X-Git-Url: http://git.sourceforge.jp/view?a=blobdiff_plain;f=libjava%2Fverify.cc;h=b002c1c0aabfb99ec7aeec267aef64a7a59876cc;hb=2c622b2eac33e9946e2ad459663f7c993d043e2a;hp=4a6ca45884965e5db7fda5cfae64a733d797ed5e;hpb=2bcd45dce03c7e08d335e305b6501acecb45111d;p=pf3gnuchains%2Fgcc-fork.git diff --git a/libjava/verify.cc b/libjava/verify.cc index 4a6ca458849..b002c1c0aab 100644 --- a/libjava/verify.cc +++ b/libjava/verify.cc @@ -1,6 +1,6 @@ -// defineclass.cc - defining a class from .class format. +// verify.cc - verify bytecode -/* Copyright (C) 2001, 2002, 2003 Free Software Foundation +/* Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006 Free Software Foundation This file is part of libgcj. @@ -14,11 +14,18 @@ details. */ #include +#include + #include #include #include #include +// On Solaris 10/x86, indirectly includes , which +// defines PC since g++ predefines __EXTENSIONS__. Undef here to avoid clash +// with PC member of class _Jv_BytecodeVerifier below. +#undef PC + #ifdef INTERPRETER #include @@ -26,17 +33,22 @@ details. */ #include #include #include +#include #ifdef VERIFY_DEBUG #include #endif /* VERIFY_DEBUG */ +// This is used to mark states which are not scheduled for +// verification. +#define INVALID_STATE ((state *) -1) + static void debug_print (const char *fmt, ...) __attribute__ ((format (printf, 1, 2))); static inline void -debug_print (const char *fmt, ...) +debug_print (MAYBE_UNUSED const char *fmt, ...) { #ifdef VERIFY_DEBUG va_list ap; @@ -46,6 +58,80 @@ debug_print (const char *fmt, ...) #endif /* VERIFY_DEBUG */ } +// This started as a fairly ordinary verifier, and for the most part +// it remains so. It works in the obvious way, by modeling the effect +// of each opcode as it is encountered. For most opcodes, this is a +// straightforward operation. +// +// This verifier does not do type merging. It used to, but this +// results in difficulty verifying some relatively simple code +// involving interfaces, and it pushed some verification work into the +// interpreter. +// +// Instead of merging reference types, when we reach a point where two +// flows of control merge, we simply keep the union of reference types +// from each branch. Then, when we need to verify a fact about a +// reference on the stack (e.g., that it is compatible with the +// argument type of a method), we check to ensure that all possible +// types satisfy the requirement. +// +// Another area this verifier differs from the norm is in its handling +// of subroutines. The JVM specification has some confusing things to +// say about subroutines. For instance, it makes claims about not +// allowing subroutines to merge and it rejects recursive subroutines. +// For the most part these are red herrings; we used to try to follow +// these things but they lead to problems. For example, the notion of +// "being in a subroutine" is not well-defined: is an exception +// handler in a subroutine? If you never execute the `ret' but +// instead `goto 1' do you remain in the subroutine? +// +// For clarity on what is really required for type safety, read +// "Simple Verification Technique for Complex Java Bytecode +// Subroutines" by Alessandro Coglio. Among other things this paper +// shows that recursive subroutines are not harmful to type safety. +// We implement something similar to what he proposes. Note that this +// means that this verifier will accept code that is rejected by some +// other verifiers. +// +// For those not wanting to read the paper, the basic observation is +// that we can maintain split states in subroutines. We maintain one +// state for each calling `jsr'. In other words, we re-verify a +// subroutine once for each caller, using the exact types held by the +// callers (as opposed to the old approach of merging types and +// keeping a bitmap registering what did or did not change). This +// approach lets us continue to verify correctly even when a +// subroutine is exited via `goto' or `athrow' and not `ret'. +// +// In some other areas the JVM specification is (mildly) incorrect, +// so we diverge. For instance, you cannot +// violate type safety by allocating an object with `new' and then +// failing to initialize it, no matter how one branches or where one +// stores the uninitialized reference. See "Improving the official +// specification of Java bytecode verification" by Alessandro Coglio. +// +// Note that there's no real point in enforcing that padding bytes or +// the mystery byte of invokeinterface must be 0, but we do that +// regardless. +// +// The verifier is currently neither completely lazy nor eager when it +// comes to loading classes. It tries to represent types by name when +// possible, and then loads them when it needs to verify a fact about +// the type. Checking types by name is valid because we only use +// names which come from the current class' constant pool. Since all +// such names are looked up using the same class loader, there is no +// danger that we might be fooled into comparing different types with +// the same name. +// +// In the future we plan to allow for a completely lazy mode of +// operation, where the verifier will construct a list of type +// assertions to be checked later. +// +// Some test cases for the verifier live in the "verify" module of the +// Mauve test suite. However, some of these are presently +// (2004-01-20) believed to be incorrect. (More precisely the notion +// of "correct" is not well-defined, and this verifier differs from +// others while remaining type-safe.) Some other tests live in the +// libgcj test suite. class _Jv_BytecodeVerifier { private: @@ -55,9 +141,15 @@ private: struct state; struct type; - struct subr_info; - struct subr_entry_info; struct linked_utf8; + struct ref_intersection; + + template + struct linked + { + T *val; + linked *next; + }; // The current PC. int PC; @@ -67,29 +159,21 @@ private: // The current state of the stack, locals, etc. state *current_state; - // We store the state at branch targets, for merging. This holds - // such states. - state **states; + // At each branch target we keep a linked list of all the states we + // can process at that point. We'll only have multiple states at a + // given PC if they both have different return-address types in the + // same stack or local slot. This array is indexed by PC and holds + // the list of all such states. + linked **states; - // We keep a linked list of all the PCs which we must reverify. - // The link is done using the PC values. This is the head of the - // list. - int next_verify_pc; + // We keep a linked list of all the states which we must reverify. + // This is the head of the list. + state *next_verify_state; // We keep some flags for each instruction. The values are the - // FLAG_* constants defined above. + // FLAG_* constants defined above. This is an array indexed by PC. char *flags; - // We need to keep track of which instructions can call a given - // subroutine. FIXME: this is inefficient. We keep a linked list - // of all calling `jsr's at at each jsr target. - subr_info **jsr_ptrs; - - // We keep a linked list of entries which map each `ret' instruction - // to its unique subroutine entry point. We expect that there won't - // be many `ret' instructions, so a linked list is ok. - subr_entry_info *entry_points; - // The bytecode itself. unsigned char *bytecode; // The exceptions. @@ -100,27 +184,21 @@ private: // This method. _Jv_InterpMethod *current_method; - // A linked list of utf8 objects we allocate. This is really ugly, - // but without this our utf8 objects would be collected. - linked_utf8 *utf8_list; + // A linked list of utf8 objects we allocate. + linked<_Jv_Utf8Const> *utf8_list; - struct linked_utf8 - { - _Jv_Utf8Const *val; - linked_utf8 *next; - }; + // A linked list of all ref_intersection objects we allocate. + ref_intersection *isect_list; + // Create a new Utf-8 constant and return it. We do this to avoid + // having our Utf-8 constants prematurely collected. _Jv_Utf8Const *make_utf8_const (char *s, int len) { - _Jv_Utf8Const *val = _Jv_makeUtf8Const (s, len); - _Jv_Utf8Const *r = (_Jv_Utf8Const *) _Jv_Malloc (sizeof (_Jv_Utf8Const) - + val->length - + 1); - r->length = val->length; - r->hash = val->hash; - memcpy (r->data, val->data, val->length + 1); - - linked_utf8 *lu = (linked_utf8 *) _Jv_Malloc (sizeof (linked_utf8)); + linked<_Jv_Utf8Const> *lu = (linked<_Jv_Utf8Const> *) + _Jv_Malloc (sizeof (linked<_Jv_Utf8Const>) + + _Jv_Utf8Const::space_needed(s, len)); + _Jv_Utf8Const *r = (_Jv_Utf8Const *) (lu + 1); + r->init(s, len); lu->val = r; lu->next = utf8_list; utf8_list = lu; @@ -128,7 +206,7 @@ private: return r; } - __attribute__ ((__noreturn__)) void verify_fail (char *s, jint pc = -1) + __attribute__ ((__noreturn__)) void verify_fail (const char *s, jint pc = -1) { using namespace java::lang; StringBuffer *buf = new StringBuffer (); @@ -146,9 +224,9 @@ private: buf->append (JvNewStringLatin1 (" in ")); buf->append (current_class->getName()); buf->append ((jchar) ':'); - buf->append (JvNewStringUTF (method->get_method()->name->data)); + buf->append (method->get_method()->name->toString()); buf->append ((jchar) '('); - buf->append (JvNewStringUTF (method->get_method()->signature->data)); + buf->append (method->get_method()->signature->toString()); buf->append ((jchar) ')'); buf->append (JvNewStringLatin1 (": ")); @@ -179,19 +257,245 @@ private: // to indicate an unusable value. unsuitable_type, return_address_type, + // This is the second word of a two-word value, i.e., a double or + // a long. continuation_type, - // There is an obscure special case which requires us to note when - // a local variable has not been used by a subroutine. See - // push_jump_merge for more information. - unused_by_subroutine_type, - // Everything after `reference_type' must be a reference type. reference_type, null_type, - unresolved_reference_type, - uninitialized_reference_type, - uninitialized_unresolved_reference_type + uninitialized_reference_type + }; + + // This represents a merged class type. Some verifiers (including + // earlier versions of this one) will compute the intersection of + // two class types when merging states. However, this loses + // critical information about interfaces implemented by the various + // classes. So instead we keep track of all the actual classes that + // have been merged. + struct ref_intersection + { + // Whether or not this type has been resolved. + bool is_resolved; + + // Actual type data. + union + { + // For a resolved reference type, this is a pointer to the class. + jclass klass; + // For other reference types, this it the name of the class. + _Jv_Utf8Const *name; + } data; + + // Link to the next reference in the intersection. + ref_intersection *ref_next; + + // This is used to keep track of all the allocated + // ref_intersection objects, so we can free them. + // FIXME: we should allocate these in chunks. + ref_intersection *alloc_next; + + ref_intersection (jclass klass, _Jv_BytecodeVerifier *verifier) + : ref_next (NULL) + { + is_resolved = true; + data.klass = klass; + alloc_next = verifier->isect_list; + verifier->isect_list = this; + } + + ref_intersection (_Jv_Utf8Const *name, _Jv_BytecodeVerifier *verifier) + : ref_next (NULL) + { + is_resolved = false; + data.name = name; + alloc_next = verifier->isect_list; + verifier->isect_list = this; + } + + ref_intersection (ref_intersection *dup, ref_intersection *tail, + _Jv_BytecodeVerifier *verifier) + : ref_next (tail) + { + is_resolved = dup->is_resolved; + data = dup->data; + alloc_next = verifier->isect_list; + verifier->isect_list = this; + } + + bool equals (ref_intersection *other, _Jv_BytecodeVerifier *verifier) + { + if (! is_resolved && ! other->is_resolved + && _Jv_equalUtf8Classnames (data.name, other->data.name)) + return true; + if (! is_resolved) + resolve (verifier); + if (! other->is_resolved) + other->resolve (verifier); + return data.klass == other->data.klass; + } + + // Merge THIS type into OTHER, returning the result. This will + // return OTHER if all the classes in THIS already appear in + // OTHER. + ref_intersection *merge (ref_intersection *other, + _Jv_BytecodeVerifier *verifier) + { + ref_intersection *tail = other; + for (ref_intersection *self = this; self != NULL; self = self->ref_next) + { + bool add = true; + for (ref_intersection *iter = other; iter != NULL; + iter = iter->ref_next) + { + if (iter->equals (self, verifier)) + { + add = false; + break; + } + } + + if (add) + tail = new ref_intersection (self, tail, verifier); + } + return tail; + } + + void resolve (_Jv_BytecodeVerifier *verifier) + { + if (is_resolved) + return; + + // This is useful if you want to see which classes have to be resolved + // while doing the class verification. + debug_print("resolving class: %s\n", data.name->chars()); + + using namespace java::lang; + java::lang::ClassLoader *loader + = verifier->current_class->getClassLoaderInternal(); + + // Due to special handling in to_array() array classes will always + // be of the "L ... ;" kind. The separator char ('.' or '/' may vary + // however. + if (data.name->limit()[-1] == ';') + { + data.klass = _Jv_FindClassFromSignature (data.name->chars(), loader); + if (data.klass == NULL) + throw new java::lang::NoClassDefFoundError(data.name->toString()); + } + else + data.klass = Class::forName (_Jv_NewStringUtf8Const (data.name), + false, loader); + is_resolved = true; + } + + // See if an object of type OTHER can be assigned to an object of + // type *THIS. This might resolve classes in one chain or the + // other. + bool compatible (ref_intersection *other, + _Jv_BytecodeVerifier *verifier) + { + ref_intersection *self = this; + + for (; self != NULL; self = self->ref_next) + { + ref_intersection *other_iter = other; + + for (; other_iter != NULL; other_iter = other_iter->ref_next) + { + // Avoid resolving if possible. + if (! self->is_resolved + && ! other_iter->is_resolved + && _Jv_equalUtf8Classnames (self->data.name, + other_iter->data.name)) + continue; + + if (! self->is_resolved) + self->resolve(verifier); + + // If the LHS of the expression is of type + // java.lang.Object, assignment will succeed, no matter + // what the type of the RHS is. Using this short-cut we + // don't need to resolve the class of the RHS at + // verification time. + if (self->data.klass == &java::lang::Object::class$) + continue; + + if (! other_iter->is_resolved) + other_iter->resolve(verifier); + + if (! is_assignable_from_slow (self->data.klass, + other_iter->data.klass)) + return false; + } + } + + return true; + } + + bool isarray () + { + // assert (ref_next == NULL); + if (is_resolved) + return data.klass->isArray (); + else + return data.name->first() == '['; + } + + bool isinterface (_Jv_BytecodeVerifier *verifier) + { + // assert (ref_next == NULL); + if (! is_resolved) + resolve (verifier); + return data.klass->isInterface (); + } + + bool isabstract (_Jv_BytecodeVerifier *verifier) + { + // assert (ref_next == NULL); + if (! is_resolved) + resolve (verifier); + using namespace java::lang::reflect; + return Modifier::isAbstract (data.klass->getModifiers ()); + } + + jclass getclass (_Jv_BytecodeVerifier *verifier) + { + if (! is_resolved) + resolve (verifier); + return data.klass; + } + + int count_dimensions () + { + int ndims = 0; + if (is_resolved) + { + jclass k = data.klass; + while (k->isArray ()) + { + k = k->getComponentType (); + ++ndims; + } + } + else + { + char *p = data.name->chars(); + while (*p++ == '[') + ++ndims; + } + return ndims; + } + + void *operator new (size_t bytes) + { + return _Jv_Malloc (bytes); + } + + void operator delete (void *mem) + { + _Jv_Free (mem); + } }; // Return the type_val corresponding to a primitive signature @@ -244,8 +548,21 @@ private: // TARGET haven't been prepared. static bool is_assignable_from_slow (jclass target, jclass source) { - // This will terminate when SOURCE==Object. - while (true) + // First, strip arrays. + while (target->isArray ()) + { + // If target is array, source must be as well. + if (! source->isArray ()) + return false; + target = target->getComponentType (); + source = source->getComponentType (); + } + + // Quick success. + if (target == &java::lang::Object::class$) + return true; + + do { if (source == target) return true; @@ -253,102 +570,56 @@ private: if (target->isPrimitive () || source->isPrimitive ()) return false; - if (target->isArray ()) - { - if (! source->isArray ()) - return false; - target = target->getComponentType (); - source = source->getComponentType (); - } - else if (target->isInterface ()) + if (target->isInterface ()) { for (int i = 0; i < source->interface_count; ++i) { // We use a recursive call because we also need to // check superinterfaces. - if (is_assignable_from_slow (target, source->interfaces[i])) - return true; - } - source = source->getSuperclass (); - if (source == NULL) - return false; - } - // We must do this check before we check to see if SOURCE is - // an interface. This way we know that any interface is - // assignable to an Object. - else if (target == &java::lang::Object::class$) - return true; - else if (source->isInterface ()) - { - for (int i = 0; i < target->interface_count; ++i) - { - // We use a recursive call because we also need to - // check superinterfaces. - if (is_assignable_from_slow (target->interfaces[i], source)) + if (is_assignable_from_slow (target, source->getInterface (i))) return true; } - target = target->getSuperclass (); - if (target == NULL) - return false; } - else if (source == &java::lang::Object::class$) - return false; - else - source = source->getSuperclass (); + source = source->getSuperclass (); } - } + while (source != NULL); - // This is used to keep track of which `jsr's correspond to a given - // jsr target. - struct subr_info - { - // PC of the instruction just after the jsr. - int