Rewrite LOAD_GLOBAL _Unbound to LOAD_IMMEDIATE in builtins module

This should save some caches and be faster. Ideally we would do this in
all Skybison-controlled modules but I don't think we have a good signal
for that yet.

runtime/bytecode.cpp

@@ -89,7 +89,8 @@ struct RewrittenOp {
   bool needs_inline_cache;
 };
 
-static RewrittenOp rewriteOperation(const Function& function, BytecodeOp op,
+static RewrittenOp rewriteOperation(Thread* thread, const Function& function,
+                                    BytecodeOp op,
                                     bool use_load_fast_reverse_unchecked) {
   auto cached_binop = [](Interpreter::BinaryOp bin_op) {
     return RewrittenOp{BINARY_OP_ANAMORPHIC, static_cast<int32_t>(bin_op),
@@ -178,6 +179,22 @@ static RewrittenOp rewriteOperation(const Function& function, BytecodeOp op,
       return cached_inplace(Interpreter::BinaryOp::XOR);
     case LOAD_ATTR:
       return RewrittenOp{LOAD_ATTR_ANAMORPHIC, op.arg, true};
+    case LOAD_GLOBAL: {
+      RawObject module = function.moduleObject();
+      if (module.isModule() &&
+          Module::cast(module).id() == thread->runtime()->builtinsModuleId()) {
+        RawTuple names = Tuple::cast(Code::cast(function.code()).names());
+        RawStr name = Str::cast(names.at(op.arg));
+        if (name.equalsCStr("_Unbound")) {
+          DCHECK(Unbound::object() ==
+                     objectFromOparg(opargFromObject(Unbound::object())),
+                 "Expected to be able to fit _Unbound in a byte");
+          return RewrittenOp{LOAD_IMMEDIATE, opargFromObject(Unbound::object()),
+                             false};
+        }
+      }
+      break;
+    }
     case LOAD_FAST: {
       CHECK(op.arg < Code::cast(function.code()).nlocals(),
             "unexpected local number");
@@ -305,7 +322,7 @@ void rewriteBytecode(Thread* thread, const Function& function) {
       use_load_fast_reverse_unchecked = false;
       continue;
     }
-    RewrittenOp rewritten = rewriteOperation(function, op, false);
+    RewrittenOp rewritten = rewriteOperation(thread, function, op, false);
     if (rewritten.needs_inline_cache) {
       num_caches++;
     }
@@ -326,7 +343,7 @@ void rewriteBytecode(Thread* thread, const Function& function) {
     BytecodeOp op = nextBytecodeOp(bytecode, &i);
     word previous_index = i - 1;
     RewrittenOp rewritten =
-        rewriteOperation(function, op, use_load_fast_reverse_unchecked);
+        rewriteOperation(thread, function, op, use_load_fast_reverse_unchecked);
     if (rewritten.bc == UNUSED_BYTECODE_0) continue;
     if (rewritten.needs_inline_cache) {
       rewrittenBytecodeOpAtPut(bytecode, previous_index, rewritten.bc);