WASM_CALL
+                                   ? ConstantOp::Kind::kRelocatableWasmCall
+                                   : ConstantOp::Kind::kRelocatableWasmStubCall,
+                               static_cast<uint64_t>(value));
+  }
 
 #define DECL_CHANGE(name, kind, from, to)                    \
   OpIndex name(OpIndex input) {                              \

src/compiler/turboshaft/deopt-data.h

@@ -5,6 +5,7 @@
 #ifndef V8_COMPILER_TURBOSHAFT_DEOPT_DATA_H_
 #define V8_COMPILER_TURBOSHAFT_DEOPT_DATA_H_
 
+#include "src/base/small-vector.h"
 #include "src/common/globals.h"
 #include "src/compiler/turboshaft/operations.h"
 

src/compiler/turboshaft/graph-builder.cc

@@ -19,7 +19,6 @@
 #include "src/compiler/machine-operator.h"
 #include "src/compiler/node-aux-data.h"
 #include "src/compiler/node-origin-table.h"
-#include "src/compiler/node-properties.h"
 #include "src/compiler/opcodes.h"
 #include "src/compiler/operator.h"
 #include "src/compiler/schedule.h"
@@ -202,6 +201,7 @@ base::Optional<BailoutReason> GraphBuilder::Run() {
       case BasicBlock::kReturn:
       case BasicBlock::kDeoptimize:
       case BasicBlock::kThrow:
+      case BasicBlock::kTailCall:
         break;
       case BasicBlock::kCall: {
         Node* call = block->control_input();
@@ -216,8 +216,6 @@ base::Optional<BailoutReason> GraphBuilder::Run() {
         op_mapping.Set(if_exception_node, catch_exception);
         break;
       }
-      case BasicBlock::kTailCall:
-        UNIMPLEMENTED();
       case BasicBlock::kNone:
         UNREACHABLE();
     }
@@ -323,7 +321,10 @@ OpIndex GraphBuilder::Process(
       return assembler.CompressedHeapConstant(HeapConstantOf(op));
     case IrOpcode::kExternalConstant:
       return assembler.ExternalConstant(OpParameter<ExternalReference>(op));
-
+    case IrOpcode::kRelocatableInt64Constant:
+      return assembler.RelocatableConstant(
+          OpParameter<RelocatablePtrConstantInfo>(op).value(),
+          OpParameter<RelocatablePtrConstantInfo>(op).rmode());
 #define BINOP_CASE(opcode, assembler_op) \
   case IrOpcode::k##opcode:              \
     return assembler.assembler_op(Map(node->InputAt(0)), Map(node->InputAt(1)));
@@ -431,6 +432,10 @@ OpIndex GraphBuilder::Process(
       UNARY_CASE(Word64ReverseBytes, Word64ReverseBytes)
       UNARY_CASE(Word32Clz, Word32CountLeadingZeros)
       UNARY_CASE(Word64Clz, Word64CountLeadingZeros)
+      UNARY_CASE(Word32Ctz, Word32CountTrailingZeros)
+      UNARY_CASE(Word64Ctz, Word64CountTrailingZeros)
+      UNARY_CASE(Word32Popcnt, Word32PopCount)
+      UNARY_CASE(Word64Popcnt, Word64PopCount)
 
