Hi all, I thought I'd look at this problem https://bugs.llvm.org/show_bug.cgi?id=43364 (Pointer cast representation problems). maybe use it as an opportunity to dig into the inner workings of the analyzer and maybe even solve a concrete problem :) But it seems I need guidance about possible paths to pursue.
Starting with a concrete case, a simple reproducer is below. I debugged this down to SimpleSValBuilder.cpp:evalCastFromLoc(). val is an SVal and castTy is a Loc, and this code path attempts to extract a concrete integer from the SVal (did I get this right?). So I think a solution to this specific case would be to dig into the SVal to see if casted data is concrete, and extract that data. Seems to me this would be the location for that (evalCastFromLoc seems appropropriate enough). If that's true, can someone point me to an example that's similar? I'll keep digging, but thought I'd ask in case this is easy for someone to drop a few helpful hints.
I've included some select dumps, state and a bt below in case this is helpful.
Best
(gdb) p val.dump() &Element{myvar,0 S64b,char}$2 = void
Ok, you are saying that we should know that `p[1]` points to an object which was initialized. And according to the target's endianness, resolve the DeclRefExpr to the appropriate constant value.
The dumps are probably from the line `char *p = &myvar;` if I'm right.
That is the reason why the right-hand side is evaluated to the location of `myvar`, but wrapped into an ElementRegion of type char representing the reinterpret cast.
IMO, it's the correct behavior to this point.
You should probably dig around the evaluation of the `DeclRefExpr` of the expression `p[1]`. That should return your concrete value.
AFAIK, the store maps the values to expressions and the problem probably resides there.
But it's just a wide guess :D
Unfortunately, that is all I can say about this right now :(
Balazs.
Vince Bridgers via cfe-dev <[hidden email]> ezt írta (időpont: 2020. dec. 7., H, 11:54):
Hi all, I thought I'd look at this problem https://bugs.llvm.org/show_bug.cgi?id=43364 (Pointer cast representation problems). maybe use it as an opportunity to dig into the inner workings of the analyzer and maybe even solve a concrete problem :) But it seems I need guidance about possible paths to pursue.
Starting with a concrete case, a simple reproducer is below. I debugged this down to SimpleSValBuilder.cpp:evalCastFromLoc(). val is an SVal and castTy is a Loc, and this code path attempts to extract a concrete integer from the SVal (did I get this right?). So I think a solution to this specific case would be to dig into the SVal to see if casted data is concrete, and extract that data. Seems to me this would be the location for that (evalCastFromLoc seems appropropriate enough). If that's true, can someone point me to an example that's similar? I'll keep digging, but thought I'd ask in case this is easy for someone to drop a few helpful hints.
I've included some select dumps, state and a bt below in case this is helpful.
Best
(gdb) p val.dump() &Element{myvar,0 S64b,char}$2 = void
Ok, first of all, the example false positive you're debugging is not
an example of the bug you're trying to look into, but an example of
a completely different long-standing bug:
https://bugs.llvm.org/show_bug.cgi?id=44114. Both bugzilla entries
contain a large number of examples to choose from. None of those are
beginner bugs though, the reason why they're not fixed is because
they require a relatively large amount of work and a relatively good
understanding of the subject.
For your specific example, no, you didn't identify the buggy
transition correctly, so you're debugging the part that has nothing
to do with the false positive. I strongly recommend finding the
buggy transition in your ExplodedGraph dump *before* jumping into
the debugger. This usually goes from bottom to top, exactly like
you're expected to read a static analyzer report. Let me explain
this procedure on your example.
I attached a screenshot of the tail of the ExplodedGraph dump for
your code (i hope this works fine with the mailing list). That's
just three last nodes of the graph but they're enough to identify
the problem.
Node 53 is the bug node, as stated in red; this is the node that was
produced by generateErrorNode() and fed into the BugReport object.
The bug report says "Branch condition evaluates to a garbage value".
