michael@0: /* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 4 -*-
michael@0:  * vim: set ts=8 sts=4 et sw=4 tw=99:
michael@0:  * This Source Code Form is subject to the terms of the Mozilla Public
michael@0:  * License, v. 2.0. If a copy of the MPL was not distributed with this
michael@0:  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
michael@0: 
michael@0: /* JS::Anchor implementation. */
michael@0: 
michael@0: #ifndef js_Anchor_h
michael@0: #define js_Anchor_h
michael@0: 
michael@0: #include "mozilla/Attributes.h"
michael@0: 
michael@0: #include "js/TypeDecls.h"
michael@0: 
michael@0: namespace JS {
michael@0: 
michael@0: /*
michael@0:  * Protecting non-Value, non-JSObject *, non-JSString * values from collection
michael@0:  *
michael@0:  * Most of the time, the garbage collector's conservative stack scanner works
michael@0:  * behind the scenes, finding all live values and protecting them from being
michael@0:  * collected. However, when JSAPI client code obtains a pointer to data the
michael@0:  * scanner does not know about, owned by an object the scanner does know about,
michael@0:  * Care Must Be Taken.
michael@0:  *
michael@0:  * The scanner recognizes only a select set of types: pointers to JSObjects and
michael@0:  * similar things (JSFunctions, and so on), pointers to JSStrings, and Values.
michael@0:  * So while the scanner finds all live |JSString| pointers, it does not notice
michael@0:  * |jschar| pointers.
michael@0:  *
michael@0:  * So suppose we have:
michael@0:  *
michael@0:  *   void f(JSString *str) {
michael@0:  *     const jschar *ch = JS_GetStringCharsZ(str);
michael@0:  *     ... do stuff with ch, but no uses of str ...;
michael@0:  *   }
michael@0:  *
michael@0:  * After the call to |JS_GetStringCharsZ|, there are no further uses of
michael@0:  * |str|, which means that the compiler is within its rights to not store
michael@0:  * it anywhere. But because the stack scanner will not notice |ch|, there
michael@0:  * is no longer any live value in this frame that would keep the string
michael@0:  * alive. If |str| is the last reference to that |JSString|, and the
michael@0:  * collector runs while we are using |ch|, the string's array of |jschar|s
michael@0:  * may be freed out from under us.
michael@0:  *
michael@0:  * Note that there is only an issue when 1) we extract a thing X the scanner
michael@0:  * doesn't recognize from 2) a thing Y the scanner does recognize, and 3) if Y
michael@0:  * gets garbage-collected, then X gets freed. If we have code like this:
michael@0:  *
michael@0:  *   void g(JSObject *obj) {
michael@0:  *     JS::Value x;
michael@0:  *     JS_GetProperty(obj, "x", &x);
michael@0:  *     ... do stuff with x ...
michael@0:  *   }
michael@0:  *
michael@0:  * there's no problem, because the value we've extracted, x, is a Value, a
michael@0:  * type that the conservative scanner recognizes.
michael@0:  *
michael@0:  * Conservative GC frees us from the obligation to explicitly root the types it
michael@0:  * knows about, but when we work with derived values like |ch|, we must root
michael@0:  * their owners, as the derived value alone won't keep them alive.
michael@0:  *
michael@0:  * A JS::Anchor is a kind of GC root that allows us to keep the owners of
michael@0:  * derived values like |ch| alive throughout the Anchor's lifetime. We could
michael@0:  * fix the above code as follows:
michael@0:  *
michael@0:  *   void f(JSString *str) {
michael@0:  *     JS::Anchor<JSString *> a_str(str);
michael@0:  *     const jschar *ch = JS_GetStringCharsZ(str);
michael@0:  *     ... do stuff with ch, but no uses of str ...;
michael@0:  *   }
michael@0:  *
michael@0:  * This simply ensures that |str| will be live until |a_str| goes out of scope.
michael@0:  * As long as we don't retain a pointer to the string's characters for longer
michael@0:  * than that, we have avoided all garbage collection hazards.