       UNARY_CASE(Float32Abs, Float32Abs)
       UNARY_CASE(Float64Abs, Float64Abs)
@@ -481,9 +486,17 @@ OpIndex GraphBuilder::Process(
       CHANGE_CASE(BitcastInt64ToFloat64, Bitcast, Word64, Float64)
       CHANGE_CASE(ChangeUint32ToUint64, ZeroExtend, Word32, Word64)
       CHANGE_CASE(ChangeInt32ToInt64, SignExtend, Word32, Word64)
-      CHANGE_CASE(ChangeInt32ToFloat64, SignedToFloat, Word32, Float64)
-      CHANGE_CASE(ChangeInt64ToFloat64, SignedToFloat, Word64, Float64)
+      CHANGE_CASE(SignExtendWord32ToInt64, SignExtend, Word32, Word64)
+      CHANGE_CASE(ChangeInt32ToFloat64, SignedNarrowing, Word32, Float64)
+      CHANGE_CASE(ChangeInt64ToFloat64, SignedNarrowing, Word64, Float64)
       CHANGE_CASE(ChangeUint32ToFloat64, UnsignedToFloat, Word32, Float64)
+      CHANGE_CASE(RoundInt64ToFloat64, SignedToFloat, Word64, Float64)
+      CHANGE_CASE(RoundUint64ToFloat64, UnsignedToFloat, Word64, Float64)
+      CHANGE_CASE(RoundInt32ToFloat32, SignedToFloat, Word32, Float32)
+      CHANGE_CASE(RoundUint32ToFloat32, UnsignedToFloat, Word32, Float32)
+      CHANGE_CASE(RoundInt64ToFloat32, SignedToFloat, Word64, Float32)
+      CHANGE_CASE(RoundUint64ToFloat32, UnsignedToFloat, Word64, Float32)
+
       CHANGE_CASE(TruncateFloat64ToWord32, JSFloatTruncate, Float64, Word32)
       CHANGE_CASE(TruncateFloat64ToFloat32, FloatConversion, Float64, Float32)
       CHANGE_CASE(ChangeFloat32ToFloat64, FloatConversion, Float32, Float64)
@@ -498,21 +511,36 @@ OpIndex GraphBuilder::Process(
     case IrOpcode::kTruncateInt64ToInt32:
       // 64- to 32-bit truncation is implicit in Turboshaft.
       return Map(node->InputAt(0));
-    case IrOpcode::kTruncateFloat64ToInt64: {
-      ChangeOp::Kind kind;
-      switch (OpParameter<TruncateKind>(op)) {
-        case TruncateKind::kArchitectureDefault:
-          kind = ChangeOp::Kind::kSignedFloatTruncate;
-          break;
-        case TruncateKind::kSetOverflowToMin:
-          kind = ChangeOp::Kind::kSignedFloatTruncateOverflowToMin;
-          break;
+    case IrOpcode::kTruncateFloat32ToInt32: {
+      ChangeOp::Kind kind =
+          OpParameter<TruncateKind>(node->op()) ==
+                  TruncateKind::kArchitectureDefault
+              ? ChangeOp::Kind::kSignedFloatTruncate
+              : ChangeOp::Kind::kSignedFloatTruncateOverflowToMin;
+      return assembler.Change(Map(node->InputAt(0)), kind,
+                              RegisterRepresentation::Float32(),
+                              RegisterRepresentation::Word32());
+    }
+    case IrOpcode::kTruncateFloat32ToUint32: {
+      ChangeOp::Kind kind =
+          OpParameter<TruncateKind>(node->op()) ==
+                  TruncateKind::kArchitectureDefault
+              ? ChangeOp::Kind::kUnsignedFloatTruncate
+              : ChangeOp::Kind::kUnsignedFloatTruncateOverflowToMin;
+      return assembler.Change(Map(node->InputAt(0)), kind,
+                              RegisterRepresentation::Float32(),
+                              RegisterRepresentation::Word32());
     }
+    case IrOpcode::kTruncateFloat64ToInt64: {
+      ChangeOp::Kind kind =
+          OpParameter<TruncateKind>(node->op()) ==
+                  TruncateKind::kArchitectureDefault
+              ? ChangeOp::Kind::kSignedFloatTruncate
+              : ChangeOp::Kind::kSignedFloatTruncateOverflowToMin;
       return assembler.Change(Map(node->InputAt(0)), kind,
                               RegisterRepresentation::Float64(),
                               RegisterRepresentation::Word64());
     }
-
     case IrOpcode::kFloat64InsertLowWord32:
       return assembler.Float64InsertWord32(
           Map(node->InputAt(0)), Map(node->InputAt(1)),
@@ -531,16 +559,27 @@ OpIndex GraphBuilder::Process(
                                      RegisterRepresentation::PointerSized(),
                                      RegisterRepresentation::Tagged());
 
