<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Fri, Aug 8, 2014 at 3:07 AM, Jordan Rose <span dir="ltr"><<a href="mailto:jordan_rose@apple.com" target="_blank">jordan_rose@apple.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word"><br><div><div><div><div>On Aug 7, 2014, at 13:01 , Manuel Klimek <<a href="mailto:klimek@google.com" target="_blank">klimek@google.com</a>> wrote:</div>
<br><blockquote type="cite"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Thu, Aug 7, 2014 at 6:57 PM, Jordan Rose <span dir="ltr"><<a href="mailto:jordan_rose@apple.com" target="_blank">jordan_rose@apple.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><div><br><div><div>
On Aug 6, 2014, at 12:51 , Manuel Klimek <<a href="mailto:klimek@google.com" target="_blank">klimek@google.com</a>> wrote:</div>
<br><blockquote type="cite"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Tue, Aug 5, 2014 at 7:28 PM, Manuel Klimek <span dir="ltr"><<a href="mailto:klimek@google.com" target="_blank">klimek@google.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">
<div>On Tue, Aug 5, 2014 at 7:13 PM, Jordan Rose <span dir="ltr"><<a href="mailto:jordan_rose@apple.com" target="_blank">jordan_rose@apple.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><br><div><div>
On Aug 1, 2014, at 4:35 , Manuel Klimek <<a href="mailto:klimek@google.com" target="_blank">klimek@google.com</a>> wrote:</div>
<br><blockquote type="cite"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Fri, Aug 1, 2014 at 1:02 AM, Jordan Rose <span dir="ltr"><<a href="mailto:jordan_rose@apple.com" target="_blank">jordan_rose@apple.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><br><div><div>On Jul 31, 2014, at 12:54 , Manuel Klimek <<a href="mailto:klimek@google.com" target="_blank">klimek@google.com</a>> wrote:</div>
<br><blockquote type="cite"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Thu, Jul 31, 2014 at 9:51 PM, Jordan Rose <span dir="ltr"><<a href="mailto:jordan_rose@apple.com" target="_blank">jordan_rose@apple.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><br><div><div>
<div>On Jul 31, 2014, at 12:46 , Manuel Klimek <<a href="mailto:klimek@google.com" target="_blank">klimek@google.com</a>> wrote:</div><br><blockquote type="cite"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">
On Thu, Jul 31, 2014 at 9:36 PM, Jordan Rose <span dir="ltr"><<a href="mailto:jordan_rose@apple.com" target="_blank">jordan_rose@apple.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><br><div><div>On Jul 31, 2014, at 1:20 , Manuel Klimek <<a href="mailto:klimek@google.com" target="_blank">klimek@google.com</a>> wrote:</div>
<br><blockquote type="cite"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Thu, Jul 31, 2014 at 5:19 AM, Jordan Rose <span dir="ltr"><<a href="mailto:jordan_rose@apple.com" target="_blank">jordan_rose@apple.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><br>
<div><div>On Jul 30, 2014, at 8:14 , Manuel Klimek <<a href="mailto:klimek@google.com" target="_blank">klimek@google.com</a>> wrote:</div><br><blockquote type="cite"><div dir="ltr">Hi Jordan,<div><br></div><div>I made some progress on inlining temporary destructors. My main problem currently is that somehow the initializer values in my constructor get lost if I don't also reference 'this' via some other way.</div>
<div><br></div><div>An examples shows that best. Consider:</div><div>struct A {</div><div> A(int* y) : z(y) {}</div><div> ~A() { *z = 23; }</div><div> int *z;</div><div>};</div><div>void f() {</div><div> int x = 0;</div>
<div> {(A(&x));}</div><div>}</div><div><br></div><div>This leads to a "Dereference of undefined pointer value" warning on *z = 23.</div><div>Funnily enough this warning goes away (and all values get correctly propagated) if change the constructor to be:</div>
<div>A(int* y) : z(y) { (void)this; }</div><div><br></div><div>Obviously this also works fine if I use a local variable instead of a temporary.</div><div><br></div><div>Any ideas what might be going wrong here would be greatly appreciated.</div>
</div></blockquote><br></div></div><div>My guess is that the temporary is an rvalue, and in the analyzer a NonLoc, and that the only reason this even appears to work is because we needed to be able to call <i>methods</i> on temporaries and so will create a temporary region for them on demand to serve as the 'this' pointer. However, I don't think we expect that to happen for <i>constructors,</i> and so we don't keep the region around after the constructor has finished executing. (That is, when it's just a method being called—and neither the constructor nor destructor are inlined—it doesn't actually matter what you do to the contents of the region, because nothing else can access it.)</div>