The AST expression p[1] (which acts as the branch condition which we
can confirm by looking at the source code) has symbolic value
"Undefined". Therefore we can conclude that the bug report was
emitted correctly and the our problem is not in the checker callback
that emitted the bug report; it was simply fed an incorrect State
1337. Since the only reason why the bug report was emitted was that
the value was Undefined, we have to find a node in which "Undefined"
was written into the state.
Node 52 doesn't update the state, so we can skip it. In fact it's
nicely grouped with node 51 so we barely even notice it.
Node 51 is the nearest node on which the AST expression p[1] was
first evaluated into Undefined. Whatever code was responsible for
evaluating node 51 is directly responsible for assigning Undefined
value to the expression - which ultimately led to the false positive
in node 53. It is easy to see what was supposed to happen at this
node: this is a memory load (which is also known in C and C++
formalism as "implicit lvalue-to-rvalue conversion" - the operation
that converts object-as-in-"location" into object-as-in-"data"). The
operation does not change contents of memory, it only reads the
existing memory.
Static analyzer represents the knowledge of existing memory in a
data structure known as the Store, aka "Region Store".
Straightforwardly, it maps "memory regions" (symbolic
representations of locations) into arbitrary SVals (symbolic
representations of data). No, Balazs, it doesn't map "values to
expressions" =/ The Store is written down in the dump and labeled
accordingly. In our case there are three entries in the Store:
(1) Variable "myvar" has 16-bit unsigned integer 4386 written into
it at byte offset 0;
(2) Variable "p" has a pointer to the first char (hence index 0) of
variable "myvar" written into it at byte offset 0.
(3) Variable "x" has 32-bit signed integer 1 written into it at byte
offset 0.
The location we're loading from is already evaluated: it's the value
of AST expression p[1] before the load. The Expressions section (aka
the Environment) contains two entries for p[1] because implicit
lvalue-to-rvalue casts do not have any visual representation in C,
however by matching statement identifiers (S780 and S784) with the
respective program point dumps (S780: ArraySubscriptExpr, S784:
ImplicitCastExpr (LValueToRValue)) you can easily see that S780 is
the sub-expression of the load and S784 is the load itself.
Therefore the location which we're loading from is the value of AST
expression p[1], namely the *second* char of variable "myvar".
Knowing these facts about the memory of our program, we need to
decide whether the load was performed correctly. If the load was
performed incorrectly, our job is done: we've found the root cause.
If the load was performed correctly that'd have meant that the state
is incorrect and we have to go further up in order to find it.
It's fairly obvious that the load was not performed correctly.
Looking at Store entry (1), it is easy to see that the second
character of variable "myvar" has value 17 rather than "Undefined".
Therefore our problem consists entirely in incorrect construction of
node 51. Now's the good time to jump into the debugger, set
conditional breakpoint on evalLoad with condition "predecessor node
has identifier 50" and debug from there.
The pointer cast that you were trying to debug was not at fault. We
can still discuss it separately but it's not what really causes this
specific false positive to show up.
This entire debugging procedure is very straightforward and requires
almost no thinking. You simply match static analyzer's simulation of
the program's behavior to the actual runtime behavior of the program
and check whether the former is a correct abstraction over the
latter. Static analyzer is just an interpreter (cf. "abstract
interpretation") so it's very easy to reason about the correctness
of every transition as long as you know the language it's trying to
interpret (in this case, C). Which is why i made this 2018
conference talk in order to explain it. The only non-trivial part is
knowing the structure of the program state (store vs environment,
what are regions and symbols and how to read them) which is why i
did this whole workbook thing on that subject in 2016. I wish i did
it properly so that all the documentation was in one place but for
now that's where we are i guess.
On 12/7/20 7:40 AM, Balázs Benics via
cfe-dev wrote:
Ok, you are saying that we should know that `p[1]` points
to an object which was initialized. And according to the
target's endianness, resolve the DeclRefExpr to the
appropriate constant value.
The dumps are probably from the line `char *p =
&myvar;` if I'm right.
