Implement optional turbofan UnalignedLoad and UnalignedStore operators

src/compiler/wasm-compiler.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/compiler/wasm-compiler.h"
 
 #include "src/isolate-inl.h"
 
 #include "src/base/platform/elapsed-timer.h"
 #include "src/base/platform/platform.h"
@@ skipping 2362 matching lines @@
     size_t limit = size - offset - memsize;
     CHECK(limit <= kMaxUInt32);
     cond = graph()->NewNode(
         jsgraph()->machine()->Uint32LessThanOrEqual(), index,
         jsgraph()->Int32Constant(static_cast<uint32_t>(limit)));
   }
 
   trap_->AddTrapIfFalse(wasm::kTrapMemOutOfBounds, cond);
 }
 
+Node* WasmGraphBuilder::BuildUnalignedLoad(wasm::LocalType type,
+                                           MachineType memtype, Node* index,
+                                           uint32_t offset) {
+  Node* result;
+  Node* load;
+  bool extendTo64Bit = false;
+
+  wasm::WasmOpcode shiftOpcode;
+  wasm::WasmOpcode orOpcode;
+  Node* constEight;
+
+  switch (type) {
+    case wasm::kAstI64:
+    case wasm::kAstF64:
+      shiftOpcode = wasm::kExprI64Shl;
+      orOpcode = wasm::kExprI64Ior;
+      result = jsgraph()->Int64Constant(0);
+      constEight = jsgraph()->Int64Constant(8);
+      extendTo64Bit = true;
+      break;
+    case wasm::kAstI32:
+    case wasm::kAstF32:
+      shiftOpcode = wasm::kExprI32Shl;
+      orOpcode = wasm::kExprI32Ior;
+      result = jsgraph()->Int32Constant(0);
+      constEight = jsgraph()->Int32Constant(8);
+      break;
+    default:
+      UNREACHABLE();
+  }
+
+  bool signExtend = memtype.IsSigned();
+
+  bool isFloat = IsFloatingPoint(memtype.representation());
+  int numberOfLoads = 1 << ElementSizeLog2Of(memtype.representation());
+
+  for (int i = 0; i < numberOfLoads; i++) {

[titzer, 2016/04/20 12:59:26]

This is good enough for a first cut, since it does byte-by-byte loads, but in
the future we can also use the alignment to generate halfword-by-halfword
loads/stores.

+    result = Binop(shiftOpcode, result, constEight);
+    load = graph()->NewNode(
+        jsgraph()->machine()->Load(signExtend ? MachineType::Int8()
+                                              : MachineType::Uint8()),
+        MemBuffer(offset + numberOfLoads - 1 - i), index, *effect_, *control_);
+    index->Print();
+    *effect_ = load;
+    if (extendTo64Bit) {
+      if (signExtend) {
+        load =
+            graph()->NewNode(jsgraph()->machine()->ChangeInt32ToInt64(), load);
+      } else {
+        load = graph()->NewNode(jsgraph()->machine()->ChangeUint32ToUint64(),
+                                load);
+      }
+    }
+    signExtend = false;
+    result = Binop(orOpcode, result, load);
+  }
+
+  // Convert to float
+  if (isFloat) {
+    switch (type) {
+      case wasm::kAstF32:
+        result = Unop(wasm::kExprF32ReinterpretI32, result);
+        break;
+      case wasm::kAstF64:
+        result = Unop(wasm::kExprF64ReinterpretI64, result);
+        break;
+      default:
+        UNREACHABLE();
+    }
+  }
+
+  return result;
+}
 
 Node* WasmGraphBuilder::LoadMem(wasm::LocalType type, MachineType memtype,
-                                Node* index, uint32_t offset) {
+                                Node* index, uint32_t offset,
+                                uint32_t alignment) {
   Node* load;
+  bool skip = false;
 
   if (module_ && module_->asm_js()) {
     // asm.js semantics use CheckedLoad (i.e. OOB reads return 0ish).
     DCHECK_EQ(0, offset);
     const Operator* op = jsgraph()->machine()->CheckedLoad(memtype);
     load = graph()->NewNode(op, MemBuffer(0), index, MemSize(0), *effect_,
                             *control_);
   } else {
     // WASM semantics throw on OOB. Introduce explicit bounds check.
     BoundsCheckMem(memtype, index, offset);
-    load = graph()->NewNode(jsgraph()->machine()->Load(memtype),
-                            MemBuffer(offset), index, *effect_, *control_);
+    // load =
+    // graph()->NewNode(jsgraph()->machine()->UnalignedLoad(memtype).op(),
+    //                         MemBuffer(offset), index, *effect_, *control_);
+    bool aligned = alignment >= ElementSizeLog2Of(memtype.representation());

[titzer, 2016/04/20 12:59:26]

It's probably best to have a
machine()->UnalignedLoadSupported(MachineRepresentation, int alignment)

which is backed by appropriate flags.

