BugReporter is used to process ALL bug reports. By using a shared map,
we are having mappings from different PathDiagnosticPieces to LocationContexts
well beyond the point where we are processing a given report. This
state is inherently error prone, and is analogous to using a global
variable. Instead, just create a temporary map, one per report,
and when we are done with it we throw it away. No extra state.
llvm-svn: 180974
Much of this patch outside of PathDiagnostics.h are just minor
syntactic changes due to the return type for operator* and the like
changing for the iterator, so the real focus should be on
PathPieces itself.
This change is motivated so that we can do efficient insertion
and removal of individual pieces from within a PathPiece, just like
this was a kind of "IR" for static analyzer diagnostics. We
currently implement path transformations by iterating over an
entire PathPiece and making a copy. This isn't very natural for
some algorithms.
We use an ilist here instead of std::list because we want operations
to rip out/insert nodes in place, just like IR manipulation. This
isn't being used yet, but opens the door for more powerful
transformation algorithms on diagnostic paths.
llvm-svn: 180741
The 2 functions were computing the same location using different logic (each one had edge case bugs that the other
one did not). Refactor them to rely on the same logic.
The location of the warning reported in text/command line output format will now match that of the plist file.
There is one change in the plist output as well. When reporting an error on a BinaryOperator, we use the location of the
operator instead of the beginning of the BinaryOperator expression. This matches our output on command line and
looks better in most cases.
llvm-svn: 180165
In this code
int getZero() {
return 0;
}
void test() {
int problem = 1 / getZero(); // expected-warning {{Division by zero}}
}
we generate these arrows:
+-----------------+
| v
int problem = 1 / getZero();
^ |
+---+
where the top one represents the control flow up to the first call, and the
bottom one represents the flow to the division.* It turns out, however, that
we were generating the top arrow twice, as if attempting to "set up context"
after we had already returned from the call. This resulted in poor
highlighting in Xcode.
* Arguably the best location for the division is the '/', but that's a
different problem.
<rdar://problem/13326040>
llvm-svn: 179350
This is important because sometimes two nodes are identical, except the
second one is a sink.
This bug has probably been around for a while, but it wouldn't have been an
issue in the old report graph algorithm. I'm ashamed to say I actually looked
at this the first time around and thought it would never be a problem...and
then didn't include an assertion to back that up.
PR15684
llvm-svn: 178944
The algorithm used here was ridiculously slow when a potential back-edge
pointed to a node that already had a lot of successors. The previous commit
makes this feature unnecessary anyway.
This reverts r177468 / f4cf6b10f863b9bc716a09b2b2a8c497dcc6aa9b.
Conflicts:
lib/StaticAnalyzer/Core/BugReporter.cpp
llvm-svn: 177765
For a given bug equivalence class, we'd like to emit the report with the
shortest path. So far to do this we've been trimming the ExplodedGraph to
only contain relevant nodes, then doing a reverse BFS (starting at all the
error nodes) to find the shortest paths from the root. However, this is
fairly expensive when we are suppressing many bug reports in the same
equivalence class.
r177468-9 tried to solve this problem by breaking cycles during graph
trimming, then updating the BFS priorities after each suppressed report
instead of recomputing the whole thing. However, breaking cycles is not
a cheap operation because an analysis graph minus cycles is still a DAG,
not a tree.
This fix changes the algorithm to do a single forward BFS (starting from the
root) and to use that to choose the report with the shortest path by looking
at the error nodes with the lowest BFS priorities. This was Anna's idea, and
has the added advantage of requiring no update step: we can just pick the
error node with the next lowest priority to produce the next bug report.
<rdar://problem/13474689>
llvm-svn: 177764
With the assurance that the trimmed graph does not contain cycles,
this patch is safe (with a few tweaks), and provides the performance
boost it was intended to.
Part of performance work for <rdar://problem/13433687>.
llvm-svn: 177469
Having a trimmed graph with no cycles (a DAG) is much more convenient for
trying to find shortest paths, which is exactly what BugReporter needs to do.
Part of the performance work for <rdar://problem/13433687>.
llvm-svn: 177468
The whole reason we were doing a BFS in the first place is because an
ExplodedGraph can have cycles. Unfortunately, my removeErrorNode "update"
doesn't work at all if there are cycles.
I'd still like to be able to avoid doing the BFS every time, but I'll come
back to it later.
This reverts r177353 / 481fa5071c203bc8ba4f88d929780f8d0f8837ba.
llvm-svn: 177448
Splitting the graph trimming and the path-finding (r177216) already
recovered quite a bit of performance lost to increased suppression.