That is the reason why the right-hand side is evaluated to
the location of `myvar`, but wrapped into an ElementRegion of
type char representing the reinterpret cast.
IMO, it's the correct behavior to this point.
You should probably dig around the evaluation of the
`DeclRefExpr` of the expression `p[1]`. That should return
your concrete value.
AFAIK, the store maps the values to expressions and the
problem probably resides there.
But it's just a wide guess :D
Unfortunately, that is all I can say about this right now
:(
Balazs.
Vince Bridgers via cfe-dev
<[hidden email]> ezt
írta (időpont: 2020. dec. 7., H, 11:54):
Hi all, I thought I'd look at this problem https://bugs.llvm.org/show_bug.cgi?id=43364
(Pointer cast representation problems). maybe use it as an
opportunity to dig into the inner workings of the analyzer
and maybe even solve a concrete problem :) But it seems I
need guidance about possible paths to pursue.
Starting with a concrete case, a simple reproducer is
below. I debugged this down to
SimpleSValBuilder.cpp:evalCastFromLoc(). val is an SVal
and castTy is a Loc, and this code path attempts to
extract a concrete integer from the SVal (did I get this
right?). So I think a solution to this specific case would
be to dig into the SVal to see if casted data is concrete,
and extract that data. Seems to me this would be
the location for that (evalCastFromLoc seems
appropropriate enough). If that's true, can someone point
me to an example that's similar? I'll keep digging, but
thought I'd ask in case this is easy for someone to drop a
few helpful hints.
I've included some select dumps, state and a bt below
in case this is helpful.
Best
(gdb) p val.dump()
&Element{myvar,0 S64b,char}$2 = void
Hi Artem, thank you for taking the time to explain the background how to chase this down. I'll carefully read through the materials and references provided and continue to build up knowledge on how to work on this part of the static analyzer.
Best - Vince
On Sun, Dec 13, 2020 at 11:55 PM Artem Dergachev <[hidden email]> wrote:
Ok, first of all, the example false positive you're debugging is not
an example of the bug you're trying to look into, but an example of
a completely different long-standing bug:
https://bugs.llvm.org/show_bug.cgi?id=44114. Both bugzilla entries
contain a large number of examples to choose from. None of those are
beginner bugs though, the reason why they're not fixed is because
they require a relatively large amount of work and a relatively good
understanding of the subject.
For your specific example, no, you didn't identify the buggy
transition correctly, so you're debugging the part that has nothing
to do with the false positive. I strongly recommend finding the
buggy transition in your ExplodedGraph dump *before* jumping into
the debugger. This usually goes from bottom to top, exactly like
you're expected to read a static analyzer report. Let me explain
this procedure on your example.
I attached a screenshot of the tail of the ExplodedGraph dump for
your code (i hope this works fine with the mailing list). That's
just three last nodes of the graph but they're enough to identify
the problem.
Node 53 is the bug node, as stated in red; this is the node that was
produced by generateErrorNode() and fed into the BugReport object.
The bug report says "Branch condition evaluates to a garbage value".
The AST expression p[1] (which acts as the branch condition which we
can confirm by looking at the source code) has symbolic value
"Undefined". Therefore we can conclude that the bug report was
emitted correctly and the our problem is not in the checker callback
that emitted the bug report; it was simply fed an incorrect State
1337. Since the only reason why the bug report was emitted was that
the value was Undefined, we have to find a node in which "Undefined"
was written into the state.
Node 52 doesn't update the state, so we can skip it. In fact it's
nicely grouped with node 51 so we barely even notice it.
Node 51 is the nearest node on which the AST expression p[1] was
first evaluated into Undefined. Whatever code was responsible for
evaluating node 51 is directly responsible for assigning Undefined
value to the expression - which ultimately led to the false positive
in node 53. It is easy to see what was supposed to happen at this
node: this is a memory load (which is also known in C and C++
formalism as "implicit lvalue-to-rvalue conversion" - the operation
that converts object-as-in-"location" into object-as-in-"data"). The
operation does not change contents of memory, it only reads the
existing memory.
