Value Type Hygiene

May 2018 (v. 0.1)

John Rose and the Valhalla Expert Group

Briefly put, types in L-world are ambiguous, leading to unhygienic mixtures of value operations with reference operations, and uncontrolled pollution from nulls infecting value code.

This note explores a promising proposal for resolving the key ambiguity. It is a cleaner design than the ad hoc mechanisms tried so far. The resulting system would seem to allow more predictable and debuggable behavior, a stronger backward compatibility story, and better optimization.

Problem statement

In the L-world design for value types, the classfile type descriptor syntax is left unchanged, and the pre-existing descriptor form "LFoo;" is overloaded to denote value types as well as object types. A previous design introoduced new descriptors for value types of the form "QFoo;", and possibly a union type "UFoo;". This design might be called Q-world. In comparison with Q-world, the L-world design approach has two advantages–compatibility and migration–but also one serious disadvantage: ambiguity.

L-world is backward compatible with tools that must parse classfile descriptors, since it leaves descriptor syntax unchanged. There have been no changes to this syntax in almost thirty years, and there is a huge volume of code that depends on its stability. The HotSpot JVM itself makes hundreds of distinct decisions based on descriptor syntax which would need careful review and testing if they were to be adapted to take account of a new descriptor type ("QFoo;", etc.).

Because of its backward compatibility, L-world also has a distinctly simpler migration story than previous designs. Some value-based classes, such as Optional and LocalTime, have been engineered to be candidates for migration to proper value types. We wish to allow such a migration without recompiling the world or forcing programmers to recode uses of the migrated types. It is very difficult to sustain the illusion in Q-world that a value type Qjava/util/Optional; can be operated on in old code under the original object type Ljava/util/Optional;, since the descriptors do not match and a myriad of adapters must be spun (one for every mention of the wrong descriptor). With L-world, we have the simpler problem (addressed in this document) of keeping straight the meaning of L-descriptors in each relevant context, whether freshly recompiled or legacy code; this is a simpler problem than spinning adapters.

But not all is well in L-world. The compatibility of descriptors implies that, when a classfile must express a semantic distinction between a reference type and an object type, it must be allowed to do so unambiguously, in a side channel outside of the descriptor.

Our first thought was, “well, just load all the value types and then you will know the list of them”. If we have a global registry of classes (such as the HotSpot system dictionary), nobody needs to express any additional distinctions, since everybody can just ask the register which are the value types.

This simple idea has a useful insight, but it goes wrong in three ways. First, for some use cases such as classfile transformation, it might be difficult to find such a global registry; in some cases we might prefer to rely on local information in the classfile. We need a way for a classfile to encode, within itself, which types it is using as value types, so that all viewers of the classfile can make consistent decisions about what’s a value and what’s not.

Second, if we are running in the JVM, the global registry of value types has to be built up by loading classfiles. In order for every classfile that uses a value type to know its status, the classfile the defines the value type must be loaded first. But there is no way to totally order these constraints, since it is easy to create circular dependencies between value types, either directly or indirectly. (N.B. Well-foundedness rules for layout don’t eliminate all the possible circularities.) And it won’t work to add more initialization phases (“first load all the classfiles, then let them all start asking questions about their contents”), because that would require preloading a classfile for every potential value type mentioned in some other classfile. That’s every name in every "LFoo;" descriptor. Loading a file for every name mentioned anywhere is very un-Java-like, and something that drastic would be required in order to make correct decisions about value types.

That leads to the third problem, which comes from our desire to make a migration story. Some classfiles need to operate on value types as if they were object references. (Below, we will see details of how operations can differ between value and reference types.) This means that, if we are to support migration, we need a way for legacy classfiles to make a local decision to treat a given type as a reference type, for backward compatibility. Luckily, this is possible, but it requires a local indication in the classfile so the JVM can adjust certain operations.