pc; - // Link. - subr_info *next; - }; - - // This is used to keep track of which subroutine entry point - // corresponds to which `ret' instruction. - struct subr_entry_info - { - // PC of the subroutine entry point. - int pc; - // PC of the `ret' instruction. - int ret_pc; - // Link. - subr_entry_info *next; - }; + return false; + } // The `type' class is used to represent a single type in the // verifier. struct type { - // The type. + // The type key. type_val key; - // Some associated data. - union - { - // For a resolved reference type, this is a pointer to the class. - jclass klass; - // For other reference types, this it the name of the class. - _Jv_Utf8Const *name; - } data; - // This is used when constructing a new object. It is the PC of the + + // For reference types, the representation of the type. + ref_intersection *klass; + + // This is used in two situations. + // + // First, when constructing a new object, it is the PC of the // `new' instruction which created the object. We use the special - // value -2 to mean that this is uninitialized, and the special - // value -1 for the case where the current method is itself the - // method. + // value UNINIT to mean that this is uninitialized. The special + // value SELF is used for the case where the current method is + // itself the method. the special value EITHER is used + // when we may optionally allow either an uninitialized or + // initialized reference to match. + // + // Second, when the key is return_address_type, this holds the PC + // of the instruction following the `jsr'. int pc; static const int UNINIT = -2; static const int SELF = -1; + static const int EITHER = -3; // Basic constructor. type () { key = unsuitable_type; - data.klass = NULL; + klass = NULL; pc = UNINIT; } @@ -357,25 +628,26 @@ private: type (type_val k) { key = k; - data.klass = NULL; - if (key == reference_type) - data.klass = &java::lang::Object::class$; + // For reference_type, if KLASS==NULL then that means we are + // looking for a generic object of any kind, including an + // uninitialized reference. + klass = NULL; pc = UNINIT; } // Make a new instance given a class. - type (jclass klass) + type (jclass k, _Jv_BytecodeVerifier *verifier) { key = reference_type; - data.klass = klass; + klass = new ref_intersection (k, verifier); pc = UNINIT; } // Make a new instance given the name of a class. - type (_Jv_Utf8Const *n) + type (_Jv_Utf8Const *n, _Jv_BytecodeVerifier *verifier) { - key = unresolved_reference_type; - data.name = n; + key = reference_type; + klass = new ref_intersection (n, verifier); pc = UNINIT; } @@ -383,7 +655,7 @@ private: type (const type &t) { key = t.key; - data = t.data; + klass = t.klass; pc = t.pc; } @@ -402,7 +674,7 @@ private: type& operator= (type_val k) { key = k; - data.klass = NULL; + klass = NULL; pc = UNINIT; return *this; } @@ -410,7 +682,7 @@ private: type& operator= (const type& t) { key = t.key; - data = t.data; + klass = t.klass; pc = t.pc; return *this; } @@ -424,35 +696,11 @@ private: return *this; } - // If *THIS is an unresolved reference type, resolve it. - void resolve (_Jv_BytecodeVerifier *verifier) - { - if (key != unresolved_reference_type - && key != uninitialized_unresolved_reference_type) - return; - - using namespace java::lang; - java::lang::ClassLoader *loader - = verifier->current_class->getClassLoaderInternal(); - // We might see either kind of name. Sigh. - if (data.name->data[0] == 'L' - && data.name->data[data.name->length - 1] == ';') - data.klass = _Jv_FindClassFromSignature (data.name->data, loader); - else - data.klass = Class::forName (_Jv_NewStringUtf8Const (data.name), - false, loader); - key = (key == unresolved_reference_type - ? reference_type - : uninitialized_reference_type); - } - // Mark this type as the uninitialized result of `new'. void set_uninitialized (int npc, _Jv_BytecodeVerifier *verifier) { if (key == reference_type) key = uninitialized_reference_type; - else if (key == unresolved_reference_type) - key = uninitialized_unresolved_reference_type; else verifier->verify_fail ("internal error in type::uninitialized"); pc = npc; @@ -461,17 +709,30 @@ private: // Mark this type as now initialized. void set_initialized (int npc) { - if (npc != UNINIT && pc == npc - && (key == uninitialized_reference_type - || key == uninitialized_unresolved_reference_type)) + if (npc != UNINIT && pc == npc && key == uninitialized_reference_type) { - key = (key == uninitialized_reference_type - ? reference_type - : unresolved_reference_type); + key = reference_type; pc = UNINIT; } } + // Mark this type as a particular return address. + void set_return_address (int npc) + { + pc = npc; + } + + // Return true if this type and type OTHER are considered + // mergeable for the purposes of state merging. This is related + // to subroutine handling. For this purpose two types are + // considered unmergeable if they are both return-addresses but + // have different PCs. + bool state_mergeable_p (const type &other) const + { + return (key != return_address_type + || other.key != return_address_type + || pc == other.pc); + } // Return true if an object of type K can be assigned to a variable // of type *THIS. Handle various special cases too. Might modify @@ -488,39 +749,60 @@ private: // The `null' type is convertible to any initialized reference // type. - if (key == null_type || k.key == null_type) - return true; + if (key == null_type) + return k.key != uninitialized_reference_type; + if (k.key == null_type) + return key != uninitialized_reference_type; - // Any reference type is convertible to Object. This is a special - // case so we don't need to unnecessarily resolve a class. - if (key == reference_type - && data.klass == &java::lang::Object::class$) + // A special case for a generic reference. + if (klass == NULL) return true; + if (k.klass == NULL) + verifier->verify_fail ("programmer error in type::compatible"); - // An initialized type and an uninitialized type are not - // compatible. - if (isinitialized () != k.isinitialized ()) - return false; - - // Two uninitialized objects are compatible if either: - // * The PCs are identical, or - // * One PC is UNINIT. - if (! isinitialized ()) + // Handle the special 'EITHER' case, which is only used in a + // special case of 'putfield'. Note that we only need to handle + // this on the LHS of a check. + if (! isinitialized () && pc == EITHER) { - if (pc != k.pc && pc != UNINIT && k.pc != UNINIT) + // If the RHS is uninitialized, it must be an uninitialized + // 'this'. + if (! k.isinitialized () && k.pc != SELF) return false; } + else if (isinitialized () != k.isinitialized ()) + { + // An initialized type and an uninitialized type are not + // otherwise compatible. + return false; + } + else + { + // Two uninitialized objects are compatible if either: + // * The PCs are identical, or + // * One PC is UNINIT. + if (! isinitialized ()) + { + if (pc != k.pc && pc != UNINIT && k.pc != UNINIT) + return false; + } + } - // Two unresolved types are equal if their names are the same. - if (! isresolved () - && ! k.isresolved () - && _Jv_equalUtf8Consts (data.name, k.data.name)) - return true; + return klass->compatible(k.klass, verifier); + } - // We must resolve both types and check assignability. - resolve (verifier); - k.resolve (verifier); - return is_assignable_from_slow (data.klass, k.data.klass); + bool equals (const type &other, _Jv_BytecodeVerifier *vfy) + { + // Only works for reference types. + if ((key != reference_type + && key != uninitialized_reference_type) + || (other.key != reference_type + && other.key != uninitialized_reference_type)) + return false; + // Only for single-valued types. + if (klass->ref_next || other.klass->ref_next) + return false; + return klass->equals (other.klass, vfy); } bool isvoid () const @@ -545,9 +827,7 @@ private: // We treat null_type as not an array. This is ok based on the // current uses of this method. if (key == reference_type) - return data.klass->isArray (); - else if (key == unresolved_reference_type) - return data.name->data[0] == '['; + return klass->isarray (); return false; } @@ -558,33 +838,28 @@ private: bool isinterface (_Jv_BytecodeVerifier *verifier) { - resolve (verifier); if (key != reference_type) return false; - return data.klass->isInterface (); + return klass->isinterface (verifier); } bool isabstract (_Jv_BytecodeVerifier *verifier) { - resolve (verifier); if (key != reference_type) return false; - using namespace java::lang::reflect; - return Modifier::isAbstract (data.klass->getModifiers ()); + return klass->isabstract (verifier); } // Return the element type of an array. type element_type (_Jv_BytecodeVerifier *verifier) { - // FIXME: maybe should do string manipulation here. - resolve (verifier); if (key != reference_type) verifier->verify_fail ("programmer error in type::element_type()", -1); - jclass k = data.klass->getComponentType (); + jclass k = klass->getclass (verifier)->getComponentType (); if (k->isPrimitive ()) return type (verifier->get_type_val_for_signature (k)); - return type (k); + return type (k, verifier); } // Return the array type corresponding to an initialized @@ -592,16 +867,73 @@ private: // types, but currently we don't need to. type to_array (_Jv_BytecodeVerifier *verifier) { - // Resolving isn't ideal, because it might force us to load - // another class, but it's easy. FIXME? - if (key == unresolved_reference_type) - resolve (verifier); + if (key != reference_type) + verifier->verify_fail ("internal error in type::to_array()"); - if (key == reference_type) - return type (_Jv_GetArrayClass (data.klass, - data.klass->getClassLoaderInternal())); + // In case the class is already resolved we can simply ask the runtime + // to give us the array version. + // If it is not resolved we prepend "[" to the classname to make the + // array usage verification more lazy. In other words: makes new Foo[300] + // pass the verifier if Foo.class is missing. + if (klass->is_resolved) + { + jclass k = klass->getclass (verifier); + + return type (_Jv_GetArrayClass (k, k->getClassLoaderInternal()), + verifier); + } else - verifier->verify_fail ("internal error in type::to_array()"); + { + int len = klass->data.name->len(); + + // If the classname is given in the Lp1/p2/cn; format we only need + // to add a leading '['. The same procedure has to be done for + // primitive arrays (ie. provided "[I", the result should be "[[I". + // If the classname is given as p1.p2.cn we have to embed it into + // "[L" and ';'. + if (klass->data.name->limit()[-1] == ';' || + _Jv_isPrimitiveOrDerived(klass->data.name)) + { + // Reserves space for leading '[' and trailing '\0' . + char arrayName[len + 2]; + + arrayName[0] = '['; + strcpy(&arrayName[1], klass->data.name->chars()); + +#ifdef VERIFY_DEBUG + // This is only needed when we want to print the string to the + // screen while debugging. + arrayName[len + 1] = '\0'; + + debug_print("len: %d - old: '%s' - new: '%s'\n", len, klass->data.name->chars(), arrayName); +#endif + + return type (verifier->make_utf8_const( arrayName, len + 1 ), + verifier); + } + else + { + // Reserves space for leading "[L" and trailing ';' and '\0' . + char arrayName[len + 4]; + + arrayName[0] = '['; + arrayName[1] = 'L'; + strcpy(&arrayName[2], klass->data.name->chars()); + arrayName[len + 2] = ';'; + +#ifdef VERIFY_DEBUG + // This is only needed when we want to print the string to the + // screen while debugging. + arrayName[len + 3] = '\0'; + + debug_print("len: %d - old: '%s' - new: '%s'\n", len, klass->data.name->chars(), arrayName); +#endif + + return type (verifier->make_utf8_const( arrayName, len + 3 ), + verifier); + } + } + } bool isreference () const @@ -616,9 +948,7 @@ private: bool isinitialized () const { - return (key == reference_type - || key == null_type - || key == unresolved_reference_type); + return key == reference_type || key == null_type; } bool isresolved () const @@ -631,28 +961,17 @@ private: void verify_dimensions (int ndims, _Jv_BytecodeVerifier *verifier) { // The way this is written, we don't need to check isarray(). - if (key == reference_type) - { - jclass k = data.klass; - while (k->isArray () && ndims > 0) - { - k = k->getComponentType (); - --ndims; - } - } - else - { - // We know KEY == unresolved_reference_type. - char *p = data.name->data; - while (*p++ == '[' && ndims-- > 0) - ; - } + if (key != reference_type) + verifier->verify_fail ("internal error in verify_dimensions:" + " not a reference type"); - if (ndims > 0) - verifier->verify_fail ("array type has fewer dimensions than required"); + if (klass->count_dimensions () < ndims) + verifier->verify_fail ("array type has fewer dimensions" + " than required"); } - // Merge OLD_TYPE into this. On error throw exception. + // Merge OLD_TYPE into this. On error throw exception. Return + // true if the merge caused a type change. bool merge (type& old_type, bool local_semantics, _Jv_BytecodeVerifier *verifier) { @@ -682,53 +1001,12 @@ private: verifier->verify_fail ("merging different uninitialized types"); } - if (! isresolved () - && ! old_type.isresolved () - && _Jv_equalUtf8Consts (data.name, old_type.data.name)) + ref_intersection *merged = old_type.klass->merge (klass, + verifier); + if (merged != klass) { - // Types are identical. - } - else - { - resolve (verifier); - old_type.resolve (verifier); - - jclass k = data.klass; - jclass oldk = old_type.data.klass; - - int arraycount = 0; - while (k->isArray () && oldk->isArray ()) - { - ++arraycount; - k = k->getComponentType (); - oldk = oldk->getComponentType (); - } - - // Ordinarily this terminates when we hit Object... - while (k != NULL) - { - if (is_assignable_from_slow (k, oldk)) - break; - k = k->getSuperclass (); - changed = true; - } - // ... but K could have been an interface, in which - // case we'll end up here. We just convert this - // into Object. - if (k == NULL) - k = &java::lang::Object::class$; - - if (changed) - { - while (arraycount > 0) - { - java::lang::ClassLoader *loader - = verifier->current_class->getClassLoaderInternal(); - k = _Jv_GetArrayClass (k, loader); - --arraycount; - } - data.klass = k; - } + klass = merged; + changed = true; } } } @@ -736,20 +1014,9 @@ private: { if (local_semantics) { - // If we're merging into an "unused" slot, then we - // simply accept whatever we're merging from. - if (key == unused_by_subroutine_type) - { - *this = old_type; - changed = true; - } - else if (old_type.key == unused_by_subroutine_type) - { - // Do nothing. - } // If we already have an `unsuitable' type, then we // don't need to change again. - else if (key != unsuitable_type) + if (key != unsuitable_type) { key = unsuitable_type; changed = true; @@ -779,12 +1046,9 @@ private: case unsuitable_type: c = '-'; break; case return_address_type: c = 'r'; break; case continuation_type: c = '+'; break; - case unused_by_subroutine_type: c = '_'; break; case reference_type: c = 'L'; break; case null_type: c = '@'; break; - case unresolved_reference_type: c = 'l'; break; case uninitialized_reference_type: c = 'U'; break; - case uninitialized_unresolved_reference_type: c = 'u'; break; } debug_print ("%c", c); } @@ -804,16 +1068,6 @@ private: type *stack; // The local variables. type *locals; - // This is used in subroutines to keep track of which local - // variables have been accessed. - bool *local_changed; - // If not 0, then we are in a subroutine. The value is the PC of - // the subroutine's entry point. We can use 0 as an exceptional - // value because PC=0 can never be a subroutine. - int subroutine; - // This is used to keep a linked list of all the states which - // require re-verification. We use the PC to keep track. - int next; // We keep track of the type of `this' specially. This is used to // ensure that an instance initializer invokes another initializer // on `this' before returning. We must keep track of this @@ -821,31 +1075,27 @@ private: // assigns to locals[0] (overwriting `this') and then returns // without really initializing. type this_type; - // This is a list of all subroutines that have been seen at this - // point. Ordinarily this is NULL; it is only allocated and used - // in relatively weird situations involving non-ret exit from a - // subroutine. We have to keep track of this in this way to avoid - // endless recursion in these cases. - subr_info *seen_subrs; - - // INVALID marks a state which is not on the linked list of states - // requiring reverification. - static const int INVALID = -1; - // NO_NEXT marks the state at the end of the reverification list. - static const int NO_NEXT = -2; - - // This is used to mark the stack depth at the instruction just - // after a `jsr' when we haven't yet processed the corresponding - // `ret'. See handle_jsr_insn for more information. - static const int NO_STACK = -1; + + // The PC for this state. This is only valid on states which are + // permanently attached to a given PC. For an object like + // `current_state', which is used transiently, this has no + // meaning. + int pc; + // We keep a linked list of all states requiring reverification. + // If this is the special value INVALID_STATE then this state is + // not on the list. NULL marks the end of the linked list. + state *next; + + // NO_NEXT is the PC value meaning that a new state must be + // acquired from the verification list. + static const int NO_NEXT = -1; state () : this_type () { stack = NULL; locals = NULL; - local_changed = NULL; - seen_subrs = NULL; + next = INVALID_STATE; } state (int max_stack, int max_locals) @@ -857,26 +1107,19 @@ private: for (int i = 0; i < max_stack; ++i) stack[i] = unsuitable_type; locals = new type[max_locals]; - local_changed = (bool *) _Jv_Malloc (sizeof (bool) * max_locals); - seen_subrs = NULL; for (int i = 0; i < max_locals; ++i) - { - locals[i] = unsuitable_type; - local_changed[i] = false; - } - next = INVALID; - subroutine = 0; + locals[i] = unsuitable_type; + pc = NO_NEXT; + next = INVALID_STATE; } - state (const state *orig, int max_stack, int max_locals, - bool ret_semantics = false) + state (const state *orig, int max_stack, int max_locals) { stack = new type[max_stack]; locals = new type[max_locals]; - local_changed = (bool *) _Jv_Malloc (sizeof (bool) * max_locals); - seen_subrs = NULL; - copy (orig, max_stack, max_locals, ret_semantics); - next = INVALID; + copy (orig, max_stack, max_locals); + pc = NO_NEXT; + next = INVALID_STATE; } ~state () @@ -885,9 +1128,6 @@ private: delete[] stack; if (locals) delete[] locals; - if (local_changed) - _Jv_Free (local_changed); - clean_subrs (); } void *operator new[] (size_t bytes) @@ -910,48 +1150,17 @@ private: _Jv_Free (mem); } - void clean_subrs () - { - subr_info *info = seen_subrs; - while (info != NULL) - { - subr_info *next = info->next; - _Jv_Free (info); - info = next; - } - } - - void copy (const state *copy, int max_stack, int max_locals, - bool ret_semantics = false) + void copy (const state *copy, int max_stack, int max_locals) { stacktop = copy->stacktop; stackdepth = copy->stackdepth; - subroutine = copy->subroutine; for (int i = 0; i < max_stack; ++i) stack[i] = copy->stack[i]; for (int i = 0; i < max_locals; ++i) - { - // See push_jump_merge to understand this case. - if (ret_semantics) - locals[i] = type (copy->local_changed[i] - ? unsuitable_type - : unused_by_subroutine_type); - else - locals[i] = copy->locals[i]; - local_changed[i] = copy->local_changed[i]; - } - - clean_subrs (); - if (copy->seen_subrs) - { - for (subr_info *info = seen_subrs; info != NULL; info = info->next) - add_subr (info->pc); - } - else - seen_subrs = NULL; + locals[i] = copy->locals[i]; this_type = copy->this_type; - // Don't modify `next'. + // Don't modify `next' or `pc'. } // Modify this state to reflect entry to an exception handler. @@ -964,32 +1173,21 @@ private: stack[i] = unsuitable_type; } - // Modify this state to reflect entry into a subroutine. - void enter_subroutine (int npc, int max_locals) + inline int get_pc () const { - subroutine = npc; - // Mark all items as unchanged. Each subroutine needs to keep - // track of its `changed' state independently. In the case of - // nested subroutines, this information will be merged back into - // parent by the `ret'. - for (int i = 0; i < max_locals; ++i) - local_changed[i] = false; + return pc; } - // Indicate that we've been in this this subroutine. - void add_subr (int pc) + void set_pc (int npc) { - subr_info *n = (subr_info *) _Jv_Malloc (sizeof (subr_info)); - n->pc = pc; - n->next = seen_subrs; - seen_subrs = n; + pc = npc; } // Merge STATE_OLD into this state. Destructively modifies this // state. Returns true if the new state was in fact changed. // Will throw an exception if the states are not mergeable. - bool merge (state *state_old, bool ret_semantics, - int max_locals, _Jv_BytecodeVerifier *verifier) + bool merge (state *state_old, int max_locals, + _Jv_BytecodeVerifier *verifier) { bool changed = false; @@ -998,118 +1196,25 @@ private: if (this_type.isinitialized ()) this_type = state_old->this_type; - // Merge subroutine states. Here we just keep track of what - // subroutine we think we're in. We only check for a merge - // (which is invalid) when we see a `ret'. - if (subroutine == state_old->subroutine) - { - // Nothing. - } - else if (subroutine == 0) - { - subroutine = state_old->subroutine; - changed = true; - } - else - { - // If the subroutines differ, and we haven't seen this - // subroutine before, indicate that the state changed. This - // is needed to detect when subroutines have merged. - bool found = false; - for (subr_info *info = seen_subrs; info != NULL; info = info->next) - { - if (info->pc == state_old->subroutine) - { - found = true; - break; - } - } - if (! found) - { - add_subr (state_old->subroutine); - changed = true; - } - } - - // Merge stacks. Special handling for NO_STACK case. - if (state_old->stacktop == NO_STACK) - { - // Nothing to do in this case; we don't care about modifying - // the old state. - } - else if (stacktop == NO_STACK) - { - stacktop = state_old->stacktop; - stackdepth = state_old->stackdepth; - for (int i = 0; i < stacktop; ++i) - stack[i] = state_old->stack[i]; - changed = true; - } - else if (state_old->stacktop != stacktop) + // Merge stacks. + if (state_old->stacktop != stacktop) // FIXME stackdepth instead? verifier->verify_fail ("stack sizes differ"); - else + for (int i = 0; i < state_old->stacktop; ++i) { - for (int i = 0; i < state_old->stacktop; ++i) - { - if (stack[i].merge (state_old->stack[i], false, verifier)) - changed = true; - } + if (stack[i].merge (state_old->stack[i], false, verifier)) + changed = true; } // Merge local variables. for (int i = 0; i < max_locals; ++i) { - // If we're not processing a `ret', then we merge every - // local variable. If we are processing a `ret', then we - // only merge locals which changed in the subroutine. When - // processing a `ret', STATE_OLD is the state at the point - // of the `ret', and THIS is the state just after the `jsr'. - if (! ret_semantics || state_old->local_changed[i]) - { - if (locals[i].merge (state_old->locals[i], true, verifier)) - { - // Note that we don't call `note_variable' here. - // This change doesn't represent a real change to a - // local, but rather a merge artifact. If we're in - // a subroutine which is called with two - // incompatible types in a slot that is unused by - // the subroutine, then we don't want to mark that - // variable as having been modified. - changed = true; - } - } - - // If we're in a subroutine, we must compute the union of - // all the changed local variables. - if (state_old->local_changed[i]) - note_variable (i); + if (locals[i].merge (state_old->locals[i], true, verifier)) + changed = true; } return changed; } - // Throw an exception if there is an uninitialized object on the - // stack or in a local variable. EXCEPTION_SEMANTICS controls - // whether we're using backwards-branch or exception-handing - // semantics. - void check_no_uninitialized_objects (int max_locals, - _Jv_BytecodeVerifier *verifier, - bool exception_semantics = false) - { - if (! exception_semantics) - { - for (int i = 0; i < stacktop; ++i) - if (stack[i].isreference () && ! stack[i].isinitialized ()) - verifier->verify_fail ("uninitialized object on stack"); - } - - for (int i = 0; i < max_locals; ++i) - if (locals[i].isreference () && ! locals[i].isinitialized ()) - verifier->verify_fail ("uninitialized object in local variable"); - - check_this_initialized (verifier); - } - // Ensure that `this' has been initialized. void check_this_initialized (_Jv_BytecodeVerifier *verifier) { @@ -1123,13 +1228,6 @@ private: this_type = k; } - // Note that a local variable was modified. - void note_variable (int index) - { - if (subroutine > 0) - local_changed[index] = true; - } - // Mark each `new'd object we know of that was allocated at PC as // initialized. void set_initialized (int pc, int max_locals) @@ -1141,17 +1239,36 @@ private: this_type.set_initialized (pc); } - // Return true if this state is the unmerged result of a `ret'. - bool is_unmerged_ret_state (int max_locals) const + // This tests to see whether two states can be considered "merge + // compatible". If both states have a return-address in the same + // slot, and the return addresses are different, then they are not + // compatible and we must not try to merge them. + bool state_mergeable_p (state *other, int max_locals, + _Jv_BytecodeVerifier *verifier) { - if (stacktop == NO_STACK) - return true; + // This is tricky: if the stack sizes differ, then not only are + // these not mergeable, but in fact we should give an error, as + // we've found two execution paths that reach a branch target + // with different stack depths. FIXME stackdepth instead? + if (stacktop != other->stacktop) + verifier->verify_fail ("stack sizes differ"); + + for (int i = 0; i < stacktop; ++i) + if (! stack[i].state_mergeable_p (other->stack[i])) + return false; for (int i = 0; i < max_locals; ++i) + if (! locals[i].state_mergeable_p (other->locals[i])) + return false; + return true; + } + + void reverify (_Jv_BytecodeVerifier *verifier) + { + if (next == INVALID_STATE) { - if (locals[i].key == unused_by_subroutine_type) - return true; + next = verifier->next_verify_state; + verifier->next_verify_state = this; } - return false; } #ifdef VERIFY_DEBUG @@ -1166,14 +1283,7 @@ private: debug_print ("."); debug_print (" [local] "); for (i = 0; i < max_locals; ++i) - { - locals[i].print (); - debug_print (local_changed[i] ? "+" : " "); - } - if (subroutine == 0) - debug_print (" | None"); - else - debug_print (" | %4d", subroutine); + locals[i].print (); debug_print (" | %p\n", this); } #else @@ -1254,18 +1364,11 @@ private: if (index > current_method->max_locals - depth) verify_fail ("invalid local