We can still do better by also performing the reverse BFS up front
(needed for shortest-path-finding) and only walking the shortest path
for each report. This does mean we have to walk back up the path and
invalidate all the BFS numbers if the report turns out to be invalid,
but it's probably still faster than redoing the full BFS every time.
More performance work for <rdar://problem/13433687>
llvm-svn: 177353
When we generate a path diagnostic for a bug report, we have to take the
full ExplodedGraph and limit it down to a single path. We do this in two
steps: "trimming", which limits the graph to all the paths that lead to
this particular bug, and "creating the report graph", which finds the
shortest path in the trimmed path to any error node.
With BugReporterVisitor false positive suppression, this becomes more
expensive: it's possible for some paths through the trimmed graph to be
invalid (i.e. likely false positives) but others to be valid. Therefore
we have to run the visitors over each path in the graph until we find one
that is valid, or until we've ruled them all out. This can become quite
expensive.
This commit separates out graph trimming from creating the report graph,
performing the first only once per bug equivalence class and the second
once per bug report. It also cleans up that portion of the code by
introducing some wrapper classes.
This seems to recover most of the performance regression described in my
last commit.
<rdar://problem/13433687>
llvm-svn: 177216
...in favor of this typedef:
typedef llvm::DenseMap<const ExplodedNode *, const ExplodedNode *>
InterExplodedGraphMap;
Use this everywhere the previous class and typedef were used.
Took the opportunity to ArrayRef-ize ExplodedGraph::trim while I'm at it.
No functionality change.
llvm-svn: 177215
I removed this check in the recursion->iteration commit, but forgot that
generatePathDiagnostic may be called multiple times if there are multiple
PathDiagnosticConsumers.
llvm-svn: 177214
The previous generatePathDiagnostic() was intended to be tail-recursive,
restarting and trying again if a report was marked invalid. However:
(1) this leaked all the cloned visitors, which weren't being deleted, and
(2) this wasn't actually tail-recursive because some local variables had
non-trivial destructors.
This was causing us to overflow the stack on inputs with large numbers of
reports in the same equivalence class, such as sqlite3.c. Being iterative
at least prevents us from blowing out the stack, but doesn't solve the
performance issue: suppressing thousands (yes, thousands) of paths in the
same equivalence class is expensive. I'm looking into that now.
<rdar://problem/13423498>
llvm-svn: 177189
We discovered that sqlite3.c currently has 2600 reports in a single
equivalence class; it would be good to know if this is a recent
development or what.
(For the curious, the different reports in an equivalence class represent
the same bug found along different paths. When we're suppressing false
positives, we need to go through /every/ path to make sure there isn't a
valid path to a bug. This is a flaw in our after-the-fact suppression,
made worse by the fact that that function isn't particularly optimized.)
llvm-svn: 177188
Use Optional<CFG*> where invalid states were needed previously. In the one case
where that's not possible (beginAutomaticObjDtorsInsert) just use a dummy
CFGAutomaticObjDtor.
Thanks for the help from Jordan Rose & discussion/feedback from Ted Kremenek
and Doug Gregor.
Post commit code review feedback on r175796 by Ted Kremenek.
llvm-svn: 175938
Suppress the warning by just not emitting the report. The sink node
would get generated, which is fine since we did reach a bad state.
Motivation
Due to the way code is structured in some of these macros, we do not
reason correctly about it and report false positives. Specifically, the
following loop reports a use-after-free. Because of the way the code is
structured inside of the macro, the analyzer assumes that the list can
have cycles, so you end up with use-after-free in the loop, that is
safely deleting elements of the list. (The user does not have a way to
teach the analyzer about shape of data structures.)
SLIST_FOREACH_SAFE(item, &ctx->example_list, example_le, tmpitem) {
if (item->index == 3) { // if you remove each time, no complaints
assert((&ctx->example_list)->slh_first == item);
SLIST_REMOVE(&ctx->example_list, item, example_s, example_le);
free(item);
}
}
llvm-svn: 172883
The issue here is that if we have 2 leaks reported at the same line for
which we cannot print the corresponding region info, they will get
treated as the same by issue_hash+description. We need to AUGMENT the
issue_hash with the allocation info to differentiate the two issues.
Add the "hash" (offset from the beginning of a function) representing
allocation site to solve the issue.
We might want to generalize solution in the future when we decide to
track more than just the 2 locations from the diagnostics.
llvm-svn: 171825
This is the case where the analyzer tries to print out source locations
for code within a synthesized function body, which of course does not have
a valid source location. The previous fix attempted to do this during
diagnostic path pruning, but some diagnostics have pruning disabled, and
so any diagnostic with a path that goes through a synthesized body will
either hit an assertion or emit invalid output.