michael@0:  */
michael@0: template<typename T> class AnchorPermitted;
michael@0: template<> class AnchorPermitted<JSObject *> { };
michael@0: template<> class AnchorPermitted<const JSObject *> { };
michael@0: template<> class AnchorPermitted<JSFunction *> { };
michael@0: template<> class AnchorPermitted<const JSFunction *> { };
michael@0: template<> class AnchorPermitted<JSString *> { };
michael@0: template<> class AnchorPermitted<const JSString *> { };
michael@0: template<> class AnchorPermitted<Value> { };
michael@0: template<> class AnchorPermitted<const JSScript *> { };
michael@0: template<> class AnchorPermitted<JSScript *> { };
michael@0: 
michael@0: template<typename T>
michael@0: class Anchor : AnchorPermitted<T>
michael@0: {
michael@0:   public:
michael@0:     Anchor() { }
michael@0:     explicit Anchor(T t) { hold = t; }
michael@0:     inline ~Anchor();
michael@0: 
michael@0:   private:
michael@0:     T hold;
michael@0: 
michael@0:     /*
michael@0:      * Rooting analysis considers use of operator= to be a use of an anchor.
michael@0:      * For simplicity, Anchor is treated as if it contained a GC thing, from
michael@0:      * construction. Thus if we had
michael@0:      *
michael@0:      *   void operator=(const T &t) { hold = t; }
michael@0:      *
michael@0:      * and this code
michael@0:      *
michael@0:      *   JS::Anchor<JSString*> anchor;
michael@0:      *   stuff that could GC, producing |str|;
michael@0:      *   anchor = str;
michael@0:      *
michael@0:      * the last line would be seen as a hazard, because the final = would "use"
michael@0:      * |anchor| that is a GC thing -- which could have been moved around by the
michael@0:      * GC. The workaround is to structure your code so that JS::Anchor is
michael@0:      * always constructed, living for however long the corresponding value must
michael@0:      * live.
michael@0:      */
michael@0:     void operator=(const T &t) MOZ_DELETE;
michael@0: 
michael@0:     Anchor(const Anchor &other) MOZ_DELETE;
michael@0:     void operator=(const Anchor &other) MOZ_DELETE;
michael@0: };
michael@0: 
michael@0: template<typename T>
michael@0: inline Anchor<T>::~Anchor()
michael@0: {
michael@0: #ifdef __GNUC__
michael@0:     /*
michael@0:      * No code is generated for this. But because this is marked 'volatile', G++ will
michael@0:      * assume it has important side-effects, and won't delete it. (G++ never looks at
michael@0:      * the actual text and notices it's empty.) And because we have passed |hold| to
michael@0:      * it, GCC will keep |hold| alive until this point.
michael@0:      *
michael@0:      * The "memory" clobber operand ensures that G++ will not move prior memory
michael@0:      * accesses after the asm --- it's a barrier. Unfortunately, it also means that
michael@0:      * G++ will assume that all memory has changed after the asm, as it would for a
michael@0:      * call to an unknown function. I don't know of a way to avoid that consequence.
michael@0:      */
michael@0:     asm volatile("":: "g" (hold) : "memory");
michael@0: #else
michael@0:     /*
michael@0:      * An adequate portable substitute, for non-structure types.
michael@0:      *
michael@0:      * The compiler promises that, by the end of an expression statement, the
michael@0:      * last-stored value to a volatile object is the same as it would be in an
michael@0:      * unoptimized, direct implementation (the "abstract machine" whose behavior the
michael@0:      * language spec describes). However, the compiler is still free to reorder
michael@0:      * non-volatile accesses across this store --- which is what we must prevent. So
michael@0:      * assigning the held value to a volatile variable, as we do here, is not enough.
michael@0:      *
michael@0:      * In our case, however, garbage collection only occurs at function calls, so it
michael@0:      * is sufficient to ensure that the destructor's store isn't moved earlier across
michael@0:      * any function calls that could collect. It is hard to imagine the compiler
michael@0:      * analyzing the program so thoroughly that it could prove that such motion was
michael@0:      * safe. In practice, compilers treat calls to the collector as opaque operations
michael@0:      * --- in particular, as operations which could access volatile variables, across
michael@0:      * which this destructor must not be moved.
michael@0:      *
michael@0:      * ("Objection, your honor!  *Alleged* killer whale!")
michael@0:      *
michael@0:      * The disadvantage of this approach is that it does generate code for the store.
michael@0:      * We do need to use Anchors in some cases where cycles are tight.
michael@0:      *
michael@0:      * Note that there is a Anchor<Value>::~Anchor() specialization in Value.h.
michael@0:      */
michael@0:     volatile T sink;
michael@0:     sink = hold;
michael@0: #endif  /* defined(__GNUC__) */
michael@0: }
michael@0: 
michael@0: } // namespace JS
michael@0: 
michael@0: #endif /* js_Anchor_h */