+    case IrOpcode::kWord32Select:
+      return assembler.Word32Select(
+          Map(node->InputAt(0)), Map(node->InputAt(1)), Map(node->InputAt(2)));
+    case IrOpcode::kWord64Select:
+      return assembler.Word64Select(
+          Map(node->InputAt(0)), Map(node->InputAt(1)), Map(node->InputAt(2)));
+
     case IrOpcode::kLoad:
+    case IrOpcode::kLoadImmutable:
+    case IrOpcode::kProtectedLoad:
     case IrOpcode::kUnalignedLoad: {
       MemoryRepresentation loaded_rep =
           MemoryRepresentation::FromMachineType(LoadRepresentationOf(op));
       RegisterRepresentation result_rep = loaded_rep.ToRegisterRepresentation();
       Node* base = node->InputAt(0);
       Node* index = node->InputAt(1);
-      LoadOp::Kind kind = opcode == IrOpcode::kLoad
-                              ? LoadOp::Kind::kRawAligned
-                              : LoadOp::Kind::kRawUnaligned;
+      // It's ok to merge LoadImmutable into Load after scheduling.
+      LoadOp::Kind kind =
+          opcode == IrOpcode::kUnalignedLoad   ? LoadOp::Kind::kRawUnaligned
+          : opcode == IrOpcode::kProtectedLoad ? LoadOp::Kind::kProtected
+                                               : LoadOp::Kind::kRawAligned;
       if (index->opcode() == IrOpcode::kInt32Constant) {
         int32_t offset = OpParameter<int32_t>(index->op());
         return assembler.Load(Map(base), kind, loaded_rep, result_rep, offset);
@@ -559,14 +598,17 @@ OpIndex GraphBuilder::Process(
     }
 
     case IrOpcode::kStore:
-    case IrOpcode::kUnalignedStore: {
-      bool aligned = opcode == IrOpcode::kStore;
+    case IrOpcode::kUnalignedStore:
+    case IrOpcode::kProtectedStore: {
+      bool aligned = opcode != IrOpcode::kUnalignedStore;
       StoreRepresentation store_rep =
           aligned ? StoreRepresentationOf(op)
                   : StoreRepresentation(UnalignedStoreRepresentationOf(op),
                                         WriteBarrierKind::kNoWriteBarrier);
       StoreOp::Kind kind =
-          aligned ? StoreOp::Kind::kRawAligned : StoreOp::Kind::kRawUnaligned;
+          opcode == IrOpcode::kStore            ? StoreOp::Kind::kRawAligned
+          : opcode == IrOpcode::kUnalignedStore ? StoreOp::Kind::kRawAligned
+                                                : StoreOp::Kind::kProtected;
       Node* base = node->InputAt(0);
       Node* index = node->InputAt(1);
       Node* value = node->InputAt(2);
@@ -654,6 +696,21 @@ OpIndex GraphBuilder::Process(
       return call;
     }
 
+    case IrOpcode::kTailCall: {
+      auto call_descriptor = CallDescriptorOf(op);
+      base::SmallVector<OpIndex, 16> arguments;
+      // The input `0` is the callee, the following value inputs are the
+      // arguments. `CallDescriptor::InputCount()` counts the callee and
+      // arguments.
+      OpIndex callee = Map(node->InputAt(0));
+      for (int i = 1; i < static_cast<int>(call_descriptor->InputCount());
+           ++i) {
+        arguments.emplace_back(Map(node->InputAt(i)));
+      }
+      return assembler.TailCall(callee, base::VectorOf(arguments),
+                                call_descriptor);
+    }
+
     case IrOpcode::kFrameState: {
       FrameState frame_state{node};
       FrameStateData::Builder builder;
@@ -673,6 +730,13 @@ OpIndex GraphBuilder::Process(
                                     &DeoptimizeParametersOf(op));
     }
 
+    case IrOpcode::kTrapIf:
+    case IrOpcode::kTrapUnless: {
+      OpIndex condition = Map(node->InputAt(0));
+      bool negated = op->opcode() == IrOpcode::kTrapUnless;
+      return assembler.TrapIf(condition, negated, TrapIdOf(op));
+    }
+
     case IrOpcode::kDeoptimize: {
       OpIndex frame_state = Map(node->InputAt(0));
       return assembler.Deoptimize(frame_state, &DeoptimizeParametersOf(op));

src/compiler/turboshaft/operations.cc

@@ -8,9 +8,7 @@
 #include <sstream>
 
 #include "src/base/platform/mutex.h"
-#include "src/base/platform/platform.h"
 #include "src/codegen/machine-type.h"
-#include "src/common/assert-scope.h"
 #include "src/common/globals.h"
 #include "src/compiler/backend/instruction-selector.h"
 #include "src/compiler/frame-states.h"
@@ -49,6 +47,10 @@ std::ostream& operator<<(std::ostream& os, WordUnaryOp::Kind kind) {
       return os << "ReverseBytes";
     case WordUnaryOp::Kind::kCountLeadingZeros:
       return os << "CountLeadingZeros";
+    case WordUnaryOp::Kind::kCountTrailingZeros:
+      return os << "CountTrailingZeros";
+    case WordUnaryOp::Kind::kPopCount:
+      return os << "PopCount";
   }
 }
 
