[cfe-dev] [analyzer] Temporaries.
Artem Dergachev via cfe-dev
cfe-dev at lists.llvm.org
Wed Feb 14 13:20:28 PST 2018
More explanations on how the analyzer keeps making its way around the
C++ AST.
== Lifetime extension ==
This is a brain dump of how (and how much) lifetime extension of
temporary objects is currently broken in the static analyzert. Spoilers:
not too much, it seems, but annoying nevertheless.
Consider an example:
1 class C {
2 public:
3 C() {}
4 ~C() {}
5 };
6
7 void foo() {
8 const C &c = C();
9 }
With the AST for the variable declaration:
DeclStmt 0x7fa5ac85cba0 <line:8:3, col:19>
`-VarDecl 0x7fa5ac85c878 <col:3, col:18> col:12 c 'const C &' cinit
`-ExprWithCleanups 0x7fa5ac85cb30 <col:16, col:18> 'const C' lvalue
`-MaterializeTemporaryExpr 0x7fa5ac85cb00 <col:16, col:18>
'const C' lvalue extended by Var 0x7fa5ac85c878 'c' 'const C &'
`-ImplicitCastExpr 0x7fa5ac85cae8 <col:16, col:18> 'const
C' <NoOp>
`-CXXBindTemporaryExpr 0x7fa5ac85cac8 <col:16, col:18>
'C' (CXXTemporary 0x7fa5ac85cac0)
`-CXXTemporaryObjectExpr 0x7fa5ac85ca88 <col:16,
col:18> 'C' 'void ()'
*here goes a periodic reminder that CXXTemporaryObjectExpr is a
sub-class of CXXConstructExpr*
Notice how MaterializeTemporaryExpr is the innermost expression (the
first one in the order of execution) that is an lvalue. Essentially, you
can think of it as the mythical "rvalue-to-lvalue" cast that takes in a
C++ object rvalue and returns the this-value for that object. Because
all C++ objects have a this-value that never changes throughout their
lifetime, it is essentially their identity. Otherwise you can't call
methods on them.
MaterializeTemporaryExpr also contains information about the lifetime
extension process: we needed the this-value in order to bind it to
variable 'c'. You see that in the AST.
In the analyzer, however, MaterializeTemporaryExpr does a different
thing, as a temporary solution (no pun intended). It constructs a new
temporary region out of thin air and binds the rvalue object to that
temporary in the Store. The respective function in our code is called
"createTemporaryRegionIfNeeded". It also has a separate purpose of
converting rvalue sub-object adjustments into lvalue sub-object
adjustments, which we wouldn't discuss this time.
Now that we learned how to inline temporary constructors and
destructors, it essentially means that the this-value in the constructor
and in the destructor would be different. Because first we construct the
object into temporary region R1, then we take lazyCompoundVal{R1} to
represent the value of CXXTemporaryObjectExpr, then we materialize
lazyCompoundVal{R1} to R2, then we bind R2 to variable 'c', then we call
the automatic(!) destructor for 'c' which contains R2. To be clear, the
region store at the time of destruction would be:
c: R2,
R2: lazyCompoundVal{R1}.
It means that fields of the object would contain the correct values,
there would be the correct number of destructors called (no temporary
destructors, just one automatic destructor), but the identity of the
object (this-value) would change in the process. Unless the object
actually makes any decisions or does any manipulations that involve its
this-value, the modeling should be correct. When the object starts to
actively use its this-value in its inlined methods, the analysis would
do wrong stuff. Additionally, it causes a mess in the checkers when they
try to track objects by their this-values - i.e. IteratorChecker has a
lot of additional code to work around the fact that the region for the
object constantly changes.
From the above it is clear that MaterializeTemporaryExpr should not
construct any new regions, at least not in this case. We already have
the correct region, namely R1, which should be re-used.