<div><br></div></div></blockquote><div><br></div><div>So, funnily enough if I remove this code:</div><div><div> if (isTemporaryPRValue(CCE, ThisV))</div><div><span style="white-space:pre-wrap"> </span>ThisV = state->getSVal(ThisV.castAs<Loc>());</div>
</div><div>which puts the temp region into a lazycompoundregion, the memory seems to be correctly handled. I have not yet looked what else this breaks though.</div></div></div></div></blockquote><br></div></div><div>Okay, analyzer design time. :-)</div>
<div><br></div><div>C++ has a distinction between lvalues and rvalues (actually glvalues and prvalues). There are a number of things that go into this distinction, but the part the analyzer cares about is that glvalues refer to true objects and prvalues do not. (In C++ parlance, "object" refers to anything with storage, a type, and lifetime, not just class instances. In the analyzer, we call such things "regions".)</div>
<div><br></div><div>If you were to build a very literal C++ compiler, every glvalue would be represented by an address. PRValues*, on the other hand, don't really need addresses; they only exist to be used transiently.</div>
</div></blockquote><div><br></div><div>I'm not sure how "exists transiently" equals "doesn't need an address". At least in C++, this seems incorrect (?) for objects.</div><div></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">
<div style="word-wrap:break-word"><div></div></div></blockquote></div></div></div></blockquote><div><br></div></div><div>IMHO it's a problem with the standard, and this extra invention of "modifiable" seems weird to me.</div>
<div><div><br></div><br><blockquote type="cite"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">
<div style="word-wrap:break-word"><div><i>* How do you capitalize "prvalue"?</i></div><div><br></div><div>The analyzer pretty much always believes the compiler when it comes to glvalues vs. prvalues. That means that the value of a glvalue expression is <i>always</i> a Loc. A prvalue expression, on the other hand, will only be a Loc if it's a pointer type*. If you're talking about a struct prvalue, what you're really talking about is the bindings in its memory region—it has no identity. This is represented by a CompoundVal or LazyCompoundVal. This model worked great for C, but it's starting to show strain in C++.</div>
<div><br></div><div><i>* or one of the other types mentioned in Loc::isLocType. But not all of those can be expression values.</i></div><div><br></div><div>Besides "regular" temporaries, this is also a problem for prvalue function arguments. When the analyzer inlines a function, it just blithely binds the arguments to the proper VarRegions created for each ParmVarDecl in the current stack frame. That doesn't really fly if we're supposed to be calling the copy constructor. Worse, the copy constructor call happens <i>outside</i> of the function call, which means it happens before we've even decided whether or not to inline, which means there are no VarRegions yet to construct into. We've/I've been putting off dealing with this one for a while; we get around it mostly because we <i>don't inline constructors for temporaries.</i></div>
<div><br></div><div>The general temporary problem goes to what Richard was saying: a temporary expression is an rvalue, but if we have to run a constructor then clearly there <i>was</i> a region at some point. The C++11 standard (at least the version I have) is not wonderful about classifying temporaries ([basic.lval]p1):</div>
<div><br></div><div><div><blockquote type="cite">- An <i>lvalue</i>...designates a function or an object.</blockquote><blockquote type="cite">- An <i>xvalue</i> (an “eXpiring” value) also refers to an object, usually near the end of its lifetime.</blockquote>
<div><blockquote type="cite">- A <i>glvalue</i> (“generalized” lvalue) is an lvalue or an xvalue.</blockquote></div><div><blockquote type="cite">- An <i>rvalue...</i>is an xvalue, a temporary object (12.2) or subobject thereof, or a value that is not associated with an object.</blockquote>