@@ -143,6 +145,23 @@ bool FloatUnaryOp::IsSupported(Kind kind, FloatRepresentation rep) {
   }
 }
 
+// static
+bool WordUnaryOp::IsSupported(Kind kind, WordRepresentation rep) {
+  switch (kind) {
+    case Kind::kCountLeadingZeros:
+    case Kind::kReverseBytes:
+      return true;
+    case Kind::kCountTrailingZeros:
+      return rep == WordRepresentation::Word32()
+                 ? SupportedOperations::word32_ctz()
+                 : SupportedOperations::word64_ctz();
+    case Kind::kPopCount:
+      return rep == WordRepresentation::Word32()
+                 ? SupportedOperations::word32_popcnt()
+                 : SupportedOperations::word64_popcnt();
+  }
+}
+
 std::ostream& operator<<(std::ostream& os, ShiftOp::Kind kind) {
   switch (kind) {
     case ShiftOp::Kind::kShiftRightArithmeticShiftOutZeros:
@@ -181,12 +200,16 @@ std::ostream& operator<<(std::ostream& os, ChangeOp::Kind kind) {
       return os << "UnsignedNarrowing";
     case ChangeOp::Kind::kFloatConversion:
       return os << "FloatConversion";
-    case ChangeOp::Kind::kSignedFloatTruncate:
-      return os << "SignedFloatTruncate";
     case ChangeOp::Kind::kJSFloatTruncate:
       return os << "JSFloatTruncate";
+    case ChangeOp::Kind::kSignedFloatTruncate:
+      return os << "SignedFloatTruncate";
     case ChangeOp::Kind::kSignedFloatTruncateOverflowToMin:
       return os << "SignedFloatTruncateOverflowToMin";
+    case ChangeOp::Kind::kUnsignedFloatTruncate:
+      return os << "UnsignedFloatTruncate";
+    case ChangeOp::Kind::kUnsignedFloatTruncateOverflowToMin:
+      return os << "UnsignedFloatTruncateOverflowToMin";
     case ChangeOp::Kind::kSignedToFloat:
       return os << "SignedToFloat";
     case ChangeOp::Kind::kUnsignedToFloat:
@@ -269,6 +292,14 @@ void ConstantOp::PrintOptions(std::ostream& os) const {
     case Kind::kCompressedHeapObject:
       os << "compressed heap object: " << handle();
       break;
+    case Kind::kRelocatableWasmCall:
+      os << "relocatable wasm call: 0x"
+         << reinterpret_cast<void*>(storage.integral);
+      break;
+    case Kind::kRelocatableWasmStubCall:
+      os << "relocatable wasm stub call: 0x"
+         << reinterpret_cast<void*>(storage.integral);
+      break;
   }
   os << "]";
 }
@@ -485,8 +516,8 @@ std::ostream& operator<<(std::ostream& os, OpProperties opProperties) {
     os << "Reading";
   } else if (opProperties == OpProperties::Writing()) {
     os << "Writing";
-  } else if (opProperties == OpProperties::CanDeopt()) {
-    os << "CanDeopt";
+  } else if (opProperties == OpProperties::CanAbort()) {
+    os << "CanAbort";
   } else if (opProperties == OpProperties::AnySideEffects()) {
     os << "AnySideEffects";
   } else if (opProperties == OpProperties::BlockTerminator()) {

src/compiler/turboshaft/optimization-phase.h

@@ -402,6 +402,11 @@ struct OptimizationPhase<Analyzer, Assembler>::Impl {
     auto arguments = MapToNewGraph<16>(op.arguments());
     return assembler.Call(callee, base::VectorOf(arguments), op.descriptor);
   }
+  OpIndex ReduceTailCall(const TailCallOp& op) {
+    OpIndex callee = MapToNewGraph(op.callee());
+    auto arguments = MapToNewGraph<16>(op.arguments());
+    return assembler.TailCall(callee, base::VectorOf(arguments), op.descriptor);
+  }
   OpIndex ReduceReturn(const ReturnOp& op) {
     // We very rarely have tuples longer than 4.
     auto return_values = MapToNewGraph<4>(op.return_values());
@@ -440,6 +445,11 @@ struct OptimizationPhase<Analyzer, Assembler>::Impl {
   OpIndex ReduceTaggedBitcast(const TaggedBitcastOp& op) {
     return assembler.TaggedBitcast(MapToNewGraph(op.input()), op.from, op.to);
   }
+  OpIndex ReduceSelect(const SelectOp& op) {
+    return assembler.Select(MapToNewGraph(op.condition()),
+                            MapToNewGraph(op.left()), MapToNewGraph(op.right()),
+                            op.rep);
+  }
   OpIndex ReduceConstant(const ConstantOp& op) {
     return assembler.Constant(op.kind, op.storage);
   }
@@ -493,6 +503,10 @@ struct OptimizationPhase<Analyzer, Assembler>::Impl {
                                   MapToNewGraph(op.frame_state()), op.negated,
                                   op.parameters);
   }
+  OpIndex ReduceTrapIf(const TrapIfOp& op) {
+    return assembler.TrapIf(MapToNewGraph(op.condition()), op.negated,
+                            op.trap_id);
+  }
   OpIndex ReduceTuple(const TupleOp& op) {
     return assembler.Tuple(base::VectorOf(MapToNewGraph<4>(op.inputs())));
   }