Static analyzer represents the knowledge of existing memory in a
data structure known as the Store, aka "Region Store".
Straightforwardly, it maps "memory regions" (symbolic
representations of locations) into arbitrary SVals (symbolic
representations of data). No, Balazs, it doesn't map "values to
expressions" =/ The Store is written down in the dump and labeled
accordingly. In our case there are three entries in the Store:
(1) Variable "myvar" has 16-bit unsigned integer 4386 written into
it at byte offset 0;
(2) Variable "p" has a pointer to the first char (hence index 0) of
variable "myvar" written into it at byte offset 0.
(3) Variable "x" has 32-bit signed integer 1 written into it at byte
offset 0.
The location we're loading from is already evaluated: it's the value
of AST expression p[1] before the load. The Expressions section (aka
the Environment) contains two entries for p[1] because implicit
lvalue-to-rvalue casts do not have any visual representation in C,
however by matching statement identifiers (S780 and S784) with the
respective program point dumps (S780: ArraySubscriptExpr, S784:
ImplicitCastExpr (LValueToRValue)) you can easily see that S780 is
the sub-expression of the load and S784 is the load itself.
Therefore the location which we're loading from is the value of AST
expression p[1], namely the *second* char of variable "myvar".
Knowing these facts about the memory of our program, we need to
decide whether the load was performed correctly. If the load was
performed incorrectly, our job is done: we've found the root cause.
If the load was performed correctly that'd have meant that the state
is incorrect and we have to go further up in order to find it.
It's fairly obvious that the load was not performed correctly.
Looking at Store entry (1), it is easy to see that the second
character of variable "myvar" has value 17 rather than "Undefined".
Therefore our problem consists entirely in incorrect construction of
node 51. Now's the good time to jump into the debugger, set
conditional breakpoint on evalLoad with condition "predecessor node
has identifier 50" and debug from there.
The pointer cast that you were trying to debug was not at fault. We
can still discuss it separately but it's not what really causes this
specific false positive to show up.
This entire debugging procedure is very straightforward and requires
almost no thinking. You simply match static analyzer's simulation of
the program's behavior to the actual runtime behavior of the program
and check whether the former is a correct abstraction over the
latter. Static analyzer is just an interpreter (cf. "abstract
interpretation") so it's very easy to reason about the correctness
of every transition as long as you know the language it's trying to
interpret (in this case, C). Which is why i made this 2018
conference talk in order to explain it. The only non-trivial part is
knowing the structure of the program state (store vs environment,
what are regions and symbols and how to read them) which is why i
did this whole workbook thing on that subject in 2016. I wish i did
it properly so that all the documentation was in one place but for
now that's where we are i guess.
On 12/7/20 7:40 AM, Balázs Benics via
cfe-dev wrote:
Ok, you are saying that we should know that `p[1]` points
to an object which was initialized. And according to the
target's endianness, resolve the DeclRefExpr to the
appropriate constant value.
The dumps are probably from the line `char *p =
&myvar;` if I'm right.
That is the reason why the right-hand side is evaluated to
the location of `myvar`, but wrapped into an ElementRegion of
type char representing the reinterpret cast.
IMO, it's the correct behavior to this point.
You should probably dig around the evaluation of the
`DeclRefExpr` of the expression `p[1]`. That should return
your concrete value.
AFAIK, the store maps the values to expressions and the
problem probably resides there.
But it's just a wide guess :D
Unfortunately, that is all I can say about this right now
:(
Balazs.
Vince Bridgers via cfe-dev
<[hidden email]> ezt
írta (időpont: 2020. dec. 7., H, 11:54):
Hi all, I thought I'd look at this problem https://bugs.llvm.org/show_bug.cgi?id=43364
(Pointer cast representation problems). maybe use it as an
opportunity to dig into the inner workings of the analyzer
and maybe even solve a concrete problem :) But it seems I
need guidance about possible paths to pursue.