A solution to these problems requires a way for each classfile to declare how it intends to use each type that is a value type, and (what is more) a way for legacy classfiles to peacefully interoperate with migrated value types. We have experimented with various partial solutions, such as adding an extra bit in a context where a value type may occur, to let the JVM know that the classfile intends a value type. (This is the famous ACC_FLATTENABLE bit on fields.) But it turns out that the number of places where value-ness is significant is hard to limit to just a few spots where we can sprinkle a mode bit. We need a comprehensive solution that can clearly and consistently define a classfile’s (local) view of the status of each type it works with, so that when the “value or reference?” question comes up, there is a clear and consistent answer. We need to prevent the values and the references from polluting each other; we need value type hygiene.

Value vs. reference operations

Value types can be thought of as simpler than reference types, because they lack two features of reference types:

This strong distinction between values and references is inspired, in part, by the design of Java’s primitive types, which also are identity free and are not nullable. Every copy of the int value 42 is completely indistinguishable from every other copy, and you can’t cast a null to int (without a null pointer exception). We hope eventually to unify value types and primitives, but even if this never comes to pass, our design slogan for value types is, codes like a class, works like an int.

By divesting themselves of identity and nullability, value types are able to enjoy new behaviors and optimizations akin to those of primitives, notably flattening in the heap and scalarization in compiled code.

To unlock these benefits, the JVM must treat values and references as operationally distinct. Some of these operational distinctions are quite subtle; some required months of discussion to elucidate, though soon (we hope) they will be obvious in hindsight.

Here is a partial list of cases where the JVM should be able to distinguish value types from reference types:

This list can be tweaked to make it shorter, by adjusting the rules in ways that lessen the impact of ambiguity in type names. The list is also incomplete. (We will add to it later.) Each point of distinction is the subject of detailed design trade-offs, many of which we are sketching here.

Some of these distinctions can be pushed to instruction link time (when resolved value classes may be present) or run time (when the actual values are on stack). A dynamic check can make a final decision, after all questions of value-ness are settled. This seems to be a good decision for acmp. The linkage of a new instruction can fail on a value class, or a checkcast instruction can reject inspected as part of the dynamic execution of operations like new.

But this delaying tactic doesn’t always work. For example, field layout must be computed during class loading, which (as was seen above) is too early to use the supposed global list of value types.

Even if some check can be delayed, like the detection of an erroneous new on a value type, we may well decide it is more useful (or “hygienic”) to detect the error earlier, such as at verification time, so that a broken program can be detected before it starts to run.

Also, some operations may be contextual, to support backward compatibility. Thus, checkcast may need to consult the local classfile about whether to reject nulls, so that legacy code won’t suddenly fail to verify or execute just because it mixes nulls with (what it thought were) references. Basically, a “legacy checkcast” should work correctly with nulls, while an “upgraded checkcast” should probably reject nulls immediately, without requiring extra tests.

We will examine these points in more detail later, but now we need to examine how to contextualize information about value types.

Towards a solution

What is to be done? The rest of this note will propose some solutions to the problem of value type hygiene, and specifically the problem of preventing nulls from mixing with values (“null hygiene”).

Both Remi Forax[1] and Frederic Parain[2] have proposed the idea of having each classfile explicitly declare the list of all value types that it is using. For the record, this author initially resisted the idea[3] as overkill: I was hoping to get away with a band-aid (ACC_FLATTENABLE), but have since realized we need a more aggressive treatment. Clean and tidy behavior from the JVM will make it easier to implement clean and tidy support for value types in the Java language.

Throughout the processing of the classfile, the list can serve as a reliable local registry of decisions about values vs. references. First we will sketch the attribute, and then revisit the points above to see how the list may be used.

The ValueTypes attribute

As proposed above, let us define a new attribute called ValueTypes which is simply a counted array of CONSTANT_Class indexes. Each indexed constant is loaded and checked to be a value type. The JVM uses this list of locally declared value types for all further decisions about value types, relative to the current class.

As a running reference, let’s call the loaded class C. C may be any class, either an object or a value. The value types locally declared by C we can call Q, Q1, Q2, etc. These are exactly the types which would get Q descriptors in Q-world.

As an attribute, ValueTypes is somewhat like the InnerClasses attribute. Both list all classes, within the context a particular classfile, which need some sort of special processing. The InnerClasses attribute includes additional data for informing the special processing (including the break down of “binary names” into outer and inner names, and extra modifier bits), but the ValueTypes attribute only needs to mention the classes which are known to be value types.