<blockquote type="cite"><font color="#0f61c8">- A <i>prvalue</i> (“pure” rvalue) is an rvalue that is not an xvalue.</font></blockquote></div></div><div><br></div><div>So temporary objects are prvalues. Even though they are modifiable (defined later in the same section). Even though with construction and destruction they end up needing a fixed region, or at least consistency in region contents.</div>
<div><br></div><div>I don't have a great answer here. My guess is we'll probably end up turning on <i>some</i> inlining for temporaries to start, but not all.</div><div><br></div><div>Oh, right, so what's that 'isTemporaryPRValue' doing? That's taking the few cases where the AST wants a prvalue but the analyzer really needs a valid Loc, and conjuring a region on the fly. That's important for things that need to have referential identity, like the 'this' pointer in a method called on a temporary, or (say) a temporary that's been lifetime-extended.</div>
</div></div></blockquote><div><br></div><div>Since this is right at the end of the constructor call, this always creates and binds the value if we hit the constructor. See my patch - it seems to work fine :) What am I missing?</div>
</div></div></div></blockquote><br></div></div><div>That breaks the invariant that a non-Loc-type prvalue is not a Loc. If you tried to assign this directly into another struct, you'd effectively be assigning an address instead of a set of bindings. IIRC there's some code in the analyzer that tries to handle this case anyway, so it may not break down immediately, but I'd rather not lose that assumption.</div>
</div></blockquote><div><br></div><div>I'm confused - isn't that already the case currently? Have you looked at <a href="http://reviews.llvm.org/D4740" target="_blank">http://reviews.llvm.org/D4740</a>? It's very small, and I don't see where I regress an assumption that's not already violated (if I remove the isTemporaryPRValue section loads of tests break, so I assume changing the layout of a temporary section will make significant changes to the analyzer necessary).</div>
</div></div></div></blockquote><br></div></div><div>Sorry, I meant removing the isTemporaryPRValue thing would break invariants. Enabling inlining doesn't break any existing invariants, but it makes things worse than they currently are. This is the test case I have in mind:</div>
<div><br></div><div>struct Test {</div><div> bool flag;</div><div> ~Test {</div><div> if (flag)</div><div> clang_analyzer_checkReached(true);</div><div> }</div><div>}</div><div><br></div><div>void use(Test arg) {</div>
<div><span style="white-space:pre-wrap"> </span>arg.flag = true;</div><div>}</div><div><br></div><div>void topLevel() {</div><div><span style="white-space:pre-wrap"> </span>Test t;</div><div><span style="white-space:pre-wrap"> </span>use(t);</div>
<div>}</div><br><div>The call 'test(t)' involves a copy constructor, which the analyzer treats as a temporary construction. Maybe that's the problem—this is really a different sort of construction. In any case, though, who's responsible for destroying 'arg'? I think it has to be the caller, because the callee may not be inlined...but maybe I'm wrong about that. Still, the value of 'this' won't match between construction at the call site and use within the 'use' method.</div>
</div></blockquote><div><br></div><div>That makes sense, but it would seem to me like that needs to be fixed when passing a parameter by value - that seems to be incorrectly modeled then.</div><div><br></div><div>I think it'd be fine to model this as a copy constructor (Richard might know more), but I think in that case we'd need to call the destructors for the temporary outside *and* the destructor for the parameter inside. That is, instead of having 2 destructor calls (one for 't' and one for the temporary), we'd have three: One for 't', one for the temporary (both with flag == false) and one for the parameter, executed inside the function (with flag == true).</div>
<div>On the other hand the more correct thing to do is probably to somehow use the memory region of the temporary we hand in for 'arg'. I have no idea where in the static analyzer this is handled though - I dug through Store.cpp and RegionStore.cpp but couldn't figure out where the copy of the regions really happens...</div>