It is tempting to take R1 directly from lazyCompoundVal{R1} - it already
has memories about once being a value of R1. I'm not sure it's not going
to work - it may work, at least i'm not ready to come up with a
counterexample. But the meaning of LazyCompoundVal's parent region is
different and coincides with what we want just accidentally. Namely,
lazyCompoundVal{R1} is a value of an object that was bound to R1 in some
particular moment of time in the past, without any explanation of when
this moment of time was - but there's no indication if R1 is the region
of the temporary we've just constructed, or a region of an unrelated
object that used to have the same value back then. As we'd see later,
MaterializeTemporaryExpr doesn't always contain a constructor within it
- there are a lot of cases to cover, and if the trick doesn't work even
once, it's going to be hard, so i'd probably not going to commit into
maintaining this invariant. Though it might be plausible to modify add
an SVal kind that does exactly what we mean here - i.e. a value that
does indeed correspond to a specific C++ object identified by region. It
might be a beautiful solution, but it may also fail miserably if tricky
cornercases show up - so i'm not ready to commit to that. Also the
correct way of dealing with such values (i.e. for how long are they
relevant?) would be extremely difficult to explain to checker developers.
The more straightforward approach here is to include
MaterializeTemporaryExpr (hereinafter MTE) into the construction
context. It means, if a temporary that we're about to construct would be
lifetime-extended later, we'd rather know about that during
construction, and maintain a map in the program state from MTE to their
respective temporary regions that were used for representing the
respective construction targets. Upon encountering the respective MTE
we'd simply retrieve the implicit temporary storage for the value from
the program state and declare that temporary region to be the value of
the MTE. This would mimic the approach we have with
CXXBindTemporaryExprs (hereinafter BTE) and their respective regions
that allows temporary destructors to work - but this time it's going to
be about MaterializeTemporaryExprs and automatic destructors. I imagine
that on the checker side this can potentially be exposed via some sort
of State->getTemporaryStorage(Expr) call, but i believe that generally
this process should be as transparent to the checkers as possible.
It sounds as if both of these maps could be eliminated by always making
sure that the target temporary is constructed "with" the MTE (for
lifetime-extended temproraries) or BTE (for temporaries that require
destruction at the end of full-expression). In this case, with the help
of construction context-assisted lookahead, we declare that the target
of the construction is CXXTempObjectRegion(MTE, LC) or
CXXTempObjectRegion(BTE, LC) respectively, rather than
CXXTempObjectRegion(CXXConstructExpr). Then during evaluation of MTE or
BTE we'd simply construct the same region with the expression we're
currently evaluating, and it's automagically going to be the correct
region. This approach, however, becomes confusing when we start dealing
with elidable constructors (explained below). So for now i believe that
it is quite irrelevant which expression is identifying the temporary region.
== Elidable constructors ==
While it doesn't sound like an immediately important task to implement
copy elision in the analyzer, it may help with making some things
easier. And it'd also make some reports fancier, as mentioned in
https://reviews.llvm.org/D43144.
Elidable copy-constructors can be explained as a form of lifetime
extension. Instead of copying the temporary, they keep using the
original value of the temporary, which in some pretty twisted sense
means that they are lifetime-extending it to be able to use it. For
example, if we modify our example by replacing the lifetime-extending
reference variable with a value-type variable:
1 class C {
2 public:
3 C() {}
4 ~C() {}
5 };
6
7 void foo() {
8 C c = C();
9 }
...then we'd still have an MTE, even though lifetime extension would
seem to be gone:
DeclStmt 0x7fb8f005afb8 <line:8:3, col:12>
`-VarDecl 0x7fb8f005ac50 <col:3, col:11> col:5 c 'C' cinit
`-ExprWithCleanups 0x7fb8f005afa0 <col:9, col:11> 'C'
`-CXXConstructExpr 0x7fb8f005af68 <col:9, col:11> 'C' 'void
(const C &) noexcept' elidable
`-MaterializeTemporaryExpr 0x7fb8f005af00 <col:9, col:11>
'const C' lvalue
`-ImplicitCastExpr 0x7fb8f005aee8 <col:9, col:11> 'const
C' <NoOp>
`-CXXBindTemporaryExpr 0x7fb8f005aec8 <col:9, col:11>
'C' (CXXTemporary 0x7fb8f005aec0)
`-CXXTemporaryObjectExpr 0x7fb8f005ae88 <col:9,
col:11> 'C' 'void ()'
In this case the MTE is expressing the fact that the temporary
constructed via CXXTemporaryObjectExpr can be "lifetime-extended" (by
actually merging it with the stack variable) rather than copied, if the
CXXConstructExpr surrounding it would be chosen to be elided. The AST
doesn't make the elision choice for us - but is compatible with both
choices. The MTE essentially overrides the necessity of immediate
destruction provided by the BTE, and lets the surrounding AST decide
upon the lifetime of the object.