variable"); current_state->locals[index] = t; - current_state->note_variable (index); if (depth == 2) - { - current_state->locals[index + 1] = continuation_type; - current_state->note_variable (index + 1); - } + current_state->locals[index + 1] = continuation_type; if (index > 0 && current_state->locals[index - 1].iswide ()) - { - current_state->locals[index - 1] = unsuitable_type; - // There's no need to call note_variable here. - } + current_state->locals[index - 1] = unsuitable_type; } type get_variable (int index, type t) @@ -1355,99 +1458,109 @@ private: return npc; } + // Add a new state to the state list at NPC. + state *add_new_state (int npc, state *old_state) + { + state *new_state = new state (old_state, current_method->max_stack, + current_method->max_locals); + debug_print ("== New state in add_new_state\n"); + new_state->print ("New", npc, current_method->max_stack, + current_method->max_locals); + linked *nlink + = (linked *) _Jv_Malloc (sizeof (linked)); + nlink->val = new_state; + nlink->next = states[npc]; + states[npc] = nlink; + new_state->set_pc (npc); + return new_state; + } + // Merge the indicated state into the state at the branch target and - // schedule a new PC if there is a change. If RET_SEMANTICS is - // true, then we are merging from a `ret' instruction into the - // instruction after a `jsr'. This is a special case with its own - // modified semantics. - void push_jump_merge (int npc, state *nstate, bool ret_semantics = false) + // schedule a new PC if there is a change. NPC is the PC of the + // branch target, and FROM_STATE is the state at the source of the + // branch. This method returns true if the destination state + // changed and requires reverification, false otherwise. + void merge_into (int npc, state *from_state) { - bool changed = true; - if (states[npc] == NULL) + // Iterate over all target states and merge our state into each, + // if applicable. FIXME one improvement we could make here is + // "state destruction". Merging a new state into an existing one + // might cause a return_address_type to be merged to + // unsuitable_type. In this case the resulting state may now be + // mergeable with other states currently held in parallel at this + // location. So in this situation we could pairwise compare and + // reduce the number of parallel states. + bool applicable = false; + for (linked *iter = states[npc]; iter != NULL; iter = iter->next) { - // There's a weird situation here. If are examining the - // branch that results from a `ret', and there is not yet a - // state available at the branch target (the instruction just - // after the `jsr'), then we have to construct a special kind - // of state at that point for future merging. This special - // state has the type `unused_by_subroutine_type' in each slot - // which was not modified by the subroutine. - states[npc] = new state (nstate, current_method->max_stack, - current_method->max_locals, ret_semantics); - debug_print ("== New state in push_jump_merge\n"); - states[npc]->print ("New", npc, current_method->max_stack, - current_method->max_locals); - } - else - { - debug_print ("== Merge states in push_jump_merge\n"); - nstate->print ("Frm", start_PC, current_method->max_stack, - current_method->max_locals); - states[npc]->print (" To", npc, current_method->max_stack, - current_method->max_locals); - changed = states[npc]->merge (nstate, ret_semantics, - current_method->max_locals, this); - states[npc]->print ("New", npc, current_method->max_stack, - current_method->max_locals); + state *new_state = iter->val; + if (new_state->state_mergeable_p (from_state, + current_method->max_locals, this)) + { + applicable = true; + + debug_print ("== Merge states in merge_into\n"); + from_state->print ("Frm", start_PC, current_method->max_stack, + current_method->max_locals); + new_state->print (" To", npc, current_method->max_stack, + current_method->max_locals); + bool changed = new_state->merge (from_state, + current_method->max_locals, + this); + new_state->print ("New", npc, current_method->max_stack, + current_method->max_locals); + + if (changed) + new_state->reverify (this); + } } - if (changed && states[npc]->next == state::INVALID) + if (! applicable) { - // The merge changed the state, and the new PC isn't yet on our - // list of PCs to re-verify. - states[npc]->next = next_verify_pc; - next_verify_pc = npc; + // Either we don't yet have a state at NPC, or we have a + // return-address type that is in conflict with all existing + // state. So, we need to create a new entry. + state *new_state = add_new_state (npc, from_state); + // A new state added in this way must always be reverified. + new_state->reverify (this); } } void push_jump (int offset) { int npc = compute_jump (offset); - if (npc < PC) - current_state->check_no_uninitialized_objects (current_method->max_locals, this); - push_jump_merge (npc, current_state); + // According to the JVM Spec, we need to check for uninitialized + // objects here. However, this does not actually affect type + // safety, and the Eclipse java compiler generates code that + // violates this constraint. + merge_into (npc, current_state); } void push_exception_jump (type t, int pc) { - current_state->check_no_uninitialized_objects (current_method->max_locals, - this, true); + // According to the JVM Spec, we need to check for uninitialized + // objects here. However, this does not actually affect type + // safety, and the Eclipse java compiler generates code that + // violates this constraint. state s (current_state, current_method->max_stack, current_method->max_locals); if (current_method->max_stack < 1) verify_fail ("stack overflow at exception handler"); s.set_exception (t, current_method->max_stack); - push_jump_merge (pc, &s); + merge_into (pc, &s); } - int pop_jump () + state *pop_jump () { - int *prev_loc = &next_verify_pc; - int npc = next_verify_pc; - - while (npc != state::NO_NEXT) + state *new_state = next_verify_state; + if (new_state == INVALID_STATE) + verify_fail ("programmer error in pop_jump"); + if (new_state != NULL) { - // If the next available PC is an unmerged `ret' state, then - // we aren't yet ready to handle it. That's because we would - // need all kind of special cases to do so. So instead we - // defer this jump until after we've processed it via a - // fall-through. This has to happen because the instruction - // before this one must be a `jsr'. - if (! states[npc]->is_unmerged_ret_state (current_method->max_locals)) - { - *prev_loc = states[npc]->next; - states[npc]->next = state::INVALID; - return npc; - } - - prev_loc = &states[npc]->next; - npc = states[npc]->next; + next_verify_state = new_state->next; + new_state->next = INVALID_STATE; } - - // Note that we might have gotten here even when there are - // remaining states to process. That can happen if we find a - // `jsr' without a `ret'. - return state::NO_NEXT; + return new_state; } void invalidate_pc () @@ -1455,7 +1568,7 @@ private: PC = state::NO_NEXT; } - void note_branch_target (int pc, bool is_jsr_target = false) + void note_branch_target (int pc) { // Don't check `pc <= PC', because we've advanced PC after // fetching the target and we haven't yet checked the next @@ -1463,14 +1576,6 @@ private: if (pc < PC && ! (flags[pc] & FLAG_INSN_START)) verify_fail ("branch not to instruction start", start_PC); flags[pc] |= FLAG_BRANCH_TARGET; - if (is_jsr_target) - { - // Record the jsr which called this instruction. - subr_info *info = (subr_info *) _Jv_Malloc (sizeof (subr_info)); - info->pc = PC; - info->next = jsr_ptrs[pc]; - jsr_ptrs[pc] = info; - } } void skip_padding () @@ -1480,108 +1585,46 @@ private: verify_fail ("found nonzero padding byte"); } - // Return the subroutine to which the instruction at PC belongs. - int get_subroutine (int pc) - { - if (states[pc] == NULL) - return 0; - return states[pc]->subroutine; - } - // Do the work for a `ret' instruction. INDEX is the index into the // local variables. void handle_ret_insn (int index) { - get_variable (index, return_address_type); - - int csub = current_state->subroutine; - if (csub == 0) - verify_fail ("no subroutine"); - - // Check to see if we've merged subroutines. - subr_entry_info *entry; - for (entry = entry_points; entry != NULL; entry = entry->next) - { - if (entry->ret_pc == start_PC) - break; - } - if (entry == NULL) - { - entry = (subr_entry_info *) _Jv_Malloc (sizeof (subr_entry_info)); - entry->pc = csub; - entry->ret_pc = start_PC; - entry->next = entry_points; - entry_points = entry; - } - else if (entry->pc != csub) - verify_fail ("subroutines merged"); - - for (subr_info *subr = jsr_ptrs[csub]; subr != NULL; subr = subr->next) - { - // We might be returning to a `jsr' that is at the end of the - // bytecode. This is ok if we never return from the called - // subroutine, but if we see this here it is an error. - if (subr->pc >= current_method->code_length) - verify_fail ("fell off end"); - - // Temporarily modify the current