src/compiler/turboshaft/recreate-schedule.cc

@@ -88,6 +88,13 @@ struct ScheduleBuilder {
     return AddNode(machine.Is64() ? machine.Word64Shl() : machine.Word32Shl(),
                    {a, b});
   }
+  Node* RelocatableIntPtrConstant(intptr_t value, RelocInfo::Mode mode) {
+    return AddNode(machine.Is64()
+                       ? common.RelocatableInt64Constant(value, mode)
+                       : common.RelocatableInt32Constant(
+                             base::checked_cast<int32_t>(value), mode),
+                   {});
+  }
   void ProcessOperation(const Operation& op);
 #define DECL_PROCESS_OPERATION(Name) Node* ProcessOperation(const Name##Op& op);
   TURBOSHAFT_OPERATION_LIST(DECL_PROCESS_OPERATION)
@@ -363,6 +370,12 @@ Node* ScheduleBuilder::ProcessOperation(const WordUnaryOp& op) {
     case WordUnaryOp::Kind::kCountLeadingZeros:
       o = word64 ? machine.Word64Clz() : machine.Word32Clz();
       break;
+    case WordUnaryOp::Kind::kCountTrailingZeros:
+      o = word64 ? machine.Word64Ctz().op() : machine.Word32Ctz().op();
+      break;
+    case WordUnaryOp::Kind::kPopCount:
+      o = word64 ? machine.Word64Popcnt().op() : machine.Word32Popcnt().op();
+      break;
   }
   return AddNode(o, {GetNode(op.input())});
 }
@@ -628,6 +641,22 @@ Node* ScheduleBuilder::ProcessOperation(const ChangeOp& op) {
         UNIMPLEMENTED();
       }
       break;
+    case Kind::kUnsignedFloatTruncate:
+      if (op.from == FloatRepresentation::Float32() &&
+          op.to == WordRepresentation::Word32()) {
+        o = machine.TruncateFloat32ToUint32(TruncateKind::kArchitectureDefault);
+      } else {
+        UNIMPLEMENTED();
+      }
+      break;
+    case Kind::kUnsignedFloatTruncateOverflowToMin:
+      if (op.from == FloatRepresentation::Float32() &&
+          op.to == WordRepresentation::Word32()) {
+        o = machine.TruncateFloat32ToUint32(TruncateKind::kSetOverflowToMin);
+      } else {
+        UNIMPLEMENTED();
+      }
+      break;
     case Kind::kJSFloatTruncate:
       if (op.from == FloatRepresentation::Float64() &&
           op.to == WordRepresentation::Word32()) {
@@ -642,7 +671,16 @@ Node* ScheduleBuilder::ProcessOperation(const ChangeOp& op) {
         o = machine.ChangeInt32ToFloat64();
       } else if (op.from == WordRepresentation::Word64() &&
                  op.to == FloatRepresentation::Float64()) {
-        o = machine.ChangeInt64ToFloat64();
+        o = machine.RoundInt64ToFloat64();
+      } else if (op.from == WordRepresentation::Word32() &&
+                 op.to == FloatRepresentation::Float32()) {
+        o = machine.RoundInt32ToFloat32();
+      } else if (op.from == WordRepresentation::Word64() &&
+                 op.to == FloatRepresentation::Float32()) {
+        o = machine.RoundInt64ToFloat32();
+      } else if (op.from == WordRepresentation::Word32() &&
+                 op.to == FloatRepresentation::Float64()) {
+        o = machine.ChangeInt32ToFloat64();
       } else {
         UNIMPLEMENTED();
       }
@@ -651,6 +689,15 @@ Node* ScheduleBuilder::ProcessOperation(const ChangeOp& op) {
       if (op.from == WordRepresentation::Word32() &&
           op.to == FloatRepresentation::Float64()) {