Already with the ACC_FLATTENABLE bit we have successfully defined logic that pre-loads a supposed value type, ensures that it is in fact a value type, and then allows the JVM to use all of the necessary properties of that value type to improve the layout of the current class. The classes mentioned in ValueTypes would be pre-loaded similarly. In fact, the ACC_FLATTENABLE bit is no longer needed, since the JVM can simply flatten all fields whose type names are mantioned in the local ValueTypes list.

We now come to the distinction between properly resolved classes (CONSTANT_Class entries) and types named in descriptors. This distinction is important to keep in mind. Once a proper class constant K is resolved by C, everything is known about it, and a permanent link to K goes into C’s constant pool. The same is not true of other type names that occur within field and method descriptors. In order for C to check whether its field type "LK;" is a value type, it must not try to resolve K. Instead it must look for K by name in the list of locally declared value types. Later on, when we examine verifier types and the components of method descriptors a similar by-name lookup will be necessary to decide whether they refer to value types. Thus, there are two ways a type can occur in a classfile and two ways to decide if it is a value type: By resolving a proper constant K and looking at the metadata, and by matching a name "LK;" against the local list. Happily, the answers will be complete and consistent if all the queries look at the same list.

So a type name can be classified as a value type without resolution, by looking for the same name in the names of the list of declared value types. And this can be done even before the list of declared value types is available. This means that any particular declared value types might not need to be loaded until “hard data” is required of it. A provisional determination of the value status of some Q can be made very early before Q’s classfile is actually located and pre-loaded. That provision answer might be enough to check some early structural constraint. It seems reasonable to actually pre-load the Q values lazily, and only when the JVM needs hard data about Q, like its actual layout, or its actual supers.

What if an element of ValueTypes turns out to be a reference type? (Perhaps someone deployed a value-type version of Optional but then got cold feet; meanwhile C is still using it under the impression it is a value type.) There are two choices here, loose and strict, either pretend the type wasn’t there anyway, or raise an error in the loading of the current classfile. The strict behavior is safer; we can loosen it later if we find a need. The case of an element failing to load at all can be treated like the previous problem, either loosely or strictly; strict is better all else being equal.

The strict treatment is also more in line with how to treat failed resolution of super-types, which are a somewhat similar kind of dependency: Super-types, like value types, are loaded as early as possible, and play a role in all phases of classfile loading, notably verification.

One corollary of making the list an attribute is that it can be easily stripped, just like InnerClasses or BootstrapMethods. Is this a bug or a feature? In the case of InnerClasses, stripping the attribute doesn’t affect execution of the classfile but it does break some Core Reflection queries. In the case of BootstrapMethods, the structural constraints on dynamic constant pool constants will break, and the classfile will promptly fail to load. The effect of removing a ValueTypes attribute is probably somewhere in between. Because L-world types are ambiguous, and because we specifically allow value types to be used as references from legacy classfiles (for migration), there’s always a way to “fake” enough reference behavior from a value type in a classfile which doesn’t make special claims about it. So it seems reasonable to allow ValueTypes to be stripped, at least in principle. At a worst case the classfile will fail to load, as in the case of a stripped BootstrapMethods, but the feature might actually prove useful (say, for before-and-after migration testing).

Note that in principle a classfile generator could choose to ignore a value type, and treat it as a (legacy) reference type. Because of migration, the JVM must support at least some such moves, but such picking and choosing is not the center of our design. In particular, we do not want the same compilation unit to treat a type as a value type in one line of code, and a reference type in the next. This may come later, if we decide to introduce concepts of nullable values and/or value boxes, but we think we can defer such features.

So for now, classfiles may differ among themselves about which types are value types, but within a single classfile there is only one source of local truth about value types. (Locally-sourced, fresh, hygienic data!)

Value types and class structure

Very early during class loading, the JVM assigns an instance layout to the new class C. Before that point it must first load the declared value types (Q1, Q2, …), and then recursively extract the layout information from each one. There is no danger of circularity in this because a value type instance cannot contain another instance of itself, directly or indirectly.

Both non-static and static fields of value type make sense (because a value “works like an int”). But static fields interact differently with the loading process than non-static fields.