Starting with a concrete case, a simple reproducer is
below. I debugged this down to
SimpleSValBuilder.cpp:evalCastFromLoc(). val is an SVal
and castTy is a Loc, and this code path attempts to
extract a concrete integer from the SVal (did I get this
right?). So I think a solution to this specific case would
be to dig into the SVal to see if casted data is concrete,
and extract that data. Seems to me this would be
the location for that (evalCastFromLoc seems
appropropriate enough). If that's true, can someone point
me to an example that's similar? I'll keep digging, but
thought I'd ask in case this is easy for someone to drop a
few helpful hints.
I've included some select dumps, state and a bt below
in case this is helpful.
Best
(gdb) p val.dump()
&Element{myvar,0 S64b,char}$2 = void
Hi Artem, thank you for taking the time to explain
the background how to chase this down. I'll carefully read
through the materials and references provided and continue to
build up knowledge on how to work on this part of the static
analyzer.
Best - Vince
On Sun, Dec 13, 2020 at 11:55
PM Artem Dergachev <[hidden email]> wrote:
Ok, first of all, the example false positive you're
debugging is not an example of the bug you're trying to look
into, but an example of a completely different long-standing
bug: https://bugs.llvm.org/show_bug.cgi?id=44114.
Both bugzilla entries contain a large number of examples to
choose from. None of those are beginner bugs though, the
reason why they're not fixed is because they require a
relatively large amount of work and a relatively good
understanding of the subject.
For your specific example, no, you didn't identify the buggy
transition correctly, so you're debugging the part that has
nothing to do with the false positive. I strongly recommend
finding the buggy transition in your ExplodedGraph dump
*before* jumping into the debugger. This usually goes from
bottom to top, exactly like you're expected to read a static
analyzer report. Let me explain this procedure on your
example.
I attached a screenshot of the tail of the ExplodedGraph
dump for your code (i hope this works fine with the mailing
list).
That's just three last nodes of the graph but they're
enough to identify the problem.
Node 53 is the bug node, as stated in red; this is the node
that was produced by generateErrorNode() and fed into the
BugReport object. The bug report says "Branch condition
evaluates to a garbage value". The AST expression p[1]
(which acts as the branch condition which we can confirm by
looking at the source code) has symbolic value "Undefined".
Therefore we can conclude that the bug report was emitted
correctly and the our problem is not in the checker callback
that emitted the bug report; it was simply fed an incorrect
State 1337. Since the only reason why the bug report was
emitted was that the value was Undefined, we have to find a
node in which "Undefined" was written into the state.
Node 52 doesn't update the state, so we can skip it. In fact
it's nicely grouped with node 51 so we barely even notice
it.
Node 51 is the nearest node on which the AST expression p[1]
was first evaluated into Undefined. Whatever code was
responsible for evaluating node 51 is directly responsible
for assigning Undefined value to the expression - which
ultimately led to the false positive in node 53. It is easy
to see what was supposed to happen at this node: this is a
memory load (which is also known in C and C++ formalism as
"implicit lvalue-to-rvalue conversion" - the operation that
converts object-as-in-"location" into object-as-in-"data").
The operation does not change contents of memory, it only
reads the existing memory.
Static analyzer represents the knowledge of existing memory
in a data structure known as the Store, aka "Region Store".
Straightforwardly, it maps "memory regions" (symbolic
representations of locations) into arbitrary SVals (symbolic
representations of data). No, Balazs, it doesn't map "values
to expressions" =/ The Store is written down in the dump and
labeled accordingly. In our case there are three entries in
the Store:
(1) Variable "myvar" has 16-bit unsigned integer 4386
written into it at byte offset 0;
(2) Variable "p" has a pointer to the first char (hence
index 0) of variable "myvar" written into it at byte offset
0.
(3) Variable "x" has 32-bit signed integer 1 written into it
at byte offset 0.