         o = machine.ChangeUint32ToFloat64();
+      } else if (op.from == WordRepresentation::Word32() &&
+                 op.to == FloatRepresentation::Float32()) {
+        o = machine.RoundUint32ToFloat32();
+      } else if (op.from == WordRepresentation::Word64() &&
+                 op.to == FloatRepresentation::Float32()) {
+        o = machine.RoundUint64ToFloat32();
+      } else if (op.from == WordRepresentation::Word64() &&
+                 op.to == FloatRepresentation::Float64()) {
+        o = machine.RoundUint64ToFloat64();
       } else {
         UNIMPLEMENTED();
       }
@@ -708,6 +755,12 @@ Node* ScheduleBuilder::ProcessOperation(const ChangeOp& op) {
       } else if (op.from == FloatRepresentation::Float64() &&
                  op.to == WordRepresentation::Word32()) {
         o = machine.ChangeFloat64ToInt32();
+      } else if (op.from == WordRepresentation::Word32() &&
+                 op.to == FloatRepresentation::Float64()) {
+        o = machine.ChangeInt32ToFloat64();
+      } else if (op.from == WordRepresentation::Word64() &&
+                 op.to == FloatRepresentation::Float64()) {
+        o = machine.ChangeInt64ToFloat64();
       } else {
         UNIMPLEMENTED();
       }
@@ -749,6 +802,14 @@ Node* ScheduleBuilder::ProcessOperation(const TaggedBitcastOp& op) {
   }
   return AddNode(o, {GetNode(op.input())});
 }
+Node* ScheduleBuilder::ProcessOperation(const SelectOp& op) {
+  const Operator* o = op.rep == WordRepresentation::Word32()
+                          ? machine.Word32Select().op()
+                          : machine.Word64Select().op();
+  return AddNode(
+      o, {GetNode(op.condition()), GetNode(op.left()), GetNode(op.right())});
+}
+
 Node* ScheduleBuilder::ProcessOperation(const PendingLoopPhiOp& op) {
   UNREACHABLE();
 }
@@ -779,6 +840,11 @@ Node* ScheduleBuilder::ProcessOperation(const ConstantOp& op) {
       return AddNode(common.Float64Constant(op.float64()), {});
     case ConstantOp::Kind::kFloat32:
       return AddNode(common.Float32Constant(op.float32()), {});
+    case ConstantOp::Kind::kRelocatableWasmCall:
+      return RelocatableIntPtrConstant(op.integral(), RelocInfo::WASM_CALL);
+    case ConstantOp::Kind::kRelocatableWasmStubCall:
+      return RelocatableIntPtrConstant(op.integral(),
+                                       RelocInfo::WASM_STUB_CALL);
   }
 }
 Node* ScheduleBuilder::ProcessOperation(const LoadOp& op) {
@@ -788,9 +854,11 @@ Node* ScheduleBuilder::ProcessOperation(const LoadOp& op) {
     offset -= kHeapObjectTag;
   }
   Node* base = GetNode(op.base());
-  return AddNode(IsAlignedAccess(op.kind)
+  return AddNode(op.kind == LoadOp::Kind::kRawAligned
                      ? machine.Load(op.loaded_rep.ToMachineType())
-                     : machine.UnalignedLoad(op.loaded_rep.ToMachineType()),
+                 : op.kind == LoadOp::Kind::kRawUnaligned
+                     ? machine.UnalignedLoad(op.loaded_rep.ToMachineType())
+                     : machine.ProtectedLoad(op.loaded_rep.ToMachineType()),
                  {base, IntPtrConstant(offset)});
 }
 Node* ScheduleBuilder::ProcessOperation(const IndexedLoadOp& op) {