Because the analyzer currently does not do copy elision, it will use the
MTE only to compute the argument for the elidable copy-constructor, and
then perform the copy-construction, and then destroy the original
temporary at the end of the full-expression. Note, however, that in this
case we need to properly support both the BTE (for the temporary
destructor to work) and the MTE (for computing its value). We need to
implement the MTE's ability to perform "rvalue-to-lvalue-cast" even if
the temporary destruction is still necessary. For this reason, if we
rely on constructing temporary regions with the correct BTEs or MTEs, at
least one of these tasks becomes impossible to perform.
If we were to support copy elision, then the CXXTemporaryObjectExpr
constructor would go directly into the variable region. For the purposes
of modeling, it'd mean that only CXXTemporaryObjectExpr would actually
need to be modeled. But this would require additional coding in the
construction context to be able to realize that the target is the
variable while modeling the CXXTemporaryObjectExpr.
For the sake of completeness, let's consider the ternary operator example:
1 class C {
2 public:
3 C(int) {}
4 ~C() {}
5 };
6
7 void foo(int coin) {
8 const C &c = coin ? C(1) : C(2);
9 }
The respective AST would be:
DeclStmt 0x7fc1e20023e0 <line:8:3, col:34>
`-VarDecl 0x7fc1e2001dc8 <col:3, col:33> col:12 c 'const C &' cinit
`-ExprWithCleanups 0x7fc1e2002370 <col:16, col:33> 'const C' lvalue
`-MaterializeTemporaryExpr 0x7fc1e2002340 <col:16, col:33>
'const C' lvalue extended by Var 0x7fc1e2001dc8 'c' 'const C &'
`-ImplicitCastExpr 0x7fc1e2002328 <col:16, col:33> 'const
C' <NoOp>
`-ConditionalOperator 0x7fc1e20022f8 <col:16, col:33> 'C'
|-ImplicitCastExpr 0x7fc1e2002170 <col:16> 'bool'
<IntegralToBoolean>
| `-ImplicitCastExpr 0x7fc1e2002158 <col:16> 'int'
<LValueToRValue>
| `-DeclRefExpr 0x7fc1e2001e28 <col:16> 'int' lvalue
ParmVar 0x7fc1e2001c18 'coin' 'int'
|-CXXBindTemporaryExpr 0x7fc1e2002248 <col:23, col:26>
'C' (CXXTemporary 0x7fc1e2002240)
| `-CXXConstructExpr 0x7fc1e2002208 <col:23, col:26>
'C' 'void (const C &) noexcept' elidable
| `-MaterializeTemporaryExpr 0x7fc1e20021a0 <col:23,
col:26> 'const C' lvalue
| `-ImplicitCastExpr 0x7fc1e2002188 <col:23,
col:26> 'const C' <NoOp>
| `-CXXFunctionalCastExpr 0x7fc1e2002078 <col:23,
col:26> 'C' functional cast to class C <ConstructorConversion>
| `-CXXBindTemporaryExpr 0x7fc1e2002058
<col:23, col:26> 'C' (CXXTemporary 0x7fc1e2002050)
| `-CXXConstructExpr 0x7fc1e2002018 <col:23,
col:26> 'C' 'void (int)'
| `-IntegerLiteral 0x7fc1e2001e60 <col:25>
'int' 1
`-CXXBindTemporaryExpr 0x7fc1e20022d8 <col:30, col:33>
'C' (CXXTemporary 0x7fc1e20022d0)
`-CXXConstructExpr 0x7fc1e2002298 <col:30, col:33>
'C' 'void (const C &) noexcept' elidable
`-MaterializeTemporaryExpr 0x7fc1e2002280 <col:30,
col:33> 'const C' lvalue
`-ImplicitCastExpr 0x7fc1e2002268 <col:30,
col:33> 'const C' <NoOp>
`-CXXFunctionalCastExpr 0x7fc1e2002130 <col:30,
col:33> 'C' functional cast to class C <ConstructorConversion>
`-CXXBindTemporaryExpr 0x7fc1e2002110
<col:30, col:33> 'C' (CXXTemporary 0x7fc1e2002108)
`-CXXConstructExpr 0x7fc1e20020d0 <col:30,
col:33> 'C' 'void (int)'
`-IntegerLiteral 0x7fc1e20020b0 <col:32>
'int' 2
Each branch contains two constructors: the temporary and the elidable
copy. The temporaries are surrounded with their respective BTEs and
copy-elision-kind MTEs, which indicates that they need to be either
destroyed as temporaries, or, if copy elision is chosen, have their
lifetime decided upon by the surrounding AST. The elidable copy
constructors also, being temporaries, have their respective BTEs. Note,
however, that there is only one MTE for both BTEs for the elidable
constructors.