A static value type field has no enclosing instance, unless the JVM chooses to make one secretly. Therefore it doesn’t need to be flattened. The JVM can make an invisible choice about how to store a static value type field:

The first option seems easiest, but the second might be more performant. The third difficult due to bootstrapping concerns.

In fact, the same implementation options apply for non-statics as for statics, but only the third one (full flattening) is desirable. The first one (immutable buffering) may be useful as a fallback implemmentation technique for special cases like jumbo values and fields which are volatile, and thus need to provide atomicity.

The root container for all of C’s statics, in HotSpot, happens to be the relevant the java.lang.Class value C.class. Presumably it’s a good place to put the invisible pointers mentioned above.

A static field of value type Q cannot make its initial value available to getfield until Q’s <clinit> method runs, (or in the case of re-entrant initialization, has at least started). Since classes can circularly refer to instances of each other via static references, Q might return the favor and require materialization of C.

The first time C requires Q’s default value, if Q has not been initialized, its <clinit> method should run. This may trigger re-entry into the initializer for C, so Q needs to get its act together before it runs its <clinit>, and immediately create Q’s own default value, storing it somewhere in Q’s own metadata (or else the Class mirror looks like a good spot). The invariant is that, before Q’s class initializer can run one bytecode, the default value for Q is created and registered for all time. Creating the default value before the initializer runs is odd but harmless, as long as no bytecode can actually access the default value without triggering Q’s initialization.

This also implies that C should create and register its own default value (if it is a value type) before it runs its own <clinit> method, lest Q come back and ask C for its value type.

The JVM may be required to bootstrap value-type statics as invisible null pointers, which are inflated (invisibly by the getstatic and/or putstatic instructions) into appropriate buffers, after ensuring the initialization of the value type class. But it seems possible that if the previous careful sequencing is observed, there is no need to do lazy inflation of nulls, which would simplify the code for getstatic and putstatic.

Value types and method linkage

A class C includes methods as well as fields, of course. A method can receive or return a value type Q simply by mentioning Q as a component of its method descriptor (as an L-descriptor "LQ;").

If a method C.m()LD; mentions some type D which is not on the declared list, then that type D will be treated, like always, as a nullable, identity-bearing reference.

Interestingly, migration compatibility requires this to be the case whether or not D is in actual fact a value type. If C is unconscious of D’s value-ness, the JVM must respect this, and preserve the illusion that D values are “just references, nothing to see here, move along”. Perhaps D is freshly upgraded to a value type, and C isn’t recompiled yet. C should not be penalized for this change, if at all possible.

This points to a core decision of the L-world design, that nearly all of the normal operations on object references “just do the right thing” when applied to value types. The two kinds of data use the same descriptor syntax. Value types can be converted to Object references, even though the resulting pseudo-reference does not expose any identity (and will never be null). Instructions like aload operate on values just as well as references, and so on.

Basically, values in L-world routinely go around lightly disguised as references, special pseudo-references which do not retain object identity. As long as nobody looks closely, the fiction that they are references is unbroken. If someone tries a monitorenter instruction, the game is over, but we think those embarassing moments will be rare.

On the other hand, if a method C.m()LQ; uses a locally-declared value type, then the JVM has some interesting options. It may choose to notice that the Q-value is not nullable, has no identity. It can adjust the calling sequence of m to work with undisguised “naked values”, which are passed on the stack, opr broken into components for transport across the method API. This would almost be a purely invisible decision, except that naked values cannot be null, and so such calling sequences are hostile to null. Again, it “works like an int”. A null Integer value will do just the same thing if you try to pass it to an int-recieving method. So we have to be prepared for an occasional embarassing NPE, when one party thinks a type is a nullable reference type and the other party knows it’s a value type.

One might think that it is straightforward to assign a value-using method a calling sequence by examining the method signature and the locally declared value types of the declaring class. But in fact there are non-local constraints. Only static and private methods can easily be adjusted to work with naked values.

Unlike fields, methods can override similar methods in some C’s super-type S. This immediately leads to the possibility of C and S differing as to the status of some type X in the method’s signature. If neither of the ValueTypes lists of C and S mentions X, then the classes are agreed that X is an object type (even if in truth it happens to be a value type). They can agree to use a reference-based calling sequence for some m that works with X.