<rdar://problem/12657843> (again)
llvm-svn: 169631
uncovered.
This required manually correcting all of the incorrect main-module
headers I could find, and running the new llvm/utils/sort_includes.py
script over the files.
I also manually added quite a few missing headers that were uncovered by
shuffling the order or moving headers up to be main-module-headers.
llvm-svn: 169237
In code like this:
void foo() {
bar();
baz();
}
...the location for the call to 'bar()' was being used as a backup location
for the call to 'baz()'. This is fine unless the call to 'bar()' is deemed
uninteresting and that part of the path deleted.
(This looks like a logic error as well, but in practice the only way 'baz()'
could have an invalid location is if the entire body of 'foo()' is
synthesized, meaning the call to 'bar()' will be using the location of the
call to 'foo()' anyway. Nevertheless, the new version better matches the
intent of the code.)
Found by Matt Beaumont-Gay using ASan. Thanks, Matt!
llvm-svn: 168080
This fixes a few cases where we'd emit path notes like this:
+---+
1| v
p = malloc(len);
^ |2
+---+
In general this should make path notes more consistent and more correct,
especially in cases where the leak happens on the false branch of an if
that jumps directly to the end of the function. There are a couple places
where the leak is reported farther away from the cause; these are usually
cases where there are several levels of nested braces before the end of
the function. This still matches our current behavior for when there /is/
a statement after all the braces, though.
llvm-svn: 168070
We do this by using the "most recent" good location: if a synthesized
function 'A' calls another function 'B', the path notes for the call to 'B'
will be placed at the same location as the path note for calling 'A'.
Similarly, the call to 'A' will have a note saying "Entered call from...",
and now we just don't emit that (since the user doesn't have a body to look
at anyway).
Previously, we were doing this for the "Calling..." notes, but not for the
"Entered call from..." or "Returning to caller". This caused a crash when
the path entered and then exiting a call within a synthesized body.
<rdar://problem/12657843>
llvm-svn: 168019
path notes for cases where a value may be assumed to be null, etc.
Instead of having redundant diagnostics, do a pass over the generated
PathDiagnostic pieces and remove notes from TrackConstraintBRVisitor
that are already covered by ConditionBRVisitor, whose notes tend
to be better.
Fixes <rdar://problem/12252783>
llvm-svn: 166728
Before, PathDiagnosticConsumers that did not support actual path output
would (sensibly) cause the generation of the full path to be skipped.
However, BugReporterVisitors may want to see the path in order to mark a
BugReport as invalid.
Now, even for a path generation scheme of 'None' we will still create a
trimmed graph and walk backwards through the bug path, doing no work other
than passing the nodes to the BugReporterVisitors. This isn't cheap, but
it's necessary to properly do suppression when the first path consumer does
not support path notes.
In the future, we should try only generating the path and visitor-provided
path notes once, or at least only creating the trimmed graph once.
llvm-svn: 164447
This is intended to allow visitors to make decisions about whether a
BugReport is likely a false positive. Currently there are no visitors
making use of this feature, so there are no tests.
When a BugReport is marked invalid, the invalidator must provide a key
that identifies the invaliation (intended to be the visitor type and a
context pointer of some kind). This allows us to reverse the decision
later on. Being able to reverse a decision about invalidation gives us more
flexibility, and allows us to formulate conditions like "this report is
invalid UNLESS the original argument is 'foo'". We can use this to
fine-tune our false-positive suppression (coming soon).
llvm-svn: 164446
their implementations are unavailable. Start by simulating dispatch_sync().
This change is largely a bunch of plumbing around something very simple. We
use AnalysisDeclContext to conjure up a fake function body (using the
current ASTContext) when one does not exist. This is controlled
under the analyzer-config option "faux-bodies", which is off by default.
The plumbing in this patch is largely to pass the necessary machinery
around. CallEvent needs the AnalysisDeclContextManager to get
the function definition, as one may get conjured up lazily.
BugReporter and PathDiagnosticLocation needed to be relaxed to handle
invalid locations, as the conjured body has no real source locations.
We do some primitive recovery in diagnostic generation to generate
some reasonable locations (for arrows and events), but it can be
improved.
llvm-svn: 164339
Again, GCC is more aggressive about reusing temporary space than we are,
leading to Release build crashes for this undefined behavior.
PR13710 (though it may not be the only problem there)
llvm-svn: 163747