So after the conditional operator is resolved (which is the first thing
we need to do, according to the CFG), we'd go ahead and perform the
constructors, and their trigger would be their respective BTE in the
non-elide case, and the single top-level MTE in the elide case. In the
non-elide case, copy constructors would be triggered by the top-level MTE.
It means that, once again, copy elision would prevent us from handling
both the BTE and the copy-elision-kind MTE in the single construction,
allowing the "predictable target region" trick to work: when we need the
temporary destructor, we construct directly into CXXTempObjectRegion of
the BTE and it gets automatically picked up during destruction, and when
we need the automatic destructor, we construct directly into
CXXTempObjectRegion of the MTE and we can easily compute the value of
the MTE. But when we don't do copy elision, we'd have to keep at least
one of those in the program state map.
== Return by value ==
Returning C++ objects by value is actually very similar to constructing
it. Consider:
1 class C {
2 public:
3 C() {}
4 ~C() {}
5 };
6
7 C bar() {
8 C c;
9 return c;
10 }
11
12 void foo() {
13 const C &c = bar();
14 }
With the respective AST for DeclStmt in foo():
DeclStmt 0x7fe62c84f8e8 <line:13:3, col:21>
`-VarDecl 0x7fe62c84f6b8 <col:3, col:20> col:12 c 'const C &' cinit
`-ExprWithCleanups 0x7fe62c84f878 <col:16, col:20> 'const C' lvalue
`-MaterializeTemporaryExpr 0x7fe62c84f848 <col:16, col:20>
'const C' lvalue extended by Var 0x7fe62c84f6b8 'c' 'const C &'
`-ImplicitCastExpr 0x7fe62c84f830 <col:16, col:20> 'const
C' <NoOp>
`-CXXBindTemporaryExpr 0x7fe62c84f810 <col:16, col:20>
'C' (CXXTemporary 0x7fe62c84f808)
`-CallExpr 0x7fe62c84f7e0 <col:16, col:20> 'C'
`-ImplicitCastExpr 0x7fe62c84f7c8 <col:16> 'C (*)()'
<FunctionToPointerDecay>
`-DeclRefExpr 0x7fe62c84f770 <col:16> 'C ()' lvalue
Function 0x7fe62c84f190 'bar' 'C ()'
And for the ReturnStmt in bar():
ReturnStmt 0x7fe62c84f5b0 <line:9:3, col:10>
`-CXXConstructExpr 0x7fe62c84f578 <col:10> 'C' 'void (const C &)
noexcept' elidable
`-ImplicitCastExpr 0x7fe62c84f518 <col:10> 'const C' lvalue <NoOp>
`-DeclRefExpr 0x7fe62c84f4f0 <col:10> 'C' lvalue Var
0x7fe62c84f280 'c' 'C'
Since https://reviews.llvm.org/D42875 we can already realize that the
constructor in bar(), assuming that we're inlining bar() during
analysis, would be constructed into something that is a return value of
bar(). This allows us, by looking that the StackFrameContext's call
site, to figure out that it is being constructed into the CallExpr in
foo(). Now if only we knew that that the call site is a
lifetime-extended temporary, i.e. if only we had a pointer to the
foo()'s MTE at the CallExpr's CFG element, we'd be able to find the
correct target region for construction: the CXXTempObjectRegion for the
MTE in the StackFrameContext of foo(). So i'm proposing to add some sort
of construction context to not only constructors, but also to functions
that return objects, and then during construction perform the lookup in
three easy steps:
1. in the callee's CFG from constructor to return statement,
2. through the location from the return statement to the call site,
3. then through the caller's CFG from the call site to the MTE.