If both lists mention some Q, then both classes agree, and in fact it must be a value type. They might be able to agree to use “naked values” for the Q type when calling the method. Or not: they still have to worry about other supers that might have another opinion about Q.

What if C doesn’t list Q but S does, and they share a method that returns Q? For example, what about C.m()Q vs. S.m()Q? In that case, the JVM may have already set up S.m to return its Q result as a naked value. Probably this happend before C was even loaded. The code for C.m will expect simply to return a normal reference. In reality, it will be unconsciously holding a JVM-assigned pseudo-reference to the buffered Q-value. The JVM must then unwrap the reference into a naked value to match the calling sequence it assigned (earlier, before C was loaded) to S.m. The bottom line is that even though C.m was loaded as a reference-returning function, the JVM may secretly rewrite it to return a naked value.

Since C.m returns a reference, it might choose to return null. What happens then? The secretly installed adaptation logic cannot extract the components of a buffer that doesn’t exist. A NullPointerException must be thrown, at the point where C.m is adapted to S.m, which has the greater knowledge that Q is value type (hence non-nullable). It will be as if the areturn instruction of C.m included a hidden null check.

Is such a hidden null check reasonable? One might explain that the C code thinks (wrongly) it is working with boxes, while the S code knows it is working with values. If the method were C.m()Integer and it were overriding S.m()int, then if C.m returns null then the adapter that converts to S.m()int must throw NPE during the implicit conversion from Integer to int. A value “works like an int”, so the result must be similar with a value type. It is as if the deficient class C were working with boxes for Q (indeed that’s all it sees) while the knowledgeable class S is working with true values. The NPE seems justifiable in such terms, although there is no visible adapter method to switch descriptors in this case.

The situation is a little odd when looked at the following way: If you view nullability as a privilege, then this privilege is enjoyed only by deficient classes, ones that have not yet been recompiled to “see” that the type Q is a value type. Ignorant classes may pass null back and forth through Q APIs, all day long, until they pass it through a class that knows Q is a value. Then an NPE will end their streak of luck. Is using null a privilege? Well, yes, but remember also that if Q started its career as an object type, it was a value-based class, and such classes are documented as being null-hostile. The null-passers were in a fool’s paradise.

What if C lists Q as a value but S doesn’t? Then the calling sequence assigned when S was loaded will use references, and these references will in fact be pseudo-references to buffered Q values (or null, as just discussed). The knowledgeable method C.m()Q will never produce a null through this API. The JVM will arrange to properly clothe the Q-value produced by C.m into a buffer whose pointer can be returned from S.m.

Class hierarchies can be much deeper than just C and S, and overrides can occur at many levels on the way down. Frederic Parain has pointed out that the net result seems to be that the first (highest) class that declares a given method (with descriptor) also gets to determine the calling sequence, which is then imposed on all overrides through that class. This leads to a workable implementation strategy, based on v-table packing. A class’s v-table is packed at during the “preparation” phase of class linking, just after loading before any subclass v-table is packed. The JVM knows, unambiguously, whether a given v-table entry is new to a class, or is being reaffirmed from a previous super-class (perhaps with an override, perhaps just with an abstract). At this point, a new v-table slot can be given a carefully selected internal calling sequence, which will then be imposed on all overrides. An old v-table slot will have the super’s calling sequence imposed on it. In this scheme, the interpreter and compiler must examine both the method descriptor and some metadata about the v-table slot when performing invokevirtual or invokespecial.

A method coming in “sideways” from an interface is harder to manage. It is reasonable to treat such a method as “owned” by the first proper class that makes a v-table entry for it. But that only works for one class hierarchy; the same method might show up in a different hierarchy with incompatible opinions about value types in the method signature. It appears that interface default methods, if not class methods, must be prepared to use more than one kind of calling sequence, in some cases. It is as if, when a class uses a default method, it imports that method and adjusts the method’s calling sequence to agree with that class’s hierarchy.

Often an interface default method is completely new to a class hierarchy. In that case, the interface can choose the calling sequence, and this is likely to provide more coherent calling sequences for that API point.

