[Custom Descriptors] Interpret ref.get_desc (#7675)

Add a descriptor to the GCData stored in a reference value. The
descriptor itself is another literal representing either a null if there
is no descriptor or otherwise a reference to the descriptor. Update
the interpretation of struct.new to set the descriptor when allocating
types with descriptors and then retrieve the descriptor when
interpreting ref.get_desc.

Author: tlively

src/literal.h

@@ -775,8 +775,13 @@ struct GCData {
   // The element or field values.
   Literals values;
 
-  GCData(HeapType type, Literals&& values)
-    : type(type), values(std::move(values)) {}
+  // The descriptor, if it exists, or null.
+  Literal desc;
+
+  GCData(HeapType type,
+         Literals&& values,
+         const Literal& desc = Literal::makeNull(HeapType::none))
+    : type(type), values(std::move(values)), desc(desc) {}
 };
 
 // The data of a (ref exn) literal.

src/wasm-interpreter.h

@@ -199,9 +199,11 @@ class ExpressionRunner : public OverriddenVisitor<SubType, Flow> {
   // this function in LSan.
   //
   // This consumes the input |data| entirely.
-  Literal makeGCData(Literals&& data, Type type) {
+  Literal makeGCData(Literals&& data,
+                     Type type,
+                     Literal desc = Literal::makeNull(HeapType::none)) {
     auto allocation =
-      std::make_shared<GCData>(type.getHeapType(), std::move(data));
+      std::make_shared<GCData>(type.getHeapType(), std::move(data), desc);
 #if __has_feature(leak_sanitizer) || __has_feature(address_sanitizer)
     // GC data with cycles will leak, since shared_ptrs do not handle cycles.
     // Binaryen is generally not used in long-running programs so we just ignore
@@ -1673,7 +1675,15 @@ class ExpressionRunner : public OverriddenVisitor<SubType, Flow> {
   }
   Flow visitRefGetDesc(RefGetDesc* curr) {
     NOTE_ENTER("RefGetDesc");
-    WASM_UNREACHABLE("unimplemented");
+    Flow ref = self()->visit(curr->ref);
+    if (ref.breaking()) {
+      return ref;
+    }
+    auto data = ref.getSingleValue().getGCData();
+    if (!data) {
+      trap("null ref");
+    }
+    return data->desc;
   }
   Flow visitBrOn(BrOn* curr) {
     NOTE_ENTER("BrOn");
@@ -1742,6 +1752,12 @@ class ExpressionRunner : public OverriddenVisitor<SubType, Flow> {
           return value;
         }
       }
+      if (curr->descriptor) {
+        auto value = self()->visit(curr->descriptor);
+        if (value.breaking()) {
+          return value;
+        }
+      }
       WASM_UNREACHABLE("unreachable but no unreachable child");
     }
     auto heapType = curr->type.getHeapType();
@@ -1759,7 +1775,14 @@ class ExpressionRunner : public OverriddenVisitor<SubType, Flow> {
         data[i] = truncateForPacking(value.getSingleValue(), field);
       }
     }
-    return makeGCData(std::move(data), curr->type);
+    if (!curr->descriptor) {
+      return makeGCData(std::move(data), curr->type);
+    }
+    auto desc = self()->visit(curr->descriptor);
+    if (desc.breaking()) {
+      return desc;
+    }
+    return makeGCData(std::move(data), curr->type, desc.getSingleValue());
   }
   Flow visitStructGet(StructGet* curr) {
     NOTE_ENTER("StructGet");

src/wasm/wasm.cpp

@@ -1216,6 +1216,9 @@ void StructNew::finalize() {
   if (handleUnreachableOperands(this)) {
     return;
   }
+  if (descriptor && descriptor->type == Type::unreachable) {
+    type = Type::unreachable;
+  }
 }
 
 void StructGet::finalize() {

test/spec/ref.get_desc.wast

@@ -3,6 +3,12 @@
     (type $struct (descriptor $desc (struct)))
     (type $desc (describes $struct (struct)))
   )
+
+  (global $desc1 (ref (exact $desc)) (struct.new $desc))
+  (global $desc2 (ref (exact $desc)) (struct.new $desc))
+  (global $struct1 (ref $struct) (struct.new $struct (global.get $desc1)))
+  (global $struct2 (ref $struct) (struct.new $struct (global.get $desc2)))
+
   (func $unreachable (param $struct (ref null $struct)) (result (ref $desc))
     (ref.get_desc $struct
       (unreachable)
@@ -23,8 +29,25 @@
       (local.get $struct)
     )
   )
+
+  (func (export "check-descs") (result i32)
+    (i32.and
+      (i32.and
+        (ref.eq (ref.get_desc $struct (global.get $struct1)) (global.get $desc1))
+        (ref.eq (ref.get_desc $struct (global.get $struct2)) (global.get $desc2))
+      )
+      (i32.eqz (ref.eq (global.get $desc1) (global.get $desc2)))
+    )
+  )
+
+  (func (export "desc-trap") (result (ref $struct))
+    (struct.new $struct (unreachable))
+  )
 )
 
+(assert_return (invoke "check-descs") (i32.const 1))
+(assert_trap (invoke "desc-trap") "unreachable")
+
 (assert_invalid
   (module
     (rec