The location we're loading from is already evaluated: it's
the value of AST expression p[1] before the load. The
Expressions section (aka the Environment) contains two
entries for p[1] because implicit lvalue-to-rvalue casts do
not have any visual representation in C, however by matching
statement identifiers (S780 and S784) with the respective
program point dumps (S780: ArraySubscriptExpr, S784:
ImplicitCastExpr (LValueToRValue)) you can easily see that
S780 is the sub-expression of the load and S784 is the load
itself. Therefore the location which we're loading from is
the value of AST expression p[1], namely the *second* char
of variable "myvar".
Knowing these facts about the memory of our program, we need
to decide whether the load was performed correctly. If the
load was performed incorrectly, our job is done: we've found
the root cause. If the load was performed correctly that'd
have meant that the state is incorrect and we have to go
further up in order to find it.
It's fairly obvious that the load was not performed
correctly. Looking at Store entry (1), it is easy to see
that the second character of variable "myvar" has value 17
rather than "Undefined".
Therefore our problem consists entirely in incorrect
construction of node 51. Now's the good time to jump into
the debugger, set conditional breakpoint on evalLoad with
condition "predecessor node has identifier 50" and debug
from there.
The pointer cast that you were trying to debug was not at
fault. We can still discuss it separately but it's not what
really causes this specific false positive to show up.
This entire debugging procedure is very straightforward and
requires almost no thinking. You simply match static
analyzer's simulation of the program's behavior to the
actual runtime behavior of the program and check whether the
former is a correct abstraction over the latter. Static
analyzer is just an interpreter (cf. "abstract
interpretation") so it's very easy to reason about the
correctness of every transition as long as you know the
language it's trying to interpret (in this case, C). Which
is why i made this 2018 conference talk in order to explain
it.
The only non-trivial part is knowing the structure of
the program state (store vs environment, what are regions
and symbols and how to read them) which is why i did this
whole workbook thing on that subject in 2016. I wish i did
it properly so that all the documentation was in one place
but for now that's where we are i guess.
On 12/7/20 7:40 AM, Balázs Benics via cfe-dev wrote:
Ok, you are saying that we should know that `p[1]`
points to an object which was initialized. And
according to the target's endianness, resolve the
DeclRefExpr to the appropriate constant value.
The dumps are probably from the line `char *p =
&myvar;` if I'm right.
That is the reason why the right-hand side is
evaluated to the location of `myvar`, but wrapped into
an ElementRegion of type char representing the
reinterpret cast.
IMO, it's the correct behavior to this point.
You should probably dig around the evaluation of
the `DeclRefExpr` of the expression `p[1]`. That
should return your concrete value.
AFAIK, the store maps the values to expressions and
the problem probably resides there.
But it's just a wide guess :D
Unfortunately, that is all I can say about this
right now :(
Balazs.
Vince Bridgers via
cfe-dev <[hidden email]>
ezt írta (időpont: 2020. dec. 7., H, 11:54):
Hi all, I thought I'd look at this
problem https://bugs.llvm.org/show_bug.cgi?id=43364
(Pointer cast representation problems). maybe use it
as an opportunity to dig into the inner workings of
the analyzer and maybe even solve a concrete problem
:) But it seems I need guidance about possible
paths to pursue.
Starting with a concrete case, a simple
reproducer is below. I debugged this down to
SimpleSValBuilder.cpp:evalCastFromLoc(). val is an
SVal and castTy is a Loc, and this code path
attempts to extract a concrete integer from the
SVal (did I get this right?). So I think a
solution to this specific case would be to dig
into the SVal to see if casted data is concrete,
and extract that data. Seems to me this would be
the location for that (evalCastFromLoc seems
appropropriate enough). If that's true, can
someone point me to an example that's similar?
I'll keep digging, but thought I'd ask in case
this is easy for someone to drop a few helpful
hints.
I've included some select dumps, state and a bt
below in case this is helpful.
Best
(gdb) p val.dump()
&Element{myvar,0 S64b,char}$2 = void
Locked
screenshot.png (273K)