These complexities will need watching as value types proliferate and begin to show up in interface-based APIs.

Value types and the verifier

Let us assume that, if the verifier sees a value type, it should flag all invalid uses of that value type immediately, rather than wait for execution.

(This assumption can be relaxed, in which case many points in this section can be dropped. We may also try to get away with implementing as few of these checks as possible, saving them for a later release.)

When verifying a method, the verifier tracks and checks types by name, mostly. Sometimes it pre-loads classes to see the class hierarchy. With the ValueTypes attribute, there is no need to pre-load value classes; the symbolic method is sufficient.

The verifier type system needs a way to distinguish value types from regular object types. To keep the changes small, this distinction can be expressed as a local predicate on type names called isValueType, implemented by referring to ValueTypes. In this way, the StackMapTable attribute does not need any change at all. Nor does the verifier type system need a change: value types go under the Object and Reference categories, despite the fact that value types are not object types, and values are not references.

The verifier rules need to consult isValueType at some points. The assignability rules for null must be adjusted to exclude value classes.

isAssignable(null, class(X, _)) :-  not(isValueType(X)).

This one change triggers widespread null rejection: wherever a value type is required, the verifier will not allow a null to be on the stack. Assuming null is on the stack and Q is a value type, the following will be rejected as a consequence of the above change:

Given comprehensive null blocking (along other paths also), the implementation of the putfield (or withfield) instruction could go ahead and pull a buffered value off the stack without first checking for null. If the verifier does not actually reject such nulls, the dynamic behavior of the bytecodes themselves should, to prevent null pollution from spreading.

The verifier rules for aastore and checkcast only check that the input type is an object reference of some sort. More narrow type checks are performed at runtime. A null may be rejected dynamically by these instructions, but the verifier logic does not need to track nulls for them.

The verifier rules for invokespecial have special cases for <init> methods, but these do not need special treatment, since such calls will fail to link when applied to a value type receiver.

The verifier could reject reference comparisons between value types other operands (including null, other value types, and reference types). This would look something like an extra pair of constraints after the main assertion that two references are on the stack:

instructionIsTypeSafe(if_acmpeq(Target), Environment, _Offset, StackFrame,
                      NextStackFrame, ExceptionStackFrame) :- 
    canPop(StackFrame, [reference, reference], NextStackFrame),
+   not( canPop(StackFrame, [_, class(X, _)], _), isValueType(X) ),
+   not( canPop(StackFrame, [class(X, _), _], _), isValueType(X) ),
    targetIsTypeSafe(Environment, NextStackFrame, Target),
    exceptionStackFrame(StackFrame, ExceptionStackFrame).

(The JVMS doesn’t use any such not operator. The actual Prolog changes would be more complex, perhaps requiring a real_reference target type instead of reference.)

This point applies equally to if_acmpeq, if_acmpne, if_null, and if_nonnull,

This doesn’t seem to be worth while, although it might be interesting to try to catch javac bugs this way. In any case, such comparisons are guaranteed to return false in L-world, and will optimize quickly in the JIT.

In a similar vein, the verifier could reject monitorenter and monitorexit instructions when they apply to value types:

instructionIsTypeSafe(monitorenter, _Environment, _Offset, StackFrame,
                      NextStackFrame, ExceptionStackFrame) :-
    canPop(StackFrame, [reference], NextStackFrame),
+   not( canPop(StackFrame, [class(X, _)], _), isValueType(X) ),
    exceptionStackFrame(StackFrame, ExceptionStackFrame).

And a new or putfield could be quickly rejected if it applies to a value type:

instructionIsTypeSafe(new(CP), Environment, Offset, StackFrame,
                      NextStackFrame, ExceptionStackFrame) :-
    StackFrame = frame(Locals, OperandStack, Flags),
    CP = class(X, _),
+   not( isValueType(X) ),

instructionIsTypeSafe(putfield(CP), Environment, _Offset, StackFrame,
                      NextStackFrame, ExceptionStackFrame) :-
    CP = field(FieldClass, FieldName, FieldDescriptor),
+   not( isValueType(FieldClass) ),

Likewise withfield could be rejected by the verifier if applied to a non-value type.