@@ -815,8 +883,10 @@ Node* ScheduleBuilder::ProcessOperation(const IndexedLoadOp& op) {
       loaded_rep = MachineType::CompressedPointer();
     }
   }
-  return AddNode(IsAlignedAccess(op.kind) ? machine.Load(loaded_rep)
-                                          : machine.UnalignedLoad(loaded_rep),
+  return AddNode(op.kind == LoadOp::Kind::kRawAligned ? machine.Load(loaded_rep)
+                 : op.kind == LoadOp::Kind::kRawUnaligned
+                     ? machine.UnalignedLoad(loaded_rep)
+                     : machine.ProtectedLoad(loaded_rep),
                  {base, index});
 }
 Node* ScheduleBuilder::ProcessOperation(const StoreOp& op) {
@@ -831,6 +901,9 @@ Node* ScheduleBuilder::ProcessOperation(const StoreOp& op) {
   if (IsAlignedAccess(op.kind)) {
     o = machine.Store(StoreRepresentation(
         op.stored_rep.ToMachineType().representation(), op.write_barrier));
+  } else if (op.kind == LoadOp::Kind::kProtected) {
+    DCHECK_EQ(op.write_barrier, WriteBarrierKind::kNoWriteBarrier);
+    o = machine.ProtectedStore(op.stored_rep.ToMachineType().representation());
   } else {
     DCHECK_EQ(op.write_barrier, WriteBarrierKind::kNoWriteBarrier);
     o = machine.UnalignedStore(op.stored_rep.ToMachineType().representation());
@@ -856,6 +929,9 @@ Node* ScheduleBuilder::ProcessOperation(const IndexedStoreOp& op) {
   if (IsAlignedAccess(op.kind)) {
     o = machine.Store(StoreRepresentation(
         op.stored_rep.ToMachineType().representation(), op.write_barrier));
+  } else if (op.kind == LoadOp::Kind::kProtected) {
+    DCHECK_EQ(op.write_barrier, WriteBarrierKind::kNoWriteBarrier);
+    o = machine.ProtectedStore(op.stored_rep.ToMachineType().representation());
   } else {
     DCHECK_EQ(op.write_barrier, WriteBarrierKind::kNoWriteBarrier);
     o = machine.UnalignedStore(op.stored_rep.ToMachineType().representation());
@@ -927,6 +1003,12 @@ Node* ScheduleBuilder::ProcessOperation(const DeoptimizeIfOp& op) {
                                                 op.parameters->feedback());
   return AddNode(o, {condition, frame_state});
 }
+Node* ScheduleBuilder::ProcessOperation(const TrapIfOp& op) {
+  Node* condition = GetNode(op.condition());
+  const Operator* o =
+      op.negated ? common.TrapUnless(op.trap_id) : common.TrapIf(op.trap_id);
+  return AddNode(o, {condition});
+}
 Node* ScheduleBuilder::ProcessOperation(const DeoptimizeOp& op) {
   Node* frame_state = GetNode(op.frame_state());
   const Operator* o =
@@ -1087,6 +1169,17 @@ Node* ScheduleBuilder::ProcessOperation(const CallOp& op) {
   }
   return AddNode(common.Call(op.descriptor), base::VectorOf(inputs));
 }
+Node* ScheduleBuilder::ProcessOperation(const TailCallOp& op) {
+  base::SmallVector<Node*, 16> inputs;
+  inputs.push_back(GetNode(op.callee()));
+  for (OpIndex i : op.arguments()) {
+    inputs.push_back(GetNode(i));
+  }
+  Node* call = MakeNode(common.TailCall(op.descriptor), base::VectorOf(inputs));
+  schedule->AddTailCall(current_block, call);
+  current_block = nullptr;
+  return nullptr;
+}
 Node* ScheduleBuilder::ProcessOperation(const UnreachableOp& op) {
   Node* node = MakeNode(common.Throw(), {});
   schedule->AddThrow(current_block, node);