</div></div></div></blockquote><br></div></div><div>I think you're looking for CallEvent::getInitialStackFrameContents (and its overrides).</div><div><br></div><div>I'd be concerned about modeling <i>more</i> constructors/destructors than actually exist at runtime—the constructor may have observable side effects. According to the standard, construction and destruction properly happens on the caller side. So the "best" way to deal with this might be to have the parameter regions somehow alias the temporary regions...or maybe for all class-typed parameters to be passed by reference.</div>
</div></blockquote><div><br></div></div><div>Yep, I assumed the idea to fix that would be to look at where we copy the region (currently without invoking the constructor) and instead creating an alias.</div><div>
<div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">
<div style="word-wrap:break-word">Anyway, <i>really</i> I think we should just continue not inlining these particular destructors at the moment, which is conservative and probably not a dealbreaker.</div></blockquote>
<div><br></div></div><div>I'm not sure I understand what you mean by "not inlining these particular destructors". Because if we inline temporary destructors, we will inline these destructors. I also think inlining is benign from the call site point of view, because all that currently seems to happen is that we inline the destructor "as if" we didn't inline the function to which the object was passed as by-value argument, or am I missing something here?</div>
</div></div></div></blockquote><div><br></div><div>In addition, after having looked at the standard, it seems to me like the standard does not require the temporary to be constructed in the slot for the by-value argument, but it can also copy-construct the by-value argument:</div>
<div>12.2-2:</div><div>"""</div><div><div>class X {</div><div>public:</div><div>X(int);</div><div>X(const X&);</div><div>X& operator=(const X&);</div><div>~X();</div><div>};</div></div><div><div>
class Y {</div><div>public:</div><div>Y(int);</div><div>Y(Y&&);</div><div>~Y();</div><div>};</div><div>X f(X);</div><div>Y g(Y);</div><div>void h() {</div><div>X a(1);</div><div>X b = f(X(2));</div><div>Y c = g(Y(3));</div>
<div>a = f(a);</div><div>}</div></div><div><div>An implementation might use a temporary in which to construct X(2) before passing it to f() using X’s copy</div><div>constructor;</div></div><div>"""</div><div>
<br></div><div>Thus, it seems to me like in the example the author of the code must deal correctly with the destructor being called on the unchanged object anyway.</div></div></div></div></blockquote><br></div></div></div>
<div>That isn't the problem. It's the copy-constructed object that then ends up unchanged.</div><div><br></div><div>struct X {</div><div><span style="white-space:pre-wrap"> </span>void **ptr;</div><div><span style="white-space:pre-wrap"> </span>X() { static void *global; ptr = &global; }</div>
<div><span style="white-space:pre-wrap"> </span>~X() { *ptr = nullptr; }</div><div>}</div><div>void nullOut(X x) {</div><div><span style="white-space:pre-wrap"> </span>x.ptr = nullptr;</div><div><span style="white-space:pre-wrap"> </span>//</div>
<div>}</div>void test() {<div><span style="white-space:pre-wrap"> </span>X local;</div><div><span style="white-space:pre-wrap"> </span>nullOut(local); // pass-by-value param constructed and destroyed here</div><div><span style="white-space:pre-wrap"> </span>// 'local' destroyed here.<br>
<div>}</div></div><div><br></div><div>In this example, there is always one copy-construction (for the value passed to 'nullOut') and one temporary destructor (for the same). However, we should see a null pointer dereference here (because we can inline 'nullOut', and that modifies the parameter region) but we don't (because the parameter region isn't actually the region that gets destroyed).</div>
</div></blockquote><div><br></div><div>If we for some reason decide not to inline the nullOut, it seems to me like we won't see the null pointer dereference either.</div><div>Given that the static analyzer has arbitrary limits on inlining etc, I don't see why some remaining false negatives are a problem (false positives in C++ code are already a huge problem, and implementing inlining for temp dtors resolves a couple (I can get more accurate numbers)).</div>
<div>I haven't yet found cases where this would lead to more false positives - examples of false positives could easily convince me that this indeed needs better handling.</div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