The effect of any or all of these verifier rule changes (if we choose to implement them) would be to prevent local code from creating a null and accidentally putting it somewhere a value type belongs, or from accidentally applying an identity-sensitive operation to an operand known statically to be a value type. These rules only work when a sharp verifier type unambiguously reports an operand as null or as a value type.

Nulls must also be rejected, and value types detected, when they are hidden, at verification time, under looser types like Object. Protecting local code from outside nulls must also be done dynamically.

Omitting all of these rules will simply shift the responsibility for null rejection and value detection fully to dynamic checks at execution time, but such dynamic checks must be implemented in any case, so the verifier’s help is mainly an earlier error check, especially to prevent null pollution inside of a single stack frame. For that reason, the only really important verifier change is the isAssignable adjustment, mentioned first.

The dynamic checks which back up or replace the other verifier checks will be discussed shortly.

Value types and legacy classfiles

We need to discuss the awkward situation of null being passed as a value type, and value types being operated on as objects, by legacy classfiles. One legacy classfile can dump null values into surprising places, even if all the other classfiles are scrupulous about containing null.

We will also observe some of the effects of having value types “invade” a legacy classfile which expects to apply identity-sensitive operations to them.

By “legacy classfile” we of course mean classfiles which lack ValueTypes attributes, and which may attempt to misuse value types in some way. (Force of habit: it’s strong.) We also can envision half-way cases where a legacy classfile has a ValueTypes attribute which is not fully up to date. In any case, there is a type Q which is not locally declared as a value type, by the legacy class C.

The first bad thing that can happen is that C declares a field of type Q. This field will be formatted as a reference field, even though the field type is a value type. Although we implementors might grumble a bit, the JVM will have to arrange to use pseudo-pointers to represent values stored in that field. (It’s as if the field were volatile, or not flattenable for some other reason.) That wasn’t too bad, but look what’s in the field to start with: It’s a null. That means that any legitmate operation on this initial value will throw an NPE. Of course, the writer of C knew Q as a value-based class, so the initial null will be discarded and replaced by a suitable non-null value, before anything else happens.

What if C makes a mistake, and passes a null to another class which does know Q is a value? At that point we have a choice, as with the verifier’s null rejection whether to do more work to detect the problem earlier, or whether to let the null flow through and eventually cause an NPE down the line. Recall that if an API point gets a calling sequence which recognizes that Q is a value type, it will probably unbuffer the value, throwing NPE immediately if C makes a mistake. This is good, because that’s the earliest we could hope to flag the mistake. But if the method accepts the boxed form of Q, then the null will sneak in, skulk around in the callee’s stack frame, and maybe cause an error later.

Meanwhile, if the JVM tries to optimize the callee, it will have to limit its optimizations somewhat, because the argument value is nullable (even if only ever by mistake). To cover this case, it may be useful to define that method entry to a method that knows about Q is null-hostile, even if the calling sequence for some reason allows references. This means that, at function entry, every known value type parameter is null-checked. This needs to be an official rule in the JVM, not just an optimization for the JIT, in order for the JIT to use it.

What if our C returns a null value to a caller who intends to use it as a value? That won’t go well either, but unless we detect the null aggressively, it might rattle around for a while, disrupting optimization, before produing an inscrutable error. (“Where’d that null come from??”). The same logic applies as with arguments: When a null is returned from a method call that purports to return Q, this can only be from a legacy file, and the calling sequences were somehow not upgraded. In that case, the JVM needs to mandate a null check on every method invocation which is known to return a value type.

The same point also applies if another class A, knowing Q as a value type, happens to load a null from one of C’s fields. The C field is formatted as a reference, and thus can hand A a surprise null, but A must refuse to see it, and throw NPE. Thus, the getfield instruction, if it is pointed at a legacy non-flattened field, will need to null-check the value loaded from the field.

Meanwhile, C is allowed to putfield and getfield null all day long into its own fields (and fields of other benighted legacy classes that it may be friends with). Thus, the getfield and putfield instructions link to slightly different behavior, not only based on the format of the field, but also based on “who’s asking”. Code in C is allowed to witness nulls in its Q fields, but code in A (upgraded) is not allowed to see them, even though it’s the same getfield to the same symbolic reference. Happily, fields are not shared widely across uncoordinated classfiles, so this is a corner case mainly for testers to worry about.