If the function is not inlined, we can still make use of the
construction context to represent the return value as a
LazyCompoundValue of the MTE's temporary. It would eliminate the need to
replace the return value with another value while evaluating the MTE,
and of course the need to re-bind the object to a different this-region.
So i believe that this is a good way to eliminate the need for the
"createTemporaryRegionIfNeeded" thing in the function calls as well.
On 06/02/2018 1:41 PM, Artem Dergachev wrote:
> A bit of an update.
>
> == Temporary destructors ==
>
> Adding some initial support for temporary destructors seems pretty
> easy and straightforward at this point, given the awesome work done by
> Manuel Klimek on our CFG a few years ago.
>
> 1. We already have the awesome CFGTemporaryDtor elements, which have
> the backreference to the CXXBindTemporaryExpr for their temporaries.
>
> 2. In simple cases CXXBindTemporaryExprs have an immediate constructor
> within them, and therefore we can provide the CXXBindTemporaryExprs as
> the construction context's trigger statements, and therefore have a
> reliable CXXTempObjectRegion for constructors.
>
> 3. Then we already track the CXXBindTemporaryExprs for the active
> temporaries in the program state. We can attach any constructor
> information to them, such as the target region, if we need to
> (probably we can reconstruct it by looking at the expression and the
> location context, not sure if we want to).
>
> 4. So when we reach the CFGTemporaryDtor element, we can just lookup
> all the info we need, and perform the destruction properly.
>
> 5. There's a bit of a liveness problem, because it seems that our
> liveness analysis tends to throw away the temporary too early. I can
> easily hack this problem away by marking all regions that correspond
> to active temporaries as live. I'll see if there's a better solution.
>
> == CXXDefaultArgExpr problems ==
>
> There's a known problem with those. Consider:
>
> void foo(const C &c = C()) {
> }
>
> void bar() {
> foo();
> foo();
> }
>
> Each call of foo() contains a CXXDefaultArgExpr for c. The default
> argument value C() is constructed before we enter foo() and destroyed
> after we leave foo(). However, c's initializer, "= C()", is *not part
> of the AST of bar()*. In particular, when foo() is called twice, the
> initializer for the two calls is the same, only CXXDefaultArgExprs are
> different. This screws a lot of invariants in the analyzer: program
> points start coinciding (causing the analysis to loop and cache out),
> Environment doesn't expect the same expression in the same location
> context have two different values (suppose calls are nested into each
> other), analysis taking wrong branches, and so on.
>
> Luckily, default-arg expressions that aren't zero integers or null
> pointers are pretty rare. Still, we'd need to eventually think how to
> avoid any really bad practical problems with them.
>
> On 25/01/2018 9:08 AM, Artem Dergachev wrote:
>> Handling C++ temporary object construction and destruction seems to
>> be the biggest cause of false positives on C++ code at the moment.
>> I'd be looking into this, even though for now i don't see the whole
>> scale of problem.
>>
>> == CFG, destructors, and ProgramPoints ==
>>
>> We should probably enable `-analyzer-config cfg-temporary-dtors=true`
>> by default soon. It is a fairly low-impact change because it only
>> alters the CFG but the analyzer rarely actually uses the new nodes.
>> Destructors for the most part are still evaluated conservatively,
>> with improper object regions. So it causes almost no changes in the
>> analyzer's positives for now, but it definitely opens up room for
>> further improvements.
>>
>> I'm observing a couple of problems with this mode at the moment,
>> namely the rich CFG destructor element hierarchy is not currently
>> complemented by an equally rich ProgramPoint hierarchy. This causes
>> the analysis to merge nodes which are not equivalent, for example two
>> implicit destructors of the same type (with the same function
>> declaration) may sometimes cause the ExplodedGraph to coil up and
>> stop analysis (lost coverage) because of having the same program
>> state at the erroneously-same program point. Because situations when
>> we can guarantee a change in the program state are pretty rare, we'd
>> have to produce more program point kinds to handle this correctly.
>>
>> CallEvent hierarchy is also very much affected in a similar manner -
>> because apparently we're constructing program points by looking at
>> CallEvents, so they'd need to carry all the information that's needed
>> to construct the pre-call/post-call program point.