state so it looks like we're - // in the enclosing context. - current_state->subroutine = get_subroutine (subr->pc); - if (subr->pc < PC) - current_state->check_no_uninitialized_objects (current_method->max_locals, this); - push_jump_merge (subr->pc, current_state, true); - } - - current_state->subroutine = csub; + type ret_addr = get_variable (index, return_address_type); + // It would be nice if we could do this. However, the JVM Spec + // doesn't say that this is what happens. It is implied that + // reusing a return address is invalid, but there's no actual + // prohibition against it. + // set_variable (index, unsuitable_type); + + int npc = ret_addr.get_pc (); + // We might be returning to a `jsr' that is at the end of the + // bytecode. This is ok if we never return from the called + // subroutine, but if we see this here it is an error. + if (npc >= current_method->code_length) + verify_fail ("fell off end"); + + // According to the JVM Spec, we need to check for uninitialized + // objects here. However, this does not actually affect type + // safety, and the Eclipse java compiler generates code that + // violates this constraint. + merge_into (npc, current_state); invalidate_pc (); } - // We're in the subroutine SUB, calling a subroutine at DEST. Make - // sure this subroutine isn't already on the stack. - void check_nonrecursive_call (int sub, int dest) - { - if (sub == 0) - return; - if (sub == dest) - verify_fail ("recursive subroutine call"); - for (subr_info *info = jsr_ptrs[sub]; info != NULL; info = info->next) - check_nonrecursive_call (get_subroutine (info->pc), dest); - } - void handle_jsr_insn (int offset) { int npc = compute_jump (offset); - if (npc < PC) - current_state->check_no_uninitialized_objects (current_method->max_locals, this); - check_nonrecursive_call (current_state->subroutine, npc); + // According to the JVM Spec, we need to check for uninitialized + // objects here. However, this does not actually affect type + // safety, and the Eclipse java compiler generates code that + // violates this constraint. // Modify our state as appropriate for entry into a subroutine. - push_type (return_address_type); - push_jump_merge (npc, current_state); - // Clean up. - pop_type (return_address_type); - - // On entry to the subroutine, the subroutine number must be set - // and the locals must be marked as cleared. We do this after - // merging state so that we don't erroneously "notice" a variable - // change merely on entry. - states[npc]->enter_subroutine (npc, current_method->max_locals); - - // Indicate that we don't know the stack depth of the instruction - // following the `jsr'. The idea here is that we need to merge - // the local variable state across the jsr, but the subroutine - // might change the stack depth, so we can't make any assumptions - // about it. So we have yet another special case. We know that - // at this point PC points to the instruction after the jsr. Note - // that it is ok to have a `jsr' at the end of the bytecode, - // provided that the called subroutine never returns. So, we have - // a special case here and another one when we handle the ret. - if (PC < current_method->code_length) - { - current_state->stacktop = state::NO_STACK; - push_jump_merge (PC, current_state); - } + type ret_addr (return_address_type); + ret_addr.set_return_address (PC); + push_type (ret_addr); + merge_into (npc, current_state); invalidate_pc (); } @@ -1621,12 +1664,9 @@ private: case unsuitable_type: case return_address_type: case continuation_type: - case unused_by_subroutine_type: case reference_type: case null_type: - case unresolved_reference_type: case uninitialized_reference_type: - case uninitialized_unresolved_reference_type: default: verify_fail ("unknown type in construct_primitive_array_type"); } @@ -1639,16 +1679,9 @@ private: void branch_prepass () { flags = (char *) _Jv_Malloc (current_method->code_length); - jsr_ptrs = (subr_info **) _Jv_Malloc (sizeof (subr_info *) - * current_method->code_length); for (int i = 0; i < current_method->code_length; ++i) - { - flags[i] = 0; - jsr_ptrs[i] = NULL; - } - - bool last_was_jsr = false; + flags[i] = 0; PC = 0; while (PC < current_method->code_length) @@ -1658,13 +1691,6 @@ private: start_PC = PC; flags[PC] |= FLAG_INSN_START; - // If the previous instruction was a jsr, then the next - // instruction is a branch target -- the branch being the - // corresponding `ret'. - if (last_was_jsr) - note_branch_target (PC); - last_was_jsr = false; - java_opcode opcode = (java_opcode) bytecode[PC++]; switch (opcode) { @@ -1858,8 +1884,6 @@ private: break; case op_jsr: - last_was_jsr = true; - // Fall through. case op_ifeq: case op_ifne: case op_iflt: @@ -1877,7 +1901,7 @@ private: case op_ifnull: case op_ifnonnull: case op_goto: - note_branch_target (compute_jump (get_short ()), last_was_jsr); + note_branch_target (compute_jump (get_short ())); break; case op_tableswitch: @@ -1924,10 +1948,8 @@ private: break; case op_jsr_w: - last_was_jsr = true; - // Fall through. case op_goto_w: - note_branch_target (compute_jump (get_int ()), last_was_jsr); + note_branch_target (compute_jump (get_int ())); break; // These are unused here, but we call them out explicitly @@ -1954,6 +1976,7 @@ private: case op_getstatic_4: case op_getstatic_8: case op_getstatic_a: + case op_breakpoint: default: verify_fail ("unrecognized instruction in branch_prepass", start_PC); @@ -1997,9 +2020,9 @@ private: check_pool_index (index); _Jv_Constants *pool = ¤t_class->constants; if (pool->tags[index] == JV_CONSTANT_ResolvedClass) - return type (pool->data[index].clazz); + return type (pool->data[index].clazz, this); else if (pool->tags[index] == JV_CONSTANT_Class) - return type (pool->data[index].utf8); + return type (pool->data[index].utf8, this); verify_fail ("expected class constant", start_PC); } @@ -2007,13 +2030,16 @@ private: { check_pool_index (index); _Jv_Constants *pool = ¤t_class->constants; - if (pool->tags[index] == JV_CONSTANT_ResolvedString - || pool->tags[index] == JV_CONSTANT_String) - return type (&java::lang::String::class$); - else if (pool->tags[index] == JV_CONSTANT_Integer) + int tag = pool->tags[index]; + if (tag == JV_CONSTANT_ResolvedString || tag == JV_CONSTANT_String) + return type (&java::lang::String::class$, this); + else if (tag == JV_CONSTANT_Integer) return type (int_type); - else if (pool->tags[index] == JV_CONSTANT_Float) + else if (tag == JV_CONSTANT_Float) return type (float_type); + else if (current_method->is_15 + && (tag == JV_CONSTANT_ResolvedClass || tag == JV_CONSTANT_Class)) + return type (&java::lang::Class::class$, this); verify_fail ("String, int, or float constant expected", start_PC); } @@ -2056,7 +2082,9 @@ private: } // Return field's type, compute class' type if requested. - type check_field_constant (int index, type *class_type = NULL) + // If PUTFIELD is true, use the special 'putfield' semantics. + type check_field_constant (int index, type *class_type = NULL, + bool putfield = false) { _Jv_Utf8Const *name, *field_type; type ct = handle_field_or_method (index, @@ -2064,9 +2092,29 @@ private: &name, &field_type); if (class_type) *class_type = ct; - if (field_type->data[0] == '[' || field_type->data[0] == 'L') - return type (field_type); - return get_type_val_for_signature (field_type->data[0]); + type result; + if (field_type->first() == '[' || field_type->first() == 'L') + result = type (field_type, this); + else + result = get_type_val_for_signature (field_type->first()); + + // We have an obscure special case here: we can use `putfield' on + // a field declared in this class, even if `this' has not yet been + // initialized. + if (putfield + && ! current_state->this_type.isinitialized () + && current_state->this_type.pc == type::SELF + && current_state->this_type.equals (ct, this) + // We don't look at the signature, figuring that if it is + // wrong we will fail during linking. FIXME? + && _Jv_Linker::has_field_p (current_class, name)) + // Note that we don't actually know whether we're going to match + // against 'this' or some other object of the same type. So, + // here we set things up so that it doesn't matter. This relies + // on knowing what our caller is up to. + class_type->set_uninitialized (type::EITHER, this); + + return result; } type check_method_constant (int index, bool is_interface, @@ -2099,7 +2147,7 @@ private: ++p; ++p; _Jv_Utf8Const *name = make_utf8_const (start, p - start); - return type (name); + return type (name, this); } // Casting to jchar here is ok since we are looking at an ASCII @@ -2116,13 +2164,14 @@ private: jclass k = construct_primitive_array_type (rt); while (--arraycount > 0) k = _Jv_GetArrayClass (k, NULL); - return type (k); + return type (k, this); } void compute_argument_types (_Jv_Utf8Const *signature, type *types) { - char *p = signature->data; + char *p = signature->chars(); + // Skip `('. ++p; @@ -2133,7 +2182,7 @@ private: type compute_return_type (_Jv_Utf8Const *signature) { - char *p = signature->data; + char *p = signature->chars(); while (*p != ')') ++p; ++p; @@ -2160,7 +2209,7 @@ private: using namespace java::lang::reflect; if (! Modifier::isStatic (current_method->self->accflags)) { - type kurr (current_class); + type kurr (current_class, this); if (is_init) { kurr.set_uninitialized (type::SELF, this); @@ -2204,77 +2253,62 @@ private: // True if we are verifying an instance initializer. bool this_is_init = initialize_stack (); - states = (state **) _Jv_Malloc (sizeof (state *) - * current_method->code_length); + states = (linked **) _Jv_Malloc (sizeof (linked *) + * current_method->code_length); for (int i = 0; i < current_method->code_length; ++i) states[i] = NULL; - next_verify_pc = state::NO_NEXT; + next_verify_state = NULL; while (true) { // If the PC was invalidated, get a new one from the work list. if (PC == state::NO_NEXT) { - PC = pop_jump (); - if (PC == state::INVALID) - verify_fail ("can't happen: saw state::INVALID"); - if (PC == state::NO_NEXT) + state *new_state = pop_jump (); + // If it is null, we're done. + if (new_state == NULL) break; + + PC = new_state->get_pc (); debug_print ("== State pop from pending list\n"); // Set up the current state. - current_state->copy (states[PC], current_method->max_stack, + current_state->copy (new_state, current_method->max_stack, current_method->max_locals); } else { - // Control can't fall off the end of the bytecode. We - // only need to check this in the fall-through case, - // because branch bounds are checked when they are - // pushed. - if (PC >= current_method->code_length) - verify_fail ("fell off end"); - // We only have to do this checking in the situation where // control flow falls through from the previous // instruction. Otherwise merging is done at the time we - // push the branch. - if (states[PC] != NULL) + // push the branch. Note that we'll catch the + // off-the-end problem just below. + if (PC < current_method->code_length && states[PC] != NULL) { // We've already visited this instruction. So merge - // the states together. If this yields no change then - // we don't have to re-verify. However, if the new - // state is an the result of an unmerged `ret', we - // must continue through it. - debug_print ("== Fall through merge\n"); - states[PC]->print ("Old", PC, current_method->max_stack, - current_method->max_locals); - current_state->print ("Cur", PC, current_method->max_stack, - current_method->max_locals); - if (! current_state->merge (states[PC], false, - current_method->max_locals, this) - && ! states[PC]->is_unmerged_ret_state (current_method->max_locals)) - { - debug_print ("== Fall through optimization\n"); - invalidate_pc (); - continue; - } - // Save a copy of it for later. - states[PC]->copy (current_state, current_method->max_stack, - current_method->max_locals); - current_state->print ("New", PC, current_method->max_stack, - current_method->max_locals); + // the states together. It is simplest, but not most + // efficient, to just always invalidate the PC here. + merge_into (PC, current_state); + invalidate_pc (); + continue; } } + // Control can't fall off the end of the bytecode. We need to + // check this in both cases, not just the fall-through case, + // because we don't check to see whether a `jsr' appears at + // the end of the bytecode until we process a `ret'. + if (PC >= current_method->code_length) + verify_fail ("fell off end"); + // We only have to keep saved state at branch targets. If // we're at a branch target and the state here hasn't been set - // yet, we set it now. + // yet, we set it now. You might notice that `ret' targets + // won't necessarily have FLAG_BRANCH_TARGET set. This + // doesn't matter, since those states will be filled in by + // merge_into. if (states[PC] == NULL && (flags[PC] & FLAG_BRANCH_TARGET)) - { - states[PC] = new state (current_state, current_method->max_stack, - current_method->max_locals); - } + add_new_state (PC, current_state); // Set this before handling exceptions so that debug output is // sane. @@ -2287,7 +2321,7 @@ private: { if (PC >= exception[i].start_pc.i && PC < exception[i].end_pc.i) { - type handler (&java::lang::Throwable::class$); + type handler (&java::lang::Throwable::class$, this); if (exception[i].handler_type.i != 0) handler = check_class_constant (exception[i].handler_type.i); push_exception_jump (handler, exception[i].handler_pc.i); @@ -2891,15 +2925,8 @@ private: case op_putfield: { type klass; - type field = check_field_constant (get_ushort (), &klass); + type field = check_field_constant (get_ushort (), &klass, true); pop_type (field); - - // We have an obscure special case here: we can use - // `putfield' on a field declared in this class, even if - // `this' has not yet been initialized. - if (! current_state->this_type.isinitialized () - && current_state->this_type.pc == type::SELF) - klass.set_uninitialized (type::SELF, this); pop_type (klass); } break; @@ -2933,7 +2960,7 @@ private: if (opcode != op_invokespecial) verify_fail ("can't invoke "); } - else if (method_name->data[0] == '<') + else if (method_name->first() == '<') verify_fail ("can't invoke method starting with `<'"); // Pop arguments and check types. @@ -2959,33 +2986,13 @@ private: { // In this case the PC doesn't matter. t.set_uninitialized (type::UNINIT, this); + // FIXME: check to make sure that the + // call is to the right class. + // It must either be super or an exact class + // match. } type raw = pop_raw (); - bool ok = false; - if (! is_init && ! raw.isinitialized ()) - { - // This is a failure. - } - else if (is_init && raw.isnull ()) - { - // Another failure. - } - else if (t.compatible (raw, this)) - { - ok = true; - } - else if (opcode == op_invokeinterface) - { - // This is a hack. We might have merged two - // items and gotten `Object'. This can happen - // because we don't keep track of where merges - // come from. This is safe as long as the - // interpreter checks interfaces at runtime. - type obj (&java::lang::Object::class$); - ok = raw.compatible (obj, this); - } - - if (! ok) + if (! t.compatible (raw, this)) verify_fail ("incompatible type on stack"); if (is_init) @@ -3002,8 +3009,8 @@ private: case op_new: { type t = check_class_constant (get_ushort ()); - if (t.isarray () || t.isinterface (this) || t.isabstract (this)) - verify_fail ("type is array, interface, or abstract"); + if (t.isarray ()) + verify_fail ("type is array"); t.set_uninitialized (start_PC, this); push_type (t); } @@ -3017,7 +3024,8 @@ private: if (atype < boolean_type || atype > long_type) verify_fail ("type not primitive", start_PC); pop_type (int_type); - push_type (construct_primitive_array_type (type_val (atype))); + type t (construct_primitive_array_type (type_val (atype)), this); + push_type (t); } break; case op_anewarray: @@ -3033,7 +3041,7 @@ private: } break; case op_athrow: - pop_type (type (&java::lang::Throwable::class$)); + pop_type (type (&java::lang::Throwable::class$, this)); invalidate_pc (); break; case op_checkcast: @@ -3146,6 +3154,7 @@ private: case op_getstatic_4: case op_getstatic_8: case op_getstatic_a: + case op_breakpoint: default: // Unrecognized opcode. verify_fail ("unrecognized instruction in verify_instructions_0", @@ -3167,7 +3176,7 @@ public: // We just print the text as utf-8. This is just for debugging // anyway. debug_print ("--------------------------------\n"); - debug_print ("-- Verifying method `%s'\n", m->self->name->data); + debug_print ("-- Verifying method `%s'\n", m->self->name->chars()); current_method = m; bytecode = m->bytecode (); @@ -3176,49 +3185,43 @@ public: states = NULL; flags = NULL; - jsr_ptrs = NULL; utf8_list = NULL; - entry_points = NULL; + isect_list = NULL; } ~_Jv_BytecodeVerifier () { - if (states) - _Jv_Free (states); if (flags) _Jv_Free (flags); - if (jsr_ptrs) - { - for (int i = 0; i < current_method->code_length; ++i) - { - if (jsr_ptrs[i] != NULL) - { - subr_info *info = jsr_ptrs[i]; - while (info != NULL) - { - subr_info *next = info->next; - _Jv_Free (info); - info = next; - } - } - } - _Jv_Free (jsr_ptrs); - } - while (utf8_list != NULL) { - linked_utf8 *n = utf8_list->next; - _Jv_Free (utf8_list->val); + linked<_Jv_Utf8Const> *n = utf8_list->next; _Jv_Free (utf8_list); utf8_list = n; } - while (entry_points != NULL) + while (isect_list != NULL) + { + ref_intersection *next = isect_list->alloc_next; + delete isect_list; + isect_list = next; + } + + if (states) { - subr_entry_info *next = entry_points->next; - _Jv_Free (entry_points); - entry_points = next; + for (int i = 0; i < current_method->code_length; ++i) + { + linked *iter = states[i]; + while (iter != NULL) + { + linked *next = iter->next; + delete iter->val; + _Jv_Free (iter); + iter = next; + } + } + _Jv_Free (states); } } }; @@ -3229,4 +3232,5 @@ _Jv_VerifyMethod (_Jv_InterpMethod *meth) _Jv_BytecodeVerifier v (meth); v.verify_instructions (); } + #endif /* INTERPRETER */