</blockquote></div></div></div></blockquote><div><br></div></div></div><div>The decision to inline nullOut is independent of the decision to inline the copy constructor/destructor for the pass-by-value parameter. If this example isn't convincing, try flipping it around:</div>
<div><br></div><div>struct X {<br><span style="white-space:pre-wrap"> </span>void **ptr = nullptr;<div><br><span style="white-space:pre-wrap"> </span>~X() { *ptr = nullptr; }<br>}<br>void nullOut(X x) {</div></div>
<div><span style="white-space:pre-wrap"> </span>static void *global1;<br><span style="white-space:pre-wrap"> </span>x.ptr = &local;<div><br>}<br>void test() {<br><span style="white-space:pre-wrap"> </span>X local;<br>
<span style="white-space:pre-wrap"> </span>nullOut(local); // pass-by-value param constructed and destroyed here<br></div></div><div><br></div><div><span style="white-space:pre-wrap"> </span>static void *global2;</div><div>
<span style="white-space:pre-wrap"> </span>local.ptr = &global2;<br><span style="white-space:pre-wrap"> </span>// 'local' destroyed here.<br>}<br><br></div><div>In this case there's <i>no</i> null pointer dereference, but unless we somehow model that nullOut can modify the pass-by-value parameter, then we'll get a null pointer dereference whether or not nullOut is inlined.</div>
</div></div></blockquote><div><br></div><div>Ok, after trying to craft a lengthy response, I think I finally understand :)</div><div><br></div><div>The problem really is that to correctly model inlining of temp dtors for parameters, we have to correctly mark the regions as "escaping" (thus all it's content is unknown) when either a) we don't inline the function call itself (perhaps because we don't actually see the implementation), or b) the data escapes within the inlined function.</div>
<div><br></div><div>I updated the patch to not inline temporary destructors that bind to parameters.</div><div>I'm not sure whether inlining the constructors in those cases is benign (it seems like it would be).</div>
<div><br></div><div>Thanks,</div><div>/Manuel</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word"><div><div><div>
<br></div><br><blockquote type="cite"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word">
<div><blockquote type="cite"><div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word">
Anyway, <i>really</i> I think we should just continue not inlining these particular destructors at the moment, which is conservative and probably not a dealbreaker.</div></blockquote><div>
<br></div></div><div>I'm not sure I understand what you mean by "not inlining these particular destructors". Because if we inline temporary destructors, we will inline these destructors. I also think inlining is benign from the call site point of view, because all that currently seems to happen is that we inline the destructor "as if" we didn't inline the function to which the object was passed as by-value argument, or am I missing something here?</div>
</blockquote><div><br></div></div><div>Anything we do or don't inline is up to us; we can have a new rule that says we don't inline dtors of temporaries copy-constructed only for use as parameters. (We've had plenty of rules like this along the way, and still have some: we don't inline constructors for 'new' expressions, we don't do array *tors properly, and we ban particular methods in the STL that do more harm than good.)</div>
</div></blockquote><div><br></div><div>That sounds fine; I'd say we would want to keep the ctor/dtor pair; either inline both, or inline none (as a ctor being inlined but a dtor not being inlined is what I've seen to lead to confusion). If that's what it takes I can take a stab at it (although I'd also be interested in implementing the memory region aliasing - my main problem is that I'm not really sure where that needs to go / how that needs to work).</div>
</div></div></div></blockquote><br></div></div><div>I'm not sure how to do the memory aliasing yet either. One hack would be to copy the result back <i>out</i> after the function call, but that doesn't help if anyone's relying on 'this' being consistent. Another option would be to treat byvalue struct parameters as references. A third would be to invent some kind of mapping from the ParmVarDecl to an existing region, rather than just synthesizing a VarRegion around it.</div>
<div><br></div><div>But yes, for now, I'd just not inline these constructors/destructors.</div><span><font color="#888888"><br><div>Jordan</div></font></span></div></blockquote></div><br></div></div>