[multibyte] Add multibyte array load instructions. (#8504)

Prototype implementation of the new multibyte array proposal that reuses
the existing memory instructions.


https://github.com/WebAssembly/multibyte-array-access/blob/944d79230b336442b1f0b3978da1908d54b9f230/proposals/multibyte-array-access/Overview.md

Author: brendandahl

src/interpreter/interpreter.cpp

@@ -262,6 +262,7 @@ struct ExpressionInterpreter : OverriddenVisitor<ExpressionInterpreter, Flow> {
   Flow visitArrayNewFixed(ArrayNewFixed* curr) { WASM_UNREACHABLE("TODO"); }
   Flow visitArrayGet(ArrayGet* curr) { WASM_UNREACHABLE("TODO"); }
   Flow visitArraySet(ArraySet* curr) { WASM_UNREACHABLE("TODO"); }
+  Flow visitArrayLoad(ArrayLoad* curr) { WASM_UNREACHABLE("TODO"); }
   Flow visitArrayStore(ArrayStore* curr) { WASM_UNREACHABLE("TODO"); }
   Flow visitArrayLen(ArrayLen* curr) { WASM_UNREACHABLE("TODO"); }
   Flow visitArrayCopy(ArrayCopy* curr) { WASM_UNREACHABLE("TODO"); }

src/ir/ReFinalize.cpp

@@ -171,6 +171,7 @@ void ReFinalize::visitArrayNewElem(ArrayNewElem* curr) { curr->finalize(); }
 void ReFinalize::visitArrayNewFixed(ArrayNewFixed* curr) { curr->finalize(); }
 void ReFinalize::visitArrayGet(ArrayGet* curr) { curr->finalize(); }
 void ReFinalize::visitArraySet(ArraySet* curr) { curr->finalize(); }
+void ReFinalize::visitArrayLoad(ArrayLoad* curr) { curr->finalize(); }
 void ReFinalize::visitArrayStore(ArrayStore* curr) { curr->finalize(); }
 void ReFinalize::visitArrayLen(ArrayLen* curr) { curr->finalize(); }
 void ReFinalize::visitArrayCopy(ArrayCopy* curr) { curr->finalize(); }

src/ir/child-typer.h

@@ -1096,6 +1096,19 @@ template<typename Subtype> struct ChildTyper : OverriddenVisitor<Subtype> {
     note(&curr->value, type);
   }
 
+  void visitArrayLoad(ArrayLoad* curr,
+                      std::optional<HeapType> ht = std::nullopt) {
+    if (!ht) {
+      if (!curr->ref->type.isRef()) {
+        self().noteUnknown();
+        return;
+      }
+      ht = curr->ref->type.getHeapType();
+    }
+    note(&curr->ref, Type(*ht, Nullable));
+    note(&curr->index, Type::i32);
+  }
+
   void visitArrayStore(ArrayStore* curr,
                        std::optional<HeapType> ht = std::nullopt,
                        std::optional<Type> valueType = std::nullopt) {

src/ir/cost.h

@@ -753,6 +753,9 @@ struct CostAnalyzer : public OverriddenVisitor<CostAnalyzer, CostType> {
     return 2 + nullCheckCost(curr->ref) + visit(curr->ref) +
            visit(curr->index) + visit(curr->value);
   }
+  CostType visitArrayLoad(ArrayLoad* curr) {
+    return 1 + nullCheckCost(curr->ref) + visit(curr->ref) + visit(curr->index);
+  }
   CostType visitArrayStore(ArrayStore* curr) {
     return 2 + nullCheckCost(curr->ref) + visit(curr->ref) +
            visit(curr->index) + visit(curr->value);

src/ir/effects.h

@@ -1147,6 +1147,15 @@ class EffectAnalyzer {
       parent.implicitTrap = true;
       writesArray(curr->ref->type.getHeapType(), curr->order);
     }
+    void visitArrayLoad(ArrayLoad* curr) {
+      if (trapOnNull(curr->ref)) {
+        return;
+      }
+      // Null refs and OOB access.
+      parent.implicitTrap = true;
+      readsArray(curr->ref->type.getHeapType(), MemoryOrder::Unordered);
+    }
+
     void visitArrayStore(ArrayStore* curr) {
       if (curr->ref->type.isNull()) {
         parent.trap = true;