What if C stores a null into somebody else’s Q field, or into an element of a Q[] array? In that case, C must throw an immediate NPE; there’s no way to reformat someone else’s data structure, however out-of-date C may be.

What if C gets a null value from somewhere and casts it to Q? Should the checkcast throw NPE (as it should in a classfile where Q is known to be a value type)? For compatibility, the answer is “no”; old code needs to be left undisturbed if possible. After all, C believes it has a legitimate need for nulls, and won’t be fixed until it is recompiled and its programmer fixes the source code.

That’s about it for null. If the above dynamic checks are implemented, then legacy classfiles will be unable to disturb upgraded classfiles with surprise null values. The goal mentioned above about controlling null on all paths is fulfilled blocking null across API calls (which might have a legacy class on one end), and by verifying that nulls never mix with values, locally within a single stack frame.

There are a few other things C’s could do to abuse Q values. Legacy code needs to be prevented immediately from making any of the following mistakes:

Happily, the above rules are not specific to legacy code but apply uniformly everywhere.

A final mistake is executing an acmp instruction on a value type. Again, this is possible everywhere, not just in legacy files, even if the verifier tries to prevent the obvious occurrences. There are several options for acmp on value types. The option which breaks the least code and preserves the O(1) performance model of acmp is to quickly detect a value type operand and just report false, even if the JVM can tell, somehow, that it’s the same buffer containing the same value, being compared to itself.

All of these mistakes can be explained by analogy, supposing that the legacy class C were working with a box type Integer where other classes had been recoded to use int. All variables under C’s control are nullable, but when it works with new code it sees only int variables. Implicit conversions sometimes throw NPE, and acmp (or monitorenter) operations on boxed Integer values yield unspecific (or nonsensical) results.

Value types and instruction linkage

Linked instructions which are clearly wrong should throw a LinkageError of some type. Examples already given are new and putfield on value types.

When a field reference of value type is linked it will have to correctly select the behavior required by both the physical layout of the field, and also the stance toward any possible null if the field is nullable. (As argued above, the stance is either lenient for legacy code or strict for new code.)

A getstatic linkage may elect to replace an invisible null with a default value.

When an invoke is linked it will have to arrange to correctly execute the calling sequence assigned to its method or its v-table.

Linkage of invokeinterface will be even more dynamic, since the calling sequence cannot be determined until the receiver class is examined.

Linkage of dynamic constants in the constant pool must reject null for value types. Value types can be determined either globally based on the resolved constant type, or locally based on the ValueTypes attribute associated with the constant pool in which the resolution occurs.

Value types and instruction execution

Most of the required dynamic behaviors to support value type hygiene have already been mentioned. Since values are identity-free and non-nullable, the basic requirement is to avoid storing nulls in value-type variables, and degrade gracefully when value types are queried about their identities. A secondary requirement is to support the needs of legacy code.

For null hygeine, the following points apply:

Linked field and invoke instructions need sufficient linkage metadata to correctly flatten instance fields and use unboxed (and/or null hostile) calling sequences.

As discussed above, the acmp must short circuit on values. This is a dynamic behavior, not affected by the ValueTypes attribute.

Generally speaking, any instruction that doesn’t refer to the constant pool cannot have contextual behavior, because there is no place to store metadata to adjust the behavior. The areturn instruction is an exception to this observation; it is a candidate for bytecode rewriting to gate the extra null check for applicable methods.

Value types and reflection

Some adjustments may be needed for the various reflection APIs, in order to bring them into alignment with the changed bytecode.


The details are complex, but the story as a whole becomes more intelligible when we require each classfile to locally declare its value types, and handle all values appropriately according to the local declaration.

Outside of legacy code, and at its boundaries, tight control of null values is feasible. Inside value-rich code, and across value-rich APIs, full optimization seems within reach.

Potential problems with ambiguity in L-world are effectively addressed by a systematic side channel for local value type declarations, assisting the interpratation of L-type descriptors. This side channel can be the ValueTypes attribute.