+    if (aligned ||
+        !jsgraph()->machine()->UnalignedLoad(memtype).IsSupported()) {
+      load = graph()->NewNode(jsgraph()->machine()->Load(memtype),
+                              MemBuffer(offset), index, *effect_, *control_);
+    } else {
+      load = BuildUnalignedLoad(type, memtype, index, offset);
+      skip = true;
+    }
   }
 
-  *effect_ = load;
+  if (!skip) {
+    *effect_ = load;
+  }
 
   if (type == wasm::kAstI64 &&
       ElementSizeLog2Of(memtype.representation()) < 3) {
     // TODO(titzer): TF zeroes the upper bits of 64-bit loads for subword sizes.
     if (memtype.IsSigned()) {
       // sign extend
       load = graph()->NewNode(jsgraph()->machine()->ChangeInt32ToInt64(), load);
     } else {
       // zero extend
       load =
           graph()->NewNode(jsgraph()->machine()->ChangeUint32ToUint64(), load);
     }
   }
 
   return load;
 }
 
+Node* WasmGraphBuilder::BuildUnalignedStore(MachineType memtype, Node* index,
+                                            uint32_t offset, uint32_t alignment,
+                                            Node* val) {
+  Node* store;
+  Node* newValue;
+
+  wasm::WasmOpcode shiftOpcode;
+
+  Node* constEight;
+  Node* mask;
+  bool extendTo64Bit = false;
+
+  if (ElementSizeLog2Of(memtype.representation()) <= 2) {
+    shiftOpcode = wasm::kExprI32ShrU;
+    constEight = jsgraph()->Int32Constant(8);
+    mask = jsgraph()->Int32Constant(0x000000ff);
+  } else {
+    shiftOpcode = wasm::kExprI64ShrU;
+    constEight = jsgraph()->Int64Constant(8);
+    mask = jsgraph()->Int64Constant(0x00000000000000ff);
+    extendTo64Bit = true;
+  }
+
+  bool isFloat = IsFloatingPoint(memtype.representation());
+  int numberOfLoads = 1 << ElementSizeLog2Of(memtype.representation());
+
+  StoreRepresentation rep(MachineType::Uint8().representation(),
+                          kNoWriteBarrier);
+
+  newValue = val;
+  if (isFloat) {
+    switch (memtype.representation()) {
+      case MachineRepresentation::kFloat64:
+        newValue = Unop(wasm::kExprI64ReinterpretF64, val);
+        break;
+      case MachineRepresentation::kFloat32:
+        newValue = Unop(wasm::kExprI32ReinterpretF32, val);
+        break;
+      default:
+        UNREACHABLE();
+    }
+  }
+
+  DCHECK(numberOfLoads >= 2);
+  for (int i = 0; i < numberOfLoads - 1; i++) {
+    store = graph()->NewNode(
+        jsgraph()->machine()->Store(rep), MemBuffer(offset + i), index,
+        extendTo64Bit ? Unop(wasm::kExprI32ConvertI64, newValue) : newValue,
+        *effect_, *control_);
+    newValue = Binop(shiftOpcode, newValue, constEight);
+    *effect_ = store;
+  }
+  store = graph()->NewNode(
+      jsgraph()->machine()->Store(rep), MemBuffer(offset + numberOfLoads - 1),
+      index,
+      extendTo64Bit ? Unop(wasm::kExprI32ConvertI64, newValue) : newValue,
+      *effect_, *control_);
+  *effect_ = store;
+  return val;
+}
 
 Node* WasmGraphBuilder::StoreMem(MachineType memtype, Node* index,
-                                 uint32_t offset, Node* val) {
+                                 uint32_t offset, uint32_t alignment,
+                                 Node* val) {
   Node* store;
+  bool skip = false;
   if (module_ && module_->asm_js()) {
     // asm.js semantics use CheckedStore (i.e. ignore OOB writes).
     DCHECK_EQ(0, offset);
     const Operator* op =
         jsgraph()->machine()->CheckedStore(memtype.representation());
     store = graph()->NewNode(op, MemBuffer(0), index, MemSize(0), val, *effect_,
                              *control_);
   } else {
     // WASM semantics throw on OOB. Introduce explicit bounds check.
     BoundsCheckMem(memtype, index, offset);
     StoreRepresentation rep(memtype.representation(), kNoWriteBarrier);
-    store =
-        graph()->NewNode(jsgraph()->machine()->Store(rep), MemBuffer(offset),
-                         index, val, *effect_, *control_);
+    bool aligned = alignment >= ElementSizeLog2Of(memtype.representation());
+
+    if (aligned ||
+        !jsgraph()
+             ->machine()
+             ->UnalignedStore(memtype.representation())
+             .IsSupported()) {
+      StoreRepresentation rep(memtype.representation(), kNoWriteBarrier);
+      store =
+          graph()->NewNode(jsgraph()->machine()->Store(rep), MemBuffer(offset),
+                           index, val, *effect_, *control_);
+    } else {
+      store = BuildUnalignedStore(memtype, index, offset, alignment, val);
+      skip = true;
+    }
   }
-  *effect_ = store;
+  if (!skip) {
+    *effect_ = store;
+  }
+
   return store;
 }
 
 
 void WasmGraphBuilder::PrintDebugName(Node* node) {
   PrintF("#%d:%s", node->id(), node->op()->mnemonic());
 }
 
 
 Node* WasmGraphBuilder::String(const char* string) {
@@ skipping 331 matching lines @@
         static_cast<int>(function.code_end_offset - function.code_start_offset),
         decode_ms, static_cast<int>(graph.NodeCount()), compile_ms);
   }
   return code;
 }
 
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8