src/ir/possible-contents.cpp

@@ -1105,6 +1105,13 @@ struct InfoCollector
     addChildParentLink(curr->ref, curr);
     addChildParentLink(curr->value, curr);
   }
+  void visitArrayLoad(ArrayLoad* curr) {
+    if (!isRelevant(curr->ref)) {
+      addRoot(curr);
+      return;
+    }
+    addChildParentLink(curr->ref, curr);
+  }
   void visitArrayStore(ArrayStore* curr) {
     if (curr->ref->type == Type::unreachable) {
       return;
@@ -1755,6 +1762,7 @@ void TNHOracle::scan(Function* func,
     }
     void visitArrayGet(ArrayGet* curr) { notePossibleTrap(curr->ref); }
     void visitArraySet(ArraySet* curr) { notePossibleTrap(curr->ref); }
+    void visitArrayLoad(ArrayLoad* curr) { notePossibleTrap(curr->ref); }
     void visitArrayStore(ArrayStore* curr) { notePossibleTrap(curr->ref); }
     void visitArrayLen(ArrayLen* curr) { notePossibleTrap(curr->ref); }
     void visitArrayCopy(ArrayCopy* curr) {
@@ -2865,6 +2873,9 @@ void Flower::flowAfterUpdate(LocationIndex locationIndex) {
     } else if (auto* set = parent->dynCast<ArraySet>()) {
       assert(set->ref == child || set->value == child);
       writeToData(set->ref, set->value, 0);
+    } else if (auto* load = parent->dynCast<ArrayLoad>()) {
+      assert(load->ref == child);
+      readFromData(load->ref->type, 0, contents, load);
     } else if (auto* store = parent->dynCast<ArrayStore>()) {
       assert(store->ref == child || store->value == child);
       // TODO: model the stored value, and handle different but equal values in
@@ -3108,15 +3119,25 @@ void Flower::filterPackedDataReads(PossibleContents& contents,
   auto signed_ = false;
   Expression* ref;
   Index index;
+  unsigned bytes = 0;
+  Type resultType = Type::none;
   if (auto* get = expr->dynCast<StructGet>()) {
     signed_ = get->signed_;
     ref = get->ref;
     index = get->index;
+    resultType = get->type;
   } else if (auto* get = expr->dynCast<ArrayGet>()) {
     signed_ = get->signed_;
     ref = get->ref;
     // Arrays are treated as having a single field.
     index = 0;
+    resultType = get->type;
+  } else if (auto* load = expr->dynCast<ArrayLoad>()) {
+    signed_ = load->signed_;
+    ref = load->ref;
+    index = 0;
+    bytes = load->bytes;
+    resultType = load->type;
   } else {
     WASM_UNREACHABLE("bad packed read");
   }
@@ -3135,13 +3156,21 @@ void Flower::filterPackedDataReads(PossibleContents& contents,
   assert(ref->type.isRef());
   auto field = GCTypeUtils::getField(ref->type.getHeapType(), index);
   assert(field);
-  if (!field->isPacked()) {
-    return;
+  if (!bytes) {
+    if (!field->isPacked()) {
+      return;
+    }
+    bytes = field->getByteSize();
   }
 
   if (contents.isLiteral()) {
     // This is a constant. We can sign-extend it and use that value.
-    auto shifts = Literal(int32_t(32 - field->getByteSize() * 8));
+    unsigned bits = resultType.getByteSize() == 8 ? 64 : 32;
+    if (bits <= bytes * 8) {
+      // No need to sign-extend for full size.
+      return;
+    }
+    auto shifts = Literal(int32_t(bits - bytes * 8));
     auto lit = contents.getLiteral();
     lit = lit.shl(shifts);
     lit = lit.shrS(shifts);

src/ir/subtype-exprs.h

@@ -412,6 +412,7 @@ struct SubtypingDiscoverer : public OverriddenVisitor<SubType> {
     auto array = curr->ref->type.getHeapType().getArray();
     self()->noteSubtype(curr->value, array.element.type);
   }
+  void visitArrayLoad(ArrayLoad* curr) {}
   void visitArrayStore(ArrayStore* curr) {}
   void visitArrayLen(ArrayLen* curr) {}
   void visitArrayCopy(ArrayCopy* curr) {

src/parser/contexts.h

@@ -570,6 +570,11 @@ struct NullInstrParserCtx {
     return Ok{};
   }
   template<typename HeapTypeT>
+  Result<> makeArrayLoad(
+    Index, const std::vector<Annotation>&, Type, int, bool, HeapTypeT) {
+    return Ok{};
+  }
+  template<typename HeapTypeT>
   Result<>
   makeArrayStore(Index, const std::vector<Annotation>&, Type, int, HeapTypeT) {
     return Ok{};
@@ -2331,6 +2336,16 @@ struct ParseDefsCtx : TypeParserCtx<ParseDefsCtx>, AnnotationParserCtx {
       pos, irBuilder.makeStore(bytes, memarg.offset, memarg.align, type, *m));
   }
 
+  Result<> makeArrayLoad(Index pos,
+                         const std::vector<Annotation>& annotations,
+                         Type type,
+                         int bytes,
+                         bool signed_,
+                         HeapTypeT arrayType) {
+    return withLoc(pos,
+                   irBuilder.makeArrayLoad(arrayType, bytes, signed_, type));
+  }
+
   Result<> makeArrayStore(Index pos,
                           const std::vector<Annotation>& annotations,
                           Type type,

src/parser/parsers.h

@@ -1783,6 +1783,18 @@ Result<> makeLoad(Ctx& ctx,
                   int bytes,
                   bool isAtomic) {
 
+  if (ctx.in.takeSExprStart("type"sv)) {
+    auto arrayType = typeidx(ctx);
+    CHECK_ERR(arrayType);
+
+    if (!ctx.in.takeRParen()) {
+      return ctx.in.err("expected end of type use");
+    }
+
+    return ctx.makeArrayLoad(
+      pos, annotations, type, bytes, signed_, *arrayType);
+  }
+
   auto mem = maybeMemidx(ctx);
   CHECK_ERR(mem);
 

src/passes/Precompute.cpp

@@ -221,6 +221,10 @@ class PrecomputingExpressionRunner
   }
   // ArrayLen is not disallowed here as it is an immutable property.
   Flow visitArrayCopy(ArrayCopy* curr) { return Flow(NONCONSTANT_FLOW); }
+  Flow visitArrayLoad(ArrayLoad* curr) {
+    // TODO: We could optimize loads from immutable data, like ArrayGet.
+    return Flow(NONCONSTANT_FLOW);
+  }
   Flow visitArrayStore(ArrayStore* curr) { return Flow(NONCONSTANT_FLOW); }
 
   // Generates heap info for a heap-allocating expression.