>>
>> == Construction contexts ==
>>
>> We are correctly modeling "this" object region during
>> construction/destruction of variables with automatic storage
>> duration, fields and base classes, and on operator new() since
>> recently, as long as these aren't arrays of objects. It was not yet
>> implemented for other cases such as temporaries, initializer lists,
>> fields or C++17 base classes within aggregates, and pass-by-value
>> from/to functions (the latter being a slightly different problem than
>> temporaries).
>>
>> First of all, by "not yet implemented" i mean that instead of
>> constructing into (destroying) the correct object (in the correct
>> memory region), we come up with a "temporary region", which looks
>> exactly like a region of a valid C++ temporary but is only used for
>> communicating that it is not the right region. Then we look at the
>> region, see that it is a temporary, and avoid inlining constructors,
>> because it would make little sense when the region is not correct.
>> However, if we are to model temporaries, we need another way of
>> communicating our failure to find the correct region, which is being
>> addressed by https://reviews.llvm.org/D42457
>>
>> Also in the cases when the correct region is used, it is being
>> computed in a funky way: in order to figure out where do we construct
>> the object, we walk forward in the CFG (or from child to parent in
>> the AST) to find the next/parent statement that would accomodate the
>> newly constructed object. This means that the CFG, while perfectly
>> telling us what to do in what order (which we, unlike CodeGen, cannot
>> easily take from AST because we cannot afford recursive evaluation of
>> statements, mainly because of inlining), discards too much context to
>> be helpful in understanding how to do it.
>>
>> I tried to add the information about such "trigger statements" for
>> constructors into CFG and it is extremely helpful and easy to both
>> use and extend. This assumes adding a new sub-class of CFGElement for
>> constructors, which maintain a "construction context" - for now it's
>> just a trigger statement. For instance, in
>>
>> class C { ... };
>> void foo() {
>> C c;
>> }
>>
>> ...the trigger for constructor C() would be DeclStmt `C c`, and once
>> we know this we can immediately figure out that the construction
>> target is the region of variable `c`. Construction trigger is not
>> necessarily a statement - it may be a CXXCtorInitializer, which is an
>> AST node kind of its own. Additionally, when constructing aggregates
>> via initializer lists, we may have the same statement trigger
>> multiple constructors, eg. in
>>
>> class C { public: C(int); ~C(); };
>> struct S { C c1, c2, c3; };
>> void foo() {
>> S s { 1, 2, 3 };
>> }
>>
>> ... we have three different constructors (actually six different
>> constructors if we include elidable copy constructors) for c1, c2, c3
>> (and lack of constructor for `s` because of the aggregate thing). It
>> would be more natural to declare that the specific field or index
>> would be a part of the CFG's construction context, as well as the
>> intermediate InitListExpr, so even in these simple cases the
>> construction context may get quite bulky. And this example is quite
>> popular on actual code - probably the second worst cause of false
>> positives after temporaries.
>>
>> For now i have no specific plan on what would construction context
>> for temporaries contain in the general case. I might not be able to
>> get the data structures right from the start. In any case, it might
>> be necessary to perform additional memory allocations for these CFG
>> elements (for analyzer's CFG only, so it wouldn't affect compilation
>> time or warnings).
>>
>> I believe that getting the correct target region in as many cases as
>> possible would be the main part of my work for the nearest future.
>> And i want to move most this work to CFG, while letting the analyzer
>> pick up the trigger statement from it and trust it as blindly as
>> possible.
>>
>> == Workflow ==
>>
>> When working on operator new, i tried hard to maintain a reasonable
>> history, making around 15 local rebases. It was not awesome because
>> it was hard for the reviewers to understand the context of the new
>> changes, and changes could have easily kicked in during rebases. A
>> few lessons learned here would be to commit more aggressively, i.e.
>> avoiding stockpiling a large history of patches (essentially a large
>> branch), which in turn would be possible through trying harder to
>> avoid unrelated hard-to-test core changes (as long as it doesn't
>> require weird workarounds) that aren't covered by a flag (such as
>> those evalCast fixes), in order to make sure reviewing take less
>> responsibility. It's fine if some parts would later be fixed
>> (assuming they would indeed be fixed), if it means making the
>> turnaround faster and the tail of patches shorter - that's also the
>> attitude i'd try to maintain when reviewing stuff myself.
>
More information about the cfe-dev
mailing list