src/compiler/turboshaft/recreate-schedule.h

@@ -7,7 +7,7 @@
 
 #include "src/compiler/compiler-source-position-table.h"
 #include "src/compiler/node-origin-table.h"
-#include "src/compiler/node.h"
+
 namespace v8::internal {
 class Zone;
 }

src/compiler/turboshaft/simplify-tf-loops.cc

+// 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.
+
+#include "src/compiler/turboshaft/simplify-tf-loops.h"
+
+#include "src/base/small-vector.h"
+#include "src/compiler/machine-graph.h"
+#include "src/compiler/node-properties.h"
+
+namespace v8::internal::compiler {
+
+Reduction SimplifyTFLoops::Reduce(Node* node) {
+  if (node->opcode() != IrOpcode::kLoop) return NoChange();
+  if (node->InputCount() <= 2) return NoChange();
+
+  Node* new_loop = mcgraph_->graph()->NewNode(mcgraph_->common()->Loop(2),
+                                              node->InputAt(0), node);
+  node->RemoveInput(0);
+  NodeProperties::ChangeOp(node, mcgraph_->common()->Merge(node->InputCount()));
+
+  base::SmallVector<Edge, 4> control_uses;
+
+  for (Edge edge : node->use_edges()) {
+    Node* use = edge.from();
+    if (!NodeProperties::IsPhi(use)) {
+      control_uses.emplace_back(edge);
+      continue;
+    }
+    Node* dominating_input = use->InputAt(0);
+    use->RemoveInput(0);
+    NodeProperties::ChangeOp(
+        use, use->opcode() == IrOpcode::kPhi
+                 ? mcgraph_->common()->Phi(PhiRepresentationOf(use->op()),
+                                           use->InputCount() - 1)
+                 : mcgraph_->common()->EffectPhi(use->InputCount() - 1));
+
+    Node* new_phi = mcgraph_->graph()->NewNode(
+        use->opcode() == IrOpcode::kPhi
+            ? mcgraph_->common()->Phi(PhiRepresentationOf(use->op()), 2)
+            : mcgraph_->common()->EffectPhi(2),
+        dominating_input, use, new_loop);
+
+    ReplaceWithValue(use, new_phi, new_phi, new_phi);
+    // Restore the use <- new_phi edge we just broke.
+    new_phi->ReplaceInput(1, use);
+  }
+
+  for (Edge edge : control_uses) {
+    if (edge.from() != new_loop) {
+      edge.from()->ReplaceInput(edge.index(), new_loop);
+    }
+  }
+
+  return NoChange();
+}
+
+}  // namespace v8::internal::compiler

src/compiler/turboshaft/simplify-tf-loops.h

+// 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.
+
+#ifndef V8_COMPILER_TURBOSHAFT_SIMPLIFY_TF_LOOPS_H_
+#define V8_COMPILER_TURBOSHAFT_SIMPLIFY_TF_LOOPS_H_
+
+#include "src/compiler/graph-reducer.h"
+
+namespace v8::internal::compiler {
+
+class MachineGraph;
+
+// Constrain loop nodes to have at most two inputs, by introducing additional
+// merges as needed.
+class SimplifyTFLoops final : public AdvancedReducer {
+ public:
+  SimplifyTFLoops(Editor* editor, MachineGraph* mcgraph)
+      : AdvancedReducer(editor), mcgraph_(mcgraph) {}
+
+  const char* reducer_name() const override { return "SimplifyTFLoops"; }
+
+  Reduction Reduce(Node* node) final;
+
+ private:
+  MachineGraph* const mcgraph_;
+};
+
+}  // namespace v8::internal::compiler
+
+#endif  // V8_COMPILER_TURBOSHAFT_SIMPLIFY_TF_LOOPS_H_

src/flags/flag-definitions.h

@@ -963,10 +963,12 @@ DEFINE_FLOAT(script_delay_fraction, 0.0,
              "busy wait after each Script::Run by the given fraction of the "
              "run's duration")
 
-DEFINE_BOOL(turboshaft, false, "enable TurboFan's Turboshaft phases")
+DEFINE_BOOL(turboshaft, false, "enable TurboFan's Turboshaft phases for JS")
 DEFINE_WEAK_IMPLICATION(future, turboshaft)
 DEFINE_BOOL(turboshaft_trace_reduction, false,
             "trace individual Turboshaft reduction steps")
+DEFINE_BOOL(turboshaft_wasm, false,
+            "enable TurboFan's Turboshaft phases for wasm")
 
 // Favor memory over execution speed.
 DEFINE_BOOL(optimize_for_size, false,

src/logging/runtime-call-stats.h

@@ -366,6 +366,7 @@ class RuntimeCallTimer final {
   ADD_THREAD_SPECIFIC_COUNTER(V, Optimize, Scheduling)                      \
   ADD_THREAD_SPECIFIC_COUNTER(V, Optimize, SelectInstructions)              \
   ADD_THREAD_SPECIFIC_COUNTER(V, Optimize, SimplifiedLowering)              \
+  ADD_THREAD_SPECIFIC_COUNTER(V, Optimize, SimplifyLoops)                   \
   ADD_THREAD_SPECIFIC_COUNTER(V, Optimize, StoreStoreElimination)           \
   ADD_THREAD_SPECIFIC_COUNTER(V, Optimize, TraceScheduleAndVerify)          \
   ADD_THREAD_SPECIFIC_COUNTER(V, Optimize, BuildTurboshaft)                 \